[FFmpeg-devel] [PATCH] NEON: put_pixels_clamped

[David Conrad]

On Apr 29, 2009, at 5:30 AM, M?ns Rullg?rd wrote:

> David Conrad <lessen42 at gmail.com> writes:
>
>> On Apr 16, 2009, at 4:23 PM, M?ns Rullg?rd wrote:
>>
>>> David Conrad <lessen42 at gmail.com> writes:
>>>
>>>> On Apr 16, 2009, at 3:44 PM, David Conrad wrote:
>>>>
>>>>> On Apr 16, 2009, at 3:32 PM, M?ns Rullg?rd wrote:
>>>>>
>>>>>> David Conrad <lessen42 at gmail.com> writes:
>>>>>>
>>>>>>> Hi,
>>>>>>>
>>>>>>> Apparently this is used for some wmv3 files in addition to the
>>>>>>> signed variant.  < 1% faster decode.
>>>>>>>
>>>>>>>
>>>>>>> diff --git a/libavcodec/arm/dsputil_neon_s.S b/libavcodec/arm/
>>>>>>> dsputil_neon_s.S
>>>>>>> index f16293d..159ee64 100644
>>>>>>> --- a/libavcodec/arm/dsputil_neon_s.S
>>>>>>> +++ b/libavcodec/arm/dsputil_neon_s.S
>>>>>>> @@ -273,6 +273,30 @@ function ff_put_h264_qpel8_mc00_neon,
>>>>>>> export=1
>>>>>>>      pixfunc2 put_ pixels8_y2,   _no_rnd, vhadd.u8
>>>>>>>      pixfunc2 put_ pixels8_xy2,  _no_rnd, vshrn.u16, 1
>>>>>>>
>>>>>>> +function ff_put_pixels_clamped_neon, export=1
>>>>>>> +        vld1.64         {d16-d19}, [r0,:128]!
>>>>>>> +        vqmovn.u16      d0, q8
>>>>>>> +        vld1.64         {d20-d23}, [r0,:128]!
>>>>>>> +        vqmovn.u16      d1, q9
>>>>>>> +        vqmovn.u16      d2, q10
>>>>>>> +        vld1.64         {d24-d27}, [r0,:128]!
>>>>>>> +        vqmovn.u16      d3, q11
>>>>>>> +        vqmovn.u16      d4, q12
>>>>>>> +        vld1.64         {d28-d31}, [r0,:128]!
>>>>>>> +        vqmovn.u16      d5, q13
>>>>>>> +        vqmovn.u16      d6, q14
>>>>>>> +        vst1.64         {d0},      [r1,:64], r2
>>>>>>> +        vqmovn.u16      d7, q15
>>>>>>> +        vst1.64         {d1},      [r1,:64], r2
>>>>>>> +        vst1.64         {d2},      [r1,:64], r2
>>>>>>> +        vst1.64         {d3},      [r1,:64], r2
>>>>>>> +        vst1.64         {d4},      [r1,:64], r2
>>>>>>> +        vst1.64         {d5},      [r1,:64], r2
>>>>>>> +        vst1.64         {d6},      [r1,:64], r2
>>>>>>> +        vst1.64         {d7},      [r1,:64], r2
>>>>>>> +        bx              lr
>>>>>>> +        .endfunc
>>>>>>
>>>>>> Shouldn't those be vqmovun.s16?  I'd also try to interleave them
>>>>>> with the loads and stores a bit more for better dual-issue
>>>>>> opportunities.
>>>>>
>>>>> Unsigned pixels; MMX does the same (packuswb for  
>>>>> put_pixels_clamped
>>>>> vs. packsswb for put_signed_pixels_clamped)
>>>>
>>>> Oops, you're right, I didn't read packuswb. New patch attached.
>>>>
>>>>> Also, the loads take two issue cycles since they're loading 4
>>>>> registers; shouldn't they be able to dual issue on both cycles?
>>>
>>> On Cortex-A8 NEON instructions with more than one issue cycle can
>>> dual-issue on the first or the last cycle but not both.
>>
>> As discussed on IRC, all the multicycle NEON instructions I've  
>> checked
>> can dual issue on both their first and last cycles.
>
> My mistake.
>
>> I can't measure any speed difference doing it this way, but it  
>> looks a
>> bit more consistent with the other put_/add_ pixels_clamped.
>
> OK.

Applied.