[x265] [PATCH 02 of 12] asm: interp_4tap_vert_ps_4x2 sse2

chen chenm003 at 163.com
Mon May 18 19:11:06 CEST 2015


 
At 2015-05-19 01:02:00,dave <dtyx265 at gmail.com> wrote:

On 05/18/2015 09:42 AM, chen wrote:


 
At 2015-05-19 00:36:01,dave <dtyx265 at gmail.com> wrote:

On 05/18/2015 07:50 AM, chen wrote:

 
At 2015-05-18 10:48:53,dtyx265 at gmail.com wrote:
># HG changeset patch
># User David T Yuen <dtyx265 at 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

unneed

add r3d, r3d unneeded?

[MC] we may replace below to [r2 + r3 * 2], it is one time code


I see, ok.


>+    psubw       m2,        [pw_2000]
>+    movh        [r2],      m2
>+    movhps      [r2 + r3], m2

MOVHPS because above PUNPCKLQDQ, we didn't need combo and split, same count of instructions and slower on Intel CPU

Interesting, I initially tried this and it performs slower on the bench test on amd, though much more so for 64-bit than 32-bit.  I will resubmit.

[MC] yes, it is faster on AMD CPU, on Intel, these instructions choke Port5, the PADD execute on Port1.  I often choice faster instrction for Intel because my PC use Intel CPU


... and of course, while I don't follow it very closely, I do believe intel still dominates the market.

Do we have any way to determine what the target build is?  Something like..

%if INTEL
    optimal intel code
%elif AMD
    optimal amd code
%endif
   

Since I difficult to find AMD documents, I suggest all of code optimize for Intel (not only your patches but also all of other code)

in future, we can review and optimize all of code for AMD


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20150519/75488026/attachment-0001.html>


More information about the x265-devel mailing list