Re: [FFmpeg-devel] avfilter/vf_interlace : add checkasm for lowpass_line and AVX2 version

[Martin Vignali]

2017-12-30 19:57 GMT+01:00 Martin Vignali :

>
> > >>
>> > >> This broke several interlace fate tests, including the new checkasm
>> one
>> > >> you added.
>> > >>
>
>
> New patch in attach for AVX2 version
> i add a process of only 1* mmsize, before the loop (who process 2 * mmsize
> at each loop)
> Pass fate test for me (osX, x86_64).
>
> Martin
>
> Pushed

Martin
___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

Re: [FFmpeg-devel] avfilter/vf_interlace : add checkasm for lowpass_line and AVX2 version

[Martin Vignali]

> > >>
> > >> This broke several interlace fate tests, including the new checkasm
> one
> > >> you added.
> > >>
> 
>

New patch in attach for AVX2 version
i add a process of only 1* mmsize, before the loop (who process 2 * mmsize
at each loop)
Pass fate test for me (osX, x86_64).

Martin


0001-avfilter-x86-vf_interlace-add-AVX2-version.patch
Description: Binary data
___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

Re: [FFmpeg-devel] avfilter/vf_interlace : add checkasm for lowpass_line and AVX2 version

[Thomas Mundt]

2017-12-19 22:57 GMT+01:00 James Almer :

> On 12/19/2017 6:40 PM, Martin Vignali wrote:
> > 2017-12-19 21:59 GMT+01:00 James Almer :
> >
> >> On 12/19/2017 5:16 PM, Martin Vignali wrote:
> 
>  LGTM, thanks.
> 
> >>>
> >>> Pushed, thanks
> >>
> >> This broke several interlace fate tests, including the new checkasm one
> >> you added.
> >>
> >> Sorry forget to run fate after AVX2 patch (only test with checkasm)
> >
> > For the checkasm fail, i can't reproduce it (on x86_64)
>
> I can reproduce it on mingw (x86_64), but not on Linux 86_64.
>

Sorry, I also should have run fate before saying LGTM.
Can´t reproduce the checkasm fail on msys2 x86_32. Only got this machine
atm.
___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

Re: [FFmpeg-devel] avfilter/vf_interlace : add checkasm for lowpass_line and AVX2 version

[James Almer]

On 12/19/2017 6:40 PM, Martin Vignali wrote:
> 2017-12-19 21:59 GMT+01:00 James Almer :
> 
>> On 12/19/2017 5:16 PM, Martin Vignali wrote:

 LGTM, thanks.

>>>
>>> Pushed, thanks
>>
>> This broke several interlace fate tests, including the new checkasm one
>> you added.
>>
>> Sorry forget to run fate after AVX2 patch (only test with checkasm)
> 
> For the checkasm fail, i can't reproduce it (on x86_64)

I can reproduce it on mingw (x86_64), but not on Linux 86_64.

> 
> If you can, revert the patchset.
> I can't take a look into this soon.

Ok, will revert the avx2 patches and the checkasm test.

> 
> Martin
> ___
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> 

___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

Re: [FFmpeg-devel] avfilter/vf_interlace : add checkasm for lowpass_line and AVX2 version

[Martin Vignali]

2017-12-19 21:59 GMT+01:00 James Almer :

> On 12/19/2017 5:16 PM, Martin Vignali wrote:
> >>
> >> LGTM, thanks.
> >>
> >
> > Pushed, thanks
>
> This broke several interlace fate tests, including the new checkasm one
> you added.
>
> Sorry forget to run fate after AVX2 patch (only test with checkasm)

For the checkasm fail, i can't reproduce it (on x86_64)

If you can, revert the patchset.
I can't take a look into this soon.

Martin
___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

Re: [FFmpeg-devel] avfilter/vf_interlace : add checkasm for lowpass_line and AVX2 version

[James Almer]

On 12/19/2017 5:16 PM, Martin Vignali wrote:
>>
>> LGTM, thanks.
>>
> 
> Pushed, thanks

This broke several interlace fate tests, including the new checkasm one
you added.

http://fate.ffmpeg.org/report.cgi?time=20171219201300=x86_64-archlinux-gcc-threads-2
___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

Re: [FFmpeg-devel] avfilter/vf_interlace : add checkasm for lowpass_line and AVX2 version

[Martin Vignali]

>
> LGTM, thanks.
>

Pushed, thanks

Do you think, that the complex low pass filter could also be improved that
> way?
>
>
>
Yes, probably. I will take a look.


Martin
___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

Re: [FFmpeg-devel] avfilter/vf_interlace : add checkasm for lowpass_line and AVX2 version

[Thomas Mundt]

2017-12-18 11:34 GMT+01:00 Martin Vignali :

> > >
> >
> > Please also add the changes you made in patch 1 and avx2 to
> vf_tinterlace.
> >
> >
> >
> > For patch 1, IMHO, it's not necessary (the modification is mainly to make
> checkasm test easier to write, and like vf_interlace and vf_tinterlace use
> the same asm
> only one is useful for checkasm)
>

I just thought it might be easier to maintain when code in both filters is
more similar.
But it´s not a big thing, so OK.

In attach new patchs, adding AVX2 lowpass_line 8 and 16 for vf_tinterlace
>

LGTM, thanks.
Do you think, that the complex low pass filter could also be improved that
way?

