virtualenv/FFmpeg
1
0
Fork 0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2024-12-12 19:18:44 +02:00
Code Issues Releases Activity
4487dd8bb5
FFmpeg/libavcodec/arm/vp9itxfm_neon.S

1689 lines
61 KiB
ArmAsm
Raw Normal View History

arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
/*
* Copyright (c) 2016 Google Inc.
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "libavutil/arm/asm.S"
#include "neon.S"
const itxfm4_coeffs, align=4
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
.short 11585, 0, 6270, 15137
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
iadst4_coeffs:
.short 5283, 15212, 9929, 13377
endconst
const iadst8_coeffs, align=4
.short 16305, 1606, 14449, 7723, 10394, 12665, 4756, 15679
idct_coeffs:
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
.short 11585, 0, 6270, 15137, 3196, 16069, 13623, 9102
.short 1606, 16305, 12665, 10394, 7723, 14449, 15679, 4756
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.short 804, 16364, 12140, 11003, 7005, 14811, 15426, 5520
.short 3981, 15893, 14053, 8423, 9760, 13160, 16207, 2404
endconst
const iadst16_coeffs, align=4
arm: vp9itxfm: Reorder iadst16 coeffs This matches the order they are in the 16 bpp version. There they are in this order, to make sure we access them in the same order they are declared, easing loading only half of the coefficients at a time. This makes the 8 bpp version match the 16 bpp version better. This is cherrypicked from libav commit 08074c092d8c97d71c5986e5325e97ffc956119d. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 22:27:13 +02:00
.short 16364, 804, 15893, 3981, 11003, 12140, 8423, 14053
.short 14811, 7005, 13160, 9760, 5520, 15426, 2404, 16207
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
endconst
@ Do four 4x4 transposes, using q registers for the subtransposes that don't
@ need to address the individual d registers.
@ r0,r1 == rq1, r2,r3 == rq1, etc
.macro transpose16_q_4x_4x4 rq0, rq1, rq2, rq3, rq4, rq5, rq6, rq7, r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15
vtrn.32 \rq0, \rq1
vtrn.32 \rq2, \rq3
vtrn.32 \rq4, \rq5
vtrn.32 \rq6, \rq7
vtrn.16 \r0, \r1
vtrn.16 \r2, \r3
vtrn.16 \r4, \r5
vtrn.16 \r6, \r7
vtrn.16 \r8, \r9
vtrn.16 \r10, \r11
vtrn.16 \r12, \r13
vtrn.16 \r14, \r15
.endm
@ out1 = ((in1 + in2) * d0[0] + (1 << 13)) >> 14
@ out2 = ((in1 - in2) * d0[0] + (1 << 13)) >> 14
@ in/out are d registers
.macro mbutterfly0 out1, out2, in1, in2, tmpd1, tmpd2, tmpq3, tmpq4, neg=0
vadd.s16 \tmpd1, \in1, \in2
vsub.s16 \tmpd2, \in1, \in2
vmull.s16 \tmpq3, \tmpd1, d0[0]
vmull.s16 \tmpq4, \tmpd2, d0[0]
.if \neg > 0
vneg.s32 \tmpq3, \tmpq3
.endif
vrshrn.s32 \out1, \tmpq3, #14
vrshrn.s32 \out2, \tmpq4, #14
.endm
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
@ Same as mbutterfly0 above, but treating the input in in2 as zero,
@ writing the same output into both out1 and out2.
.macro mbutterfly0_h out1, out2, in1, in2, tmpd1, tmpd2, tmpq3, tmpq4
vmull.s16 \tmpq3, \in1, d0[0]
vrshrn.s32 \out1, \tmpq3, #14
vrshrn.s32 \out2, \tmpq3, #14
.endm
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
@ out1,out2 = ((in1 + in2) * d0[0] + (1 << 13)) >> 14
@ out3,out4 = ((in1 - in2) * d0[0] + (1 << 13)) >> 14
@ Same as mbutterfly0, but with input being 2 q registers, output
@ being 4 d registers.
@ This can do with either 4 or 6 temporary q registers.
.macro dmbutterfly0 out1, out2, out3, out4, in1, in2, tmpq1, tmpq2, tmpd11, tmpd12, tmpd21, tmpd22, tmpq3, tmpq4, tmpq5, tmpq6
vadd.s16 \tmpq1, \in1, \in2
vsub.s16 \tmpq2, \in1, \in2
vmull.s16 \tmpq3, \tmpd11, d0[0]
vmull.s16 \tmpq4, \tmpd12, d0[0]
.ifb \tmpq5
vrshrn.s32 \out1, \tmpq3, #14
vrshrn.s32 \out2, \tmpq4, #14
vmull.s16 \tmpq3, \tmpd21, d0[0]
vmull.s16 \tmpq4, \tmpd22, d0[0]
vrshrn.s32 \out3, \tmpq3, #14
vrshrn.s32 \out4, \tmpq4, #14
.else
vmull.s16 \tmpq5, \tmpd21, d0[0]
vmull.s16 \tmpq6, \tmpd22, d0[0]
vrshrn.s32 \out1, \tmpq3, #14
vrshrn.s32 \out2, \tmpq4, #14
vrshrn.s32 \out3, \tmpq5, #14
vrshrn.s32 \out4, \tmpq6, #14
.endif
.endm
@ out1 = in1 * coef1 - in2 * coef2
@ out2 = in1 * coef2 + in2 * coef1
@ out are 2 q registers, in are 2 d registers
.macro mbutterfly_l out1, out2, in1, in2, coef1, coef2
vmull.s16 \out1, \in1, \coef1
vmlsl.s16 \out1, \in2, \coef2
vmull.s16 \out2, \in1, \coef2
vmlal.s16 \out2, \in2, \coef1
.endm
@ out1,out2 = in1,in2 * coef1 - in3,in4 * coef2
@ out3,out4 = in1,in2 * coef2 + in3,in4 * coef1
@ out are 4 q registers, in are 4 d registers
.macro dmbutterfly_l out1, out2, out3, out4, in1, in2, in3, in4, coef1, coef2
vmull.s16 \out1, \in1, \coef1
vmull.s16 \out2, \in2, \coef1
vmull.s16 \out3, \in1, \coef2
vmull.s16 \out4, \in2, \coef2
vmlsl.s16 \out1, \in3, \coef2
vmlsl.s16 \out2, \in4, \coef2
vmlal.s16 \out3, \in3, \coef1
vmlal.s16 \out4, \in4, \coef1
.endm
arm: vp9itxfm: Rename a macro parameter to fit better Since the same parameter is used for both input and output, the name inout is more fitting. This matches the naming used below in the dmbutterfly macro. This is cherrypicked from libav commit 79566ec8c77969d5f9be533de04b1349834cca62. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:12 +02:00
@ inout1 = (inout1 * coef1 - inout2 * coef2 + (1 << 13)) >> 14
@ inout2 = (inout1 * coef2 + inout2 * coef1 + (1 << 13)) >> 14
@ inout are 2 d registers, tmp are 2 q registers
.macro mbutterfly inout1, inout2, coef1, coef2, tmp1, tmp2, neg=0
mbutterfly_l \tmp1, \tmp2, \inout1, \inout2, \coef1, \coef2
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.if \neg > 0
vneg.s32 \tmp2, \tmp2
.endif
arm: vp9itxfm: Rename a macro parameter to fit better Since the same parameter is used for both input and output, the name inout is more fitting. This matches the naming used below in the dmbutterfly macro. This is cherrypicked from libav commit 79566ec8c77969d5f9be533de04b1349834cca62. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:12 +02:00
vrshrn.s32 \inout1, \tmp1, #14
vrshrn.s32 \inout2, \tmp2, #14
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.endm
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
@ Same as mbutterfly above, but treating the input in inout2 as zero
.macro mbutterfly_h1 inout1, inout2, coef1, coef2, tmp1, tmp2
vmull.s16 \tmp1, \inout1, \coef1
vmull.s16 \tmp2, \inout1, \coef2
vrshrn.s32 \inout1, \tmp1, #14
vrshrn.s32 \inout2, \tmp2, #14
.endm
@ Same as mbutterfly above, but treating the input in inout1 as zero
.macro mbutterfly_h2 inout1, inout2, coef1, coef2, tmp1, tmp2
vmull.s16 \tmp1, \inout2, \coef2
vmull.s16 \tmp2, \inout2, \coef1
vneg.s32 \tmp1, \tmp1
vrshrn.s32 \inout2, \tmp2, #14
vrshrn.s32 \inout1, \tmp1, #14
.endm
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
@ inout1,inout2 = (inout1,inout2 * coef1 - inout3,inout4 * coef2 + (1 << 13)) >> 14
@ inout3,inout4 = (inout1,inout2 * coef2 + inout3,inout4 * coef1 + (1 << 13)) >> 14
@ inout are 4 d registers, tmp are 4 q registers
.macro dmbutterfly inout1, inout2, inout3, inout4, coef1, coef2, tmp1, tmp2, tmp3, tmp4
dmbutterfly_l \tmp1, \tmp2, \tmp3, \tmp4, \inout1, \inout2, \inout3, \inout4, \coef1, \coef2
vrshrn.s32 \inout1, \tmp1, #14
vrshrn.s32 \inout2, \tmp2, #14
vrshrn.s32 \inout3, \tmp3, #14
vrshrn.s32 \inout4, \tmp4, #14
.endm
@ out1 = in1 + in2
@ out2 = in1 - in2
.macro butterfly out1, out2, in1, in2
vadd.s16 \out1, \in1, \in2
vsub.s16 \out2, \in1, \in2
.endm
@ out1 = in1 - in2
@ out2 = in1 + in2
.macro butterfly_r out1, out2, in1, in2
vsub.s16 \out1, \in1, \in2
vadd.s16 \out2, \in1, \in2
.endm
@ out1 = (in1 + in2 + (1 << 13)) >> 14
@ out2 = (in1 - in2 + (1 << 13)) >> 14
@ out are 2 d registers, in are 2 q registers, tmp are 2 q registers
.macro butterfly_n out1, out2, in1, in2, tmp1, tmp2
vadd.s32 \tmp1, \in1, \in2
vsub.s32 \tmp2, \in1, \in2
vrshrn.s32 \out1, \tmp1, #14
vrshrn.s32 \out2, \tmp2, #14
.endm
@ out1,out2 = (in1,in2 + in3,in4 + (1 << 13)) >> 14
@ out3,out4 = (in1,in2 - in3,in4 + (1 << 13)) >> 14
@ out are 4 d registers, in are 4 q registers, tmp are 4 q registers
.macro dbutterfly_n out1, out2, out3, out4, in1, in2, in3, in4, tmp1, tmp2, tmp3, tmp4
vadd.s32 \tmp1, \in1, \in3
vadd.s32 \tmp2, \in2, \in4
vsub.s32 \tmp3, \in1, \in3
vsub.s32 \tmp4, \in2, \in4
vrshrn.s32 \out1, \tmp1, #14
vrshrn.s32 \out2, \tmp2, #14
vrshrn.s32 \out3, \tmp3, #14
vrshrn.s32 \out4, \tmp4, #14
.endm
.macro iwht4 c0, c1, c2, c3
vadd.i16 \c0, \c0, \c1
vsub.i16 d17, \c2, \c3
vsub.i16 d16, \c0, d17
vshr.s16 d16, d16, #1
vsub.i16 \c2, d16, \c1
vsub.i16 \c1, d16, \c3
vadd.i16 \c3, d17, \c2
vsub.i16 \c0, \c0, \c1
.endm
.macro idct4 c0, c1, c2, c3
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
vmull.s16 q13, \c1, d0[3]
vmull.s16 q11, \c1, d0[2]
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
vadd.i16 d16, \c0, \c2
vsub.i16 d17, \c0, \c2
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
vmlal.s16 q13, \c3, d0[2]
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
vmull.s16 q9, d16, d0[0]
vmull.s16 q10, d17, d0[0]
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
vmlsl.s16 q11, \c3, d0[3]
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
vrshrn.s32 d26, q13, #14
vrshrn.s32 d18, q9, #14
vrshrn.s32 d20, q10, #14
vrshrn.s32 d22, q11, #14
vadd.i16 \c0, d18, d26
vsub.i16 \c3, d18, d26
vadd.i16 \c1, d20, d22
vsub.i16 \c2, d20, d22
.endm
.macro iadst4 c0, c1, c2, c3
vmull.s16 q10, \c0, d1[0]
vmlal.s16 q10, \c2, d1[1]
vmlal.s16 q10, \c3, d1[2]
vmull.s16 q11, \c0, d1[2]
vmlsl.s16 q11, \c2, d1[0]
vsub.s16 \c0, \c0, \c2
vmlsl.s16 q11, \c3, d1[1]
vadd.s16 \c0, \c0, \c3
vmull.s16 q13, \c1, d1[3]
vmull.s16 q12, \c0, d1[3]
vadd.s32 q14, q10, q13
vadd.s32 q1, q11, q13
vrshrn.s32 \c0, q14, #14
vadd.s32 q10, q10, q11
vrshrn.s32 \c1, q1, #14
vsub.s32 q10, q10, q13
vrshrn.s32 \c2, q12, #14
vrshrn.s32 \c3, q10, #14
.endm
@ The public functions in this file have got the following signature:
@ void itxfm_add(uint8_t *dst, ptrdiff_t stride, int16_t *block, int eob);
.macro itxfm_func4x4 txfm1, txfm2
function ff_vp9_\txfm1\()_\txfm2\()_4x4_add_neon, export=1
.ifc \txfm1,\txfm2
.ifc \txfm1,idct
movrel r12, itxfm4_coeffs
vld1.16 {d0}, [r12,:64]
.endif
.ifc \txfm1,iadst
movrel r12, iadst4_coeffs
vld1.16 {d1}, [r12,:64]
.endif
.else
movrel r12, itxfm4_coeffs
vld1.16 {q0}, [r12,:128]
.endif
vmov.i16 q15, #0
.ifc \txfm1\()_\txfm2,idct_idct
cmp r3, #1
bne 1f
@ DC-only for idct/idct
vld1.16 {d4[]}, [r2,:16]
vmull.s16 q2, d4, d0[0]
vrshrn.s32 d4, q2, #14
vmull.s16 q2, d4, d0[0]
vrshrn.s32 d4, q2, #14
vst1.16 {d30[0]}, [r2,:16]
vdup.16 q2, d4[0]
vmov q3, q2
b 2f
.endif
1:
vld1.16 {d4-d7}, [r2,:128]
vst1.16 {q15}, [r2,:128]!
.ifc \txfm1,iwht
vshr.s16 q2, q2, #2
vshr.s16 q3, q3, #2
.endif
\txfm1\()4 d4, d5, d6, d7
vst1.16 {q15}, [r2,:128]!
@ Transpose 4x4 with 16 bit elements
vtrn.16 d4, d5
vtrn.16 d6, d7
vtrn.32 q2, q3
\txfm2\()4 d4, d5, d6, d7
2:
vld1.32 {d0[]}, [r0,:32], r1
vld1.32 {d0[1]}, [r0,:32], r1
.ifnc \txfm1,iwht
vrshr.s16 q2, q2, #4
vrshr.s16 q3, q3, #4
.endif
vaddw.u8 q2, q2, d0
vld1.32 {d1[]}, [r0,:32], r1
vld1.32 {d1[1]}, [r0,:32], r1
vqmovun.s16 d0, q2
sub r0, r0, r1, lsl #2
vaddw.u8 q3, q3, d1
vst1.32 {d0[0]}, [r0,:32], r1
vqmovun.s16 d1, q3
vst1.32 {d0[1]}, [r0,:32], r1
vst1.32 {d1[0]}, [r0,:32], r1
vst1.32 {d1[1]}, [r0,:32], r1
bx lr
endfunc
.endm
itxfm_func4x4 idct, idct
itxfm_func4x4 iadst, idct
itxfm_func4x4 idct, iadst
itxfm_func4x4 iadst, iadst
itxfm_func4x4 iwht, iwht
.macro idct8
dmbutterfly0 d16, d17, d24, d25, q8, q12, q2, q4, d4, d5, d8, d9, q3, q2, q5, q4 @ q8 = t0a, q12 = t1a
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
dmbutterfly d20, d21, d28, d29, d0[2], d0[3], q2, q3, q4, q5 @ q10 = t2a, q14 = t3a
dmbutterfly d18, d19, d30, d31, d1[0], d1[1], q2, q3, q4, q5 @ q9 = t4a, q15 = t7a
dmbutterfly d26, d27, d22, d23, d1[2], d1[3], q2, q3, q4, q5 @ q13 = t5a, q11 = t6a
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
butterfly q2, q14, q8, q14 @ q2 = t0, q14 = t3
butterfly q3, q10, q12, q10 @ q3 = t1, q10 = t2
butterfly q4, q13, q9, q13 @ q4 = t4, q13 = t5a
butterfly q5, q11, q15, q11 @ q5 = t7, q11 = t6a
butterfly q8, q15, q2, q5 @ q8 = out[0], q15 = out[7]
dmbutterfly0 d4, d5, d10, d11, q11, q13, q9, q13, d18, d19, d26, d27, q2, q5, q11, q12 @ q2 = t6, q5 = t5
butterfly q11, q12, q14, q4 @ q11 = out[3], q12 = out[4]
butterfly q9, q14, q3, q2 @ q9 = out[1], q14 = out[6]
butterfly_r q13, q10, q10, q5 @ q13 = out[5], q10 = out[2]
.endm
.macro iadst8
dmbutterfly_l q4, q5, q2, q3, d30, d31, d16, d17, d2[1], d2[0] @ q4,q5 = t1a, q2,q3 = t0a
dmbutterfly_l q8, q15, q6, q7, d22, d23, d24, d25, d3[1], d3[0] @ q8,q15 = t5a, q6,q7 = t4a
dbutterfly_n d22, d23, d4, d5, q2, q3, q6, q7, q11, q12, q2, q3 @ q11 = t0, q2 = t4
dbutterfly_n d24, d25, d6, d7, q4, q5, q8, q15, q12, q3, q6, q7 @ q12 = t1, q3 = t5
dmbutterfly_l q6, q7, q4, q5, d26, d27, d20, d21, d2[3], d2[2] @ q6,q7 = t3a, q4,q5 = t2a
dmbutterfly_l q10, q13, q8, q15, d18, d19, d28, d29, d3[3], d3[2] @ q10,q13 = t7a, q8,q15 = t6a
dbutterfly_n d18, d19, d8, d9, q4, q5, q8, q15, q9, q14, q4, q5 @ q9 = t2, q4 = t6
dbutterfly_n d16, d17, d12, d13, q6, q7, q10, q13, q8, q15, q6, q7 @ q8 = t3, q6 = t7
butterfly q15, q12, q12, q8 @ q15 = -out[7], q12 = t3
vneg.s16 q15, q15 @ q15 = out[7]
butterfly q8, q9, q11, q9 @ q8 = out[0], q9 = t2
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
dmbutterfly_l q10, q11, q5, q7, d4, d5, d6, d7, d0[2], d0[3] @ q10,q11 = t5a, q5,q7 = t4a
dmbutterfly_l q2, q3, q13, q14, d12, d13, d8, d9, d0[3], d0[2] @ q2,q3 = t6a, q13,q14 = t7a
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
dbutterfly_n d28, d29, d8, d9, q10, q11, q13, q14, q4, q6, q10, q11 @ q14 = out[6], q4 = t7
dmbutterfly0 d22, d23, d24, d25, q9, q12, q6, q13, d12, d13, d26, d27, q9, q10 @ q11 = -out[3], q12 = out[4]
vneg.s16 q11, q11 @ q11 = out[3]
dbutterfly_n d18, d19, d4, d5, q5, q7, q2, q3, q9, q10, q2, q3 @ q9 = -out[1], q2 = t6
vneg.s16 q9, q9 @ q9 = out[1]
dmbutterfly0 d20, d21, d26, d27, q2, q4, q3, q5, d6, d7, d10, d11, q6, q7 @ q10 = out[2], q13 = -out[5]
vneg.s16 q13, q13 @ q13 = out[5]
.endm
.macro itxfm_func8x8 txfm1, txfm2
function ff_vp9_\txfm1\()_\txfm2\()_8x8_add_neon, export=1
@ Push q4-q7 if iadst is used, idct requires
@ a few scratch registers less, so only push q4-q5
@ if only idct is involved.
