[x265] [PATCH] asm: filterPixelToShort 8-bit and 10-bit sse2
dtyx265 at gmail.com
dtyx265 at gmail.com
Fri Jun 5 03:14:14 CEST 2015
# HG changeset patch
# User David T Yuen <dtyx265 at gmail.com>
# Date 1433466246 25200
# Node ID a828fd4b5f48f7c493fe48e5e12a1d34498c1433
# Parent 0f0d88319f7cc96661eef3c3dcc1befcf60354f3
asm: filterPixelToShort 8-bit and 10-bit sse2
This replaces c code for all of filterPixelToShort for 8 and 10 bit.
64-bit
./test/TestBench --testbench interp | grep p2s
convert_p2s[ 8x8] 2.82x 397.50 1119.92
convert_p2s[16x16] 3.04x 1347.50 4102.63
convert_p2s[32x32] 1.41x 5197.50 7332.50
convert_p2s[64x64] 1.26x 20588.66 25962.67
convert_p2s[ 8x4] 2.52x 229.99 580.00
convert_p2s[ 4x8] 2.22x 279.99 622.38
convert_p2s[ 16x8] 2.74x 710.00 1944.96
convert_p2s[ 8x16] 3.09x 730.00 2254.97
convert_p2s[32x16] 1.43x 2630.20 3774.20
convert_p2s[16x32] 3.07x 2630.01 8064.97
convert_p2s[64x32] 1.28x 10307.52 13162.50
convert_p2s[32x64] 1.40x 10307.60 14382.50
convert_p2s[16x12] 3.05x 1027.50 3134.97
convert_p2s[12x16] 2.78x 1115.00 3100.73
convert_p2s[ 16x4] 2.56x 387.50 990.06
convert_p2s[ 4x16] 2.43x 490.02 1190.04
convert_p2s[32x24] 1.42x 3909.22 5532.57
convert_p2s[24x32] 3.94x 3907.50 15387.65
convert_p2s[ 32x8] 1.37x 1347.49 1848.57
convert_p2s[ 8x32] 3.31x 1390.00 4596.10
convert_p2s[64x48] 1.27x 15455.30 19562.58
convert_p2s[48x64] 1.30x 15428.06 20132.50
convert_p2s[64x16] 1.28x 5192.50 6669.05
convert_p2s[16x64] 3.05x 5197.80 15855.29
chroma_p2s[ 4x4] 1.89x 177.50 334.95
chroma_p2s[ 8x8] 2.82x 397.50 1119.83
chroma_p2s[16x16] 3.05x 1347.51 4105.62
chroma_p2s[32x32] 1.41x 5187.50 7334.46
chroma_p2s[ 4x2] 1.67x 115.04 192.48
chroma_p2s[ 2x4] 1.16x 184.99 214.98
chroma_p2s[ 8x4] 2.55x 227.50 580.01
chroma_p2s[ 4x8] 2.24x 277.50 622.42
chroma_p2s[ 16x8] 2.75x 707.50 1945.02
chroma_p2s[ 8x16] 2.92x 772.45 2254.91
chroma_p2s[32x16] 1.44x 2627.50 3772.50
chroma_p2s[16x32] 3.07x 2627.50 8065.07
chroma_p2s[ 8x6] 2.74x 307.59 842.44
chroma_p2s[ 6x8] 1.71x 507.50 870.00
chroma_p2s[ 8x2] 2.02x 147.50 297.50
chroma_p2s[ 2x8] 1.12x 307.50 344.96
chroma_p2s[16x12] 3.05x 1027.50 3134.97
chroma_p2s[12x16] 2.79x 1112.50 3100.21
chroma_p2s[ 16x4] 2.56x 387.50 990.19
chroma_p2s[ 4x16] 2.43x 489.99 1192.44
chroma_p2s[32x24] 1.42x 3907.50 5533.21
chroma_p2s[24x32] 3.93x 3923.72 15427.02
chroma_p2s[ 32x8] 1.37x 1347.50 1850.34
chroma_p2s[ 8x32] 3.31x 1387.50 4595.20
chroma_p2s[ 4x8] 2.22x 277.50 617.42
chroma_p2s[ 8x16] 3.09x 727.50 2247.50
chroma_p2s[16x32] 3.07x 2627.50 8065.15
chroma_p2s[32x64] 1.40x 10307.58 14385.50
chroma_p2s[ 4x4] 1.87x 177.50 332.51
chroma_p2s[ 2x8] 1.11x 307.50 342.46
chroma_p2s[ 8x8] 2.80x 397.50 1112.50
chroma_p2s[ 4x16] 2.43x 489.99 1192.49
chroma_p2s[16x16] 3.05x 1347.74 4104.98
chroma_p2s[ 8x32] 3.31x 1387.50 4595.44
chroma_p2s[32x32] 1.41x 5197.55 7332.50
chroma_p2s[16x64] 3.05x 5197.60 15855.21
chroma_p2s[ 8x12] 2.92x 557.50 1627.41
chroma_p2s[ 6x16] 1.79x 1002.50 1797.47
chroma_p2s[ 8x4] 2.55x 227.50 580.01
chroma_p2s[ 2x16] 1.09x 602.49 657.49
chroma_p2s[16x24] 3.05x 1987.50 6054.97
chroma_p2s[12x32] 2.81x 2170.00 6095.56
chroma_p2s[ 16x8] 2.75x 707.50 1944.97
chroma_p2s[ 4x32] 3.09x 877.50 2707.95
chroma_p2s[32x48] 1.40x 7757.54 10862.72
chroma_p2s[24x64] 3.95x 7757.50 30663.70
chroma_p2s[32x16] 1.44x 2627.50 3773.21
chroma_p2s[ 8x64] 3.30x 2717.50 8955.97
chroma_p2s[ 4x4] 1.89x 177.50 334.94
chroma_p2s[ 8x8] 2.82x 397.50 1119.95
chroma_p2s[16x16] 3.05x 1347.50 4105.23
chroma_p2s[32x32] 1.41x 5197.52 7332.50
chroma_p2s[64x64] 1.25x 20722.45 25962.96
chroma_p2s[ 8x4] 2.57x 227.50 584.01
chroma_p2s[ 4x8] 2.23x 277.49 617.44
chroma_p2s[ 16x8] 2.75x 707.57 1945.83
chroma_p2s[ 8x16] 3.08x 729.99 2247.50
chroma_p2s[32x16] 1.44x 2627.50 3772.50
chroma_p2s[16x32] 3.07x 2627.50 8064.97
chroma_p2s[64x32] 1.28x 10307.86 13162.50
chroma_p2s[32x64] 1.40x 10307.68 14385.40
chroma_p2s[16x12] 3.05x 1027.50 3135.10
chroma_p2s[12x16] 2.79x 1112.50 3100.94
chroma_p2s[ 16x4] 2.57x 387.50 994.88
chroma_p2s[ 4x16] 2.43x 489.99 1192.44
chroma_p2s[32x24] 1.42x 3907.82 5532.55
chroma_p2s[24x32] 3.95x 3907.50 15422.51
chroma_p2s[ 32x8] 1.37x 1347.50 1849.83
chroma_p2s[ 8x32] 3.31x 1387.50 4594.97
chroma_p2s[64x48] 1.27x 15458.83 19562.50
chroma_p2s[48x64] 1.30x 15427.81 20132.50
chroma_p2s[64x16] 1.28x 5187.50 6662.50
chroma_p2s[16x64] 3.05x 5197.50 15855.09
32-bit
./test/TestBench --testbench interp | grep p2s
