Add MSA optimized I422ToYUY2Row, I422ToUYVYRow functions

source/row_msa.cc

Index: source/row_msa.cc
diff --git a/source/row_msa.cc b/source/row_msa.cc
index b86865cf315307f1c8ddf0e09444819f5849974b..52a246cdb2c6a0779e6789dd46f01bbde29077ff 100644
--- a/source/row_msa.cc
+++ b/source/row_msa.cc
@@ -53,6 +53,54 @@ void ARGBMirrorRow_MSA(const uint8* src, uint8* dst, int width) {
   }
 }
 
+void I422ToYUY2Row_MSA(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_yuy2,
+                       int width) {
+  int x;
+  v16u8 src_u0, src_v0, src_y0, src_y1, vec_uv0, vec_uv1;
+  v16u8 dst_yuy2_0, dst_yuy2_1, dst_yuy2_2, dst_yuy2_3;
+
+  for (x = 0; x < width; x += 32) {
+    src_u0 = LD_UB(src_u);
+    src_v0 = LD_UB(src_v);
+    LD_UB2(src_y, 16, src_y0, src_y1);
+    ILVRL_B2_UB(src_v0, src_u0, vec_uv0, vec_uv1);

[fbarchard1, 2016/09/29 17:10:43]

Did you try just 1 vector at a time instead of 2?  Is it improving performance
much to unroll all instructions to 2?

[manojkumar.bhosale, 2016/09/30 09:00:29]

all the loop unrolling we have done is based on load latencies and to reduce
un-necessary stalls. Here, byte vector holds 16 elements of u and v after load,
so we needed 2 loads of y and hence all subsequent operations with 2
instructions.
Operating 1 y vector at a time is possible but then we will be wasting half of
the u and v elements getting loaded anyway, will insert more stalls which will
be un-optimal.

+    ILVRL_B2_UB(vec_uv0, src_y0, dst_yuy2_0, dst_yuy2_1);
+    ILVRL_B2_UB(vec_uv1, src_y1, dst_yuy2_2, dst_yuy2_3);
+    ST_UB4(dst_yuy2_0, dst_yuy2_1, dst_yuy2_2, dst_yuy2_3, dst_yuy2, 16);
+    src_u += 16;
+    src_v += 16;
+    src_y += 32;
+    dst_yuy2 += 64;
+  }
+}
+
+void I422ToUYVYRow_MSA(const uint8* src_y,
+                       const uint8* src_u,
+                       const uint8* src_v,
+                       uint8* dst_uyvy,
+                       int width) {
+  int x;
+  v16u8 src_u0, src_v0, src_y0, src_y1, vec_uv0, vec_uv1;
+  v16u8 dst_uyvy0, dst_uyvy1, dst_uyvy2, dst_uyvy3;
+
+  for (x = 0; x < width; x += 32) {
+    src_u0 = LD_UB(src_u);
+    src_v0 = LD_UB(src_v);
+    LD_UB2(src_y, 16, src_y0, src_y1);
+    ILVRL_B2_UB(src_v0, src_u0, vec_uv0, vec_uv1);
+    ILVRL_B2_UB(src_y0, vec_uv0, dst_uyvy0, dst_uyvy1);
+    ILVRL_B2_UB(src_y1, vec_uv1, dst_uyvy2, dst_uyvy3);
+    ST_UB4(dst_uyvy0, dst_uyvy1, dst_uyvy2, dst_uyvy3, dst_uyvy, 16);
+    src_u += 16;
+    src_v += 16;
+    src_y += 32;
+    dst_uyvy += 64;
+  }
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 }  // namespace libyuv