@ The iadst also uses a few coefficients from
@ idct, so those always need to be loaded.
.ifc \txfm1\()_\txfm2,idct_idct
movrel r12, idct_coeffs
vpush {q4-q5}
.else
movrel r12, iadst8_coeffs
vld1.16 {q1}, [r12,:128]!
vpush {q4-q7}
.endif
arm: vp9itxfm: Share instructions for loading idct coeffs in the 8x8 function This is cherrypicked from libav commit 3933b86bb93aca47f29fbd493075b0f110c1e3f5. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-03 16:38:56 +02:00
vld1.16 {q0}, [r12,:128]
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
vmov.i16 q2, #0
vmov.i16 q3, #0
.ifc \txfm1\()_\txfm2,idct_idct
cmp r3, #1
bne 1f
@ DC-only for idct/idct
vld1.16 {d16[]}, [r2,:16]
vmull.s16 q8, d16, d0[0]
vrshrn.s32 d16, q8, #14
vmull.s16 q8, d16, d0[0]
vrshrn.s32 d16, q8, #14
vdup.16 q8, d16[0]
vmov q9, q8
vmov q10, q8
vmov q11, q8
vmov q12, q8
vmov q13, q8
vmov q14, q8
vmov q15, q8
vst1.16 {d4[0]}, [r2,:16]
b 2f
.endif
1:
vld1.16 {q8-q9}, [r2,:128]!
vld1.16 {q10-q11}, [r2,:128]!
vld1.16 {q12-q13}, [r2,:128]!
vld1.16 {q14-q15}, [r2,:128]!
sub r2, r2, #128
vst1.16 {q2-q3}, [r2,:128]!
vst1.16 {q2-q3}, [r2,:128]!
vst1.16 {q2-q3}, [r2,:128]!
vst1.16 {q2-q3}, [r2,:128]!
\txfm1\()8
@ Transpose 8x8 with 16 bit elements
vswp d17, d24
vswp d19, d26
vswp d21, d28
vswp d23, d30
transpose16_4x4 q8, q9, q10, q11, q12, q13, q14, q15
\txfm2\()8
2:
mov r3, r0
@ Add into the destination
vld1.8 {d4}, [r0,:64], r1
vrshr.s16 q8, q8, #5
vld1.8 {d5}, [r0,:64], r1
vrshr.s16 q9, q9, #5
vld1.8 {d6}, [r0,:64], r1
vrshr.s16 q10, q10, #5
vaddw.u8 q8, q8, d4
vld1.8 {d7}, [r0,:64], r1
vrshr.s16 q11, q11, #5
vaddw.u8 q9, q9, d5
vld1.8 {d8}, [r0,:64], r1
vrshr.s16 q12, q12, #5
vaddw.u8 q10, q10, d6
vqmovun.s16 d4, q8
vld1.8 {d9}, [r0,:64], r1
vrshr.s16 q13, q13, #5
vaddw.u8 q11, q11, d7
vqmovun.s16 d5, q9
vld1.8 {d10}, [r0,:64], r1
vrshr.s16 q14, q14, #5
vaddw.u8 q12, q12, d8
vqmovun.s16 d6, q10
vld1.8 {d11}, [r0,:64], r1
vrshr.s16 q15, q15, #5
vaddw.u8 q13, q13, d9
vqmovun.s16 d7, q11
vst1.8 {d4}, [r3,:64], r1
vaddw.u8 q14, q14, d10
vst1.8 {d5}, [r3,:64], r1
vqmovun.s16 d8, q12
vst1.8 {d6}, [r3,:64], r1
vaddw.u8 q15, q15, d11
vst1.8 {d7}, [r3,:64], r1
vqmovun.s16 d9, q13
vst1.8 {d8}, [r3,:64], r1
vqmovun.s16 d10, q14
vst1.8 {d9}, [r3,:64], r1
vqmovun.s16 d11, q15
vst1.8 {d10}, [r3,:64], r1
vst1.8 {d11}, [r3,:64], r1
.ifc \txfm1\()_\txfm2,idct_idct
vpop {q4-q5}
.else
vpop {q4-q7}
.endif
bx lr
endfunc
.endm
itxfm_func8x8 idct, idct
itxfm_func8x8 iadst, idct
.ltorg
itxfm_func8x8 idct, iadst
itxfm_func8x8 iadst, iadst
function idct16x16_dc_add_neon
movrel r12, idct_coeffs
vld1.16 {d0}, [r12,:64]
arm/aarch64: vp9: Fix vertical alignment Align the second/third operands as they usually are. Due to the wildly varying sizes of the written out operands in aarch64 assembly, the column alignment is usually not as clear as in arm assembly. This is cherrypicked from libav commit 7995ebfad12002033c73feed422a1cfc62081e8f. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-09 00:04:19 +02:00
vmov.i16 q2, #0
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
vld1.16 {d16[]}, [r2,:16]
vmull.s16 q8, d16, d0[0]
vrshrn.s32 d16, q8, #14
vmull.s16 q8, d16, d0[0]
vrshrn.s32 d16, q8, #14
vdup.16 q8, d16[0]
vst1.16 {d4[0]}, [r2,:16]
vrshr.s16 q8, q8, #6
arm: vp9itxfm: Optimize 16x16 and 32x32 idct dc by unrolling This work is sponsored by, and copyright, Google. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_32x32_sub1_add_neon: 752.0 459.2 862.2 553.9 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 226.5 145.0 225.1 171.8 vp9_inv_dct_dct_32x32_sub1_add_neon: 721.2 415.7 727.6 475.0 This is cherrypicked from libav commit a76bf8cf1277ef6feb1580b578f5e6ca327e713c. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-04 13:08:51 +02:00
mov r3, r0
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
mov r12, #16
1:
@ Loop to add the constant from q8 into all 16x16 outputs
arm: vp9itxfm: Optimize 16x16 and 32x32 idct dc by unrolling This work is sponsored by, and copyright, Google. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_32x32_sub1_add_neon: 752.0 459.2 862.2 553.9 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 226.5 145.0 225.1 171.8 vp9_inv_dct_dct_32x32_sub1_add_neon: 721.2 415.7 727.6 475.0 This is cherrypicked from libav commit a76bf8cf1277ef6feb1580b578f5e6ca327e713c. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-04 13:08:51 +02:00
subs r12, r12, #2
vld1.8 {q2}, [r0,:128], r1
vaddw.u8 q10, q8, d4
vld1.8 {q3}, [r0,:128], r1
vaddw.u8 q11, q8, d5
vaddw.u8 q12, q8, d6
vaddw.u8 q13, q8, d7
vqmovun.s16 d4, q10
vqmovun.s16 d5, q11
vqmovun.s16 d6, q12
vst1.8 {q2}, [r3,:128], r1
vqmovun.s16 d7, q13
vst1.8 {q3}, [r3,:128], r1
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
bne 1b
bx lr
endfunc
.ltorg
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.macro idct16_end
butterfly d18, d7, d4, d7 @ d18 = t0a, d7 = t7a
butterfly d19, d22, d5, d22 @ d19 = t1a, d22 = t6
butterfly d4, d26, d20, d26 @ d4 = t2a, d26 = t5
butterfly d5, d6, d28, d6 @ d5 = t3a, d6 = t4
butterfly d20, d28, d16, d24 @ d20 = t8a, d28 = t11a
butterfly d24, d21, d23, d21 @ d24 = t9, d21 = t10
butterfly d23, d27, d25, d27 @ d23 = t14, d27 = t13
butterfly d25, d29, d29, d17 @ d25 = t15a, d29 = t12a
mbutterfly0 d27, d21, d27, d21, d16, d30, q8, q15 @ d27 = t13a, d21 = t10a
mbutterfly0 d29, d28, d29, d28, d16, d30, q8, q15 @ d29 = t12, d28 = t11
vswp d27, d29 @ d27 = t12, d29 = t13a
vswp d28, d27 @ d28 = t12, d27 = t11
butterfly d16, d31, d18, d25 @ d16 = out[0], d31 = out[15]
butterfly d17, d30, d19, d23 @ d17 = out[1], d30 = out[14]
butterfly_r d25, d22, d22, d24 @ d25 = out[9], d22 = out[6]
butterfly d23, d24, d7, d20 @ d23 = out[7], d24 = out[8]
butterfly d18, d29, d4, d29 @ d18 = out[2], d29 = out[13]
butterfly d19, d28, d5, d28 @ d19 = out[3], d28 = out[12]
vmov d4, d21 @ d4 = t10a
butterfly d20, d27, d6, d27 @ d20 = out[4], d27 = out[11]
butterfly d21, d26, d26, d4 @ d21 = out[5], d26 = out[10]
bx lr
.endm
arm: vp9itxfm: Make the larger core transforms standalone functions This work is sponsored by, and copyright, Google. This reduces the code size of libavcodec/arm/vp9itxfm_neon.o from 15324 to 12388 bytes. This gives a small slowdown of a couple tens of cycles, up to around 150 cycles for the full case of the largest transform, but makes it more feasible to add more optimized versions of these transforms. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub4_add_neon: 2063.4 1516.0 1719.5 1245.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3279.3 2454.5 2525.2 1982.3 vp9_inv_dct_dct_32x32_sub4_add_neon: 10750.0 7955.4 8525.6 6754.2 vp9_inv_dct_dct_32x32_sub32_add_neon: 18574.0 17108.4 14216.7 12010.2 After: vp9_inv_dct_dct_16x16_sub4_add_neon: 2060.8 1608.5 1735.7 1262.0 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.2 2443.5 2546.1 1999.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10682.0 8043.8 8581.3 6810.1 vp9_inv_dct_dct_32x32_sub32_add_neon: 18522.4 17277.4 14286.7 12087.9 This is cherrypicked from libav commit 0331c3f5e8cb6e6b53fab7893e91d1be1bfa979c. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-23 10:56:12 +02:00
function idct16
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
mbutterfly0 d16, d24, d16, d24, d4, d6, q2, q3 @ d16 = t0a, d24 = t1a
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
mbutterfly d20, d28, d0[2], d0[3], q2, q3 @ d20 = t2a, d28 = t3a
mbutterfly d18, d30, d1[0], d1[1], q2, q3 @ d18 = t4a, d30 = t7a
mbutterfly d26, d22, d1[2], d1[3], q2, q3 @ d26 = t5a, d22 = t6a
mbutterfly d17, d31, d2[0], d2[1], q2, q3 @ d17 = t8a, d31 = t15a
mbutterfly d25, d23, d2[2], d2[3], q2, q3 @ d25 = t9a, d23 = t14a
mbutterfly d21, d27, d3[0], d3[1], q2, q3 @ d21 = t10a, d27 = t13a
mbutterfly d29, d19, d3[2], d3[3], q2, q3 @ d29 = t11a, d19 = t12a
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
butterfly d4, d28, d16, d28 @ d4 = t0, d28 = t3
butterfly d5, d20, d24, d20 @ d5 = t1, d20 = t2
butterfly d6, d26, d18, d26 @ d6 = t4, d26 = t5
butterfly d7, d22, d30, d22 @ d7 = t7, d22 = t6
butterfly d16, d25, d17, d25 @ d16 = t8, d25 = t9
butterfly d24, d21, d29, d21 @ d24 = t11, d21 = t10
butterfly d17, d27, d19, d27 @ d17 = t12, d27 = t13
butterfly d29, d23, d31, d23 @ d29 = t15, d23 = t14
mbutterfly0 d22, d26, d22, d26, d18, d30, q9, q15 @ d22 = t6a, d26 = t5a
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
mbutterfly d23, d25, d0[2], d0[3], q9, q15 @ d23 = t9a, d25 = t14a
mbutterfly d27, d21, d0[2], d0[3], q9, q15, neg=1 @ d27 = t13a, d21 = t10a
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
idct16_end
endfunc
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
function idct16_half
mbutterfly0_h d16, d24, d16, d24, d4, d6, q2, q3 @ d16 = t0a, d24 = t1a
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
mbutterfly_h1 d20, d28, d0[2], d0[3], q2, q3 @ d20 = t2a, d28 = t3a
mbutterfly_h1 d18, d30, d1[0], d1[1], q2, q3 @ d18 = t4a, d30 = t7a
mbutterfly_h2 d26, d22, d1[2], d1[3], q2, q3 @ d26 = t5a, d22 = t6a
mbutterfly_h1 d17, d31, d2[0], d2[1], q2, q3 @ d17 = t8a, d31 = t15a
mbutterfly_h2 d25, d23, d2[2], d2[3], q2, q3 @ d25 = t9a, d23 = t14a
mbutterfly_h1 d21, d27, d3[0], d3[1], q2, q3 @ d21 = t10a, d27 = t13a
mbutterfly_h2 d29, d19, d3[2], d3[3], q2, q3 @ d29 = t11a, d19 = t12a
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
butterfly d4, d28, d16, d28 @ d4 = t0, d28 = t3
butterfly d5, d20, d24, d20 @ d5 = t1, d20 = t2
butterfly d6, d26, d18, d26 @ d6 = t4, d26 = t5
butterfly d7, d22, d30, d22 @ d7 = t7, d22 = t6
butterfly d16, d25, d17, d25 @ d16 = t8, d25 = t9
butterfly d24, d21, d29, d21 @ d24 = t11, d21 = t10
butterfly d17, d27, d19, d27 @ d17 = t12, d27 = t13
butterfly d29, d23, d31, d23 @ d29 = t15, d23 = t14
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
mbutterfly0 d22, d26, d22, d26, d18, d30, q9, q15 @ d22 = t6a, d26 = t5a
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
mbutterfly d23, d25, d0[2], d0[3], q9, q15 @ d23 = t9a, d25 = t14a
mbutterfly d27, d21, d0[2], d0[3], q9, q15, neg=1 @ d27 = t13a, d21 = t10a
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
idct16_end
endfunc
function idct16_quarter
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
vmull.s16 q12, d19, d3[3]
vmull.s16 q2, d17, d2[0]
vmull.s16 q3, d18, d1[1]
vmull.s16 q15, d18, d1[0]
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
vneg.s32 q12, q12
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
vmull.s16 q14, d17, d2[1]
vmull.s16 q13, d19, d3[2]
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
vmull.s16 q11, d16, d0[0]
vrshrn.s32 d24, q12, #14
vrshrn.s32 d16, q2, #14
vrshrn.s32 d7, q3, #14
vrshrn.s32 d6, q15, #14
vrshrn.s32 d29, q14, #14
vrshrn.s32 d17, q13, #14
vrshrn.s32 d28, q11, #14
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
mbutterfly_l q10, q11, d17, d24, d0[2], d0[3]
mbutterfly_l q9, q15, d29, d16, d0[2], d0[3]
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
vneg.s32 q11, q11
vrshrn.s32 d27, q10, #14
vrshrn.s32 d21, q11, #14
vrshrn.s32 d23, q9, #14
vrshrn.s32 d25, q15, #14
vmov d4, d28
vmov d5, d28
mbutterfly0 d22, d26, d7, d6, d18, d30, q9, q15
vmov d20, d28
idct16_end
arm: vp9itxfm: Make the larger core transforms standalone functions This work is sponsored by, and copyright, Google. This reduces the code size of libavcodec/arm/vp9itxfm_neon.o from 15324 to 12388 bytes. This gives a small slowdown of a couple tens of cycles, up to around 150 cycles for the full case of the largest transform, but makes it more feasible to add more optimized versions of these transforms. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub4_add_neon: 2063.4 1516.0 1719.5 1245.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3279.3 2454.5 2525.2 1982.3 vp9_inv_dct_dct_32x32_sub4_add_neon: 10750.0 7955.4 8525.6 6754.2 vp9_inv_dct_dct_32x32_sub32_add_neon: 18574.0 17108.4 14216.7 12010.2 After: vp9_inv_dct_dct_16x16_sub4_add_neon: 2060.8 1608.5 1735.7 1262.0 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.2 2443.5 2546.1 1999.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10682.0 8043.8 8581.3 6810.1 vp9_inv_dct_dct_32x32_sub32_add_neon: 18522.4 17277.4 14286.7 12087.9 This is cherrypicked from libav commit 0331c3f5e8cb6e6b53fab7893e91d1be1bfa979c. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-23 10:56:12 +02:00
endfunc
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
arm: vp9itxfm: Make the larger core transforms standalone functions This work is sponsored by, and copyright, Google. This reduces the code size of libavcodec/arm/vp9itxfm_neon.o from 15324 to 12388 bytes. This gives a small slowdown of a couple tens of cycles, up to around 150 cycles for the full case of the largest transform, but makes it more feasible to add more optimized versions of these transforms. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub4_add_neon: 2063.4 1516.0 1719.5 1245.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3279.3 2454.5 2525.2 1982.3 vp9_inv_dct_dct_32x32_sub4_add_neon: 10750.0 7955.4 8525.6 6754.2 vp9_inv_dct_dct_32x32_sub32_add_neon: 18574.0 17108.4 14216.7 12010.2 After: vp9_inv_dct_dct_16x16_sub4_add_neon: 2060.8 1608.5 1735.7 1262.0 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.2 2443.5 2546.1 1999.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10682.0 8043.8 8581.3 6810.1 vp9_inv_dct_dct_32x32_sub32_add_neon: 18522.4 17277.4 14286.7 12087.9 This is cherrypicked from libav commit 0331c3f5e8cb6e6b53fab7893e91d1be1bfa979c. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-23 10:56:12 +02:00
function iadst16
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
movrel r12, iadst16_coeffs
vld1.16 {q0-q1}, [r12,:128]
mbutterfly_l q3, q2, d31, d16, d0[1], d0[0] @ q3 = t1, q2 = t0
arm: vp9itxfm: Reorder iadst16 coeffs This matches the order they are in the 16 bpp version. There they are in this order, to make sure we access them in the same order they are declared, easing loading only half of the coefficients at a time. This makes the 8 bpp version match the 16 bpp version better. This is cherrypicked from libav commit 08074c092d8c97d71c5986e5325e97ffc956119d. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 22:27:13 +02:00
mbutterfly_l q5, q4, d23, d24, d1[1], d1[0] @ q5 = t9, q4 = t8
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
butterfly_n d31, d24, q3, q5, q6, q5 @ d31 = t1a, d24 = t9a
mbutterfly_l q7, q6, d29, d18, d0[3], d0[2] @ q7 = t3, q6 = t2
butterfly_n d16, d23, q2, q4, q3, q4 @ d16 = t0a, d23 = t8a
arm: vp9itxfm: Reorder iadst16 coeffs This matches the order they are in the 16 bpp version. There they are in this order, to make sure we access them in the same order they are declared, easing loading only half of the coefficients at a time. This makes the 8 bpp version match the 16 bpp version better. This is cherrypicked from libav commit 08074c092d8c97d71c5986e5325e97ffc956119d. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 22:27:13 +02:00
mbutterfly_l q3, q2, d21, d26, d1[3], d1[2] @ q3 = t11, q2 = t10
