[x265] [PATCH] asm: interp_8tap_horiz pp and ps sse2

[chen]

在 2015-04-29 09:30:36，dave <dtyx265 at gmail.com> 写道：

On 04/28/2015 06:13 PM, chen wrote:



在 2015-04-29 07:49:46，dave <dtyx265 at gmail.com> 写道：

On 04/28/2015 03:32 PM, chen wrote:

Most part are fine now, just modify about r5, see below comment


At 2015-04-29 06:27:27,dtyx265 at gmail.com wrote:
># HG changeset patch
># User David T Yuen dtyx265 at gmail.com>
># Date 1430259967 25200
># Node ID 6108fbda1be654a481a78f7ef593518033919674
># Parent  e9df93f380664932e7d6c7e85b2cae16cd5e1dcd
>asm: interp_8tap_horiz pp and ps sse2
>
>This replaces c code and covers
>
+;----------------------------------------------------------------------------------------------------------------------------
>+; void interp_8tap_horiz_%3_%1x%2(pixel *src, intptr_t srcStride, pixel *dst, intptr_t dstStride, int coeffIdx, int isRowExt)
>+;----------------------------------------------------------------------------------------------------------------------------
>+%macro IPFILTER_LUMA_sse2 3
>+INIT_XMM sse2
>+cglobal interp_8tap_horiz_%3_%1x%2, 4,7,8
>+
>+    mov       r4d, r4m
>+    add       r4d, r4d
>+    pxor      m6, m6
>+%ifdef PIC
>+    lea       r6, [tabw_LumaCoeff]
>+    movu      m3, [r6 + r4 * 8]
>+%else
>+    movu      m3, [tabw_LumaCoeff + r4 * 8]
>+%endif
>+
>+    mov       r4d, %2
>+%ifidn %3, pp
>+    mova      m2, [pw_32]
>+%else
>+    mova      m2, [pw_2000]
>+    add       r3d, r3d
>+    cmp       r5m, byte 0
if we move above 2 lines to up, we can reduce r6 and reuse r5.

I am not sure if this can be done.  r4 is used to to set m3 then it is reused and modified depending on r5 and r5 can't be used for something else before it's cmp'ed.

'mov, lea' didn't affect eflags register
but r4 is needed for the lea instruction and r4 is later reused and modified depending on r5.  Only one of them can be reused, not both, so r6 is needed.


How about below code:

    pxor      m6, m6
    add       r3d, r3d

    cmp       r5m, byte 0
%ifdef PIC
    lea       r6, [tabw_LumaCoeff]
    movu      m3, [r6 + r4 * 8]
%else
    movu      m3, [tabw_LumaCoeff + r4 * 8]
%endif
    mov       r4d, %2
%ifidn %3, pp
    mova      m2, [pw_32]
%else
    mova      m2, [pw_2000]
   jnz ...
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.videolan.org/pipermail/x265-devel/attachments/20150429/4cc05122/attachment.html>