convert_p2s[ 8x8] 2.55x 484.99 1237.46
convert_p2s[16x16] 2.89x 1444.98 4174.96
convert_p2s[32x32] 6.70x 5295.00 35485.83
convert_p2s[64x64] 6.41x 20695.04 132575.69
convert_p2s[ 8x4] 2.06x 325.00 669.93
convert_p2s[ 4x8] 1.93x 360.08 694.95
convert_p2s[ 16x8] 2.52x 805.00 2032.47
convert_p2s[ 8x16] 2.83x 820.00 2319.99
convert_p2s[32x16] 6.55x 2725.00 17855.75
convert_p2s[16x32] 3.00x 2725.00 8185.23
convert_p2s[64x32] 6.36x 10405.41 66205.67
convert_p2s[32x64] 6.82x 10404.99 70948.87
convert_p2s[16x12] 2.84x 1124.99 3197.47
convert_p2s[12x16] 2.61x 1207.52 3155.00
convert_p2s[ 16x4] 2.29x 485.00 1112.42
convert_p2s[ 4x16] 2.22x 564.99 1252.46
convert_p2s[32x24] 6.66x 4005.00 26655.58
convert_p2s[24x32] 6.99x 4005.00 28003.99
convert_p2s[ 32x8] 6.27x 1445.00 9054.99
convert_p2s[ 8x32] 3.18x 1475.00 4687.54
convert_p2s[64x48] 6.40x 15525.12 99348.35
convert_p2s[48x64] 6.54x 15525.42 101593.84
convert_p2s[64x16] 6.33x 5285.10 33457.09
convert_p2s[16x64] 3.01x 5295.02 15944.96
chroma_p2s[ 4x4] 1.64x 257.49 422.54
chroma_p2s[ 8x8] 2.55x 484.99 1237.48
chroma_p2s[16x16] 2.89x 1444.99 4174.94
chroma_p2s[32x32] 6.70x 5295.02 35485.08
chroma_p2s[ 4x2] 1.33x 212.48 282.52
chroma_p2s[ 2x4] 0.98x 272.49 267.50
chroma_p2s[ 8x4] 2.06x 325.00 669.99
chroma_p2s[ 4x8] 1.93x 360.00 694.98
chroma_p2s[ 16x8] 2.52x 805.00 2032.51
chroma_p2s[ 8x16] 2.83x 820.15 2319.98
chroma_p2s[32x16] 6.55x 2725.00 17855.50
chroma_p2s[16x32] 3.00x 2725.00 8184.98
chroma_p2s[ 8x6] 2.33x 402.50 937.48
chroma_p2s[ 6x8] 1.71x 585.00 1000.03
chroma_p2s[ 8x2] 1.62x 242.49 392.50
chroma_p2s[ 2x8] 1.15x 375.00 432.50
chroma_p2s[16x12] 2.84x 1124.99 3194.98
chroma_p2s[12x16] 2.63x 1200.00 3155.18
chroma_p2s[ 16x4] 2.29x 485.07 1112.51
chroma_p2s[ 4x16] 2.18x 562.54 1224.98
chroma_p2s[32x24] 6.66x 4005.01 26660.21
chroma_p2s[24x32] 6.94x 4005.05 27805.24
chroma_p2s[ 32x8] 6.27x 1445.00 9054.99
chroma_p2s[ 8x32] 3.18x 1475.00 4687.52
chroma_p2s[ 4x8] 2.01x 360.00 722.48
chroma_p2s[ 8x16] 2.83x 820.00 2319.98
chroma_p2s[16x32] 3.00x 2725.07 8185.31
chroma_p2s[32x64] 6.79x 10405.07 70686.27
chroma_p2s[ 4x4] 1.63x 259.98 422.48
chroma_p2s[ 2x8] 1.12x 385.00 432.50
chroma_p2s[ 8x8] 2.55x 485.00 1237.48
chroma_p2s[ 4x16] 2.18x 562.49 1224.98
chroma_p2s[16x16] 2.89x 1445.10 4174.98
chroma_p2s[ 8x32] 3.18x 1475.00 4687.56
chroma_p2s[32x32] 6.70x 5295.13 35485.12
chroma_p2s[16x64] 3.01x 5295.00 15945.08
chroma_p2s[ 8x12] 2.71x 655.09 1774.92
chroma_p2s[ 6x16] 1.80x 1057.50 1899.96
chroma_p2s[ 8x4] 2.06x 325.00 669.99
chroma_p2s[ 2x16] 1.06x 659.99 697.48
chroma_p2s[16x24] 2.93x 2085.00 6115.20
chroma_p2s[12x32] 2.72x 2250.00 6125.25
chroma_p2s[ 16x8] 2.52x 805.03 2032.57
chroma_p2s[ 4x32] 2.85x 957.54 2725.27
chroma_p2s[32x48] 6.79x 7855.06 53302.55
chroma_p2s[24x64] 7.04x 7855.01 55325.50
chroma_p2s[32x16] 6.55x 2725.00 17855.12
chroma_p2s[ 8x64] 3.23x 2805.32 9050.32
chroma_p2s[ 4x4] 1.63x 259.99 422.49
chroma_p2s[ 8x8] 2.55x 487.49 1245.00
chroma_p2s[16x16] 2.89x 1445.00 4175.20
chroma_p2s[32x32] 6.74x 5295.00 35670.43
chroma_p2s[64x64] 6.40x 20696.68 132486.70
chroma_p2s[ 8x4] 2.06x 325.00 669.99
chroma_p2s[ 4x8] 1.93x 360.03 694.99
chroma_p2s[ 16x8] 2.52x 805.00 2032.49
chroma_p2s[ 8x16] 2.83x 820.14 2320.16
chroma_p2s[32x16] 6.55x 2724.98 17855.27
chroma_p2s[16x32] 3.00x 2725.10 8185.09
chroma_p2s[64x32] 6.39x 10405.02 66479.45
chroma_p2s[32x64] 6.81x 10414.99 70945.01
chroma_p2s[16x12] 2.84x 1125.00 3194.98
chroma_p2s[12x16] 2.63x 1200.13 3155.00
chroma_p2s[ 16x4] 2.29x 485.00 1112.48
chroma_p2s[ 4x16] 2.18x 562.49 1224.98
chroma_p2s[32x24] 6.66x 4005.17 26655.93
chroma_p2s[24x32] 6.99x 4005.01 28008.09
chroma_p2s[ 32x8] 6.27x 1444.99 9060.05
chroma_p2s[ 8x32] 3.18x 1474.99 4687.52
chroma_p2s[64x48] 6.40x 15525.01 99400.07
chroma_p2s[48x64] 6.55x 15525.09 101699.22
chroma_p2s[64x16] 6.28x 5285.00 33215.04
chroma_p2s[16x64] 3.01x 5295.16 15944.98
10-bit
./test/TestBench --testbench interp | grep p2s
convert_p2s[ 8x8] 3.48x 369.50 1284.66
convert_p2s[16x16] 3.13x 1309.98 4095.75
convert_p2s[32x32] 3.63x 5150.01 18697.50
convert_p2s[64x64] 3.61x 20510.21 74113.83
convert_p2s[ 8x4] 2.80x 219.99 614.97
convert_p2s[ 4x8] 3.22x 212.49 684.96
convert_p2s[ 16x8] 3.02x 669.99 2022.54
convert_p2s[ 8x16] 3.51x 674.98 2366.56
convert_p2s[32x16] 3.67x 2589.99 9497.50
convert_p2s[16x32] 3.09x 2589.98 8015.87
convert_p2s[64x32] 3.58x 10270.03 36767.52
convert_p2s[32x64] 3.61x 10271.21 37097.50
