diff options
| author | Mark Reid <mindmark@gmail.com> | 2021-11-24 13:15:21 -0800 |
|---|---|---|
| committer | James Almer <jamrial@gmail.com> | 2022-01-11 16:34:33 -0300 |
| commit | 52f70261642725f2b1af6e1a5a0b2c9d868997fa (patch) | |
| tree | da784581434b55fca80101e38325a45c38446577 /libswscale/x86 | |
| parent | 9e445a5be2dca30a1f1103c73440648ccf5af9b1 (diff) | |
| download | ffmpeg-52f70261642725f2b1af6e1a5a0b2c9d868997fa.tar.gz | |
swscale/x86/input.asm: add x86-optimized planer rgb2yuv functions
sse2 only operates on 2 lanes per loop for to_y and to_uv functions, due
to the lack of pmulld instruction. Emulating pmulld with 2 pmuludq and shuffles
proved too costly and made to_uv functions slower then the c implementation.
For to_y on sse2 only float functions are generated,
I was are not able outperform the c implementation on the integer pixel formats.
For to_a on see4 only the float functions are generated.
sse2 and sse4 generated nearly identical performing code on integer pixel formats,
so only sse2/avx2 versions are generated.
planar_gbrp_to_y_512_c: 1197.5
planar_gbrp_to_y_512_sse4: 444.5
planar_gbrp_to_y_512_avx2: 287.5
planar_gbrap_to_y_512_c: 1204.5
planar_gbrap_to_y_512_sse4: 447.5
planar_gbrap_to_y_512_avx2: 289.5
planar_gbrp9be_to_y_512_c: 1380.0
planar_gbrp9be_to_y_512_sse4: 543.5
planar_gbrp9be_to_y_512_avx2: 340.0
planar_gbrp9le_to_y_512_c: 1200.5
planar_gbrp9le_to_y_512_sse4: 442.0
planar_gbrp9le_to_y_512_avx2: 282.0
planar_gbrp10be_to_y_512_c: 1378.5
planar_gbrp10be_to_y_512_sse4: 544.0
planar_gbrp10be_to_y_512_avx2: 337.5
planar_gbrp10le_to_y_512_c: 1200.0
planar_gbrp10le_to_y_512_sse4: 448.0
planar_gbrp10le_to_y_512_avx2: 285.5
planar_gbrap10be_to_y_512_c: 1380.0
planar_gbrap10be_to_y_512_sse4: 542.0
planar_gbrap10be_to_y_512_avx2: 340.5
planar_gbrap10le_to_y_512_c: 1199.0
planar_gbrap10le_to_y_512_sse4: 446.0
planar_gbrap10le_to_y_512_avx2: 289.5
planar_gbrp12be_to_y_512_c: 10563.0
planar_gbrp12be_to_y_512_sse4: 542.5
planar_gbrp12be_to_y_512_avx2: 339.0
planar_gbrp12le_to_y_512_c: 1201.0
planar_gbrp12le_to_y_512_sse4: 440.5
planar_gbrp12le_to_y_512_avx2: 286.0
planar_gbrap12be_to_y_512_c: 1701.5
planar_gbrap12be_to_y_512_sse4: 917.0
planar_gbrap12be_to_y_512_avx2: 338.5
planar_gbrap12le_to_y_512_c: 1201.0
planar_gbrap12le_to_y_512_sse4: 444.5
planar_gbrap12le_to_y_512_avx2: 288.0
planar_gbrp14be_to_y_512_c: 1370.5
planar_gbrp14be_to_y_512_sse4: 545.0
planar_gbrp14be_to_y_512_avx2: 338.5
planar_gbrp14le_to_y_512_c: 1199.0
planar_gbrp14le_to_y_512_sse4: 444.0
planar_gbrp14le_to_y_512_avx2: 279.5
planar_gbrp16be_to_y_512_c: 1364.0
planar_gbrp16be_to_y_512_sse4: 544.5
planar_gbrp16be_to_y_512_avx2: 339.5
planar_gbrp16le_to_y_512_c: 1201.0
planar_gbrp16le_to_y_512_sse4: 445.5
planar_gbrp16le_to_y_512_avx2: 280.5
planar_gbrap16be_to_y_512_c: 1377.0
planar_gbrap16be_to_y_512_sse4: 545.0
planar_gbrap16be_to_y_512_avx2: 338.5
planar_gbrap16le_to_y_512_c: 1201.0
planar_gbrap16le_to_y_512_sse4: 442.0
planar_gbrap16le_to_y_512_avx2: 279.0
planar_gbrpf32be_to_y_512_c: 4113.0
planar_gbrpf32be_to_y_512_sse2: 2438.0
planar_gbrpf32be_to_y_512_sse4: 1068.0
planar_gbrpf32be_to_y_512_avx2: 904.5
planar_gbrpf32le_to_y_512_c: 3818.5
planar_gbrpf32le_to_y_512_sse2: 2024.5
planar_gbrpf32le_to_y_512_sse4: 1241.5
planar_gbrpf32le_to_y_512_avx2: 657.0
planar_gbrapf32be_to_y_512_c: 3707.0
