1
0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2024-12-07 11:13:41 +02:00
FFmpeg/libavcodec/arm/vp9mc_neon.S

721 lines
24 KiB
ArmAsm
Raw Normal View History

arm: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. The filter coefficients are signed values, where the product of the multiplication with one individual filter coefficient doesn't overflow a 16 bit signed value (the largest filter coefficient is 127). But when the products are accumulated, the resulting sum can overflow the 16 bit signed range. Instead of accumulating in 32 bit, we accumulate the largest product (either index 3 or 4) last with a saturated addition. (The VP8 MC asm does something similar, but slightly simpler, by accumulating each half of the filter separately. In the VP9 MC filters, each half of the filter can also overflow though, so the largest component has to be handled individually.) Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_avg4_neon: 1.71 1.15 1.42 1.49 vp9_avg8_neon: 2.51 3.63 3.14 2.58 vp9_avg16_neon: 2.95 6.76 3.01 2.84 vp9_avg32_neon: 3.29 6.64 2.85 3.00 vp9_avg64_neon: 3.47 6.67 3.14 2.80 vp9_avg_8tap_smooth_4h_neon: 3.22 4.73 2.76 4.67 vp9_avg_8tap_smooth_4hv_neon: 3.67 4.76 3.28 4.71 vp9_avg_8tap_smooth_4v_neon: 5.52 7.60 4.60 6.31 vp9_avg_8tap_smooth_8h_neon: 6.22 9.04 5.12 9.32 vp9_avg_8tap_smooth_8hv_neon: 6.38 8.21 5.72 8.17 vp9_avg_8tap_smooth_8v_neon: 9.22 12.66 8.15 11.10 vp9_avg_8tap_smooth_64h_neon: 7.02 10.23 5.54 11.58 vp9_avg_8tap_smooth_64hv_neon: 6.76 9.46 5.93 9.40 vp9_avg_8tap_smooth_64v_neon: 10.76 14.13 9.46 13.37 vp9_put4_neon: 1.11 1.47 1.00 1.21 vp9_put8_neon: 1.23 2.17 1.94 1.48 vp9_put16_neon: 1.63 4.02 1.73 1.97 vp9_put32_neon: 1.56 4.92 2.00 1.96 vp9_put64_neon: 2.10 5.28 2.03 2.35 vp9_put_8tap_smooth_4h_neon: 3.11 4.35 2.63 4.35 vp9_put_8tap_smooth_4hv_neon: 3.67 4.69 3.25 4.71 vp9_put_8tap_smooth_4v_neon: 5.45 7.27 4.49 6.52 vp9_put_8tap_smooth_8h_neon: 5.97 8.18 4.81 8.56 vp9_put_8tap_smooth_8hv_neon: 6.39 7.90 5.64 8.15 vp9_put_8tap_smooth_8v_neon: 9.03 11.84 8.07 11.51 vp9_put_8tap_smooth_64h_neon: 6.78 9.48 4.88 10.89 vp9_put_8tap_smooth_64hv_neon: 6.99 8.87 5.94 9.56 vp9_put_8tap_smooth_64v_neon: 10.69 13.30 9.43 14.34 For the larger 8tap filters, the speedup vs C code is around 5-14x. This is significantly faster than libvpx's implementation of the same functions, at least when comparing the put_8tap_smooth_64 functions (compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from libvpx). Absolute runtimes from checkasm: Cortex A7 A8 A9 A53 vp9_put_8tap_smooth_64h_neon: 20150.3 14489.4 19733.6 10863.7 libvpx vpx_convolve8_horiz_neon: 52623.3 19736.4 21907.7 25027.7 vp9_put_8tap_smooth_64v_neon: 14455.0 12303.9 13746.4 9628.9 libvpx vpx_convolve8_vert_neon: 42090.0 17706.2 17659.9 16941.2 Thus, on the A9, the horizontal filter is only marginally faster than libvpx, while our version is significantly faster on the other cores, and the vertical filter is significantly faster on all cores. The difference is especially large on the A7. The libvpx implementation does the accumulation in 32 bit, which probably explains most of the differences. This is an adapted cherry-pick from libav commits ffbd1d2b0002576ef0d976a41ff959c635373fdc, 392caa65df3efa8b2d48a80f08a6af4892c61c08, 557c1675cf0e803b2fee43b4c8b58433842c84d0 and 11623217e3c9b859daee544e31acdd0821b61039. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:21 +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"
@ All public functions in this file have the following signature:
@ typedef void (*vp9_mc_func)(uint8_t *dst, ptrdiff_t dst_stride,
@ const uint8_t *ref, ptrdiff_t ref_stride,
@ int h, int mx, int my);
function ff_vp9_copy64_neon, export=1
ldr r12, [sp]
sub r1, r1, #32
sub r3, r3, #32
1:
vld1.8 {q0, q1}, [r2]!
vst1.8 {q0, q1}, [r0, :128]!
vld1.8 {q2, q3}, [r2], r3
subs r12, r12, #1
vst1.8 {q2, q3}, [r0, :128], r1
bne 1b
bx lr
endfunc
function ff_vp9_avg64_neon, export=1
push {lr}
ldr r12, [sp, #4]
sub r1, r1, #32
sub r3, r3, #32
mov lr, r0
1:
vld1.8 {q8, q9}, [r2]!
vld1.8 {q0, q1}, [r0, :128]!
vld1.8 {q10, q11}, [r2], r3
vrhadd.u8 q0, q0, q8
vld1.8 {q2, q3}, [r0, :128], r1
vrhadd.u8 q1, q1, q9
vrhadd.u8 q2, q2, q10
vst1.8 {q0, q1}, [lr, :128]!
vrhadd.u8 q3, q3, q11
vst1.8 {q2, q3}, [lr, :128], r1
subs r12, r12, #1
bne 1b
pop {pc}
endfunc
function ff_vp9_copy32_neon, export=1
ldr r12, [sp]
1:
vld1.8 {q0, q1}, [r2], r3
subs r12, r12, #1
vst1.8 {q0, q1}, [r0, :128], r1
bne 1b
bx lr
endfunc
function ff_vp9_avg32_neon, export=1
ldr r12, [sp]
1:
vld1.8 {q2, q3}, [r2], r3
vld1.8 {q0, q1}, [r0, :128]
vrhadd.u8 q0, q0, q2
vrhadd.u8 q1, q1, q3
subs r12, r12, #1
vst1.8 {q0, q1}, [r0, :128], r1
arm: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. The filter coefficients are signed values, where the product of the multiplication with one individual filter coefficient doesn't overflow a 16 bit signed value (the largest filter coefficient is 127). But when the products are accumulated, the resulting sum can overflow the 16 bit signed range. Instead of accumulating in 32 bit, we accumulate the largest product (either index 3 or 4) last with a saturated addition. (The VP8 MC asm does something similar, but slightly simpler, by accumulating each half of the filter separately. In the VP9 MC filters, each half of the filter can also overflow though, so the largest component has to be handled individually.) Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_avg4_neon: 1.71 1.15 1.42 1.49 vp9_avg8_neon: 2.51 3.63 3.14 2.58 vp9_avg16_neon: 2.95 6.76 3.01 2.84 vp9_avg32_neon: 3.29 6.64 2.85 3.00 vp9_avg64_neon: 3.47 6.67 3.14 2.80 vp9_avg_8tap_smooth_4h_neon: 3.22 4.73 2.76 4.67 vp9_avg_8tap_smooth_4hv_neon: 3.67 4.76 3.28 4.71 vp9_avg_8tap_smooth_4v_neon: 5.52 7.60 4.60 6.31 vp9_avg_8tap_smooth_8h_neon: 6.22 9.04 5.12 9.32 vp9_avg_8tap_smooth_8hv_neon: 6.38 8.21 5.72 8.17 vp9_avg_8tap_smooth_8v_neon: 9.22 12.66 8.15 11.10 vp9_avg_8tap_smooth_64h_neon: 7.02 10.23 5.54 11.58 vp9_avg_8tap_smooth_64hv_neon: 6.76 9.46 5.93 9.40 vp9_avg_8tap_smooth_64v_neon: 10.76 14.13 9.46 13.37 vp9_put4_neon: 1.11 1.47 1.00 1.21 vp9_put8_neon: 1.23 2.17 1.94 1.48 vp9_put16_neon: 1.63 4.02 1.73 1.97 vp9_put32_neon: 1.56 4.92 2.00 1.96 vp9_put64_neon: 2.10 5.28 2.03 2.35 vp9_put_8tap_smooth_4h_neon: 3.11 4.35 2.63 4.35 vp9_put_8tap_smooth_4hv_neon: 3.67 4.69 3.25 4.71 vp9_put_8tap_smooth_4v_neon: 5.45 7.27 4.49 6.52 vp9_put_8tap_smooth_8h_neon: 5.97 8.18 4.81 8.56 vp9_put_8tap_smooth_8hv_neon: 6.39 7.90 5.64 8.15 vp9_put_8tap_smooth_8v_neon: 9.03 11.84 8.07 11.51 vp9_put_8tap_smooth_64h_neon: 6.78 9.48 4.88 10.89 vp9_put_8tap_smooth_64hv_neon: 6.99 8.87 5.94 9.56 vp9_put_8tap_smooth_64v_neon: 10.69 13.30 9.43 14.34 For the larger 8tap filters, the speedup vs C code is around 5-14x. This is significantly faster than libvpx's implementation of the same functions, at least when comparing the put_8tap_smooth_64 functions (compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from libvpx). Absolute runtimes from checkasm: Cortex A7 A8 A9 A53 vp9_put_8tap_smooth_64h_neon: 20150.3 14489.4 19733.6 10863.7 libvpx vpx_convolve8_horiz_neon: 52623.3 19736.4 21907.7 25027.7 vp9_put_8tap_smooth_64v_neon: 14455.0 12303.9 13746.4 9628.9 libvpx vpx_convolve8_vert_neon: 42090.0 17706.2 17659.9 16941.2 Thus, on the A9, the horizontal filter is only marginally faster than libvpx, while our version is significantly faster on the other cores, and the vertical filter is significantly faster on all cores. The difference is especially large on the A7. The libvpx implementation does the accumulation in 32 bit, which probably explains most of the differences. This is an adapted cherry-pick from libav commits ffbd1d2b0002576ef0d976a41ff959c635373fdc, 392caa65df3efa8b2d48a80f08a6af4892c61c08, 557c1675cf0e803b2fee43b4c8b58433842c84d0 and 11623217e3c9b859daee544e31acdd0821b61039. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:21 +02:00