convert_p2s[16x12] 3.17x 990.00 3136.46
convert_p2s[12x16] 3.15x 987.49 3112.45
convert_p2s[ 16x4] 2.86x 357.50 1022.46
convert_p2s[ 4x16] 3.73x 354.95 1324.90
convert_p2s[32x24] 3.64x 3869.97 14097.50
convert_p2s[24x32] 3.63x 3872.45 14068.67
convert_p2s[ 32x8] 3.60x 1309.98 4720.05
convert_p2s[ 8x32] 3.79x 1314.99 4980.84
convert_p2s[64x48] 3.57x 15391.05 55007.52
convert_p2s[48x64] 3.60x 15393.61 55424.42
convert_p2s[64x16] 3.61x 5150.39 18616.89
convert_p2s[16x64] 3.06x 5149.98 15776.41
chroma_p2s[ 4x4] 2.53x 137.50 347.48
chroma_p2s[ 8x8] 3.50x 367.50 1284.99
chroma_p2s[16x16] 3.13x 1307.50 4094.98
chroma_p2s[32x32] 3.67x 5147.55 18867.50
chroma_p2s[ 4x2] 2.27x 89.29 203.13
chroma_p2s[ 2x4] 1.51x 147.50 222.48
chroma_p2s[ 8x4] 2.83x 217.50 615.14
chroma_p2s[ 4x8] 3.30x 207.61 684.99
chroma_p2s[ 16x8] 2.98x 670.02 1999.99
chroma_p2s[ 8x16] 3.52x 672.50 2365.32
chroma_p2s[32x16] 3.67x 2587.50 9497.50
chroma_p2s[16x32] 3.10x 2587.50 8015.44
chroma_p2s[ 8x6] 3.18x 297.50 944.97
chroma_p2s[ 6x8] 2.70x 347.52 937.57
chroma_p2s[ 8x2] 2.25x 147.57 332.45
chroma_p2s[ 2x8] 1.76x 217.50 382.49
chroma_p2s[16x12] 3.18x 987.50 3135.74
chroma_p2s[12x16] 3.16x 985.00 3112.70
chroma_p2s[ 16x4] 2.84x 357.50 1014.98
chroma_p2s[ 4x16] 3.86x 343.33 1324.94
chroma_p2s[32x24] 3.65x 3867.50 14097.50
chroma_p2s[24x32] 3.63x 3869.99 14065.64
chroma_p2s[ 32x8] 3.61x 1307.50 4720.05
chroma_p2s[ 8x32] 3.79x 1312.50 4980.54
chroma_p2s[ 4x8] 3.30x 207.50 684.95
chroma_p2s[ 8x16] 3.52x 672.50 2365.13
chroma_p2s[16x32] 2.91x 2756.49 8015.58
chroma_p2s[32x64] 3.65x 10267.50 37511.57
chroma_p2s[ 4x4] 2.53x 137.50 347.49
chroma_p2s[ 2x8] 1.76x 217.50 382.49
chroma_p2s[ 8x8] 3.50x 367.50 1284.99
chroma_p2s[ 4x16] 3.86x 343.33 1324.96
chroma_p2s[16x16] 3.13x 1307.58 4095.93
chroma_p2s[ 8x32] 2.89x 1750.02 5052.14
chroma_p2s[32x32] 3.63x 5147.50 18697.50
chroma_p2s[16x64] 3.06x 5147.52 15775.37
chroma_p2s[ 8x12] 3.65x 517.50 1887.48
chroma_p2s[ 6x16] 2.90x 660.01 1912.45
chroma_p2s[ 8x4] 2.83x 217.50 614.98
chroma_p2s[ 2x16] 1.82x 376.25 684.98
chroma_p2s[16x24] 3.11x 1947.58 6054.96
chroma_p2s[12x32] 3.08x 1945.00 5993.06
chroma_p2s[ 16x8] 2.98x 670.15 1999.99
chroma_p2s[ 4x32] 3.74x 741.68 2772.84
chroma_p2s[32x48] 3.62x 7707.50 27897.86
chroma_p2s[24x64] 3.61x 7707.50 27826.29
chroma_p2s[32x16] 3.67x 2587.50 9499.59
chroma_p2s[ 8x64] 3.76x 2592.50 9741.75
chroma_p2s[ 4x4] 2.53x 137.50 347.47
chroma_p2s[ 8x8] 3.50x 367.50 1284.47
chroma_p2s[16x16] 3.13x 1307.50 4095.38
chroma_p2s[32x32] 3.63x 5147.51 18697.65
chroma_p2s[64x64] 3.58x 20511.86 73401.03
chroma_p2s[ 8x4] 2.83x 217.50 615.10
chroma_p2s[ 4x8] 3.30x 207.50 684.95
chroma_p2s[ 16x8] 2.99x 670.00 1999.99
chroma_p2s[ 8x16] 3.52x 672.50 2365.32
chroma_p2s[32x16] 3.67x 2587.50 9497.50
chroma_p2s[16x32] 3.10x 2587.53 8015.30
chroma_p2s[64x32] 3.58x 10267.54 36767.52
chroma_p2s[32x64] 3.60x 10311.38 37097.81
chroma_p2s[16x12] 3.17x 987.62 3135.36
chroma_p2s[12x16] 3.16x 985.00 3112.94
chroma_p2s[ 16x4] 2.84x 357.50 1014.97
chroma_p2s[ 4x16] 3.87x 342.50 1324.90
chroma_p2s[32x24] 3.67x 3867.50 14201.55
chroma_p2s[24x32] 3.64x 3869.99 14068.54
chroma_p2s[ 32x8] 3.58x 1307.50 4678.90
chroma_p2s[ 8x32] 3.85x 1312.54 5052.43
chroma_p2s[64x48] 3.57x 15387.58 55007.52
chroma_p2s[48x64] 3.57x 15431.11 55017.58
chroma_p2s[64x16] 3.60x 5147.50 18527.71
chroma_p2s[16x64] 3.06x 5148.45 15775.25
diff -r 0f0d88319f7c -r a828fd4b5f48 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp Tue Jun 02 12:35:53 2015 +0530
+++ b/source/common/x86/asm-primitives.cpp Thu Jun 04 18:04:06 2015 -0700
@@ -934,6 +934,20 @@
LUMA_VSS_FILTERS(sse2);
p.frameInitLowres = x265_frame_init_lowres_core_sse2;
+
+ ALL_CHROMA_420_PU(p2s, filterPixelToShort, sse2);
+ ALL_CHROMA_422_PU(p2s, filterPixelToShort, sse2);
+ ALL_CHROMA_444_PU(p2s, filterPixelToShort, sse2);
+ ALL_LUMA_PU(convert_p2s, filterPixelToShort, sse2);
+ }
+ if (cpuMask & X265_CPU_SSE3)
+ {
+ ALL_CHROMA_420_PU(filter_hpp, interp_4tap_horiz_pp, sse3);
+ ALL_CHROMA_422_PU(filter_hpp, interp_4tap_horiz_pp, sse3);
+ ALL_CHROMA_444_PU(filter_hpp, interp_4tap_horiz_pp, sse3);
+ ALL_CHROMA_420_PU(filter_hps, interp_4tap_horiz_ps, sse3);
+ ALL_CHROMA_422_PU(filter_hps, interp_4tap_horiz_ps, sse3);
+ ALL_CHROMA_444_PU(filter_hps, interp_4tap_horiz_ps, sse3);
}
if (cpuMask & X265_CPU_SSSE3)
{
@@ -1851,6 +1865,10 @@
p.idst4x4 = x265_idst4_sse2;
p.planecopy_sp = x265_downShift_16_sse2;
