[x265] [PATCH] asm: interp_8tap_horiz pp and ps sse2
dtyx265 at gmail.com
dtyx265 at gmail.com
Wed Apr 29 00:27:27 CEST 2015
# HG changeset patch
# User David T Yuen <dtyx265 at gmail.com>
# Date 1430259967 25200
# Node ID 6108fbda1be654a481a78f7ef593518033919674
# Parent e9df93f380664932e7d6c7e85b2cae16cd5e1dcd
asm: interp_8tap_horiz pp and ps sse2
This replaces c code and covers
4x4, 4x8, 4x16, 8x4, 8x8, 8x16, 8x32, 12x16, 16x8, 16x12, 16x16, 16x32, 16x64,
24x32, 32x8, 32x16, 32x24, 32x32, 32x64, 48x64, 64x16, 64x32, 64x48, 64x64
64-bit
./test/TestBench --testbench interp | grep luma_h
luma_hpp[ 4x4] 1.93x 1785.20 3440.44
luma_hps[ 4x4] 1.79x 4657.95 8317.65
luma_hpp[ 8x8] 2.22x 6005.27 13358.12
luma_hps[ 8x8] 1.93x 10947.71 21138.45
luma_hpp[16x16] 2.56x 23857.76 61174.20
luma_hps[16x16] 2.28x 32898.86 75057.14
luma_hpp[32x32] 2.56x 94326.77 241235.27
luma_hps[32x32] 2.21x 110846.29 244798.20
luma_hpp[64x64] 2.53x 376398.12 951384.19
luma_hps[64x64] 2.30x 402133.00 925341.62
luma_hpp[ 8x4] 2.37x 3025.66 7175.86
luma_hps[ 8x4] 1.92x 8127.56 15578.26
luma_hpp[ 4x8] 1.94x 3562.82 6902.69
luma_hps[ 4x8] 1.81x 6217.55 11237.91
luma_hpp[ 16x8] 2.55x 11910.52 30372.57
luma_hps[ 16x8] 2.28x 21563.62 49186.05
luma_hpp[ 8x16] 2.22x 12171.61 27009.06
luma_hps[ 8x16] 1.94x 16615.39 32258.58
luma_hpp[32x16] 2.54x 47296.66 120222.60
luma_hps[32x16] 2.21x 65472.99 144738.33
luma_hpp[16x32] 2.55x 47522.79 121164.35
luma_hps[16x32] 2.28x 55655.69 126994.38
luma_hpp[64x32] 2.53x 188087.05 476171.22
luma_hps[64x32] 2.31x 221113.20 511619.09
luma_hpp[32x64] 2.56x 188168.89 482277.28
luma_hps[32x64] 2.21x 201237.75 445382.62
luma_hpp[16x12] 2.53x 17977.69 45532.54
luma_hps[16x12] 2.28x 27197.39 62106.50
luma_hpp[12x16] 2.34x 18912.56 44205.58
luma_hps[12x16] 1.90x 25558.93 48526.43
luma_hpp[ 16x4] 2.55x 5973.14 15212.48
luma_hps[ 16x4] 2.26x 15899.24 35920.28
luma_hpp[ 4x16] 2.06x 7217.65 14895.11
luma_hps[ 4x16] 1.83x 9337.61 17077.48
luma_hpp[32x24] 2.55x 70767.15 180183.33
luma_hps[32x24] 2.21x 88166.45 194734.70
luma_hpp[24x32] 2.56x 70825.14 181610.52
luma_hps[24x32] 2.23x 83238.99 185634.08
luma_hpp[ 32x8] 2.55x 23588.33 60156.39
luma_hps[ 32x8] 2.20x 42897.08 94313.20
luma_hpp[ 8x32] 2.40x 23999.96 57624.74
luma_hps[ 8x32] 1.94x 28095.17 54599.12
luma_hpp[64x48] 2.53x 282031.94 712775.44
luma_hps[64x48] 2.30x 311860.69 716810.94
luma_hpp[48x64] 2.53x 282089.56 712646.94
luma_hps[48x64] 2.19x 301830.38 662193.25
luma_hpp[64x16] 2.52x 94224.15 237673.86
luma_hps[64x16] 2.30x 130537.17 299657.62
luma_hpp[16x64] 2.57x 94691.69 243003.95
luma_hps[16x64] 2.29x 100925.21 231004.50
32-bit
./test/TestBench --testbench interp | grep luma_h
luma_hpp[ 4x4] 2.03x 1868.23 3794.21
luma_hps[ 4x4] 1.85x 4657.62 8630.03
luma_hpp[ 8x8] 2.35x 6195.20 14560.19
luma_hps[ 8x8] 1.94x 11035.12 21390.28
luma_hpp[16x16] 2.71x 24079.95 65327.17
luma_hps[16x16] 2.32x 32982.48 76565.53
luma_hpp[32x32] 2.66x 94731.72 252013.30
luma_hps[32x32] 2.24x 111078.02 248742.02
luma_hpp[64x64] 2.61x 377356.12 985574.31
luma_hps[64x64] 2.19x 403961.25 885788.44
