lavr: Add x86-optimized function for s16 to s32 conversion

libavresample/x86/audio_convert.asm

 ;******************************************************************************
 ;* x86 optimized Format Conversion Utils
 ;* Copyright (c) 2008 Loren Merritt
+;* Copyright (c) 2012 Justin Ruggles <justin.ruggles@gmail.com>
 ;*
 ;* This file is part of Libav.
 ;*
@@ -24,6 +25,28 @@
 
 SECTION_TEXT
 
+;------------------------------------------------------------------------------
+; void ff_conv_s16_to_s32(int32_t *dst, const int16_t *src, int len);
+;------------------------------------------------------------------------------
+
+INIT_XMM sse2
+cglobal conv_s16_to_s32, 3,3,3, dst, src, len
+    lea      lenq, [2*lend]
+    lea      dstq, [dstq+2*lenq]
+    add      srcq, lenq
+    neg      lenq
+.loop:
+    mova       m2, [srcq+lenq]
+    pxor       m0, m0
+    pxor       m1, m1
+    punpcklwd  m0, m2
+    punpckhwd  m1, m2
+    mova  [dstq+2*lenq       ], m0
+    mova  [dstq+2*lenq+mmsize], m1
+    add      lenq, mmsize
+    jl .loop
+    REP_RET
+
 ;-----------------------------------------------------------------------------
 ; void ff_conv_fltp_to_flt_6ch(float *dst, float *const *src, int len,
 ;                              int channels);

libavresample/x86/audio_convert_init.c

@@ -22,6 +22,8 @@
 #include "libavutil/cpu.h"
 #include "libavresample/audio_convert.h"
 
+extern void ff_conv_s16_to_s32_sse2(int16_t *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_sse4(float *dst, float *const *src, int len);
 extern void ff_conv_fltp_to_flt_6ch_avx (float *dst, float *const *src, int len);
@@ -43,5 +45,9 @@ av_cold void ff_audio_convert_init_x86(AudioConvert *ac)
         ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_FLT, AV_SAMPLE_FMT_FLTP,
                                   6, 16, 4, "AVX", ff_conv_fltp_to_flt_6ch_avx);
     }
+    if (mm_flags & AV_CPU_FLAG_SSE2 && HAVE_SSE) {
+        ff_audio_convert_set_func(ac, AV_SAMPLE_FMT_S32, AV_SAMPLE_FMT_S16,
+                                  0, 16, 8, "SSE2", ff_conv_s16_to_s32_sse2);
+    }
 #endif
 }