+ ALL_CHROMA_420_PU(p2s, filterPixelToShort, sse2);
+ ALL_CHROMA_422_PU(p2s, filterPixelToShort, sse2);
+ ALL_CHROMA_444_PU(p2s, filterPixelToShort, sse2);
+ ALL_LUMA_PU(convert_p2s, filterPixelToShort, sse2);
}
if (cpuMask & X265_CPU_SSE3)
{
diff -r 0f0d88319f7c -r a828fd4b5f48 source/common/x86/const-a.asm
--- a/source/common/x86/const-a.asm Tue Jun 02 12:35:53 2015 +0530
+++ b/source/common/x86/const-a.asm Thu Jun 04 18:04:06 2015 -0700
@@ -80,6 +80,7 @@
const pw_1023, times 16 dw 1023
const pw_1024, times 16 dw 1024
const pw_4096, times 16 dw 4096
+const pw_8192, times 8 dw 8192
const pw_00ff, times 16 dw 0x00ff
const pw_ff00, times 8 dw 0xff00
const pw_2000, times 16 dw 0x2000
diff -r 0f0d88319f7c -r a828fd4b5f48 source/common/x86/ipfilter16.asm
--- a/source/common/x86/ipfilter16.asm Tue Jun 02 12:35:53 2015 +0530
+++ b/source/common/x86/ipfilter16.asm Thu Jun 04 18:04:06 2015 -0700
@@ -564,6 +564,560 @@
FILTER_VER_LUMA_sse2 ps, 64, 16
FILTER_VER_LUMA_sse2 ps, 16, 64
+%macro FILTERH_W2_4_sse3 2
+ movh m3, [r0 + %1]
+ movhps m3, [r0 + %1 + 2]
+ pmaddwd m3, m0
+ movh m4, [r0 + r1 + %1]
+ movhps m4, [r0 + r1 + %1 + 2]
+ pmaddwd m4, m0
+ pshufd m2, m3, q2301
+ paddd m3, m2
+ pshufd m2, m4, q2301
+ paddd m4, m2
+ pshufd m3, m3, q3120
+ pshufd m4, m4, q3120
+ punpcklqdq m3, m4
+ paddd m3, m1
+ movh m5, [r0 + 2 * r1 + %1]
+ movhps m5, [r0 + 2 * r1 + %1 + 2]
+ pmaddwd m5, m0
+ movh m4, [r0 + r4 + %1]
+ movhps m4, [r0 + r4 + %1 + 2]
+ pmaddwd m4, m0
+ pshufd m2, m5, q2301
+ paddd m5, m2
+ pshufd m2, m4, q2301
+ paddd m4, m2
+ pshufd m5, m5, q3120
+ pshufd m4, m4, q3120
+ punpcklqdq m5, m4
+ paddd m5, m1
+%ifidn %2, pp
+ psrad m3, 6
+ psrad m5, 6
+ packssdw m3, m5
+ CLIPW m3, m7, m6
+%else
+ psrad m3, 2
+ psrad m5, 2
+ packssdw m3, m5
+%endif
+ movd [r2 + %1], m3
+ psrldq m3, 4
+ movd [r2 + r3 + %1], m3
+ psrldq m3, 4
+ movd [r2 + r3 * 2 + %1], m3
+ psrldq m3, 4
+ movd [r2 + r5 + %1], m3
+%endmacro
+
+%macro FILTERH_W2_3_sse3 1
+ movh m3, [r0 + %1]
+ movhps m3, [r0 + %1 + 2]
+ pmaddwd m3, m0
+ movh m4, [r0 + r1 + %1]
+ movhps m4, [r0 + r1 + %1 + 2]
+ pmaddwd m4, m0
+ pshufd m2, m3, q2301
+ paddd m3, m2
+ pshufd m2, m4, q2301
+ paddd m4, m2
+ pshufd m3, m3, q3120
+ pshufd m4, m4, q3120
+ punpcklqdq m3, m4
+ paddd m3, m1
+
+ movh m5, [r0 + 2 * r1 + %1]
+ movhps m5, [r0 + 2 * r1 + %1 + 2]
+ pmaddwd m5, m0
+
+ pshufd m2, m5, q2301
+ paddd m5, m2
+ pshufd m5, m5, q3120
+ paddd m5, m1
+
+ psrad m3, 2
+ psrad m5, 2
+ packssdw m3, m5
+
+ movd [r2 + %1], m3
+ psrldq m3, 4
+ movd [r2 + r3 + %1], m3
+ psrldq m3, 4
+ movd [r2 + r3 * 2 + %1], m3
+%endmacro
+
+%macro FILTERH_W4_2_sse3 2
+ movh m3, [r0 + %1]
+ movhps m3, [r0 + %1 + 2]
+ pmaddwd m3, m0
+ movh m4, [r0 + %1 + 4]
+ movhps m4, [r0 + %1 + 6]
+ pmaddwd m4, m0
+ pshufd m2, m3, q2301
+ paddd m3, m2
+ pshufd m2, m4, q2301
+ paddd m4, m2
+ pshufd m3, m3, q3120
+ pshufd m4, m4, q3120
+ punpcklqdq m3, m4
+ paddd m3, m1
+
+ movh m5, [r0 + r1 + %1]
+ movhps m5, [r0 + r1 + %1 + 2]
+ pmaddwd m5, m0
+ movh m4, [r0 + r1 + %1 + 4]
+ movhps m4, [r0 + r1 + %1 + 6]
+ pmaddwd m4, m0
+ pshufd m2, m5, q2301
+ paddd m5, m2
+ pshufd m2, m4, q2301
+ paddd m4, m2
+ pshufd m5, m5, q3120
+ pshufd m4, m4, q3120
+ punpcklqdq m5, m4
+ paddd m5, m1
+%ifidn %2, pp
+ psrad m3, 6
+ psrad m5, 6
+ packssdw m3, m5
+ CLIPW m3, m7, m6
+%else
+ psrad m3, 2
+ psrad m5, 2
+ packssdw m3, m5
+%endif
+ movh [r2 + %1], m3
+ movhps [r2 + r3 + %1], m3
+%endmacro
+
+%macro FILTERH_W4_1_sse3 1
+ movh m3, [r0 + 2 * r1 + %1]
+ movhps m3, [r0 + 2 * r1 + %1 + 2]
+ pmaddwd m3, m0
+ movh m4, [r0 + 2 * r1 + %1 + 4]
+ movhps m4, [r0 + 2 * r1 + %1 + 6]
+ pmaddwd m4, m0
+ pshufd m2, m3, q2301
+ paddd m3, m2
+ pshufd m2, m4, q2301
+ paddd m4, m2
+ pshufd m3, m3, q3120
+ pshufd m4, m4, q3120
+ punpcklqdq m3, m4
+ paddd m3, m1
+
+ psrad m3, 2
+ packssdw m3, m3
+ movh [r2 + r3 * 2 + %1], m3
+%endmacro
+
+%macro FILTERH_W8_1_sse3 2
+ movh m3, [r0 + %1]
+ movhps m3, [r0 + %1 + 2]
+ pmaddwd m3, m0
+ movh m4, [r0 + %1 + 4]
+ movhps m4, [r0 + %1 + 6]
+ pmaddwd m4, m0
+ pshufd m2, m3, q2301
+ paddd m3, m2
+ pshufd m2, m4, q2301
+ paddd m4, m2
+ pshufd m3, m3, q3120
+ pshufd m4, m4, q3120
+ punpcklqdq m3, m4
+ paddd m3, m1
+
+ movh m5, [r0 + %1 + 8]
+ movhps m5, [r0 + %1 + 10]
+ pmaddwd m5, m0
+ movh m4, [r0 + %1 + 12]
+ movhps m4, [r0 + %1 + 14]
+ pmaddwd m4, m0
+ pshufd m2, m5, q2301
+ paddd m5, m2
+ pshufd m2, m4, q2301
+ paddd m4, m2