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
butterfly_n d29, d26, q7, q3, q4, q3 @ d29 = t3a, d26 = t11a
arm: vp9itxfm: Reorder iadst16 coeffs This matches the order they are in the 16 bpp version. There they are in this order, to make sure we access them in the same order they are declared, easing loading only half of the coefficients at a time. This makes the 8 bpp version match the 16 bpp version better. This is cherrypicked from libav commit 08074c092d8c97d71c5986e5325e97ffc956119d. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 22:27:13 +02:00
mbutterfly_l q5, q4, d27, d20, d2[1], d2[0] @ q5 = t5, q4 = t4
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
butterfly_n d18, d21, q6, q2, q3, q2 @ d18 = t2a, d21 = t10a
mbutterfly_l q7, q6, d19, d28, d3[1], d3[0] @ q7 = t13, q6 = t12
butterfly_n d20, d28, q5, q7, q2, q7 @ d20 = t5a, d28 = t13a
arm: vp9itxfm: Reorder iadst16 coeffs This matches the order they are in the 16 bpp version. There they are in this order, to make sure we access them in the same order they are declared, easing loading only half of the coefficients at a time. This makes the 8 bpp version match the 16 bpp version better. This is cherrypicked from libav commit 08074c092d8c97d71c5986e5325e97ffc956119d. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 22:27:13 +02:00
mbutterfly_l q3, q2, d25, d22, d2[3], d2[2] @ q3 = t7, q2 = t6
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
butterfly_n d27, d19, q4, q6, q5, q6 @ d27 = t4a, d19 = t12a
mbutterfly_l q5, q4, d17, d30, d3[3], d3[2] @ q5 = t15, q4 = t14
movrel r12, idct_coeffs
vld1.16 {q0}, [r12,:128]
butterfly_n d22, d30, q3, q5, q6, q5 @ d22 = t7a, d30 = t15a
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
mbutterfly_l q7, q6, d23, d24, d1[0], d1[1] @ q7 = t9, q6 = t8
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
butterfly_n d25, d17, q2, q4, q3, q4 @ d25 = t6a, d17 = t14a
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
mbutterfly_l q2, q3, d28, d19, d1[1], d1[0] @ q2 = t12, q3 = t13
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
butterfly_n d23, d19, q6, q2, q4, q2 @ d23 = t8a, d19 = t12a
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
mbutterfly_l q5, q4, d21, d26, d1[2], d1[3] @ q5 = t11, q4 = t10
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
butterfly_r d4, d27, d16, d27 @ d4 = t4, d27 = t0
butterfly_n d24, d28, q7, q3, q6, q3 @ d24 = t9a, d28 = t13a
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
mbutterfly_l q6, q7, d30, d17, d1[3], d1[2] @ q6 = t14, q7 = t15
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
butterfly_r d5, d20, d31, d20 @ d5 = t5, d20 = t1
butterfly_n d21, d17, q4, q6, q3, q6 @ d21 = t10a, d17 = t14a
butterfly_n d26, d30, q5, q7, q4, q7 @ d26 = t11a, d30 = t15a
butterfly_r d6, d25, d18, d25 @ d6 = t6, d25 = t2
butterfly_r d7, d22, d29, d22 @ d7 = t7, d22 = t3
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
mbutterfly_l q5, q4, d19, d28, d0[2], d0[3] @ q5 = t13, q4 = t12
mbutterfly_l q6, q7, d30, d17, d0[3], d0[2] @ q6 = t14, q7 = t15
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
butterfly_n d18, d30, q4, q6, q8, q6 @ d18 = out[2], d30 = t14a
butterfly_n d29, d17, q5, q7, q6, q7 @ d29 = -out[13], d17 = t15a
vneg.s16 d29, d29 @ d29 = out[13]
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
mbutterfly_l q5, q4, d4, d5, d0[2], d0[3] @ q5 = t5a, q4 = t4a
mbutterfly_l q6, q7, d7, d6, d0[3], d0[2] @ q6 = t6a, q7 = t7a
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
butterfly d2, d6, d27, d25 @ d2 = out[0], d6 = t2a
butterfly d3, d7, d23, d21 @ d3 =-out[1], d7 = t10
butterfly_n d19, d31, q4, q6, q2, q4 @ d19 = -out[3], d31 = t6
vneg.s16 d19, d19 @ d19 = out[3]
butterfly_n d28, d16, q5, q7, q2, q5 @ d28 = out[12], d16 = t7
butterfly d5, d8, d20, d22 @ d5 =-out[15],d8 = t3a
butterfly d4, d9, d24, d26 @ d4 = out[14],d9 = t11
mbutterfly0 d23, d24, d6, d8, d10, d11, q6, q7, 1 @ d23 = out[7], d24 = out[8]
mbutterfly0 d20, d27, d16, d31, d10, d11, q6, q7 @ d20 = out[4], d27 = out[11]
mbutterfly0 d22, d25, d9, d7, d10, d11, q6, q7 @ d22 = out[6], d25 = out[9]
mbutterfly0 d21, d26, d30, d17, d10, d11, q6, q7, 1 @ d21 = out[5], d26 = out[10]
vneg.s16 d31, d5 @ d31 = out[15]
vneg.s16 d17, d3 @ d17 = out[1]
vmov d16, d2
vmov d30, d4
arm: vp9itxfm: Make the larger core transforms standalone functions This work is sponsored by, and copyright, Google. This reduces the code size of libavcodec/arm/vp9itxfm_neon.o from 15324 to 12388 bytes. This gives a small slowdown of a couple tens of cycles, up to around 150 cycles for the full case of the largest transform, but makes it more feasible to add more optimized versions of these transforms. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub4_add_neon: 2063.4 1516.0 1719.5 1245.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3279.3 2454.5 2525.2 1982.3 vp9_inv_dct_dct_32x32_sub4_add_neon: 10750.0 7955.4 8525.6 6754.2 vp9_inv_dct_dct_32x32_sub32_add_neon: 18574.0 17108.4 14216.7 12010.2 After: vp9_inv_dct_dct_16x16_sub4_add_neon: 2060.8 1608.5 1735.7 1262.0 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.2 2443.5 2546.1 1999.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10682.0 8043.8 8581.3 6810.1 vp9_inv_dct_dct_32x32_sub32_add_neon: 18522.4 17277.4 14286.7 12087.9 This is cherrypicked from libav commit 0331c3f5e8cb6e6b53fab7893e91d1be1bfa979c. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-23 10:56:12 +02:00
bx lr
endfunc
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
arm: vp9itxfm: Move the load_add_store macro out from the itxfm16 pass2 function This allows reusing the macro for a separate implementation of the pass2 function. This is cherrypicked from libav commit 47b3c2c18d1897f3c753ba0cec4b2d7aa24526af. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-05 22:55:20 +02:00
.macro load_add_store coef0, coef1, coef2, coef3
vrshr.s16 \coef0, \coef0, #6
vrshr.s16 \coef1, \coef1, #6
vld1.32 {d4[]}, [r0,:32], r1
vld1.32 {d4[1]}, [r3,:32], r1
vrshr.s16 \coef2, \coef2, #6
vrshr.s16 \coef3, \coef3, #6
vld1.32 {d5[]}, [r0,:32], r1
vld1.32 {d5[1]}, [r3,:32], r1
vaddw.u8 \coef0, \coef0, d4
vld1.32 {d6[]}, [r0,:32], r1
vld1.32 {d6[1]}, [r3,:32], r1
vaddw.u8 \coef1, \coef1, d5
vld1.32 {d7[]}, [r0,:32], r1
vld1.32 {d7[1]}, [r3,:32], r1
vqmovun.s16 d4, \coef0
vqmovun.s16 d5, \coef1
sub r0, r0, r1, lsl #2
sub r3, r3, r1, lsl #2
vaddw.u8 \coef2, \coef2, d6
vaddw.u8 \coef3, \coef3, d7
vst1.32 {d4[0]}, [r0,:32], r1
vst1.32 {d4[1]}, [r3,:32], r1
vqmovun.s16 d6, \coef2
vst1.32 {d5[0]}, [r0,:32], r1
vst1.32 {d5[1]}, [r3,:32], r1
vqmovun.s16 d7, \coef3
vst1.32 {d6[0]}, [r0,:32], r1
vst1.32 {d6[1]}, [r3,:32], r1
vst1.32 {d7[0]}, [r0,:32], r1
vst1.32 {d7[1]}, [r3,:32], r1
.endm
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.macro itxfm16_1d_funcs txfm
@ Read a vertical 4x16 slice out of a 16x16 matrix, do a transform on it,
@ transpose into a horizontal 16x4 slice and store.
@ r0 = dst (temp buffer)
arm: vp9itxfm: Skip empty slices in the first pass of idct_idct 16x16 and 32x32 This work is sponsored by, and copyright, Google. Previously all subpartitions except the eob=1 (DC) case ran with the same runtime: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub16_add_neon: 3188.1 2435.4 2499.0 1969.0 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.7 16582.3 14207.6 12000.3 By skipping individual 4x16 or 4x32 pixel slices in the first pass, we reduce the runtime of these functions like this: vp9_inv_dct_dct_16x16_sub1_add_neon: 274.6 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2064.0 1534.8 1719.4 1248.7 vp9_inv_dct_dct_16x16_sub4_add_neon: 2135.0 1477.2 1736.3 1249.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2446.7 1828.7 1993.6 1494.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2832.4 2118.3 2266.5 1735.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.7 2475.3 2523.5 1983.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 756.2 456.7 862.0 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10682.2 8190.4 8539.2 6762.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10813.5 8014.9 8518.3 6762.8 vp9_inv_dct_dct_32x32_sub8_add_neon: 11859.6 9313.0 9347.4 7514.5 vp9_inv_dct_dct_32x32_sub12_add_neon: 12946.6 10752.4 10192.2 8280.2 vp9_inv_dct_dct_32x32_sub16_add_neon: 14074.6 11946.5 11001.4 9008.6 vp9_inv_dct_dct_32x32_sub20_add_neon: 15269.9 13662.7 11816.1 9762.6 vp9_inv_dct_dct_32x32_sub24_add_neon: 16327.9 14940.1 12626.7 10516.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17462.7 15776.1 13446.2 11264.7 vp9_inv_dct_dct_32x32_sub32_add_neon: 18575.5 17157.0 14249.3 12015.1 I.e. in general a very minor overhead for the full subpartition case due to the additional loads and cmps, but a significant speedup for the cases when we only need to process a small part of the actual input data. In common VP9 content in a few inspected clips, 70-90% of the non-dc-only 16x16 and 32x32 IDCTs only have nonzero coefficients in the upper left 8x8 or 16x16 subpartitions respectively. This is cherrypicked from libav commit 9c8bc74c2b40537b0997f646c87c008042d788c2. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:15 +02:00
@ r1 = slice offset
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
@ r2 = src
function \txfm\()16_1d_4x16_pass1_neon
arm: vp9itxfm: Make the larger core transforms standalone functions This work is sponsored by, and copyright, Google. This reduces the code size of libavcodec/arm/vp9itxfm_neon.o from 15324 to 12388 bytes. This gives a small slowdown of a couple tens of cycles, up to around 150 cycles for the full case of the largest transform, but makes it more feasible to add more optimized versions of these transforms. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub4_add_neon: 2063.4 1516.0 1719.5 1245.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3279.3 2454.5 2525.2 1982.3 vp9_inv_dct_dct_32x32_sub4_add_neon: 10750.0 7955.4 8525.6 6754.2 vp9_inv_dct_dct_32x32_sub32_add_neon: 18574.0 17108.4 14216.7 12010.2 After: vp9_inv_dct_dct_16x16_sub4_add_neon: 2060.8 1608.5 1735.7 1262.0 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.2 2443.5 2546.1 1999.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10682.0 8043.8 8581.3 6810.1 vp9_inv_dct_dct_32x32_sub32_add_neon: 18522.4 17277.4 14286.7 12087.9 This is cherrypicked from libav commit 0331c3f5e8cb6e6b53fab7893e91d1be1bfa979c. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-23 10:56:12 +02:00
push {lr}
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
mov r12, #32
arm/aarch64: vp9: Fix vertical alignment Align the second/third operands as they usually are. Due to the wildly varying sizes of the written out operands in aarch64 assembly, the column alignment is usually not as clear as in arm assembly. This is cherrypicked from libav commit 7995ebfad12002033c73feed422a1cfc62081e8f. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-09 00:04:19 +02:00
vmov.s16 q2, #0
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
vld1.16 {d\i}, [r2,:64]
vst1.16 {d4}, [r2,:64], r12
.endr
arm: vp9itxfm: Make the larger core transforms standalone functions This work is sponsored by, and copyright, Google. This reduces the code size of libavcodec/arm/vp9itxfm_neon.o from 15324 to 12388 bytes. This gives a small slowdown of a couple tens of cycles, up to around 150 cycles for the full case of the largest transform, but makes it more feasible to add more optimized versions of these transforms. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub4_add_neon: 2063.4 1516.0 1719.5 1245.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3279.3 2454.5 2525.2 1982.3 vp9_inv_dct_dct_32x32_sub4_add_neon: 10750.0 7955.4 8525.6 6754.2 vp9_inv_dct_dct_32x32_sub32_add_neon: 18574.0 17108.4 14216.7 12010.2 After: vp9_inv_dct_dct_16x16_sub4_add_neon: 2060.8 1608.5 1735.7 1262.0 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.2 2443.5 2546.1 1999.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10682.0 8043.8 8581.3 6810.1 vp9_inv_dct_dct_32x32_sub32_add_neon: 18522.4 17277.4 14286.7 12087.9 This is cherrypicked from libav commit 0331c3f5e8cb6e6b53fab7893e91d1be1bfa979c. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-23 10:56:12 +02:00
bl \txfm\()16
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
@ Do four 4x4 transposes. Originally, d16-d31 contain the
@ 16 rows. Afterwards, d16-d19, d20-d23, d24-d27, d28-d31
@ contain the transposed 4x4 blocks.
transpose16_q_4x_4x4 q8, q9, q10, q11, q12, q13, q14, q15, d16, d17, d18, d19, d20, d21, d22, d23, d24, d25, d26, d27, d28, d29, d30, d31
@ Store the transposed 4x4 blocks horizontally.
arm: vp9itxfm: Skip empty slices in the first pass of idct_idct 16x16 and 32x32 This work is sponsored by, and copyright, Google. Previously all subpartitions except the eob=1 (DC) case ran with the same runtime: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub16_add_neon: 3188.1 2435.4 2499.0 1969.0 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.7 16582.3 14207.6 12000.3 By skipping individual 4x16 or 4x32 pixel slices in the first pass, we reduce the runtime of these functions like this: vp9_inv_dct_dct_16x16_sub1_add_neon: 274.6 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2064.0 1534.8 1719.4 1248.7 vp9_inv_dct_dct_16x16_sub4_add_neon: 2135.0 1477.2 1736.3 1249.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2446.7 1828.7 1993.6 1494.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2832.4 2118.3 2266.5 1735.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.7 2475.3 2523.5 1983.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 756.2 456.7 862.0 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10682.2 8190.4 8539.2 6762.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10813.5 8014.9 8518.3 6762.8 vp9_inv_dct_dct_32x32_sub8_add_neon: 11859.6 9313.0 9347.4 7514.5 vp9_inv_dct_dct_32x32_sub12_add_neon: 12946.6 10752.4 10192.2 8280.2 vp9_inv_dct_dct_32x32_sub16_add_neon: 14074.6 11946.5 11001.4 9008.6 vp9_inv_dct_dct_32x32_sub20_add_neon: 15269.9 13662.7 11816.1 9762.6 vp9_inv_dct_dct_32x32_sub24_add_neon: 16327.9 14940.1 12626.7 10516.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17462.7 15776.1 13446.2 11264.7 vp9_inv_dct_dct_32x32_sub32_add_neon: 18575.5 17157.0 14249.3 12015.1 I.e. in general a very minor overhead for the full subpartition case due to the additional loads and cmps, but a significant speedup for the cases when we only need to process a small part of the actual input data. In common VP9 content in a few inspected clips, 70-90% of the non-dc-only 16x16 and 32x32 IDCTs only have nonzero coefficients in the upper left 8x8 or 16x16 subpartitions respectively. This is cherrypicked from libav commit 9c8bc74c2b40537b0997f646c87c008042d788c2. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:15 +02:00
cmp r1, #12
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
beq 1f
.irp i, 16, 20, 24, 28, 17, 21, 25, 29, 18, 22, 26, 30, 19, 23, 27, 31
vst1.16 {d\i}, [r0,:64]!