luma_hpp[ 8x4] 2.30x 3175.29 7296.47
luma_hps[ 8x4] 1.92x 8215.03 15791.43
luma_hpp[ 4x8] 2.10x 3647.63 7652.95
luma_hps[ 4x8] 1.88x 6177.56 11629.97
luma_hpp[ 16x8] 2.71x 12005.05 32576.90
luma_hps[ 16x8] 2.32x 21651.47 50298.37
luma_hpp[ 8x16] 2.38x 12287.62 29195.14
luma_hps[ 8x16] 1.95x 16675.90 32590.87
luma_hpp[32x16] 2.66x 47354.46 126068.27
luma_hps[32x16] 2.25x 65691.70 147645.44
luma_hpp[16x32] 2.74x 47591.86 130209.16
luma_hps[16x32] 2.36x 55761.86 131400.09
luma_hpp[64x32] 2.62x 188984.22 494236.72
luma_hps[64x32] 2.19x 222335.78 486210.91
luma_hpp[32x64] 2.64x 189273.75 500425.28
luma_hps[32x64] 2.25x 201834.66 453847.62
luma_hpp[16x12] 2.73x 18077.49 49263.77
luma_hps[16x12] 2.32x 27379.47 63601.13
luma_hpp[12x16] 2.30x 19017.79 43714.32
luma_hps[12x16] 1.92x 25617.42 49177.80
luma_hpp[ 16x4] 2.87x 6068.10 17385.54
luma_hps[ 16x4] 2.33x 15973.71 37155.74
luma_hpp[ 4x16] 2.11x 7328.28 15452.48
luma_hps[ 4x16] 1.94x 9074.02 17630.57
luma_hpp[32x24] 2.66x 70822.45 188204.25
luma_hps[32x24] 2.24x 88687.81 198895.23
luma_hpp[24x32] 2.66x 71223.68 189182.20
luma_hps[24x32] 2.28x 83453.07 190015.16
luma_hpp[ 32x8] 2.65x 23702.91 62783.11
luma_hps[ 32x8] 2.26x 42828.17 96832.39
luma_hpp[ 8x32] 2.44x 23880.27 58202.21
luma_hps[ 8x32] 1.96x 28077.69 55090.10
luma_hpp[64x48] 2.61x 283508.34 739777.50
luma_hps[64x48] 2.23x 313392.56 699904.75
luma_hpp[48x64] 2.63x 283169.62 744934.31
luma_hps[48x64] 2.21x 303515.28 670147.75
luma_hpp[64x16] 2.60x 94806.54 246188.88
luma_hps[64x16] 2.19x 131033.46 287519.62
luma_hpp[16x64] 2.72x 95223.20 259393.41
luma_hps[16x64] 2.35x 101401.92 238135.25
diff -r e9df93f38066 -r 6108fbda1be6 source/common/x86/asm-primitives.cpp
--- a/source/common/x86/asm-primitives.cpp Tue Apr 28 20:24:06 2015 +0800
+++ b/source/common/x86/asm-primitives.cpp Tue Apr 28 15:26:07 2015 -0700
@@ -1343,6 +1343,11 @@
CHROMA_422_VSP_FILTERS(_sse2);
CHROMA_444_VSP_FILTERS(_sse2);
+ ALL_LUMA_PU(luma_hpp, interp_8tap_horiz_pp, sse2);
+ p.pu[LUMA_4x4].luma_hpp = x265_interp_8tap_horiz_pp_4x4_sse2;
+ ALL_LUMA_PU(luma_hps, interp_8tap_horiz_ps, sse2);
+ p.pu[LUMA_4x4].luma_hps = x265_interp_8tap_horiz_ps_4x4_sse2;
+
//p.frameInitLowres = x265_frame_init_lowres_core_mmx2;
p.frameInitLowres = x265_frame_init_lowres_core_sse2;
diff -r e9df93f38066 -r 6108fbda1be6 source/common/x86/ipfilter8.asm
--- a/source/common/x86/ipfilter8.asm Tue Apr 28 20:24:06 2015 +0800
+++ b/source/common/x86/ipfilter8.asm Tue Apr 28 15:26:07 2015 -0700
@@ -151,6 +151,12 @@
db -1, 4, -11, 40, 40, -11, 4, -1
db 0, 1, -5, 17, 58, -10, 4, -1
+const tabw_LumaCoeff, dw 0, 0, 0, 64, 0, 0, 0, 0
+ dw -1, 4, -10, 58, 17, -5, 1, 0
+ dw -1, 4, -11, 40, 40, -11, 4, -1
+ dw 0, 1, -5, 17, 58, -10, 4, -1
+
+
const tab_LumaCoeffV, times 4 dw 0, 0
times 4 dw 0, 64
times 4 dw 0, 0
@@ -807,6 +813,229 @@
IPFILTER_CHROMA_W_sse3 48, 64
IPFILTER_CHROMA_W_sse3 64, 16
+%macro FILTER_H8_W8_sse2 0
+ movh m1, [r0 + x - 3]
+ movh m4, [r0 + x - 2]
+ punpcklbw m1, m6
+ punpcklbw m4, m6
+ movh m5, [r0 + x - 1]
+ movh m0, [r0 + x]
+ punpcklbw m5, m6
+ punpcklbw m0, m6
+ pmaddwd m1, m3
+ pmaddwd m4, m3
