Add MSA optimized ARGBToRGB565Row_MSA, ARGBToARGB1555Row_MSA,…

Add MSA optimized ARGBToRGB565Row_MSA, ARGBToARGB1555Row_MSA, ARGBToARGB4444Row_MSA, ARGBToUV444Row_MSA functions

R=fbarchard@google.com
BUG=libyuv:634

Performance Gain (vs C vectorized)
ARGBToRGB565Row_MSA       - ~1.6x
ARGBToRGB565Row_Any_MSA   - ~1.6x
ARGBToARGB1555Row_MSA     - ~1.3x
ARGBToARGB1555Row_Any_MSA - ~1.3x
ARGBToARGB4444Row_MSA     - ~3.8x
ARGBToARGB4444Row_Any_MSA - ~3.8x
ARGBToUV444Row_MSA        - ~2.4x
ARGBToUV444Row_Any_MSA    - ~2.4x

Performance Gain (vs C non-vectorized)
ARGBToRGB565Row_MSA       - ~2.8x
ARGBToRGB565Row_Any_MSA   - ~2.8x
ARGBToARGB1555Row_MSA     - ~2.2x
ARGBToARGB1555Row_Any_MSA - ~2.2x
ARGBToARGB4444Row_MSA     - ~6.8x
ARGBToARGB4444Row_Any_MSA - ~6.6x
ARGBToUV444Row_MSA        - ~6.7x
ARGBToUV444Row_Any_MSA    - ~6.7x

Review URL: https://codereview.chromium.org/2520003004 .

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1634
+Version: 1635
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -386,6 +386,10 @@ extern "C" {
 #define HAS_I422TORGB24ROW_MSA
 #define HAS_ARGBTORGB24ROW_MSA
 #define HAS_ARGBTORAWROW_MSA
+#define HAS_ARGBTORGB565ROW_MSA
+#define HAS_ARGBTOARGB1555ROW_MSA
+#define HAS_ARGBTOARGB4444ROW_MSA
+#define HAS_ARGBTOUV444ROW_MSA
 #endif
 
 #if defined(_MSC_VER) && !defined(__CLR_VER) && !defined(__clang__)
@@ -714,6 +718,10 @@ void ARGBToUVRow_NEON(const uint8* src_argb,
                       uint8* dst_u,
                       uint8* dst_v,
                       int width);
+void ARGBToUV444Row_MSA(const uint8* src_argb,
+                        uint8* dst_u,
+                        uint8* dst_v,
+                        int width);
 void ARGBToUVRow_MSA(const uint8* src_argb,
                      int src_stride_argb,
                      uint8* dst_u,
@@ -884,6 +892,10 @@ void ARGBToUVRow_Any_NEON(const uint8* src_argb,
                           uint8* dst_u,
                           uint8* dst_v,
                           int width);
+void ARGBToUV444Row_Any_MSA(const uint8* src_argb,
+                            uint8* dst_u,
+                            uint8* dst_v,
+                            int width);
 void ARGBToUVRow_Any_MSA(const uint8* src_argb,
                          int src_stride_argb,
                          uint8* dst_u,
@@ -1308,6 +1320,9 @@ void ARGBToRGB565DitherRow_NEON(const uint8* src_argb,
                                 int width);
 void ARGBToRGB24Row_MSA(const uint8* src_argb, uint8* dst_rgb, int width);
 void ARGBToRAWRow_MSA(const uint8* src_argb, uint8* dst_rgb, int width);
+void ARGBToRGB565Row_MSA(const uint8* src_argb, uint8* dst_rgb, int width);
+void ARGBToARGB1555Row_MSA(const uint8* src_argb, uint8* dst_rgb, int width);
+void ARGBToARGB4444Row_MSA(const uint8* src_argb, uint8* dst_rgb, int width);
 
 void ARGBToRGBARow_C(const uint8* src_argb, uint8* dst_rgb, int width);
 void ARGBToRGB24Row_C(const uint8* src_argb, uint8* dst_rgb, int width);
@@ -1898,6 +1913,13 @@ void ARGBToRGB565DitherRow_Any_NEON(const uint8* src_argb,
                                     int width);
 void ARGBToRGB24Row_Any_MSA(const uint8* src_argb, uint8* dst_rgb, int width);
 void ARGBToRAWRow_Any_MSA(const uint8* src_argb, uint8* dst_rgb, int width);
+void ARGBToRGB565Row_Any_MSA(const uint8* src_argb, uint8* dst_rgb, int width);
+void ARGBToARGB1555Row_Any_MSA(const uint8* src_argb,
+                               uint8* dst_rgb,
+                               int width);
+void ARGBToARGB4444Row_Any_MSA(const uint8* src_argb,
+                               uint8* dst_rgb,
+                               int width);
 
 void I444ToARGBRow_Any_NEON(const uint8* src_y,
                             const uint8* src_u,

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1634
+#define LIBYUV_VERSION 1635
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_

source/convert_from_argb.cc

@@ -68,6 +68,14 @@ int ARGBToI444(const uint8* src_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOUV444ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToUV444Row = ARGBToUV444Row_Any_MSA;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV444Row = ARGBToUV444Row_MSA;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
     ARGBToYRow = ARGBToYRow_Any_SSSE3;
@@ -1029,6 +1037,14 @@ int ARGBToRGB565(const uint8* src_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTORGB565ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToRGB565Row = ARGBToRGB565Row_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToRGB565Row = ARGBToRGB565Row_MSA;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGBToRGB565Row(src_argb, dst_rgb565, width);
@@ -1087,6 +1103,14 @@ int ARGBToARGB1555(const uint8* src_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOARGB1555ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToARGB1555Row = ARGBToARGB1555Row_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToARGB1555Row = ARGBToARGB1555Row_MSA;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGBToARGB1555Row(src_argb, dst_argb1555, width);
@@ -1145,6 +1169,14 @@ int ARGBToARGB4444(const uint8* src_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOARGB4444ROW_MSA)
+  if (TestCpuFlag(kCpuHasMSA)) {
+    ARGBToARGB4444Row = ARGBToARGB4444Row_Any_MSA;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToARGB4444Row = ARGBToARGB4444Row_MSA;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGBToARGB4444Row(src_argb, dst_argb4444, width);

source/row_any.cc

@@ -383,6 +383,9 @@ ANY11(I400ToARGBRow_Any_NEON, I400ToARGBRow_NEON, 0, 1, 4, 7)
 #if defined(HAS_ARGBTORGB24ROW_MSA)
 ANY11(ARGBToRGB24Row_Any_MSA, ARGBToRGB24Row_MSA, 0, 4, 3, 15)
 ANY11(ARGBToRAWRow_Any_MSA, ARGBToRAWRow_MSA, 0, 4, 3, 15)
+ANY11(ARGBToRGB565Row_Any_MSA, ARGBToRGB565Row_MSA, 0, 4, 2, 7)
+ANY11(ARGBToARGB1555Row_Any_MSA, ARGBToARGB1555Row_MSA, 0, 4, 2, 7)
+ANY11(ARGBToARGB4444Row_Any_MSA, ARGBToARGB4444Row_MSA, 0, 4, 2, 7)
 #endif
 #if defined(HAS_RAWTORGB24ROW_NEON)
 ANY11(RAWToRGB24Row_Any_NEON, RAWToRGB24Row_NEON, 0, 3, 3, 7)
@@ -786,6 +789,7 @@ ANY12(YUY2ToUV422Row_Any_NEON, YUY2ToUV422Row_NEON, 1, 4, 1, 15)
 ANY12(UYVYToUV422Row_Any_NEON, UYVYToUV422Row_NEON, 1, 4, 1, 15)
 #endif
 #ifdef HAS_YUY2TOUV422ROW_MSA
+ANY12(ARGBToUV444Row_Any_MSA, ARGBToUV444Row_MSA, 0, 4, 0, 15)
 ANY12(YUY2ToUV422Row_Any_MSA, YUY2ToUV422Row_MSA, 1, 4, 1, 31)
 ANY12(UYVYToUV422Row_Any_MSA, UYVYToUV422Row_MSA, 1, 4, 1, 31)
 #endif

source/row_msa.cc

@@ -788,6 +788,175 @@ void ARGBToRAWRow_MSA(const uint8* src_argb, uint8* dst_rgb, int width) {
   }
 }
 
