SSE2 version of vp9_short_fdct32x32_rd.

43,000 -> 5,750 cycles, about 7.5x faster.

Change-Id: Ibfd92821b9603f4ed9c256e0ececec14fa4565d0

vp9/common/vp9_idct.h

@@ -25,7 +25,10 @@
 #define pair_set_epi16(a, b) \
   _mm_set1_epi32(((uint16_t)(a)) + (((uint16_t)(b)) << 16))
 
-// Constants are round(16384 * cos(k*Pi/64)) where k = 1 to 31.
+// Constants:
+//  for (int i = 1; i< 32; ++i)
+//    printf("static const int cospi_%d_64 = %.0f;\n", i,
+//           round(16384 * cos(i*M_PI/64)));
 // Note: sin(k*Pi/64) = cos((32-k)*Pi/64)
 static const int cospi_1_64  = 16364;
 static const int cospi_2_64  = 16305;

vp9/common/vp9_rtcd_defs.sh

@@ -600,7 +600,7 @@ prototype void vp9_short_fdct32x32 "int16_t *InputData, int16_t *OutputData, int
 specialize vp9_short_fdct32x32
 
 prototype void vp9_short_fdct32x32_rd "int16_t *InputData, int16_t *OutputData, int pitch"
-specialize vp9_short_fdct32x32_rd
+specialize vp9_short_fdct32x32_rd sse2
 
 prototype void vp9_short_fdct16x16 "int16_t *InputData, int16_t *OutputData, int pitch"
 specialize vp9_short_fdct16x16 sse2

vp9/encoder/vp9_dct.c

@@ -1366,6 +1366,9 @@ void vp9_short_fdct32x32_rd_c(int16_t *input, int16_t *out, int pitch) {
       temp_in[j] = input[j * shortpitch + i] << 2;
     dct32_1d(temp_in, temp_out, 0);
     for (j = 0; j < 32; ++j)
+      // TODO(cd): see quality impact of only doing
+      //           output[j * 32 + i] = (temp_out[j] + 1) >> 2;
+      //           PS: also change code in vp9/encoder/x86/vp9_dct_sse2.c
       output[j * 32 + i] = (temp_out[j] + 1 + (temp_out[j] > 0)) >> 2;
   }