swresample/resample: mmx2/sse2 int16 linear interpolation

About three times faster
Signed-off-by: James Almer <jamrial@gmail.com>
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

libswresample/resample_template.c

@@ -81,9 +81,11 @@
 #        define RENAME(N) N ## _int16
 #    elif defined(TEMPLATE_RESAMPLE_S16_MMX2)
 #        define COMMON_CORE COMMON_CORE_INT16_MMX2
+#        define LINEAR_CORE LINEAR_CORE_INT16_MMX2
 #        define RENAME(N) N ## _int16_mmx2
 #    elif defined(TEMPLATE_RESAMPLE_S16_SSE2)
 #        define COMMON_CORE COMMON_CORE_INT16_SSE2
+#        define LINEAR_CORE LINEAR_CORE_INT16_SSE2
 #        define RENAME(N) N ## _int16_sse2
 #    endif
 
@@ -163,10 +165,14 @@ int RENAME(swri_resample)(ResampleContext *c, DELEM *dst, const DELEM *src, int
                 OUT(dst[dst_index], val);
             }else if(c->linear){
                 FELEM2 v2=0;
+#ifdef LINEAR_CORE
+                LINEAR_CORE
+#else
                 for(i=0; i<c->filter_length; i++){
                     val += src[sample_index + i] * (FELEM2)filter[i];
                     v2  += src[sample_index + i] * (FELEM2)filter[i + c->filter_alloc];
                 }
+#endif
                 val+=(v2-val)*(FELEML)frac / c->src_incr;
                 OUT(dst[dst_index], val);
             }else{
@@ -213,6 +219,7 @@ int RENAME(swri_resample)(ResampleContext *c, DELEM *dst, const DELEM *src, int
 }
 
 #undef COMMON_CORE
+#undef LINEAR_CORE
 #undef RENAME
 #undef FILTER_SHIFT
 #undef DELEM

libswresample/x86/resample_mmx.h

@@ -50,6 +50,34 @@ __asm__ volatile(\
       NAMED_CONSTRAINTS_ADD(ff_resample_int16_rounder)\
 );
 
+#define LINEAR_CORE_INT16_MMX2 \
+    x86_reg len= -2*c->filter_length;\
+__asm__ volatile(\
+    "pxor          %%mm0, %%mm0 \n\t"\
+    "pxor          %%mm2, %%mm2 \n\t"\
+    "1:                         \n\t"\
+    "movq       (%3, %0), %%mm1 \n\t"\
+    "movq          %%mm1, %%mm3 \n\t"\
+    "pmaddwd    (%4, %0), %%mm1 \n\t"\
+    "pmaddwd    (%5, %0), %%mm3 \n\t"\
+    "paddd         %%mm1, %%mm0 \n\t"\
+    "paddd         %%mm3, %%mm2 \n\t"\
+    "add              $8, %0    \n\t"\
+    " js 1b                     \n\t"\
+    "pshufw $0x0E, %%mm0, %%mm1 \n\t"\
+    "pshufw $0x0E, %%mm2, %%mm3 \n\t"\
+    "paddd         %%mm1, %%mm0 \n\t"\
+    "paddd         %%mm3, %%mm2 \n\t"\
+    "movd          %%mm0, %1    \n\t"\
+    "movd          %%mm2, %2    \n\t"\
+    : "+r" (len),\
+      "=r" (val),\
+      "=r" (v2)\
+    : "r" (((uint8_t*)(src+sample_index))-len),\
+      "r" (((uint8_t*)filter)-len),\
+      "r" (((uint8_t*)(filter+c->filter_alloc))-len)\
+);
+
 #define COMMON_CORE_INT16_SSE2 \
     x86_reg len= -2*c->filter_length;\
 __asm__ volatile(\
@@ -74,6 +102,39 @@ __asm__ volatile(\
       NAMED_CONSTRAINTS_ADD(ff_resample_int16_rounder)\
 );
 
+#define LINEAR_CORE_INT16_SSE2 \
+    x86_reg len= -2*c->filter_length;\
+__asm__ volatile(\
+    "pxor          %%xmm0, %%xmm0 \n\t"\
+    "pxor          %%xmm2, %%xmm2 \n\t"\
+    "1:                           \n\t"\
+    "movdqu      (%3, %0), %%xmm1 \n\t"\
+    "movdqa        %%xmm1, %%xmm3 \n\t"\
+    "pmaddwd     (%4, %0), %%xmm1 \n\t"\
+    "pmaddwd     (%5, %0), %%xmm3 \n\t"\
+    "paddd         %%xmm1, %%xmm0 \n\t"\
+    "paddd         %%xmm3, %%xmm2 \n\t"\
+    "add              $16, %0     \n\t"\
+    " js 1b                       \n\t"\
+    "pshufd $0x0E, %%xmm0, %%xmm1 \n\t"\
+    "pshufd $0x0E, %%xmm2, %%xmm3 \n\t"\
+    "paddd         %%xmm1, %%xmm0 \n\t"\
+    "paddd         %%xmm3, %%xmm2 \n\t"\
+    "pshufd $0x01, %%xmm0, %%xmm1 \n\t"\
+    "pshufd $0x01, %%xmm2, %%xmm3 \n\t"\
+    "paddd         %%xmm1, %%xmm0 \n\t"\
+    "paddd         %%xmm3, %%xmm2 \n\t"\
+    "movd          %%xmm0, %1     \n\t"\
+    "movd          %%xmm2, %2     \n\t"\
+    : "+r" (len),\
+      "=r" (val),\
+      "=r" (v2)\
+    : "r" (((uint8_t*)(src+sample_index))-len),\
+      "r" (((uint8_t*)filter)-len),\
+      "r" (((uint8_t*)(filter+c->filter_alloc))-len)\
+    XMM_CLOBBERS_ONLY("%xmm0", "%xmm1", "%xmm2", "%xmm3")\
+);
+
 #define COMMON_CORE_FLT_SSE \
     x86_reg len= -4*c->filter_length;\
 __asm__ volatile(\