+ pmaddwd m5, m3
+ pmaddwd m0, m3
+ packssdw m1, m4
+ packssdw m5, m0
+ pshuflw m4, m1, q2301
+ pshufhw m4, m4, q2301
+ pshuflw m0, m5, q2301
+ pshufhw m0, m0, q2301
+ paddw m1, m4
+ paddw m5, m0
+ psrldq m1, 2
+ psrldq m5, 2
+ pshufd m1, m1, q3120
+ pshufd m5, m5, q3120
+ punpcklqdq m1, m5
+ movh m7, [r0 + x + 1]
+ movh m4, [r0 + x + 2]
+ punpcklbw m7, m6
+ punpcklbw m4, m6
+ movh m5, [r0 + x + 3]
+ movh m0, [r0 + x + 4]
+ punpcklbw m5, m6
+ punpcklbw m0, m6
+ pmaddwd m7, m3
+ pmaddwd m4, m3
+ pmaddwd m5, m3
+ pmaddwd m0, m3
+ packssdw m7, m4
+ packssdw m5, m0
+ pshuflw m4, m7, q2301
+ pshufhw m4, m4, q2301
+ pshuflw m0, m5, q2301
+ pshufhw m0, m0, q2301
+ paddw m7, m4
+ paddw m5, m0
+ psrldq m7, 2
+ psrldq m5, 2
+ pshufd m7, m7, q3120
+ pshufd m5, m5, q3120
+ punpcklqdq m7, m5
+ pshuflw m4, m1, q2301
+ pshufhw m4, m4, q2301
+ pshuflw m0, m7, q2301
+ pshufhw m0, m0, q2301
+ paddw m1, m4
+ paddw m7, m0
+ psrldq m1, 2
+ psrldq m7, 2
+ pshufd m1, m1, q3120
+ pshufd m7, m7, q3120
+ punpcklqdq m1, m7
+%endmacro
+
+%macro FILTER_H8_W4_sse2 0
+ movh m1, [r0 + x - 3]
+ movh m0, [r0 + x - 2]
+ punpcklbw m1, m6
+ punpcklbw m0, m6
+ movh m4, [r0 + x - 1]
+ movh m5, [r0 + x]
+ punpcklbw m4, m6
+ punpcklbw m5, m6
+ pmaddwd m1, m3
+ pmaddwd m0, m3
+ pmaddwd m4, m3
+ pmaddwd m5, m3
+ packssdw m1, m0
+ packssdw m4, m5
+ pshuflw m0, m1, q2301
+ pshufhw m0, m0, q2301
+ pshuflw m5, m4, q2301
+ pshufhw m5, m5, q2301
+ paddw m1, m0
+ paddw m4, m5
+ psrldq m1, 2
+ psrldq m4, 2
+ pshufd m1, m1, q3120
+ pshufd m4, m4, q3120
+ punpcklqdq m1, m4
+ pshuflw m0, m1, q2301
+ pshufhw m0, m0, q2301
+ paddw m1, m0
+ psrldq m1, 2
+ pshufd m1, m1, q3120
+%endmacro
+
+;----------------------------------------------------------------------------------------------------------------------------
+; void interp_8tap_horiz_%3_%1x%2(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx, int isRowExt)
+;----------------------------------------------------------------------------------------------------------------------------
+%macro IPFILTER_LUMA_sse2 3
+INIT_XMM sse2
+cglobal interp_8tap_horiz_%3_%1x%2, 4,7,8
+
+ mov r4d, r4m
+ add r4d, r4d
+ pxor m6, m6
+%ifdef PIC
+ lea r6, [tabw_LumaCoeff]
+ movu m3, [r6 + r4 * 8]
+%else
+ movu m3, [tabw_LumaCoeff + r4 * 8]
+%endif
+
+ mov r4d, %2
+%ifidn %3, pp
+ mova m2, [pw_32]
+%else
+ mova m2, [pw_2000]
+ add r3d, r3d
+ cmp r5m, byte 0
+ je .loopH
+ lea r6, [r1 + 2 * r1]
+ sub r0d, r6d
+ add r4d, 7
+%endif
+
+.loopH:
+%assign x 0
+%rep %1 / 8
+ FILTER_H8_W8_sse2
+ %ifidn %3, pp
+ paddw m1, m2
+ psraw m1, 6
+ packuswb m1, m1
+ movh [r2 + x], m1
+ %else
+ psubw m1, m2
+ movu [r2 + 2 * x], m1
+ %endif
+%assign x x+8
+%endrep
+
+%rep (%1 % 8) / 4
+ FILTER_H8_W4_sse2
+ %ifidn %3, pp
+ paddw m1, m2
+ psraw m1, 6
+ packuswb m1, m1
+ movd [r2 + x], m1
+ %else
+ psubw m1, m2
+ movh [r2 + 2 * x], m1
+ %endif
+%endrep
+
+ add r0d, r1d
+ add r2d, r3d
+
+ dec r4d
+ jnz .loopH
+ RET
+
+%endmacro
+
+;--------------------------------------------------------------------------------------------------------------
