[x265] [PATCH] blockcopy_pp_6x8 sse2 asm code optimization

[Praveen Tiwari]

maskmovdqu, [i420] copy_pp[  6x8]  0.79x    670.04          527.63
even slower than C code as well as high maintenance code, need to maintain
separate code for separate platform as we have mapped rdi differently.


Regards,
Praveen

On Tue, Feb 3, 2015 at 2:57 PM, chen <chenm003 at 163.com> wrote:

> How about maskmovdqu
>
> At 2015-02-03 17:00:54,praveen at multicorewareinc.com wrote:
>
> ># HG changeset patch
> ># User Praveen Tiwari
> ># Date 1422954042 -19800
> ># Node ID d212ce9fa3705b9e7d4d23f14412bd28fe3bbfde
> ># Parent  059892f65db3e4c70017241ea847717e11be0124
> >blockcopy_pp_6x8 sse2 asm code optimization
> >
> >improved, 248.67c -> 212.56c
> >
> >diff -r 059892f65db3 -r d212ce9fa370 source/common/x86/blockcopy8.asm
> >--- a/source/common/x86/blockcopy8.asm Tue Feb 03 11:58:18 2015 +0530
> >+++ b/source/common/x86/blockcopy8.asm Tue Feb 03 14:30:42 2015 +0530
> >@@ -224,65 +224,51 @@
>
> > ; void blockcopy_pp_6x8(pixel* dst, intptr_t dstStride, const pixel* src, intptr_t srcStride)
>
> > ;-----------------------------------------------------------------------------
> > INIT_XMM sse2
> >-cglobal blockcopy_pp_6x8, 4, 7, 8
> >-
> >-    movd     m0,     [r2]
> >-    movd     m1,     [r2 + r3]
> >-    movd     m2,     [r2 + 2 * r3]
> >-    lea      r5,     [r2 + 2 * r3]
> >-    movd     m3,     [r5 + r3]
> >-
> >-    movd     m4,     [r5 + 2 * r3]
> >-    lea      r5,     [r5 + 2 * r3]
> >-    movd     m5,     [r5 + r3]
> >-    movd     m6,     [r5 + 2 * r3]
> >-    lea      r5,     [r5 + 2 * r3]
> >-    movd     m7,     [r5 + r3]
> >-
> >-    movd     [r0],                m0
> >-    movd     [r0 + r1],           m1
> >-    movd     [r0 + 2 * r1],       m2
> >-    lea      r6,                  [r0 + 2 * r1]
> >-    movd     [r6 + r1],           m3
> >-
> >-    movd     [r6 + 2 * r1],        m4
> >-    lea      r6,                   [r6 + 2 * r1]
> >-    movd     [r6 + r1],            m5
> >-    movd     [r6 + 2 * r1],        m6
> >-    lea      r6,                   [r6 + 2 * r1]
> >-    movd     [r6 + r1],            m7
> >-
> >-    mov     r4w,     [r2 + 4]
> >-    mov     r5w,     [r2 + r3 + 4]
> >-    mov     r6w,     [r2 + 2 * r3 + 4]
> >-
> >-    mov     [r0 + 4],            r4w
> >-    mov     [r0 + r1 + 4],       r5w
> >-    mov     [r0 + 2 * r1 + 4],   r6w
> >-
> >-    lea     r0,              [r0 + 2 * r1]
> >-    lea     r2,              [r2 + 2 * r3]
> >-
> >-    mov     r4w,             [r2 + r3 + 4]
> >-    mov     r5w,             [r2 + 2 * r3 + 4]
> >-
> >-    mov     [r0 + r1 + 4],       r4w
> >-    mov     [r0 + 2 * r1 + 4],   r5w
> >-
> >-    lea     r0,              [r0 + 2 * r1]
> >-    lea     r2,              [r2 + 2 * r3]
> >-
> >-    mov     r4w,             [r2 + r3 + 4]
> >-    mov     r5w,             [r2 + 2 * r3 + 4]
> >-
> >-    mov     [r0 + r1 + 4],       r4w
> >-    mov     [r0 + 2 * r1 + 4],   r5w
> >-
> >-    lea     r0,              [r0 + 2 * r1]
> >-    lea     r2,              [r2 + 2 * r3]
> >-
> >-    mov     r4w,             [r2 + r3 + 4]
> >-    mov     [r0 + r1 + 4],       r4w
> >+cglobal blockcopy_pp_6x8, 4, 7, 3
> >+
> >+    movd     m0,  [r2]
> >+    mov      r4w, [r2 + 4]
> >+    movd     m1,  [r2 + r3]
> >+    mov      r5w, [r2 + r3 + 4]
> >+    movd     m2,  [r2 + 2 * r3]
> >+    mov      r6w, [r2 + 2 * r3 + 4]
> >+
> >+    movd     [r0],              m0
> >+    mov      [r0 + 4],          r4w
> >+    movd     [r0 + r1],         m1
> >+    mov      [r0 + r1 + 4],     r5w
> >+    movd     [r0 + 2 * r1],     m2
> >+    mov      [r0 + 2 * r1 + 4], r6w
> >+
> >+    lea      r2,  [r2 + 2 * r3]
> >+    movd     m0,  [r2 + r3]
> >+    mov      r4w, [r2 + r3 + 4]
> >+    movd     m1,  [r2 + 2 * r3]
> >+    mov      r5w, [r2 + 2 * r3 + 4]
> >+    lea      r2,  [r2 + 2 * r3]
> >+    movd     m2,  [r2 + r3]
> >+    mov      r6w, [r2 + r3 + 4]
> >+
> >+    lea      r0,                [r0 + 2 * r1]
> >+    movd     [r0 + r1],         m0
> >+    mov      [r0 + r1 + 4],     r4w
> >+    movd     [r0 + 2 * r1],     m1
> >+    mov      [r0 + 2 * r1 + 4], r5w
> >+    lea      r0,                [r0 + 2 * r1]
> >+    movd     [r0 + r1],         m2
> >+    mov      [r0 + r1 + 4],     r6w
> >+
> >+    lea      r2,                [r2 + 2 * r3]
> >+    movd     m0,                [r2]
> >+    mov      r4w,               [r2 + 4]
> >+    movd     m1,                [r2 + r3]
> >+    mov      r5w,               [r2 + r3 + 4]
> >+
> >+    lea      r0,            [r0 + 2 * r1]
> >+    movd     [r0],          m0
> >+    mov      [r0 + 4],      r4w
> >+    movd     [r0 + r1],     m1
> >+    mov      [r0 + r1 + 4], r5w
> >     RET
> >
>
> > ;-----------------------------------------------------------------------------
> >_______________________________________________
> >x265-devel mailing list
> >x265-devel at videolan.org
> >https://mailman.videolan.org/listinfo/x265-devel
>
> _______________________________________________
> x265-devel mailing list
> x265-devel at videolan.org
> https://mailman.videolan.org/listinfo/x265-devel
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20150203/0532d94f/attachment-0001.html>