[FFmpeg-devel] [PATCH] Moves yuv2yuvX_sse3 to yasm, unrolls main loop and other small optimizations for ~20% speedup.
Alan Kelly
alankelly at google.com
Tue Jan 5 14:31:25 EET 2021
Ping!
On Thu, Dec 17, 2020 at 11:42 AM Alan Kelly <alankelly at google.com> wrote:
> ---
> Fixes memory alignment problem in checkasm-sw_scale
> Tested on Linux 32 and 64 bit and mingw32
> libswscale/x86/Makefile | 1 +
> libswscale/x86/swscale.c | 106 +++++++++-----------------------
> libswscale/x86/yuv2yuvX.asm | 117 ++++++++++++++++++++++++++++++++++++
> tests/checkasm/sw_scale.c | 98 ++++++++++++++++++++++++++++++
> 4 files changed, 246 insertions(+), 76 deletions(-)
> create mode 100644 libswscale/x86/yuv2yuvX.asm
>
> diff --git a/libswscale/x86/Makefile b/libswscale/x86/Makefile
> index 831d5359aa..bfe383364e 100644
> --- a/libswscale/x86/Makefile
> +++ b/libswscale/x86/Makefile
> @@ -13,3 +13,4 @@ X86ASM-OBJS += x86/input.o
> \
> x86/scale.o \
> x86/rgb_2_rgb.o \
> x86/yuv_2_rgb.o \
> + x86/yuv2yuvX.o \
> diff --git a/libswscale/x86/swscale.c b/libswscale/x86/swscale.c
> index 3160fedf04..8cd8713705 100644
> --- a/libswscale/x86/swscale.c
> +++ b/libswscale/x86/swscale.c
> @@ -197,81 +197,30 @@ void ff_updateMMXDitherTables(SwsContext *c, int
> dstY)
> }
>
> #if HAVE_MMXEXT
> -static void yuv2yuvX_sse3(const int16_t *filter, int filterSize,
> - const int16_t **src, uint8_t *dest, int dstW,
> - const uint8_t *dither, int offset)
> -{
> - if(((uintptr_t)dest) & 15){
> - yuv2yuvX_mmxext(filter, filterSize, src, dest, dstW, dither,
> offset);
> - return;
> - }
> - filterSize--;
> -#define MAIN_FUNCTION \
> - "pxor %%xmm0, %%xmm0 \n\t" \
> - "punpcklbw %%xmm0, %%xmm3 \n\t" \
> - "movd %4, %%xmm1 \n\t" \
> - "punpcklwd %%xmm1, %%xmm1 \n\t" \
> - "punpckldq %%xmm1, %%xmm1 \n\t" \
> - "punpcklqdq %%xmm1, %%xmm1 \n\t" \
> - "psllw $3, %%xmm1 \n\t" \
> - "paddw %%xmm1, %%xmm3 \n\t" \
> - "psraw $4, %%xmm3 \n\t" \
> - "movdqa %%xmm3, %%xmm4 \n\t" \
> - "movdqa %%xmm3, %%xmm7 \n\t" \
> - "movl %3, %%ecx \n\t" \
> - "mov %0, %%"FF_REG_d"
> \n\t"\
> - "mov (%%"FF_REG_d"), %%"FF_REG_S"
> \n\t"\
> - ".p2align 4 \n\t" /*
> FIXME Unroll? */\
> - "1: \n\t"\
> - "movddup 8(%%"FF_REG_d"), %%xmm0 \n\t" /*
> filterCoeff */\
> - "movdqa (%%"FF_REG_S", %%"FF_REG_c", 2), %%xmm2
> \n\t" /* srcData */\
> - "movdqa 16(%%"FF_REG_S", %%"FF_REG_c", 2), %%xmm5
> \n\t" /* srcData */\
> - "add $16, %%"FF_REG_d"
> \n\t"\
> - "mov (%%"FF_REG_d"), %%"FF_REG_S"
> \n\t"\
> - "test %%"FF_REG_S", %%"FF_REG_S"
> \n\t"\
> - "pmulhw %%xmm0, %%xmm2 \n\t"\
> - "pmulhw %%xmm0, %%xmm5 \n\t"\
> - "paddw %%xmm2, %%xmm3 \n\t"\
> - "paddw %%xmm5, %%xmm4 \n\t"\
> - " jnz 1b \n\t"\
> - "psraw $3, %%xmm3 \n\t"\
> - "psraw $3, %%xmm4 \n\t"\
> - "packuswb %%xmm4, %%xmm3 \n\t"\
> - "movntdq %%xmm3, (%1, %%"FF_REG_c")
> \n\t"\
> - "add $16, %%"FF_REG_c" \n\t"\
> - "cmp %2, %%"FF_REG_c" \n\t"\
> - "movdqa %%xmm7, %%xmm3 \n\t" \
> - "movdqa %%xmm7, %%xmm4 \n\t" \
> - "mov %0, %%"FF_REG_d"
> \n\t"\
> - "mov (%%"FF_REG_d"), %%"FF_REG_S"
> \n\t"\
> - "jb 1b \n\t"
> -
> - if (offset) {
> - __asm__ volatile(
> - "movq %5, %%xmm3 \n\t"
> - "movdqa %%xmm3, %%xmm4 \n\t"
> - "psrlq $24, %%xmm3 \n\t"
> - "psllq $40, %%xmm4 \n\t"
> - "por %%xmm4, %%xmm3 \n\t"
> - MAIN_FUNCTION
> - :: "g" (filter),
> - "r" (dest-offset), "g" ((x86_reg)(dstW+offset)), "m"
> (offset),
