| Message ID | 20240604190742.1742-1-jamrial@gmail.com |
|---|---|
| State | New |
| Headers | show |
| Series | [FFmpeg-devel] swscale/x86/input: add AVX2 optimized RGB24 to YUV functions | expand |
| Context | Check | Description |
|---|---|---|
| yinshiyou/make_loongarch64 | success | Make finished |
| yinshiyou/make_fate_loongarch64 | success | Make fate finished |
| andriy/make_x86 | success | Make finished |
| andriy/make_fate_x86 | success | Make fate finished |
James Almer: > rgb24_to_uv_8_c: 39.3 > rgb24_to_uv_8_sse2: 14.3 > rgb24_to_uv_8_ssse3: 13.3 > rgb24_to_uv_8_avx: 12.8 > rgb24_to_uv_8_avx2: 14.3 Worse than avx and ssse3 > rgb24_to_uv_128_c: 582.8 > rgb24_to_uv_128_sse2: 127.3 > rgb24_to_uv_128_ssse3: 107.3 > rgb24_to_uv_128_avx: 111.3 Worse than ssse3 > rgb24_to_uv_128_avx2: 62.3 > rgb24_to_uv_1080_c: 4981.3 > rgb24_to_uv_1080_sse2: 1048.3 > rgb24_to_uv_1080_ssse3: 876.8 > rgb24_to_uv_1080_avx: 887.8 Worse than ssse3 > rgb24_to_uv_1080_avx2: 492.3 > rgb24_to_uv_1280_c: 5906.8 > rgb24_to_uv_1280_sse2: 1263.3 > rgb24_to_uv_1280_ssse3: 1048.3 > rgb24_to_uv_1280_avx: 1045.8 Not really better than ssse3 > rgb24_to_uv_1280_avx2: 579.8 > rgb24_to_uv_1920_c: 8665.3 > rgb24_to_uv_1920_sse2: 1888.8 > rgb24_to_uv_1920_ssse3: 1571.8 > rgb24_to_uv_1920_avx: 1558.8 Not really better than ssse3 > rgb24_to_uv_1920_avx2: 869.3 > rgb24_to_y_8_c: 20.3 > rgb24_to_y_8_sse2: 11.8 > rgb24_to_y_8_ssse3: 10.3 > rgb24_to_y_8_avx: 10.3 Not better than ssse3 > rgb24_to_y_8_avx2: 10.8 Worse than ssse3 > rgb24_to_y_128_c: 284.8 > rgb24_to_y_128_sse2: 83.3 > rgb24_to_y_128_ssse3: 66.8 > rgb24_to_y_128_avx: 64.8 Not really better than ssse3 > rgb24_to_y_128_avx2: 39.3 > rgb24_to_y_1080_c: 2451.3 > rgb24_to_y_1080_sse2: 696.3 > rgb24_to_y_1080_ssse3: 516.8 > rgb24_to_y_1080_avx: 518.8 Worse than ssse3 > rgb24_to_y_1080_avx2: 301.8 > rgb24_to_y_1280_c: 2892.8 > rgb24_to_y_1280_sse2: 816.8 > rgb24_to_y_1280_ssse3: 623.3 > rgb24_to_y_1280_avx: 616.3 Not really better than ssse3 > rgb24_to_y_1280_avx2: 350.8 > rgb24_to_y_1920_c: 4338.8 > rgb24_to_y_1920_sse2: 1210.8 > rgb24_to_y_1920_ssse3: 928.3 > rgb24_to_y_1920_avx: 920.3 Not really better than ssse3 > rgb24_to_y_1920_avx2: 534.8 > > Signed-off-by: James Almer <jamrial@gmail.com> > --- > libswscale/x86/input.asm | 49 ++++++++++++++++++++++++++++++++++++---- > libswscale/x86/swscale.c | 7 ++++++ > 2 files changed, 51 insertions(+), 5 deletions(-) > > diff --git a/libswscale/x86/input.asm b/libswscale/x86/input.asm > index a197183f1f..e79fe11405 100644 > --- a/libswscale/x86/input.asm > +++ b/libswscale/x86/input.asm > @@ -23,7 +23,7 @@ > > %include "libavutil/x86/x86util.asm" > > -SECTION_RODATA > +SECTION_RODATA 32 > > %define RY 0x20DE > %define GY 0x4087 > @@ -90,8 +90,12 @@ rgb_UVrnd: times 4 dd 0x400100 ; 128.5 << 15 > ; rgba_Vcoeff_ag: times 4 dw 0, GV > > shuf_rgb_12x4: db 0, 0x80, 1, 0x80, 2, 0x80, 3, 0x80, \ > + 6, 0x80, 7, 0x80, 8, 0x80, 9, 0x80, \ > + 0, 0x80, 1, 0x80, 2, 0x80, 3, 0x80, \ > 6, 0x80, 7, 0x80, 8, 0x80, 9, 0x80 > shuf_rgb_3x56: db 2, 0x80, 3, 0x80, 4, 0x80, 5, 0x80, \ > + 8, 0x80, 9, 0x80, 10, 0x80, 11, 0x80, \ > + 2, 0x80, 3, 0x80, 4, 0x80, 5, 0x80, \ > 8, 0x80, 9, 0x80, 10, 0x80, 11, 0x80 > pd_65535f: times 8 dd 65535.0 > pb_pack_shuffle16le: db 0, 1, 4, 5, \ > @@ -134,8 +138,13 @@ SECTION .text > %macro RGB24_TO_Y_FN 2-3 > cglobal %2 %+ 24ToY, 6, 6, %1, dst, src, u1, u2, w, table > %if ARCH_X86_64 > +%if mmsize == 32 > + vbroadcasti128 m8, [%2_Ycoeff_12x4] > + vbroadcasti128 m9, [%2_Ycoeff_3x56] > +%else > mova m8, [%2_Ycoeff_12x4] > mova m9, [%2_Ycoeff_3x56] > +%endif > %define coeff1 m8 > %define coeff2 m9 > %else ; x86-32 > @@ -165,11 +174,19 @@ cglobal %2 %+ 24ToY, 6, 6, %1, dst, src, u1, u2, w, table > %if notcpuflag(ssse3) > pxor m7, m7 > %endif ; !cpuflag(ssse3) > +%if mmsize == 32 > + vbroadcasti128 m4, [rgb_Yrnd] > +%else > mova m4, [rgb_Yrnd] > +%endif > .loop: > %if cpuflag(ssse3) > - movu m0, [srcq+0] ; (byte) { Bx, Gx, Rx }[0-3] > - movu m2, [srcq+12] ; (byte) { Bx, Gx, Rx }[4-7] > + movu xm0, [srcq+0] ; (byte) { Bx, Gx, Rx }[0-3] > + movu xm2, [srcq+12] ; (byte) { Bx, Gx, Rx }[4-7] > +%if mmsize == 32 > + vinserti128 m0, m0, [srcq+24], 1 > + vinserti128 m2, m2, [srcq+36], 1 > +%endif > pshufb m1, m0, shuf_rgb2 ; (word) { R0, B1, G1, R1, R2, B3, G3, R3 } > pshufb m0, shuf_rgb1 ; (word) { B0, G0, R0, B1, B2, G2, R2, B3 } > pshufb m3, m2, shuf_rgb2 ; (word) { R4, B5, G5, R5, R6, B7, G7, R7 } > @@ -216,10 +233,17 @@ cglobal %2 %+ 24ToY, 6, 6, %1, dst, src, u1, u2, w, table > %macro RGB24_TO_UV_FN 2-3 > cglobal %2 %+ 24ToUV, 7, 7, %1, dstU, dstV, u1, src, u2, w, table > %if ARCH_X86_64 > +%if mmsize == 32 > + vbroadcasti128 m8, [%2_Ucoeff_12x4] > + vbroadcasti128 m9, [%2_Ucoeff_3x56] > + vbroadcasti128 m10, [%2_Vcoeff_12x4] > + vbroadcasti128 m11, [%2_Vcoeff_3x56] > +%else > mova m8, [%2_Ucoeff_12x4] > mova