bne 1b
bx lr
endfunc
function ff_vp9_copy16_neon, export=1
push {r4,lr}
ldr r12, [sp, #8]
add r4, r0, r1
add lr, r2, r3
add r1, r1, r1
add r3, r3, r3
1:
vld1.8 {q0}, [r2], r3
vld1.8 {q1}, [lr], r3
subs r12, r12, #2
vst1.8 {q0}, [r0, :128], r1
vst1.8 {q1}, [r4, :128], r1
bne 1b
pop {r4,pc}
endfunc
function ff_vp9_avg16_neon, export=1
push {lr}
ldr r12, [sp, #4]
mov lr, r0
1:
vld1.8 {q2}, [r2], r3
vld1.8 {q0}, [r0, :128], r1
vld1.8 {q3}, [r2], r3
vrhadd.u8 q0, q0, q2
vld1.8 {q1}, [r0, :128], r1
vrhadd.u8 q1, q1, q3
subs r12, r12, #2
vst1.8 {q0}, [lr, :128], r1
vst1.8 {q1}, [lr, :128], r1
bne 1b
pop {pc}
endfunc
function ff_vp9_copy8_neon, export=1
ldr r12, [sp]
1:
vld1.8 {d0}, [r2], r3
vld1.8 {d1}, [r2], r3
subs r12, r12, #2
vst1.8 {d0}, [r0, :64], r1
vst1.8 {d1}, [r0, :64], r1
bne 1b
bx lr
endfunc
function ff_vp9_avg8_neon, export=1
ldr r12, [sp]
1:
vld1.8 {d2}, [r2], r3
vld1.8 {d0}, [r0, :64], r1
vld1.8 {d3}, [r2], r3
vrhadd.u8 d0, d0, d2
vld1.8 {d1}, [r0, :64]
sub r0, r0, r1
vrhadd.u8 d1, d1, d3
subs r12, r12, #2
vst1.8 {d0}, [r0, :64], r1
vst1.8 {d1}, [r0, :64], r1
bne 1b
bx lr
endfunc
function ff_vp9_copy4_neon, export=1
ldr r12, [sp]
1:
vld1.32 {d0[]}, [r2], r3
vld1.32 {d1[]}, [r2], r3
vst1.32 {d0[0]}, [r0, :32], r1
vld1.32 {d2[]}, [r2], r3
vst1.32 {d1[0]}, [r0, :32], r1
vld1.32 {d3[]}, [r2], r3
subs r12, r12, #4
vst1.32 {d2[0]}, [r0, :32], r1
vst1.32 {d3[0]}, [r0, :32], r1
bne 1b
bx lr
endfunc
function ff_vp9_avg4_neon, export=1
push {lr}
ldr r12, [sp, #4]
mov lr, r0
1:
vld1.32 {d4[]}, [r2], r3
vld1.32 {d0[]}, [r0, :32], r1
vld1.32 {d5[]}, [r2], r3
vrhadd.u8 d0, d0, d4
vld1.32 {d1[]}, [r0, :32], r1
vld1.32 {d6[]}, [r2], r3
vrhadd.u8 d1, d1, d5
vld1.32 {d2[]}, [r0, :32], r1
vld1.32 {d7[]}, [r2], r3
vrhadd.u8 d2, d2, d6
vld1.32 {d3[]}, [r0, :32], r1
subs r12, r12, #4
vst1.32 {d0[0]}, [lr, :32], r1
vrhadd.u8 d3, d3, d7
vst1.32 {d1[0]}, [lr, :32], r1
vst1.32 {d2[0]}, [lr, :32], r1
vst1.32 {d3[0]}, [lr, :32], r1
bne 1b
pop {pc}
endfunc
@ Helper macros for vmul/vmla with a constant from either d0 or d1 depending on index
.macro vmul_lane dst, src, idx
.if \idx < 4
vmul.s16 \dst, \src, d0[\idx]
.else
vmul.s16 \dst, \src, d1[\idx - 4]
.endif
.endm
.macro vmla_lane dst, src, idx
.if \idx < 4
vmla.s16 \dst, \src, d0[\idx]
.else
vmla.s16 \dst, \src, d1[\idx - 4]
.endif
.endm
@ Extract a vector from src1-src2 and src4-src5 (src1-src3 and src4-src6
@ for size >= 16), and multiply-accumulate into dst1 and dst3 (or
@ dst1-dst2 and dst3-dst4 for size >= 16)
.macro extmla dst1, dst2, dst3, dst4, dst1d, dst3d, src1, src2, src3, src4, src5, src6, offset, size
arm: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. The filter coefficients are signed values, where the product of the multiplication with one individual filter coefficient doesn't overflow a 16 bit signed value (the largest filter coefficient is 127). But when the products are accumulated, the resulting sum can overflow the 16 bit signed range. Instead of accumulating in 32 bit, we accumulate the largest product (either index 3 or 4) last with a saturated addition. (The VP8 MC asm does something similar, but slightly simpler, by accumulating each half of the filter separately. In the VP9 MC filters, each half of the filter can also overflow though, so the largest component has to be handled individually.) Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_avg4_neon: 1.71 1.15 1.42 1.49 vp9_avg8_neon: 2.51 3.63 3.14 2.58 vp9_avg16_neon: 2.95 6.76 3.01 2.84 vp9_avg32_neon: 3.29 6.64 2.85 3.00 vp9_avg64_neon: 3.47 6.67 3.14 2.80 vp9_avg_8tap_smooth_4h_neon: 3.22 4.73 2.76 4.67 vp9_avg_8tap_smooth_4hv_neon: 3.67 4.76 3.28 4.71 vp9_avg_8tap_smooth_4v_neon: 5.52 7.60 4.60 6.31 vp9_avg_8tap_smooth_8h_neon: 6.22 9.04 5.12 9.32 vp9_avg_8tap_smooth_8hv_neon: 6.38 8.21 5.72 8.17 vp9_avg_8tap_smooth_8v_neon: 9.22 12.66 8.15 11.10 vp9_avg_8tap_smooth_64h_neon: 7.02 10.23 5.54 11.58 vp9_avg_8tap_smooth_64hv_neon: 6.76 9.46 5.93 9.40 vp9_avg_8tap_smooth_64v_neon: 10.76 14.13 9.46 13.37 vp9_put4_neon: 1.11 1.47 1.00 1.21 vp9_put8_neon: 1.23 2.17 1.94 1.48 vp9_put16_neon: 1.63 4.02 1.73 1.97 vp9_put32_neon: 1.56 4.92 2.00 1.96 vp9_put64_neon: 2.10 5.28 2.03 2.35 vp9_put_8tap_smooth_4h_neon: 3.11 4.35 2.63 4.35 vp9_put_8tap_smooth_4hv_neon: 3.67 4.69 3.25 4.71 vp9_put_8tap_smooth_4v_neon: 5.45 7.27 4.49 6.52 vp9_put_8tap_smooth_8h_neon: 5.97 8.18 4.81 8.56 vp9_put_8tap_smooth_8hv_neon: 6.39 7.90 5.64 8.15 vp9_put_8tap_smooth_8v_neon: 9.03 11.84 8.07 11.51 vp9_put_8tap_smooth_64h_neon: 6.78 9.48 4.88 10.89 vp9_put_8tap_smooth_64hv_neon: 6.99 8.87 5.94 9.56 vp9_put_8tap_smooth_64v_neon: 10.69 13.30 9.43 14.34 For the larger 8tap filters, the speedup vs C code is around 5-14x. This is significantly faster than libvpx's implementation of the same functions, at least when comparing the put_8tap_smooth_64 functions (compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from libvpx). Absolute runtimes from checkasm: Cortex A7 A8 A9 A53 vp9_put_8tap_smooth_64h_neon: 20150.3 14489.4 19733.6 10863.7 libvpx vpx_convolve8_horiz_neon: 52623.3 19736.4 21907.7 25027.7 vp9_put_8tap_smooth_64v_neon: 14455.0 12303.9 13746.4 9628.9 libvpx vpx_convolve8_vert_neon: 42090.0 17706.2 17659.9 16941.2 Thus, on the A9, the horizontal filter is only marginally faster than libvpx, while our version is significantly faster on the other cores, and the vertical filter is significantly faster on all cores. The difference is especially large on the A7. The libvpx implementation does the accumulation in 32 bit, which probably explains most of the differences. This is an adapted cherry-pick from libav commits ffbd1d2b0002576ef0d976a41ff959c635373fdc, 392caa65df3efa8b2d48a80f08a6af4892c61c08, 557c1675cf0e803b2fee43b4c8b58433842c84d0 and 11623217e3c9b859daee544e31acdd0821b61039. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:21 +02:00
vext.8 q14, \src1, \src2, #(2*\offset)
vext.8 q15, \src4, \src5, #(2*\offset)
.if \size >= 16
vmla_lane \dst1, q14, \offset
vext.8 q5, \src2, \src3, #(2*\offset)
vmla_lane \dst3, q15, \offset
vext.8 q6, \src5, \src6, #(2*\offset)
vmla_lane \dst2, q5, \offset
vmla_lane \dst4, q6, \offset
.elseif \size == 8
arm: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. The filter coefficients are signed values, where the product of the multiplication with one individual filter coefficient doesn't overflow a 16 bit signed value (the largest filter coefficient is 127). But when the products are accumulated, the resulting sum can overflow the 16 bit signed range. Instead of accumulating in 32 bit, we accumulate the largest product (either index 3 or 4) last with a saturated addition. (The VP8 MC asm does something similar, but slightly simpler, by accumulating each half of the filter separately. In the VP9 MC filters, each half of the filter can also overflow though, so the largest component has to be handled individually.) Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_avg4_neon: 1.71 1.15 1.42 1.49 vp9_avg8_neon: 2.51 3.63 3.14 2.58 vp9_avg16_neon: 2.95 6.76 3.01 2.84 vp9_avg32_neon: 3.29 6.64 2.85 3.00 vp9_avg64_neon: 3.47 6.67 3.14 2.80 vp9_avg_8tap_smooth_4h_neon: 3.22 4.73 2.76 4.67 vp9_avg_8tap_smooth_4hv_neon: 3.67 4.76 3.28 4.71 vp9_avg_8tap_smooth_4v_neon: 5.52 7.60 4.60 6.31 vp9_avg_8tap_smooth_8h_neon: 6.22 9.04 5.12 9.32 vp9_avg_8tap_smooth_8hv_neon: 6.38 8.21 5.72 8.17 vp9_avg_8tap_smooth_8v_neon: 9.22 12.66 8.15 11.10 vp9_avg_8tap_smooth_64h_neon: 7.02 10.23 5.54 11.58 vp9_avg_8tap_smooth_64hv_neon: 6.76 9.46 5.93 9.40 vp9_avg_8tap_smooth_64v_neon: 10.76 14.13 9.46 13.37 vp9_put4_neon: 1.11 1.47 1.00 1.21 vp9_put8_neon: 1.23 2.17 1.94 1.48 vp9_put16_neon: 1.63 4.02 1.73 1.97 vp9_put32_neon: 1.56 4.92 2.00 1.96 vp9_put64_neon: 2.10 5.28 2.03 2.35 vp9_put_8tap_smooth_4h_neon: 3.11 4.35 2.63 4.35 vp9_put_8tap_smooth_4hv_neon: 3.67 4.69 3.25 4.71 vp9_put_8tap_smooth_4v_neon: 5.45 7.27 4.49 6.52 vp9_put_8tap_smooth_8h_neon: 5.97 8.18 4.81 8.56 vp9_put_8tap_smooth_8hv_neon: 6.39 7.90 5.64 8.15 vp9_put_8tap_smooth_8v_neon: 9.03 11.84 8.07 11.51 vp9_put_8tap_smooth_64h_neon: 6.78 9.48 4.88 10.89 vp9_put_8tap_smooth_64hv_neon: 6.99 8.87 5.94 9.56 vp9_put_8tap_smooth_64v_neon: 10.69 13.30 9.43 14.34 For the larger 8tap filters, the speedup vs C code is around 5-14x. This is significantly faster than libvpx's implementation of the same functions, at least when comparing the put_8tap_smooth_64 functions (compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from libvpx). Absolute runtimes from checkasm: Cortex A7 A8 A9 A53 vp9_put_8tap_smooth_64h_neon: 20150.3 14489.4 19733.6 10863.7 libvpx vpx_convolve8_horiz_neon: 52623.3 19736.4 21907.7 25027.7 vp9_put_8tap_smooth_64v_neon: 14455.0 12303.9 13746.4 9628.9 libvpx vpx_convolve8_vert_neon: 42090.0 17706.2 17659.9 16941.2 Thus, on the A9, the horizontal filter is only marginally faster than libvpx, while our version is significantly faster on the other cores, and the vertical filter is significantly faster on all cores. The difference is especially large on the A7. The libvpx implementation does the accumulation in 32 bit, which probably explains most of the differences. This is an adapted cherry-pick from libav commits ffbd1d2b0002576ef0d976a41ff959c635373fdc, 392caa65df3efa8b2d48a80f08a6af4892c61c08, 557c1675cf0e803b2fee43b4c8b58433842c84d0 and 11623217e3c9b859daee544e31acdd0821b61039. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:21 +02:00
vmla_lane \dst1, q14, \offset
vmla_lane \dst3, q15, \offset
.else
vmla_lane \dst1d, d28, \offset
vmla_lane \dst3d, d30, \offset
arm: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. The filter coefficients are signed values, where the product of the multiplication with one individual filter coefficient doesn't overflow a 16 bit signed value (the largest filter coefficient is 127). But when the products are accumulated, the resulting sum can overflow the 16 bit signed range. Instead of accumulating in 32 bit, we accumulate the largest product (either index 3 or 4) last with a saturated addition. (The VP8 MC asm does something similar, but slightly simpler, by accumulating each half of the filter separately. In the VP9 MC filters, each half of the filter can also overflow though, so the largest component has to be handled individually.) Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_avg4_neon: 1.71 1.15 1.42 1.49 vp9_avg8_neon: 2.51 3.63 3.14 2.58 vp9_avg16_neon: 2.95 6.76 3.01 2.84 vp9_avg32_neon: 3.29 6.64 2.85 3.00 vp9_avg64_neon: 3.47 6.67 3.14 2.80 vp9_avg_8tap_smooth_4h_neon: 3.22 4.73 2.76 4.67 vp9_avg_8tap_smooth_4hv_neon: 3.67 4.76 3.28 4.71 vp9_avg_8tap_smooth_4v_neon: 5.52 7.60 4.60 6.31 vp9_avg_8tap_smooth_8h_neon: 6.22 9.04 5.12 9.32 vp9_avg_8tap_smooth_8hv_neon: 6.38 8.21 5.72 8.17 vp9_avg_8tap_smooth_8v_neon: 9.22 12.66 8.15 11.10 vp9_avg_8tap_smooth_64h_neon: 7.02 10.23 5.54 11.58 vp9_avg_8tap_smooth_64hv_neon: 6.76 9.46 5.93 9.40 vp9_avg_8tap_smooth_64v_neon: 10.76 14.13 9.46 13.37 vp9_put4_neon: 1.11 1.47 1.00 1.21 vp9_put8_neon: 1.23 2.17 1.94 1.48 vp9_put16_neon: 1.63 4.02 1.73 1.97 vp9_put32_neon: 1.56 4.92 2.00 1.96 vp9_put64_neon: 2.10 5.28 2.03 2.35 vp9_put_8tap_smooth_4h_neon: 3.11 4.35 2.63 4.35 vp9_put_8tap_smooth_4hv_neon: 3.67 4.69 3.25 4.71 vp9_put_8tap_smooth_4v_neon: 5.45 7.27 4.49 6.52 vp9_put_8tap_smooth_8h_neon: 5.97 8.18 4.81 8.56 vp9_put_8tap_smooth_8hv_neon: 6.39 7.90 5.64 8.15 vp9_put_8tap_smooth_8v_neon: 9.03 11.84 8.07 11.51 vp9_put_8tap_smooth_64h_neon: 6.78 9.48 4.88 10.89 vp9_put_8tap_smooth_64hv_neon: 6.99 8.87 5.94 9.56 vp9_put_8tap_smooth_64v_neon: 10.69 13.30 9.43 14.34 For the larger 8tap filters, the speedup vs C code is around 5-14x. This is significantly faster than libvpx's implementation of the same functions, at least when comparing the put_8tap_smooth_64 functions (compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from libvpx). Absolute runtimes from checkasm: Cortex A7 A8 A9 A53 vp9_put_8tap_smooth_64h_neon: 20150.3 14489.4 19733.6 10863.7 libvpx vpx_convolve8_horiz_neon: 52623.3 19736.4 21907.7 25027.7 vp9_put_8tap_smooth_64v_neon: 14455.0 12303.9 13746.4 9628.9 libvpx vpx_convolve8_vert_neon: 42090.0 17706.2 17659.9 16941.2 Thus, on the A9, the horizontal filter is only marginally faster than libvpx, while our version is significantly faster on the other cores, and the vertical filter is significantly faster on all cores. The difference is especially large on the A7. The libvpx implementation does the accumulation in 32 bit, which probably explains most of the differences. This is an adapted cherry-pick from libav commits ffbd1d2b0002576ef0d976a41ff959c635373fdc, 392caa65df3efa8b2d48a80f08a6af4892c61c08, 557c1675cf0e803b2fee43b4c8b58433842c84d0 and 11623217e3c9b859daee544e31acdd0821b61039. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:21 +02:00
.endif
.endm
@ The same as above, but don't accumulate straight into the
@ destination, but use a temp register and accumulate with saturation.