+ pshufd m5, m5, q3120
+ pshufd m4, m4, q3120
+ punpcklqdq m5, m4
+ paddd m5, m1
+%ifidn %2, pp
+ psrad m3, 6
+ psrad m5, 6
+ packssdw m3, m5
+ CLIPW m3, m7, m6
+%else
+ psrad m3, 2
+ psrad m5, 2
+ packssdw m3, m5
+%endif
+ movdqu [r2 + %1], m3
+%endmacro
+
+;-----------------------------------------------------------------------------
+; void interp_4tap_horiz_%3_%1x%2(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx)
+;-----------------------------------------------------------------------------
+%macro FILTER_HOR_CHROMA_sse3 3
+INIT_XMM sse3
+cglobal interp_4tap_horiz_%3_%1x%2, 4, 7, 8
+ add r3, r3
+ add r1, r1
+ sub r0, 2
+ mov r4d, r4m
+ add r4d, r4d
+
+%ifdef PIC
+ lea r6, [tab_ChromaCoeff]
+ movddup m0, [r6 + r4 * 4]
+%else
+ movddup m0, [tab_ChromaCoeff + r4 * 4]
+%endif
+
+%ifidn %3, ps
+ mova m1, [tab_c_n32768]
+ cmp r5m, byte 0
+%if %1 <= 6
+ lea r4, [r1 * 3]
+ lea r5, [r3 * 3]
+%endif
+ je .skip
+ sub r0, r1
+%if %1 <= 6
+%assign y 1
+%else
+%assign y 3
+%endif
+%assign z 0
+%rep y
+%assign x 0
+%rep %1/8
+ FILTERH_W8_1_sse3 x, %3
+%assign x x+16
+%endrep
+%if %1 == 4 || (%1 == 6 && z == 0) || (%1 == 12 && z == 0)
+ FILTERH_W4_2_sse3 x, %3
+ FILTERH_W4_1_sse3 x
+%assign x x+8
+%endif
+%if %1 == 2 || (%1 == 6 && z == 0)
+ FILTERH_W2_3_sse3 x
+%endif
+%if %1 <= 6
+ lea r0, [r0 + r4]
+ lea r2, [r2 + r5]
+%else
+ lea r0, [r0 + r1]
+ lea r2, [r2 + r3]
+%endif
+%assign z z+1
+%endrep
+.skip:
+%elifidn %3, pp
+ pxor m7, m7
+ mova m6, [pw_pixel_max]
+ mova m1, [tab_c_32]
+%if %1 == 2 || %1 == 6
+ lea r4, [r1 * 3]
+ lea r5, [r3 * 3]
+%endif
+%endif
+
+%if %1 == 2
+%assign y %2/4
+%elif %1 <= 6
+%assign y %2/2
+%else
+%assign y %2
+%endif
+%assign z 0
+%rep y
+%assign x 0
+%rep %1/8
+ FILTERH_W8_1_sse3 x, %3
+%assign x x+16
+%endrep
+%if %1 == 4 || %1 == 6 || (%1 == 12 && (z % 2) == 0)
+ FILTERH_W4_2_sse3 x, %3
+%assign x x+8
+%endif
+%if %1 == 2 || (%1 == 6 && (z % 2) == 0)
+ FILTERH_W2_4_sse3 x, %3
+%endif
+%assign z z+1
+%if z < y
+%if %1 == 2
+ lea r0, [r0 + 4 * r1]
+ lea r2, [r2 + 4 * r3]
+%elif %1 <= 6
+ lea r0, [r0 + 2 * r1]
+ lea r2, [r2 + 2 * r3]
+%else
+ lea r0, [r0 + r1]
+ lea r2, [r2 + r3]
+%endif
+%endif ;z < y
+%endrep
+
+RET
+%endmacro
+
+;-----------------------------------------------------------------------------
+; void interp_4tap_horiz_pp_%1x%2(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx)
+;-----------------------------------------------------------------------------
+
+FILTER_HOR_CHROMA_sse3 2, 4, pp
+FILTER_HOR_CHROMA_sse3 2, 8, pp
+FILTER_HOR_CHROMA_sse3 2, 16, pp
+FILTER_HOR_CHROMA_sse3 4, 2, pp
+FILTER_HOR_CHROMA_sse3 4, 4, pp
+FILTER_HOR_CHROMA_sse3 4, 8, pp
+FILTER_HOR_CHROMA_sse3 4, 16, pp
+FILTER_HOR_CHROMA_sse3 4, 32, pp
+FILTER_HOR_CHROMA_sse3 6, 8, pp
+FILTER_HOR_CHROMA_sse3 6, 16, pp
+FILTER_HOR_CHROMA_sse3 8, 2, pp
+FILTER_HOR_CHROMA_sse3 8, 4, pp
+FILTER_HOR_CHROMA_sse3 8, 6, pp
+FILTER_HOR_CHROMA_sse3 8, 8, pp
+FILTER_HOR_CHROMA_sse3 8, 12, pp
+FILTER_HOR_CHROMA_sse3 8, 16, pp
+FILTER_HOR_CHROMA_sse3 8, 32, pp
+FILTER_HOR_CHROMA_sse3 8, 64, pp
+FILTER_HOR_CHROMA_sse3 12, 16, pp
+FILTER_HOR_CHROMA_sse3 12, 32, pp
+FILTER_HOR_CHROMA_sse3 16, 4, pp
+FILTER_HOR_CHROMA_sse3 16, 8, pp
+FILTER_HOR_CHROMA_sse3 16, 12, pp
+FILTER_HOR_CHROMA_sse3 16, 16, pp
+FILTER_HOR_CHROMA_sse3 16, 24, pp
+FILTER_HOR_CHROMA_sse3 16, 32, pp
+FILTER_HOR_CHROMA_sse3 16, 64, pp
+FILTER_HOR_CHROMA_sse3 24, 32, pp
+FILTER_HOR_CHROMA_sse3 24, 64, pp
+FILTER_HOR_CHROMA_sse3 32, 8, pp
+FILTER_HOR_CHROMA_sse3 32, 16, pp
+FILTER_HOR_CHROMA_sse3 32, 24, pp
+FILTER_HOR_CHROMA_sse3 32, 32, pp
+FILTER_HOR_CHROMA_sse3 32, 48, pp
+FILTER_HOR_CHROMA_sse3 32, 64, pp
+FILTER_HOR_CHROMA_sse3 48, 64, pp
+FILTER_HOR_CHROMA_sse3 64, 16, pp
+FILTER_HOR_CHROMA_sse3 64, 32, pp
+FILTER_HOR_CHROMA_sse3 64, 48, pp
+FILTER_HOR_CHROMA_sse3 64, 64, pp
+
+;-----------------------------------------------------------------------------
+; void interp_4tap_horiz_ps_%1x%2(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx)
+;-----------------------------------------------------------------------------
+
+FILTER_HOR_CHROMA_sse3 2, 4, ps
+FILTER_HOR_CHROMA_sse3 2, 8, ps
+FILTER_HOR_CHROMA_sse3 2, 16, ps
+FILTER_HOR_CHROMA_sse3 4, 2, ps
+FILTER_HOR_CHROMA_sse3 4, 4, ps
+FILTER_HOR_CHROMA_sse3 4, 8, ps
+FILTER_HOR_CHROMA_sse3 4, 16, ps
+FILTER_HOR_CHROMA_sse3 4, 32, ps
+FILTER_HOR_CHROMA_sse3 6, 8, ps
+FILTER_HOR_CHROMA_sse3 6, 16, ps