m9, [%2_Ucoeff_3x56] > mova m10, [%2_Vcoeff_12x4] > mova m11, [%2_Vcoeff_3x56] > +%endif > %define coeffU1 m8 > %define coeffU2 m9 > %define coeffV1 m10 > @@ -253,14 +277,22 @@ cglobal %2 %+ 24ToUV, 7, 7, %1, dstU, dstV, u1, src, u2, w, table > add dstUq, wq > add dstVq, wq > neg wq > +%if mmsize == 32 > + vbroadcasti128 m6, [rgb_UVrnd] > +%else > mova m6, [rgb_UVrnd] > +%endif > %if notcpuflag(ssse3) > pxor m7, m7 > %endif > .loop: > %if cpuflag(ssse3) > - movu m0, [srcq+0] ; (byte) { Bx, Gx, Rx }[0-3] > - movu m4, [srcq+12] ; (byte) { Bx, Gx, Rx }[4-7] > + movu xm0, [srcq+0] ; (byte) { Bx, Gx, Rx }[0-3] > + movu xm4, [srcq+12] ; (byte) { Bx, Gx, Rx }[4-7] > +%if mmsize == 32 > + vinserti128 m0, m0, [srcq+24], 1 > + vinserti128 m4, m4, [srcq+36], 1 > +%endif > pshufb m1, m0, shuf_rgb2 ; (word) { R0, B1, G1, R1, R2, B3, G3, R3 } > pshufb m0, shuf_rgb1 ; (word) { B0, G0, R0, B1, B2, G2, R2, B3 } > %else ; !cpuflag(ssse3) > @@ -337,6 +369,13 @@ INIT_XMM avx > RGB24_FUNCS 11, 13 > %endif > > +%if ARCH_X86_64 > +%if HAVE_AVX2_EXTERNAL > +INIT_YMM avx2 > +RGB24_FUNCS 11, 13 > +%endif > +%endif > + > ; %1 = nr. of XMM registers > ; %2-5 = rgba, bgra, argb or abgr (in individual characters) > %macro RGB32_TO_Y_FN 5-6 > diff --git a/libswscale/x86/swscale.c b/libswscale/x86/swscale.c > index fff8bb4396..1438c077e6 100644 > --- a/libswscale/x86/swscale.c > +++ b/libswscale/x86/swscale.c > @@ -321,6 +321,8 @@ void ff_ ## fmt ## ToUV_ ## opt(uint8_t *dstU, uint8_t *dstV, \ > INPUT_FUNCS(sse2); > INPUT_FUNCS(ssse3); > INPUT_FUNCS(avx); > +INPUT_FUNC(rgb24, avx2); > +INPUT_FUNC(bgr24, avx2); > > #if ARCH_X86_64 > #define YUV2NV_DECL(fmt, opt) \ > @@ -634,6 +636,11 @@ switch(c->dstBpc){ \ > } > > if (EXTERNAL_AVX2_FAST(cpu_flags)) { > + if (ARCH_X86_64) > + switch (c->srcFormat) { > + case_rgb(rgb24, RGB24, avx2); > + case_rgb(bgr24, BGR24, avx2); > + } > switch (c->dstFormat) { > case AV_PIX_FMT_NV12: > case AV_PIX_FMT_NV24:
On 6/4/2024 4:25 PM, Andreas Rheinhardt wrote: > James Almer: >> rgb24_to_uv_8_c: 39.3 >> rgb24_to_uv_8_sse2: 14.3 >> rgb24_to_uv_8_ssse3: 13.3 >> rgb24_to_uv_8_avx: 12.8 >> rgb24_to_uv_8_avx2: 14.3 > > Worse than avx and ssse3 Can't be disabled for small sizes. > >> rgb24_to_uv_128_c: 582.8 >> rgb24_to_uv_128_sse2: 127.3 >> rgb24_to_uv_128_ssse3: 107.3 >> rgb24_to_uv_128_avx: 111.3 > > Worse than ssse3 I know, but it's not related to this patch. Same for all the comments below.
