Adding SSE2 optimized vp9_short_idct32x32_1_add function.

Change-Id: I4b1c6bb9ff615f5872b96ed07dbf0f5e18e63643

vp9/common/vp9_rtcd_defs.sh

@@ -311,7 +311,7 @@ prototype void vp9_short_idct32x32_add "int16_t *input, uint8_t *dest, int dest_
 specialize vp9_short_idct32x32_add sse2 neon
 
 prototype void vp9_short_idct32x32_1_add "int16_t *input, uint8_t *dest, int dest_stride"
-specialize vp9_short_idct32x32_1_add
+specialize vp9_short_idct32x32_1_add sse2
 
 prototype void vp9_short_iht4x4_add "int16_t *input, uint8_t *dest, int dest_stride, int tx_type"
 specialize vp9_short_iht4x4_add sse2 neon

vp9/common/x86/vp9_idct_intrin_sse2.c

@@ -3549,3 +3549,51 @@ void vp9_short_idct32x32_add_sse2(int16_t *input, uint8_t *dest, int stride) {
     }
   }
 }  //NOLINT
+
+void vp9_short_idct32x32_1_add_sse2(int16_t *input, uint8_t *dest, int stride) {
+  __m128i dc_value;
+  const __m128i zero = _mm_setzero_si128();
+  int a, i;
+
+  a = dct_const_round_shift(input[0] * cospi_16_64);
+  a = dct_const_round_shift(a * cospi_16_64);
+  a = ROUND_POWER_OF_TWO(a, 6);
+
+  dc_value = _mm_set1_epi16(a);
+
+  for (i = 0; i < 4; ++i) {
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    RECON_AND_STORE(dest, dc_value);
+    dest += 8 - (stride * 32);
+  }
+}