swresample: add swri_resample_float_sse

At least two times faster than the C version.
Signed-off-by: James Almer <jamrial@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

libswresample/resample.c

@@ -302,6 +302,12 @@ static int set_compensation(ResampleContext *c, int sample_delta, int compensati
 #include "resample_template.c"
 #undef TEMPLATE_RESAMPLE_S16_MMX2
 
+#if HAVE_SSE_INLINE
+#define TEMPLATE_RESAMPLE_FLT_SSE
+#include "resample_template.c"
+#undef TEMPLATE_RESAMPLE_FLT_SSE
+#endif
+
 #if HAVE_SSE2_INLINE
 #define TEMPLATE_RESAMPLE_S16_SSE2
 #include "resample_template.c"
@@ -328,6 +334,10 @@ static int multiple_resample(ResampleContext *c, AudioData *dst, int dst_size, A
 #endif
              if(c->format == AV_SAMPLE_FMT_S16P) ret= swri_resample_int16(c, (int16_t*)dst->ch[i], (const int16_t*)src->ch[i], consumed, src_size, dst_size, i+1==dst->ch_count);
         else if(c->format == AV_SAMPLE_FMT_S32P) ret= swri_resample_int32(c, (int32_t*)dst->ch[i], (const int32_t*)src->ch[i], consumed, src_size, dst_size, i+1==dst->ch_count);
+#if HAVE_SSE_INLINE
+        else if(c->format == AV_SAMPLE_FMT_FLTP && (mm_flags&AV_CPU_FLAG_SSE))
+                                                 ret= swri_resample_float_sse (c, (float*)dst->ch[i], (const float*)src->ch[i], consumed, src_size, dst_size, i+1==dst->ch_count);
+#endif
         else if(c->format == AV_SAMPLE_FMT_FLTP) ret= swri_resample_float(c, (float  *)dst->ch[i], (const float  *)src->ch[i], consumed, src_size, dst_size, i+1==dst->ch_count);
         else if(c->format == AV_SAMPLE_FMT_DBLP) ret= swri_resample_double(c,(double *)dst->ch[i], (const double *)src->ch[i], consumed, src_size, dst_size, i+1==dst->ch_count);
     }

libswresample/resample_template.c

@@ -34,8 +34,9 @@
 #    define FELEML double
 #    define OUT(d, v) d = v
 
-#elif defined(TEMPLATE_RESAMPLE_FLT)
-#    define RENAME(N) N ## _float
+#elif    defined(TEMPLATE_RESAMPLE_FLT)     \
+      || defined(TEMPLATE_RESAMPLE_FLT_SSE)
+
 #    define FILTER_SHIFT 0
 #    define DELEM  float
 #    define FELEM  float
@@ -43,6 +44,13 @@
 #    define FELEML float
 #    define OUT(d, v) d = v
 
+#    if defined(TEMPLATE_RESAMPLE_FLT)
+#        define RENAME(N) N ## _float
+#    elif defined(TEMPLATE_RESAMPLE_FLT_SSE)
+#        define COMMON_CORE COMMON_CORE_FLT_SSE
+#        define RENAME(N) N ## _float_sse
+#    endif
+
 #elif defined(TEMPLATE_RESAMPLE_S32)
 #    define RENAME(N) N ## _int32
 #    define FILTER_SHIFT 30

libswresample/x86/resample_mmx.h

@@ -24,6 +24,7 @@
 
 int swri_resample_int16_mmx2 (struct ResampleContext *c, int16_t *dst, const int16_t *src, int *consumed, int src_size, int dst_size, int update_ctx);
 int swri_resample_int16_sse2 (struct ResampleContext *c, int16_t *dst, const int16_t *src, int *consumed, int src_size, int dst_size, int update_ctx);
+int swri_resample_float_sse  (struct ResampleContext *c,   float *dst, const   float *src, int *consumed, int src_size, int dst_size, int update_ctx);
 
 DECLARE_ALIGNED(16, const uint64_t, ff_resample_int16_rounder)[2]    = { 0x0000000000004000ULL, 0x0000000000000000ULL};
 
@@ -72,3 +73,25 @@ __asm__ volatile(\
       "r" (dst+dst_index)\
       NAMED_CONSTRAINTS_ADD(ff_resample_int16_rounder)\
 );
+
+#define COMMON_CORE_FLT_SSE \
+    x86_reg len= -4*c->filter_length;\
+__asm__ volatile(\
+    "xorps     %%xmm0, %%xmm0     \n\t"\
+    "1:                           \n\t"\
+    "movups  (%1, %0), %%xmm1     \n\t"\
+    "mulps   (%2, %0), %%xmm1     \n\t"\
+    "addps     %%xmm1, %%xmm0     \n\t"\
+    "add       $16, %0            \n\t"\
+    " js 1b                       \n\t"\
+    "movhlps   %%xmm0, %%xmm1     \n\t"\
+    "addps     %%xmm1, %%xmm0     \n\t"\
+    "movss     %%xmm0, %%xmm1     \n\t"\
+    "shufps $1, %%xmm0, %%xmm0    \n\t"\
+    "addps     %%xmm1, %%xmm0     \n\t"\
+    "movss     %%xmm0, (%3)       \n\t"\
+    : "+r" (len)\
+    : "r" (((uint8_t*)(src+sample_index))-len),\
+      "r" (((uint8_t*)filter)-len),\
+      "r" (dst+dst_index)\
+);