James Almer: > On 6/4/2024 4:25 PM, Andreas Rheinhardt wrote: >> James Almer: >>> rgb24_to_uv_8_c: 39.3 >>> rgb24_to_uv_8_sse2: 14.3 >>> rgb24_to_uv_8_ssse3: 13.3 >>> rgb24_to_uv_8_avx: 12.8 >>> rgb24_to_uv_8_avx2: 14.3 >> >> Worse than avx and ssse3 > > Can't be disabled for small sizes. > Why can't this be disabled? >> >>> rgb24_to_uv_128_c: 582.8 >>> rgb24_to_uv_128_sse2: 127.3 >>> rgb24_to_uv_128_ssse3: 107.3 >>> rgb24_to_uv_128_avx: 111.3 >> >> Worse than ssse3 > > I know, but it's not related to this patch. > Same for all the comments below.
On 6/4/2024 4:32 PM, Andreas Rheinhardt wrote: > James Almer: >> On 6/4/2024 4:25 PM, Andreas Rheinhardt wrote: >>> James Almer: >>>> rgb24_to_uv_8_c: 39.3 >>>> rgb24_to_uv_8_sse2: 14.3 >>>> rgb24_to_uv_8_ssse3: 13.3 >>>> rgb24_to_uv_8_avx: 12.8 >>>> rgb24_to_uv_8_avx2: 14.3 >>> >>> Worse than avx and ssse3 >> >> Can't be disabled for small sizes. >> > > Why can't this be disabled? The function pointer is for all sizes. And adding a check + jmp in the assembly is not going to be better. Also, this being for 8 pixel wide buffers, it's not going to be anyone's bottleneck. > >>> >>>> rgb24_to_uv_128_c: 582.8 >>>> rgb24_to_uv_128_sse2: 127.3 >>>> rgb24_to_uv_128_ssse3: 107.3 >>>> rgb24_to_uv_128_avx: 111.3 >>> >>> Worse than ssse3 >> >> I know, but it's not related to this patch. >> Same for all the comments below. > > _______________________________________________ > ffmpeg-devel mailing list > ffmpeg-devel@ffmpeg.org > https://ffmpeg.org/mailman/listinfo/ffmpeg-devel > > To unsubscribe, visit link above, or email > ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
diff --git a/libswscale/x86/input.asm b/libswscale/x86/input.asm index a197183f1f..e79fe11405 100644 --- a/libswscale/x86/input.asm +++ b/libswscale/x86/input.asm @@ -23,7 +23,7 @@ %include "libavutil/x86/x86util.asm" -SECTION_RODATA +SECTION_RODATA 32 %define RY 0x20DE %define GY 0x4087 @@ -90,8 +90,12 @@ rgb_UVrnd: times 4 dd 0x400100 ; 128.5 << 15 ; rgba_Vcoeff_ag: times 4 dw 0, GV shuf_rgb_12x4: db 0, 0x80, 1, 0x80, 2, 0x80, 3, 0x80, \ + 6, 0x80, 7, 0x80, 8, 0x80, 9, 0x80, \ + 0, 0x80, 1, 0x80, 2, 0x80, 3, 0x80, \ 6, 0x80, 7, 0x80, 8, 0x80, 9, 0x80 shuf_rgb_3x56: db 2, 0x80, 3, 0x80, 4, 0x80, 5, 0x80, \ + 8, 0x80, 9, 0x80, 10, 0x80, 11, 0x80, \ + 2, 0x80, 3, 0x80, 4, 0x80, 5, 0x80, \ 8, 0x80, 9, 0x80, 10, 0x80, 11, 0x80 pd_65535f: times 8 dd 65535.0 pb_pack_shuffle16le: db 0, 1, 4, 