planar_gbrapf32be_to_y_512_sse2: 2444.0
planar_gbrapf32be_to_y_512_sse4: 1077.0
planar_gbrapf32be_to_y_512_avx2: 909.0
planar_gbrapf32le_to_y_512_c: 3822.0
planar_gbrapf32le_to_y_512_sse2: 2024.5
planar_gbrapf32le_to_y_512_sse4: 1176.0
planar_gbrapf32le_to_y_512_avx2: 658.5
planar_gbrp_to_uv_512_c: 2325.8
planar_gbrp_to_uv_512_sse2: 1726.8
planar_gbrp_to_uv_512_sse4: 771.8
planar_gbrp_to_uv_512_avx2: 506.8
planar_gbrap_to_uv_512_c: 2281.8
planar_gbrap_to_uv_512_sse2: 1726.3
planar_gbrap_to_uv_512_sse4: 768.3
planar_gbrap_to_uv_512_avx2: 496.3
planar_gbrp9be_to_uv_512_c: 2336.8
planar_gbrp9be_to_uv_512_sse2: 1924.8
planar_gbrp9be_to_uv_512_sse4: 852.3
planar_gbrp9be_to_uv_512_avx2: 552.8
planar_gbrp9le_to_uv_512_c: 2270.3
planar_gbrp9le_to_uv_512_sse2: 1512.3
planar_gbrp9le_to_uv_512_sse4: 764.3
planar_gbrp9le_to_uv_512_avx2: 491.3
planar_gbrp10be_to_uv_512_c: 2281.8
planar_gbrp10be_to_uv_512_sse2: 1917.8
planar_gbrp10be_to_uv_512_sse4: 855.3
planar_gbrp10be_to_uv_512_avx2: 541.3
planar_gbrp10le_to_uv_512_c: 2269.8
planar_gbrp10le_to_uv_512_sse2: 1515.3
planar_gbrp10le_to_uv_512_sse4: 759.8
planar_gbrp10le_to_uv_512_avx2: 487.8
planar_gbrap10be_to_uv_512_c: 2382.3
planar_gbrap10be_to_uv_512_sse2: 1924.8
planar_gbrap10be_to_uv_512_sse4: 855.3
planar_gbrap10be_to_uv_512_avx2: 540.8
planar_gbrap10le_to_uv_512_c: 2382.3
planar_gbrap10le_to_uv_512_sse2: 1512.3
planar_gbrap10le_to_uv_512_sse4: 759.3
planar_gbrap10le_to_uv_512_avx2: 484.8
planar_gbrp12be_to_uv_512_c: 2283.8
planar_gbrp12be_to_uv_512_sse2: 1936.8
planar_gbrp12be_to_uv_512_sse4: 858.3
planar_gbrp12be_to_uv_512_avx2: 541.3
planar_gbrp12le_to_uv_512_c: 2278.8
planar_gbrp12le_to_uv_512_sse2: 1507.3
planar_gbrp12le_to_uv_512_sse4: 760.3
planar_gbrp12le_to_uv_512_avx2: 485.8
planar_gbrap12be_to_uv_512_c: 2385.3
planar_gbrap12be_to_uv_512_sse2: 1927.8
planar_gbrap12be_to_uv_512_sse4: 855.3
planar_gbrap12be_to_uv_512_avx2: 539.8
planar_gbrap12le_to_uv_512_c: 2377.3
planar_gbrap12le_to_uv_512_sse2: 1516.3
planar_gbrap12le_to_uv_512_sse4: 759.3
planar_gbrap12le_to_uv_512_avx2: 484.8
planar_gbrp14be_to_uv_512_c: 2283.8
planar_gbrp14be_to_uv_512_sse2: 1935.3
planar_gbrp14be_to_uv_512_sse4: 852.3
planar_gbrp14be_to_uv_512_avx2: 540.3
planar_gbrp14le_to_uv_512_c: 2276.8
planar_gbrp14le_to_uv_512_sse2: 1514.8
planar_gbrp14le_to_uv_512_sse4: 762.3
planar_gbrp14le_to_uv_512_avx2: 484.8
planar_gbrp16be_to_uv_512_c: 2383.3
planar_gbrp16be_to_uv_512_sse2: 1881.8
planar_gbrp16be_to_uv_512_sse4: 852.3
planar_gbrp16be_to_uv_512_avx2: 541.8
planar_gbrp16le_to_uv_512_c: 2378.3
planar_gbrp16le_to_uv_512_sse2: 1476.8
planar_gbrp16le_to_uv_512_sse4: 765.3
planar_gbrp16le_to_uv_512_avx2: 485.8
planar_gbrap16be_to_uv_512_c: 2382.3
planar_gbrap16be_to_uv_512_sse2: 1886.3
planar_gbrap16be_to_uv_512_sse4: 853.8
planar_gbrap16be_to_uv_512_avx2: 550.8
planar_gbrap16le_to_uv_512_c: 2381.8
planar_gbrap16le_to_uv_512_sse2: 1488.3
planar_gbrap16le_to_uv_512_sse4: 765.3
planar_gbrap16le_to_uv_512_avx2: 491.8
planar_gbrpf32be_to_uv_512_c: 4863.0
planar_gbrpf32be_to_uv_512_sse2: 3347.5
planar_gbrpf32be_to_uv_512_sse4: 1800.0
planar_gbrpf32be_to_uv_512_avx2: 1199.0
planar_gbrpf32le_to_uv_512_c: 4725.0
planar_gbrpf32le_to_uv_512_sse2: 2753.0
planar_gbrpf32le_to_uv_512_sse4: 1474.5
planar_gbrpf32le_to_uv_512_avx2: 927.5
planar_gbrapf32be_to_uv_512_c: 4859.0
planar_gbrapf32be_to_uv_512_sse2: 3269.0
