Attenuate and Unattenuate Any variations for sse2, sss3 and avx2

BUG=190
TESTED=out\release\libyuv_unittest --gtest_filter=*Unatt*
Review URL: https://webrtc-codereview.appspot.com/1121005

git-svn-id: http://libyuv.googlecode.com/svn/trunk@579 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 578
+Version: 579
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -1318,6 +1318,8 @@ void ARGBAttenuateRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
                                int width);
 void ARGBAttenuateRow_Any_SSSE3(const uint8* src_argb, uint8* dst_argb,
                                 int width);
+void ARGBAttenuateRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+                               int width);
 void ARGBAttenuateRow_Any_NEON(const uint8* src_argb, uint8* dst_argb,
                                int width);
 
@@ -1326,6 +1328,10 @@ extern uint32 fixed_invtbl8[256];
 void ARGBUnattenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width);
 void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
 void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
+void ARGBUnattenuateRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
+                                 int width);
+void ARGBUnattenuateRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
+                                 int width);
 
 void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width);
 void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width);

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 578
+#define LIBYUV_VERSION 579
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/planar_functions.cc

@@ -1034,9 +1034,12 @@ int ARGBAttenuate(const uint8* src_argb, int src_stride_argb,
 #endif
 #if defined(HAS_ARGBATTENUATEROW_AVX2)
   bool clear = false;
-  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 8)) {
-    bool clear = true;
-    ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    clear = true;
+    ARGBAttenuateRow = ARGBAttenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBAttenuateRow = ARGBAttenuateRow_AVX2;
+    }
   }
 #endif
 #if defined(HAS_ARGBATTENUATEROW_NEON)
@@ -1077,19 +1080,25 @@ int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
     src_stride_argb = -src_stride_argb;
   }
   void (*ARGBUnattenuateRow)(const uint8* src_argb, uint8* dst_argb,
-                             int width) = ARGBUnattenuateRow_C;
+                           int width) = ARGBUnattenuateRow_C;
 #if defined(HAS_ARGBUNATTENUATEROW_SSE2)
-  if (TestCpuFlag(kCpuHasSSE2) && IS_ALIGNED(width, 4) &&
+  if (TestCpuFlag(kCpuHasSSE2) && width >= 4 &&
       IS_ALIGNED(src_argb, 16) && IS_ALIGNED(src_stride_argb, 16) &&
       IS_ALIGNED(dst_argb, 16) && IS_ALIGNED(dst_stride_argb, 16)) {
-    ARGBUnattenuateRow = ARGBUnattenuateRow_SSE2;
+    ARGBUnattenuateRow = ARGBUnattenuateRow_Any_SSE2;
+    if (IS_ALIGNED(width, 4)) {
+      ARGBUnattenuateRow = ARGBUnattenuateRow_SSE2;
+    }
   }
 #endif
 #if defined(HAS_ARGBUNATTENUATEROW_AVX2)
   bool clear = false;
-  if (TestCpuFlag(kCpuHasAVX2) && IS_ALIGNED(width, 8)) {
-    bool clear = true;
-    ARGBUnattenuateRow = ARGBUnattenuateRow_AVX2;
+  if (TestCpuFlag(kCpuHasAVX2) && width >= 8) {
+    clear = true;
+    ARGBUnattenuateRow = ARGBUnattenuateRow_Any_AVX2;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBUnattenuateRow = ARGBUnattenuateRow_AVX2;
+    }
   }
 #endif
 // TODO(fbarchard): Neon version.
@@ -1099,6 +1108,13 @@ int ARGBUnattenuate(const uint8* src_argb, int src_stride_argb,
     src_argb += src_stride_argb;
     dst_argb += dst_stride_argb;
   }
+
+#if defined(HAS_ARGBUNATTENUATEROW_AVX2)
+  if (clear) {
+    __asm vzeroupper;
+  }
+#endif
+
   return 0;
 }
 

source/row_any.cc

@@ -250,6 +250,18 @@ YANY(ARGBAttenuateRow_Any_SSSE3, ARGBAttenuateRow_SSSE3, ARGBAttenuateRow_C,
 YANY(ARGBAttenuateRow_Any_SSE2, ARGBAttenuateRow_SSE2, ARGBAttenuateRow_C,
      4, 4, 3)
 #endif
+#ifdef HAS_ARGBUNATTENUATEROW_SSE2
+YANY(ARGBUnattenuateRow_Any_SSE2, ARGBUnattenuateRow_SSE2, ARGBUnattenuateRow_C,
+     4, 4, 3)
+#endif
+#ifdef HAS_ARGBATTENUATEROW_AVX2
+YANY(ARGBAttenuateRow_Any_AVX2, ARGBAttenuateRow_AVX2, ARGBAttenuateRow_C,
+     4, 4, 7)
+#endif
+#ifdef HAS_ARGBUNATTENUATEROW_AVX2
+YANY(ARGBUnattenuateRow_Any_AVX2, ARGBUnattenuateRow_AVX2, ARGBUnattenuateRow_C,
+     4, 4, 7)
+#endif
 #ifdef HAS_ARGBATTENUATEROW_NEON
 YANY(ARGBAttenuateRow_Any_NEON, ARGBAttenuateRow_NEON, ARGBAttenuateRow_C,
      4, 4, 7)

unit_test/planar_test.cc

@@ -59,9 +59,9 @@ TEST_F(libyuvTest, TestAttenuate) {
   EXPECT_EQ(255u, unatten_pixels[0][1]);
   EXPECT_EQ(254u, unatten_pixels[0][2]);
   EXPECT_EQ(128u, unatten_pixels[0][3]);
-  EXPECT_EQ(16u, unatten_pixels[1][0]);
-  EXPECT_EQ(64u, unatten_pixels[1][1]);
-  EXPECT_EQ(192u, unatten_pixels[1][2]);
+  EXPECT_EQ(0u, unatten_pixels[1][0]);
+  EXPECT_EQ(0u, unatten_pixels[1][1]);
+  EXPECT_EQ(0u, unatten_pixels[1][2]);
   EXPECT_EQ(0u, unatten_pixels[1][3]);
   EXPECT_EQ(16u, unatten_pixels[2][0]);
   EXPECT_EQ(64u, unatten_pixels[2][1]);