[FFmpeg-devel] [PATCH 2/2] x86/vf_w3fdif: simplify w3fdif_simple_high

[Hendrik Leppkes]

On Thu, Jan 7, 2016 at 3:13 AM, Hendrik Leppkes <h.leppkes at gmail.com> wrote:
> On Mon, Oct 12, 2015 at 1:21 AM, James Almer <jamrial at gmail.com> wrote:
>> On 10/11/2015 3:11 PM, Ronald S. Bultje wrote:
>>> Hi,
>>>
>>> On Sun, Oct 11, 2015 at 1:17 PM, James Almer <jamrial at gmail.com> wrote:
>>>
>>>> On 10/11/2015 4:31 AM, Paul B Mahol wrote:
>>>>> On 10/11/15, James Almer <jamrial at gmail.com> wrote:
>>>>>> Signed-off-by: James Almer <jamrial at gmail.com>
>>>>>> ---
>>>>>>  libavfilter/x86/vf_w3fdif.asm | 16 +++++++---------
>>>>>>  1 file changed, 7 insertions(+), 9 deletions(-)
>>>>>>
>>>>>> diff --git a/libavfilter/x86/vf_w3fdif.asm
>>>> b/libavfilter/x86/vf_w3fdif.asm
>>>>>> index f02319b..f2001a4 100644
>>>>>> --- a/libavfilter/x86/vf_w3fdif.asm
>>>>>> +++ b/libavfilter/x86/vf_w3fdif.asm
>>>>>> @@ -103,13 +103,11 @@ REP_RET
>>>>>>
>>>>>>  %if ARCH_X86_64
>>>>>>
>>>>>> -cglobal w3fdif_simple_high, 5, 9, 9, 0, work_line, in_lines_cur0,
>>>>>> in_lines_adj0, coef, linesize
>>>>>> +cglobal w3fdif_simple_high, 5, 9, 8, 0, work_line, in_lines_cur0,
>>>>>> in_lines_adj0, coef, linesize
>>>>>>      movq                  m2, [coefq]
>>>>>>      DEFINE_ARGS    work_line, in_lines_cur0, in_lines_adj0,
>>>> in_lines_cur1,
>>>>>> linesize, offset, in_lines_cur2, in_lines_adj1, in_lines_adj2
>>>>>> -    SPLATW                m0, m2, 0
>>>>>> -    SPLATW                m1, m2, 1
>>>>>> +    pshufd                m0, m2, q0000
>>>>>>      SPLATW                m2, m2, 2
>>>>>> -    SBUTTERFLY            wd, 0, 1, 7
>>>>>>      pxor                  m7, m7
>>>>>>      mov              offsetq, 0
>>>>>>      mov       in_lines_cur2q, [in_lines_cur0q+gprsize*2]
>>>>>> @@ -124,23 +122,23 @@ cglobal w3fdif_simple_high, 5, 9, 9, 0, work_line,
>>>>>> in_lines_cur0, in_lines_adj0,
>>>>>>      movh                                   m4, [in_lines_cur1q+offsetq]
>>>>>>      punpcklbw                              m3, m7
>>>>>>      punpcklbw                              m4, m7
>>>>>> -    SBUTTERFLY                             wd, 3, 4, 8
>>>>>> +    SBUTTERFLY                             wd, 3, 4, 1
>>>>>>      pmaddwd                                m3, m0
>>>>>> -    pmaddwd                                m4, m1
>>>>>> +    pmaddwd                                m4, m0
>>>>>>      movh                                   m5, [in_lines_adj0q+offsetq]
>>>>>>      movh                                   m6, [in_lines_adj1q+offsetq]
>>>>>>      punpcklbw                              m5, m7
>>>>>>      punpcklbw                              m6, m7
>>>>>> -    SBUTTERFLY                             wd, 5, 6, 8
>>>>>> +    SBUTTERFLY                             wd, 5, 6, 1
>>>>>>      pmaddwd                                m5, m0
>>>>>> -    pmaddwd                                m6, m1
>>>>>> +    pmaddwd                                m6, m0
>>>>>>      paddd                                  m3, m5
>>>>>>      paddd                                  m4, m6
>>>>>>      movh                                   m5, [in_lines_cur2q+offsetq]
>>>>>>      movh                                   m6, [in_lines_adj2q+offsetq]
>>>>>>      punpcklbw                              m5, m7
>>>>>>      punpcklbw                              m6, m7
>>>>>> -    SBUTTERFLY                             wd, 5, 6, 8
>>>>>> +    SBUTTERFLY                             wd, 5, 6, 1
>>>>>>      pmaddwd                                m5, m2
>>>>>>      pmaddwd                                m6, m2
>>>>>>      paddd                                  m3, m5
>>>>>> --
>>>>>> 2.6.0
>>>>>>
>>>>>> _______________________________________________
>>>>>> ffmpeg-devel mailing list
>>>>>> ffmpeg-devel at ffmpeg.org
>>>>>> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>>>>>>
>>>>>
>>>>> Cant this now be used on x32?
>>>>
>>>
>>> Add to the data pointers directly (in_lines_cur0q and work_lineq). Then sub
>>> all other curXq/adjXq from cur0q (on 32bit only) before the loop and you
>>> have to adds (on 32bit) instead of one (on 64bit), but one reg less
>>> (offset), making it 7, which means it works.
>>>
>>> Ronald
>>
>> Ah, like it's being done in PACK_6CH from swr's audio_convert.asm
>> For complex_high some stack ab/use will be needed (see PACK_8CH), but it should
>> be doable.
>> This way w3fdif will be able to fully dethrone yadif :P
>
> Are you still working on w3fdif_simple_high for 32bit?
> I would be interested in that. Otherwise I might try to do it myself,
> although it sounds like a lot of #if'ery

I was bored and it was easy, so patch coming up anyway!