.endr
arm: vp9itxfm: Make the larger core transforms standalone functions This work is sponsored by, and copyright, Google. This reduces the code size of libavcodec/arm/vp9itxfm_neon.o from 15324 to 12388 bytes. This gives a small slowdown of a couple tens of cycles, up to around 150 cycles for the full case of the largest transform, but makes it more feasible to add more optimized versions of these transforms. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub4_add_neon: 2063.4 1516.0 1719.5 1245.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3279.3 2454.5 2525.2 1982.3 vp9_inv_dct_dct_32x32_sub4_add_neon: 10750.0 7955.4 8525.6 6754.2 vp9_inv_dct_dct_32x32_sub32_add_neon: 18574.0 17108.4 14216.7 12010.2 After: vp9_inv_dct_dct_16x16_sub4_add_neon: 2060.8 1608.5 1735.7 1262.0 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.2 2443.5 2546.1 1999.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10682.0 8043.8 8581.3 6810.1 vp9_inv_dct_dct_32x32_sub32_add_neon: 18522.4 17277.4 14286.7 12087.9 This is cherrypicked from libav commit 0331c3f5e8cb6e6b53fab7893e91d1be1bfa979c. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-23 10:56:12 +02:00
pop {pc}
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
1:
arm: vp9itxfm: Skip empty slices in the first pass of idct_idct 16x16 and 32x32 This work is sponsored by, and copyright, Google. Previously all subpartitions except the eob=1 (DC) case ran with the same runtime: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub16_add_neon: 3188.1 2435.4 2499.0 1969.0 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.7 16582.3 14207.6 12000.3 By skipping individual 4x16 or 4x32 pixel slices in the first pass, we reduce the runtime of these functions like this: vp9_inv_dct_dct_16x16_sub1_add_neon: 274.6 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2064.0 1534.8 1719.4 1248.7 vp9_inv_dct_dct_16x16_sub4_add_neon: 2135.0 1477.2 1736.3 1249.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2446.7 1828.7 1993.6 1494.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2832.4 2118.3 2266.5 1735.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.7 2475.3 2523.5 1983.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 756.2 456.7 862.0 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10682.2 8190.4 8539.2 6762.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10813.5 8014.9 8518.3 6762.8 vp9_inv_dct_dct_32x32_sub8_add_neon: 11859.6 9313.0 9347.4 7514.5 vp9_inv_dct_dct_32x32_sub12_add_neon: 12946.6 10752.4 10192.2 8280.2 vp9_inv_dct_dct_32x32_sub16_add_neon: 14074.6 11946.5 11001.4 9008.6 vp9_inv_dct_dct_32x32_sub20_add_neon: 15269.9 13662.7 11816.1 9762.6 vp9_inv_dct_dct_32x32_sub24_add_neon: 16327.9 14940.1 12626.7 10516.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17462.7 15776.1 13446.2 11264.7 vp9_inv_dct_dct_32x32_sub32_add_neon: 18575.5 17157.0 14249.3 12015.1 I.e. in general a very minor overhead for the full subpartition case due to the additional loads and cmps, but a significant speedup for the cases when we only need to process a small part of the actual input data. In common VP9 content in a few inspected clips, 70-90% of the non-dc-only 16x16 and 32x32 IDCTs only have nonzero coefficients in the upper left 8x8 or 16x16 subpartitions respectively. This is cherrypicked from libav commit 9c8bc74c2b40537b0997f646c87c008042d788c2. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:15 +02:00
@ Special case: For the last input column (r1 == 12),
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
@ which would be stored as the last row in the temp buffer,
@ don't store the first 4x4 block, but keep it in registers
@ for the first slice of the second pass (where it is the
@ last 4x4 block).
add r0, r0, #8
arm: vp9itxfm: Avoid .irp when it doesn't save any lines This makes it more readable. This is cherrypicked from libav commit 3bc5b28d5a191864c54bba60646933a63da31656. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-04 22:16:09 +02:00
vst1.16 {d20}, [r0,:64]!
vst1.16 {d24}, [r0,:64]!
vst1.16 {d28}, [r0,:64]!
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
add r0, r0, #8
arm: vp9itxfm: Avoid .irp when it doesn't save any lines This makes it more readable. This is cherrypicked from libav commit 3bc5b28d5a191864c54bba60646933a63da31656. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-04 22:16:09 +02:00
vst1.16 {d21}, [r0,:64]!
vst1.16 {d25}, [r0,:64]!
vst1.16 {d29}, [r0,:64]!
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
add r0, r0, #8
arm: vp9itxfm: Avoid .irp when it doesn't save any lines This makes it more readable. This is cherrypicked from libav commit 3bc5b28d5a191864c54bba60646933a63da31656. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-04 22:16:09 +02:00
vst1.16 {d22}, [r0,:64]!
vst1.16 {d26}, [r0,:64]!
vst1.16 {d30}, [r0,:64]!
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
add r0, r0, #8
arm: vp9itxfm: Avoid .irp when it doesn't save any lines This makes it more readable. This is cherrypicked from libav commit 3bc5b28d5a191864c54bba60646933a63da31656. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-04 22:16:09 +02:00
vst1.16 {d23}, [r0,:64]!
vst1.16 {d27}, [r0,:64]!
vst1.16 {d31}, [r0,:64]!
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
vmov d28, d16
vmov d29, d17
vmov d30, d18
vmov d31, d19
arm: vp9itxfm: Make the larger core transforms standalone functions This work is sponsored by, and copyright, Google. This reduces the code size of libavcodec/arm/vp9itxfm_neon.o from 15324 to 12388 bytes. This gives a small slowdown of a couple tens of cycles, up to around 150 cycles for the full case of the largest transform, but makes it more feasible to add more optimized versions of these transforms. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub4_add_neon: 2063.4 1516.0 1719.5 1245.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3279.3 2454.5 2525.2 1982.3 vp9_inv_dct_dct_32x32_sub4_add_neon: 10750.0 7955.4 8525.6 6754.2 vp9_inv_dct_dct_32x32_sub32_add_neon: 18574.0 17108.4 14216.7 12010.2 After: vp9_inv_dct_dct_16x16_sub4_add_neon: 2060.8 1608.5 1735.7 1262.0 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.2 2443.5 2546.1 1999.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10682.0 8043.8 8581.3 6810.1 vp9_inv_dct_dct_32x32_sub32_add_neon: 18522.4 17277.4 14286.7 12087.9 This is cherrypicked from libav commit 0331c3f5e8cb6e6b53fab7893e91d1be1bfa979c. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-23 10:56:12 +02:00
pop {pc}
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
endfunc
@ Read a vertical 4x16 slice out of a 16x16 matrix, do a transform on it,
@ load the destination pixels (from a similar 4x16 slice), add and store back.
@ r0 = dst
@ r1 = dst stride
@ r2 = src (temp buffer)
@ r3 = slice offset
function \txfm\()16_1d_4x16_pass2_neon
arm: vp9itxfm: Make the larger core transforms standalone functions This work is sponsored by, and copyright, Google. This reduces the code size of libavcodec/arm/vp9itxfm_neon.o from 15324 to 12388 bytes. This gives a small slowdown of a couple tens of cycles, up to around 150 cycles for the full case of the largest transform, but makes it more feasible to add more optimized versions of these transforms. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub4_add_neon: 2063.4 1516.0 1719.5 1245.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3279.3 2454.5 2525.2 1982.3 vp9_inv_dct_dct_32x32_sub4_add_neon: 10750.0 7955.4 8525.6 6754.2 vp9_inv_dct_dct_32x32_sub32_add_neon: 18574.0 17108.4 14216.7 12010.2 After: vp9_inv_dct_dct_16x16_sub4_add_neon: 2060.8 1608.5 1735.7 1262.0 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.2 2443.5 2546.1 1999.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10682.0 8043.8 8581.3 6810.1 vp9_inv_dct_dct_32x32_sub32_add_neon: 18522.4 17277.4 14286.7 12087.9 This is cherrypicked from libav commit 0331c3f5e8cb6e6b53fab7893e91d1be1bfa979c. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-23 10:56:12 +02:00
push {lr}
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
mov r12, #32
.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27
vld1.16 {d\i}, [r2,:64], r12
.endr
cmp r3, #0
beq 1f
.irp i, 28, 29, 30, 31
vld1.16 {d\i}, [r2,:64], r12
.endr
1:
add r3, r0, r1
lsl r1, r1, #1
arm: vp9itxfm: Make the larger core transforms standalone functions This work is sponsored by, and copyright, Google. This reduces the code size of libavcodec/arm/vp9itxfm_neon.o from 15324 to 12388 bytes. This gives a small slowdown of a couple tens of cycles, up to around 150 cycles for the full case of the largest transform, but makes it more feasible to add more optimized versions of these transforms. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub4_add_neon: 2063.4 1516.0 1719.5 1245.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3279.3 2454.5 2525.2 1982.3 vp9_inv_dct_dct_32x32_sub4_add_neon: 10750.0 7955.4 8525.6 6754.2 vp9_inv_dct_dct_32x32_sub32_add_neon: 18574.0 17108.4 14216.7 12010.2 After: vp9_inv_dct_dct_16x16_sub4_add_neon: 2060.8 1608.5 1735.7 1262.0 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.2 2443.5 2546.1 1999.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10682.0 8043.8 8581.3 6810.1 vp9_inv_dct_dct_32x32_sub32_add_neon: 18522.4 17277.4 14286.7 12087.9 This is cherrypicked from libav commit 0331c3f5e8cb6e6b53fab7893e91d1be1bfa979c. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-23 10:56:12 +02:00
bl \txfm\()16
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
load_add_store q8, q9, q10, q11
load_add_store q12, q13, q14, q15
arm: vp9itxfm: Make the larger core transforms standalone functions This work is sponsored by, and copyright, Google. This reduces the code size of libavcodec/arm/vp9itxfm_neon.o from 15324 to 12388 bytes. This gives a small slowdown of a couple tens of cycles, up to around 150 cycles for the full case of the largest transform, but makes it more feasible to add more optimized versions of these transforms. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub4_add_neon: 2063.4 1516.0 1719.5 1245.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3279.3 2454.5 2525.2 1982.3 vp9_inv_dct_dct_32x32_sub4_add_neon: 10750.0 7955.4 8525.6 6754.2 vp9_inv_dct_dct_32x32_sub32_add_neon: 18574.0 17108.4 14216.7 12010.2 After: vp9_inv_dct_dct_16x16_sub4_add_neon: 2060.8 1608.5 1735.7 1262.0 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.2 2443.5 2546.1 1999.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10682.0 8043.8 8581.3 6810.1 vp9_inv_dct_dct_32x32_sub32_add_neon: 18522.4 17277.4 14286.7 12087.9 This is cherrypicked from libav commit 0331c3f5e8cb6e6b53fab7893e91d1be1bfa979c. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-23 10:56:12 +02:00
pop {pc}
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
endfunc
.endm
itxfm16_1d_funcs idct
itxfm16_1d_funcs iadst
arm: vp9itxfm: Skip empty slices in the first pass of idct_idct 16x16 and 32x32 This work is sponsored by, and copyright, Google. Previously all subpartitions except the eob=1 (DC) case ran with the same runtime: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub16_add_neon: 3188.1 2435.4 2499.0 1969.0 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.7 16582.3 14207.6 12000.3 By skipping individual 4x16 or 4x32 pixel slices in the first pass, we reduce the runtime of these functions like this: vp9_inv_dct_dct_16x16_sub1_add_neon: 274.6 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2064.0 1534.8 1719.4 1248.7 vp9_inv_dct_dct_16x16_sub4_add_neon: 2135.0 1477.2 1736.3 1249.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2446.7 1828.7 1993.6 1494.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2832.4 2118.3 2266.5 1735.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.7 2475.3 2523.5 1983.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 756.2 456.7 862.0 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10682.2 8190.4 8539.2 6762.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10813.5 8014.9 8518.3 6762.8 vp9_inv_dct_dct_32x32_sub8_add_neon: 11859.6 9313.0 9347.4 7514.5 vp9_inv_dct_dct_32x32_sub12_add_neon: 12946.6 10752.4 10192.2 8280.2 vp9_inv_dct_dct_32x32_sub16_add_neon: 14074.6 11946.5 11001.4 9008.6 vp9_inv_dct_dct_32x32_sub20_add_neon: 15269.9 13662.7 11816.1 9762.6 vp9_inv_dct_dct_32x32_sub24_add_neon: 16327.9 14940.1 12626.7 10516.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17462.7 15776.1 13446.2 11264.7 vp9_inv_dct_dct_32x32_sub32_add_neon: 18575.5 17157.0 14249.3 12015.1 I.e. in general a very minor overhead for the full subpartition case due to the additional loads and cmps, but a significant speedup for the cases when we only need to process a small part of the actual input data. In common VP9 content in a few inspected clips, 70-90% of the non-dc-only 16x16 and 32x32 IDCTs only have nonzero coefficients in the upper left 8x8 or 16x16 subpartitions respectively. This is cherrypicked from libav commit 9c8bc74c2b40537b0997f646c87c008042d788c2. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:15 +02:00
@ This is the minimum eob value for each subpartition, in increments of 4
const min_eob_idct_idct_16, align=4
.short 0, 10, 38, 89
endconst
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.macro itxfm_func16x16 txfm1, txfm2
function ff_vp9_\txfm1\()_\txfm2\()_16x16_add_neon, export=1
.ifc \txfm1\()_\txfm2,idct_idct
cmp r3, #1
beq idct16x16_dc_add_neon
.endif
arm: vp9itxfm: Skip empty slices in the first pass of idct_idct 16x16 and 32x32 This work is sponsored by, and copyright, Google. Previously all subpartitions except the eob=1 (DC) case ran with the same runtime: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub16_add_neon: 3188.1 2435.4 2499.0 1969.0 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.7 16582.3 14207.6 12000.3 By skipping individual 4x16 or 4x32 pixel slices in the first pass, we reduce the runtime of these functions like this: vp9_inv_dct_dct_16x16_sub1_add_neon: 274.6 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2064.0 1534.8 1719.4 1248.7 vp9_inv_dct_dct_16x16_sub4_add_neon: 2135.0 1477.2 1736.3 1249.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2446.7 1828.7 1993.6 1494.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2832.4 2118.3 2266.5 1735.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.7 2475.3 2523.5 1983.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 756.2 456.7 862.0 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10682.2 8190.4 8539.2 6762.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10813.5 8014.9 8518.3 6762.8 vp9_inv_dct_dct_32x32_sub8_add_neon: 11859.6 9313.0 9347.4 7514.5 vp9_inv_dct_dct_32x32_sub12_add_neon: 12946.6 10752.4 10192.2 8280.2 vp9_inv_dct_dct_32x32_sub16_add_neon: 14074.6 11946.5 11001.4 9008.6 vp9_inv_dct_dct_32x32_sub20_add_neon: 15269.9 13662.7 11816.1 9762.6 vp9_inv_dct_dct_32x32_sub24_add_neon: 16327.9 14940.1 12626.7 10516.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17462.7 15776.1 13446.2 11264.7 vp9_inv_dct_dct_32x32_sub32_add_neon: 18575.5 17157.0 14249.3 12015.1 I.e. in general a very minor overhead for the full subpartition case due to the additional loads and cmps, but a significant speedup for the cases when we only need to process a small part of the actual input data. In common VP9 content in a few inspected clips, 70-90% of the non-dc-only 16x16 and 32x32 IDCTs only have nonzero coefficients in the upper left 8x8 or 16x16 subpartitions respectively. This is cherrypicked from libav commit 9c8bc74c2b40537b0997f646c87c008042d788c2. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:15 +02:00
push {r4-r8,lr}
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.ifnc \txfm1\()_\txfm2,idct_idct
vpush {q4-q7}
.endif
@ Align the stack, allocate a temp buffer
arm: vp9itxfm: Simplify the stack alignment code This is one instruction less for thumb, and only have got 1/2 arm/thumb specific instructions. This is cherrypicked from libav commit e5b0fc170f85b00f7dd0ac514918fb5c95253d39. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:09 +02:00
T mov r7, sp
T and r7, r7, #15
A and r7, sp, #15
add r7, r7, #512
sub sp, sp, r7
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
mov r4, r0
mov r5, r1
mov r6, r2
.ifc \txfm1,idct
movrel r12, idct_coeffs
vld1.16 {q0-q1}, [r12,:128]
.endif
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.ifc \txfm1\()_\txfm2,idct_idct
cmp r3, #10
ble idct16x16_quarter_add_neon
cmp r3, #38
ble idct16x16_half_add_neon
arm/aarch64: vp9itxfm: Skip loading the min_eob pointer when it won't be used In the half/quarter cases where we don't use the min_eob array, defer loading the pointer until we know it will be needed. This is cherrypicked from libav commit 3a0d5e206d24d41d87a25ba16a79b2ea04c39d4c. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-26 22:13:10 +02:00
movrel r8, min_eob_idct_idct_16 + 2
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.endif
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.irp i, 0, 4, 8, 12
add r0, sp, #(\i*32)
arm: vp9itxfm: Skip empty slices in the first pass of idct_idct 16x16 and 32x32 This work is sponsored by, and copyright, Google. Previously all subpartitions except the eob=1 (DC) case ran with the same runtime: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub16_add_neon: 3188.1 2435.4 2499.0 1969.0 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.7 16582.3 14207.6 12000.3 By skipping individual 4x16 or 4x32 pixel slices in the first pass, we reduce the runtime of these functions like this: vp9_inv_dct_dct_16x16_sub1_add_neon: 274.6 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2064.0 1534.8 1719.4 1248.7 vp9_inv_dct_dct_16x16_sub4_add_neon: 2135.0 1477.2 1736.3 1249.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2446.7 1828.7 1993.6 1494.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2832.4 2118.3 2266.5 1735.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.7 2475.3 2523.5 1983.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 756.2 456.7 862.0 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10682.2 8190.4 8539.2 6762.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10813.5 8014.9 8518.3 6762.8 vp9_inv_dct_dct_32x32_sub8_add_neon: 11859.6 9313.0 9347.4 7514.5 vp9_inv_dct_dct_32x32_sub12_add_neon: 12946.6 10752.4 10192.2 8280.2 vp9_inv_dct_dct_32x32_sub16_add_neon: 14074.6 11946.5 11001.4 9008.6 vp9_inv_dct_dct_32x32_sub20_add_neon: 15269.9 13662.7 11816.1 9762.6 vp9_inv_dct_dct_32x32_sub24_add_neon: 16327.9 14940.1 12626.7 10516.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17462.7 15776.1 13446.2 11264.7 vp9_inv_dct_dct_32x32_sub32_add_neon: 18575.5 17157.0 14249.3 12015.1 I.e. in general a very minor overhead for the full subpartition case due to the additional loads and cmps, but a significant speedup for the cases when we only need to process a small part of the actual input data. In common VP9 content in a few inspected clips, 70-90% of the non-dc-only 16x16 and 32x32 IDCTs only have nonzero coefficients in the upper left 8x8 or 16x16 subpartitions respectively. This is cherrypicked from libav commit 9c8bc74c2b40537b0997f646c87c008042d788c2. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:15 +02:00
.ifc \txfm1\()_\txfm2,idct_idct
.if \i > 0
ldrh_post r1, r8, #2
cmp r3, r1
it le
movle r1, #(16 - \i)/4
ble 1f
.endif
.endif
mov r1, #\i
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
add r2, r6, #(\i*2)
bl \txfm1\()16_1d_4x16_pass1_neon
.endr
arm: vp9itxfm: Skip empty slices in the first pass of idct_idct 16x16 and 32x32 This work is sponsored by, and copyright, Google. Previously all subpartitions except the eob=1 (DC) case ran with the same runtime: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub16_add_neon: 3188.1 2435.4 2499.0 1969.0 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.7 16582.3 14207.6 12000.3 By skipping individual 4x16 or 4x32 pixel slices in the first pass, we reduce the runtime of these functions like this: vp9_inv_dct_dct_16x16_sub1_add_neon: 274.6 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2064.0 1534.8 1719.4 1248.7 vp9_inv_dct_dct_16x16_sub4_add_neon: 2135.0 1477.2 1736.3 1249.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2446.7 1828.7 1993.6 1494.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2832.4 2118.3 2266.5 1735.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.7 2475.3 2523.5 1983.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 756.2 456.7 862.0 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10682.2 8190.4 8539.2 6762.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10813.5 8014.9 8518.3 6762.8 vp9_inv_dct_dct_32x32_sub8_add_neon: 11859.6 9313.0 9347.4 7514.5 vp9_inv_dct_dct_32x32_sub12_add_neon: 12946.6 10752.4 10192.2 8280.2 vp9_inv_dct_dct_32x32_sub16_add_neon: 14074.6 11946.5 11001.4 9008.6 vp9_inv_dct_dct_32x32_sub20_add_neon: 15269.9 13662.7 11816.1 9762.6 vp9_inv_dct_dct_32x32_sub24_add_neon: 16327.9 14940.1 12626.7 10516.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17462.7 15776.1 13446.2 11264.7 vp9_inv_dct_dct_32x32_sub32_add_neon: 18575.5 17157.0 14249.3 12015.1 I.e. in general a very minor overhead for the full subpartition case due to the additional loads and cmps, but a significant speedup for the cases when we only need to process a small part of the actual input data. In common VP9 content in a few inspected clips, 70-90% of the non-dc-only 16x16 and 32x32 IDCTs only have nonzero coefficients in the upper left 8x8 or 16x16 subpartitions respectively. This is cherrypicked from libav commit 9c8bc74c2b40537b0997f646c87c008042d788c2. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:15 +02:00
.ifc \txfm1\()_\txfm2,idct_idct
b 3f
1:
@ For all-zero slices in pass 1, set d28-d31 to zero, for the in-register
@ passthrough of coefficients to pass 2 and clear the end of the temp buffer
vmov.i16 q14, #0
vmov.i16 q15, #0
2:
subs r1, r1, #1
.rept 4
vst1.16 {q14-q15}, [r0,:128]!