planar_gbrapf32be_to_uv_512_sse4: 1802.0
planar_gbrapf32be_to_uv_512_avx2: 1201.5
planar_gbrapf32le_to_uv_512_c: 6338.0
planar_gbrapf32le_to_uv_512_sse2: 2756.5
planar_gbrapf32le_to_uv_512_sse4: 1476.0
planar_gbrapf32le_to_uv_512_avx2: 908.5
planar_gbrap_to_a_512_c: 383.3
planar_gbrap_to_a_512_sse2: 66.8
planar_gbrap_to_a_512_avx2: 43.8
planar_gbrap10be_to_a_512_c: 601.8
planar_gbrap10be_to_a_512_sse2: 86.3
planar_gbrap10be_to_a_512_avx2: 34.8
planar_gbrap10le_to_a_512_c: 602.3
planar_gbrap10le_to_a_512_sse2: 48.8
planar_gbrap10le_to_a_512_avx2: 31.3
planar_gbrap12be_to_a_512_c: 601.8
planar_gbrap12be_to_a_512_sse2: 111.8
planar_gbrap12be_to_a_512_avx2: 41.3
planar_gbrap12le_to_a_512_c: 385.8
planar_gbrap12le_to_a_512_sse2: 75.3
planar_gbrap12le_to_a_512_avx2: 39.8
planar_gbrap16be_to_a_512_c: 386.8
planar_gbrap16be_to_a_512_sse2: 79.8
planar_gbrap16be_to_a_512_avx2: 31.3
planar_gbrap16le_to_a_512_c: 600.3
planar_gbrap16le_to_a_512_sse2: 40.3
planar_gbrap16le_to_a_512_avx2: 30.3
planar_gbrapf32be_to_a_512_c: 1148.8
planar_gbrapf32be_to_a_512_sse2: 611.3
planar_gbrapf32be_to_a_512_sse4: 234.8
planar_gbrapf32be_to_a_512_avx2: 183.3
planar_gbrapf32le_to_a_512_c: 851.3
planar_gbrapf32le_to_a_512_sse2: 263.3
planar_gbrapf32le_to_a_512_sse4: 199.3
planar_gbrapf32le_to_a_512_avx2: 156.8
Reviewed-by: Paul B Mahol <onemda@gmail.com>
Signed-off-by: James Almer <jamrial@gmail.com>
Diffstat (limited to 'libswscale/x86')
| -rw-r--r-- | libswscale/x86/input.asm | 484 | ||||
| -rw-r--r-- | libswscale/x86/swscale.c | 170 |
2 files changed, 653 insertions, 1 deletions
diff --git a/libswscale/x86/input.asm b/libswscale/x86/input.asm index af9afcaa53..fcdfe2fcd8 100644 --- a/libswscale/x86/input.asm +++ b/libswscale/x86/input.asm @@ -93,7 +93,31 @@ shuf_rgb_12x4: db 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 - +pd_65535f: times 8 dd 65535.0 +pb_pack_shuffle16le: db 0, 1, 4, 5, \ + 8, 9, 12, 13, \ + -1, -1, -1, -1, \ + -1, -1, -1, -1, \ + -1, -1, -1, -1, \ + -1, -1, -1, -1, \ + 0, 1, 4, 5, \ + 8, 9, 12, 13 +pb_shuffle32be: db 3, 2, 1, 0, \ + 7, 6, 5, 4, \ + 11, 10, 9, 8, \ + 15, 14, 13, 12, \ + 3, 2, 1, 0, \ + 7, 6, 5, 4, \ + 11, 10, 9, 8, \ + 15, 14, 13, 12 +pb_shuffle16be: db 1, 0, 3, 2, \ + 5, 4, 7, 6, \ + 9, 8, 11, 10, \ + 13, 12, 15, 14, \ + 1, 0, 3, 2, \ + 5, 4, 7, 6, \ + 9, 8, 11, 10, \ + 13, 12, 15, 14 SECTION .text ;----------------------------------------------------------------------------- @@ -738,3 +762,461 @@ YUYV_TO_UV_FN 3, uyvy, 1 NVXX_TO_UV_FN 5, nv12 NVXX_TO_UV_FN 5, nv21 %endif + +%if ARCH_X86_64 +%define RY_IDX 0 +%define GY_IDX 1 +%define BY_IDX 2 +%define RU_IDX 3 +%define GU_IDX 4 +%define BU_IDX 5 +%define RV_IDX 6 +%define GV_IDX 7 +%define BV_IDX 8 +%define RGB2YUV_SHIFT 15 + +%define R m0 +%define G m1 +%define B m2 + +%macro SWAP32 1 +%if mmsize > 16 || cpuflag(sse4) + pshufb m%1, [pb_shuffle32be] +%else + psrlw xm7, xm%1, 8 + psllw xm%1, 8 + por xm%1, xm7 + pshuflw xm%1, xm%1, (2 << 6 | 3 << 4 | 0 << 2 | 1 << 0) + pshufhw xm%1, xm%1, (2 << 6 | 3 << 4 | 0 << 2 | 1 << 0) +%endif +%endmacro + +; 1 - dest +; 2 - source +; 3 - is big endian +; 4 - load only 2 values on sse2 +%macro LOADF32 4 + %if notcpuflag(sse4) && %4 + %if %3 ; big endian + mov tmp1q, %2 + bswap tmp1q + movq xm%1, tmp1q + %else + movq m%1, %2 + %endif + %else + movu m%1, %2 + %if %3 + SWAP32 %1 + %endif + %endif + maxps m%1, m9 ; 0.0 (nan, -inf) -> 0.0 + mulps m%1, m8 ; [pd_65535f] + minps m%1, m8 ; +inf -> 65535 + ; cvtps2dq rounds to nearest int + ; assuming mxcsr register is default rounding + ; 0.40 -> 0.0, 0.50 -> 0.0, 0.51 -> 1.0 + cvtps2dq m%1, m%1 + + %if notcpuflag(sse4) && %4 + ; line up the 2 values in lanes 0,2 + %if %3 ; big endian + pshufd m%1, m%1, (3 << 6 | 0 << 4 | 2 << 2 | 1 << 0) + %else + pshufd m%1, m%1, (3 << 6 | 1 << 4 | 2 << 2 | 0 << 0) + %endif + %endif +%endmacro + +; 1 - dest +; 2 - source +; 3 - is big endian +%macro LOAD16 3 + %if cpuflag(sse4) || mmsize > 16 + pmovzxwd m%1, %2 + %if %3 ; bigendian + pshufb m%1, m8 ; [pb_shuffle16be] + %endif + %else + %if %3 ; bigendian + mov tmp1d, dword %2 + bswap tmp1d + movd xm%1, tmp1d + pshuflw m%1, m%1, (3 << 6 | 0 << 4 | 3 << 2 | 1 << 0) + pshufd m%1, m%1, (3 << 6 | 1 << 4 | 2 << 2 | 0 << 0) + %else + movd xm%1, %2 + punpcklwd m%1, m9 ; interleave words with zero + pshufd m%1, m%1, (3 << 6 | 1 << 4 | 2 << 2 | 0 << 0) + %endif + %endif +%endmacro + +%macro LOAD8_RGB 0 + %if cpuflag(sse4) || mmsize > 16 + pmovzxbd R, [srcRq + xq] + pmovzxbd G, [srcGq + xq] + pmovzxbd B, [srcBq + xq] + %else + ; thought this would be faster but from my measurments its not + ; movd m0, [srcRq + xq + 0]; overeads by 2 bytes + ; punpcklbw m0, m9 ; interleave bytes with zero + ; punpcklwd m0, m9 ; interleave words with zero + ; pshufd m0, m0, (3 << 6 | 1 << 4 | 2 << 2 | 0 << 0) + + movzx tmp2q, byte [srcRq + xq + 1] + movzx tmp1q, byte [srcRq + xq + 0] + shl tmp2q, 32 + or tmp1q, tmp2q + movq xm0, tmp1q + + movzx tmp2q, byte [srcGq + xq + 1] + movzx tmp3q, byte [srcGq + xq + 0] + shl tmp2q, 32 + or tmp3q, tmp2q + movq xm1, tmp3q + + movzx tmp2q, byte [srcBq + xq + 1] + movzx tmp1q, byte [srcBq + xq + 0] + shl tmp2q, 32 + or tmp1q, tmp2q + movq xm2, tmp1q + + pshufd m0, m0, (3 << 6 | 1 << 4 | 2 << 2 | 0 << 0) + pshufd m1, m1, (3 << 6 | 1 << 4 | 2 << 2 | 0 << 0) + pshufd m2, m2, (3 << 6 | 1 << 4 | 2 << 2 | 0 << 0) + %endif +%endmacro + +; 1 - dest +; 2 - source +; 3 - store only 2 values on sse2 +%macro STORE16 3 + %if %3 && notcpuflag(sse4) + pshufd m%2, m%2, (3 << 6 | 1 << 4 | 2 << 2 | 0 << 0) + pshuflw m%2, m%2, (3 << 6 | 1 << 4 | 2 << 2 | 0 << 0) + movd %1, m%2 + %elif mmsize > 16 + pshufb m%2, m7 ; [pb_pack_shuffle16le] + vpermq m%2, m%2, (3 << 6 | 0 << 4 | 3 << 2 | 0 << 0) + movu %1, xm%2 + %else + %if cpuflag(sse4) + pshufb m%2, m7 ; [pb_pack_shuffle16le] + %else + pshuflw m%2, m%2, (1 << 6 | 1 << 4 | 2 << 2 | 0 << 0) + pshufhw m%2, m%2, (1 << 6 | 1 << 4 | 2 << 2 | 0 << 0) + pshufd m%2, m%2, (3 << 6 | 3 << 4 | 2 << 2 | 0 << 0) + %endif + movq %1, m%2 + %endif +%endmacro + +%macro PMUL 3 +%if cpuflag(sse4) || mmsize > 16 + pmulld %1, %2, %3 +%else + pmuludq %1, %2, %3 +%endif +%endmacro + +; 1 - name +; 2 - depth +; 3 - is big endian +; 4 - is float +; in sse2 mode only 2 values are done per loop, due to lack of pmulld instruction +%macro planar_rgb_to_y_fn 4 +%if %2 == 8 + %define OFFSET (0x801<<(RGB2YUV_SHIFT-7)) + %define RSHIFT (RGB2YUV_SHIFT-6) +%else + %if %2 < 16 + %define SHIFT %2 + %define BPC %2 + %else + %define SHIFT 14 + %define BPC 16 + %endif + %define OFFSET ((16 << (RGB2YUV_SHIFT + BPC - 8)) + (1 << (RGB2YUV_SHIFT + SHIFT - 15))) + %define RSHIFT (RGB2YUV_SHIFT + SHIFT - 14) +%endif +cglobal planar_%1_to_y, 4, 12, 