On 2/21/15, James Almer <jamr...@gmail.com> wrote:
> Zeroing them allows us to use scalarproduct_int16 even if the length of
> vectors is not multiple of 16.
>
> Signed-off-by: James Almer <jamr...@gmail.com>
> ---
> libavcodec/takdec.c | 28 +++++-----------------------
> 1 file changed, 5 insertions(+), 23 deletions(-)
>
> diff --git a/libavcodec/takdec.c b/libavcodec/takdec.c
> index 2f0155d..0f808e0 100644
> --- a/libavcodec/takdec.c
> +++ b/libavcodec/takdec.c
> @@ -451,6 +451,7 @@ static int decode_subframe(TAKDecContext *s, int32_t
> *decoded,
> s->filter[j] = x - ((tfilter[i] + y) >> (15 - filter_quant));
> s->filter[i] = x - ((tfilter[j] + y) >> (15 - filter_quant));
> }
> + memset(s->filter + filter_order, 0, sizeof(s->filter) -
> filter_order*2);
>
> if ((ret = decode_residues(s, &decoded[filter_order],
> subframe_size - filter_order)) < 0)
> @@ -467,15 +468,8 @@ static int decode_subframe(TAKDecContext *s, int32_t
> *decoded,
> for (i = 0; i < tmp; i++) {
> int v = 1 << (filter_quant - 1);
>
> - if (filter_order & -16)
> - v += s->adsp.scalarproduct_int16(&s->residues[i],
> s->filter,
> - filter_order & -16);
> - for (j = filter_order & -16; j < filter_order; j += 4) {
> - v += s->residues[i + j + 3] * s->filter[j + 3] +
> - s->residues[i + j + 2] * s->filter[j + 2] +
> - s->residues[i + j + 1] * s->filter[j + 1] +
> - s->residues[i + j ] * s->filter[j ];
> - }
> + v += s->adsp.scalarproduct_int16(&s->residues[i], s->filter,
> + FFALIGN(filter_order, 16));
> v = (av_clip(v >> filter_quant, -8192, 8191) << dshift) -
> *decoded;
> *decoded++ = v;
> s->residues[filter_order + i] = v >> dshift;
> @@ -601,6 +595,7 @@ static int decorrelate(TAKDecContext *s, int c1, int c2,
> int length)
> code_size = 14 - get_bits(gb, 3);
> s->filter[i] = get_sbits(gb, code_size);
> }
> + memset(s->filter + filter_order, 0, sizeof(s->filter) -
> filter_order*2);
>
> order_half = filter_order / 2;
> length2 = length - (filter_order - 1);
> @@ -638,20 +633,7 @@ static int decorrelate(TAKDecContext *s, int c1, int
> c2, int length)
> for (i = 0; i < tmp; i++) {
> int v = 1 << 9;
>
> - if (filter_order == 16) {
> - v += s->adsp.scalarproduct_int16(&s->residues[i],
> s->filter,
> - filter_order);
> - } else {
> - v += s->residues[i + 7] * s->filter[7] +
> - s->residues[i + 6] * s->filter[6] +
> - s->residues[i + 5] * s->filter[5] +
> - s->residues[i + 4] * s->filter[4] +
> - s->residues[i + 3] * s->filter[3] +
> - s->residues[i + 2] * s->filter[2] +
> - s->residues[i + 1] * s->filter[1] +
> - s->residues[i ] * s->filter[0];
> - }
> -
> + v += s->adsp.scalarproduct_int16(&s->residues[i],
> s->filter, 16);
> v = (av_clip(v >> 10, -8192, 8191) << dshift) - *p1;
> *p1++ = v;
> }
> --
> 2.3.0
>
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
Have you measured performance drop before and after?
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
