[FFmpeg-devel] [PATCH 05/12] WMA: use vector_fmul_step_scalar

[Måns Rullgård]

Michael Niedermayer <michaelni at gmx.at> writes:

> On Tue, Sep 29, 2009 at 12:04:06PM +0100, M?ns Rullg?rd wrote:
>> Michael Niedermayer <michaelni at gmx.at> writes:
>> 
>> > On Sun, Sep 27, 2009 at 11:49:21AM +0100, Mans Rullgard wrote:
>> >> ---
>> >>  libavcodec/wmadec.c |    6 +++---
>> >>  1 files changed, 3 insertions(+), 3 deletions(-)
>> >> 
>> >> diff --git a/libavcodec/wmadec.c b/libavcodec/wmadec.c
>> >> index a4d5ad6..9367e97 100644
>> >> --- a/libavcodec/wmadec.c
>> >> +++ b/libavcodec/wmadec.c
>> >> @@ -601,9 +601,9 @@ static int wma_decode_block(WMACodecContext *s)
>> >>                  for(i = 0;i < s->coefs_start; i++)
>> >>                      *coefs++ = 0.0;
>> >>                  n = nb_coefs[ch];
>> >> -                for(i = 0;i < n; i++) {
>> >> -                    *coefs++ = coefs1[i] * exponents[i<<bsize>>esize] * mult;
>> >> -                }
>> >> +                s->dsp.vector_fmul_step_scalar(coefs, coefs1, exponents,
>> >> +                                               mult, 1<<bsize>>esize, n);
>> >> +                coefs += n;
>> >>                  n = s->block_len - s->coefs_end[bsize];
>> >>                  for(i = 0;i < n; i++)
>> >>                      *coefs++ = 0.0;
>> >
>> > i suspect that highlevel optimizations are possible and should be done first
>> > as they likely would render this optimized code obsolete
>> >
>> > for example: what value does 1<<bsize>>esize generally have?
>> > if its commonly 1 the others could be convertet ...
>> 
>> On a randomly chosen test file, 1 is the most common value, roughly
>> 50%, with 2, 3, and 4 also occurring frequently.
>
> i assume you speak of bsize-esize, because 1<<bsize>>esize cant be 3
> i also assume that the % are scaled so that smaller n are weighted
> accordingly less
>
> ive not checked at all but from the hunk above it looks as if the original
> code was written to also handle bsize < esize, the new code would not ...

I just noticed that too.  Working on it.

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