+void ARGBToRGB565Row_MSA(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  v16u8 src0, src1, dst0;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v16i8 zero = {0};
+
+  for (x = 0; x < width; x += 8) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_argb, 16);
+    vec0 = (v16u8)__msa_srai_b((v16i8)src0, 3);
+    vec1 = (v16u8)__msa_slli_b((v16i8)src0, 3);
+    vec2 = (v16u8)__msa_srai_b((v16i8)src0, 5);
+    vec4 = (v16u8)__msa_srai_b((v16i8)src1, 3);
+    vec5 = (v16u8)__msa_slli_b((v16i8)src1, 3);
+    vec6 = (v16u8)__msa_srai_b((v16i8)src1, 5);
+    vec1 = (v16u8)__msa_sldi_b(zero, (v16i8)vec1, 1);
+    vec2 = (v16u8)__msa_sldi_b(zero, (v16i8)vec2, 1);
+    vec5 = (v16u8)__msa_sldi_b(zero, (v16i8)vec5, 1);
+    vec6 = (v16u8)__msa_sldi_b(zero, (v16i8)vec6, 1);
+    vec3 = (v16u8)__msa_sldi_b(zero, (v16i8)src0, 2);
+    vec7 = (v16u8)__msa_sldi_b(zero, (v16i8)src1, 2);
+    vec0 = __msa_binsli_b(vec0, vec1, 2);
+    vec1 = __msa_binsli_b(vec2, vec3, 4);
+    vec4 = __msa_binsli_b(vec4, vec5, 2);
+    vec5 = __msa_binsli_b(vec6, vec7, 4);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)vec1, (v16i8)vec0);
+    vec4 = (v16u8)__msa_ilvev_b((v16i8)vec5, (v16i8)vec4);
+    dst0 = (v16u8)__msa_pckev_h((v8i16)vec4, (v8i16)vec0);
+    ST_UB(dst0, dst_rgb);
+    src_argb += 32;
+    dst_rgb += 16;
+  }
+}
+
+void ARGBToARGB1555Row_MSA(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  v16u8 src0, src1, dst0;
+  v16u8 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7, vec8, vec9;
+  v16i8 zero = {0};
+
+  for (x = 0; x < width; x += 8) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_argb, 16);
+    vec0 = (v16u8)__msa_srai_b((v16i8)src0, 3);
+    vec1 = (v16u8)__msa_slli_b((v16i8)src0, 2);
+    vec2 = (v16u8)__msa_srai_b((v16i8)vec0, 3);
+    vec1 = (v16u8)__msa_sldi_b(zero, (v16i8)vec1, 1);
+    vec2 = (v16u8)__msa_sldi_b(zero, (v16i8)vec2, 1);
+    vec3 = (v16u8)__msa_srai_b((v16i8)src0, 1);
+    vec5 = (v16u8)__msa_srai_b((v16i8)src1, 3);
+    vec6 = (v16u8)__msa_slli_b((v16i8)src1, 2);
+    vec7 = (v16u8)__msa_srai_b((v16i8)vec5, 3);
+    vec6 = (v16u8)__msa_sldi_b(zero, (v16i8)vec6, 1);
+    vec7 = (v16u8)__msa_sldi_b(zero, (v16i8)vec7, 1);
+    vec8 = (v16u8)__msa_srai_b((v16i8)src1, 1);
+    vec3 = (v16u8)__msa_sldi_b(zero, (v16i8)vec3, 2);
+    vec8 = (v16u8)__msa_sldi_b(zero, (v16i8)vec8, 2);
+    vec4 = (v16u8)__msa_sldi_b(zero, (v16i8)src0, 3);
+    vec9 = (v16u8)__msa_sldi_b(zero, (v16i8)src1, 3);
+    vec0 = __msa_binsli_b(vec0, vec1, 2);
+    vec5 = __msa_binsli_b(vec5, vec6, 2);
+    vec1 = __msa_binsli_b(vec2, vec3, 5);
+    vec6 = __msa_binsli_b(vec7, vec8, 5);
+    vec1 = __msa_binsli_b(vec1, vec4, 0);
+    vec6 = __msa_binsli_b(vec6, vec9, 0);
+    vec0 = (v16u8)__msa_ilvev_b((v16i8)vec1, (v16i8)vec0);
+    vec1 = (v16u8)__msa_ilvev_b((v16i8)vec6, (v16i8)vec5);
+    dst0 = (v16u8)__msa_pckev_h((v8i16)vec1, (v8i16)vec0);
+    ST_UB(dst0, dst_rgb);
+    src_argb += 32;
+    dst_rgb += 16;
+  }
+}
+
+void ARGBToARGB4444Row_MSA(const uint8* src_argb, uint8* dst_rgb, int width) {
+  int x;
+  v16u8 src0, src1;
+  v16u8 vec0, vec1;
+  v16u8 dst0;
+  v16i8 zero = {0};
+
+  for (x = 0; x < width; x += 8) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_argb, 16);