.endr
bne 2b
3:
.endif
arm: vp9itxfm: Only reload the idct coeffs for the iadst_idct combination This avoids reloading them if they haven't been clobbered, if the first pass also was idct. This is similar to what was done in the aarch64 version. This is cherrypicked from libav commit 3c87039a404c5659ae9bf7454a04e186532eb40b. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:14 +02:00
.ifc \txfm1\()_\txfm2,iadst_idct
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
movrel r12, idct_coeffs
vld1.16 {q0-q1}, [r12,:128]
.endif
.irp i, 0, 4, 8, 12
add r0, r4, #(\i)
mov r1, r5
add r2, sp, #(\i*2)
mov r3, #\i
bl \txfm2\()16_1d_4x16_pass2_neon
.endr
arm: vp9itxfm: Simplify the stack alignment code This is one instruction less for thumb, and only have got 1/2 arm/thumb specific instructions. This is cherrypicked from libav commit e5b0fc170f85b00f7dd0ac514918fb5c95253d39. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:09 +02:00
add sp, sp, r7
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.ifnc \txfm1\()_\txfm2,idct_idct
vpop {q4-q7}
.endif
arm: vp9itxfm: Skip empty slices in the first pass of idct_idct 16x16 and 32x32 This work is sponsored by, and copyright, Google. Previously all subpartitions except the eob=1 (DC) case ran with the same runtime: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub16_add_neon: 3188.1 2435.4 2499.0 1969.0 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.7 16582.3 14207.6 12000.3 By skipping individual 4x16 or 4x32 pixel slices in the first pass, we reduce the runtime of these functions like this: vp9_inv_dct_dct_16x16_sub1_add_neon: 274.6 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2064.0 1534.8 1719.4 1248.7 vp9_inv_dct_dct_16x16_sub4_add_neon: 2135.0 1477.2 1736.3 1249.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2446.7 1828.7 1993.6 1494.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2832.4 2118.3 2266.5 1735.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.7 2475.3 2523.5 1983.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 756.2 456.7 862.0 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10682.2 8190.4 8539.2 6762.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10813.5 8014.9 8518.3 6762.8 vp9_inv_dct_dct_32x32_sub8_add_neon: 11859.6 9313.0 9347.4 7514.5 vp9_inv_dct_dct_32x32_sub12_add_neon: 12946.6 10752.4 10192.2 8280.2 vp9_inv_dct_dct_32x32_sub16_add_neon: 14074.6 11946.5 11001.4 9008.6 vp9_inv_dct_dct_32x32_sub20_add_neon: 15269.9 13662.7 11816.1 9762.6 vp9_inv_dct_dct_32x32_sub24_add_neon: 16327.9 14940.1 12626.7 10516.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17462.7 15776.1 13446.2 11264.7 vp9_inv_dct_dct_32x32_sub32_add_neon: 18575.5 17157.0 14249.3 12015.1 I.e. in general a very minor overhead for the full subpartition case due to the additional loads and cmps, but a significant speedup for the cases when we only need to process a small part of the actual input data. In common VP9 content in a few inspected clips, 70-90% of the non-dc-only 16x16 and 32x32 IDCTs only have nonzero coefficients in the upper left 8x8 or 16x16 subpartitions respectively. This is cherrypicked from libav commit 9c8bc74c2b40537b0997f646c87c008042d788c2. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:15 +02:00
pop {r4-r8,pc}
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
endfunc
.endm
itxfm_func16x16 idct, idct
itxfm_func16x16 iadst, idct
itxfm_func16x16 idct, iadst
itxfm_func16x16 iadst, iadst
.ltorg
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
function idct16_1d_4x16_pass1_quarter_neon
push {lr}
mov r12, #32
vmov.s16 q2, #0
.irp i, 16, 17, 18, 19
vld1.16 {d\i}, [r2,:64]
vst1.16 {d4}, [r2,:64], r12
.endr
bl idct16_quarter
@ Do four 4x4 transposes. Originally, d16-d31 contain the
@ 16 rows. Afterwards, d16-d19, d20-d23, d24-d27, d28-d31
@ contain the transposed 4x4 blocks.
transpose16_q_4x_4x4 q8, q9, q10, q11, q12, q13, q14, q15, d16, d17, d18, d19, d20, d21, d22, d23, d24, d25, d26, d27, d28, d29, d30, d31
@ Store the transposed 4x4 blocks horizontally.
@ The first 4x4 block is kept in registers for the second pass,
@ store the rest in the temp buffer.
add r0, r0, #8
vst1.16 {d20}, [r0,:64]!
vst1.16 {d24}, [r0,:64]!
vst1.16 {d28}, [r0,:64]!
add r0, r0, #8
vst1.16 {d21}, [r0,:64]!
vst1.16 {d25}, [r0,:64]!
vst1.16 {d29}, [r0,:64]!
add r0, r0, #8
vst1.16 {d22}, [r0,:64]!
vst1.16 {d26}, [r0,:64]!
vst1.16 {d30}, [r0,:64]!
add r0, r0, #8
vst1.16 {d23}, [r0,:64]!
vst1.16 {d27}, [r0,:64]!
vst1.16 {d31}, [r0,:64]!
pop {pc}
endfunc
function idct16_1d_4x16_pass2_quarter_neon
push {lr}
@ Only load the top 4 lines, and only do it for the later slices.
@ For the first slice, d16-d19 is kept in registers from the first pass.
cmp r3, #0
beq 1f
mov r12, #32
.irp i, 16, 17, 18, 19
vld1.16 {d\i}, [r2,:64], r12
.endr
1:
add r3, r0, r1
lsl r1, r1, #1
bl idct16_quarter
load_add_store q8, q9, q10, q11
load_add_store q12, q13, q14, q15
pop {pc}
endfunc
function idct16_1d_4x16_pass1_half_neon
push {lr}
mov r12, #32
vmov.s16 q2, #0
.irp i, 16, 17, 18, 19, 20, 21, 22, 23
vld1.16 {d\i}, [r2,:64]
vst1.16 {d4}, [r2,:64], r12
.endr
bl idct16_half
@ Do four 4x4 transposes. Originally, d16-d31 contain the
@ 16 rows. Afterwards, d16-d19, d20-d23, d24-d27, d28-d31
@ contain the transposed 4x4 blocks.
transpose16_q_4x_4x4 q8, q9, q10, q11, q12, q13, q14, q15, d16, d17, d18, d19, d20, d21, d22, d23, d24, d25, d26, d27, d28, d29, d30, d31
@ Store the transposed 4x4 blocks horizontally.
cmp r1, #4
beq 1f
.irp i, 16, 20, 24, 28, 17, 21, 25, 29, 18, 22, 26, 30, 19, 23, 27, 31
vst1.16 {d\i}, [r0,:64]!
.endr
pop {pc}
1:
@ Special case: For the second input column (r1 == 4),
@ which would be stored as the second row in the temp buffer,
@ don't store the first 4x4 block, but keep it in registers
@ for the first slice of the second pass (where it is the
@ second 4x4 block).
add r0, r0, #8
vst1.16 {d20}, [r0,:64]!
vst1.16 {d24}, [r0,:64]!
vst1.16 {d28}, [r0,:64]!
add r0, r0, #8
vst1.16 {d21}, [r0,:64]!
vst1.16 {d25}, [r0,:64]!
vst1.16 {d29}, [r0,:64]!
add r0, r0, #8
vst1.16 {d22}, [r0,:64]!
vst1.16 {d26}, [r0,:64]!
vst1.16 {d30}, [r0,:64]!
add r0, r0, #8
vst1.16 {d23}, [r0,:64]!
vst1.16 {d27}, [r0,:64]!
vst1.16 {d31}, [r0,:64]!
vmov d20, d16
vmov d21, d17
vmov d22, d18
vmov d23, d19
pop {pc}
endfunc
function idct16_1d_4x16_pass2_half_neon
push {lr}
mov r12, #32
cmp r3, #0
.irp i, 16, 17, 18, 19
vld1.16 {d\i}, [r2,:64], r12
.endr
beq 1f
.irp i, 20, 21, 22, 23
vld1.16 {d\i}, [r2,:64], r12
.endr
1:
add r3, r0, r1
lsl r1, r1, #1
bl idct16_half
load_add_store q8, q9, q10, q11
load_add_store q12, q13, q14, q15
pop {pc}
endfunc
.purgem load_add_store
.macro idct16_partial size
function idct16x16_\size\()_add_neon
add r0, sp, #(0*32)
mov r1, #0
add r2, r6, #(0*2)
bl idct16_1d_4x16_pass1_\size\()_neon
.ifc \size,half
add r0, sp, #(4*32)
mov r1, #4
add r2, r6, #(4*2)
bl idct16_1d_4x16_pass1_\size\()_neon
.endif
.irp i, 0, 4, 8, 12
add r0, r4, #(\i)
mov r1, r5
add r2, sp, #(\i*2)
mov r3, #\i
bl idct16_1d_4x16_pass2_\size\()_neon
.endr
add sp, sp, r7
pop {r4-r8,pc}
endfunc
.endm
idct16_partial quarter
idct16_partial half
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
function idct32x32_dc_add_neon
movrel r12, idct_coeffs
vld1.16 {d0}, [r12,:64]
arm/aarch64: vp9: Fix vertical alignment Align the second/third operands as they usually are. Due to the wildly varying sizes of the written out operands in aarch64 assembly, the column alignment is usually not as clear as in arm assembly. This is cherrypicked from libav commit 7995ebfad12002033c73feed422a1cfc62081e8f. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-09 00:04:19 +02:00
vmov.i16 q2, #0
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
vld1.16 {d16[]}, [r2,:16]
vmull.s16 q8, d16, d0[0]
vrshrn.s32 d16, q8, #14
vmull.s16 q8, d16, d0[0]
vrshrn.s32 d16, q8, #14
vdup.16 q8, d16[0]
vst1.16 {d4[0]}, [r2,:16]
vrshr.s16 q8, q8, #6
arm: vp9itxfm: Optimize 16x16 and 32x32 idct dc by unrolling This work is sponsored by, and copyright, Google. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_32x32_sub1_add_neon: 752.0 459.2 862.2 553.9 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 226.5 145.0 225.1 171.8 vp9_inv_dct_dct_32x32_sub1_add_neon: 721.2 415.7 727.6 475.0 This is cherrypicked from libav commit a76bf8cf1277ef6feb1580b578f5e6ca327e713c. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-04 13:08:51 +02:00
mov r3, r0
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
mov r12, #32
1:
@ Loop to add the constant from q8 into all 32x32 outputs
arm: vp9itxfm: Optimize 16x16 and 32x32 idct dc by unrolling This work is sponsored by, and copyright, Google. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_32x32_sub1_add_neon: 752.0 459.2 862.2 553.9 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 226.5 145.0 225.1 171.8 vp9_inv_dct_dct_32x32_sub1_add_neon: 721.2 415.7 727.6 475.0 This is cherrypicked from libav commit a76bf8cf1277ef6feb1580b578f5e6ca327e713c. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-04 13:08:51 +02:00
subs r12, r12, #2
vld1.8 {q0-q1}, [r0,:128], r1
vaddw.u8 q9, q8, d0
vaddw.u8 q10, q8, d1
vld1.8 {q2-q3}, [r0,:128], r1
vaddw.u8 q11, q8, d2
vaddw.u8 q12, q8, d3
vaddw.u8 q13, q8, d4
vaddw.u8 q14, q8, d5
vaddw.u8 q15, q8, d6
vqmovun.s16 d0, q9
vaddw.u8 q9, q8, d7
vqmovun.s16 d1, q10
vqmovun.s16 d2, q11
vqmovun.s16 d3, q12
vqmovun.s16 d4, q13
vqmovun.s16 d5, q14
vst1.8 {q0-q1}, [r3,:128], r1
vqmovun.s16 d6, q15
vqmovun.s16 d7, q9
vst1.8 {q2-q3}, [r3,:128], r1
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
bne 1b
bx lr
endfunc
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.macro idct32_end
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
butterfly d16, d9, d8, d9 @ d16 = t16a, d9 = t19a
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
butterfly d17, d20, d23, d20 @ d17 = t17, d20 = t18
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
butterfly d18, d10, d11, d10 @ d18 = t23a, d10 = t20a
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
butterfly d19, d21, d22, d21 @ d19 = t22, d21 = t21
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
butterfly d8, d28, d28, d30 @ d8 = t24a, d28 = t27a
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
butterfly d23, d26, d25, d26 @ d23 = t25, d26 = t26
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
butterfly d11, d29, d29, d31 @ d11 = t31a, d29 = t28a
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
butterfly d22, d27, d24, d27 @ d22 = t30, d27 = t29
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
mbutterfly d27, d20, d0[2], d0[3], q12, q15 @ d27 = t18a, d20 = t29a
mbutterfly d29, d9, d0[2], d0[3], q12, q15 @ d29 = t19, d5 = t28
mbutterfly d28, d10, d0[2], d0[3], q12, q15, neg=1 @ d28 = t27, d6 = t20
mbutterfly d26, d21, d0[2], d0[3], q12, q15, neg=1 @ d26 = t26a, d21 = t21a
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
butterfly d31, d24, d11, d8 @ d31 = t31, d24 = t24
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
butterfly d30, d25, d22, d23 @ d30 = t30a, d25 = t25a
butterfly_r d23, d16, d16, d18 @ d23 = t23, d16 = t16
butterfly_r d22, d17, d17, d19 @ d22 = t22a, d17 = t17a
butterfly d18, d21, d27, d21 @ d18 = t18, d21 = t21
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
butterfly_r d27, d28, d9, d28 @ d27 = t27a, d28 = t28a
butterfly d8, d26, d20, d26 @ d8 = t29, d26 = t26
butterfly d19, d20, d29, d10 @ d19 = t19a, d20 = t20
vmov d29, d8 @ d29 = t29
mbutterfly0 d27, d20, d27, d20, d8, d10, q4, q5 @ d27 = t27, d20 = t20
mbutterfly0 d26, d21, d26, d21, d8, d10, q4, q5 @ d26 = t26a, d21 = t21a
mbutterfly0 d25, d22, d25, d22, d8, d10, q4, q5 @ d25 = t25, d22 = t22
mbutterfly0 d24, d23, d24, d23, d8, d10, q4, q5 @ d24 = t24a, d23 = t23a
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
bx lr
.endm
arm: vp9itxfm: Make the larger core transforms standalone functions This work is sponsored by, and copyright, Google. This reduces the code size of libavcodec/arm/vp9itxfm_neon.o from 15324 to 12388 bytes. This gives a small slowdown of a couple tens of cycles, up to around 150 cycles for the full case of the largest transform, but makes it more feasible to add more optimized versions of these transforms. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub4_add_neon: 2063.4 1516.0 1719.5 1245.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3279.3 2454.5 2525.2 1982.3 vp9_inv_dct_dct_32x32_sub4_add_neon: 10750.0 7955.4 8525.6 6754.2 vp9_inv_dct_dct_32x32_sub32_add_neon: 18574.0 17108.4 14216.7 12010.2 After: vp9_inv_dct_dct_16x16_sub4_add_neon: 2060.8 1608.5 1735.7 1262.0 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.2 2443.5 2546.1 1999.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10682.0 8043.8 8581.3 6810.1 vp9_inv_dct_dct_32x32_sub32_add_neon: 18522.4 17277.4 14286.7 12087.9 This is cherrypicked from libav commit 0331c3f5e8cb6e6b53fab7893e91d1be1bfa979c. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-23 10:56:12 +02:00
function idct32_odd
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
mbutterfly d16, d31, d4[0], d4[1], q4, q5 @ d16 = t16a, d31 = t31a
mbutterfly d24, d23, d4[2], d4[3], q4, q5 @ d24 = t17a, d23 = t30a
mbutterfly d20, d27, d5[0], d5[1], q4, q5 @ d20 = t18a, d27 = t29a
mbutterfly d28, d19, d5[2], d5[3], q4, q5 @ d28 = t19a, d19 = t28a
mbutterfly d18, d29, d6[0], d6[1], q4, q5 @ d18 = t20a, d29 = t27a
mbutterfly d26, d21, d6[2], d6[3], q4, q5 @ d26 = t21a, d21 = t26a
mbutterfly d22, d25, d7[0], d7[1], q4, q5 @ d22 = t22a, d25 = t25a
mbutterfly d30, d17, d7[2], d7[3], q4, q5 @ d30 = t23a, d17 = t24a
butterfly d8, d24, d16, d24 @ d8 = t16, d24 = t17
butterfly d9, d20, d28, d20 @ d9 = t19, d20 = t18
butterfly d10, d26, d18, d26 @ d10 = t20, d26 = t21
butterfly d11, d22, d30, d22 @ d11 = t23, d22 = t22
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
butterfly d28, d25, d17, d25 @ d28 = t24, d25 = t25
butterfly d30, d21, d29, d21 @ d30 = t27, d21 = t26
butterfly d29, d23, d31, d23 @ d29 = t31, d23 = t30
