[FFmpeg-devel] [PATCH] avcodec/v210: add avx2 version of the line encoder

[James Darnley]

On 2016-01-14 20:21, Henrik Gramner wrote:
> On Wed, Jan 13, 2016 at 4:55 PM, James Darnley <james.darnley at gmail.com> wrote:
>> diff --git a/libavcodec/x86/v210enc.asm b/libavcodec/x86/v210enc.asm
>> index 859e2d9..a8f3d3c 100644
>> --- a/libavcodec/x86/v210enc.asm
>> +++ b/libavcodec/x86/v210enc.asm
>> -cextern pb_FE
>> -%define v210_enc_max_8 pb_FE
>> +;cextern pb_FE
>> +local_pb_FE: times 32 db 0xfe
>> +%define v210_enc_max_8 local_pb_FE
> 
> You could change ff_pb_FE to be 32-byte instead of duplicating it.

I can change that.

>> +%if cpuflag(avx2)
>> +    movu        xm1, [yq+widthq*2]
>> +    vinserti128 m1,   m1, [yq+widthq*2+12], 1
>> +%else
>>      movu    m1, [yq+2*widthq]
>> +%endif
> 
> xmN can be used unconditionally which gets rid of the %else. E.g.
> 
>     movu       xm1, [yq+widthq*2]
> %if cpuflag(avx2)
>     vinserti128 m1, m1, [yq+widthq*2+12], 1
> %endif

I can change that.  I slightly prefer to not mix register sizes like
that but it seems unavoidable with avx2.

>> +%if cpuflag(avx2)
>> +    movq         xm3, [uq+widthq]
>> +    movhps       xm3, [vq+widthq]
>> +    movq         xm7, [uq+widthq+6]
>> +    movhps       xm7, [vq+widthq+6]
>> +    vinserti128  m3,   m3, xm7, 1
>> +%else
>>      movq    m3, [uq+widthq]
>>      movhps  m3, [vq+widthq]
>> +%endif
> 
> Ditto. Also use xm2 instead of xm7 since it's unused at this point and
> it avoids having to use an extra vector register in the AVX2 version.

Thanks.

>> +%if cpuflag(avx2)
>> +    movu         [dstq],    xm0
>> +    movu         [dstq+16], xm1
>> +    vextracti128 [dstq+32], m0, 1
>> +    vextracti128 [dstq+48], m1, 1
>> +%else
>>      movu    [dstq], m0
>>      movu    [dstq+mmsize], m1
>> +%endif
> 
> Ditto.
> 
> Otherwise LGTM.

Noted.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 603 bytes
Desc: OpenPGP digital signature
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20160114/c49bcf11/attachment.sig>