.macro extmulqadd dst1, dst2, dst3, dst4, dst1d, dst3d, src1, src2, src3, src4, src5, src6, offset, size
arm: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. The filter coefficients are signed values, where the product of the multiplication with one individual filter coefficient doesn't overflow a 16 bit signed value (the largest filter coefficient is 127). But when the products are accumulated, the resulting sum can overflow the 16 bit signed range. Instead of accumulating in 32 bit, we accumulate the largest product (either index 3 or 4) last with a saturated addition. (The VP8 MC asm does something similar, but slightly simpler, by accumulating each half of the filter separately. In the VP9 MC filters, each half of the filter can also overflow though, so the largest component has to be handled individually.) Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_avg4_neon: 1.71 1.15 1.42 1.49 vp9_avg8_neon: 2.51 3.63 3.14 2.58 vp9_avg16_neon: 2.95 6.76 3.01 2.84 vp9_avg32_neon: 3.29 6.64 2.85 3.00 vp9_avg64_neon: 3.47 6.67 3.14 2.80 vp9_avg_8tap_smooth_4h_neon: 3.22 4.73 2.76 4.67 vp9_avg_8tap_smooth_4hv_neon: 3.67 4.76 3.28 4.71 vp9_avg_8tap_smooth_4v_neon: 5.52 7.60 4.60 6.31 vp9_avg_8tap_smooth_8h_neon: 6.22 9.04 5.12 9.32 vp9_avg_8tap_smooth_8hv_neon: 6.38 8.21 5.72 8.17 vp9_avg_8tap_smooth_8v_neon: 9.22 12.66 8.15 11.10 vp9_avg_8tap_smooth_64h_neon: 7.02 10.23 5.54 11.58 vp9_avg_8tap_smooth_64hv_neon: 6.76 9.46 5.93 9.40 vp9_avg_8tap_smooth_64v_neon: 10.76 14.13 9.46 13.37 vp9_put4_neon: 1.11 1.47 1.00 1.21 vp9_put8_neon: 1.23 2.17 1.94 1.48 vp9_put16_neon: 1.63 4.02 1.73 1.97 vp9_put32_neon: 1.56 4.92 2.00 1.96 vp9_put64_neon: 2.10 5.28 2.03 2.35 vp9_put_8tap_smooth_4h_neon: 3.11 4.35 2.63 4.35 vp9_put_8tap_smooth_4hv_neon: 3.67 4.69 3.25 4.71 vp9_put_8tap_smooth_4v_neon: 5.45 7.27 4.49 6.52 vp9_put_8tap_smooth_8h_neon: 5.97 8.18 4.81 8.56 vp9_put_8tap_smooth_8hv_neon: 6.39 7.90 5.64 8.15 vp9_put_8tap_smooth_8v_neon: 9.03 11.84 8.07 11.51 vp9_put_8tap_smooth_64h_neon: 6.78 9.48 4.88 10.89 vp9_put_8tap_smooth_64hv_neon: 6.99 8.87 5.94 9.56 vp9_put_8tap_smooth_64v_neon: 10.69 13.30 9.43 14.34 For the larger 8tap filters, the speedup vs C code is around 5-14x. This is significantly faster than libvpx's implementation of the same functions, at least when comparing the put_8tap_smooth_64 functions (compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from libvpx). Absolute runtimes from checkasm: Cortex A7 A8 A9 A53 vp9_put_8tap_smooth_64h_neon: 20150.3 14489.4 19733.6 10863.7 libvpx vpx_convolve8_horiz_neon: 52623.3 19736.4 21907.7 25027.7 vp9_put_8tap_smooth_64v_neon: 14455.0 12303.9 13746.4 9628.9 libvpx vpx_convolve8_vert_neon: 42090.0 17706.2 17659.9 16941.2 Thus, on the A9, the horizontal filter is only marginally faster than libvpx, while our version is significantly faster on the other cores, and the vertical filter is significantly faster on all cores. The difference is especially large on the A7. The libvpx implementation does the accumulation in 32 bit, which probably explains most of the differences. This is an adapted cherry-pick from libav commits ffbd1d2b0002576ef0d976a41ff959c635373fdc, 392caa65df3efa8b2d48a80f08a6af4892c61c08, 557c1675cf0e803b2fee43b4c8b58433842c84d0 and 11623217e3c9b859daee544e31acdd0821b61039. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:21 +02:00
vext.8 q14, \src1, \src2, #(2*\offset)
vext.8 q15, \src4, \src5, #(2*\offset)
.if \size >= 16
vmul_lane q14, q14, \offset
vext.8 q5, \src2, \src3, #(2*\offset)
vmul_lane q15, q15, \offset
vext.8 q6, \src5, \src6, #(2*\offset)
vmul_lane q5, q5, \offset
vmul_lane q6, q6, \offset
.elseif \size == 8
arm: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. The filter coefficients are signed values, where the product of the multiplication with one individual filter coefficient doesn't overflow a 16 bit signed value (the largest filter coefficient is 127). But when the products are accumulated, the resulting sum can overflow the 16 bit signed range. Instead of accumulating in 32 bit, we accumulate the largest product (either index 3 or 4) last with a saturated addition. (The VP8 MC asm does something similar, but slightly simpler, by accumulating each half of the filter separately. In the VP9 MC filters, each half of the filter can also overflow though, so the largest component has to be handled individually.) Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_avg4_neon: 1.71 1.15 1.42 1.49 vp9_avg8_neon: 2.51 3.63 3.14 2.58 vp9_avg16_neon: 2.95 6.76 3.01 2.84 vp9_avg32_neon: 3.29 6.64 2.85 3.00 vp9_avg64_neon: 3.47 6.67 3.14 2.80 vp9_avg_8tap_smooth_4h_neon: 3.22 4.73 2.76 4.67 vp9_avg_8tap_smooth_4hv_neon: 3.67 4.76 3.28 4.71 vp9_avg_8tap_smooth_4v_neon: 5.52 7.60 4.60 6.31 vp9_avg_8tap_smooth_8h_neon: 6.22 9.04 5.12 9.32 vp9_avg_8tap_smooth_8hv_neon: 6.38 8.21 5.72 8.17 vp9_avg_8tap_smooth_8v_neon: 9.22 12.66 8.15 11.10 vp9_avg_8tap_smooth_64h_neon: 7.02 10.23 5.54 11.58 vp9_avg_8tap_smooth_64hv_neon: 6.76 9.46 5.93 9.40 vp9_avg_8tap_smooth_64v_neon: 10.76 14.13 9.46 13.37 vp9_put4_neon: 1.11 1.47 1.00 1.21 vp9_put8_neon: 1.23 2.17 1.94 1.48 vp9_put16_neon: 1.63 4.02 1.73 1.97 vp9_put32_neon: 1.56 4.92 2.00 1.96 vp9_put64_neon: 2.10 5.28 2.03 2.35 vp9_put_8tap_smooth_4h_neon: 3.11 4.35 2.63 4.35 vp9_put_8tap_smooth_4hv_neon: 3.67 4.69 3.25 4.71 vp9_put_8tap_smooth_4v_neon: 5.45 7.27 4.49 6.52 vp9_put_8tap_smooth_8h_neon: 5.97 8.18 4.81 8.56 vp9_put_8tap_smooth_8hv_neon: 6.39 7.90 5.64 8.15 vp9_put_8tap_smooth_8v_neon: 9.03 11.84 8.07 11.51 vp9_put_8tap_smooth_64h_neon: 6.78 9.48 4.88 10.89 vp9_put_8tap_smooth_64hv_neon: 6.99 8.87 5.94 9.56 vp9_put_8tap_smooth_64v_neon: 10.69 13.30 9.43 14.34 For the larger 8tap filters, the speedup vs C code is around 5-14x. This is significantly faster than libvpx's implementation of the same functions, at least when comparing the put_8tap_smooth_64 functions (compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from libvpx). Absolute runtimes from checkasm: Cortex A7 A8 A9 A53 vp9_put_8tap_smooth_64h_neon: 20150.3 14489.4 19733.6 10863.7 libvpx vpx_convolve8_horiz_neon: 52623.3 19736.4 21907.7 25027.7 vp9_put_8tap_smooth_64v_neon: 14455.0 12303.9 13746.4 9628.9 libvpx vpx_convolve8_vert_neon: 42090.0 17706.2 17659.9 16941.2 Thus, on the A9, the horizontal filter is only marginally faster than libvpx, while our version is significantly faster on the other cores, and the vertical filter is significantly faster on all cores. The difference is especially large on the A7. The libvpx implementation does the accumulation in 32 bit, which probably explains most of the differences. This is an adapted cherry-pick from libav commits ffbd1d2b0002576ef0d976a41ff959c635373fdc, 392caa65df3efa8b2d48a80f08a6af4892c61c08, 557c1675cf0e803b2fee43b4c8b58433842c84d0 and 11623217e3c9b859daee544e31acdd0821b61039. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:21 +02:00
vmul_lane q14, q14, \offset
vmul_lane q15, q15, \offset
.else
vmul_lane d28, d28, \offset
vmul_lane d30, d30, \offset
arm: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. The filter coefficients are signed values, where the product of the multiplication with one individual filter coefficient doesn't overflow a 16 bit signed value (the largest filter coefficient is 127). But when the products are accumulated, the resulting sum can overflow the 16 bit signed range. Instead of accumulating in 32 bit, we accumulate the largest product (either index 3 or 4) last with a saturated addition. (The VP8 MC asm does something similar, but slightly simpler, by accumulating each half of the filter separately. In the VP9 MC filters, each half of the filter can also overflow though, so the largest component has to be handled individually.) Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_avg4_neon: 1.71 1.15 1.42 1.49 vp9_avg8_neon: 2.51 3.63 3.14 2.58 vp9_avg16_neon: 2.95 6.76 3.01 2.84 vp9_avg32_neon: 3.29 6.64 2.85 3.00 vp9_avg64_neon: 3.47 6.67 3.14 2.80 vp9_avg_8tap_smooth_4h_neon: 3.22 4.73 2.76 4.67 vp9_avg_8tap_smooth_4hv_neon: 3.67 4.76 3.28 4.71 vp9_avg_8tap_smooth_4v_neon: 5.52 7.60 4.60 6.31 vp9_avg_8tap_smooth_8h_neon: 6.22 9.04 5.12 9.32 vp9_avg_8tap_smooth_8hv_neon: 6.38 8.21 5.72 8.17 vp9_avg_8tap_smooth_8v_neon: 9.22 12.66 8.15 11.10 vp9_avg_8tap_smooth_64h_neon: 7.02 10.23 5.54 11.58 vp9_avg_8tap_smooth_64hv_neon: 6.76 9.46 5.93 9.40 vp9_avg_8tap_smooth_64v_neon: 10.76 14.13 9.46 13.37 vp9_put4_neon: 1.11 1.47 1.00 1.21 vp9_put8_neon: 1.23 2.17 1.94 1.48 vp9_put16_neon: 1.63 4.02 1.73 1.97 vp9_put32_neon: 1.56 4.92 2.00 1.96 vp9_put64_neon: 2.10 5.28 2.03 2.35 vp9_put_8tap_smooth_4h_neon: 3.11 4.35 2.63 4.35 vp9_put_8tap_smooth_4hv_neon: 3.67 4.69 3.25 4.71 vp9_put_8tap_smooth_4v_neon: 5.45 7.27 4.49 6.52 vp9_put_8tap_smooth_8h_neon: 5.97 8.18 4.81 8.56 vp9_put_8tap_smooth_8hv_neon: 6.39 7.90 5.64 8.15 vp9_put_8tap_smooth_8v_neon: 9.03 11.84 8.07 11.51 vp9_put_8tap_smooth_64h_neon: 6.78 9.48 4.88 10.89 vp9_put_8tap_smooth_64hv_neon: 6.99 8.87 5.94 9.56 vp9_put_8tap_smooth_64v_neon: 10.69 13.30 9.43 14.34 For the larger 8tap filters, the speedup vs C code is around 5-14x. This is significantly faster than libvpx's implementation of the same functions, at least when comparing the put_8tap_smooth_64 functions (compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from libvpx). Absolute runtimes from checkasm: Cortex A7 A8 A9 A53 vp9_put_8tap_smooth_64h_neon: 20150.3 14489.4 19733.6 10863.7 libvpx vpx_convolve8_horiz_neon: 52623.3 19736.4 21907.7 25027.7 vp9_put_8tap_smooth_64v_neon: 14455.0 12303.9 13746.4 9628.9 libvpx vpx_convolve8_vert_neon: 42090.0 17706.2 17659.9 16941.2 Thus, on the A9, the horizontal filter is only marginally faster than libvpx, while our version is significantly faster on the other cores, and the vertical filter is significantly faster on all cores. The difference is especially large on the A7. The libvpx implementation does the accumulation in 32 bit, which probably explains most of the differences. This is an adapted cherry-pick from libav commits ffbd1d2b0002576ef0d976a41ff959c635373fdc, 392caa65df3efa8b2d48a80f08a6af4892c61c08, 557c1675cf0e803b2fee43b4c8b58433842c84d0 and 11623217e3c9b859daee544e31acdd0821b61039. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:21 +02:00
.endif
.if \size == 4
vqadd.s16 \dst1d, \dst1d, d28
vqadd.s16 \dst3d, \dst3d, d30
.else
arm: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. The filter coefficients are signed values, where the product of the multiplication with one individual filter coefficient doesn't overflow a 16 bit signed value (the largest filter coefficient is 127). But when the products are accumulated, the resulting sum can overflow the 16 bit signed range. Instead of accumulating in 32 bit, we accumulate the largest product (either index 3 or 4) last with a saturated addition. (The VP8 MC asm does something similar, but slightly simpler, by accumulating each half of the filter separately. In the VP9 MC filters, each half of the filter can also overflow though, so the largest component has to be handled individually.) Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_avg4_neon: 1.71 1.15 1.42 1.49 vp9_avg8_neon: 2.51 3.63 3.14 2.58 vp9_avg16_neon: 2.95 6.76 3.01 2.84 vp9_avg32_neon: 3.29 6.64 2.85 3.00 vp9_avg64_neon: 3.47 6.67 3.14 2.80 vp9_avg_8tap_smooth_4h_neon: 3.22 4.73 2.76 4.67 vp9_avg_8tap_smooth_4hv_neon: 3.67 4.76 3.28 4.71 vp9_avg_8tap_smooth_4v_neon: 5.52 7.60 4.60 6.31 vp9_avg_8tap_smooth_8h_neon: 6.22 9.04 5.12 9.32 vp9_avg_8tap_smooth_8hv_neon: 6.38 8.21 5.72 8.17 vp9_avg_8tap_smooth_8v_neon: 9.22 12.66 8.15 11.10 vp9_avg_8tap_smooth_64h_neon: 7.02 10.23 5.54 11.58 vp9_avg_8tap_smooth_64hv_neon: 6.76 9.46 5.93 9.40 vp9_avg_8tap_smooth_64v_neon: 10.76 14.13 9.46 13.37 vp9_put4_neon: 1.11 1.47 1.00 1.21 vp9_put8_neon: 1.23 2.17 1.94 1.48 vp9_put16_neon: 1.63 4.02 1.73 1.97 vp9_put32_neon: 1.56 4.92 2.00 1.96 vp9_put64_neon: 2.10 5.28 2.03 2.35 vp9_put_8tap_smooth_4h_neon: 3.11 4.35 2.63 4.35 vp9_put_8tap_smooth_4hv_neon: 3.67 4.69 3.25 4.71 vp9_put_8tap_smooth_4v_neon: 5.45 7.27 4.49 6.52 vp9_put_8tap_smooth_8h_neon: 5.97 8.18 4.81 8.56 vp9_put_8tap_smooth_8hv_neon: 6.39 7.90 5.64 8.15 vp9_put_8tap_smooth_8v_neon: 9.03 11.84 8.07 11.51 vp9_put_8tap_smooth_64h_neon: 6.78 9.48 4.88 10.89 vp9_put_8tap_smooth_64hv_neon: 6.99 8.87 5.94 9.56 vp9_put_8tap_smooth_64v_neon: 10.69 13.30 9.43 14.34 For the larger 8tap filters, the speedup vs C code is around 5-14x. This is significantly faster than libvpx's implementation of the same functions, at least when comparing the put_8tap_smooth_64 functions (compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from libvpx). Absolute runtimes from checkasm: Cortex A7 A8 A9 A53 vp9_put_8tap_smooth_64h_neon: 20150.3 14489.4 19733.6 10863.7 libvpx vpx_convolve8_horiz_neon: 52623.3 19736.4 21907.7 25027.7 vp9_put_8tap_smooth_64v_neon: 14455.0 12303.9 13746.4 9628.9 libvpx vpx_convolve8_vert_neon: 42090.0 17706.2 17659.9 16941.2 Thus, on the A9, the horizontal filter is only marginally faster than libvpx, while our version is significantly faster on the other cores, and the vertical filter is significantly faster on all cores. The difference is especially large on the A7. The libvpx implementation does the accumulation in 32 bit, which probably explains most of the differences. This is an adapted cherry-pick from libav commits ffbd1d2b0002576ef0d976a41ff959c635373fdc, 392caa65df3efa8b2d48a80f08a6af4892c61c08, 557c1675cf0e803b2fee43b4c8b58433842c84d0 and 11623217e3c9b859daee544e31acdd0821b61039. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:21 +02:00
vqadd.s16 \dst1, \dst1, q14
vqadd.s16 \dst3, \dst3, q15
.if \size >= 16
vqadd.s16 \dst2, \dst2, q5
vqadd.s16 \dst4, \dst4, q6
.endif
.endif
arm: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. The filter coefficients are signed values, where the product of the multiplication with one individual filter coefficient doesn't overflow a 16 bit signed value (the largest filter coefficient is 127). But when the products are accumulated, the resulting sum can overflow the 16 bit signed range. Instead of accumulating in 32 bit, we accumulate the largest product (either index 3 or 4) last with a saturated addition. (The VP8 MC asm does something similar, but slightly simpler, by accumulating each half of the filter separately. In the VP9 MC filters, each half of the filter can also overflow though, so the largest component has to be handled individually.) Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_avg4_neon: 1.71 1.15 1.42 1.49 vp9_avg8_neon: 2.51 3.63 3.14 2.58 vp9_avg16_neon: 2.95 6.76 3.01 2.84 vp9_avg32_neon: 3.29 6.64 2.85 3.00 vp9_avg64_neon: 3.47 6.67 3.14 2.80 vp9_avg_8tap_smooth_4h_neon: 3.22 4.73 2.76 4.67 vp9_avg_8tap_smooth_4hv_neon: 3.67 4.76 3.28 4.71 vp9_avg_8tap_smooth_4v_neon: 5.52 7.60 4.60 6.31 vp9_avg_8tap_smooth_8h_neon: 6.22 9.04 5.12 9.32 vp9_avg_8tap_smooth_8hv_neon: 6.38 8.21 5.72 8.17 vp9_avg_8tap_smooth_8v_neon: 9.22 12.66 8.15 11.10 vp9_avg_8tap_smooth_64h_neon: 7.02 10.23 5.54 11.58 vp9_avg_8tap_smooth_64hv_neon: 6.76 9.46 5.93 9.40 vp9_avg_8tap_smooth_64v_neon: 10.76 14.13 9.46 13.37 vp9_put4_neon: 1.11 1.47 1.00 1.21 vp9_put8_neon: 1.23 2.17 1.94 1.48 vp9_put16_neon: 1.63 4.02 1.73 1.97 vp9_put32_neon: 1.56 4.92 2.00 1.96 vp9_put64_neon: 2.10 5.28 2.03 2.35 vp9_put_8tap_smooth_4h_neon: 3.11 4.35 2.63 4.35 vp9_put_8tap_smooth_4hv_neon: 3.67 4.69 3.25 4.71 vp9_put_8tap_smooth_4v_neon: 5.45 7.27 4.49 6.52 vp9_put_8tap_smooth_8h_neon: 5.97 8.18 4.81 8.56 vp9_put_8tap_smooth_8hv_neon: 6.39 7.90 5.64 8.15 vp9_put_8tap_smooth_8v_neon: 9.03 11.84 8.07 11.51 vp9_put_8tap_smooth_64h_neon: 6.78 9.48 4.88 10.89 vp9_put_8tap_smooth_64hv_neon: 6.99 8.87 5.94 9.56 vp9_put_8tap_smooth_64v_neon: 10.69 13.30 9.43 14.34 For the larger 8tap filters, the speedup vs C code is around 5-14x. This is significantly faster than libvpx's implementation of the same functions, at least when comparing the put_8tap_smooth_64 functions (compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from libvpx). Absolute runtimes from checkasm: Cortex A7 A8 A9 A53 vp9_put_8tap_smooth_64h_neon: 20150.3 14489.4 19733.6 10863.7 libvpx vpx_convolve8_horiz_neon: 52623.3 19736.4 21907.7 25027.7 vp9_put_8tap_smooth_64v_neon: 14455.0 12303.9 13746.4 9628.9 libvpx vpx_convolve8_vert_neon: 42090.0 17706.2 17659.9 16941.2 Thus, on the A9, the horizontal filter is only marginally faster than libvpx, while our version is significantly faster on the other cores, and the vertical filter is significantly faster on all cores. The difference is especially large on the A7. The libvpx implementation does the accumulation in 32 bit, which probably explains most of the differences. This is an adapted cherry-pick from libav commits ffbd1d2b0002576ef0d976a41ff959c635373fdc, 392caa65df3efa8b2d48a80f08a6af4892c61c08, 557c1675cf0e803b2fee43b4c8b58433842c84d0 and 11623217e3c9b859daee544e31acdd0821b61039. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:21 +02:00
.endm
@ Instantiate a horizontal filter function for the given size.
@ This can work on 4, 8 or 16 pixels in parallel; for larger
@ widths it will do 16 pixels at a time and loop horizontally.
