x86/sbrdsp: add ff_sbr_autocorrelate_{sse,sse3}

2 to 2.5 times faster.

Signed-off-by: James Almer <jamrial@gmail.com>

1 files changed, 8 insertions, 0 deletions

diff --git a/libavcodec/x86/sbrdsp_init.c b/libavcodec/x86/sbrdsp_init.c
index a2aca742cf..6911a1a515 100644
--- a/libavcodec/x86/sbrdsp_init.c
+++ b/libavcodec/x86/sbrdsp_init.c
@@ -53,6 +53,9 @@ void ff_sbr_hf_apply_noise_3_sse2(float (*Y)[2], const float *s_m,
 
 void ff_sbr_qmf_deint_neg_sse(float *v, const float *src);
 
+void ff_sbr_autocorrelate_sse (const float x[40][2], float phi[3][2][2]);
+void ff_sbr_autocorrelate_sse3(const float x[40][2], float phi[3][2][2]);
+
 av_cold void ff_sbrdsp_init_x86(SBRDSPContext *s)
 {
     int cpu_flags = av_get_cpu_flags();
@@ -66,6 +69,7 @@ av_cold void ff_sbrdsp_init_x86(SBRDSPContext *s)
         s->qmf_post_shuffle = ff_sbr_qmf_post_shuffle_sse;
         s->qmf_deint_bfly   = ff_sbr_qmf_deint_bfly_sse;
         s->qmf_deint_neg    = ff_sbr_qmf_deint_neg_sse;
+        s->autocorrelate    = ff_sbr_autocorrelate_sse;
     }
 
     if (EXTERNAL_SSE2(cpu_flags)) {
@@ -76,4 +80,8 @@ av_cold void ff_sbrdsp_init_x86(SBRDSPContext *s)
         s->hf_apply_noise[2] = ff_sbr_hf_apply_noise_2_sse2;
         s->hf_apply_noise[3] = ff_sbr_hf_apply_noise_3_sse2;
     }
+
+    if (EXTERNAL_SSE3(cpu_flags)) {
+        s->autocorrelate = ff_sbr_autocorrelate_sse3;
+    }
 }