Optimize functions for LASX in row_lasx.cc.

1. Optimize 18 functions in source/row_lasx.cc file.
2. Make small modifications to LSX.
3. Remove some unnecessary content.

Bug: libyuv:912
Change-Id: Ifd1d85366efb9cdb3b99491e30fa450ff1848661
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/3507640
Reviewed-by: Mirko Bonadei <mbonadei@chromium.org>
Reviewed-by: Frank Barchard <fbarchard@chromium.org>
Commit-Queue: Frank Barchard <fbarchard@chromium.org>

1 files changed, 64 insertions, 0 deletions

diff --git a/source/convert_argb.cc b/source/convert_argb.cc
index f23b5d11..90f6c947 100644
--- a/source/convert_argb.cc
+++ b/source/convert_argb.cc
@@ -2886,6 +2886,14 @@ int RGB24ToARGB(const uint8_t* src_rgb24,
     }
   }
 #endif
+#if defined(HAS_RGB24TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RGB24ToARGBRow = RGB24ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RGB24ToARGBRow = RGB24ToARGBRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     RGB24ToARGBRow(src_rgb24, dst_argb, width);
@@ -2953,6 +2961,14 @@ int RAWToARGB(const uint8_t* src_raw,
     }
   }
 #endif
+#if defined(HAS_RAWTOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RAWToARGBRow = RAWToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RAWToARGBRow = RAWToARGBRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     RAWToARGBRow(src_raw, dst_argb, width);
@@ -3079,6 +3095,14 @@ int RGB565ToARGB(const uint8_t* src_rgb565,
     }
   }
 #endif
+#if defined(HAS_RGB565TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RGB565ToARGBRow = RGB565ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RGB565ToARGBRow = RGB565ToARGBRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     RGB565ToARGBRow(src_rgb565, dst_argb, width);
@@ -3154,6 +3178,14 @@ int ARGB1555ToARGB(const uint8_t* src_argb1555,
     }
   }
 #endif
+#if defined(HAS_ARGB1555TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGB1555ToARGBRow = ARGB1555ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGB1555ToARGBRow = ARGB1555ToARGBRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGB1555ToARGBRow(src_argb1555, dst_argb, width);
@@ -3229,6 +3261,14 @@ int ARGB4444ToARGB(const uint8_t* src_argb4444,
     }
   }
 #endif
+#if defined(HAS_ARGB4444TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ARGB4444ToARGBRow = ARGB4444ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ARGB4444ToARGBRow = ARGB4444ToARGBRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ARGB4444ToARGBRow(src_argb4444, dst_argb, width);
@@ -3516,6 +3556,14 @@ int NV12ToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_NV12TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    NV12ToARGBRow = NV12ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      NV12ToARGBRow = NV12ToARGBRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     NV12ToARGBRow(src_y, src_uv, dst_argb, yuvconstants, width);
@@ -3592,6 +3640,14 @@ int NV21ToARGBMatrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_NV21TOARGBROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    NV21ToARGBRow = NV21ToARGBRow_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      NV21ToARGBRow = NV21ToARGBRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     NV21ToARGBRow(src_y, src_vu, dst_argb, yuvconstants, width);
@@ -4340,6 +4396,14 @@ int NV12ToRGB565Matrix(const uint8_t* src_y,
     }
   }
 #endif
+#if defined(HAS_NV12TORGB565ROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    NV12ToRGB565Row = NV12ToRGB565Row_Any_LASX;
+    if (IS_ALIGNED(width, 16)) {
+      NV12ToRGB565Row = NV12ToRGB565Row_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     NV12ToRGB565Row(src_y, src_uv, dst_rgb565, yuvconstants, width);