generic-library/ffmpeg
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

x86: sbrdsp: implement SSE2 hf_apply_noise

Browse Source 
233 to 105 cycles on Arrandale and Win64.
Replacing the multiplication by s_m[m] by a pand and a pxor with
appropriate vectors is slower. Unrolling is a 15 cycles win.
A SSE version was 4 cycles slower.

Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
This commit is contained in:
Christophe Gisquet
2013-04-09 22:16:36 +02:00
committed by Michael Niedermayer
parent 380cfce2b2
commit 76c7277385
3 changed files with 136 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
libavcodec/x86/sbrdsp_init.c
View File

@@ -38,6 +38,19 @@ void ff_sbr_qmf_deint_bfly_sse(float *v, const float *src0, const float *src1);
void ff_sbr_qmf_deint_bfly_sse2(float *v, const float *src0, const float *src1);
void ff_sbr_qmf_pre_shuffle_sse2(float *z);
void ff_sbr_hf_apply_noise_0_sse2(float (*Y)[2], const float *s_m,
const float *q_filt, int noise,
int kx, int m_max);
void ff_sbr_hf_apply_noise_1_sse2(float (*Y)[2], const float *s_m,
const float *q_filt, int noise,
int kx, int m_max);
void ff_sbr_hf_apply_noise_2_sse2(float (*Y)[2], const float *s_m,
const float *q_filt, int noise,
int kx, int m_max);
void ff_sbr_hf_apply_noise_3_sse2(float (*Y)[2], const float *s_m,
const float *q_filt, int noise,
int kx, int m_max);
av_cold void ff_sbrdsp_init_x86(SBRDSPContext *s)
{
int mm_flags = av_get_cpu_flags();
@@ -55,5 +68,9 @@ av_cold void ff_sbrdsp_init_x86(SBRDSPContext *s)
if (EXTERNAL_SSE2(mm_flags)) {
s->qmf_deint_bfly = ff_sbr_qmf_deint_bfly_sse2;
s->qmf_pre_shuffle = ff_sbr_qmf_pre_shuffle_sse2;
s->hf_apply_noise[0] = ff_sbr_hf_apply_noise_0_sse2;
s->hf_apply_noise[1] = ff_sbr_hf_apply_noise_1_sse2;
s->hf_apply_noise[2] = ff_sbr_hf_apply_noise_2_sse2;
s->hf_apply_noise[3] = ff_sbr_hf_apply_noise_3_sse2;
}
}