13, dst, src, w, rgb2yuv, srcR, srcG, srcB, x, tmp1, tmp2, tmp3, tmp4 + VBROADCASTSS m10, dword [rgb2yuvq + RY_IDX*4] ; ry + VBROADCASTSS m11, dword [rgb2yuvq + GY_IDX*4] ; gy + VBROADCASTSS m12, dword [rgb2yuvq + BY_IDX*4] ; by + pxor m9, m9 + + %if %4 + movu m8, [pd_65535f] + %endif + + %if cpuflag(sse4) || mmsize > 16 + movu m7, [pb_pack_shuffle16le] + %if %3 && %2 > 8 && %2 <= 16 + movu m8, [pb_shuffle16be] + %endif + %endif + + mov xq, OFFSET + movq xm6, xq + VBROADCASTSS m6, xm6 + + mov srcGq, [srcq + 0] + mov srcBq, [srcq + 8] + mov srcRq, [srcq + 16] + + xor xq, xq + %%loop_x: + %if %4 + LOADF32 0, [srcRq + xq*4], %3, 1 + LOADF32 1, [srcGq + xq*4], %3, 1 + LOADF32 2, [srcBq + xq*4], %3, 1 + %elif %2 == 8 + LOAD8_RGB + %else + LOAD16 0, [srcRq + xq*2], %3 + LOAD16 1, [srcGq + xq*2], %3 + LOAD16 2, [srcBq + xq*2], %3 + %endif + + PMUL R, R, m10 ; r*ry + PMUL G, G, m11 ; g*gy + PMUL B, B, m12 ; b*by + paddd m0, m6 ; + OFFSET + paddd B, G + paddd m0, B + psrad m0, RSHIFT + STORE16 [dstq + 2*xq], 0, 1 + + %if cpuflag(avx2) || cpuflag(sse4) + add xq, mmsize/4 + %else + add xd, 2 + %endif + cmp xd, wd + jl %%loop_x +RET +%endmacro + +; 1 - name +; 2 - depth +; 3 - is big endian +; 4 - is float +; in sse2 mode only 2 values are done per loop, due to lack of pmulld instruction +%macro planar_rgb_to_uv_fn 4 +%if %2 == 8 + %define OFFSET (0x4001<<(RGB2YUV_SHIFT-7)) + %define RSHIFT (RGB2YUV_SHIFT-6) +%else + %if %2 < 16 + %define SHIFT %2 + %define BPC %2 + %else + %define SHIFT 14 + %define BPC 16 + %endif + %define OFFSET ((128 << (RGB2YUV_SHIFT + BPC - 8)) + (1 << (RGB2YUV_SHIFT + SHIFT - 15))) + %define RSHIFT (RGB2YUV_SHIFT + SHIFT - 14) +%endif +cglobal planar_%1_to_uv, 5, 12, 16, dstU, dstV, src, w, rgb2yuv, srcR, srcG, srcB, x, tmp1, tmp2, tmp3 + VBROADCASTSS m10, dword [rgb2yuvq + RU_IDX*4] ; ru + VBROADCASTSS m11, dword [rgb2yuvq + GU_IDX*4] ; gu + VBROADCASTSS m12, dword [rgb2yuvq + BU_IDX*4] ; bu + VBROADCASTSS m13, dword [rgb2yuvq + RV_IDX*4] ; rv + VBROADCASTSS m14, dword [rgb2yuvq + GV_IDX*4] ; gv + VBROADCASTSS m15, dword [rgb2yuvq + BV_IDX*4] ; bv + pxor m9, m9 + + %if %4 + movu m8, [pd_65535f] + %endif + + %if cpuflag(sse4) || mmsize > 16 + movu m7, [pb_pack_shuffle16le] + %if %3 && %2 > 8 && %2 <= 16 + movu m8, [pb_shuffle16be] + %endif + %endif + + mov xq, OFFSET + movq xm6, xq + VBROADCASTSS m6, xm6 + + mov srcGq, [srcq + 0] + mov srcBq, [srcq + 8] + mov srcRq, [srcq + 16] + + xor xq, xq + %%loop_x: + %if %4 + LOADF32 0, [srcRq + xq*4], %3, 1 + LOADF32 1, [srcGq + xq*4], %3, 1 + LOADF32 2, [srcBq + xq*4], %3, 1 + %elif %2 == 8 + LOAD8_RGB + %else + LOAD16 0, [srcRq + xq*2], %3 + LOAD16 1, [srcGq + xq*2], %3 + LOAD16 2, [srcBq + xq*2], %3 + %endif + + PMUL m5, R, m10 ; r*ru + PMUL m4, G, m11 ; b*gu + paddd m4, m5 + PMUL m5, B, m12 ; b*bu + paddd m4, m6 ; + OFFSET + paddd m4, m5 + psrad m4, RSHIFT + STORE16 [dstUq + 2*xq], 4, 1 + + PMUL R, R, m13 ; r*rv + PMUL G, G, m14 ; g*gv*g + PMUL B, B, m15 ; b*bv + paddd m0, m6 ; + OFFSET + paddd B, G + paddd m0, B + psrad m0, RSHIFT + STORE16 [dstVq + 2*xq], 0, 1 + + %if cpuflag(avx2) || cpuflag(sse4) + add xd, mmsize/4 + %else + add xd, 2 + %endif + cmp xd, wd + jl %%loop_x +RET +%endmacro + +; 1 - name +; 2 - depth +; 3 - is big endian +; 4 - is float +%macro planar_rgb_to_a_fn 4 +cglobal planar_%1_to_a, 4, 6, 10, dst, src, w, rgb2yuv, srcA, x + %if %4 && (cpuflag(sse4) || mmsize > 16) + movu m7, [pb_pack_shuffle16le] + %elif %3 && (cpuflag(sse4) || mmsize > 16) + movu m7, [pb_shuffle16be] + %endif + + %if %4 + movu m8, [pd_65535f] + %endif + + pxor m9, m9 + mov srcAq, [srcq + 24] + xor xq, xq + %%loop_x: + %if %4 ; float + LOADF32 0, [srcAq + xq*4], %3, 0 + STORE16 [dstq + xq*2], 0, 0 + add xq, mmsize/4 + %elif %2 == 8 + ; only need to convert 8bit value to 16bit + %if cpuflag(sse4) || mmsize > 16 + pmovzxbw m0, [srcAq + xq] + %else + movsd m0, [srcAq + xq] + punpcklbw m0, m9 ; interleave bytes with zero + %endif + psllw m0, 6 + movu [dstq + xq*2], m0 + add xq, mmsize/2 + %else + ; only need to convert 16bit format to 16le + movu m0, [srcAq + xq*2] + %if %3 ; bigendian + %if cpuflag(sse4) || mmsize > 16 + pshufb m0, m7 ; [pb_shuffle16be] + %else + psrlw m7, m0, 8 + psllw m0, 8 + por m0, m7 + %endif + %endif + %if %2 < 16 + psllw m0, (14 - %2) + %endif + movu [dstq + xq*2], m0 + add xq, mmsize/2 + %endif + cmp xd, wd + jl %%loop_x +RET +%endmacro + +; 1 - name +; 2 - depth +; 3 - is float +%macro planer_rgbxx_y_fn_decl 3 +planar_rgb_to_y_fn %1le, %2, 0, %3 +planar_rgb_to_y_fn %1be, %2, 1, %3 +%endmacro + +; 1 - name +; 2 - depth +; 3 - is float +%macro planer_rgbxx_uv_fn_decl 3 +planar_rgb_to_uv_fn %1le, %2, 0, %3 +planar_rgb_to_uv_fn %1be, %2, 1, %3 +%endmacro + +; 1 - name +; 2 - depth +; 3 - is float +%macro planer_rgbxx_a_fn_decl 3 +planar_rgb_to_a_fn %1le, %2, 0, %3 +planar_rgb_to_a_fn %1be, %2, 1, %3 +%endmacro + +%macro planar_rgb_y_all_fn_decl 0 +planar_rgb_to_y_fn rgb, 8, 0, 0 +planer_rgbxx_y_fn_decl rgb9, 9, 0 +planer_rgbxx_y_fn_decl rgb10, 10, 0 +planer_rgbxx_y_fn_decl rgb12, 12, 0 +planer_rgbxx_y_fn_decl rgb14, 14, 0 +planer_rgbxx_y_fn_decl rgb16, 16, 0 +planer_rgbxx_y_fn_decl rgbf32, 32, 1 +%endmacro + +%macro planar_rgb_uv_all_fn_decl 0 +planar_rgb_to_uv_fn rgb, 8, 0, 0 +planer_rgbxx_uv_fn_decl rgb9, 9, 0 +planer_rgbxx_uv_fn_decl rgb10, 10, 0 +planer_rgbxx_uv_fn_decl rgb12, 12, 0 +planer_rgbxx_uv_fn_decl rgb14, 14, 0 +planer_rgbxx_uv_fn_decl rgb16, 16, 0 +planer_rgbxx_uv_fn_decl rgbf32, 32, 1 +%endmacro + +%macro planar_rgb_a_all_fn_decl 0 +planar_rgb_to_a_fn rgb, 8, 0, 0 +planer_rgbxx_a_fn_decl rgb10, 10, 0 +planer_rgbxx_a_fn_decl rgb12, 12, 0 +planer_rgbxx_a_fn_decl rgb16, 16, 0 +planer_rgbxx_a_fn_decl rgbf32, 32, 1 +%endmacro + +; sse2 to_y only matches c speed with current implementation +; except on floating point formats +INIT_XMM sse2 +planer_rgbxx_y_fn_decl rgbf32, 32, 1 +planar_rgb_uv_all_fn_decl +planar_rgb_a_all_fn_decl + +; sse4 to_a conversions are just the sse2 ones +; except on floating point formats +INIT_XMM sse4 +planar_rgb_y_all_fn_decl +planar_rgb_uv_all_fn_decl +planer_rgbxx_a_fn_decl rgbf32, 32, 1 + +%if HAVE_AVX2_EXTERNAL +INIT_YMM avx2 +planar_rgb_y_all_fn_decl +planar_rgb_uv_all_fn_decl +planar_rgb_a_all_fn_decl +%endif + +%endif ; ARCH_X86_64 diff --git a/libswscale/x86/swscale.c b/libswscale/x86/swscale.c index d5a467da0e..3cf0c419fd 100644 --- a/libswscale/x86/swscale.c +++ b/libswscale/x86/swscale.c @@ -390,6 +390,80 @@ YUV2GBRP_DECL(sse2); YUV2GBRP_DECL(sse4); YUV2GBRP_DECL(avx2); +#define INPUT_PLANAR_RGB_Y_FN_DECL(fmt, opt) \ +void ff_planar_##fmt##_to_y_##opt(uint8_t *dst, \ + const uint8_t *src[4], int w, int32_t *rgb2yuv) + +#define INPUT_PLANAR_RGB_UV_FN_DECL(fmt, opt) \ +void