ARGBExtractAlpha 16 pixels at a time for ARM

arm64   8     TestARGBExtractAlpha (10019 ms) <-original 64 bit code
arm64   8 x2  TestARGBExtractAlpha (7639 ms)
arm64   16    TestARGBExtractAlpha (7369 ms) <- new 64 bit code
thumb32 8     TestARGBExtractAlpha (9505 ms) <- original 32 bit code
thumb32 8 x2  TestARGBExtractAlpha (7400 ms)
thumb32 8 x2i TestARGBExtractAlpha (7266 ms) <- new 32 bit code
arm32   8     TestARGBExtractAlpha (10002 ms)

BUG=libyuv:572
TESTED=local test on nexus 9
R=harryjin@google.com, wangcheng@google.com

Review URL: https://codereview.chromium.org/2035573002 .

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1593
+Version: 1594
 License: BSD
 License File: LICENSE
 

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1593
+#define LIBYUV_VERSION 1594
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

libyuv.gyp

@@ -61,6 +61,7 @@
             '-mfpu=vfp',
             '-mfpu=vfpv3',
             '-mfpu=vfpv3-d16',
+            # '-mthumb',  # arm32 not thumb
           ],
           'conditions': [
             # Disable LTO in libyuv_neon target due to gcc 4.9 compiler bug.
@@ -74,6 +75,7 @@
             ['target_arch != "arm64"', {
               'cflags': [
                 '-mfpu=neon',
+                # '-marm',  # arm32 not thumb
               ],
             }],
           ],

source/planar_functions.cc

@@ -2404,8 +2404,8 @@ int ARGBExtractAlpha(const uint8* src_argb, int src_stride,
 #endif
 #if defined(HAS_ARGBEXTRACTALPHAROW_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
-    ARGBExtractAlphaRow = IS_ALIGNED(width, 8) ? ARGBExtractAlphaRow_NEON
-                                               : ARGBExtractAlphaRow_Any_NEON;
+    ARGBExtractAlphaRow = IS_ALIGNED(width, 16) ? ARGBExtractAlphaRow_NEON
+                                                : ARGBExtractAlphaRow_Any_NEON;
   }
 #endif
 

source/row_any.cc

@@ -470,7 +470,7 @@ ANY11(ARGBAttenuateRow_Any_NEON, ARGBAttenuateRow_NEON, 0, 4, 4, 7)
 ANY11(ARGBExtractAlphaRow_Any_SSE2, ARGBExtractAlphaRow_SSE2, 0, 4, 1, 7)
 #endif
 #ifdef HAS_ARGBEXTRACTALPHAROW_NEON
-ANY11(ARGBExtractAlphaRow_Any_NEON, ARGBExtractAlphaRow_NEON, 0, 4, 1, 7)
+ANY11(ARGBExtractAlphaRow_Any_NEON, ARGBExtractAlphaRow_NEON, 0, 4, 1, 15)
 #endif
 #undef ANY11
 

source/row_neon.cc

@@ -1302,16 +1302,17 @@ void ARGBExtractAlphaRow_NEON(const uint8* src_argb, uint8* dst_a, int width) {
   asm volatile (
   "1:                                          \n"
     MEMACCESS(0)
-    "vld4.8     {d0, d1, d2, d3}, [%0]!        \n"  // load row 8 pixels
-    "subs       %2, %2, #8                     \n"  // 8 processed per loop
+    "vld4.8     {d0, d2, d4, d6}, [%0]!        \n"  // load 8 ARGB pixels
+    "vld4.8     {d1, d3, d5, d7}, [%0]!        \n"  // load next 8 ARGB pixels
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop
     MEMACCESS(1)
-    "vst1.8     {d3}, [%1]!                    \n"  // store 8 A's.
+    "vst1.8     {q3}, [%1]!                    \n"  // store 16 A's.
     "bgt       1b                              \n"
   : "+r"(src_argb),   // %0
     "+r"(dst_a),      // %1
     "+r"(width)       // %2
   :
-  : "cc", "memory", "d0", "d1", "d2", "d3"  // Clobber List
+  : "cc", "memory", "q0", "q1", "q2", "q3"  // Clobber List
   );
 }
 

source/row_neon64.cc

@@ -450,7 +450,6 @@ void I422ToARGB4444Row_NEON(const uint8* src_y,
 void I400ToARGBRow_NEON(const uint8* src_y,
                         uint8* dst_argb,
                         int width) {
-  int64 width64 = (int64)(width);
   asm volatile (
     YUVTORGB_SETUP
     "movi       v23.8b, #255                   \n"
@@ -463,7 +462,7 @@ void I400ToARGBRow_NEON(const uint8* src_y,
     "b.gt       1b                             \n"
     : "+r"(src_y),     // %0
       "+r"(dst_argb),  // %1
-      "+r"(width64)    // %2
+      "+r"(width)      // %2
     : [kUVToRB]"r"(&kYuvI601Constants.kUVToRB),
       [kUVToG]"r"(&kYuvI601Constants.kUVToG),
       [kUVBiasBGR]"r"(&kYuvI601Constants.kUVBiasBGR),
@@ -1404,10 +1403,10 @@ void ARGBExtractAlphaRow_NEON(const uint8* src_argb, uint8* dst_a, int width) {
   asm volatile (
   "1:                                          \n"
     MEMACCESS(0)
-    "ld4        {v0.8b,v1.8b,v2.8b,v3.8b}, [%0], #32 \n"  // load row 8 pixels
-    "subs       %w2, %w2, #8                   \n"  // 8 processed per loop
+    "ld4        {v0.16b,v1.16b,v2.16b,v3.16b}, [%0], #64 \n"  // load row 16 pixels
+    "subs       %w2, %w2, #16                  \n"  // 16 processed per loop
     MEMACCESS(1)
-    "st1        {v3.8b}, [%1], #8              \n"  // store 8 A's.
+    "st1        {v3.16b}, [%1], #16            \n"  // store 16 A's.
     "b.gt       1b                             \n"
   : "+r"(src_argb),   // %0
     "+r"(dst_a),      // %1