@ The actual width is passed in r5, the height in r4 and
@ the filter coefficients in r12. idx2 is the index of the largest
@ filter coefficient (3 or 4) and idx1 is the other one of them.
.macro do_8tap_h type, size, idx1, idx2
function \type\()_8tap_\size\()h_\idx1\idx2
sub r2, r2, #3
add r6, r0, r1
add r7, r2, r3
add r1, r1, r1
add r3, r3, r3
@ Only size >= 16 loops horizontally and needs
@ reduced dst stride
.if \size >= 16
sub r1, r1, r5
.endif
@ size >= 16 loads two qwords and increments r2,
@ for size 4/8 it's enough with one qword and no
@ postincrement
.if \size >= 16
sub r3, r3, r5
sub r3, r3, #8
.endif
@ Load the filter vector
vld1.16 {q0}, [r12,:128]
1:
.if \size >= 16
mov r12, r5
.endif
@ Load src
.if \size >= 16
vld1.8 {d18, d19, d20}, [r2]!
vld1.8 {d24, d25, d26}, [r7]!
.else
vld1.8 {q9}, [r2]
vld1.8 {q12}, [r7]
.endif
vmovl.u8 q8, d18
vmovl.u8 q9, d19
vmovl.u8 q11, d24
vmovl.u8 q12, d25
.if \size >= 16
vmovl.u8 q10, d20
vmovl.u8 q13, d26
.endif
2:
@ Accumulate, adding idx2 last with a separate
@ saturating add. The positive filter coefficients
@ for all indices except idx2 must add up to less
@ than 127 for this not to overflow.
vmul.s16 q1, q8, d0[0]
vmul.s16 q3, q11, d0[0]
.if \size >= 16
vmul.s16 q2, q9, d0[0]
vmul.s16 q4, q12, d0[0]
.endif
extmla q1, q2, q3, q4, d2, d6, q8, q9, q10, q11, q12, q13, 1, \size
extmla q1, q2, q3, q4, d2, d6, q8, q9, q10, q11, q12, q13, 2, \size
extmla q1, q2, q3, q4, d2, d6, q8, q9, q10, q11, q12, q13, \idx1, \size
extmla q1, q2, q3, q4, d2, d6, q8, q9, q10, q11, q12, q13, 5, \size
extmla q1, q2, q3, q4, d2, d6, q8, q9, q10, q11, q12, q13, 6, \size
extmla q1, q2, q3, q4, d2, d6, q8, q9, q10, q11, q12, q13, 7, \size
extmulqadd q1, q2, q3, q4, d2, d6, q8, q9, q10, q11, q12, q13, \idx2, \size
arm: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. The filter coefficients are signed values, where the product of the multiplication with one individual filter coefficient doesn't overflow a 16 bit signed value (the largest filter coefficient is 127). But when the products are accumulated, the resulting sum can overflow the 16 bit signed range. Instead of accumulating in 32 bit, we accumulate the largest product (either index 3 or 4) last with a saturated addition. (The VP8 MC asm does something similar, but slightly simpler, by accumulating each half of the filter separately. In the VP9 MC filters, each half of the filter can also overflow though, so the largest component has to be handled individually.) Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_avg4_neon: 1.71 1.15 1.42 1.49 vp9_avg8_neon: 2.51 3.63 3.14 2.58 vp9_avg16_neon: 2.95 6.76 3.01 2.84 vp9_avg32_neon: 3.29 6.64 2.85 3.00 vp9_avg64_neon: 3.47 6.67 3.14 2.80 vp9_avg_8tap_smooth_4h_neon: 3.22 4.73 2.76 4.67 vp9_avg_8tap_smooth_4hv_neon: 3.67 4.76 3.28 4.71 vp9_avg_8tap_smooth_4v_neon: 5.52 7.60 4.60 6.31 vp9_avg_8tap_smooth_8h_neon: 6.22 9.04 5.12 9.32 vp9_avg_8tap_smooth_8hv_neon: 6.38 8.21 5.72 8.17 vp9_avg_8tap_smooth_8v_neon: 9.22 12.66 8.15 11.10 vp9_avg_8tap_smooth_64h_neon: 7.02 10.23 5.54 11.58 vp9_avg_8tap_smooth_64hv_neon: 6.76 9.46 5.93 9.40 vp9_avg_8tap_smooth_64v_neon: 10.76 14.13 9.46 13.37 vp9_put4_neon: 1.11 1.47 1.00 1.21 vp9_put8_neon: 1.23 2.17 1.94 1.48 vp9_put16_neon: 1.63 4.02 1.73 1.97 vp9_put32_neon: 1.56 4.92 2.00 1.96 vp9_put64_neon: 2.10 5.28 2.03 2.35 vp9_put_8tap_smooth_4h_neon: 3.11 4.35 2.63 4.35 vp9_put_8tap_smooth_4hv_neon: 3.67 4.69 3.25 4.71 vp9_put_8tap_smooth_4v_neon: 5.45 7.27 4.49 6.52 vp9_put_8tap_smooth_8h_neon: 5.97 8.18 4.81 8.56 vp9_put_8tap_smooth_8hv_neon: 6.39 7.90 5.64 8.15 vp9_put_8tap_smooth_8v_neon: 9.03 11.84 8.07 11.51 vp9_put_8tap_smooth_64h_neon: 6.78 9.48 4.88 10.89 vp9_put_8tap_smooth_64hv_neon: 6.99 8.87 5.94 9.56 vp9_put_8tap_smooth_64v_neon: 10.69 13.30 9.43 14.34 For the larger 8tap filters, the speedup vs C code is around 5-14x. This is significantly faster than libvpx's implementation of the same functions, at least when comparing the put_8tap_smooth_64 functions (compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from libvpx). Absolute runtimes from checkasm: Cortex A7 A8 A9 A53 vp9_put_8tap_smooth_64h_neon: 20150.3 14489.4 19733.6 10863.7 libvpx vpx_convolve8_horiz_neon: 52623.3 19736.4 21907.7 25027.7 vp9_put_8tap_smooth_64v_neon: 14455.0 12303.9 13746.4 9628.9 libvpx vpx_convolve8_vert_neon: 42090.0 17706.2 17659.9 16941.2 Thus, on the A9, the horizontal filter is only marginally faster than libvpx, while our version is significantly faster on the other cores, and the vertical filter is significantly faster on all cores. The difference is especially large on the A7. The libvpx implementation does the accumulation in 32 bit, which probably explains most of the differences. This is an adapted cherry-pick from libav commits ffbd1d2b0002576ef0d976a41ff959c635373fdc, 392caa65df3efa8b2d48a80f08a6af4892c61c08, 557c1675cf0e803b2fee43b4c8b58433842c84d0 and 11623217e3c9b859daee544e31acdd0821b61039. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:21 +02:00
@ Round, shift and saturate
vqrshrun.s16 d2, q1, #7
vqrshrun.s16 d6, q3, #7
.if \size >= 16
vqrshrun.s16 d3, q2, #7
vqrshrun.s16 d7, q4, #7
.endif
@ Average
.ifc \type,avg
.if \size >= 16
vld1.8 {q14}, [r0,:128]
vld1.8 {q15}, [r6,:128]
vrhadd.u8 q1, q1, q14
vrhadd.u8 q3, q3, q15
.elseif \size == 8
vld1.8 {d28}, [r0,:64]
vld1.8 {d30}, [r6,:64]
vrhadd.u8 d2, d2, d28
vrhadd.u8 d6, d6, d30
.else
@ We only need d28[0], but [] is faster on some cores
vld1.32 {d28[]}, [r0,:32]
vld1.32 {d30[]}, [r6,:32]
vrhadd.u8 d2, d2, d28
vrhadd.u8 d6, d6, d30
.endif
.endif
@ Store and loop horizontally (for size >= 16)
.if \size >= 16
subs r12, r12, #16
vst1.8 {q1}, [r0,:128]!
vst1.8 {q3}, [r6,:128]!
beq 3f
vmov q8, q10
vmov q11, q13
vld1.8 {q10}, [r2]!
vld1.8 {q13}, [r7]!
