Refine some functions on the Longarch platform.

Add ARGBToYMatrixRow_LSX/LASX, RGBAToYMatrixRow_LSX/LASX and
RGBToYMatrixRow_LSX/LASX functions with RgbConstants argument.

Bug: libyuv:912
Change-Id: I956e639d1f0da4a47a55b79c9d41dcd29e29bdc5
Reviewed-on: https://chromium-review.googlesource.com/c/libyuv/libyuv/+/4167860
Reviewed-by: Frank Barchard <fbarchard@chromium.org>
Reviewed-by: Mirko Bonadei <mbonadei@chromium.org>
Commit-Queue: Frank Barchard <fbarchard@chromium.org>

1 files changed, 157 insertions, 10 deletions

diff --git a/source/convert_from_argb.cc b/source/convert_from_argb.cc
index f7eab0c6..d548aec2 100644
--- a/source/convert_from_argb.cc
+++ b/source/convert_from_argb.cc
@@ -116,6 +116,14 @@ int ARGBToI444(const uint8_t* src_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYRow = ARGBToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_LSX;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYROW_LASX)
   if (TestCpuFlag(kCpuHasLASX)) {
     ARGBToYRow = ARGBToYRow_Any_LASX;
@@ -230,7 +238,14 @@ int ARGBToI422(const uint8_t* src_argb,
     }
   }
 #endif
-
+#if defined(HAS_ARGBTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYRow = ARGBToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_LSX;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
   if (TestCpuFlag(kCpuHasLASX)) {
     ARGBToYRow = ARGBToYRow_Any_LASX;
@@ -340,6 +355,14 @@ int ARGBToNV12(const uint8_t* src_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYRow = ARGBToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_LSX;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
   if (TestCpuFlag(kCpuHasLASX)) {
     ARGBToYRow = ARGBToYRow_Any_LASX;
@@ -502,6 +525,14 @@ int ARGBToNV21(const uint8_t* src_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYRow = ARGBToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_LSX;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
   if (TestCpuFlag(kCpuHasLASX)) {
     ARGBToYRow = ARGBToYRow_Any_LASX;
@@ -663,6 +694,22 @@ int ABGRToNV12(const uint8_t* src_abgr,
     }
   }
 #endif
+#if defined(HAS_ABGRTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ABGRToYRow = ABGRToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYRow = ABGRToYRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ABGRToYRow = ABGRToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToYRow = ABGRToYRow_LASX;
+    }
+  }
+#endif
 #if defined(HAS_MERGEUVROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     MergeUVRow_ = MergeUVRow_Any_SSE2;
@@ -815,6 +862,22 @@ int ABGRToNV21(const uint8_t* src_abgr,
     }
   }
 #endif
+#if defined(HAS_ABGRTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ABGRToYRow = ABGRToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYRow = ABGRToYRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ABGRToYRow = ABGRToYRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToYRow = ABGRToYRow_LASX;
+    }
+  }
+#endif
 #if defined(HAS_MERGEUVROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2)) {
     MergeUVRow_ = MergeUVRow_Any_SSE2;
@@ -972,6 +1035,14 @@ int ARGBToYUY2(const uint8_t* src_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYRow = ARGBToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_LSX;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
   if (TestCpuFlag(kCpuHasLASX)) {
     ARGBToYRow = ARGBToYRow_Any_LASX;
@@ -1135,6 +1206,14 @@ int ARGBToUYVY(const uint8_t* src_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYRow = ARGBToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_LSX;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYROW_LASX) && defined(HAS_ARGBTOUVROW_LASX)
   if (TestCpuFlag(kCpuHasLASX)) {
     ARGBToYRow = ARGBToYRow_Any_LASX;
@@ -1262,6 +1341,14 @@ int ARGBToI400(const uint8_t* src_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOYROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYRow = ARGBToYRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYRow = ARGBToYRow_LSX;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYROW_LASX)
