vpx_fdct16x16_1_sse2: improve load pattern

load the full row rather than doing 2 8-wide columns

Change-Id: I7a1c0cba06b0dc1ae86046410922b1efccb95c95

diff --git a/vpx_dsp/x86/fwd_txfm_sse2.c b/vpx_dsp/x86/fwd_txfm_sse2.c
index 62c18d3b4ad6bf9423400df5b9fdff4b6e4fdde1..e4deeecaeb0530acfcf601cbc98148ac1fffea86 100644 (file)
--- a/vpx_dsp/x86/fwd_txfm_sse2.c
+++ b/vpx_dsp/x86/fwd_txfm_sse2.c
@@ -91,40 +91,39 @@ void vpx_fdct16x16_1_sse2(const int16_t *input, tran_low_t *output,
   int i;
 
   for (i = 0; i < 2; ++i) {
-    input += 8 * i;
-    in0  = _mm_load_si128((const __m128i *)(input +  0 * stride));
-    in1  = _mm_load_si128((const __m128i *)(input +  1 * stride));
-    in2  = _mm_load_si128((const __m128i *)(input +  2 * stride));
-    in3  = _mm_load_si128((const __m128i *)(input +  3 * stride));
+    in0  = _mm_load_si128((const __m128i *)(input + 0 * stride + 0));
+    in1  = _mm_load_si128((const __m128i *)(input + 0 * stride + 8));
+    in2  = _mm_load_si128((const __m128i *)(input + 1 * stride + 0));
+    in3  = _mm_load_si128((const __m128i *)(input + 1 * stride + 8));
 
     u0 = _mm_add_epi16(in0, in1);
     u1 = _mm_add_epi16(in2, in3);
     sum = _mm_add_epi16(sum, u0);
 
-    in0  = _mm_load_si128((const __m128i *)(input +  4 * stride));
-    in1  = _mm_load_si128((const __m128i *)(input +  5 * stride));
-    in2  = _mm_load_si128((const __m128i *)(input +  6 * stride));
-    in3  = _mm_load_si128((const __m128i *)(input +  7 * stride));
+    in0  = _mm_load_si128((const __m128i *)(input + 2 * stride + 0));
+    in1  = _mm_load_si128((const __m128i *)(input + 2 * stride + 8));
+    in2  = _mm_load_si128((const __m128i *)(input + 3 * stride + 0));
+    in3  = _mm_load_si128((const __m128i *)(input + 3 * stride + 8));
 
     sum = _mm_add_epi16(sum, u1);
     u0  = _mm_add_epi16(in0, in1);
     u1  = _mm_add_epi16(in2, in3);
     sum = _mm_add_epi16(sum, u0);
 
-    in0  = _mm_load_si128((const __m128i *)(input +  8 * stride));
-    in1  = _mm_load_si128((const __m128i *)(input +  9 * stride));
-    in2  = _mm_load_si128((const __m128i *)(input + 10 * stride));
-    in3  = _mm_load_si128((const __m128i *)(input + 11 * stride));
+    in0  = _mm_load_si128((const __m128i *)(input + 4 * stride + 0));
+    in1  = _mm_load_si128((const __m128i *)(input + 4 * stride + 8));
+    in2  = _mm_load_si128((const __m128i *)(input + 5 * stride + 0));
+    in3  = _mm_load_si128((const __m128i *)(input + 5 * stride + 8));
 
     sum = _mm_add_epi16(sum, u1);
     u0  = _mm_add_epi16(in0, in1);
     u1  = _mm_add_epi16(in2, in3);
     sum = _mm_add_epi16(sum, u0);
 
-    in0  = _mm_load_si128((const __m128i *)(input + 12 * stride));
-    in1  = _mm_load_si128((const __m128i *)(input + 13 * stride));
-    in2  = _mm_load_si128((const __m128i *)(input + 14 * stride));
-    in3  = _mm_load_si128((const __m128i *)(input + 15 * stride));
+    in0  = _mm_load_si128((const __m128i *)(input + 6 * stride + 0));
+    in1  = _mm_load_si128((const __m128i *)(input + 6 * stride + 8));
+    in2  = _mm_load_si128((const __m128i *)(input + 7 * stride + 0));
+    in3  = _mm_load_si128((const __m128i *)(input + 7 * stride + 8));
 
     sum = _mm_add_epi16(sum, u1);
     u0  = _mm_add_epi16(in0, in1);
@@ -132,6 +131,7 @@ void vpx_fdct16x16_1_sse2(const int16_t *input, tran_low_t *output,
     sum = _mm_add_epi16(sum, u0);
 
     sum = _mm_add_epi16(sum, u1);
+    input += 8 * stride;
   }
 
   u0  = _mm_setzero_si128();