> - "m"(filterSize), "m"(((uint64_t *) dither)[0])
> - : XMM_CLOBBERS("%xmm0" , "%xmm1" , "%xmm2" , "%xmm3" ,
> "%xmm4" , "%xmm5" , "%xmm7" ,)
> - "%"FF_REG_d, "%"FF_REG_S, "%"FF_REG_c
> - );
> - } else {
> - __asm__ volatile(
> - "movq %5, %%xmm3 \n\t"
> - MAIN_FUNCTION
> - :: "g" (filter),
> - "r" (dest-offset), "g" ((x86_reg)(dstW+offset)), "m"
> (offset),
> - "m"(filterSize), "m"(((uint64_t *) dither)[0])
> - : XMM_CLOBBERS("%xmm0" , "%xmm1" , "%xmm2" , "%xmm3" ,
> "%xmm4" , "%xmm5" , "%xmm7" ,)
> - "%"FF_REG_d, "%"FF_REG_S, "%"FF_REG_c
> - );
> - }
> +#define YUV2YUVX_FUNC(opt, step) \
> +void ff_yuv2yuvX_ ##opt(const int16_t *filter, long filterSize, const
> int16_t **src, \
> + uint8_t *dest, int dstW, \
> + const uint8_t *dither, int offset); \
> +static void yuv2yuvX_ ##opt(const int16_t *filter, int filterSize, \
> + const int16_t **src, uint8_t *dest, int dstW, \
> + const uint8_t *dither, int offset) \
> +{ \
> + int remainder = (dstW % step); \
> + int pixelsProcessed = dstW - remainder; \
> + if(((uintptr_t)dest) & 15){ \
> + yuv2yuvX_mmx(filter, filterSize, src, dest, dstW, dither,
> offset); \
> + return; \
> + } \
> + ff_yuv2yuvX_ ##opt(filter, filterSize - 1, src, dest - offset,
> pixelsProcessed + offset, dither, offset); \
> + if(remainder > 0){ \
> + yuv2yuvX_mmx(filter, filterSize, src, dest + pixelsProcessed,
> remainder, dither, offset + pixelsProcessed); \
> + } \
> + return; \
> }
> +
> +YUV2YUVX_FUNC(sse3, 32)
> +YUV2YUVX_FUNC(avx2, 64)
> +
> #endif
>
> #endif /* HAVE_INLINE_ASM */
> @@ -402,9 +351,14 @@ av_cold void ff_sws_init_swscale_x86(SwsContext *c)
> #if HAVE_MMXEXT_INLINE
> if (INLINE_MMXEXT(cpu_flags))
> sws_init_swscale_mmxext(c);
> - if (cpu_flags & AV_CPU_FLAG_SSE3){
> - if(c->use_mmx_vfilter && !(c->flags & SWS_ACCURATE_RND))
> + if (cpu_flags & AV_CPU_FLAG_AVX2){
> + if(c->use_mmx_vfilter && !(c->flags & SWS_ACCURATE_RND)){
> + c->yuv2planeX = yuv2yuvX_avx2;
> + }
> + } else if (cpu_flags & AV_CPU_FLAG_SSE3){
> + if(c->use_mmx_vfilter && !(c->flags & SWS_ACCURATE_RND)){
> c->yuv2planeX = yuv2yuvX_sse3;
> + }
> }
> #endif
>
> diff --git a/libswscale/x86/yuv2yuvX.asm b/libswscale/x86/yuv2yuvX.asm
> new file mode 100644
> index 0000000000..899b84c50b
> --- /dev/null
> +++ b/libswscale/x86/yuv2yuvX.asm
> @@ -0,0 +1,117 @@
>
> +;******************************************************************************
> +;* x86-optimized yuv2yuvX
> +;* Copyright 2020 Google LLC
> +;* Copyright (C) 2001-2011 Michael Niedermayer <michaelni at gmx.at>
> +;*
> +;* 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/x86/x86util.asm"
> +
> +SECTION .text
> +
>
> +;-----------------------------------------------------------------------------
> +; yuv2yuvX
> +;
> +; void ff_yuv2yuvX_<opt>(const int16_t *filter, int filterSize,
> +; uint8_t *dest, int dstW,
> +; const uint8_t *dither, int offset);
> +;
>
> +;-----------------------------------------------------------------------------
> +
> +%macro YUV2YUVX_FUNC 0
> +cglobal yuv2yuvX, 7, 7, 8, filter, filterSize, src, dest, dstW, dither,
> offset
> +%if ARCH_X86_64
> + movsxd dstWq, dstWd
> + movsxd offsetq, offsetd
> +%endif ; x86-64
> + movddup m0, [filterq + 8]
> +%if cpuflag(avx2)
> + vpbroadcastq m3, [ditherq]
> +%else
> + movq xmm3, [ditherq]
> +%endif ; avx2
> + cmp offsetd, 0
> + jz .offset
> +
> + ; offset != 0 path.