butterfly d31, d27, d19, d27 @ d31 = t28, d27 = t29
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
mbutterfly d23, d24, d1[0], d1[1], q8, q9 @ d23 = t17a, d24 = t30a
mbutterfly d27, d20, d1[0], d1[1], q8, q9, neg=1 @ d27 = t29a, d20 = t18a
mbutterfly d21, d26, d1[2], d1[3], q8, q9 @ d21 = t21a, d26 = t26a
mbutterfly d25, d22, d1[2], d1[3], q8, q9, neg=1 @ d25 = t25a, d22 = t22a
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
idct32_end
endfunc
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
function idct32_odd_half
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
mbutterfly_h1 d16, d31, d4[0], d4[1], q4, q5 @ d16 = t16a, d31 = t31a
mbutterfly_h2 d24, d23, d4[2], d4[3], q4, q5 @ d24 = t17a, d23 = t30a
mbutterfly_h1 d20, d27, d5[0], d5[1], q4, q5 @ d20 = t18a, d27 = t29a
mbutterfly_h2 d28, d19, d5[2], d5[3], q4, q5 @ d28 = t19a, d19 = t28a
mbutterfly_h1 d18, d29, d6[0], d6[1], q4, q5 @ d18 = t20a, d29 = t27a
mbutterfly_h2 d26, d21, d6[2], d6[3], q4, q5 @ d26 = t21a, d21 = t26a
mbutterfly_h1 d22, d25, d7[0], d7[1], q4, q5 @ d22 = t22a, d25 = t25a
mbutterfly_h2 d30, d17, d7[2], d7[3], q4, q5 @ d30 = t23a, d17 = t24a
butterfly d8, d24, d16, d24 @ d8 = t16, d24 = t17
butterfly d9, d20, d28, d20 @ d9 = t19, d20 = t18
butterfly d10, d26, d18, d26 @ d10 = t20, d26 = t21
butterfly d11, d22, d30, d22 @ d11 = t23, d22 = t22
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
butterfly d28, d25, d17, d25 @ d28 = t24, d25 = t25
butterfly d30, d21, d29, d21 @ d30 = t27, d21 = t26
butterfly d29, d23, d31, d23 @ d29 = t31, d23 = t30
butterfly d31, d27, d19, d27 @ d31 = t28, d27 = t29
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
mbutterfly d23, d24, d1[0], d1[1], q8, q9 @ d23 = t17a, d24 = t30a
mbutterfly d27, d20, d1[0], d1[1], q8, q9, neg=1 @ d27 = t29a, d20 = t18a
mbutterfly d21, d26, d1[2], d1[3], q8, q9 @ d21 = t21a, d26 = t26a
mbutterfly d25, d22, d1[2], d1[3], q8, q9, neg=1 @ d25 = t25a, d22 = t22a
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
idct32_end
arm: vp9itxfm: Make the larger core transforms standalone functions This work is sponsored by, and copyright, Google. This reduces the code size of libavcodec/arm/vp9itxfm_neon.o from 15324 to 12388 bytes. This gives a small slowdown of a couple tens of cycles, up to around 150 cycles for the full case of the largest transform, but makes it more feasible to add more optimized versions of these transforms. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub4_add_neon: 2063.4 1516.0 1719.5 1245.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3279.3 2454.5 2525.2 1982.3 vp9_inv_dct_dct_32x32_sub4_add_neon: 10750.0 7955.4 8525.6 6754.2 vp9_inv_dct_dct_32x32_sub32_add_neon: 18574.0 17108.4 14216.7 12010.2 After: vp9_inv_dct_dct_16x16_sub4_add_neon: 2060.8 1608.5 1735.7 1262.0 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.2 2443.5 2546.1 1999.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10682.0 8043.8 8581.3 6810.1 vp9_inv_dct_dct_32x32_sub32_add_neon: 18522.4 17277.4 14286.7 12087.9 This is cherrypicked from libav commit 0331c3f5e8cb6e6b53fab7893e91d1be1bfa979c. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-23 10:56:12 +02:00
endfunc
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
function idct32_odd_quarter
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vmull.s16 q4, d16, d4[0]
vmull.s16 q14, d19, d5[3]
vmull.s16 q15, d16, d4[1]
vmull.s16 q11, d17, d7[2]
vmull.s16 q5, d17, d7[3]
vmull.s16 q13, d19, d5[2]
vmull.s16 q10, d18, d6[0]
vmull.s16 q12, d18, d6[1]
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
vneg.s32 q14, q14
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vneg.s32 q5, q5
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vrshrn.s32 d8, q4, #14
vrshrn.s32 d9, q14, #14
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
vrshrn.s32 d29, q15, #14
vrshrn.s32 d28, q11, #14
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vrshrn.s32 d11, q5, #14
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
vrshrn.s32 d31, q13, #14
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vrshrn.s32 d10, q10, #14
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
vrshrn.s32 d30, q12, #14
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
mbutterfly_l q8, q9, d29, d8, d1[0], d1[1]
mbutterfly_l q13, q10, d31, d9, d1[0], d1[1]
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
vrshrn.s32 d23, q8, #14
vrshrn.s32 d24, q9, #14
vneg.s32 q10, q10
vrshrn.s32 d27, q13, #14
vrshrn.s32 d20, q10, #14
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
mbutterfly_l q8, q9, d30, d10, d1[2], d1[3]
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
vrshrn.s32 d21, q8, #14
vrshrn.s32 d26, q9, #14
arm: vp9itxfm: Reorder the idct coefficients for better pairing All elements are used pairwise, except for the first one. Previously, the 16th element was unused. Move the unused element to the second slot, to make the later element pairs not split across registers. This simplifies loading only parts of the coefficients, reducing the difference to the 16 bpp version. This is cherrypicked from libav commit de06bdfe6c8abd8266d5c6f5c68e4df0060b61fc. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-12-31 14:05:44 +02:00
mbutterfly_l q8, q9, d28, d11, d1[2], d1[3]
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
vrshrn.s32 d25, q8, #14
vneg.s32 q9, q9
vrshrn.s32 d22, q9, #14
idct32_end
endfunc
.macro idct32_funcs suffix
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
@ Do an 32-point IDCT of a 4x32 slice out of a 32x32 matrix.
@ We don't have register space to do a single pass IDCT of 4x32 though,
@ but the 32-point IDCT can be decomposed into two 16-point IDCTs;
@ a normal IDCT16 with every other input component (the even ones, with
@ each output written twice), followed by a separate 16-point IDCT
@ of the odd inputs, added/subtracted onto the outputs of the first idct16.
@ r0 = dst (temp buffer)
@ r1 = unused
@ r2 = src
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
function idct32_1d_4x32_pass1\suffix\()_neon
arm: vp9itxfm: Make the larger core transforms standalone functions This work is sponsored by, and copyright, Google. This reduces the code size of libavcodec/arm/vp9itxfm_neon.o from 15324 to 12388 bytes. This gives a small slowdown of a couple tens of cycles, up to around 150 cycles for the full case of the largest transform, but makes it more feasible to add more optimized versions of these transforms. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub4_add_neon: 2063.4 1516.0 1719.5 1245.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3279.3 2454.5 2525.2 1982.3 vp9_inv_dct_dct_32x32_sub4_add_neon: 10750.0 7955.4 8525.6 6754.2 vp9_inv_dct_dct_32x32_sub32_add_neon: 18574.0 17108.4 14216.7 12010.2 After: vp9_inv_dct_dct_16x16_sub4_add_neon: 2060.8 1608.5 1735.7 1262.0 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.2 2443.5 2546.1 1999.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10682.0 8043.8 8581.3 6810.1 vp9_inv_dct_dct_32x32_sub32_add_neon: 18522.4 17277.4 14286.7 12087.9 This is cherrypicked from libav commit 0331c3f5e8cb6e6b53fab7893e91d1be1bfa979c. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-23 10:56:12 +02:00
push {lr}
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
@ idct16 clobbers q2-q3 (since it doesn't clobber q4-q7 at all
@ when doing the normal 16x16 idct), so move the idct32_odd coeffs
@ to q4-q5
vmov q4, q2
vmov q5, q3
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
@ Double stride of the input, since we only read every other line
mov r12, #128
arm/aarch64: vp9: Fix vertical alignment Align the second/third operands as they usually are. Due to the wildly varying sizes of the written out operands in aarch64 assembly, the column alignment is usually not as clear as in arm assembly. This is cherrypicked from libav commit 7995ebfad12002033c73feed422a1cfc62081e8f. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-09 00:04:19 +02:00
vmov.s16 d4, #0
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
@ d16 = IN(0), d17 = IN(2) ... d31 = IN(30)
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.ifb \suffix
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
vld1.16 {d\i}, [r2,:64]
vst1.16 {d4}, [r2,:64], r12
.endr
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.endif
.ifc \suffix,_quarter
.irp i, 16, 17, 18, 19
vld1.16 {d\i}, [r2,:64]
vst1.16 {d4}, [r2,:64], r12
.endr
.endif
.ifc \suffix,_half
.irp i, 16, 17, 18, 19, 20, 21, 22, 23
vld1.16 {d\i}, [r2,:64]
vst1.16 {d4}, [r2,:64], r12
.endr
.endif
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
bl idct16\suffix
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
@ Move the idct32_odd coeffs back into q2-q3 for idct32_odd;
@ the constants for a vmul with a lane must be in q0-q3.
vmov q2, q4
vmov q3, q5
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
@ Do four 4x4 transposes. Originally, d16-d31 contain the
@ 16 rows. Afterwards, d16-d19, d20-d23, d24-d27, d28-d31
@ contain the transposed 4x4 blocks.
transpose16_q_4x_4x4 q8, q9, q10, q11, q12, q13, q14, q15, d16, d17, d18, d19, d20, d21, d22, d23, d24, d25, d26, d27, d28, d29, d30, d31
@ Store the registers a, b, c, d horizontally, followed
@ by the same registers d, c, b, a mirrored.
.macro store_rev a, b, c, d
.irp i, \a, \b, \c, \d
vst1.16 {d\i}, [r0,:64]!
vrev64.16 d\i, d\i
.endr
.irp i, \d, \c, \b, \a
vst1.16 {d\i}, [r0,:64]!
.endr
.endm
store_rev 16, 20, 24, 28
store_rev 17, 21, 25, 29
store_rev 18, 22, 26, 30
store_rev 19, 23, 27, 31
sub r0, r0, #256
.purgem store_rev
@ Move r2 back to the start of the input, and move
@ to the first odd row
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.ifb \suffix
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
sub r2, r2, r12, lsl #4
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.endif
.ifc \suffix,_quarter
sub r2, r2, r12, lsl #2
.endif
.ifc \suffix,_half
sub r2, r2, r12, lsl #3
.endif
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
add r2, r2, #64
arm/aarch64: vp9: Fix vertical alignment Align the second/third operands as they usually are. Due to the wildly varying sizes of the written out operands in aarch64 assembly, the column alignment is usually not as clear as in arm assembly. This is cherrypicked from libav commit 7995ebfad12002033c73feed422a1cfc62081e8f. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-09 00:04:19 +02:00
vmov.s16 d8, #0
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
@ d16 = IN(1), d17 = IN(3) ... d31 = IN(31)
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.ifb \suffix
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
vld1.16 {d\i}, [r2,:64]
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vst1.16 {d8}, [r2,:64], r12
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.endr
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.endif
.ifc \suffix,_quarter
.irp i, 16, 17, 18, 19
vld1.16 {d\i}, [r2,:64]
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vst1.16 {d8}, [r2,:64], r12
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.endr
.endif
.ifc \suffix,_half
.irp i, 16, 17, 18, 19, 20, 21, 22, 23
vld1.16 {d\i}, [r2,:64]
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vst1.16 {d8}, [r2,:64], r12
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.endr
.endif
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
bl idct32_odd\suffix
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
transpose16_q_4x_4x4 q15, q14, q13, q12, q11, q10, q9, q8, d31, d30, d29, d28, d27, d26, d25, d24, d23, d22, d21, d20, d19, d18, d17, d16
@ Store the registers a, b, c, d horizontally,
@ adding into the output first, and then mirrored, subtracted
@ from the output.
.macro store_rev a, b, c, d
.irp i, \a, \b, \c, \d
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vld1.16 {d8}, [r0,:64]
vadd.s16 d8, d8, d\i
vst1.16 {d8}, [r0,:64]!
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
vrev64.16 d\i, d\i
.endr
.irp i, \d, \c, \b, \a
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vld1.16 {d8}, [r0,:64]
vsub.s16 d8, d8, d\i
vst1.16 {d8}, [r0,:64]!
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.endr
.endm
arm/aarch64: vp9itxfm: Fix indentation of macro arguments This is cherrypicked from libav commit 721bc37522c5c1d6a8c3cea5e9c3fcde8d256c05. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:11 +02:00
store_rev 31, 27, 23, 19
store_rev 30, 26, 22, 18
store_rev 29, 25, 21, 17
store_rev 28, 24, 20, 16
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.purgem store_rev
arm: vp9itxfm: Make the larger core transforms standalone functions This work is sponsored by, and copyright, Google. This reduces the code size of libavcodec/arm/vp9itxfm_neon.o from 15324 to 12388 bytes. This gives a small slowdown of a couple tens of cycles, up to around 150 cycles for the full case of the largest transform, but makes it more feasible to add more optimized versions of these transforms. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub4_add_neon: 2063.4 1516.0 1719.5 1245.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3279.3 2454.5 2525.2 1982.3 vp9_inv_dct_dct_32x32_sub4_add_neon: 10750.0 7955.4 8525.6 6754.2 vp9_inv_dct_dct_32x32_sub32_add_neon: 18574.0 17108.4 14216.7 12010.2 After: vp9_inv_dct_dct_16x16_sub4_add_neon: 2060.8 1608.5 1735.7 1262.0 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.2 2443.5 2546.1 1999.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10682.0 8043.8 8581.3 6810.1 vp9_inv_dct_dct_32x32_sub32_add_neon: 18522.4 17277.4 14286.7 12087.9 This is cherrypicked from libav commit 0331c3f5e8cb6e6b53fab7893e91d1be1bfa979c. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-23 10:56:12 +02:00
pop {pc}
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
endfunc
.ltorg
@ This is mostly the same as 4x32_pass1, but without the transpose,
@ and use the source as temp buffer between the two idct passes, and
@ add into the destination.