Regards,
Thomas
___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

Re: [FFmpeg-devel] avfilter/vf_interlace : add checkasm for lowpass_line and AVX2 version

[Martin Vignali]

> >
>
> Please also add the changes you made in patch 1 and avx2 to vf_tinterlace.
>
>
>
> For patch 1, IMHO, it's not necessary (the modification is mainly to make
checkasm test easier to write, and like vf_interlace and vf_tinterlace use
the same asm
only one is useful for checkasm)

In attach new patchs, adding AVX2 lowpass_line 8 and 16 for vf_tinterlace

Martin


0001-avfilter-vf_interlace-move-func-init-in.patch
Description: Binary data


0002-checkasm-vf_interlace-add-test-for-lowpass_line-8-an.patch
Description: Binary data


0003-avfilter-vf_interlace-add-AVX2-for-lowpass_line-8-an.patch
Description: Binary data


0004-avfilter-vf_tinterlace-add-AVX2-func-for-lowpass_lin.patch
Description: Binary data
___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

Re: [FFmpeg-devel] avfilter/vf_interlace : add checkasm for lowpass_line and AVX2 version

[Thomas Mundt]

Hi,

2017-12-16 18:35 GMT+01:00 Martin Vignali :

> 2017-12-16 14:48 GMT+01:00 Carl Eugen Hoyos :
>
> > 2017-12-16 14:17 GMT+01:00 Martin Vignali :
> >
> > > 002 : Checkasm test for lowpass_line
> >
> > The change to checkasm.mak contains unexpected tabs iiuc.
> >
>
> New patch in attach
>

Please also add the changes you made in patch 1 and avx2 to vf_tinterlace.

Regards,
Thomas
___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

Re: [FFmpeg-devel] avfilter/vf_interlace : add checkasm for lowpass_line and AVX2 version

[Paul B Mahol]

On 12/16/17, Martin Vignali  wrote:
> 2017-12-16 14:48 GMT+01:00 Carl Eugen Hoyos :
>
>> 2017-12-16 14:17 GMT+01:00 Martin Vignali :
>>
>> > 002 : Checkasm test for lowpass_line
>>
>> The change to checkasm.mak contains unexpected tabs iiuc.
>>
>
> New patch in attach
>
>
>> I can confirm that avx is not faster than sse2 here.
>>
>>
>> Thanks for testing, i will post a patch to remove avx version

Please don't.
___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

Re: [FFmpeg-devel] avfilter/vf_interlace : add checkasm for lowpass_line and AVX2 version

[Martin Vignali]

2017-12-16 14:48 GMT+01:00 Carl Eugen Hoyos :

> 2017-12-16 14:17 GMT+01:00 Martin Vignali :
>
> > 002 : Checkasm test for lowpass_line
>
> The change to checkasm.mak contains unexpected tabs iiuc.
>

New patch in attach


> I can confirm that avx is not faster than sse2 here.
>
>
> Thanks for testing, i will post a patch to remove avx version

Martin


0001-avfilter-vf_interlace-move-func-init-in.patch
Description: Binary data


0002-checkasm-vf_interlace-add-test-for-lowpass_line-8-an.patch
Description: Binary data


0003-avfilter-vf_interlace-add-AVX2-for-lowpass_line-8-an.patch
Description: Binary data
___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

Re: [FFmpeg-devel] avfilter/vf_interlace : add checkasm for lowpass_line and AVX2 version

[Carl Eugen Hoyos]

2017-12-16 14:17 GMT+01:00 Martin Vignali :

> 002 : Checkasm test for lowpass_line

The change to checkasm.mak contains unexpected tabs iiuc.

I can confirm that avx is not faster than sse2 here.

Carl Eugen
___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel

[FFmpeg-devel] avfilter/vf_interlace : add checkasm for lowpass_line and AVX2 version

[Martin Vignali]

Hello,

In attach patch to add a checkasm test, for lowpass_line (8 and 16)
and AVX2 version for lowpass_line (8 and 16)


001 : Modify init part of vf_interlace (add ff_interlace_init
and modify ff_interlace_init_x86)

002 : Checkasm test for lowpass_line
can be test with
./tests/checkasm/checkasm --test=vf_interlace --bench
or
make fate-checkasm

003 : AVX2 version for lowpass_line (8 and 16)


For the checkasm test, i add an align test (related to ticket 6491)
but i only "unalign" the "mref value" (if i unalign src in this test, i
have a crash)
I can't find a "real" filter command line who reproduce the same crash
(if someone have a command line example for that, i will modify the
checkasm and the asm)

Checkasm result : (in my tests the AVX version is always slower, than SSE
version (not sure it's interested to keep AVX version)

./tests/checkasm/checkasm --test=vf_interlace --bench

benchmarking with native FFmpeg timers
nop: 36.2
checkasm: using random seed 1323076929
SSE2:
 - vf_interlace.lowpass_line_8  [OK]
 - vf_interlace.lowpass_line_16 [OK]
AVX:
 - vf_interlace.lowpass_line_8  [OK]
 - vf_interlace.lowpass_line_16 [OK]
AVX2:
 - vf_interlace.lowpass_line_8  [OK]
 - vf_interlace.lowpass_line_16 [OK]
checkasm: all 6 tests passed
lowpass_line_8_c: 311.2
lowpass_line_8_sse2: 44.4
lowpass_line_8_avx: 46.9
lowpass_line_8_avx2: 26.7
lowpass_line_16_c: 182.9
lowpass_line_16_sse2: 44.4
lowpass_line_16_avx: 48.4
lowpass_line_16_avx2: 28.2


Martin


0001-avfilter-vf_interlace-move-func-init-in.patch
Description: Binary data


0002-checkasm-vf_interlace-add-test-for-lowpass_line-8-an.patch
Description: Binary data


0003-avfilter-vf_interlace-add-AVX2-for-lowpass_line-8-an.patch
Description: Binary data
___
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel