---
Athlon64
C - 6216
SSE2 - 1333
SandyBridge
C - 8209
SSE2 - 470
AVX - 385
libavresample/x86/audio_convert.asm | 38 +++++++++++++++++++++++++++++++-
libavresample/x86/audio_convert_init.c | 9 +++++++
2 files changed, 46 insertions(+), 1 deletions(-)
diff --git a/libavresample/x86/audio_convert.asm
b/libavresample/x86/audio_convert.asm
index 3e3c781..ee71d42 100644
--- a/libavresample/x86/audio_convert.asm
+++ b/libavresample/x86/audio_convert.asm
@@ -25,6 +25,7 @@
SECTION_RODATA 32
+pf_s32_inv_scale: times 8 dd 0x30000000
pf_s16_inv_scale: times 4 dd 0x38000000
SECTION_TEXT
@@ -54,7 +55,6 @@ cglobal conv_s16_to_s32, 3,3,3, dst, src, len
;------------------------------------------------------------------------------
; void ff_conv_s16_to_flt(float *dst, const int16_t *src, int len);
;------------------------------------------------------------------------------
-
%macro CONV_S16_TO_FLT 0
cglobal conv_s16_to_flt, 3,3,3, dst, src, len
lea lenq, [2*lend]
@@ -128,6 +128,42 @@ CONV_S32_TO_S16
INIT_XMM sse2
CONV_S32_TO_S16
+;------------------------------------------------------------------------------
+; void ff_conv_s32_to_flt(float *dst, const int32_t *src, int len);
+;------------------------------------------------------------------------------
+
+%macro CONV_S32_TO_FLT 0
+cglobal conv_s32_to_flt, 3,3,3, dst, src, len
+ lea lenq, [4*lend]
+ add srcq, lenq
+ add dstq, lenq
+ neg lenq
+ mova m0, [pf_s32_inv_scale]
+ ALIGN 16
+.loop:
+ cvtdq2ps m1, [srcq+lenq ]
+ cvtdq2ps m2, [srcq+lenq+mmsize]
+ mulps m1, m1, m0
+ mulps m2, m2, m0
+ mova [dstq+lenq ], m1
+ mova [dstq+lenq+mmsize], m2
+ add lenq, mmsize*2
+ jl .loop
+%if mmsize == 32
+ vzeroupper
+ RET
+%else
+ REP_RET
+%endif
+%endmacro
+
+INIT_XMM sse2
+CONV_S32_TO_FLT
+%if HAVE_AVX
+INIT_YMM avx
+CONV_S32_TO_FLT
+%endif
+
;-----------------------------------------------------------------------------
; 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 e236626..2ef8e08 100644
--- a/libavresample/x86/audio_convert_init.c
+++ b/libavresample/x86/audio_convert_init.c
@@ -30,6 +30,9 @@ extern void ff_conv_s16_to_flt_sse4(float *dst, const int16_t
*src, int len);
extern void ff_conv_s32_to_s16_mmx (int16_t *dst, const int32_t *src, int len);
extern void ff_conv_s32_to_s16_sse2(int16_t *dst, const int32_t *src, int len);
+extern void ff_conv_s32_to_flt_sse2(float *dst, const int32_t *src, int len);
+extern void ff_conv_s32_to_flt_avx (float *dst, const int32_t *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_sse(float *dst, float *const *src, int
len);
@@ -57,10 +60,16 @@ av_cold void ff_audio_convert_init_x86(AudioConvert *ac)
0, 16, 8, "SSE2", ff_conv_s16_to_s32_sse2);
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_S16,
0, 16, 8, "SSE2", ff_conv_s16_to_flt_sse2);
+ ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_S32,
+ 0, 16, 8, "SSE2", ff_conv_s32_to_flt_sse2);
}
if (mm_flags & AV_CPU_FLAG_SSE4 && HAVE_SSE) {
ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_S16,
0, 16, 8, "SSE4", ff_conv_s16_to_flt_sse4);
}
+ if (mm_flags & AV_CPU_FLAG_AVX && HAVE_AVX) {
+ ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_S32,
+ 0, 32, 16, "AVX", ff_conv_s32_to_flt_avx);
+ }
#endif
}
--
1.7.1
_______________________________________________
libav-devel mailing list
[email protected]
https://lists.libav.org/mailman/listinfo/libav-devel