+; void interp_8tap_horiz_pp_%1x%2(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx)
+;--------------------------------------------------------------------------------------------------------------
+ IPFILTER_LUMA_sse2 4, 4, pp
+ IPFILTER_LUMA_sse2 4, 8, pp
+ IPFILTER_LUMA_sse2 8, 4, pp
+ IPFILTER_LUMA_sse2 8, 8, pp
+ IPFILTER_LUMA_sse2 16, 16, pp
+ IPFILTER_LUMA_sse2 16, 8, pp
+ IPFILTER_LUMA_sse2 8, 16, pp
+ IPFILTER_LUMA_sse2 16, 12, pp
+ IPFILTER_LUMA_sse2 12, 16, pp
+ IPFILTER_LUMA_sse2 16, 4, pp
+ IPFILTER_LUMA_sse2 4, 16, pp
+ IPFILTER_LUMA_sse2 32, 32, pp
+ IPFILTER_LUMA_sse2 32, 16, pp
+ IPFILTER_LUMA_sse2 16, 32, pp
+ IPFILTER_LUMA_sse2 32, 24, pp
+ IPFILTER_LUMA_sse2 24, 32, pp
+ IPFILTER_LUMA_sse2 32, 8, pp
+ IPFILTER_LUMA_sse2 8, 32, pp
+ IPFILTER_LUMA_sse2 64, 64, pp
+ IPFILTER_LUMA_sse2 64, 32, pp
+ IPFILTER_LUMA_sse2 32, 64, pp
+ IPFILTER_LUMA_sse2 64, 48, pp
+ IPFILTER_LUMA_sse2 48, 64, pp
+ IPFILTER_LUMA_sse2 64, 16, pp
+ IPFILTER_LUMA_sse2 16, 64, pp
+
+;----------------------------------------------------------------------------------------------------------------------------
+; void interp_8tap_horiz_ps_%1x%2(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx, int isRowExt)
+;----------------------------------------------------------------------------------------------------------------------------
+ IPFILTER_LUMA_sse2 4, 4, ps
+ IPFILTER_LUMA_sse2 8, 8, ps
+ IPFILTER_LUMA_sse2 8, 4, ps
+ IPFILTER_LUMA_sse2 4, 8, ps
+ IPFILTER_LUMA_sse2 16, 16, ps
+ IPFILTER_LUMA_sse2 16, 8, ps
+ IPFILTER_LUMA_sse2 8, 16, ps
+ IPFILTER_LUMA_sse2 16, 12, ps
+ IPFILTER_LUMA_sse2 12, 16, ps
+ IPFILTER_LUMA_sse2 16, 4, ps
+ IPFILTER_LUMA_sse2 4, 16, ps
+ IPFILTER_LUMA_sse2 32, 32, ps
+ IPFILTER_LUMA_sse2 32, 16, ps
+ IPFILTER_LUMA_sse2 16, 32, ps
+ IPFILTER_LUMA_sse2 32, 24, ps
+ IPFILTER_LUMA_sse2 24, 32, ps
+ IPFILTER_LUMA_sse2 32, 8, ps
+ IPFILTER_LUMA_sse2 8, 32, ps
+ IPFILTER_LUMA_sse2 64, 64, ps
+ IPFILTER_LUMA_sse2 64, 32, ps
+ IPFILTER_LUMA_sse2 32, 64, ps
+ IPFILTER_LUMA_sse2 64, 48, ps
+ IPFILTER_LUMA_sse2 48, 64, ps
+ IPFILTER_LUMA_sse2 64, 16, ps
+ IPFILTER_LUMA_sse2 16, 64, ps
+
;-----------------------------------------------------------------------------
; void interp_4tap_horiz_pp_2x4(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx)
;-----------------------------------------------------------------------------
diff -r e9df93f38066 -r 6108fbda1be6 source/common/x86/ipfilter8.h
--- a/source/common/x86/ipfilter8.h Tue Apr 28 20:24:06 2015 +0800
+++ b/source/common/x86/ipfilter8.h Tue Apr 28 15:26:07 2015 -0700
@@ -850,6 +850,56 @@
void x265_interp_4tap_horiz_pp_64x32_sse3(const pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
void x265_interp_4tap_horiz_pp_64x48_sse3(const pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