> + psrlq m5, m3, $18
> + psllq m3, m3, $28
> + por m3, m3, m5
> +
> +.offset:
> + movd xmm1, filterSized
> +%if cpuflag(avx2)
> + vpbroadcastw m1, xmm1
> +%else
> + pshuflw m1, m1, q0000
> + punpcklqdq m1, m1
> +%endif ; avx2
> + pxor m0, m0, m0
> + mov filterSizeq, filterq
> + mov srcq, [filterSizeq]
> + punpcklbw m3, m0
> + psllw m1, m1, 3
> + paddw m3, m3, m1
> + psraw m7, m3, 4
> +.outerloop:
> + mova m4, m7
> + mova m3, m7
> + mova m6, m7
> + mova m1, m7
> +.loop:
> +%if cpuflag(avx2)
> + vpbroadcastq m0, [filterSizeq + 8]
> +%else
> + movddup m0, [filterSizeq + 8]
> +%endif
> + pmulhw m2, m0, [srcq + offsetq * 2]
> + pmulhw m5, m0, [srcq + offsetq * 2 + mmsize]
> + paddw m3, m3, m2
> + paddw m4, m4, m5
> + pmulhw m2, m0, [srcq + offsetq * 2 + 2 * mmsize]
> + pmulhw m5, m0, [srcq + offsetq * 2 + 3 * mmsize]
> + paddw m6, m6, m2
> + paddw m1, m1, m5
> + add filterSizeq, $10
> + mov srcq, [filterSizeq]
> + test srcq, srcq
> + jnz .loop
> + psraw m3, m3, 3
> + psraw m4, m4, 3
> + psraw m6, m6, 3
> + psraw m1, m1, 3
> + packuswb m3, m3, m4
> + packuswb m6, m6, m1
> + mov srcq, [filterq]
> +%if cpuflag(avx2)
> + vpermq m3, m3, 216
> + vpermq m6, m6, 216
> +%endif
> + mova [destq + offsetq], m3
> + mova [destq + offsetq + mmsize], m6
> + add offsetq, mmsize * 2
> + mov filterSizeq, filterq
> + cmp offsetq, dstWq
> + jb .outerloop
> + REP_RET
> +%endmacro
> +
> +INIT_XMM sse3
> +YUV2YUVX_FUNC
> +INIT_YMM avx2
> +YUV2YUVX_FUNC
> diff --git a/tests/checkasm/sw_scale.c b/tests/checkasm/sw_scale.c
> index 9efa2b4def..3bc6066f35 100644
> --- a/tests/checkasm/sw_scale.c
> +++ b/tests/checkasm/sw_scale.c
> @@ -37,6 +37,102 @@
>
> #define SRC_PIXELS 128
>
> +
> +// This reference function is the same approximate algorithm employed by
> the
> +// SIMD functions
> +static void ref_function(const int16_t *filter, int filterSize,
> + const int16_t **src,
> uint8_t *dest, int dstW,
> + const uint8_t *dither,
> int offset)
> +{
> + int i, d;
> + d = ((filterSize - 1) * 8 + dither[0]) >> 4;
> + for (i=0; i<dstW; i++) {
> + int16_t val = d;
> + int j;
> + union {
> + int val;
> + int16_t v[2];
> + } t;
> + for (j=0; j<filterSize; j++){
> + t.val = (int)src[j][i + offset] * (int)filter[j];
> + val += t.v[1];
> + }
> + dest[i]= av_clip_uint8(val>>3);
> + }
> +}
> +
> +static void check_yuv2yuvX(void)
> +{
> + struct SwsContext *ctx;
> + int fsi, osi, i, j;
> +#define LARGEST_FILTER 16
> +#define FILTER_SIZES 4
> + static const int filter_sizes[FILTER_SIZES] = {1, 4, 8, 16};
> +
> + declare_func_emms(AV_CPU_FLAG_MMX, void, const int16_t *filter,
> + int filterSize, const int16_t **src, uint8_t *dest,
> + int dstW, const uint8_t *dither, int offset);
