<div style="line-height:1.7;color:#000000;font-size:14px;font-family:arial"><div> </div><pre><br>At 2015-05-18 10:48:53,dtyx265@gmail.com wrote:
># HG changeset patch
># User David T Yuen <dtyx265@gmail.com>
># Date 1431911615 25200
># Node ID 72bba6b9e99739599d040000be62c7e02a3c8faa
># Parent 465fb4340a241e501b53a6241f5ae81c29ba073a
>asm: interp_4tap_vert_ps_4x2 sse2
>
>Converted vert_pp_4x2 primitive to macro that also creates ps. This replaces c code for ps with minimal impact on pp.
>
>64-bit
>
>/test/TestBench --testbench interp | grep vp | grep " 4x2"
>chroma_vpp[ 4x2] 2.13x 524.99 1117.40
>chroma_vps[ 4x2] 1.87x 457.54 854.98
>
>32-bit
>
>./test/TestBench --testbench interp | grep vp | grep " 4x2"
>chroma_vpp[ 4x2] 2.34x 592.50 1387.29
>chroma_vps[ 4x2] 2.41x 542.48 1304.96
>
>diff -r 465fb4340a24 -r 72bba6b9e997 source/common/x86/asm-primitives.cpp
>--- a/source/common/x86/asm-primitives.cpp Sun May 17 18:03:29 2015 -0700
>+++ b/source/common/x86/asm-primitives.cpp Sun May 17 18:13:35 2015 -0700
>@@ -1450,6 +1450,7 @@
> p.chroma[X265_CSP_I444].pu[LUMA_4x16].filter_vpp = x265_interp_4tap_vert_pp_4x16_sse2;
> p.chroma[X265_CSP_I420].pu[CHROMA_420_2x4].filter_vps = x265_interp_4tap_vert_ps_2x4_sse2;
> p.chroma[X265_CSP_I420].pu[CHROMA_420_2x8].filter_vps = x265_interp_4tap_vert_ps_2x8_sse2;
>+ p.chroma[X265_CSP_I420].pu[CHROMA_420_4x2].filter_vps = x265_interp_4tap_vert_ps_4x2_sse2;
> p.chroma[X265_CSP_I422].pu[CHROMA_422_2x16].filter_vps = x265_interp_4tap_vert_ps_2x16_sse2;
> #if X86_64
> p.chroma[X265_CSP_I420].pu[CHROMA_420_6x8].filter_vpp = x265_interp_4tap_vert_pp_6x8_sse2;
>diff -r 465fb4340a24 -r 72bba6b9e997 source/common/x86/ipfilter8.asm
>--- a/source/common/x86/ipfilter8.asm Sun May 17 18:03:29 2015 -0700
>+++ b/source/common/x86/ipfilter8.asm Sun May 17 18:13:35 2015 -0700
>@@ -1025,11 +1025,11 @@
> FILTER_V4_W2_H4_sse2 ps, 16
>
> ;-----------------------------------------------------------------------------
>-; void interp_4tap_vert_pp_4x2(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx)
>-;-----------------------------------------------------------------------------
>+; void interp_4tap_vert_%1_4x2(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx)
>+;-----------------------------------------------------------------------------
>+%macro FILTER_V2_W4_H4_sse2 1
> INIT_XMM sse2
>-cglobal interp_4tap_vert_pp_4x2, 4, 6, 8
>-
>+cglobal interp_4tap_vert_%1_4x2, 4, 6, 8
> mov r4d, r4m
> sub r0, r1
> pxor m7, m7
>@@ -1078,6 +1078,8 @@
> pshuflw m5, m3, q2301
> pshufhw m5, m5, q2301
> paddw m3, m5
>+
>+%ifidn %1, pp
> psrld m2, 16
> psrld m3, 16
> packssdw m2, m3
>@@ -1089,7 +1091,24 @@
> movd [r2], m2
> psrldq m2, 4
> movd [r2 + r3], m2
>- RET
>+%elifidn %1, ps
>+ psrldq m2, 2
>+ psrldq m3, 2
>+ pshufd m2, m2, q3120
>+ pshufd m3, m3, q3120
>+ punpcklqdq m2, m3
>+
>+ add r3d, r3d
</pre><pre>unneed</pre><pre>>+ psubw m2, [pw_2000]
>+ movh [r2], m2
>+ movhps [r2 + r3], m2
</pre><pre>MOVHPS because above PUNPCKLQDQ, we didn't need combo and split, same count of instructions and slower on Intel CPU</pre><pre>>+%endif
>+ RET
>+
>+%endmacro
>+
</pre></div>