[x265] [PATCH 3 of 8] asm:intra_pred_ang4_4_sse2 improved ~2% 647.49 -> 634.98 with nits and tweaks

[chen]

At 2015-03-29 06:42:32,dave <dtyx265 at gmail.com> wrote:

On 03/28/2015 03:20 PM, chen wrote:

 
At 2015-03-29 05:35:21,dtyx265 at gmail.com wrote:
># HG changeset patch
># User David T Yuen <dtyx265 at gmail.com>
># Date 1427576216 25200
># Node ID 0a75e3d50518e73f5a199d7519f800a9ff1c2e2c
># Parent  6595ba5f989fdd521e268911ddf027665a610e25
>asm:intra_pred_ang4_4_sse2 improved ~2% 647.49 -> 634.98 with nits and tweaks
>
>Corrected parameter count
>Changed r3 and r4 to r3d and r4d
>tweaked unpacking for performance
>
>diff -r 6595ba5f989f -r 0a75e3d50518 source/common/x86/intrapred8.asm
>--- a/source/common/x86/intrapred8.asm	Sat Mar 28 13:40:20 2015 -0700
>+++ b/source/common/x86/intrapred8.asm	Sat Mar 28 13:56:56 2015 -0700
>@@ -1413,23 +1413,22 @@
>     movd        [r0 + r1], m0
>     RET
> 
>-cglobal intra_pred_ang4_4, 3,5,8
>-    xor         r4, r4
>-    inc         r4
>+cglobal intra_pred_ang4_4, 4,5,8

load parameter into r3, and overwrite in below?
>+    xor         r4d, r4d
>+    inc         r4d
>     cmp         r3m, byte 32

After cmp sets flags, r3 is free to use.

[MC] this is problem, on x86, r3m is memory, so we didn't need load it into register, on x64, it have been in register. so auto load unneed




>-    mov         r3, 9
>-    cmove       r3, r4
>+    mov         r3d, 9
>+    cmove       r3d, r4d
This is to allow reused for transposed mode 32.

> 
>     movh        m0, [r2 + r3]    ; [8 7 6 5 4 3 2 1]
>+    punpcklbw   m0, m0
>+    psrldq      m0, 1
>+    mova        m2, m0
>+    psrldq      m2, 2           ; [x x x x x x x x 6 5 5 4 4 3 3 2]
>     mova        m1, m0
>-    psrldq      m1, 1           ; [x 8 7 6 5 4 3 2]
>-    punpcklbw   m0, m1          ; [x 8 8 7 7 6 6 5 5 4 4 3 3 2 2 1]
>-    mova        m1, m0
>-    psrldq      m1, 2           ; [x x x x x x x x 6 5 5 4 4 3 3 2]
>-    mova        m3, m0
>-    psrldq      m3, 4           ; [x x x x x x x x 7 6 6 5 5 4 4 3]
>-    punpcklqdq  m0, m1
>-    punpcklqdq  m2, m1, m3
>+    psrldq      m1, 4           ; [x x x x x x x x 7 6 6 5 5 4 4 3]
>+    punpcklqdq  m0, m2
>+    punpcklqdq  m2, m1
> 
>     lea         r3, [pw_ang_table + 18 * 16]
>     mova        m4, [r3 +  3 * 16]  ; [21]
>_______________________________________________
>x265-devel mailing list
>x265-devel at videolan.org
>https://mailman.videolan.org/listinfo/x265-devel



_______________________________________________
x265-devel mailing list
x265-devel at videolan.orghttps://mailman.videolan.org/listinfo/x265-devel

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20150329/da03c085/attachment.html>