+FILTER_HOR_CHROMA_sse3 8, 2, ps
+FILTER_HOR_CHROMA_sse3 8, 4, ps
+FILTER_HOR_CHROMA_sse3 8, 6, ps
+FILTER_HOR_CHROMA_sse3 8, 8, ps
+FILTER_HOR_CHROMA_sse3 8, 12, ps
+FILTER_HOR_CHROMA_sse3 8, 16, ps
+FILTER_HOR_CHROMA_sse3 8, 32, ps
+FILTER_HOR_CHROMA_sse3 8, 64, ps
+FILTER_HOR_CHROMA_sse3 12, 16, ps
+FILTER_HOR_CHROMA_sse3 12, 32, ps
+FILTER_HOR_CHROMA_sse3 16, 4, ps
+FILTER_HOR_CHROMA_sse3 16, 8, ps
+FILTER_HOR_CHROMA_sse3 16, 12, ps
+FILTER_HOR_CHROMA_sse3 16, 16, ps
+FILTER_HOR_CHROMA_sse3 16, 24, ps
+FILTER_HOR_CHROMA_sse3 16, 32, ps
+FILTER_HOR_CHROMA_sse3 16, 64, ps
+FILTER_HOR_CHROMA_sse3 24, 32, ps
+FILTER_HOR_CHROMA_sse3 24, 64, ps
+FILTER_HOR_CHROMA_sse3 32, 8, ps
+FILTER_HOR_CHROMA_sse3 32, 16, ps
+FILTER_HOR_CHROMA_sse3 32, 24, ps
+FILTER_HOR_CHROMA_sse3 32, 32, ps
+FILTER_HOR_CHROMA_sse3 32, 48, ps
+FILTER_HOR_CHROMA_sse3 32, 64, ps
+FILTER_HOR_CHROMA_sse3 48, 64, ps
+FILTER_HOR_CHROMA_sse3 64, 16, ps
+FILTER_HOR_CHROMA_sse3 64, 32, ps
+FILTER_HOR_CHROMA_sse3 64, 48, ps
+FILTER_HOR_CHROMA_sse3 64, 64, ps
+
+%macro FILTER_P2S_2_4_sse2 1
+ movd m0, [r0 + %1]
+ movd m2, [r0 + r1 * 2 + %1]
+ movhps m0, [r0 + r1 + %1]
+ movhps m2, [r0 + r4 + %1]
+ psllw m0, 4
+ psllw m2, 4
+ psubw m0, m1
+ psubw m2, m1
+
+ movd [r2 + r3 * 0 + %1], m0
+ movd [r2 + r3 * 2 + %1], m2
+ movhlps m0, m0
+ movhlps m2, m2
+ movd [r2 + r3 * 1 + %1], m0
+ movd [r2 + r5 + %1], m2
+%endmacro
+
+%macro FILTER_P2S_4_4_sse2 1
+ movh m0, [r0 + %1]
+ movhps m0, [r0 + r1 + %1]
+ psllw m0, 4
+ psubw m0, m1
+ movh [r2 + r3 * 0 + %1], m0
+ movhps [r2 + r3 * 1 + %1], m0
+
+ movh m2, [r0 + r1 * 2 + %1]
+ movhps m2, [r0 + r4 + %1]
+ psllw m2, 4
+ psubw m2, m1
+ movh [r2 + r3 * 2 + %1], m2
+ movhps [r2 + r5 + %1], m2
+%endmacro
+
+%macro FILTER_P2S_4_2_sse2 0
+ movh m0, [r0]
+ movhps m0, [r0 + r1 * 2]
+ psllw m0, 4
+ psubw m0, [pw_2000]
+ movh [r2 + r3 * 0], m0
+ movhps [r2 + r3 * 2], m0
+%endmacro
+
+%macro FILTER_P2S_8_4_sse2 1
+ movu m0, [r0 + %1]
+ movu m2, [r0 + r1 + %1]
+ psllw m0, 4
+ psllw m2, 4
+ psubw m0, m1
+ psubw m2, m1
+ movu [r2 + r3 * 0 + %1], m0
+ movu [r2 + r3 * 1 + %1], m2
+
+ movu m3, [r0 + r1 * 2 + %1]
+ movu m4, [r0 + r4 + %1]
+ psllw m3, 4
+ psllw m4, 4
+ psubw m3, m1
+ psubw m4, m1
+ movu [r2 + r3 * 2 + %1], m3
+ movu [r2 + r5 + %1], m4
+%endmacro
+
+%macro FILTER_P2S_8_2_sse2 1
+ movu m0, [r0 + %1]
+ movu m2, [r0 + r1 + %1]
+ psllw m0, 4
+ psllw m2, 4
+ psubw m0, m1
+ psubw m2, m1
+ movu [r2 + r3 * 0 + %1], m0
+ movu [r2 + r3 * 1 + %1], m2
+%endmacro
+
+;-----------------------------------------------------------------------------
+; void filterPixelToShort(pixel *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride)
+;-----------------------------------------------------------------------------
+%macro FILTER_PIX_TO_SHORT_sse2 2
+INIT_XMM sse2
+cglobal filterPixelToShort_%1x%2, 4, 6, 3
+%if %2 == 2
+%if %1 == 4
+ FILTER_P2S_4_2_sse2
+%elif %1 == 8
+ add r1d, r1d
+ add r3d, r3d
+ mova m1, [pw_2000]
+ FILTER_P2S_8_2_sse2 0
+%endif
+%else
+ add r1d, r1d
+ add r3d, r3d
+ mova m1, [pw_2000]
+ lea r4, [r1 * 3]
+ lea r5, [r3 * 3]
+%assign y 1
+%rep %2/4
+%assign x 0
+%rep %1/8
+ FILTER_P2S_8_4_sse2 x
+%if %2 == 6
+ lea r0, [r0 + 4 * r1]
+ lea r2, [r2 + 4 * r3]
+ FILTER_P2S_8_2_sse2 x
+%endif
+%assign x x+16
+%endrep
+%rep (%1 % 8)/4
+ FILTER_P2S_4_4_sse2 x
+%assign x x+8
+%endrep
+%rep (%1 % 4)/2
+ FILTER_P2S_2_4_sse2 x
+%endrep
+%if y < %2/4
+ lea r0, [r0 + 4 * r1]
+ lea r2, [r2 + 4 * r3]
+%assign y y+1
+%endif
+%endrep
+%endif
+RET
+%endmacro
+
+ FILTER_PIX_TO_SHORT_sse2 2, 4
+ FILTER_PIX_TO_SHORT_sse2 2, 8
+ FILTER_PIX_TO_SHORT_sse2 2, 16
+ FILTER_PIX_TO_SHORT_sse2 4, 2
+ FILTER_PIX_TO_SHORT_sse2 4, 4
+ FILTER_PIX_TO_SHORT_sse2 4, 8
+ FILTER_PIX_TO_SHORT_sse2 4, 16
+ FILTER_PIX_TO_SHORT_sse2 4, 32
+ FILTER_PIX_TO_SHORT_sse2 6, 8
+ FILTER_PIX_TO_SHORT_sse2 6, 16
+ FILTER_PIX_TO_SHORT_sse2 8, 2
+ FILTER_PIX_TO_SHORT_sse2 8, 4
+ FILTER_PIX_TO_SHORT_sse2 8, 6
+ FILTER_PIX_TO_SHORT_sse2 8, 8
+ FILTER_PIX_TO_SHORT_sse2 8, 12
+ FILTER_PIX_TO_SHORT_sse2 8, 16
+ FILTER_PIX_TO_SHORT_sse2 8, 32
+ FILTER_PIX_TO_SHORT_sse2 8, 64
+ FILTER_PIX_TO_SHORT_sse2 12, 16
+ FILTER_PIX_TO_SHORT_sse2 12, 32
+ FILTER_PIX_TO_SHORT_sse2 16, 4
+ FILTER_PIX_TO_SHORT_sse2 16, 8
+ FILTER_PIX_TO_SHORT_sse2 16, 12
+ FILTER_PIX_TO_SHORT_sse2 16, 16
+ FILTER_PIX_TO_SHORT_sse2 16, 24
+ FILTER_PIX_TO_SHORT_sse2 16, 32
+ FILTER_PIX_TO_SHORT_sse2 16, 64
