---
Separated loads from muls. It's slightly faster for SSE2. If making this an
AVX function is ever faster on some system, we can change it to 3-arg mulps.
libavresample/x86/audio_convert.asm | 33 ++++++++++++++++++++++++++++++++
libavresample/x86/audio_convert_init.c | 4 +++
2 files changed, 37 insertions(+), 0 deletions(-)
diff --git a/libavresample/x86/audio_convert.asm
b/libavresample/x86/audio_convert.asm
index a281981..9f32d04 100644
--- a/libavresample/x86/audio_convert.asm
+++ b/libavresample/x86/audio_convert.asm
@@ -27,6 +27,7 @@
SECTION_RODATA 32
pf_s32_inv_scale: times 8 dd 0x30000000
+pf_s32_scale: times 8 dd 0x4f000000
pf_s16_inv_scale: times 4 dd 0x38000000
pf_s16_scale: times 4 dd 0x47000000
@@ -191,6 +192,38 @@ cglobal conv_flt_to_s16, 3,3,5, dst, src, len
jl .loop
REP_RET
+;------------------------------------------------------------------------------
+; void ff_conv_flt_to_s32(int32_t *dst, const float *src, int len);
+;------------------------------------------------------------------------------
+
+INIT_XMM sse2
+cglobal conv_flt_to_s32, 3,3,5, dst, src, len
+ lea lenq, [lend*4]
+ add srcq, lenq
+ add dstq, lenq
+ neg lenq
+ mova m4, [pf_s32_scale]
+.loop:
+ mova m0, [srcq+lenq ]
+ mova m1, [srcq+lenq+1*mmsize]
+ mova m2, [srcq+lenq+2*mmsize]
+ mova m3, [srcq+lenq+3*mmsize]
+ mulps m0, m4
+ mulps m1, m4
+ mulps m2, m4
+ mulps m3, m4
+ cvtps2dq m0, m0
+ cvtps2dq m1, m1
+ cvtps2dq m2, m2
+ cvtps2dq m3, m3
+ mova [dstq+lenq ], m0
+ mova [dstq+lenq+1*mmsize], m1
+ mova [dstq+lenq+2*mmsize], m2
+ mova [dstq+lenq+3*mmsize], m3
+ add lenq, mmsize*4
+ jl .loop
+ REP_RET
+
;-----------------------------------------------------------------------------
; void ff_conv_fltp_to_flt_6ch(float *dst, float *const *src, int len,
; int channels);
diff --git a/libavresample/x86/audio_convert_init.c
b/libavresample/x86/audio_convert_init.c
index 1b2f93c..cc48c10 100644
--- a/libavresample/x86/audio_convert_init.c
+++ b/libavresample/x86/audio_convert_init.c
@@ -35,6 +35,8 @@ extern void ff_conv_s32_to_flt_avx (float *dst, const int32_t
*src, int len);
extern void ff_conv_flt_to_s16_sse2(int16_t *dst, const float *src, int len);
+extern void ff_conv_flt_to_s32_sse2(int32_t *dst, const float *src, int len);
+
extern void ff_conv_fltp_to_flt_6ch_mmx (float *dst, float *const *src, int
len);
extern void ff_conv_fltp_to_flt_6ch_sse4(float *dst, float *const *src, int
len);
extern void ff_conv_fltp_to_flt_6ch_avx (float *dst, float *const *src, int
len);
@@ -71,6 +73,8 @@ av_cold void ff_audio_convert_init_x86(AudioConvert *ac)
0, 16, 8, "SSE2", ff_conv_s32_to_flt_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_FLT,
0, 16, 16, "SSE2", ff_conv_flt_to_s16_sse2);
+ ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S32, AV_SAMPLE_FMT_FLT,
+ 0, 16, 16, "SSE2", ff_conv_flt_to_s32_sse2);
}
if (mm_flags & AV_CPU_FLAG_SSE4 && HAVE_SSE) {
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_S16,
--
1.7.1
_______________________________________________
libav-devel mailing list
[email protected]
https://lists.libav.org/mailman/listinfo/libav-devel