5, \ @@ -134,8 +138,13 @@ SECTION .text %macro RGB24_TO_Y_FN 2-3 cglobal %2 %+ 24ToY, 6, 6, %1, dst, src, u1, u2, w, table %if ARCH_X86_64 +%if mmsize == 32 + vbroadcasti128 m8, [%2_Ycoeff_12x4] + vbroadcasti128 m9, [%2_Ycoeff_3x56] +%else mova m8, [%2_Ycoeff_12x4] mova m9, [%2_Ycoeff_3x56] +%endif %define coeff1 m8 %define coeff2 m9 %else ; x86-32 @@ -165,11 +174,19 @@ cglobal %2 %+ 24ToY, 6, 6, %1, dst, src, u1, u2, w, table %if notcpuflag(ssse3) pxor m7, m7 %endif ; !cpuflag(ssse3) +%if mmsize == 32 + vbroadcasti128 m4, [rgb_Yrnd] +%else mova m4, [rgb_Yrnd] +%endif .loop: %if cpuflag(ssse3) - movu m0, [srcq+0] ; (byte) { Bx, Gx, Rx }[0-3] - movu m2, [srcq+12] ; (byte) { Bx, Gx, Rx }[4-7] + movu xm0, [srcq+0] ; (byte) { Bx, Gx, Rx }[0-3] + movu xm2, [srcq+12] ; (byte) { Bx, Gx, Rx }[4-7] +%if mmsize == 32 + vinserti128 m0, m0, [srcq+24], 1 + vinserti128 m2, m2, [srcq+36], 1 +%endif pshufb m1, m0, shuf_rgb2 ; (word) { R0, B1, G1, R1, R2, B3, G3, R3 } pshufb m0, shuf_rgb1 ; (word) { B0, G0, R0, B1, B2, G2, R2, B3 } pshufb m3, m2, shuf_rgb2 ; (word) { R4, B5, G5, R5, R6, B7, G7, R7 } @@ -216,10 +233,17 @@ cglobal %2 %+ 24ToY, 6, 6, %1, dst, src, u1, u2, w, table %macro RGB24_TO_UV_FN 2-3 cglobal %2 %+ 24ToUV, 7, 7, %1, dstU, dstV, u1, src, u2, w, table %if ARCH_X86_64 +%if mmsize == 32 + vbroadcasti128 m8, [%2_Ucoeff_12x4] + vbroadcasti128 m9, [%2_Ucoeff_3x56] + vbroadcasti128 m10, [%2_Vcoeff_12x4] + vbroadcasti128 m11, [%2_Vcoeff_3x56] +%else mova m8, [%2_Ucoeff_12x4] mova m9, [%2_Ucoeff_3x56] mova m10, [%2_Vcoeff_12x4] mova m11, [%2_Vcoeff_3x56] +%endif %define coeffU1 m8 %define coeffU2 m9 %define coeffV1 m10 @@ -253,14 +277,22 @@ cglobal %2 %+ 24ToUV, 7, 7, %1, dstU, dstV, u1, src, u2, w, table add dstUq, wq add dstVq, wq neg wq +%if mmsize == 32 + vbroadcasti128 m6, [rgb_UVrnd] +%else mova m6, [rgb_UVrnd] +%endif %if notcpuflag(ssse3) pxor m7, m7 %endif .loop: %if cpuflag(ssse3) - movu m0, [srcq+0] ; (byte) { Bx, Gx, Rx }[0-3] - movu m4, [srcq+12] ; (byte) { Bx, Gx, Rx }[4-7] + movu xm0, [srcq+0] ; (byte) { Bx, Gx, Rx }[0-3] + movu xm4, [srcq+12] ; (byte) { Bx, Gx, Rx }[4-7] +%if mmsize == 32 + vinserti128 m0, m0, [srcq+24], 1 + vinserti128 m4, m4, [srcq+36], 1 +%endif pshufb m1, m0, shuf_rgb2 ; (word) { R0, B1, G1, R1, R2, B3, G3, R3 } pshufb m0, shuf_rgb1 ; (word) { B0, G0, R0, B1, B2, G2, R2, B3 } %else ; !cpuflag(ssse3) @@ -337,6 +369,13 @@ INIT_XMM avx RGB24_FUNCS 11, 13 %endif +%if ARCH_X86_64 +%if HAVE_AVX2_EXTERNAL +INIT_YMM avx2 +RGB24_FUNCS 11, 13 +%endif +%endif + ; %1 = nr. of XMM registers ; %2-5 = rgba, bgra, argb or abgr (in individual characters) %macro RGB32_TO_Y_FN 5-6 diff --git a/libswscale/x86/swscale.c b/libswscale/x86/swscale.c index fff8bb4396..1438c077e6 100644 --- a/libswscale/x86/swscale.c +++ b/libswscale/x86/swscale.c @@ -321,6 +321,8 @@ void ff_ ## fmt ## ToUV_ ## opt(uint8_t *dstU, uint8_t *dstV, \ INPUT_FUNCS(sse2); INPUT_FUNCS(ssse3); INPUT_FUNCS(avx); +INPUT_FUNC(rgb24, avx2); +INPUT_FUNC(bgr24, avx2); #if ARCH_X86_64 #define YUV2NV_DECL(fmt, opt) \ @@ -634,6 +636,11 @@ switch(c->dstBpc){ \ } if (EXTERNAL_AVX2_FAST(cpu_flags)) { + if (ARCH_X86_64) + switch (c->srcFormat) { + case_rgb(rgb24, RGB24, avx2); + case_rgb(bgr24, BGR24, avx2); + } switch (c->dstFormat) { case AV_PIX_FMT_NV12: case AV_PIX_FMT_NV24:
rgb24_to_uv_8_c: 39.3 rgb24_to_uv_8_sse2: 14.3 rgb24_to_uv_8_ssse3: 13.3 rgb24_to_uv_8_avx: 12.8 rgb24_to_uv_8_avx2: 14.3 rgb24_to_uv_128_c: 582.8 rgb24_to_uv_128_sse2: 127.3 rgb24_to_uv_128_ssse3: 107.3 rgb24_to_uv_128_avx: 111.3 rgb24_to_uv_128_avx2: 62.3 rgb24_to_uv_1080_c: 4981.3 rgb24_to_uv_1080_sse2: 1048.3 rgb24_to_uv_1080_ssse3: 876.8 rgb24_to_uv_1080_avx: 887.8 rgb24_to_uv_1080_avx2: 492.3 rgb24_to_uv_1280_c: 5906.8 rgb24_to_uv_1280_sse2: 1263.3 rgb24_to_uv_1280_ssse3: 1048.3 rgb24_to_uv_1280_avx: 1045.8 rgb24_to_uv_1280_avx2: 579.8 rgb24_to_uv_1920_c: 8665.3 rgb24_to_uv_1920_sse2: 1888.8 rgb24_to_uv_1920_ssse3: 1571.8 rgb24_to_uv_1920_avx: 1558.8 rgb24_to_uv_1920_avx2: 869.3 rgb24_to_y_8_c: 20.3 rgb24_to_y_8_sse2: 11.8 rgb24_to_y_8_ssse3: 10.3 rgb24_to_y_8_avx: 10.3 rgb24_to_y_8_avx2: 10.8 rgb24_to_y_128_c: 284.8 rgb24_to_y_128_sse2: 83.3 rgb24_to_y_128_ssse3: 66.8 rgb24_to_y_128_avx: 64.8 rgb24_to_y_128_avx2: 39.3 rgb24_to_y_1080_c: 2451.3 rgb24_to_y_1080_sse2: 696.3 rgb24_to_y_1080_ssse3: 516.8 rgb24_to_y_1080_avx: 518.8 rgb24_to_y_1080_avx2: 301.8 rgb24_to_y_1280_c: 2892.8 rgb24_to_y_1280_sse2: 816.8 rgb24_to_y_1280_ssse3: 623.3 rgb24_to_y_1280_avx: 616.3 rgb24_to_y_1280_avx2: 350.8 rgb24_to_y_1920_c: 4338.8 rgb24_to_y_1920_sse2: 1210.8 rgb24_to_y_1920_ssse3: 928.3 rgb24_to_y_1920_avx: 920.3 rgb24_to_y_1920_avx2: 534.8 Signed-off-by: James Almer <jamrial@gmail.com> --- libswscale/x86/input.asm | 49 ++++++++++++++++++++++++++++++++++++---- libswscale/x86/swscale.c | 7 ++++++ 2 files changed, 51 insertions(+), 5 deletions(-)