[x265] [PATCH Review only] asm: code for transpose_32x32 routine

[chen]

At 2013-11-20 20:21:41,murugan at multicorewareinc.com wrote:
># HG changeset patch
># User Murugan Vairavel <murugan at multicorewareinc.com>
># Date 1384950080 -19800
>#      Wed Nov 20 17:51:20 2013 +0530
># Node ID 2ea432a36a03a7ee7e0f788181ff35ac8db20337
># Parent  a059cd5c08133486fb0499aeb77ac34677759d9b
>asm: code for transpose_32x32 routine
>
>diff -r a059cd5c0813 -r 2ea432a36a03 source/common/x86/asm-primitives.cpp
>--- a/source/common/x86/asm-primitives.cpp Wed Nov 20 15:13:48 2013 +0530
>+++ b/source/common/x86/asm-primitives.cpp Wed Nov 20 17:51:20 2013 +0530
>@@ -547,6 +547,7 @@
>         p.transpose[BLOCK_4x4] = x265_transpose4_sse2;
>         p.transpose[BLOCK_8x8] = x265_transpose8_sse2;
>         p.transpose[BLOCK_16x16] = x265_transpose16_sse2;
>+        p.transpose[BLOCK_32x32] = x265_transpose32_sse2;
>     }
>     if (cpuMask & X265_CPU_SSSE3)
>     {
>diff -r a059cd5c0813 -r 2ea432a36a03 source/common/x86/pixel-a.asm
>--- a/source/common/x86/pixel-a.asm Wed Nov 20 15:13:48 2013 +0530
>+++ b/source/common/x86/pixel-a.asm Wed Nov 20 17:51:20 2013 +0530
>@@ -7079,6 +7079,48 @@
> 
>     RET
> 
>+%macro TRANSPOSE_16x16 1
>+    TRANSPOSE_8x8 %1
>+    lea    r1,    [r1 + 2 * r2]
>+    lea    r0,    [r5 + 8]
>+    TRANSPOSE_8x8 %1
>+    lea    r1,    [r6 + 8]
>+    lea    r0,    [r5 + 8 * %1]
>+    TRANSPOSE_8x8 %1
>+    lea    r1,    [r1 + 2 * r2]
>+    lea    r0,    [r5 + 8 * %1 + 8]
>+    TRANSPOSE_8x8 %1
>+%endmacro
>+
>+;-----------------------------------------------------------------
>+; void transpose_32x32(pixel *dst, pixel *src, intptr_t stride)
>+;-----------------------------------------------------------------
>+INIT_XMM sse2
>+cglobal transpose32, 3, 7, 8, dest, src, stride
>+
>+    mov    r3,    r0
>+    mov    r4,    r1
>+    mov    r5,    r0
>+    mov    r6,    r1
>+    TRANSPOSE_16x16 32
>+    lea    r1,    [r1 - 8 + 2 * r2]
>+    lea    r0,    [r3 + 16]
>+    mov    r5,    r0
>+    mov    r6,    r1
>+    TRANSPOSE_16x16 32
>+    lea    r1,    [r4 + 16]
>+    lea    r0,    [r3 + 16 * 32]
>+    mov    r5,    r0
>+    mov    r6,    r1
>+    TRANSPOSE_16x16 32
>+    lea    r1,    [r1 - 8 + 2 * r2]
>+    lea    r0,    [r3 + 16 * 32 + 16]
>+    mov    r5,    r0
>+    mov    r6,    r1
>+    TRANSPOSE_16x16 32
>+
>+    RET

The code is right, but all of macro extended, so the obj is large and lower cache performance
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20131120/5b6b9eae/attachment-0001.html>