+    vec0 = (v16u8)__msa_srai_b((v16i8)src0, 4);
+    vec1 = (v16u8)__msa_srai_b((v16i8)src1, 4);
+    src0 = (v16u8)__msa_sldi_b(zero, (v16i8)src0, 1);
+    src1 = (v16u8)__msa_sldi_b(zero, (v16i8)src1, 1);
+    vec0 = __msa_binsli_b(vec0, src0, 3);
+    vec1 = __msa_binsli_b(vec1, src1, 3);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    ST_UB(dst0, dst_rgb);
+    src_argb += 32;
+    dst_rgb += 16;
+  }
+}
+
+void ARGBToUV444Row_MSA(const uint8* src_argb,
+                        uint8* dst_u,
+                        uint8* dst_v,
+                        int32 width) {
+  int32 x;
+  v16u8 src0, src1, src2, src3, reg0, reg1, reg2, reg3, dst0, dst1;
+  v8u16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+  v8u16 vec8, vec9, vec10, vec11;
+  v8u16 const_112 = (v8u16)__msa_ldi_h(112);
+  v8u16 const_74 = (v8u16)__msa_ldi_h(74);
+  v8u16 const_38 = (v8u16)__msa_ldi_h(38);
+  v8u16 const_94 = (v8u16)__msa_ldi_h(94);
+  v8u16 const_18 = (v8u16)__msa_ldi_h(18);
+  v8u16 const_32896 = (v8u16)__msa_fill_h(32896);
+  v16i8 zero = {0};
+
+  for (x = width; x > 0; x -= 16) {
+    src0 = (v16u8)__msa_ld_b((v16i8*)src_argb, 0);
+    src1 = (v16u8)__msa_ld_b((v16i8*)src_argb, 16);
+    src2 = (v16u8)__msa_ld_b((v16i8*)src_argb, 32);
+    src3 = (v16u8)__msa_ld_b((v16i8*)src_argb, 48);
+    reg0 = (v16u8)__msa_pckev_b((v16i8)src1, (v16i8)src0);
+    reg1 = (v16u8)__msa_pckev_b((v16i8)src3, (v16i8)src2);
+    reg2 = (v16u8)__msa_pckod_b((v16i8)src1, (v16i8)src0);
+    reg3 = (v16u8)__msa_pckod_b((v16i8)src3, (v16i8)src2);
+    src0 = (v16u8)__msa_pckev_b((v16i8)reg1, (v16i8)reg0);
+    src1 = (v16u8)__msa_pckev_b((v16i8)reg3, (v16i8)reg2);
+    src2 = (v16u8)__msa_pckod_b((v16i8)reg1, (v16i8)reg0);
+    vec0 = (v8u16)__msa_ilvr_b(zero, (v16i8)src0);
+    vec1 = (v8u16)__msa_ilvl_b(zero, (v16i8)src0);
+    vec2 = (v8u16)__msa_ilvr_b(zero, (v16i8)src1);
+    vec3 = (v8u16)__msa_ilvl_b(zero, (v16i8)src1);
+    vec4 = (v8u16)__msa_ilvr_b(zero, (v16i8)src2);
+    vec5 = (v8u16)__msa_ilvl_b(zero, (v16i8)src2);
+    vec10 = vec0 * const_18;
+    vec11 = vec1 * const_18;
+    vec8 = vec2 * const_94;
+    vec9 = vec3 * const_94;
+    vec6 = vec4 * const_112;
+    vec7 = vec5 * const_112;
+    vec0 *= const_112;
+    vec1 *= const_112;
+    vec2 *= const_74;
+    vec3 *= const_74;
+    vec4 *= const_38;
+    vec5 *= const_38;
+    vec8 += vec10;
+    vec9 += vec11;
+    vec6 += const_32896;
+    vec7 += const_32896;
+    vec0 += const_32896;
+    vec1 += const_32896;
+    vec2 += vec4;
+    vec3 += vec5;
+    vec0 -= vec2;
+    vec1 -= vec3;
+    vec6 -= vec8;
+    vec7 -= vec9;
+    vec0 = (v8u16)__msa_srai_h((v8i16)vec0, 8);
+    vec1 = (v8u16)__msa_srai_h((v8i16)vec1, 8);
+    vec6 = (v8u16)__msa_srai_h((v8i16)vec6, 8);
+    vec7 = (v8u16)__msa_srai_h((v8i16)vec7, 8);
+    dst0 = (v16u8)__msa_pckev_b((v16i8)vec1, (v16i8)vec0);
+    dst1 = (v16u8)__msa_pckev_b((v16i8)vec7, (v16i8)vec6);
+    ST_UB(dst0, dst_u);
+    ST_UB(dst1, dst_v);
+    src_argb += 64;
+    dst_u += 16;
+    dst_v += 16;
+  }
+}
+
 void ARGB4444ToARGBRow_MSA(const uint8* src_argb4444,
                            uint8* dst_argb,
                            int width) {