1
0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2024-12-23 12:43:46 +02:00
FFmpeg/libavcodec/arm/vp9itxfm_neon.S

1203 lines
45 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
.short 11585, 6270, 15137, 0
iadst4_coeffs:
.short 5283, 15212, 9929, 13377
endconst
const iadst8_coeffs, align=4
.short 16305, 1606, 14449, 7723, 10394, 12665, 4756, 15679
idct_coeffs:
.short 11585, 6270, 15137, 3196, 16069, 13623, 9102, 1606
.short 16305, 12665, 10394, 7723, 14449, 15679, 4756, 0
.short 804, 16364, 12140, 11003, 7005, 14811, 15426, 5520
.short 3981, 15893, 14053, 8423, 9760, 13160, 16207, 2404
endconst
const iadst16_coeffs, align=4
.short 16364, 804, 15893, 3981, 14811, 7005, 13160, 9760
.short 11003, 12140, 8423, 14053, 5520, 15426, 2404, 16207
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
@ 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
@ 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
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
@ 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
vmull.s16 q13, \c1, d0[2]
vmull.s16 q11, \c1, d0[1]
vadd.i16 d16, \c0, \c2
vsub.i16 d17, \c0, \c2
vmlal.s16 q13, \c3, d0[1]
vmull.s16 q9, d16, d0[0]
vmull.s16 q10, d17, d0[0]
vmlsl.s16 q11, \c3, d0[2]
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
dmbutterfly d20, d21, d28, d29, d0[1], d0[2], q2, q3, q4, q5 @ q10 = t2a, q14 = t3a
dmbutterfly d18, d19, d30, d31, d0[3], d1[0], q2, q3, q4, q5 @ q9 = t4a, q15 = t7a
dmbutterfly d26, d27, d22, d23, d1[1], d1[2], q2, q3, q4, q5 @ q13 = t5a, q11 = t6a
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
dmbutterfly_l q10, q11, q5, q7, d4, d5, d6, d7, d0[1], d0[2] @ q10,q11 = t5a, q5,q7 = t4a
dmbutterfly_l q2, q3, q13, q14, d12, d13, d8, d9, d0[2], d0[1] @ q2,q3 = t6a, q13,q14 = t7a
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}
vld1.16 {q0}, [r12,:128]
.else
movrel r12, iadst8_coeffs
vld1.16 {q1}, [r12,:128]!
vpush {q4-q7}
vld1.16 {q0}, [r12,:128]
.endif
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]
vmov.i16 q2, #0
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
mov r12, #16
1:
@ Loop to add the constant from q8 into all 16x16 outputs
vld1.8 {q3}, [r0,:128]
vaddw.u8 q10, q8, d6
vaddw.u8 q11, q8, d7
vqmovun.s16 d6, q10
vqmovun.s16 d7, q11
vst1.8 {q3}, [r0,:128], r1
subs r12, r12, #1
bne 1b
bx lr
endfunc
.ltorg
.macro idct16
mbutterfly0 d16, d24, d16, d24, d4, d6, q2, q3 @ d16 = t0a, d24 = t1a
mbutterfly d20, d28, d0[1], d0[2], q2, q3 @ d20 = t2a, d28 = t3a
mbutterfly d18, d30, d0[3], d1[0], q2, q3 @ d18 = t4a, d30 = t7a
mbutterfly d26, d22, d1[1], d1[2], q2, q3 @ d26 = t5a, d22 = t6a
mbutterfly d17, d31, d1[3], d2[0], q2, q3 @ d17 = t8a, d31 = t15a
mbutterfly d25, d23, d2[1], d2[2], q2, q3 @ d25 = t9a, d23 = t14a
mbutterfly d21, d27, d2[3], d3[0], q2, q3 @ d21 = t10a, d27 = t13a
mbutterfly d29, d19, d3[1], d3[2], q2, q3 @ d29 = t11a, d19 = t12a
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
mbutterfly d23, d25, d0[1], d0[2], q9, q15 @ d23 = t9a, d25 = t14a
mbutterfly d27, d21, d0[1], d0[2], q9, q15, neg=1 @ d27 = t13a, d21 = t10a
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]
.endm
.macro iadst16
movrel r12, iadst16_coeffs
vld1.16 {q0-q1}, [r12,:128]
mbutterfly_l q3, q2, d31, d16, d0[1], d0[0] @ q3 = t1, q2 = t0
mbutterfly_l q5, q4, d23, d24, d2[1], d2[0] @ q5 = t9, q4 = t8
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
mbutterfly_l q3, q2, d21, d26, d2[3], d2[2] @ q3 = t11, q2 = t10
butterfly_n d29, d26, q7, q3, q4, q3 @ d29 = t3a, d26 = t11a
mbutterfly_l q5, q4, d27, d20, d1[1], d1[0] @ q5 = t5, q4 = t4
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
mbutterfly_l q3, q2, d25, d22, d1[3], d1[2] @ q3 = t7, q2 = t6
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
mbutterfly_l q7, q6, d23, d24, d0[3], d1[0] @ q7 = t9, q6 = t8
