SSE2 optimizations for Monkey's Audio decoder vector functions

Originally committed as revision 14161 to svn://svn.ffmpeg.org/ffmpeg/trunk

libavcodec/i386/dsputil_mmx.c

@@ -2092,6 +2092,79 @@ extern void ff_snow_inner_add_yblock_sse2(const uint8_t *obmc, const int obmc_st
 extern void ff_snow_inner_add_yblock_mmx(const uint8_t *obmc, const int obmc_stride, uint8_t * * block, int b_w, int b_h,
                           int src_x, int src_y, int src_stride, slice_buffer * sb, int add, uint8_t * dst8);
 
+
+static void add_int16_sse2(int16_t * v1, int16_t * v2, int order)
+{
+    x86_reg o = -(order << 1);
+    v1 += order;
+    v2 += order;
+    asm volatile(
+        "1:                          \n\t"
+        "movdqu   (%1,%2),   %%xmm0  \n\t"
+        "movdqu 16(%1,%2),   %%xmm1  \n\t"
+        "paddw    (%0,%2),   %%xmm0  \n\t"
+        "paddw  16(%0,%2),   %%xmm1  \n\t"
+        "movdqa   %%xmm0,    (%0,%2) \n\t"
+        "movdqa   %%xmm1,  16(%0,%2) \n\t"
+        "add      $32,       %2      \n\t"
+        "js       1b                 \n\t"
+        : "+r"(v1), "+r"(v2), "+r"(o)
+    );
+}
+
+static void sub_int16_sse2(int16_t * v1, int16_t * v2, int order)
+{
+    x86_reg o = -(order << 1);
+    v1 += order;
+    v2 += order;
+    asm volatile(
+        "1:                           \n\t"
+        "movdqa    (%0,%2),   %%xmm0  \n\t"
+        "movdqa  16(%0,%2),   %%xmm2  \n\t"
+        "movdqu    (%1,%2),   %%xmm1  \n\t"
+        "movdqu  16(%1,%2),   %%xmm3  \n\t"
+        "psubw     %%xmm1,    %%xmm0  \n\t"
+        "psubw     %%xmm3,    %%xmm2  \n\t"
+        "movdqa    %%xmm0,    (%0,%2) \n\t"
+        "movdqa    %%xmm2,  16(%0,%2) \n\t"
+        "add       $32,       %2      \n\t"
+        "js        1b                 \n\t"
+        : "+r"(v1), "+r"(v2), "+r"(o)
+    );
+}
+
+static int32_t scalarproduct_int16_sse2(int16_t * v1, int16_t * v2, int order, int shift)
+{
+    int res = 0;
+    DECLARE_ALIGNED_16(int64_t, sh);
+    x86_reg o = -(order << 1);
+
+    v1 += order;
+    v2 += order;
+    sh = shift;
+    asm volatile(
+        "pxor      %%xmm7,  %%xmm7        \n\t"
+        "1:                               \n\t"
+        "movdqu    (%0,%3), %%xmm0        \n\t"
+        "movdqu  16(%0,%3), %%xmm1        \n\t"
+        "pmaddwd   (%1,%3), %%xmm0        \n\t"
+        "pmaddwd 16(%1,%3), %%xmm1        \n\t"
+        "paddd     %%xmm0,  %%xmm7        \n\t"
+        "paddd     %%xmm1,  %%xmm7        \n\t"
+        "add       $32,     %3            \n\t"
+        "js        1b                     \n\t"
+        "movhlps   %%xmm7,  %%xmm2        \n\t"
+        "paddd     %%xmm2,  %%xmm7        \n\t"
+        "psrad     %4,      %%xmm7        \n\t"
+        "pshuflw   $0x4E,   %%xmm7,%%xmm2 \n\t"
+        "paddd     %%xmm2,  %%xmm7        \n\t"
+        "movd      %%xmm7,  %2            \n\t"
+        : "+r"(v1), "+r"(v2), "=r"(res), "+r"(o)
+        : "m"(sh)
+    );
+    return res;
+}
+
 void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
 {
     mm_flags = mm_support();
@@ -2463,6 +2536,11 @@ void dsputil_init_mmx(DSPContext* c, AVCodecContext *avctx)
         }
         if(mm_flags & MM_3DNOW)
             c->vector_fmul_add_add = vector_fmul_add_add_3dnow; // faster than sse
+        if(mm_flags & MM_SSE2){
+            c->add_int16 = add_int16_sse2;
+            c->sub_int16 = sub_int16_sse2;
+            c->scalarproduct_int16 = scalarproduct_int16_sse2;
+        }
     }
 
     if (ENABLE_ENCODERS)