@ r0 = dst
@ r1 = dst stride
@ r2 = src (temp buffer)
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
function idct32_1d_4x32_pass2\suffix\()_neon
arm: vp9itxfm: Make the larger core transforms standalone functions This work is sponsored by, and copyright, Google. This reduces the code size of libavcodec/arm/vp9itxfm_neon.o from 15324 to 12388 bytes. This gives a small slowdown of a couple tens of cycles, up to around 150 cycles for the full case of the largest transform, but makes it more feasible to add more optimized versions of these transforms. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub4_add_neon: 2063.4 1516.0 1719.5 1245.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3279.3 2454.5 2525.2 1982.3 vp9_inv_dct_dct_32x32_sub4_add_neon: 10750.0 7955.4 8525.6 6754.2 vp9_inv_dct_dct_32x32_sub32_add_neon: 18574.0 17108.4 14216.7 12010.2 After: vp9_inv_dct_dct_16x16_sub4_add_neon: 2060.8 1608.5 1735.7 1262.0 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.2 2443.5 2546.1 1999.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10682.0 8043.8 8581.3 6810.1 vp9_inv_dct_dct_32x32_sub32_add_neon: 18522.4 17277.4 14286.7 12087.9 This is cherrypicked from libav commit 0331c3f5e8cb6e6b53fab7893e91d1be1bfa979c. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-23 10:56:12 +02:00
push {lr}
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vmov q4, q2
vmov q5, q3
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
mov r12, #128
@ d16 = IN(0), d17 = IN(2) ... d31 = IN(30)
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.ifb \suffix
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
vld1.16 {d\i}, [r2,:64], r12
.endr
sub r2, r2, r12, lsl #4
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.endif
.ifc \suffix,_quarter
.irp i, 16, 17, 18, 19
vld1.16 {d\i}, [r2,:64], r12
.endr
sub r2, r2, r12, lsl #2
.endif
.ifc \suffix,_half
.irp i, 16, 17, 18, 19, 20, 21, 22, 23
vld1.16 {d\i}, [r2,:64], r12
.endr
sub r2, r2, r12, lsl #3
.endif
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
bl idct16\suffix
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vmov q2, q4
vmov q3, q5
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
vst1.16 {d\i}, [r2,:64], r12
.endr
sub r2, r2, r12, lsl #4
add r2, r2, #64
@ d16 = IN(1), d17 = IN(3) ... d31 = IN(31)
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.ifb \suffix
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
vld1.16 {d\i}, [r2,:64], r12
.endr
sub r2, r2, r12, lsl #4
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.endif
.ifc \suffix,_quarter
.irp i, 16, 17, 18, 19
vld1.16 {d\i}, [r2,:64], r12
.endr
sub r2, r2, r12, lsl #2
.endif
.ifc \suffix,_half
.irp i, 16, 17, 18, 19, 20, 21, 22, 23
vld1.16 {d\i}, [r2,:64], r12
.endr
sub r2, r2, r12, lsl #3
.endif
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
sub r2, r2, #64
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
bl idct32_odd\suffix
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
mov r12, #128
.macro load_acc_store a, b, c, d, neg=0
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vld1.16 {d8}, [r2,:64], r12
vld1.16 {d9}, [r2,:64], r12
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.if \neg == 0
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vadd.s16 d8, d8, d\a
vld1.16 {d10}, [r2,:64], r12
vadd.s16 d9, d9, d\b
vld1.16 {d11}, [r2,:64], r12
vadd.s16 d10, d10, d\c
vadd.s16 d11, d11, d\d
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.else
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vsub.s16 d8, d8, d\a
vld1.16 {d10}, [r2,:64], r12
vsub.s16 d9, d9, d\b
vld1.16 {d11}, [r2,:64], r12
vsub.s16 d10, d10, d\c
vsub.s16 d11, d11, d\d
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.endif
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vld1.32 {d12[]}, [r0,:32], r1
vld1.32 {d12[1]}, [r0,:32], r1
arm/aarch64: vp9: Fix vertical alignment Align the second/third operands as they usually are. Due to the wildly varying sizes of the written out operands in aarch64 assembly, the column alignment is usually not as clear as in arm assembly. This is cherrypicked from libav commit 7995ebfad12002033c73feed422a1cfc62081e8f. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-09 00:04:19 +02:00
vrshr.s16 q4, q4, #6
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vld1.32 {d13[]}, [r0,:32], r1
arm/aarch64: vp9: Fix vertical alignment Align the second/third operands as they usually are. Due to the wildly varying sizes of the written out operands in aarch64 assembly, the column alignment is usually not as clear as in arm assembly. This is cherrypicked from libav commit 7995ebfad12002033c73feed422a1cfc62081e8f. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-09 00:04:19 +02:00
vrshr.s16 q5, q5, #6
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vld1.32 {d13[1]}, [r0,:32], r1
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
sub r0, r0, r1, lsl #2
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vaddw.u8 q4, q4, d12
vaddw.u8 q5, q5, d13
vqmovun.s16 d8, q4
vqmovun.s16 d9, q5
vst1.32 {d8[0]}, [r0,:32], r1
vst1.32 {d8[1]}, [r0,:32], r1
vst1.32 {d9[0]}, [r0,:32], r1
vst1.32 {d9[1]}, [r0,:32], r1
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.endm
load_acc_store 31, 30, 29, 28
load_acc_store 27, 26, 25, 24
load_acc_store 23, 22, 21, 20
load_acc_store 19, 18, 17, 16
sub r2, r2, r12
neg r12, r12
load_acc_store 16, 17, 18, 19, 1
load_acc_store 20, 21, 22, 23, 1
load_acc_store 24, 25, 26, 27, 1
load_acc_store 28, 29, 30, 31, 1
.purgem load_acc_store
arm: vp9itxfm: Make the larger core transforms standalone functions This work is sponsored by, and copyright, Google. This reduces the code size of libavcodec/arm/vp9itxfm_neon.o from 15324 to 12388 bytes. This gives a small slowdown of a couple tens of cycles, up to around 150 cycles for the full case of the largest transform, but makes it more feasible to add more optimized versions of these transforms. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub4_add_neon: 2063.4 1516.0 1719.5 1245.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3279.3 2454.5 2525.2 1982.3 vp9_inv_dct_dct_32x32_sub4_add_neon: 10750.0 7955.4 8525.6 6754.2 vp9_inv_dct_dct_32x32_sub32_add_neon: 18574.0 17108.4 14216.7 12010.2 After: vp9_inv_dct_dct_16x16_sub4_add_neon: 2060.8 1608.5 1735.7 1262.0 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.2 2443.5 2546.1 1999.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10682.0 8043.8 8581.3 6810.1 vp9_inv_dct_dct_32x32_sub32_add_neon: 18522.4 17277.4 14286.7 12087.9 This is cherrypicked from libav commit 0331c3f5e8cb6e6b53fab7893e91d1be1bfa979c. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-23 10:56:12 +02:00
pop {pc}
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
endfunc
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.endm
idct32_funcs
idct32_funcs _quarter
idct32_funcs _half
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
arm: vp9itxfm: Skip empty slices in the first pass of idct_idct 16x16 and 32x32 This work is sponsored by, and copyright, Google. Previously all subpartitions except the eob=1 (DC) case ran with the same runtime: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub16_add_neon: 3188.1 2435.4 2499.0 1969.0 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.7 16582.3 14207.6 12000.3 By skipping individual 4x16 or 4x32 pixel slices in the first pass, we reduce the runtime of these functions like this: vp9_inv_dct_dct_16x16_sub1_add_neon: 274.6 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2064.0 1534.8 1719.4 1248.7 vp9_inv_dct_dct_16x16_sub4_add_neon: 2135.0 1477.2 1736.3 1249.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2446.7 1828.7 1993.6 1494.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2832.4 2118.3 2266.5 1735.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.7 2475.3 2523.5 1983.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 756.2 456.7 862.0 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10682.2 8190.4 8539.2 6762.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10813.5 8014.9 8518.3 6762.8 vp9_inv_dct_dct_32x32_sub8_add_neon: 11859.6 9313.0 9347.4 7514.5 vp9_inv_dct_dct_32x32_sub12_add_neon: 12946.6 10752.4 10192.2 8280.2 vp9_inv_dct_dct_32x32_sub16_add_neon: 14074.6 11946.5 11001.4 9008.6 vp9_inv_dct_dct_32x32_sub20_add_neon: 15269.9 13662.7 11816.1 9762.6 vp9_inv_dct_dct_32x32_sub24_add_neon: 16327.9 14940.1 12626.7 10516.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17462.7 15776.1 13446.2 11264.7 vp9_inv_dct_dct_32x32_sub32_add_neon: 18575.5 17157.0 14249.3 12015.1 I.e. in general a very minor overhead for the full subpartition case due to the additional loads and cmps, but a significant speedup for the cases when we only need to process a small part of the actual input data. In common VP9 content in a few inspected clips, 70-90% of the non-dc-only 16x16 and 32x32 IDCTs only have nonzero coefficients in the upper left 8x8 or 16x16 subpartitions respectively. This is cherrypicked from libav commit 9c8bc74c2b40537b0997f646c87c008042d788c2. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:15 +02:00
const min_eob_idct_idct_32, align=4
.short 0, 9, 34, 70, 135, 240, 336, 448
endconst
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
function ff_vp9_idct_idct_32x32_add_neon, export=1
cmp r3, #1
beq idct32x32_dc_add_neon
arm: vp9itxfm: Skip empty slices in the first pass of idct_idct 16x16 and 32x32 This work is sponsored by, and copyright, Google. Previously all subpartitions except the eob=1 (DC) case ran with the same runtime: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub16_add_neon: 3188.1 2435.4 2499.0 1969.0 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.7 16582.3 14207.6 12000.3 By skipping individual 4x16 or 4x32 pixel slices in the first pass, we reduce the runtime of these functions like this: vp9_inv_dct_dct_16x16_sub1_add_neon: 274.6 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2064.0 1534.8 1719.4 1248.7 vp9_inv_dct_dct_16x16_sub4_add_neon: 2135.0 1477.2 1736.3 1249.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2446.7 1828.7 1993.6 1494.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2832.4 2118.3 2266.5 1735.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.7 2475.3 2523.5 1983.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 756.2 456.7 862.0 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10682.2 8190.4 8539.2 6762.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10813.5 8014.9 8518.3 6762.8 vp9_inv_dct_dct_32x32_sub8_add_neon: 11859.6 9313.0 9347.4 7514.5 vp9_inv_dct_dct_32x32_sub12_add_neon: 12946.6 10752.4 10192.2 8280.2 vp9_inv_dct_dct_32x32_sub16_add_neon: 14074.6 11946.5 11001.4 9008.6 vp9_inv_dct_dct_32x32_sub20_add_neon: 15269.9 13662.7 11816.1 9762.6 vp9_inv_dct_dct_32x32_sub24_add_neon: 16327.9 14940.1 12626.7 10516.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17462.7 15776.1 13446.2 11264.7 vp9_inv_dct_dct_32x32_sub32_add_neon: 18575.5 17157.0 14249.3 12015.1 I.e. in general a very minor overhead for the full subpartition case due to the additional loads and cmps, but a significant speedup for the cases when we only need to process a small part of the actual input data. In common VP9 content in a few inspected clips, 70-90% of the non-dc-only 16x16 and 32x32 IDCTs only have nonzero coefficients in the upper left 8x8 or 16x16 subpartitions respectively. This is cherrypicked from libav commit 9c8bc74c2b40537b0997f646c87c008042d788c2. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:15 +02:00
push {r4-r8,lr}
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vpush {q4-q6}
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
@ Align the stack, allocate a temp buffer
arm: vp9itxfm: Simplify the stack alignment code This is one instruction less for thumb, and only have got 1/2 arm/thumb specific instructions. This is cherrypicked from libav commit e5b0fc170f85b00f7dd0ac514918fb5c95253d39. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:09 +02:00
T mov r7, sp
T and r7, r7, #15
A and r7, sp, #15
add r7, r7, #2048
sub sp, sp, r7
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
mov r4, r0
mov r5, r1
mov r6, r2
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
movrel r12, idct_coeffs
vld1.16 {q0-q1}, [r12,:128]!
vld1.16 {q2-q3}, [r12,:128]
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
cmp r3, #34
ble idct32x32_quarter_add_neon
cmp r3, #135
ble idct32x32_half_add_neon
arm: vp9itxfm: Template the quarter/half idct32 function This reduces the number of lines and reduces the duplication. Also simplify the eob check for the half case. If we are in the half case, we know we at least will need to do the first three slices, we only need to check eob for the fourth one, so we can hardcode the value to check against instead of loading from the min_eob array. Since at most one slice can be skipped in the first pass, we can unroll the loop for filling zeros completely, as it was done for the quarter case before. This allows skipping loading the min_eob pointer when using the quarter/half cases. This is cherrypicked from libav commit 98ee855ae0cc118bd1d20921d6bdb14731832462. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-26 14:02:35 +02:00
movrel r8, min_eob_idct_idct_32 + 2
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.irp i, 0, 4, 8, 12, 16, 20, 24, 28
add r0, sp, #(\i*64)
arm: vp9itxfm: Skip empty slices in the first pass of idct_idct 16x16 and 32x32 This work is sponsored by, and copyright, Google. Previously all subpartitions except the eob=1 (DC) case ran with the same runtime: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub16_add_neon: 3188.1 2435.4 2499.0 1969.0 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.7 16582.3 14207.6 12000.3 By skipping individual 4x16 or 4x32 pixel slices in the first pass, we reduce the runtime of these functions like this: vp9_inv_dct_dct_16x16_sub1_add_neon: 274.6 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2064.0 1534.8 1719.4 1248.7 vp9_inv_dct_dct_16x16_sub4_add_neon: 2135.0 1477.2 1736.3 1249.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2446.7 1828.7 1993.6 1494.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2832.4 2118.3 2266.5 1735.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.7 2475.3 2523.5 1983.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 756.2 456.7 862.0 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10682.2 8190.4 8539.2 6762.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10813.5 8014.9 8518.3 6762.8 vp9_inv_dct_dct_32x32_sub8_add_neon: 11859.6 9313.0 9347.4 7514.5 vp9_inv_dct_dct_32x32_sub12_add_neon: 12946.6 10752.4 10192.2 8280.2 vp9_inv_dct_dct_32x32_sub16_add_neon: 14074.6 11946.5 11001.4 9008.6 vp9_inv_dct_dct_32x32_sub20_add_neon: 15269.9 13662.7 11816.1 9762.6 vp9_inv_dct_dct_32x32_sub24_add_neon: 16327.9 14940.1 12626.7 10516.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17462.7 15776.1 13446.2 11264.7 vp9_inv_dct_dct_32x32_sub32_add_neon: 18575.5 17157.0 14249.3 12015.1 I.e. in general a very minor overhead for the full subpartition case due to the additional loads and cmps, but a significant speedup for the cases when we only need to process a small part of the actual input data. In common VP9 content in a few inspected clips, 70-90% of the non-dc-only 16x16 and 32x32 IDCTs only have nonzero coefficients in the upper left 8x8 or 16x16 subpartitions respectively. This is cherrypicked from libav commit 9c8bc74c2b40537b0997f646c87c008042d788c2. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:15 +02:00
.if \i > 0
ldrh_post r1, r8, #2
cmp r3, r1
it le
movle r1, #(32 - \i)/2
ble 1f
.endif
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
add r2, r6, #(\i*2)
bl idct32_1d_4x32_pass1_neon
.endr
arm: vp9itxfm: Skip empty slices in the first pass of idct_idct 16x16 and 32x32 This work is sponsored by, and copyright, Google. Previously all subpartitions except the eob=1 (DC) case ran with the same runtime: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub16_add_neon: 3188.1 2435.4 2499.0 1969.0 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.7 16582.3 14207.6 12000.3 By skipping individual 4x16 or 4x32 pixel slices in the first pass, we reduce the runtime of these functions like this: vp9_inv_dct_dct_16x16_sub1_add_neon: 274.6 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2064.0 1534.8 1719.4 1248.7 vp9_inv_dct_dct_16x16_sub4_add_neon: 2135.0 1477.2 1736.3 1249.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2446.7 1828.7 1993.6 1494.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2832.4 2118.3 2266.5 1735.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.7 2475.3 2523.5 1983.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 756.2 456.7 862.0 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10682.2 8190.4 8539.2 6762.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10813.5 8014.9 8518.3 6762.8 vp9_inv_dct_dct_32x32_sub8_add_neon: 11859.6 9313.0 9347.4 7514.5 vp9_inv_dct_dct_32x32_sub12_add_neon: 12946.6 10752.4 10192.2 8280.2 vp9_inv_dct_dct_32x32_sub16_add_neon: 14074.6 11946.5 11001.4 9008.6 vp9_inv_dct_dct_32x32_sub20_add_neon: 15269.9 13662.7 11816.1 9762.6 vp9_inv_dct_dct_32x32_sub24_add_neon: 16327.9 14940.1 12626.7 10516.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17462.7 15776.1 13446.2 11264.7 vp9_inv_dct_dct_32x32_sub32_add_neon: 18575.5 17157.0 14249.3 12015.1 I.e. in general a very minor overhead for the full subpartition case due to the additional loads and cmps, but a significant speedup for the cases when we only need to process a small part of the actual input data. In common VP9 content in a few inspected clips, 70-90% of the non-dc-only 16x16 and 32x32 IDCTs only have nonzero coefficients in the upper left 8x8 or 16x16 subpartitions respectively. This is cherrypicked from libav commit 9c8bc74c2b40537b0997f646c87c008042d788c2. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:15 +02:00
b 3f
1:
@ Write zeros to the temp buffer for pass 2
vmov.i16 q14, #0
vmov.i16 q15, #0
2:
subs r1, r1, #1
.rept 4
vst1.16 {q14-q15}, [r0,:128]!