butterfly_n d25, d17, q2, q4, q3, q4 @ d25 = t6a, d17 = t14a
mbutterfly_l q2, q3, d28, d19, d1[0], d0[3] @ q2 = t12, q3 = t13
butterfly_n d23, d19, q6, q2, q4, q2 @ d23 = t8a, d19 = t12a
mbutterfly_l q5, q4, d21, d26, d1[1], d1[2] @ q5 = t11, q4 = t10
butterfly_r d4, d27, d16, d27 @ d4 = t4, d27 = t0
butterfly_n d24, d28, q7, q3, q6, q3 @ d24 = t9a, d28 = t13a
mbutterfly_l q6, q7, d30, d17, d1[2], d1[1] @ q6 = t14, q7 = t15
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
mbutterfly_l q5, q4, d19, d28, d0[1], d0[2] @ q5 = t13, q4 = t12
mbutterfly_l q6, q7, d30, d17, d0[2], d0[1] @ q6 = t14, q7 = t15
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]
mbutterfly_l q5, q4, d4, d5, d0[1], d0[2] @ q5 = t5a, q4 = t4a
mbutterfly_l q6, q7, d7, d6, d0[2], d0[1] @ q6 = t6a, q7 = t7a
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
.endm
.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
mov r12, #32
vmov.s16 q2, #0
.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
\txfm\()16
@ 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
bx lr
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
.irp i, 20, 24, 28
vst1.16 {d\i}, [r0,:64]!
.endr
add r0, r0, #8
.irp i, 21, 25, 29
vst1.16 {d\i}, [r0,:64]!
.endr
add r0, r0, #8
.irp i, 22, 26, 30
vst1.16 {d\i}, [r0,:64]!
.endr
add r0, r0, #8
.irp i, 23, 27, 31
vst1.16 {d\i}, [r0,:64]!
.endr
vmov d28, d16
vmov d29, d17
vmov d30, d18
vmov d31, d19
bx lr
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
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
\txfm\()16
.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
load_add_store q8, q9, q10, q11
load_add_store q12, q13, q14, q15
.purgem load_add_store
bx lr
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}
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
.else
movrel r8, min_eob_idct_idct_16 + 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
.endif
@ Align the stack, allocate a temp buffer
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
.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
.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
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
function idct32x32_dc_add_neon
movrel r12, idct_coeffs
vld1.16 {d0}, [r12,:64]
vmov.i16 q2, #0
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
mov r12, #32
1:
@ Loop to add the constant from q8 into all 32x32 outputs
vld1.8 {q2-q3}, [r0,:128]
vaddw.u8 q10, q8, d4
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
vqmovun.s16 d7, q13
vst1.8 {q2-q3}, [r0,:128], r1
subs r12, r12, #1
bne 1b
bx lr
endfunc
.macro idct32_odd
movrel r12, idct_coeffs
add r12, r12, #32
vld1.16 {q0-q1}, [r12,:128]
mbutterfly d16, d31, d0[0], d0[1], q2, q3 @ d16 = t16a, d31 = t31a
mbutterfly d24, d23, d0[2], d0[3], q2, q3 @ d24 = t17a, d23 = t30a
mbutterfly d20, d27, d1[0], d1[1], q2, q3 @ d20 = t18a, d27 = t29a
mbutterfly d28, d19, d1[2], d1[3], q2, q3 @ d28 = t19a, d19 = t28a
mbutterfly d18, d29, d2[0], d2[1], q2, q3 @ d18 = t20a, d29 = t27a
mbutterfly d26, d21, d2[2], d2[3], q2, q3 @ d26 = t21a, d21 = t26a
mbutterfly d22, d25, d3[0], d3[1], q2, q3 @ d22 = t22a, d25 = t25a
mbutterfly d30, d17, d3[2], d3[3], q2, q3 @ d30 = t23a, d17 = t24a
sub r12, r12, #32
vld1.16 {q0}, [r12,:128]
butterfly d4, d24, d16, d24 @ d4 = t16, d24 = t17
butterfly d5, d20, d28, d20 @ d5 = t19, d20 = t18
butterfly d6, d26, d18, d26 @ d6 = t20, d26 = t21
butterfly d7, d22, d30, d22 @ d7 = t23, d22 = t22
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
mbutterfly d23, d24, d0[3], d1[0], q8, q9 @ d23 = t17a, d24 = t30a
mbutterfly d27, d20, d0[3], d1[0], q8, q9, neg=1 @ d27 = t29a, d20 = t18a
mbutterfly d21, d26, d1[1], d1[2], q8, q9 @ d21 = t21a, d26 = t26a
mbutterfly d25, d22, d1[1], d1[2], q8, q9, neg=1 @ d25 = t25a, d22 = t22a
butterfly d16, d5, d4, d5 @ d16 = t16a, d5 = t19a
butterfly d17, d20, d23, d20 @ d17 = t17, d20 = t18
