---
libavresample/x86/audio_mix.asm | 55 ++++++++++++++++++++++++++++++++++++
libavresample/x86/audio_mix_init.c | 13 ++++++++
2 files changed, 68 insertions(+), 0 deletions(-)
diff --git a/libavresample/x86/audio_mix.asm b/libavresample/x86/audio_mix.asm
index dbc79e5..961f7c3 100644
--- a/libavresample/x86/audio_mix.asm
+++ b/libavresample/x86/audio_mix.asm
@@ -64,3 +64,58 @@ MIX_2_TO_1_FLTP_FLT
INIT_YMM avx
MIX_2_TO_1_FLTP_FLT
%endif
+
+;-----------------------------------------------------------------------------
+; void ff_mix_2_to_1_s16p_flt(int16_t **src, float **matrix, int len,
+; int out_ch, int in_ch);
+;-----------------------------------------------------------------------------
+
+%macro MIX_2_TO_1_S16P_FLT 0
+cglobal mix_2_to_1_s16p_flt, 3,4,6, src, matrix, len, src1
+ mov src1q, [srcq+gprsize]
+ mov srcq, [srcq]
+ sub src1q, srcq
+ mov matrixq, [matrixq ]
+ VBROADCASTSS m4, [matrixq ]
+ VBROADCASTSS m5, [matrixq+4]
+ ALIGN 16
+.loop:
+%if cpuflag(sse4)
+ pmovsxwd m0, [srcq ]
+ pmovsxwd m1, [srcq+ mmsize/2]
+ pmovsxwd m2, [srcq+src1q ]
+ pmovsxwd m3, [srcq+src1q+mmsize/2]
+%else
+ mova m0, [srcq ]
+ mova m2, [srcq+src1q]
+ punpckhwd m1, m0
+ punpcklwd m0, m0
+ psrad m1, 16
+ psrad m0, 16
+ punpckhwd m3, m2
+ punpcklwd m2, m2
+ psrad m3, 16
+ psrad m2, 16
+%endif
+ cvtdq2ps m0, m0
+ cvtdq2ps m1, m1
+ cvtdq2ps m2, m2
+ cvtdq2ps m3, m3
+ addps m0, m2
+ addps m1, m3
+ mulps m0, m4
+ mulps m1, m5
+ cvtps2dq m0, m0
+ cvtps2dq m1, m1
+ packssdw m0, m1
+ mova [srcq], m0
+ add srcq, mmsize
+ sub lend, mmsize/2
+ jg .loop
+ REP_RET
+%endmacro
+
+INIT_XMM sse2
+MIX_2_TO_1_S16P_FLT
+INIT_XMM sse4
+MIX_2_TO_1_S16P_FLT
diff --git a/libavresample/x86/audio_mix_init.c
b/libavresample/x86/audio_mix_init.c
index 8f8930f..947c314 100644
--- a/libavresample/x86/audio_mix_init.c
+++ b/libavresample/x86/audio_mix_init.c
@@ -27,6 +27,11 @@ extern void ff_mix_2_to_1_fltp_flt_sse(float **src, float
**matrix, int len,
extern void ff_mix_2_to_1_fltp_flt_avx(float **src, float **matrix, int len,
int out_ch, int in_ch);
+extern void ff_mix_2_to_1_s16p_flt_sse2(int16_t **src, float **matrix, int len,
+ int out_ch, int in_ch);
+extern void ff_mix_2_to_1_s16p_flt_sse4(int16_t **src, float **matrix, int len,
+ int out_ch, int in_ch);
+
av_cold void ff_audio_mix_init_x86(AudioMix *am)
{
#if HAVE_YASM
@@ -36,6 +41,14 @@ av_cold void ff_audio_mix_init_x86(AudioMix *am)
ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,
2, 1, 16, 8, "SSE", ff_mix_2_to_1_fltp_flt_sse);
}
+ if (mm_flags & AV_CPU_FLAG_SSE2 && HAVE_SSE) {
+ ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,
+ 2, 1, 16, 8, "SSE2",
ff_mix_2_to_1_s16p_flt_sse2);
+ }
+ if (mm_flags & AV_CPU_FLAG_SSE4 && HAVE_SSE) {
+ ff_audio_mix_set_func(am, AV_SAMPLE_FMT_S16P, AV_MIX_COEFF_TYPE_FLT,
+ 2, 1, 16, 8, "SSE4",
ff_mix_2_to_1_s16p_flt_sse4);
+ }
if (mm_flags & AV_CPU_FLAG_AVX && HAVE_AVX) {
ff_audio_mix_set_func(am, AV_SAMPLE_FMT_FLTP, AV_MIX_COEFF_TYPE_FLT,
2, 1, 32, 16, "AVX", ff_mix_2_to_1_fltp_flt_avx);
--
1.7.1
_______________________________________________
libav-devel mailing list
[email protected]
https://lists.libav.org/mailman/listinfo/libav-devel
