[FFmpeg-devel] [PATCH 0/5] Integer scalar product DSP functions

[Michael Niedermayer]

On Wed, May 08, 2013 at 04:53:49PM +0200, Christophe Gisquet wrote:
> 2013/5/8 Christophe Gisquet <christophe.gisquet at gmail.com>:
> > Those functions generally work in batch of 16 elements, but some cases
> > need batches of 8. The first 2 patches only matter for ra144, and alternate
> > patch ra144-pad-buffer-and-use-scalarproduct_int16.patch is probably a better
> > idea.
> 
> Here it is.
> 
> Btw, fate-ra-144 and fate-lossless-wma do pass here.

>  ra144.c |   12 +++++-------
>  ra144.h |    4 +++-
>  2 files changed, 8 insertions(+), 8 deletions(-)
> c15a9d6d52cfe700ac99d4a4df6a7213b177e02d  ra144-pad-buffer-and-use-scalarproduct_int16.patch
> diff --git a/libavcodec/ra144.c b/libavcodec/ra144.c
> index fe9a5bc..2874b78 100644
> --- a/libavcodec/ra144.c
> +++ b/libavcodec/ra144.c
> @@ -1681,12 +1681,9 @@ unsigned int ff_rescale_rms(unsigned int rms, unsigned int energy)
>  }
>  
>  /** inverse root mean square */
> -int ff_irms(const int16_t *data)
> +int ff_irms(DSPContext *dsp, const int16_t *data)
>  {
> -    unsigned int i, sum = 0;
> -
> -    for (i=0; i < BLOCKSIZE; i++)
> -        sum += data[i] * data[i];
> +    unsigned int sum = dsp->scalarproduct_int16(data, data, BLOCKSIZE);
>  
>      if (sum == 0)
>          return 0; /* OOPS - division by zero */
> @@ -1698,14 +1695,15 @@ void ff_subblock_synthesis(RA144Context *ractx, const int16_t *lpc_coefs,
>                             int cba_idx, int cb1_idx, int cb2_idx,
>                             int gval, int gain)
>  {
> -    int16_t buffer_a[BLOCKSIZE];
> +    LOCAL_ALIGNED(16, int16_t, buffer_a, [FFALIGN(BLOCKSIZE,16)]);
>      int16_t *block;
>      int m[3];
>

> +    AV_ZERO128(buffer_a+BLOCKSIZE);

if buffer_a is moved to the context then this could be avoided
also it would avoid the need for LOCAL_ALIGNED()

also trying to build the code with the patch results in:

libavcodec/ra144enc.c: In function ‘ra144_encode_subblock’:
libavcodec/ra144enc.c:377:9: warning: passing argument 1 of ‘ff_irms’ from incompatible pointer type [enabled by default]
libavcodec/ra144.h:73:5: note: expected ‘struct DSPContext *’ but argument is of type ‘int16_t *’
libavcodec/ra144enc.c:377:9: error: too few arguments to function ‘ff_irms’
libavcodec/ra144.h:73:5: note: declared here

[...]

thanks
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Dictatorship naturally arises out of democracy, and the most aggravated
form of tyranny and slavery out of the most extreme liberty. -- Plato
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 198 bytes
Desc: Digital signature
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20130508/86440db8/attachment.asc>