vmovl.u8 q9, d20
vmovl.u8 q10, d21
vmovl.u8 q12, d26
vmovl.u8 q13, d27
b 2b
.elseif \size == 8
vst1.8 {d2}, [r0,:64]
vst1.8 {d6}, [r6,:64]
.else @ \size == 4
vst1.32 {d2[0]}, [r0,:32]
vst1.32 {d6[0]}, [r6,:32]
.endif
3:
@ Loop vertically
add r0, r0, r1
add r6, r6, r1
add r2, r2, r3
add r7, r7, r3
subs r4, r4, #2
bne 1b
.if \size >= 16
vpop {q4-q6}
.endif
pop {r4-r7}
bx lr
endfunc
.endm
.macro do_8tap_h_size size
do_8tap_h put, \size, 3, 4
do_8tap_h avg, \size, 3, 4
do_8tap_h put, \size, 4, 3
do_8tap_h avg, \size, 4, 3
.endm
do_8tap_h_size 4
do_8tap_h_size 8
do_8tap_h_size 16
.macro do_8tap_h_func type, filter, offset, size
function ff_vp9_\type\()_\filter\()\size\()_h_neon, export=1
push {r4-r7}
.if \size >= 16
vpush {q4-q6}
ldr r4, [sp, #64]
ldr r5, [sp, #68]
.else
ldr r4, [sp, #16]
ldr r5, [sp, #20]
.endif
movrelx r12, X(ff_vp9_subpel_filters), r6
add r12, r12, 256*\offset
cmp r5, #8
add r12, r12, r5, lsl #4
mov r5, #\size
arm: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. The filter coefficients are signed values, where the product of the multiplication with one individual filter coefficient doesn't overflow a 16 bit signed value (the largest filter coefficient is 127). But when the products are accumulated, the resulting sum can overflow the 16 bit signed range. Instead of accumulating in 32 bit, we accumulate the largest product (either index 3 or 4) last with a saturated addition. (The VP8 MC asm does something similar, but slightly simpler, by accumulating each half of the filter separately. In the VP9 MC filters, each half of the filter can also overflow though, so the largest component has to be handled individually.) Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_avg4_neon: 1.71 1.15 1.42 1.49 vp9_avg8_neon: 2.51 3.63 3.14 2.58 vp9_avg16_neon: 2.95 6.76 3.01 2.84 vp9_avg32_neon: 3.29 6.64 2.85 3.00 vp9_avg64_neon: 3.47 6.67 3.14 2.80 vp9_avg_8tap_smooth_4h_neon: 3.22 4.73 2.76 4.67 vp9_avg_8tap_smooth_4hv_neon: 3.67 4.76 3.28 4.71 vp9_avg_8tap_smooth_4v_neon: 5.52 7.60 4.60 6.31 vp9_avg_8tap_smooth_8h_neon: 6.22 9.04 5.12 9.32 vp9_avg_8tap_smooth_8hv_neon: 6.38 8.21 5.72 8.17 vp9_avg_8tap_smooth_8v_neon: 9.22 12.66 8.15 11.10 vp9_avg_8tap_smooth_64h_neon: 7.02 10.23 5.54 11.58 vp9_avg_8tap_smooth_64hv_neon: 6.76 9.46 5.93 9.40 vp9_avg_8tap_smooth_64v_neon: 10.76 14.13 9.46 13.37 vp9_put4_neon: 1.11 1.47 1.00 1.21 vp9_put8_neon: 1.23 2.17 1.94 1.48 vp9_put16_neon: 1.63 4.02 1.73 1.97 vp9_put32_neon: 1.56 4.92 2.00 1.96 vp9_put64_neon: 2.10 5.28 2.03 2.35 vp9_put_8tap_smooth_4h_neon: 3.11 4.35 2.63 4.35 vp9_put_8tap_smooth_4hv_neon: 3.67 4.69 3.25 4.71 vp9_put_8tap_smooth_4v_neon: 5.45 7.27 4.49 6.52 vp9_put_8tap_smooth_8h_neon: 5.97 8.18 4.81 8.56 vp9_put_8tap_smooth_8hv_neon: 6.39 7.90 5.64 8.15 vp9_put_8tap_smooth_8v_neon: 9.03 11.84 8.07 11.51 vp9_put_8tap_smooth_64h_neon: 6.78 9.48 4.88 10.89 vp9_put_8tap_smooth_64hv_neon: 6.99 8.87 5.94 9.56 vp9_put_8tap_smooth_64v_neon: 10.69 13.30 9.43 14.34 For the larger 8tap filters, the speedup vs C code is around 5-14x. This is significantly faster than libvpx's implementation of the same functions, at least when comparing the put_8tap_smooth_64 functions (compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from libvpx). Absolute runtimes from checkasm: Cortex A7 A8 A9 A53 vp9_put_8tap_smooth_64h_neon: 20150.3 14489.4 19733.6 10863.7 libvpx vpx_convolve8_horiz_neon: 52623.3 19736.4 21907.7 25027.7 vp9_put_8tap_smooth_64v_neon: 14455.0 12303.9 13746.4 9628.9 libvpx vpx_convolve8_vert_neon: 42090.0 17706.2 17659.9 16941.2 Thus, on the A9, the horizontal filter is only marginally faster than libvpx, while our version is significantly faster on the other cores, and the vertical filter is significantly faster on all cores. The difference is especially large on the A7. The libvpx implementation does the accumulation in 32 bit, which probably explains most of the differences. This is an adapted cherry-pick from libav commits ffbd1d2b0002576ef0d976a41ff959c635373fdc, 392caa65df3efa8b2d48a80f08a6af4892c61c08, 557c1675cf0e803b2fee43b4c8b58433842c84d0 and 11623217e3c9b859daee544e31acdd0821b61039. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:21 +02:00
.if \size >= 16
bge \type\()_8tap_16h_34
b \type\()_8tap_16h_43
.else
bge \type\()_8tap_\size\()h_34
b \type\()_8tap_\size\()h_43
.endif
endfunc
.endm
.macro do_8tap_h_filters size
do_8tap_h_func put, regular, 1, \size
do_8tap_h_func avg, regular, 1, \size
do_8tap_h_func put, sharp, 2, \size
do_8tap_h_func avg, sharp, 2, \size
do_8tap_h_func put, smooth, 0, \size
do_8tap_h_func avg, smooth, 0, \size
.endm
do_8tap_h_filters 64
do_8tap_h_filters 32
do_8tap_h_filters 16
do_8tap_h_filters 8
do_8tap_h_filters 4
.ltorg
@ Vertical filters
@ Round, shift and saturate and store qreg1-2 over 4 lines
.macro do_store4 qreg1, dreg1, qreg2, dreg2, tmp1, tmp2, type
vqrshrun.s16 \dreg1, \qreg1, #7
vqrshrun.s16 \dreg2, \qreg2, #7
.ifc \type,avg
vld1.32 {\tmp1[]}, [r0,:32], r1
vld1.32 {\tmp2[]}, [r0,:32], r1
vld1.32 {\tmp1[1]}, [r0,:32], r1
vld1.32 {\tmp2[1]}, [r0,:32], r1
vrhadd.u8 \dreg1, \dreg1, \tmp1
vrhadd.u8 \dreg2, \dreg2, \tmp2
sub r0, r0, r1, lsl #2
.endif
vst1.32 {\dreg1[0]}, [r0,:32], r1
vst1.32 {\dreg2[0]}, [r0,:32], r1
vst1.32 {\dreg1[1]}, [r0,:32], r1
vst1.32 {\dreg2[1]}, [r0,:32], r1
.endm
@ Round, shift and saturate and store qreg1-4
.macro do_store qreg1, dreg1, qreg2, dreg2, qreg3, dreg3, qreg4, dreg4, tmp1, tmp2, tmp3, tmp4, type
vqrshrun.s16 \dreg1, \qreg1, #7
vqrshrun.s16 \dreg2, \qreg2, #7
vqrshrun.s16 \dreg3, \qreg3, #7
vqrshrun.s16 \dreg4, \qreg4, #7
.ifc \type,avg
vld1.8 {\tmp1}, [r0,:64], r1
vld1.8 {\tmp2}, [r0,:64], r1
vld1.8 {\tmp3}, [r0,:64], r1
vld1.8 {\tmp4}, [r0,:64], r1
vrhadd.u8 \dreg1, \dreg1, \tmp1
vrhadd.u8 \dreg2, \dreg2, \tmp2
vrhadd.u8 \dreg3, \dreg3, \tmp3
vrhadd.u8 \dreg4, \dreg4, \tmp4
sub r0, r0, r1, lsl #2
.endif
vst1.8 {\dreg1}, [r0,:64], r1
vst1.8 {\dreg2}, [r0,:64], r1
vst1.8 {\dreg3}, [r0,:64], r1
vst1.8 {\dreg4}, [r0,:64], r1
.endm
@ Evaluate the filter twice in parallel, from the inputs src1-src9 into dst1-dst2
@ (src1-src8 into dst1, src2-src9 into dst2), adding idx2 separately
@ at the end with saturation. Indices 0 and 7 always have negative or zero
@ coefficients, so they can be accumulated into tmp1-tmp2 together with the
@ largest coefficient.
.macro convolve dst1, dst2, src1, src2, src3, src4, src5, src6, src7, src8, src9, idx1, idx2, tmp1, tmp2
vmul.s16 \dst1, \src2, d0[1]
vmul.s16 \dst2, \src3, d0[1]
vmul.s16 \tmp1, \src1, d0[0]
vmul.s16 \tmp2, \src2, d0[0]
vmla.s16 \dst1, \src3, d0[2]
vmla.s16 \dst2, \src4, d0[2]
.if \idx1 == 3
vmla.s16 \dst1, \src4, d0[3]
vmla.s16 \dst2, \src5, d0[3]
.else
vmla.s16 \dst1, \src5, d1[0]
vmla.s16 \dst2, \src6, d1[0]
.endif
vmla.s16 \dst1, \src6, d1[1]
vmla.s16 \dst2, \src7, d1[1]
vmla.s16 \tmp1, \src8, d1[3]
vmla.s16 \tmp2, \src9, d1[3]
vmla.s16 \dst1, \src7, d1[2]
vmla.s16 \dst2, \src8, d1[2]
.if \idx2 == 3
vmla.s16 \tmp1, \src4, d0[3]
vmla.s16 \tmp2, \src5, d0[3]
.else
vmla.s16 \tmp1, \src5, d1[0]
vmla.s16 \tmp2, \src6, d1[0]
.endif
vqadd.s16 \dst1, \dst1, \tmp1
vqadd.s16 \dst2, \dst2, \tmp2
.endm
@ Load pixels and extend them to 16 bit
.macro loadl dst1, dst2, dst3, dst4
vld1.8 {d2}, [r2], r3
vld1.8 {d3}, [r2], r3
vld1.8 {d4}, [r2], r3
.ifnb \dst4
vld1.8 {d5}, [r2], r3
.endif
vmovl.u8 \dst1, d2
vmovl.u8 \dst2, d3
vmovl.u8 \dst3, d4
.ifnb \dst4
vmovl.u8 \dst4, d5
.endif
.endm
@ Instantiate a vertical filter function for filtering 8 pixels at a time.
@ The height is passed in r4, the width in r5 and the filter coefficients
@ in r12. idx2 is the index of the largest filter coefficient (3 or 4)
@ and idx1 is the other one of them.