+ FILTER_PIX_TO_SHORT_sse2 24, 32
+ FILTER_PIX_TO_SHORT_sse2 24, 64
+ FILTER_PIX_TO_SHORT_sse2 32, 8
+ FILTER_PIX_TO_SHORT_sse2 32, 16
+ FILTER_PIX_TO_SHORT_sse2 32, 24
+ FILTER_PIX_TO_SHORT_sse2 32, 32
+ FILTER_PIX_TO_SHORT_sse2 32, 48
+ FILTER_PIX_TO_SHORT_sse2 32, 64
+ FILTER_PIX_TO_SHORT_sse2 48, 64
+ FILTER_PIX_TO_SHORT_sse2 64, 16
+ FILTER_PIX_TO_SHORT_sse2 64, 32
+ FILTER_PIX_TO_SHORT_sse2 64, 48
+ FILTER_PIX_TO_SHORT_sse2 64, 64
+
;------------------------------------------------------------------------------------------------------------
; void interp_8tap_horiz_pp_4x4(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx)
;------------------------------------------------------------------------------------------------------------
diff -r 0f0d88319f7c -r a828fd4b5f48 source/common/x86/ipfilter8.asm
--- a/source/common/x86/ipfilter8.asm Tue Jun 02 12:35:53 2015 +0530
+++ b/source/common/x86/ipfilter8.asm Thu Jun 04 18:04:06 2015 -0700
@@ -301,6 +301,7 @@
cextern pw_32
cextern pw_512
cextern pw_2000
+cextern pw_8192
%macro FILTER_H4_w2_2_sse2 0
pxor m3, m3
@@ -3283,6 +3284,183 @@
FILTER_V4_W16n_H2_sse2 ps, 64, 16
%endif
+%macro FILTER_P2S_2_4_sse2 1
+ movd m2, [r0 + %1]
+ movd m3, [r0 + r1 + %1]
+ punpcklwd m2, m3
+ movd m3, [r0 + r1 * 2 + %1]
+ movd m4, [r0 + r4 + %1]
+ punpcklwd m3, m4
+ punpckldq m2, m3
+ punpcklbw m2, m0
+ psllw m2, 6
+ psubw m2, m1
+
+ movd [r2 + r3 * 0 + %1 * 2], m2
+ psrldq m2, 4
+ movd [r2 + r3 * 1 + %1 * 2], m2
+ psrldq m2, 4
+ movd [r2 + r3 * 2 + %1 * 2], m2
+ psrldq m2, 4
+ movd [r2 + r5 + %1 * 2], m2
+%endmacro
+
+%macro FILTER_P2S_4_4_sse2 1
+ movd m2, [r0 + %1]
+ movd m3, [r0 + r1 + %1]
+ movd m4, [r0 + r1 * 2 + %1]
+ movd m5, [r0 + r4 + %1]
+ punpckldq m2, m3
+ punpcklbw m2, m0
+ punpckldq m4, m5
+ punpcklbw m4, m0
+ psllw m2, 6
+ psllw m4, 6
+ psubw m2, m1
+ psubw m4, m1
+ movh [r2 + r3 * 0 + %1 * 2], m2
+ movh [r2 + r3 * 2 + %1 * 2], m4
+ movhps [r2 + r3 * 1 + %1 * 2], m2
+ movhps [r2 + r5 + %1 * 2], m4
+%endmacro
+
+%macro FILTER_P2S_4_2_sse2 0
+ movd m2, [r0]
+ movd m3, [r0 + r1]
+ punpckldq m2, m3
+ punpcklbw m2, m0
+ psllw m2, 6
+ psubw m2, [pw_8192]
+ movh [r2], m2
+ movhps [r2 + r3 * 2], m2
+%endmacro
+
+%macro FILTER_P2S_8_4_sse2 1
+ movh m2, [r0 + %1]
+ movh m3, [r0 + r1 + %1]
+ movh m4, [r0 + r1 * 2 + %1]
+ movh m5, [r0 + r4 + %1]
+ punpcklbw m2, m0
+ punpcklbw m3, m0
+ punpcklbw m5, m0
+ punpcklbw m4, m0
+ psllw m2, 6
+ psllw m3, 6
+ psllw m5, 6
+ psllw m4, 6
+ psubw m2, m1
+ psubw m3, m1
+ psubw m4, m1
+ psubw m5, m1
+ movu [r2 + r3 * 0 + %1 * 2], m2
+ movu [r2 + r3 * 1 + %1 * 2], m3
+ movu [r2 + r3 * 2 + %1 * 2], m4
+ movu [r2 + r5 + %1 * 2], m5
+%endmacro
+
+%macro FILTER_P2S_8_2_sse2 1
+ movh m2, [r0 + %1]
+ movh m3, [r0 + r1 + %1]
+ punpcklbw m2, m0
+ punpcklbw m3, m0
+ psllw m2, 6
+ psllw m3, 6
+ psubw m2, m1
+ psubw m3, m1
+ movu [r2 + r3 * 0 + %1 * 2], m2
+ movu [r2 + r3 * 1 + %1 * 2], m3
+%endmacro
+
+;-----------------------------------------------------------------------------
+; void filterPixelToShort(pixel *src, intptr_t srcStride, int16_t *dst, intptr_t dstStride)
+;-----------------------------------------------------------------------------
+%macro FILTER_PIX_TO_SHORT_sse2 2
+INIT_XMM sse2
+cglobal filterPixelToShort_%1x%2, 4, 6, 6
+ pxor m0, m0
+%if %2 == 2
+%if %1 == 4
+ FILTER_P2S_4_2_sse2
+%elif %1 == 8
+ add r3d, r3d
+ mova m1, [pw_8192]
+ FILTER_P2S_8_2_sse2 0
+%endif
+%else
+ add r3d, r3d
+ mova m1, [pw_8192]
+ lea r4, [r1 * 3]
+ lea r5, [r3 * 3]
+%assign y 1
+%rep %2/4
+%assign x 0
+%rep %1/8
+ FILTER_P2S_8_4_sse2 x
+%if %2 == 6
+ lea r0, [r0 + 4 * r1]
+ lea r2, [r2 + 4 * r3]
+ FILTER_P2S_8_2_sse2 x
+%endif
+%assign x x+8
+%endrep
+%rep (%1 % 8)/4
+ FILTER_P2S_4_4_sse2 x
+%assign x x+4
+%endrep
+%rep (%1 % 4)/2
+ FILTER_P2S_2_4_sse2 x
+%endrep
+%if y < %2/4
+ lea r0, [r0 + 4 * r1]
+ lea r2, [r2 + 4 * r3]
+%assign y y+1
+%endif
+%endrep
+%endif
+RET
+%endmacro
+
+ FILTER_PIX_TO_SHORT_sse2 2, 4
+ FILTER_PIX_TO_SHORT_sse2 2, 8
+ FILTER_PIX_TO_SHORT_sse2 2, 16
+ FILTER_PIX_TO_SHORT_sse2 4, 2
+ FILTER_PIX_TO_SHORT_sse2 4, 4
+ FILTER_PIX_TO_SHORT_sse2 4, 8
+ FILTER_PIX_TO_SHORT_sse2 4, 16
+ FILTER_PIX_TO_SHORT_sse2 4, 32
+ FILTER_PIX_TO_SHORT_sse2 6, 8
+ FILTER_PIX_TO_SHORT_sse2 6, 16
+ FILTER_PIX_TO_SHORT_sse2 8, 2
+ FILTER_PIX_TO_SHORT_sse2 8, 4
+ FILTER_PIX_TO_SHORT_sse2 8, 6
+ FILTER_PIX_TO_SHORT_sse2 8, 8
+ FILTER_PIX_TO_SHORT_sse2 8, 12