   if (TestCpuFlag(kCpuHasLASX)) {
     ARGBToYRow = ARGBToYRow_Any_LASX;
@@ -1939,6 +2026,16 @@ int ARGBToJ420(const uint8_t* src_argb,
     }
   }
 #endif
+#if defined(HAS_ARGBTOYJROW_LSX) && defined(HAS_ARGBTOUVJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ARGBToYJRow = ARGBToYJRow_Any_LSX;
+    ARGBToUVJRow = ARGBToUVJRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToYJRow = ARGBToYJRow_LSX;
+      ARGBToUVJRow = ARGBToUVJRow_LSX;
+    }
+  }
+#endif
 #if defined(HAS_ARGBTOYJROW_LASX) && defined(HAS_ARGBTOUVJROW_LASX)
   if (TestCpuFlag(kCpuHasLASX)) {
     ARGBToYJRow = ARGBToYJRow_Any_LASX;
@@ -2215,6 +2312,22 @@ int RGBAToJ400(const uint8_t* src_rgba,
     }
   }
 #endif
+#if defined(HAS_RGBATOYJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RGBAToYJRow = RGBAToYJRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RGBAToYJRow = RGBAToYJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RGBATOYJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RGBAToYJRow = RGBAToYJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RGBAToYJRow = RGBAToYJRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     RGBAToYJRow(src_rgba, dst_yj, width);
@@ -2309,13 +2422,19 @@ int ABGRToJ420(const uint8_t* src_abgr,
     }
   }
 #endif
-#if defined(HAS_ABGRTOYJROW_LSX) && defined(HAS_ABGRTOUVJROW_LSX)
+#if defined(HAS_ABGRTOYJROW_LSX)
   if (TestCpuFlag(kCpuHasLSX)) {
     ABGRToYJRow = ABGRToYJRow_Any_LSX;
-    ABGRToUVJRow = ABGRToUVJRow_Any_LSX;
     if (IS_ALIGNED(width, 16)) {
       ABGRToYJRow = ABGRToYJRow_LSX;
-      ABGRToUVJRow = ABGRToUVJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ABGRToYJRow = ABGRToYJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToYJRow = ABGRToYJRow_LASX;
     }
   }
 #endif
@@ -2430,23 +2549,19 @@ int ABGRToJ422(const uint8_t* src_abgr,
     }
   }
 #endif
-#if defined(HAS_ABGRTOYJROW_LSX) && defined(HAS_ABGRTOUVJROW_LSX)
+#if defined(HAS_ABGRTOYJROW_LSX)
   if (TestCpuFlag(kCpuHasLSX)) {
     ABGRToYJRow = ABGRToYJRow_Any_LSX;
-    ABGRToUVJRow = ABGRToUVJRow_Any_LSX;
     if (IS_ALIGNED(width, 16)) {
       ABGRToYJRow = ABGRToYJRow_LSX;
-      ABGRToUVJRow = ABGRToUVJRow_LSX;
     }
   }
 #endif
-#if defined(HAS_ABGRTOYJROW_LASX) && defined(HAS_ABGRTOUVJROW_LASX)
+#if defined(HAS_ABGRTOYJROW_LASX)
   if (TestCpuFlag(kCpuHasLASX)) {
     ABGRToYJRow = ABGRToYJRow_Any_LASX;
-    ABGRToUVJRow = ABGRToUVJRow_Any_LASX;
     if (IS_ALIGNED(width, 32)) {
       ABGRToYJRow = ABGRToYJRow_LASX;
-      ABGRToUVJRow = ABGRToUVJRow_LASX;
     }
   }
 #endif
@@ -2519,6 +2634,22 @@ int ABGRToJ400(const uint8_t* src_abgr,
     }
   }
 #endif
+#if defined(HAS_ABGRTOYJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    ABGRToYJRow = ABGRToYJRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      ABGRToYJRow = ABGRToYJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_ABGRTOYJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    ABGRToYJRow = ABGRToYJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      ABGRToYJRow = ABGRToYJRow_LASX;
+    }
+  }
+#endif
 
   for (y = 0; y < height; ++y) {
     ABGRToYJRow(src_abgr, dst_yj, width);
@@ -2713,6 +2844,22 @@ int RAWToJNV21(const uint8_t* src_raw,
     }
   }
 #endif
+#if defined(HAS_RAWTOYJROW_LSX)
+  if (TestCpuFlag(kCpuHasLSX)) {
+    RAWToYJRow = RAWToYJRow_Any_LSX;
+    if (IS_ALIGNED(width, 16)) {
+      RAWToYJRow = RAWToYJRow_LSX;
+    }
+  }
+#endif
+#if defined(HAS_RAWTOYJROW_LASX)
+  if (TestCpuFlag(kCpuHasLASX)) {
+    RAWToYJRow = RAWToYJRow_Any_LASX;
+    if (IS_ALIGNED(width, 32)) {
+      RAWToYJRow = RAWToYJRow_LASX;
+    }
+  }
+#endif
 
 // Other platforms do intermediate conversion from RAW to ARGB.
 #else  // HAS_RAWTOYJROW