.macro do_8tap_8v type, idx1, idx2
function \type\()_8tap_8v_\idx1\idx2
sub r2, r2, r3, lsl #1
sub r2, r2, r3
vld1.16 {q0}, [r12, :128]
1:
mov r12, r4
arm: vp9: Add NEON optimizations of VP9 MC functions This work is sponsored by, and copyright, Google. The filter coefficients are signed values, where the product of the multiplication with one individual filter coefficient doesn't overflow a 16 bit signed value (the largest filter coefficient is 127). But when the products are accumulated, the resulting sum can overflow the 16 bit signed range. Instead of accumulating in 32 bit, we accumulate the largest product (either index 3 or 4) last with a saturated addition. (The VP8 MC asm does something similar, but slightly simpler, by accumulating each half of the filter separately. In the VP9 MC filters, each half of the filter can also overflow though, so the largest component has to be handled individually.) Examples of relative speedup compared to the C version, from checkasm: Cortex A7 A8 A9 A53 vp9_avg4_neon: 1.71 1.15 1.42 1.49 vp9_avg8_neon: 2.51 3.63 3.14 2.58 vp9_avg16_neon: 2.95 6.76 3.01 2.84 vp9_avg32_neon: 3.29 6.64 2.85 3.00 vp9_avg64_neon: 3.47 6.67 3.14 2.80 vp9_avg_8tap_smooth_4h_neon: 3.22 4.73 2.76 4.67 vp9_avg_8tap_smooth_4hv_neon: 3.67 4.76 3.28 4.71 vp9_avg_8tap_smooth_4v_neon: 5.52 7.60 4.60 6.31 vp9_avg_8tap_smooth_8h_neon: 6.22 9.04 5.12 9.32 vp9_avg_8tap_smooth_8hv_neon: 6.38 8.21 5.72 8.17 vp9_avg_8tap_smooth_8v_neon: 9.22 12.66 8.15 11.10 vp9_avg_8tap_smooth_64h_neon: 7.02 10.23 5.54 11.58 vp9_avg_8tap_smooth_64hv_neon: 6.76 9.46 5.93 9.40 vp9_avg_8tap_smooth_64v_neon: 10.76 14.13 9.46 13.37 vp9_put4_neon: 1.11 1.47 1.00 1.21 vp9_put8_neon: 1.23 2.17 1.94 1.48 vp9_put16_neon: 1.63 4.02 1.73 1.97 vp9_put32_neon: 1.56 4.92 2.00 1.96 vp9_put64_neon: 2.10 5.28 2.03 2.35 vp9_put_8tap_smooth_4h_neon: 3.11 4.35 2.63 4.35 vp9_put_8tap_smooth_4hv_neon: 3.67 4.69 3.25 4.71 vp9_put_8tap_smooth_4v_neon: 5.45 7.27 4.49 6.52 vp9_put_8tap_smooth_8h_neon: 5.97 8.18 4.81 8.56 vp9_put_8tap_smooth_8hv_neon: 6.39 7.90 5.64 8.15 vp9_put_8tap_smooth_8v_neon: 9.03 11.84 8.07 11.51 vp9_put_8tap_smooth_64h_neon: 6.78 9.48 4.88 10.89 vp9_put_8tap_smooth_64hv_neon: 6.99 8.87 5.94 9.56 vp9_put_8tap_smooth_64v_neon: 10.69 13.30 9.43 14.34 For the larger 8tap filters, the speedup vs C code is around 5-14x. This is significantly faster than libvpx's implementation of the same functions, at least when comparing the put_8tap_smooth_64 functions (compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from libvpx). Absolute runtimes from checkasm: Cortex A7 A8 A9 A53 vp9_put_8tap_smooth_64h_neon: 20150.3 14489.4 19733.6 10863.7 libvpx vpx_convolve8_horiz_neon: 52623.3 19736.4 21907.7 25027.7 vp9_put_8tap_smooth_64v_neon: 14455.0 12303.9 13746.4 9628.9 libvpx vpx_convolve8_vert_neon: 42090.0 17706.2 17659.9 16941.2 Thus, on the A9, the horizontal filter is only marginally faster than libvpx, while our version is significantly faster on the other cores, and the vertical filter is significantly faster on all cores. The difference is especially large on the A7. The libvpx implementation does the accumulation in 32 bit, which probably explains most of the differences. This is an adapted cherry-pick from libav commits ffbd1d2b0002576ef0d976a41ff959c635373fdc, 392caa65df3efa8b2d48a80f08a6af4892c61c08, 557c1675cf0e803b2fee43b4c8b58433842c84d0 and 11623217e3c9b859daee544e31acdd0821b61039. Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>
2016-11-14 12:32:21 +02:00
loadl q5, q6, q7
loadl q8, q9, q10, q11
2:
loadl q12, q13, q14, q15
convolve q1, q2, q5, q6, q7, q8, q9, q10, q11, q12, q13, \idx1, \idx2, q4, q5
convolve q3, q4, q7, q8, q9, q10, q11, q12, q13, q14, q15, \idx1, \idx2, q5, q6
do_store q1, d2, q2, d4, q3, d6, q4, d8, d3, d5, d7, d9, \type
subs r12, r12, #4
beq 8f
loadl q4, q5, q6, q7
convolve q1, q2, q9, q10, q11, q12, q13, q14, q15, q4, q5, \idx1, \idx2, q8, q9
convolve q3, q8, q11, q12, q13, q14, q15, q4, q5, q6, q7, \idx1, \idx2, q9, q10
do_store q1, d2, q2, d4, q3, d6, q8, d16, d3, d5, d7, d17, \type
subs r12, r12, #4
beq 8f
loadl q8, q9, q10, q11
convolve q1, q2, q13, q14, q15, q4, q5, q6, q7, q8, q9, \idx1, \idx2, q12, q13
convolve q3, q12, q15, q4, q5, q6, q7, q8, q9, q10, q11, \idx1, \idx2, q13, q14
do_store q1, d2, q2, d4, q3, d6, q12, d24, d3, d5, d7, d25, \type
subs r12, r12, #4
bne 2b
8:
subs r5, r5, #8
beq 9f
@ r0 -= h * dst_stride
mls r0, r1, r4, r0
@ r2 -= h * src_stride
mls r2, r3, r4, r2
@ r2 -= 8 * src_stride
sub r2, r2, r3, lsl #3
@ r2 += 1 * src_stride
add r2, r2, r3
add r2, r2, #8
add r0, r0, #8
b 1b
9:
vpop {q4-q7}
pop {r4-r5}
bx lr
endfunc
.endm
do_8tap_8v put, 3, 4
do_8tap_8v put, 4, 3
do_8tap_8v avg, 3, 4
do_8tap_8v avg, 4, 3
@ Instantiate a vertical filter function for filtering a 4 pixels wide
@ slice. The first half of the registers contain one row, while the second
@ half of a register contains the second-next row (also stored in the first
@ half of the register two steps ahead). The convolution does two outputs
@ at a time; the output of q5-q12 into one, and q4-q13 into another one.
@ The first half of first output is the first output row, the first half
@ of the other output is the second output row. The second halves of the
@ registers are rows 3 and 4.
@ This only is designed to work for 4 or 8 output lines.
.macro do_8tap_4v type, idx1, idx2
function \type\()_8tap_4v_\idx1\idx2
sub r2, r2, r3, lsl #1
sub r2, r2, r3
vld1.16 {q0}, [r12, :128]
vld1.32 {d2[]}, [r2], r3
vld1.32 {d3[]}, [r2], r3
vld1.32 {d4[]}, [r2], r3
vld1.32 {d5[]}, [r2], r3
vld1.32 {d6[]}, [r2], r3
vld1.32 {d7[]}, [r2], r3
vext.8 d2, d2, d4, #4
vld1.32 {d8[]}, [r2], r3
vext.8 d3, d3, d5, #4
vld1.32 {d9[]}, [r2], r3
vmovl.u8 q5, d2
vext.8 d4, d4, d6, #4
vld1.32 {d28[]}, [r2], r3
vmovl.u8 q6, d3
vext.8 d5, d5, d7, #4
vld1.32 {d29[]}, [r2], r3
vmovl.u8 q7, d4
vext.8 d6, d6, d8, #4
vld1.32 {d30[]}, [r2], r3
vmovl.u8 q8, d5
vext.8 d7, d7, d9, #4
vmovl.u8 q9, d6
vext.8 d8, d8, d28, #4
vmovl.u8 q10, d7
vext.8 d9, d9, d29, #4
vmovl.u8 q11, d8
vext.8 d28, d28, d30, #4
vmovl.u8 q12, d9
vmovl.u8 q13, d28
convolve q1, q2, q5, q6, q7, q8, q9, q10, q11, q12, q13, \idx1, \idx2, q4, q3
do_store4 q1, d2, q2, d4, d3, d5, \type
subs r4, r4, #4
beq 9f
vld1.32 {d2[]}, [r2], r3
vld1.32 {d3[]}, [r2], r3
vext.8 d29, d29, d2, #4
vext.8 d30, d30, d3, #4
vld1.32 {d2[1]}, [r2], r3
vmovl.u8 q14, d29
vld1.32 {d3[1]}, [r2], r3
vmovl.u8 q15, d30
vmovl.u8 q5, d2
vmovl.u8 q6, d3
convolve q1, q2, q9, q10, q11, q12, q13, q14, q15, q5, q6, \idx1, \idx2, q4, q3
do_store4 q1, d2, q2, d4, d3, d5, \type
9:
vpop {q4-q7}
pop {r4-r5}
bx lr
endfunc
.endm
do_8tap_4v put, 3, 4
do_8tap_4v put, 4, 3
do_8tap_4v avg, 3, 4
do_8tap_4v avg, 4, 3
.macro do_8tap_v_func type, filter, offset, size
function ff_vp9_\type\()_\filter\()\size\()_v_neon, export=1
push {r4-r5}
vpush {q4-q7}
ldr r4, [sp, #72]
movrelx r12, X(ff_vp9_subpel_filters), r5
ldr r5, [sp, #80]
add r12, r12, 256*\offset
add r12, r12, r5, lsl #4
cmp r5, #8
mov r5, #\size
.if \size >= 8
bge \type\()_8tap_8v_34
b \type\()_8tap_8v_43
.else
bge \type\()_8tap_4v_34
b \type\()_8tap_4v_43
.endif
endfunc
.endm
.macro do_8tap_v_filters size
do_8tap_v_func put, regular, 1, \size
do_8tap_v_func avg, regular, 1, \size
do_8tap_v_func put, sharp, 2, \size
do_8tap_v_func avg, sharp, 2, \size
do_8tap_v_func put, smooth, 0, \size
do_8tap_v_func avg, smooth, 0, \size
.endm
do_8tap_v_filters 64
do_8tap_v_filters 32
do_8tap_v_filters 16
do_8tap_v_filters 8
do_8tap_v_filters 4