void x265_interp_4tap_horiz_pp_64x64_sse3(const pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_4x4_sse2(const pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_4x8_sse2(const pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_4x16_sse2(const pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_8x4_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_8x8_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_8x16_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_8x32_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_12x16_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_16x4_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_16x8_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_16x12_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_16x16_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_16x32_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_16x64_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_24x32_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_32x8_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_32x16_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_32x24_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_32x32_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_32x64_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_48x64_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_64x16_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_64x32_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_64x48_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_pp_64x64_sse2(const pixel* src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx);
+void x265_interp_8tap_horiz_ps_4x4_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_4x8_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_4x16_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_8x4_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_8x8_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_8x16_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_8x32_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_12x16_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_16x4_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_16x8_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_16x12_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_16x16_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_16x32_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_16x64_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_24x32_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_32x8_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_32x16_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_32x24_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_32x32_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_32x64_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_48x64_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_64x16_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_64x32_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_64x48_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
+void x265_interp_8tap_horiz_ps_64x64_sse2(const pixel* src, intptr_t srcStride, int16_t* dst, intptr_t dstStride, int coeffIdx, int isRowExt);
#undef LUMA_FILTERS
#undef LUMA_SP_FILTERS
#undef LUMA_SS_FILTERS
More information about the x265-devel
mailing list