Row AArch64 Neon implementation - Part 5

BUG=319
TESTED=libyuv_unittest
R=fbarchard@chromium.org, fbarchard@google.com

Change-Id: Ia76096088ddd771388f01dd86110089db2faedfc
Signed-off-by: Ashok Bhat <ashok.bhat@arm.com>

Review URL: https://webrtc-codereview.appspot.com/21189004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@1055 16f28f9a-4ce2-e073-06de-1de4eb20be90

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1054
+Version: 1055
 License: BSD
 License File: LICENSE
 

include/libyuv/row.h

@@ -301,8 +301,8 @@ extern "C" {
 // #define HAS_ARGBTORGB565ROW_NEON
 // #define HAS_ARGBTOARGB1555ROW_NEON
 // #define HAS_ARGBTOARGB4444ROW_NEON
-// #define HAS_ARGBTOYROW_NEON
-// #define HAS_ARGBTOYJROW_NEON
+#define HAS_ARGBTOYROW_NEON
+#define HAS_ARGBTOYJROW_NEON
 // #define HAS_ARGBTOUV444ROW_NEON
 // #define HAS_ARGBTOUV422ROW_NEON
 // #define HAS_ARGBTOUV411ROW_NEON

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1054
+#define LIBYUV_VERSION 1055
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/convert.cc

@@ -711,13 +711,15 @@ int ARGBToI420(const uint8* src_argb, int src_stride_argb,
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
-    if (width >= 16) {
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
     ARGBToUVRow = ARGBToUVRow_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       ARGBToUVRow = ARGBToUVRow_NEON;
     }
   }
-  }
 #endif
 
   for (y = 0; y < height - 1; y += 2) {

source/convert_from_argb.cc

@@ -60,6 +60,13 @@ int ARGBToI444(const uint8* src_argb, int src_stride_argb,
         }
       }
   }
+#elif defined(HAS_ARGBTOUV444ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
+    ARGBToUV444Row = ARGBToUV444Row_Any_NEON;
+    if (IS_ALIGNED(width, 8)) {
+      ARGBToUV444Row = ARGBToUV444Row_NEON;
+    }
+  }
 #endif
 #if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
@@ -76,10 +83,8 @@ int ARGBToI444(const uint8* src_argb, int src_stride_argb,
 #elif defined(HAS_ARGBTOYROW_NEON)
   if (TestCpuFlag(kCpuHasNEON) && width >= 8) {
     ARGBToYRow = ARGBToYRow_Any_NEON;
-    ARGBToUV444Row = ARGBToUV444Row_Any_NEON;
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
-      ARGBToUV444Row = ARGBToUV444Row_NEON;
     }
   }
 #endif
@@ -134,6 +139,13 @@ int ARGBToI422(const uint8* src_argb, int src_stride_argb,
       }
     }
   }
+#elif defined(HAS_ARGBTOUV422ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_NEON;
+    }
+  }
 #endif
 
 #if defined(HAS_ARGBTOYROW_SSSE3)
@@ -153,12 +165,6 @@ int ARGBToI422(const uint8* src_argb, int src_stride_argb,
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
-    if (width >= 16) {
-      ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        ARGBToUV422Row = ARGBToUV422Row_NEON;
-      }
-    }
   }
 #endif
 
@@ -228,13 +234,15 @@ int ARGBToI411(const uint8* src_argb, int src_stride_argb,
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
-    if (width >= 32) {
+  }
+#endif
+#if defined(HAS_ARGBTOUV411ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 32) {
     ARGBToUV411Row = ARGBToUV411Row_Any_NEON;
     if (IS_ALIGNED(width, 32)) {
       ARGBToUV411Row = ARGBToUV411Row_NEON;
     }
   }
-  }
 #endif
 
   for (y = 0; y < height; ++y) {
@@ -296,13 +304,15 @@ int ARGBToNV12(const uint8* src_argb, int src_stride_argb,
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
-    if (width >= 16) {
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
     ARGBToUVRow = ARGBToUVRow_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       ARGBToUVRow = ARGBToUVRow_NEON;
     }
   }
-  }
 #endif
 #if defined(HAS_MERGEUVROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2) && halfwidth >= 16) {
@@ -399,13 +409,15 @@ int ARGBToNV21(const uint8* src_argb, int src_stride_argb,
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
-    if (width >= 16) {
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
     ARGBToUVRow = ARGBToUVRow_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       ARGBToUVRow = ARGBToUVRow_NEON;
     }
   }
-  }
 #endif
 #if defined(HAS_MERGEUVROW_SSE2)
   if (TestCpuFlag(kCpuHasSSE2) && halfwidth >= 16) {
@@ -493,6 +505,13 @@ int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
       }
     }
   }
+#elif defined(HAS_ARGBTOUV422ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_NEON;
+    }
+  }
 #endif
 #if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
@@ -510,12 +529,6 @@ int ARGBToYUY2(const uint8* src_argb, int src_stride_argb,
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
-    if (width >= 16) {
-      ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        ARGBToUV422Row = ARGBToUV422Row_NEON;
-      }
-    }
   }
 #endif
 
@@ -594,6 +607,13 @@ int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
       }
     }
   }
+#elif defined(HAS_ARGBTOUV422ROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
+    ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
+    if (IS_ALIGNED(width, 16)) {
+      ARGBToUV422Row = ARGBToUV422Row_NEON;
+    }
+  }
 #endif
 #if defined(HAS_ARGBTOYROW_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3) && width >= 16) {
@@ -611,12 +631,6 @@ int ARGBToUYVY(const uint8* src_argb, int src_stride_argb,
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
-    if (width >= 16) {
-      ARGBToUV422Row = ARGBToUV422Row_Any_NEON;
-      if (IS_ALIGNED(width, 16)) {
-        ARGBToUV422Row = ARGBToUV422Row_NEON;
-      }
-    }
   }
 #endif
 
