Add optimization functions in row_lsx.cc file.

Optimize 44 functions in source/row_lsx.cc file.
All test cases passed on loongarch platform.

Bug: libyuv:913
Change-Id: Ic80a5751314adc2e9bd435f2bbd928ab017a90f9
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/3351467
Reviewed-by: Frank Barchard <fbarchard@chromium.org>

1 files changed, 104 insertions, 0 deletions

diff --git a/source/convert_argb.cc b/source/convert_argb.cc
index 7128e9f9..1da455c2 100644
--- a/source/convert_argb.cc
+++ b/source/convert_argb.cc
@@ -605,6 +605,14 @@ int I444ToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_I444TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    I444ToARGBRow = I444ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      I444ToARGBRow = I444ToARGBRow_LSX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     I444ToARGBRow(src_y, src_u, src_v, dst_argb, yuvconstants, width);
@@ -2731,6 +2739,14 @@ int I400ToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_I400TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    I400ToARGBRow = I400ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      I400ToARGBRow = I400ToARGBRow_LSX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     I400ToARGBRow(src_y, dst_argb, yuvconstants, width);
@@ -2818,6 +2834,14 @@ int J400ToARGB(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_J400TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    J400ToARGBRow = J400ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      J400ToARGBRow = J400ToARGBRow_LSX;
+    }
+  }
+#endif
   for (y = 0; y < height; ++y) {
     J400ToARGBRow(src_y, dst_argb, width);
     src_y += src_stride_y;
@@ -2972,6 +2996,14 @@ int RGB24ToARGB(const uint8_t* src_rgb24,
     }
   }
 #endif
+#if defined(HAS_RGB24TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_LSX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     RGB24ToARGBRow(src_rgb24, dst_argb, width);
@@ -3039,6 +3071,14 @@ int RAWToARGB(const uint8_t* src_raw,
     }
   }
 #endif
+#if defined(HAS_RAWTOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RAWToARGBRow = RAWToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToARGBRow = RAWToARGBRow_LSX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     RAWToARGBRow(src_raw, dst_argb, width);
@@ -3165,6 +3205,14 @@ int RGB565ToARGB(const uint8_t* src_rgb565,
     }
   }
 #endif
+#if defined(HAS_RGB565TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_LSX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     RGB565ToARGBRow(src_rgb565, dst_argb, width);
@@ -3240,6 +3288,14 @@ int ARGB1555ToARGB(const uint8_t* src_argb1555,
     }
   }
 #endif
+#if defined(HAS_ARGB1555TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_LSX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGB1555ToARGBRow(src_argb1555, dst_argb, width);
@@ -3315,6 +3371,14 @@ int ARGB4444ToARGB(const uint8_t* src_argb4444,
     }
   }
 #endif
+#if defined(HAS_ARGB4444TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_LSX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGB4444ToARGBRow(src_argb4444, dst_argb, width);
@@ -3602,6 +3666,14 @@ int NV12ToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_NV12TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToARGBRow = NV12ToARGBRow_LSX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     NV12ToARGBRow(src_y, src_uv, dst_argb, yuvconstants, width);
@@ -3678,6 +3750,14 @@ int NV21ToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_NV21TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    NV21ToARGBRow = NV21ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 8)) {
+      NV21ToARGBRow = NV21ToARGBRow_LSX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     NV21ToARGBRow(src_y, src_vu, dst_argb, yuvconstants, width);
@@ -4066,6 +4146,14 @@ int YUY2ToARGB(const uint8_t* src_yuy2,
     }
   }
 #endif
+#if defined(HAS_YUY2TOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    YUY2ToARGBRow = YUY2ToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 8)) {
+      YUY2ToARGBRow = YUY2ToARGBRow_LSX;
+    }
+  }
+#endif
   for (y = 0; y < height; ++y) {
     YUY2ToARGBRow(src_yuy2, dst_argb, &kYuvI601Constants, width);
     src_yuy2 += src_stride_yuy2;
@@ -4141,6 +4229,14 @@ int UYVYToARGB(const uint8_t* src_uyvy,
     }
   }
 #endif
+#if defined(HAS_UYVYTOARGBROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    UYVYToARGBRow = UYVYToARGBRow_Any_LSX;
+    if (IS_ALIGNED(width, 8)) {
+      UYVYToARGBRow = UYVYToARGBRow_LSX;
+    }
+  }
+#endif
   for (y = 0; y < height; ++y) {
     UYVYToARGBRow(src_uyvy, dst_argb, &kYuvI601Constants, width);
     src_uyvy += src_stride_uyvy;
@@ -4450,6 +4546,14 @@ int NV12ToRGB565Matrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_NV12TORGB565ROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_LSX;
+    if (IS_ALIGNED(width, 8)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_LSX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     NV12ToRGB565Row(src_y, src_uv, dst_rgb565, yuvconstants, width);