butterfly d18, d6, d7, d6 @ d18 = t23a, d6 = t20a
butterfly d19, d21, d22, d21 @ d19 = t22, d21 = t21
butterfly d4, d28, d28, d30 @ d4 = t24a, d28 = t27a
butterfly d23, d26, d25, d26 @ d23 = t25, d26 = t26
butterfly d7, d29, d29, d31 @ d7 = t31a, d29 = t28a
butterfly d22, d27, d24, d27 @ d22 = t30, d27 = t29
mbutterfly d27, d20, d0[1], d0[2], q12, q15 @ d27 = t18a, d20 = t29a
mbutterfly d29, d5, d0[1], d0[2], q12, q15 @ d29 = t19, d5 = t28
mbutterfly d28, d6, d0[1], d0[2], q12, q15, neg=1 @ d28 = t27, d6 = t20
mbutterfly d26, d21, d0[1], d0[2], q12, q15, neg=1 @ d26 = t26a, d21 = t21a
butterfly d31, d24, d7, d4 @ d31 = t31, d24 = t24
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
butterfly_r d27, d28, d5, d28 @ d27 = t27a, d28 = t28a
butterfly d4, d26, d20, d26 @ d4 = t29, d26 = t26
butterfly d19, d20, d29, d6 @ d19 = t19a, d20 = t20
vmov d29, d4 @ d29 = t29
mbutterfly0 d27, d20, d27, d20, d4, d6, q2, q3 @ d27 = t27, d20 = t20
mbutterfly0 d26, d21, d26, d21, d4, d6, q2, q3 @ d26 = t26a, d21 = t21a
mbutterfly0 d25, d22, d25, d22, d4, d6, q2, q3 @ d25 = t25, d22 = t22
mbutterfly0 d24, d23, d24, d23, d4, d6, q2, q3 @ d24 = t24a, d23 = t23a
.endm
@ 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
function idct32_1d_4x32_pass1_neon
movrel r12, idct_coeffs
vld1.16 {q0-q1}, [r12,:128]
@ Double stride of the input, since we only read every other line
mov r12, #128
vmov.s16 d4, #0
@ d16 = IN(0), d17 = IN(2) ... d31 = IN(30)
.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
idct16
@ 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
sub r2, r2, r12, lsl #4
add r2, r2, #64
vmov.s16 d4, #0
@ d16 = IN(1), d17 = IN(3) ... d31 = IN(31)
.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
idct32_odd
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
vld1.16 {d4}, [r0,:64]
vadd.s16 d4, d4, d\i
vst1.16 {d4}, [r0,:64]!
vrev64.16 d\i, d\i
.endr
.irp i, \d, \c, \b, \a
vld1.16 {d4}, [r0,:64]
vsub.s16 d4, d4, d\i
vst1.16 {d4}, [r0,:64]!
.endr
.endm
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
bx lr
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)
function idct32_1d_4x32_pass2_neon
movrel r12, idct_coeffs
vld1.16 {q0-q1}, [r12,:128]
mov r12, #128
@ d16 = IN(0), d17 = IN(2) ... d31 = IN(30)
.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
idct16
.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)
.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
sub r2, r2, #64
idct32_odd
mov r12, #128
.macro load_acc_store a, b, c, d, neg=0
vld1.16 {d4}, [r2,:64], r12
vld1.16 {d5}, [r2,:64], r12
.if \neg == 0
vadd.s16 d4, d4, d\a
vld1.16 {d6}, [r2,:64], r12
vadd.s16 d5, d5, d\b
vld1.16 {d7}, [r2,:64], r12
vadd.s16 d6, d6, d\c
vadd.s16 d7, d7, d\d
.else
vsub.s16 d4, d4, d\a
vld1.16 {d6}, [r2,:64], r12
vsub.s16 d5, d5, d\b
vld1.16 {d7}, [r2,:64], r12
vsub.s16 d6, d6, d\c
vsub.s16 d7, d7, d\d
.endif
vld1.32 {d2[]}, [r0,:32], r1
vld1.32 {d2[1]}, [r0,:32], r1
vrshr.s16 q2, q2, #6
vld1.32 {d3[]}, [r0,:32], r1
vrshr.s16 q3, q3, #6
vld1.32 {d3[1]}, [r0,:32], r1
sub r0, r0, r1, lsl #2
vaddw.u8 q2, q2, d2
vaddw.u8 q3, q3, d3
vqmovun.s16 d4, q2
vqmovun.s16 d5, q3
vst1.32 {d4[0]}, [r0,:32], r1
vst1.32 {d4[1]}, [r0,:32], r1
vst1.32 {d5[0]}, [r0,:32], r1
vst1.32 {d5[1]}, [r0,:32], r1
.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
bx lr
endfunc
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: 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
vpush {q4-q7}
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
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
@ Align the stack, allocate a temp buffer
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
.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
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
vpop {q4-q7}
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