+ FILTER_PIX_TO_SHORT_sse2 8, 16
+ FILTER_PIX_TO_SHORT_sse2 8, 32
+ FILTER_PIX_TO_SHORT_sse2 8, 64
+ FILTER_PIX_TO_SHORT_sse2 12, 16
+ FILTER_PIX_TO_SHORT_sse2 12, 32
+ FILTER_PIX_TO_SHORT_sse2 16, 4
+ FILTER_PIX_TO_SHORT_sse2 16, 8
+ FILTER_PIX_TO_SHORT_sse2 16, 12
+ FILTER_PIX_TO_SHORT_sse2 16, 16
+ FILTER_PIX_TO_SHORT_sse2 16, 24
+ FILTER_PIX_TO_SHORT_sse2 16, 32
+ FILTER_PIX_TO_SHORT_sse2 16, 64
+ FILTER_PIX_TO_SHORT_sse2 24, 32
+ FILTER_PIX_TO_SHORT_sse2 24, 64
+ FILTER_PIX_TO_SHORT_sse2 32, 8
+ FILTER_PIX_TO_SHORT_sse2 32, 16
+ FILTER_PIX_TO_SHORT_sse2 32, 24
+ FILTER_PIX_TO_SHORT_sse2 32, 32
+ FILTER_PIX_TO_SHORT_sse2 32, 48
+ FILTER_PIX_TO_SHORT_sse2 32, 64
+ FILTER_PIX_TO_SHORT_sse2 48, 64
+ FILTER_PIX_TO_SHORT_sse2 64, 16
+ FILTER_PIX_TO_SHORT_sse2 64, 32
+ FILTER_PIX_TO_SHORT_sse2 64, 48
+ FILTER_PIX_TO_SHORT_sse2 64, 64
+
%macro FILTER_H4_w2_2 3
movh %2, [srcq - 1]
pshufb %2, %2, Tm0
@@ -3299,6 +3477,7 @@
mov [dstq + dststrideq], r4w
%endmacro
+
;-----------------------------------------------------------------------------
; void interp_4tap_horiz_pp_2x4(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx)
;-----------------------------------------------------------------------------
diff -r 0f0d88319f7c -r a828fd4b5f48 source/common/x86/ipfilter8.h
--- a/source/common/x86/ipfilter8.h Tue Jun 02 12:35:53 2015 +0530
+++ b/source/common/x86/ipfilter8.h Thu Jun 04 18:04:06 2015 -0700
@@ -965,6 +965,46 @@
void x265_interp_4tap_vert_ps_4x8_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx);
void x265_interp_4tap_vert_ps_4x16_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx);
void x265_interp_4tap_vert_ps_4x32_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx);
+void x265_filterPixelToShort_2x4_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_2x8_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_2x16_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_4x2_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_4x4_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_4x8_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_4x16_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_4x32_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_6x8_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_6x16_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_8x2_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_8x4_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_8x6_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_8x8_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_8x12_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_8x16_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_8x32_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_8x64_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_12x32_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_16x4_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_16x8_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_16x12_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_16x16_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_16x24_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_16x32_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_16x64_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_24x64_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_32x8_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_32x16_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_32x24_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_32x32_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_32x48_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_32x64_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_64x16_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_64x32_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_64x48_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_64x64_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_12x16_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_24x32_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
+void x265_filterPixelToShort_48x64_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride);
#ifdef X86_64
void x265_interp_4tap_vert_pp_6x8_sse2(const pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
void x265_interp_4tap_vert_pp_6x16_sse2(const pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
More information about the x265-devel
mailing list