[FFmpeg-devel] [PATCH 2/2] ac3enc: add SIMD-optimized shifting functions for use with the fixed-point AC3 encoder.

[Måns Rullgård]

Ivan Kalvachev <ikalvachev at gmail.com> writes:

> On 3/12/11, M?ns Rullg?rd <mans at mansr.com> wrote:
>> Justin Ruggles <justin.ruggles at gmail.com> writes:
>>
>>> ---
>>>  libavcodec/ac3dsp.c         |   24 +++++++++++++++++++++
>>>  libavcodec/ac3dsp.h         |   22 +++++++++++++++++++
>>>  libavcodec/ac3enc_fixed.c   |   40 ++---------------------------------
>>>  libavcodec/x86/ac3dsp.asm   |   48
>>> +++++++++++++++++++++++++++++++++++++++++++
>>>  libavcodec/x86/ac3dsp_mmx.c |   12 ++++++++++
>>>  5 files changed, 109 insertions(+), 37 deletions(-)
>>>
>>>
>>> diff --git a/libavcodec/ac3dsp.c b/libavcodec/ac3dsp.c
>>> index da3a123..83f48a6 100644
>>> --- a/libavcodec/ac3dsp.c
>>> +++ b/libavcodec/ac3dsp.c
>>> @@ -50,10 +50,34 @@ static int ac3_max_msb_abs_int16_c(const int16_t *src,
>>> int len)
>>>      return v;
>>>  }
>>>
>>> +static void ac3_lshift_int16_c(int16_t *src, unsigned int len,
>>> +                               unsigned int shift)
>>> +{
>>> +    int i;
>>> +
>>> +    if (shift > 0) {
>>> +        for (i = 0; i < len; i++)
>>> +            src[i] <<= shift;
>>> +    }
>>> +}
>
> You can shift 2 values at once if you read them as int32_t, shift and
> then mask (masking could even be skipped if you are sure they will
> never overflow).

There's no point optimising the C version beyond comprehension.  We have
the asm hooks for that.

-- 
M?ns Rullg?rd
mans at mansr.com