ff_planar_##fmt##_to_uv_##opt(uint8_t *dstU, uint8_t *dstV, \ + const uint8_t *src[4], int w, int32_t *rgb2yuv) + +#define INPUT_PLANAR_RGB_A_FN_DECL(fmt, opt) \ +void ff_planar_##fmt##_to_a_##opt(uint8_t *dst, \ + const uint8_t *src[4], int w, int32_t *rgb2yuv) + + +#define INPUT_PLANAR_RGBXX_A_DECL(fmt, opt) \ +INPUT_PLANAR_RGB_A_FN_DECL(fmt##le, opt); \ +INPUT_PLANAR_RGB_A_FN_DECL(fmt##be, opt); + +#define INPUT_PLANAR_RGBXX_Y_DECL(fmt, opt) \ +INPUT_PLANAR_RGB_Y_FN_DECL(fmt##le, opt); \ +INPUT_PLANAR_RGB_Y_FN_DECL(fmt##be, opt); + +#define INPUT_PLANAR_RGBXX_UV_DECL(fmt, opt) \ +INPUT_PLANAR_RGB_UV_FN_DECL(fmt##le, opt); \ +INPUT_PLANAR_RGB_UV_FN_DECL(fmt##be, opt); + +#define INPUT_PLANAR_RGBXX_YUVA_DECL(fmt, opt) \ +INPUT_PLANAR_RGBXX_Y_DECL(fmt, opt); \ +INPUT_PLANAR_RGBXX_UV_DECL(fmt, opt); \ +INPUT_PLANAR_RGBXX_A_DECL(fmt, opt); + +#define INPUT_PLANAR_RGBXX_YUV_DECL(fmt, opt) \ +INPUT_PLANAR_RGBXX_Y_DECL(fmt, opt); \ +INPUT_PLANAR_RGBXX_UV_DECL(fmt, opt); + +#define INPUT_PLANAR_RGBXX_UVA_DECL(fmt, opt) \ +INPUT_PLANAR_RGBXX_UV_DECL(fmt, opt); \ +INPUT_PLANAR_RGBXX_A_DECL(fmt, opt); + +#define INPUT_PLANAR_RGB_A_ALL_DECL(opt) \ +INPUT_PLANAR_RGB_A_FN_DECL(rgb, opt); \ +INPUT_PLANAR_RGBXX_A_DECL(rgb10, opt); \ +INPUT_PLANAR_RGBXX_A_DECL(rgb12, opt); \ +INPUT_PLANAR_RGBXX_A_DECL(rgb16, opt); \ +INPUT_PLANAR_RGBXX_A_DECL(rgbf32, opt); + +#define INPUT_PLANAR_RGB_Y_ALL_DECL(opt) \ +INPUT_PLANAR_RGB_Y_FN_DECL(rgb, opt); \ +INPUT_PLANAR_RGBXX_Y_DECL(rgb9, opt); \ +INPUT_PLANAR_RGBXX_Y_DECL(rgb10, opt); \ +INPUT_PLANAR_RGBXX_Y_DECL(rgb12, opt); \ +INPUT_PLANAR_RGBXX_Y_DECL(rgb14, opt); \ +INPUT_PLANAR_RGBXX_Y_DECL(rgb16, opt); \ +INPUT_PLANAR_RGBXX_Y_DECL(rgbf32, opt); + +#define INPUT_PLANAR_RGB_UV_ALL_DECL(opt) \ +INPUT_PLANAR_RGB_UV_FN_DECL(rgb, opt); \ +INPUT_PLANAR_RGBXX_UV_DECL(rgb9, opt); \ +INPUT_PLANAR_RGBXX_UV_DECL(rgb10, opt); \ +INPUT_PLANAR_RGBXX_UV_DECL(rgb12, opt); \ +INPUT_PLANAR_RGBXX_UV_DECL(rgb14, opt); \ +INPUT_PLANAR_RGBXX_UV_DECL(rgb16, opt); \ +INPUT_PLANAR_RGBXX_UV_DECL(rgbf32, opt); + +INPUT_PLANAR_RGBXX_Y_DECL(rgbf32, sse2); +INPUT_PLANAR_RGB_UV_ALL_DECL(sse2); +INPUT_PLANAR_RGB_A_ALL_DECL(sse2); + +INPUT_PLANAR_RGB_Y_ALL_DECL(sse4); +INPUT_PLANAR_RGB_UV_ALL_DECL(sse4); +INPUT_PLANAR_RGBXX_A_DECL(rgbf32, sse4); + +INPUT_PLANAR_RGB_Y_ALL_DECL(avx2); +INPUT_PLANAR_RGB_UV_ALL_DECL(avx2); +INPUT_PLANAR_RGB_A_ALL_DECL(avx2); #endif av_cold void ff_sws_init_swscale_x86(SwsContext *c) @@ -639,6 +713,102 @@ switch(c->dstBpc){ \ } } + +#define INPUT_PLANER_RGB_A_FUNC_CASE(fmt, name, opt) \ + case fmt: \ + c->readAlpPlanar = ff_planar_##name##_to_a_##opt; + +#define INPUT_PLANER_RGBA_YUV_FUNC_CASE(rgb_fmt, rgba_fmt, name, opt) \ + case rgba_fmt: \ + case rgb_fmt: \ + c->readLumPlanar = ff_planar_##name##_to_y_##opt; \ + c->readChrPlanar = ff_planar_##name##_to_uv_##opt; \ + break; + +#define INPUT_PLANER_RGB_YUV_FUNC_CASE(fmt, name, opt) \ + case fmt: \ + c->readLumPlanar = ff_planar_##name##_to_y_##opt; \ + c->readChrPlanar = ff_planar_##name##_to_uv_##opt; \ + break; + +#define INPUT_PLANER_RGB_UV_FUNC_CASE(fmt, name, opt) \ + case fmt: \ + c->readChrPlanar = ff_planar_##name##_to_uv_##opt; \ + break; + +#define INPUT_PLANER_RGBAXX_YUVA_FUNC_CASE(rgb_fmt, rgba_fmt, name, opt) \ + INPUT_PLANER_RGB_A_FUNC_CASE(rgba_fmt##LE, name##le, opt) \ + INPUT_PLANER_RGB_YUV_FUNC_CASE(rgb_fmt##LE, name##le, opt) \ + INPUT_PLANER_RGB_A_FUNC_CASE(rgba_fmt##BE, name##be, opt) \ + INPUT_PLANER_RGB_YUV_FUNC_CASE(rgb_fmt##BE, name##be, opt) + +#define INPUT_PLANER_RGBAXX_UVA_FUNC_CASE(rgb_fmt, rgba_fmt, name, opt) \ + INPUT_PLANER_RGB_A_FUNC_CASE(rgba_fmt##LE, name##le, opt) \ + INPUT_PLANER_RGB_UV_FUNC_CASE(rgb_fmt##LE, name##le, opt) \ + INPUT_PLANER_RGB_A_FUNC_CASE(rgba_fmt##BE, name##be, opt) \ + INPUT_PLANER_RGB_UV_FUNC_CASE(rgb_fmt##BE, name##be, opt) + +#define INPUT_PLANER_RGBAXX_YUV_FUNC_CASE(rgb_fmt, rgba_fmt, name, opt) \ + INPUT_PLANER_RGBA_YUV_FUNC_CASE(rgb_fmt##LE, rgba_fmt##LE, name##le, opt) \ + INPUT_PLANER_RGBA_YUV_FUNC_CASE(rgb_fmt##BE, rgba_fmt##BE, name##be, opt) + +#define INPUT_PLANER_RGBXX_YUV_FUNC_CASE(rgb_fmt, name, opt) \ + INPUT_PLANER_RGB_YUV_FUNC_CASE(rgb_fmt##LE, name##le, opt) \ + INPUT_PLANER_RGB_YUV_FUNC_CASE(rgb_fmt##BE, name##be, opt) + +#define INPUT_PLANER_RGBXX_UV_FUNC_CASE(rgb_fmt, name, opt) \ + INPUT_PLANER_RGB_UV_FUNC_CASE(rgb_fmt##LE, name##le, opt) \ + INPUT_PLANER_RGB_UV_FUNC_CASE(rgb_fmt##BE, name##be, opt) + +#define INPUT_PLANER_RGB_YUVA_ALL_CASES(opt) \ + INPUT_PLANER_RGB_A_FUNC_CASE( AV_PIX_FMT_GBRAP, rgb, opt) \ + INPUT_PLANER_RGB_YUV_FUNC_CASE( AV_PIX_FMT_GBRP, rgb, opt) \ + INPUT_PLANER_RGBXX_YUV_FUNC_CASE( AV_PIX_FMT_GBRP9, rgb9, opt) \ + INPUT_PLANER_RGBAXX_YUVA_FUNC_CASE(AV_PIX_FMT_GBRP10, AV_PIX_FMT_GBRAP10, rgb10, opt) \ + INPUT_PLANER_RGBAXX_YUVA_FUNC_CASE(AV_PIX_FMT_GBRP12, AV_PIX_FMT_GBRAP12, rgb12, opt) \ + INPUT_PLANER_RGBXX_YUV_FUNC_CASE( AV_PIX_FMT_GBRP14, rgb14, opt) \ + INPUT_PLANER_RGBAXX_YUVA_FUNC_CASE(AV_PIX_FMT_GBRP16, AV_PIX_FMT_GBRAP16, rgb16, opt) \ + INPUT_PLANER_RGBAXX_YUVA_FUNC_CASE(AV_PIX_FMT_GBRPF32, AV_PIX_FMT_GBRAPF32, rgbf32, opt) + + + if (EXTERNAL_SSE2(cpu_flags)) { + switch (c->srcFormat) { + INPUT_PLANER_RGB_A_FUNC_CASE( AV_PIX_FMT_GBRAP, rgb, sse2); + INPUT_PLANER_RGB_UV_FUNC_CASE( AV_PIX_FMT_GBRP, rgb, sse2); + INPUT_PLANER_RGBXX_UV_FUNC_CASE( AV_PIX_FMT_GBRP9, rgb9, sse2); + INPUT_PLANER_RGBAXX_UVA_FUNC_CASE( AV_PIX_FMT_GBRP10, AV_PIX_FMT_GBRAP10, rgb10, sse2); + INPUT_PLANER_RGBAXX_UVA_FUNC_CASE( AV_PIX_FMT_GBRP12, AV_PIX_FMT_GBRAP12, rgb12, sse2); + INPUT_PLANER_RGBXX_UV_FUNC_CASE( AV_PIX_FMT_GBRP14, rgb14, sse2); + INPUT_PLANER_RGBAXX_UVA_FUNC_CASE( AV_PIX_FMT_GBRP16, AV_PIX_FMT_GBRAP16, rgb16, sse2); + INPUT_PLANER_RGBAXX_YUVA_FUNC_CASE(AV_PIX_FMT_GBRPF32, AV_PIX_FMT_GBRAPF32, rgbf32, sse2); + default: + break; + } + } + + if (EXTERNAL_SSE4(cpu_flags)) { + switch (c->srcFormat) { + case AV_PIX_FMT_GBRAP: + INPUT_PLANER_RGB_YUV_FUNC_CASE( AV_PIX_FMT_GBRP, rgb, sse4); + INPUT_PLANER_RGBXX_YUV_FUNC_CASE( AV_PIX_FMT_GBRP9, rgb9, sse4); + INPUT_PLANER_RGBAXX_YUV_FUNC_CASE( AV_PIX_FMT_GBRP10, AV_PIX_FMT_GBRAP10, rgb10, sse4); + INPUT_PLANER_RGBAXX_YUV_FUNC_CASE( AV_PIX_FMT_GBRP12, AV_PIX_FMT_GBRAP12, rgb12, sse4); + INPUT_PLANER_RGBXX_YUV_FUNC_CASE( AV_PIX_FMT_GBRP14, rgb14, sse4); + INPUT_PLANER_RGBAXX_YUV_FUNC_CASE( AV_PIX_FMT_GBRP16, AV_PIX_FMT_GBRAP16, rgb16, sse4); + INPUT_PLANER_RGBAXX_YUVA_FUNC_CASE(AV_PIX_FMT_GBRPF32, AV_PIX_FMT_GBRAPF32, rgbf32, sse4); + default: + break; + } + } + + if (EXTERNAL_AVX2_FAST(cpu_flags)) { + switch (c->srcFormat) { + INPUT_PLANER_RGB_YUVA_ALL_CASES(avx2) + default: + break; + } + } + if(c->flags & SWS_FULL_CHR_H_INT) { /* yuv2gbrp uses the SwsContext for yuv coefficients |