[x265] [PATCH] asm: intrapred dc8 sse2

[chen]

在 2015-02-26 07:06:06，dave <dtyx265 at gmail.com> 写道：

On 02/25/2015 05:55 AM, chen wrote:

At 2015-02-25 08:15:59,dtyx265 at gmail.com wrote:
># HG changeset patch
># User David T Yuen <dtyx265 at gmail.com>
># Date 1424818021 28800
># Node ID 644d27ca0b197455393171ba705b1190f3d9b420
># Parent  1a703601f7c8b85f1e6e680caa281b9edada89ab
>asm: intrapred dc8 sse2
>
>This replaces c code for systems using ssse3 to sse2 processors
>The code is backported from intrapred dc8 sse4
>
>./test/TestBench --testbench intrapred | grep 8x8
>intra_dc_8x8[f=0]	3.86x 	 235.09   	 906.94
>intra_dc_8x8[f=1]	2.44x 	 525.19   	 1280.05
>
>Also a white space nit
>
>diff -r 1a703601f7c8 -r 644d27ca0b19 source/common/x86/asm-primitives.cpp
>--- a/source/common/x86/asm-primitives.cpp	Tue Feb 24 12:51:43 2015 -0800
>+++ b/source/common/x86/asm-primitives.cpp	Tue Feb 24 14:47:01 2015 -0800
>@@ -1074,6 +1074,7 @@
>         ALL_LUMA_TU_S(ssd_s, pixel_ssd_s_, sse2);
> 
>         p.cu[BLOCK_4x4].intra_pred[DC_IDX] = x265_intra_pred_dc4_sse2;
>+        p.cu[BLOCK_8x8].intra_pred[DC_IDX] = x265_intra_pred_dc8_sse2;
> 
>         p.cu[BLOCK_4x4].calcresidual = x265_getResidual4_sse2;
>         p.cu[BLOCK_8x8].calcresidual = x265_getResidual8_sse2;
>diff -r 1a703601f7c8 -r 644d27ca0b19 source/common/x86/intrapred.h
>--- a/source/common/x86/intrapred.h	Tue Feb 24 12:51:43 2015 -0800
>+++ b/source/common/x86/intrapred.h	Tue Feb 24 14:47:01 2015 -0800
>@@ -26,8 +26,9 @@
> #ifndef X265_INTRAPRED_H
> #define X265_INTRAPRED_H
> 
>-void x265_intra_pred_dc4_sse2 (pixel* dst, intptr_t dstStride, const pixel*srcPix, int, int filter);
>-void x265_intra_pred_dc4_sse4 (pixel* dst, intptr_t dstStride, const pixel*srcPix, int, int filter);
>+void x265_intra_pred_dc4_sse2(pixel* dst, intptr_t dstStride, const pixel*srcPix, int, int filter);
>+void x265_intra_pred_dc8_sse2(pixel* dst, intptr_t dstStride, const pixel*srcPix, int, int filter);
>+void x265_intra_pred_dc4_sse4(pixel* dst, intptr_t dstStride, const pixel*srcPix, int, int filter);
> void x265_intra_pred_dc8_sse4(pixel* dst, intptr_t dstStride, const pixel* srcPix, int, int filter);
> void x265_intra_pred_dc16_sse4(pixel* dst, intptr_t dstStride, const pixel* srcPix, int, int filter);
> void x265_intra_pred_dc32_sse4(pixel* dst, intptr_t dstStride, const pixel* srcPix, int, int filter);
>diff -r 1a703601f7c8 -r 644d27ca0b19 source/common/x86/intrapred8.asm
>--- a/source/common/x86/intrapred8.asm	Tue Feb 24 12:51:43 2015 -0800
>+++ b/source/common/x86/intrapred8.asm	Tue Feb 24 14:47:01 2015 -0800
>@@ -76,6 +76,7 @@
> 
> SECTION .text
> 
>+cextern pw_2
> cextern pw_4
> cextern pw_8
> cextern pw_16
>@@ -156,6 +157,90 @@
> ;---------------------------------------------------------------------------------------------
> ; void intra_pred_dc(pixel* dst, intptr_t dstStride, pixel *srcPix, int dirMode, int bFilter)
> ;---------------------------------------------------------------------------------------------
>+INIT_XMM sse2
>+cglobal intra_pred_dc8, 5, 7, 3

maximum register is r6
I will adjust 7 to 9
>+    pxor            m0,            m0
>+    movh            m1,            [r2 + 1]
>+    movh            m2,            [r2 + 17]
>+    punpcklqdq      m1,            m2
>+    psadbw          m1,            m0
>+    pshufd          m2,            m1, 2
>+    paddw           m1,            m2
>+
>+    paddw           m1,            [pw_8]
>+    psraw           m1,            4
>+    pmullw          m1,            [pw_257]
>+    pshuflw         m1,            m1, 0x00       ; m1 = byte [dc_val ...]
>+
>+    test            r4d,           r4d
>+
>+    ; store DC 8x8
>+    lea             r6,            [r1 + r1 * 2]
>+    lea             r7,            [r6 + r1 * 2]
>+    lea             r8,            [r6 + r1 * 4]
r7, r8 is invalid

>+    movh            [r0],          m1
>+    movh            [r0 + r1],     m1
>+    movh            [r0 + r1 * 2], m1
>+    movh            [r0 + r6],     m1
>+    movh            [r0 + r1 * 4], m1
>+    movh            [r0 + r7],     m1
>+    movh            [r0 + r6 * 2], m1
>+    movh            [r0 + r8],     m1
>+
>+    ; Do DC Filter
>+    jz              .end
>+    psrlw           m1,            8
>+    mova            m2,            [pw_2]
>+    pmullw          m2,            m1
why not PSLL?


Do you mean something like

pslldq  m1, 56
mova    m2, [pdq_2]
pmuludq m2, m1
paddq    m2, [pdq_2] ?

once pshuflw  m1, m1, 0 happens everything above 16 bits is wiped out and we are only interested in the bottom 8 bits.  Besides, I tried this and there was no difference in performance and 16 bits should be enough to handle it.  I can change it if that is how you want it.

[MC] "pmullw m2,m1" = "pmullw pw_2, m1" = "paddw m2, m1, m1" or "psllw m2, m1, 1"

 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20150226/a0837c70/attachment.html>