From b66e20d2aab58689f33af38a3ef5570c99f86295 Mon Sep 17 00:00:00 2001
From: Justin Ruggles <justin.ruggles@gmail.com>
Date: Wed, 2 May 2012 16:32:08 -0400
Subject: [PATCH] lavr: x86: optimized 2-channel fltp to flt conversion

---
 libavresample/x86/audio_convert.asm    | 37 ++++++++++++++++++++++++++
 libavresample/x86/audio_convert_init.c |  7 +++++
 2 files changed, 44 insertions(+)

diff --git a/libavresample/x86/audio_convert.asm b/libavresample/x86/audio_convert.asm
index acc6bb0dc1..48508af81c 100644
--- a/libavresample/x86/audio_convert.asm
+++ b/libavresample/x86/audio_convert.asm
@@ -695,6 +695,43 @@ INIT_XMM avx
 CONV_FLTP_TO_S16_6CH
 %endif
 
+;------------------------------------------------------------------------------
+; void ff_conv_fltp_to_flt_2ch(float *dst, float *const *src, int len,
+;                              int channels);
+;------------------------------------------------------------------------------
+
+%macro CONV_FLTP_TO_FLT_2CH 0
+cglobal conv_fltp_to_flt_2ch, 3,4,5, dst, src0, len, src1
+    mov  src1q, [src0q+gprsize]
+    mov  src0q, [src0q]
+    lea   lenq, [4*lend]
+    add  src0q, lenq
+    add  src1q, lenq
+    lea   dstq, [dstq+2*lenq]
+    neg   lenq
+.loop
+    mova    m0, [src0q+lenq       ]
+    mova    m1, [src1q+lenq       ]
+    mova    m2, [src0q+lenq+mmsize]
+    mova    m3, [src1q+lenq+mmsize]
+    SBUTTERFLYPS 0, 1, 4
+    SBUTTERFLYPS 2, 3, 4
+    mova  [dstq+2*lenq+0*mmsize], m0
+    mova  [dstq+2*lenq+1*mmsize], m1
+    mova  [dstq+2*lenq+2*mmsize], m2
+    mova  [dstq+2*lenq+3*mmsize], m3
+    add   lenq, 2*mmsize
+    jl .loop
+    REP_RET
+%endmacro
+
+INIT_XMM sse
+CONV_FLTP_TO_FLT_2CH
+%if HAVE_AVX
+INIT_XMM avx
+CONV_FLTP_TO_FLT_2CH
+%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 51e51f5762..6e78338059 100644
--- a/libavresample/x86/audio_convert_init.c
+++ b/libavresample/x86/audio_convert_init.c
@@ -78,6 +78,11 @@ extern void ff_conv_fltp_to_s16_6ch_sse2(int16_t *dst, float *const *src,
 extern void ff_conv_fltp_to_s16_6ch_avx (int16_t *dst, float *const *src,
                                          int len, int channels);
 
+extern void ff_conv_fltp_to_flt_2ch_sse(float *dst, float *const *src, int len,
+                                        int channels);
+extern void ff_conv_fltp_to_flt_2ch_avx(float *dst, float *const *src, int len,
+                                        int channels);
+
 extern void ff_conv_fltp_to_flt_6ch_mmx (float *dst, float *const *src, int len,
                                          int channels);
 extern void ff_conv_fltp_to_flt_6ch_sse4(float *dst, float *const *src, int len,
@@ -99,6 +104,8 @@ av_cold void ff_audio_convert_init_x86(AudioConvert *ac)
     if (mm_flags & AV_CPU_FLAG_SSE && HAVE_SSE) {
         ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S16, AV_SAMPLE_FMT_FLTP,
                                   6, 1, 2, "SSE", ff_conv_fltp_to_s16_6ch_sse);
+        ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_FLTP,
+                                  2, 16, 8, "SSE", ff_conv_fltp_to_flt_2ch_sse);
     }
     if (mm_flags & AV_CPU_FLAG_SSE2 && HAVE_SSE) {
         if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) {