> +
> + int dstW = SRC_PIXELS;
> + const int16_t **src;
> + LOCAL_ALIGNED_32(int16_t, src_pixels, [LARGEST_FILTER * SRC_PIXELS]);
> + LOCAL_ALIGNED_32(int16_t, filter_coeff, [LARGEST_FILTER]);
> + LOCAL_ALIGNED_32(uint8_t, dst0, [SRC_PIXELS]);
> + LOCAL_ALIGNED_32(uint8_t, dst1, [SRC_PIXELS]);
> + LOCAL_ALIGNED_32(uint8_t, dither, [SRC_PIXELS]);
> + union VFilterData{
> + const int16_t *src;
> + uint16_t coeff[8];
> + } *vFilterData;
> + uint8_t d_val = rnd();
> + randomize_buffers(filter_coeff, LARGEST_FILTER);
> + randomize_buffers(src_pixels, LARGEST_FILTER * SRC_PIXELS);
> + ctx = sws_alloc_context();
> + if (sws_init_context(ctx, NULL, NULL) < 0)
> + fail();
> +
> + ff_getSwsFunc(ctx);
> + for(i = 0; i < SRC_PIXELS; ++i){
> + dither[i] = d_val;
> + }
> + for(osi = 0; osi < 64; osi += 16){
> + for(fsi = 0; fsi < FILTER_SIZES; ++fsi){
> + src = av_malloc(sizeof(int16_t*) * filter_sizes[fsi]);
> + vFilterData = av_malloc((filter_sizes[fsi] + 2) *
> sizeof(union VFilterData));
> + memset(vFilterData, 0, (filter_sizes[fsi] + 2) * sizeof(union
> VFilterData));
> + for(i = 0; i < filter_sizes[fsi]; ++i){
> + src[i] = &src_pixels[i * SRC_PIXELS];
> + vFilterData[i].src = src[i];
> + for(j = 0; j < 4; ++j){
> + vFilterData[i].coeff[j + 4] = filter_coeff[i];
> + }
> + }
> + if (check_func(ctx->yuv2planeX, "yuv2yuvX_%d_%d",
> filter_sizes[fsi], osi)){
> + memset(dst0, 0, SRC_PIXELS * sizeof(dst0[0]));
> + memset(dst1, 0, SRC_PIXELS * sizeof(dst1[0]));
> +
> + // The reference function is not the scalar function
> selected when mmx
> + // is deactivated as the SIMD functions do not give the
> same result as
> + // the scalar ones due to rounding. The SIMD functions
> are activated by
> + // the flag SWS_ACCURATE_RND
> + ref_function(&filter_coeff[0], filter_sizes[fsi], src,
> dst0, dstW - osi, dither, osi);
> + // There's no point in calling new for the reference
> function
> + if(ctx->use_mmx_vfilter){
> + call_new((const int16_t*)vFilterData,
> filter_sizes[fsi], src, dst1, dstW - osi, dither, osi);
> + if (memcmp(dst0, dst1, SRC_PIXELS * sizeof(dst0[0]))){
> + fail();
> + }
> + bench_new((const int16_t*)vFilterData,
> filter_sizes[fsi], src, dst1, dstW - osi, dither, osi);
> + }
> + }
> + free(src);
> + free(vFilterData);
> + }
> + }
> + sws_freeContext(ctx);
> +#undef FILTER_SIZES
> +}
> +
> static void check_hscale(void)
> {
> #define MAX_FILTER_WIDTH 40
> @@ -131,4 +227,6 @@ void checkasm_check_sw_scale(void)
> {
> check_hscale();
> report("hscale");
> + check_yuv2yuvX();
> + report("yuv2yuvX");
> }
> --
> 2.29.2.684.gfbc64c5ab5-goog
>
>
More information about the ffmpeg-devel
mailing list