@@ -1022,13 +1036,15 @@ int ARGBToJ420(const uint8* src_argb, int src_stride_argb,
     if (IS_ALIGNED(width, 8)) {
       ARGBToYJRow = ARGBToYJRow_NEON;
     }
-    if (width >= 16) {
+  }
+#endif
+#if defined(HAS_ARGBTOUVJROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
     ARGBToUVJRow = ARGBToUVJRow_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       ARGBToUVJRow = ARGBToUVJRow_NEON;
     }
   }
-  }
 #endif
 
   for (y = 0; y < height - 1; y += 2) {

source/format_conversion.cc

@@ -332,13 +332,15 @@ int BayerToI420(const uint8* src_bayer, int src_stride_bayer,
     if (IS_ALIGNED(width, 8)) {
       ARGBToYRow = ARGBToYRow_NEON;
     }
-    if (width >= 16) {
+  }
+#endif
+#if defined(HAS_ARGBTOUVROW_NEON)
+  if (TestCpuFlag(kCpuHasNEON) && width >= 16) {
     ARGBToUVRow = ARGBToUVRow_Any_NEON;
     if (IS_ALIGNED(width, 16)) {
       ARGBToUVRow = ARGBToUVRow_NEON;
     }
   }
-  }
 #endif
 
   switch (src_fourcc_bayer) {

source/row_neon64.cc

@@ -1580,28 +1580,28 @@ void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_argb4444,
 #ifdef HAS_ARGBTOYROW_NEON
 void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
   asm volatile (
-    "vmov.u8    d24, #13                       \n"  // B * 0.1016 coefficient
-    "vmov.u8    d25, #65                       \n"  // G * 0.5078 coefficient
-    "vmov.u8    d26, #33                       \n"  // R * 0.2578 coefficient
-    "vmov.u8    d27, #16                       \n"  // Add 16 constant
+    "movi       v4.8b, #13                     \n"  // B * 0.1016 coefficient
+    "movi       v5.8b, #65                     \n"  // G * 0.5078 coefficient
+    "movi       v6.8b, #33                     \n"  // R * 0.2578 coefficient
+    "movi       v7.8b, #16                     \n"  // Add 16 constant
     ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "ld4        {v0.8b-v3.8b}, [%0], #32       \n"  // load 8 ARGB pixels.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q2, d0, d24                    \n"  // B
-    "vmlal.u8   q2, d1, d25                    \n"  // G
-    "vmlal.u8   q2, d2, d26                    \n"  // R
-    "vqrshrun.s16 d0, q2, #7                   \n"  // 16 bit to 8 bit Y
-    "vqadd.u8   d0, d27                        \n"
+    "umull      v3.8h, v0.8b, v4.8b            \n"  // B
+    "umlal      v3.8h, v1.8b, v5.8b            \n"  // G
+    "umlal      v3.8h, v2.8b, v6.8b            \n"  // R
+    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 16 bit to 8 bit Y
+    "uqadd      v0.8b, v0.8b, v7.8b            \n"
     MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
     "bgt        1b                             \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_y),     // %1
     "+r"(pix)        // %2
   :
-  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7"
   );
 }
 #endif  // HAS_ARGBTOYROW_NEON
@@ -1609,26 +1609,26 @@ void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
 #ifdef HAS_ARGBTOYJROW_NEON
 void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix) {
   asm volatile (
-    "vmov.u8    d24, #15                       \n"  // B * 0.11400 coefficient
-    "vmov.u8    d25, #75                       \n"  // G * 0.58700 coefficient
-    "vmov.u8    d26, #38                       \n"  // R * 0.29900 coefficient
+    "movi       v4.8b, #15                     \n"  // B * 0.11400 coefficient
+    "movi       v5.8b, #75                     \n"  // G * 0.58700 coefficient
+    "movi       v6.8b, #38                     \n"  // R * 0.29900 coefficient
     ".p2align   2                              \n"
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load 8 ARGB pixels.
+    "ld4        {v0.8b-v3.8b}, [%0], #32       \n"  // load 8 ARGB pixels.
     "subs       %2, %2, #8                     \n"  // 8 processed per loop.
-    "vmull.u8   q2, d0, d24                    \n"  // B
-    "vmlal.u8   q2, d1, d25                    \n"  // G
-    "vmlal.u8   q2, d2, d26                    \n"  // R
-    "vqrshrun.s16 d0, q2, #7                   \n"  // 15 bit to 8 bit Y
+    "umull      v3.8h, v0.8b, v4.8b            \n"  // B
+    "umlal      v3.8h, v1.8b, v5.8b            \n"  // G
+    "umlal      v3.8h, v2.8b, v6.8b            \n"  // R
+    "sqrshrun   v0.8b, v3.8h, #7               \n"  // 15 bit to 8 bit Y
     MEMACCESS(1)
-    "vst1.8     {d0}, [%1]!                    \n"  // store 8 pixels Y.
+    "st1        {v0.8b}, [%1], #8              \n"  // store 8 pixels Y.
     "bgt        1b                             \n"
   : "+r"(src_argb),  // %0
     "+r"(dst_y),     // %1
     "+r"(pix)        // %2
   :
-  : "cc", "memory", "q0", "q1", "q2", "q12", "q13"
+  : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6"
   );
 }
 #endif  // HAS_ARGBTOYJROW_NEON