[x265] [PATCH] Fixed 32 bit bug in intrapred dc4 sse2

[dave]

FYI,

I forgot to comment in the commit message:

I kept the original code for 64 bit because while using r2b works in 64 
bits, using it severely hurt  performance to the point that it was well 
below c code.

This is probably due to how things like instruction order, length and 
layout in memory can affect performance(see agner docs).

This probably leaves some room for performance improvements since most 
X265 assembler is based primarily on algorithms implementing x265 
functionality but not all aspects of processor function. While writing 
optimized assembler for every processor is unrealistic, the assembler of 
each simd level could be optimized for the latest processor that 
supports it.(i.e. the sse4 assembler could be optimized to support the 
latest processor that support only up to sse4).

One downside is these types of optimizations are more likely to generate 
code that looks like the jumbled code generated by a compiler and thus 
be less easy to read, understand and maintain. Of course more comments 
can help here.

Is x265 interested in such optimizations?

On 02/25/2015 08:31 PM, Deepthi Nandakumar wrote:
> Thanks, pushed.
>
> On Thu, Feb 26, 2015 at 8:03 AM, <dtyx265 at gmail.com 
> <mailto:dtyx265 at gmail.com>> wrote:
>
>     # HG changeset patch
>     # User David T Yuen <dtyx265 at gmail.com <mailto:dtyx265 at gmail.com>>
>     # Date 1424917924 28800
>     # Node ID 13346cb90bff040492f0688226f44182bb6b97d8
>     # Parent  74c716607444c77b9d5ea1dce5b99c875f0b20fe
>     Fixed 32 bit bug in intrapred dc4 sse2
>
>     Changed register written from to one that supports low byte access
>     in 32 bit
>     Also moved pw_257 constant to const-a.asm
>
>     diff -r 74c716607444 -r 13346cb90bff source/common/x86/const-a.asm
>     --- a/source/common/x86/const-a.asm     Tue Feb 24 13:39:16 2015 +0530
>     +++ b/source/common/x86/const-a.asm     Wed Feb 25 18:32:04 2015 -0800
>     @@ -37,6 +37,7 @@
>      const pw_32,       times 16 dw 32
>      const pw_128,      times 16 dw 128
>      const pw_256,      times 16 dw 256
>     +const pw_257,      times 16 dw 257
>      const pw_512,      times 16 dw 512
>      const pw_1023,     times 8  dw 1023
>      const pw_1024,     times 16 dw 1024
>     diff -r 74c716607444 -r 13346cb90bff source/common/x86/intrapred8.asm
>     --- a/source/common/x86/intrapred8.asm  Tue Feb 24 13:39:16 2015 +0530
>     +++ b/source/common/x86/intrapred8.asm  Wed Feb 25 18:32:04 2015 -0800
>     @@ -65,8 +65,6 @@
>      pw_planar32_L:        dw 31, 30, 29, 28, 27, 26, 25, 24
>      pw_planar32_H:        dw 23, 22, 21, 20, 19, 18, 17, 16
>
>     -pw_257:         times 8 dw 257
>     -
>      const ang_table
>      %assign x 0
>      %rep 32
>     @@ -80,6 +78,7 @@
>      cextern pw_8
>      cextern pw_16
>      cextern pw_32
>     +cextern pw_257
>      cextern pw_1024
>      cextern pb_unpackbd1
>      cextern multiL
>     @@ -144,12 +143,21 @@
>          paddw       m2, m1
>          psraw       m2, 2
>          packuswb    m2, m2
>     +%if ARCH_X86_64
>          movq        r4, m2
>          mov         [r0], r4b
>          shr         r4, 8
>          mov         [r0 + r1], r4b
>          shr         r4, 8
>          mov         [r0 + r1 * 2], r4b
>     +%else
>     +    movd        r2d, m2
>     +    mov         [r0], r2b
>     +    shr         r2, 8
>     +    mov         [r0 + r1], r2b
>     +    shr         r2, 8
>     +    mov         [r0 + r1 * 2], r2b
>     +%endif
>      .end:
>          RET
>
>     _______________________________________________
>     x265-devel mailing list
>     x265-devel at videolan.org <mailto: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/20150226/25d5fb16/attachment-0001.html>