Add MSA optimized TransposeWx8_MSA and TransposeUVWx8_MSA functions

include/libyuv/macros_msa.h

Index: include/libyuv/macros_msa.h
diff --git a/include/libyuv/macros_msa.h b/include/libyuv/macros_msa.h
index 9be804e11cb3514a32170cbb6274a88dbf8990da..53de2b8a39599f934282e4a62bdbff82100a8f95 100644
--- a/include/libyuv/macros_msa.h
+++ b/include/libyuv/macros_msa.h
@@ -50,6 +50,42 @@
     val_m;                                                             \
   })
 #endif  // (__mips == 64)
+
+#define SW(val, pdst)                                \
+  ({                                                 \
+    uint8_t* pdst_sw_m = (uint8_t*)(pdst);           \
+    uint32_t val_m = (val);                          \
+                                                     \

[fbarchard1, 2016/12/12 19:59:18]

remove blank lines from macros.

[manojkumar.bhosale, 2016/12/13 08:52:22]

Done.

+    asm volatile("sw  %[val_m],  %[pdst_sw_m]  \n\t" \
+                                                     \
+                 : [pdst_sw_m] "=m"(*pdst_sw_m)      \
+                 : [val_m] "r"(val_m));              \
+  })
+
+#if (__mips == 64)
+#define SD(val, pdst)                                \
+  ({                                                 \
+    uint8_t* pdst_sd_m = (uint8_t*)(pdst);           \
+    uint64_t val_m = (val);                          \
+                                                     \
+    asm volatile("sd  %[val_m],  %[pdst_sd_m]  \n\t" \
+                                                     \
+                 : [pdst_sd_m] "=m"(*pdst_sd_m)      \
+                 : [val_m] "r"(val_m));              \
+  })
+#else  // !(__mips == 64)
+#define SD(val, pdst)                                        \
+  ({                                                         \
+    uint8_t* pdst_sd_m = (uint8_t*)(pdst);                   \
+    uint32_t val0_m, val1_m;                                 \
+                                                             \
+    val0_m = (uint32_t)((val)&0x00000000FFFFFFFF);           \
+    val1_m = (uint32_t)(((val) >> 32) & 0x00000000FFFFFFFF); \
+                                                             \
+    SW(val0_m, pdst_sd_m);                                   \
+    SW(val1_m, pdst_sd_m + 4);                               \
+  })
+#endif  // !(__mips == 64)
 #else   // !(__mips_isa_rev >= 6)
 #define LW(psrc)                                      \
   ({                                                  \
@@ -85,6 +121,29 @@
     val_m;                                                             \
   })
 #endif  // (__mips == 64)
+
+#define SW(val, pdst)                                 \
+  ({                                                  \
+    uint8_t* pdst_sw_m = (uint8_t*)(pdst);            \
+    uint32_t val_m = (val);                           \
+                                                      \
+    asm volatile("usw  %[val_m],  %[pdst_sw_m]  \n\t" \

[fbarchard1, 2016/12/12 19:59:18]

remove \t ?  I've never found it necessary.

[manojkumar.bhosale, 2016/12/13 08:52:22]

Done.

+                                                      \
+                 : [pdst_sw_m] "=m"(*pdst_sw_m)       \
+                 : [val_m] "r"(val_m));               \
+  })
+
+#define SD(val, pdst)                                        \
+  ({                                                         \
+    uint8_t* pdst_sd_m = (uint8_t*)(pdst);                   \
+    uint32_t val0_m, val1_m;                                 \
+                                                             \
+    val0_m = (uint32_t)((val)&0x00000000FFFFFFFF);           \
+    val1_m = (uint32_t)(((val) >> 32) & 0x00000000FFFFFFFF); \
+                                                             \
+    SW(val0_m, pdst_sd_m);                                   \
+    SW(val1_m, pdst_sd_m + 4);                               \
+  })
 #endif  // (__mips_isa_rev >= 6)
 
 // TODO(fbarchard): Consider removing __VAR_ARGS versions.