.endr
bne 2b
3:
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
.irp i, 0, 4, 8, 12, 16, 20, 24, 28
add r0, r4, #(\i)
mov r1, r5
add r2, sp, #(\i*2)
bl idct32_1d_4x32_pass2_neon
.endr
arm: vp9itxfm: Simplify the stack alignment code This is one instruction less for thumb, and only have got 1/2 arm/thumb specific instructions. This is cherrypicked from libav commit e5b0fc170f85b00f7dd0ac514918fb5c95253d39. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:09 +02:00
add sp, sp, r7
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vpop {q4-q6}
arm: vp9itxfm: Skip empty slices in the first pass of idct_idct 16x16 and 32x32 This work is sponsored by, and copyright, Google. Previously all subpartitions except the eob=1 (DC) case ran with the same runtime: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub16_add_neon: 3188.1 2435.4 2499.0 1969.0 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.7 16582.3 14207.6 12000.3 By skipping individual 4x16 or 4x32 pixel slices in the first pass, we reduce the runtime of these functions like this: vp9_inv_dct_dct_16x16_sub1_add_neon: 274.6 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2064.0 1534.8 1719.4 1248.7 vp9_inv_dct_dct_16x16_sub4_add_neon: 2135.0 1477.2 1736.3 1249.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2446.7 1828.7 1993.6 1494.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2832.4 2118.3 2266.5 1735.1 vp9_inv_dct_dct_16x16_sub16_add_neon: 3211.7 2475.3 2523.5 1983.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 756.2 456.7 862.0 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10682.2 8190.4 8539.2 6762.5 vp9_inv_dct_dct_32x32_sub4_add_neon: 10813.5 8014.9 8518.3 6762.8 vp9_inv_dct_dct_32x32_sub8_add_neon: 11859.6 9313.0 9347.4 7514.5 vp9_inv_dct_dct_32x32_sub12_add_neon: 12946.6 10752.4 10192.2 8280.2 vp9_inv_dct_dct_32x32_sub16_add_neon: 14074.6 11946.5 11001.4 9008.6 vp9_inv_dct_dct_32x32_sub20_add_neon: 15269.9 13662.7 11816.1 9762.6 vp9_inv_dct_dct_32x32_sub24_add_neon: 16327.9 14940.1 12626.7 10516.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17462.7 15776.1 13446.2 11264.7 vp9_inv_dct_dct_32x32_sub32_add_neon: 18575.5 17157.0 14249.3 12015.1 I.e. in general a very minor overhead for the full subpartition case due to the additional loads and cmps, but a significant speedup for the cases when we only need to process a small part of the actual input data. In common VP9 content in a few inspected clips, 70-90% of the non-dc-only 16x16 and 32x32 IDCTs only have nonzero coefficients in the upper left 8x8 or 16x16 subpartitions respectively. This is cherrypicked from libav commit 9c8bc74c2b40537b0997f646c87c008042d788c2. Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
2017-01-10 00:15:15 +02:00
pop {r4-r8,pc}
arm: vp9: Add NEON itxfm routines This work is sponsored by, and copyright, Google. For the transforms up to 8x8, we can fit all the data (including temporaries) in registers and just do a straightforward transform of all the data. For 16x16, we do a transform of 4x16 pixels in 4 slices, using a temporary buffer. For 32x32, we transform 4x32 pixels at a time, in two steps of 4x16 pixels each. Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_inv_adst_adst_4x4_add_neon: 3.39 5.83 4.17 4.01 vp9_inv_adst_adst_8x8_add_neon: 3.79 4.86 4.23 3.98 vp9_inv_adst_adst_16x16_add_neon: 3.33 4.36 4.11 4.16 vp9_inv_dct_dct_4x4_add_neon: 4.06 6.16 4.59 4.46 vp9_inv_dct_dct_8x8_add_neon: 4.61 6.01 4.98 4.86 vp9_inv_dct_dct_16x16_add_neon: 3.35 3.44 3.36 3.79 vp9_inv_dct_dct_32x32_add_neon: 3.89 3.50 3.79 4.42 vp9_inv_wht_wht_4x4_add_neon: 3.22 5.13 3.53 3.77 Thus, the speedup vs C code is around 3-6x. This is mostly marginally faster than the corresponding routines in libvpx on most cores, tested with their 32x32 idct (compared to vpx_idct32x32_1024_add_neon). These numbers are slightly in libvpx's favour since their version doesn't clear the input buffer like ours do (although the effect of that on the total runtime probably is negligible.) Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_add_neon: 18436.8 16874.1 14235.1 11988.9 libvpx vpx_idct32x32_1024_add_neon 20789.0 13344.3 15049.9 13030.5 Only on the Cortex A8, the libvpx function is faster. On the other cores, ours is slightly faster even though ours has got source block clearing integrated. This is an adapted cherry-pick from libav commits a67ae67083151f2f9595a1f2d17b601da19b939e and 52d196fb30fb6628921b5f1b31e7bd11eb7e1d9a. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:22 +02:00
endfunc
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
arm: vp9itxfm: Template the quarter/half idct32 function This reduces the number of lines and reduces the duplication. Also simplify the eob check for the half case. If we are in the half case, we know we at least will need to do the first three slices, we only need to check eob for the fourth one, so we can hardcode the value to check against instead of loading from the min_eob array. Since at most one slice can be skipped in the first pass, we can unroll the loop for filling zeros completely, as it was done for the quarter case before. This allows skipping loading the min_eob pointer when using the quarter/half cases. This is cherrypicked from libav commit 98ee855ae0cc118bd1d20921d6bdb14731832462. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-26 14:02:35 +02:00
.macro idct32_partial size
function idct32x32_\size\()_add_neon
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.irp i, 0, 4
add r0, sp, #(\i*64)
arm: vp9itxfm: Template the quarter/half idct32 function This reduces the number of lines and reduces the duplication. Also simplify the eob check for the half case. If we are in the half case, we know we at least will need to do the first three slices, we only need to check eob for the fourth one, so we can hardcode the value to check against instead of loading from the min_eob array. Since at most one slice can be skipped in the first pass, we can unroll the loop for filling zeros completely, as it was done for the quarter case before. This allows skipping loading the min_eob pointer when using the quarter/half cases. This is cherrypicked from libav commit 98ee855ae0cc118bd1d20921d6bdb14731832462. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-26 14:02:35 +02:00
.ifc \size,quarter
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.if \i == 4
cmp r3, #9
ble 1f
arm: vp9itxfm: Template the quarter/half idct32 function This reduces the number of lines and reduces the duplication. Also simplify the eob check for the half case. If we are in the half case, we know we at least will need to do the first three slices, we only need to check eob for the fourth one, so we can hardcode the value to check against instead of loading from the min_eob array. Since at most one slice can be skipped in the first pass, we can unroll the loop for filling zeros completely, as it was done for the quarter case before. This allows skipping loading the min_eob pointer when using the quarter/half cases. This is cherrypicked from libav commit 98ee855ae0cc118bd1d20921d6bdb14731832462. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-26 14:02:35 +02:00
.endif
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.endif
add r2, r6, #(\i*2)
arm: vp9itxfm: Template the quarter/half idct32 function This reduces the number of lines and reduces the duplication. Also simplify the eob check for the half case. If we are in the half case, we know we at least will need to do the first three slices, we only need to check eob for the fourth one, so we can hardcode the value to check against instead of loading from the min_eob array. Since at most one slice can be skipped in the first pass, we can unroll the loop for filling zeros completely, as it was done for the quarter case before. This allows skipping loading the min_eob pointer when using the quarter/half cases. This is cherrypicked from libav commit 98ee855ae0cc118bd1d20921d6bdb14731832462. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-26 14:02:35 +02:00
bl idct32_1d_4x32_pass1_\size\()_neon
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.endr
arm: vp9itxfm: Template the quarter/half idct32 function This reduces the number of lines and reduces the duplication. Also simplify the eob check for the half case. If we are in the half case, we know we at least will need to do the first three slices, we only need to check eob for the fourth one, so we can hardcode the value to check against instead of loading from the min_eob array. Since at most one slice can be skipped in the first pass, we can unroll the loop for filling zeros completely, as it was done for the quarter case before. This allows skipping loading the min_eob pointer when using the quarter/half cases. This is cherrypicked from libav commit 98ee855ae0cc118bd1d20921d6bdb14731832462. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-26 14:02:35 +02:00
.ifc \size,half
.irp i, 8, 12
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
add r0, sp, #(\i*64)
arm: vp9itxfm: Template the quarter/half idct32 function This reduces the number of lines and reduces the duplication. Also simplify the eob check for the half case. If we are in the half case, we know we at least will need to do the first three slices, we only need to check eob for the fourth one, so we can hardcode the value to check against instead of loading from the min_eob array. Since at most one slice can be skipped in the first pass, we can unroll the loop for filling zeros completely, as it was done for the quarter case before. This allows skipping loading the min_eob pointer when using the quarter/half cases. This is cherrypicked from libav commit 98ee855ae0cc118bd1d20921d6bdb14731832462. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-26 14:02:35 +02:00
.if \i == 12
cmp r3, #70
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
ble 1f
.endif
add r2, r6, #(\i*2)
arm: vp9itxfm: Template the quarter/half idct32 function This reduces the number of lines and reduces the duplication. Also simplify the eob check for the half case. If we are in the half case, we know we at least will need to do the first three slices, we only need to check eob for the fourth one, so we can hardcode the value to check against instead of loading from the min_eob array. Since at most one slice can be skipped in the first pass, we can unroll the loop for filling zeros completely, as it was done for the quarter case before. This allows skipping loading the min_eob pointer when using the quarter/half cases. This is cherrypicked from libav commit 98ee855ae0cc118bd1d20921d6bdb14731832462. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-26 14:02:35 +02:00
bl idct32_1d_4x32_pass1_\size\()_neon
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.endr
arm: vp9itxfm: Template the quarter/half idct32 function This reduces the number of lines and reduces the duplication. Also simplify the eob check for the half case. If we are in the half case, we know we at least will need to do the first three slices, we only need to check eob for the fourth one, so we can hardcode the value to check against instead of loading from the min_eob array. Since at most one slice can be skipped in the first pass, we can unroll the loop for filling zeros completely, as it was done for the quarter case before. This allows skipping loading the min_eob pointer when using the quarter/half cases. This is cherrypicked from libav commit 98ee855ae0cc118bd1d20921d6bdb14731832462. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-26 14:02:35 +02:00
.endif
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
b 3f
1:
@ Write zeros to the temp buffer for pass 2
vmov.i16 q14, #0
vmov.i16 q15, #0
arm: vp9itxfm: Template the quarter/half idct32 function This reduces the number of lines and reduces the duplication. Also simplify the eob check for the half case. If we are in the half case, we know we at least will need to do the first three slices, we only need to check eob for the fourth one, so we can hardcode the value to check against instead of loading from the min_eob array. Since at most one slice can be skipped in the first pass, we can unroll the loop for filling zeros completely, as it was done for the quarter case before. This allows skipping loading the min_eob pointer when using the quarter/half cases. This is cherrypicked from libav commit 98ee855ae0cc118bd1d20921d6bdb14731832462. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-26 14:02:35 +02:00
.rept 8
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
vst1.16 {q14-q15}, [r0,:128]!
.endr
arm: vp9itxfm: Template the quarter/half idct32 function This reduces the number of lines and reduces the duplication. Also simplify the eob check for the half case. If we are in the half case, we know we at least will need to do the first three slices, we only need to check eob for the fourth one, so we can hardcode the value to check against instead of loading from the min_eob array. Since at most one slice can be skipped in the first pass, we can unroll the loop for filling zeros completely, as it was done for the quarter case before. This allows skipping loading the min_eob pointer when using the quarter/half cases. This is cherrypicked from libav commit 98ee855ae0cc118bd1d20921d6bdb14731832462. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-26 14:02:35 +02:00
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
3:
.irp i, 0, 4, 8, 12, 16, 20, 24, 28
add r0, r4, #(\i)
mov r1, r5
add r2, sp, #(\i*2)
arm: vp9itxfm: Template the quarter/half idct32 function This reduces the number of lines and reduces the duplication. Also simplify the eob check for the half case. If we are in the half case, we know we at least will need to do the first three slices, we only need to check eob for the fourth one, so we can hardcode the value to check against instead of loading from the min_eob array. Since at most one slice can be skipped in the first pass, we can unroll the loop for filling zeros completely, as it was done for the quarter case before. This allows skipping loading the min_eob pointer when using the quarter/half cases. This is cherrypicked from libav commit 98ee855ae0cc118bd1d20921d6bdb14731832462. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-26 14:02:35 +02:00
bl idct32_1d_4x32_pass2_\size\()_neon
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
.endr
add sp, sp, r7
arm: vp9itxfm: Avoid reloading the idct32 coefficients The idct32x32 function actually pushed q4-q7 onto the stack even though it didn't clobber them; there are plenty of registers that can be used to allow keeping all the idct coefficients in registers without having to reload different subsets of them at different stages in the transform. Since the idct16 core transform avoids clobbering q4-q7 (but clobbers q2-q3 instead, to avoid needing to back up and restore q4-q7 at all in the idct16 function), and the lanewise vmul needs a register in the q0-q3 range, we move the stored coefficients from q2-q3 into q4-q5 while doing idct16. While keeping these coefficients in registers, we still can skip pushing q7. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_32x32_sub32_add_neon: 18553.8 17182.7 14303.3 12089.7 After: vp9_inv_dct_dct_32x32_sub32_add_neon: 18470.3 16717.7 14173.6 11860.8 This is cherrypicked from libav commit 402546a17233a8815307df9e14ff88cd70424537. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-01-02 22:50:38 +02:00
vpop {q4-q6}
arm: vp9itxfm: Do a simpler half/quarter idct16/idct32 when possible This work is sponsored by, and copyright, Google. This avoids loading and calculating coefficients that we know will be zero, and avoids filling the temp buffer with zeros in places where we know the second pass won't read. This gives a pretty substantial speedup for the smaller subpartitions. The code size increases from 12388 bytes to 19784 bytes. The idct16/32_end macros are moved above the individual functions; the instructions themselves are unchanged, but since new functions are added at the same place where the code is moved from, the diff looks rather messy. Before: Cortex A7 A8 A9 A53 vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 212.0 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 2102.1 1521.7 1736.2 1265.8 vp9_inv_dct_dct_16x16_sub4_add_neon: 2104.5 1533.0 1736.6 1265.5 vp9_inv_dct_dct_16x16_sub8_add_neon: 2484.8 1828.7 2014.4 1506.5 vp9_inv_dct_dct_16x16_sub12_add_neon: 2851.2 2117.8 2294.8 1753.2 vp9_inv_dct_dct_16x16_sub16_add_neon: 3239.4 2408.3 2543.5 1994.9 vp9_inv_dct_dct_32x32_sub1_add_neon: 758.3 456.7 864.5 553.9 vp9_inv_dct_dct_32x32_sub2_add_neon: 10776.7 7949.8 8567.7 6819.7 vp9_inv_dct_dct_32x32_sub4_add_neon: 10865.6 8131.5 8589.6 6816.3 vp9_inv_dct_dct_32x32_sub8_add_neon: 12053.9 9271.3 9387.7 7564.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 13328.3 10463.2 10217.0 8321.3 vp9_inv_dct_dct_32x32_sub16_add_neon: 14176.4 11509.5 11018.7 9062.3 vp9_inv_dct_dct_32x32_sub20_add_neon: 15301.5 12999.9 11855.1 9828.2 vp9_inv_dct_dct_32x32_sub24_add_neon: 16482.7 14931.5 12650.1 10575.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17589.5 15811.9 13482.8 11333.4 vp9_inv_dct_dct_32x32_sub32_add_neon: 18696.2 17049.2 14355.6 12089.7 After: vp9_inv_dct_dct_16x16_sub1_add_neon: 273.0 189.5 211.7 235.8 vp9_inv_dct_dct_16x16_sub2_add_neon: 1203.5 998.2 1035.3 763.0 vp9_inv_dct_dct_16x16_sub4_add_neon: 1203.5 998.1 1035.5 760.8 vp9_inv_dct_dct_16x16_sub8_add_neon: 1926.1 1610.6 1722.1 1271.7 vp9_inv_dct_dct_16x16_sub12_add_neon: 2873.2 2129.7 2285.1 1757.3 vp9_inv_dct_dct_16x16_sub16_add_neon: 3221.4 2520.3 2557.6 2002.1 vp9_inv_dct_dct_32x32_sub1_add_neon: 753.0 457.5 866.6 554.6 vp9_inv_dct_dct_32x32_sub2_add_neon: 7554.6 5652.4 6048.4 4920.2 vp9_inv_dct_dct_32x32_sub4_add_neon: 7549.9 5685.0 6046.9 4925.7 vp9_inv_dct_dct_32x32_sub8_add_neon: 8336.9 6704.5 6604.0 5478.0 vp9_inv_dct_dct_32x32_sub12_add_neon: 10914.0 9777.2 9240.4 7416.9 vp9_inv_dct_dct_32x32_sub16_add_neon: 11859.2 11223.3 9966.3 8095.1 vp9_inv_dct_dct_32x32_sub20_add_neon: 15237.1 13029.4 11838.3 9829.4 vp9_inv_dct_dct_32x32_sub24_add_neon: 16293.2 14379.8 12644.9 10572.0 vp9_inv_dct_dct_32x32_sub28_add_neon: 17424.3 15734.7 13473.0 11326.9 vp9_inv_dct_dct_32x32_sub32_add_neon: 18531.3 17457.0 14298.6 12080.0 This is cherrypicked from libav commit 5eb5aec475aabc884d083566f902876ecbc072cb. Signed-off-by: Martin Storsjö <martin@martin.st>
2016-11-22 11:07:38 +02:00
pop {r4-r8,pc}
endfunc
arm: vp9itxfm: Template the quarter/half idct32 function This reduces the number of lines and reduces the duplication. Also simplify the eob check for the half case. If we are in the half case, we know we at least will need to do the first three slices, we only need to check eob for the fourth one, so we can hardcode the value to check against instead of loading from the min_eob array. Since at most one slice can be skipped in the first pass, we can unroll the loop for filling zeros completely, as it was done for the quarter case before. This allows skipping loading the min_eob pointer when using the quarter/half cases. This is cherrypicked from libav commit 98ee855ae0cc118bd1d20921d6bdb14731832462. Signed-off-by: Martin Storsjö <martin@martin.st>
2017-02-26 14:02:35 +02:00
.endm
idct32_partial quarter
idct32_partial half
Reference in New Issue Copy Permalink