Add ScaleAddRows_NEON for ARM32/64

ARM32/64 NEON versions of ScaleAddRows_NEON are implemented.

BUG=319
TESTED=libyuvTest.Scale* on ARM32/64 with Android
R=fbarchard@google.com

Change-Id: I45b88c2b5f576042ba5b3d8d6f8851257fdb7218

Review URL: https://webrtc-codereview.appspot.com/46379004

git-svn-id: http://libyuv.googlecode.com/svn/trunk@1326 16f28f9a-4ce2-e073-06de-1de4eb20be90

include/libyuv/scale_row.h

@@ -51,6 +51,7 @@ extern "C" {
 #define HAS_SCALEROWDOWN38_NEON
 #define HAS_SCALEARGBROWDOWNEVEN_NEON
 #define HAS_SCALEARGBROWDOWN2_NEON
+#define HAS_SCALEADDROWS_NEON
 #endif
 
 // The following are available on Mips platforms:
@@ -305,6 +306,9 @@ void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
                                ptrdiff_t src_stride,
                                uint8* dst_ptr, int dst_width);
 
+void ScaleAddRows_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                       uint16* dst_ptr, int src_width, int src_height);
+
 void ScaleRowDown2_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,
                               uint8* dst, int dst_width);
 void ScaleRowDown2Box_MIPS_DSPR2(const uint8* src_ptr, ptrdiff_t src_stride,

source/scale.cc

@@ -758,6 +758,12 @@ static void ScalePlaneBox(int src_width, int src_height,
     }
 #endif
 
+#if defined(HAS_SCALEADDROWS_NEON)
+    if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(src_width, 16)) {
+      ScaleAddRows = ScaleAddRows_NEON;
+    }
+#endif
+
     for (j = 0; j < dst_height; ++j) {
       int boxheight;
       int iy = y >> 16;

source/scale_neon.cc

@@ -541,6 +541,40 @@ void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
   );
 }
 
+void ScaleAddRows_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                    uint16* dst_ptr, int src_width, int src_height) {
+  const uint8* src_tmp = NULL;
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    "mov       %0, %1                          \n"
+    "mov       r12, %5                         \n"
+    "veor      q2, q2, q2                      \n"
+    "veor      q3, q3, q3                      \n"
+  "2:                                          \n"
+    // load 16 pixels into q0
+    MEMACCESS(0)
+    "vld1.8     {q0}, [%0], %3                 \n"
+    "vaddw.u8   q3, q3, d1                     \n"
+    "vaddw.u8   q2, q2, d0                     \n"
+    "subs       r12, r12, #1                   \n"
+    "bgt        2b                             \n"
+    MEMACCESS(2)
+    "vst1.16    {q2, q3}, [%2]!                \n"  // store pixels
+    "add        %1, %1, #16                    \n"
+    "subs       %4, %4, #16                    \n"  // 16 processed per loop
+    "bgt        1b                             \n"
+  : "+r"(src_tmp),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(dst_ptr),          // %2
+    "+r"(src_stride),       // %3
+    "+r"(src_width),        // %4
+    "+r"(src_height)        // %5
+  :
+  : "memory", "cc", "r12", "q0", "q1", "q2", "q3"  // Clobber List
+  );
+}
+
 // 16x2 -> 16x1
 void ScaleFilterRows_NEON(uint8* dst_ptr,
                           const uint8* src_ptr, ptrdiff_t src_stride,

source/scale_neon64.cc

@@ -545,6 +545,39 @@ void ScaleRowDown38_2_Box_NEON(const uint8* src_ptr,
   );
 }
 
+void ScaleAddRows_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                    uint16* dst_ptr, int src_width, int src_height) {
+  const uint8* src_tmp = NULL;
+  asm volatile (
+  "1:                                          \n"
+    "mov       %0, %1                          \n"
+    "mov       x12, %5                         \n"
+    "eor       v2.16b, v2.16b, v2.16b          \n"
+    "eor       v3.16b, v3.16b, v3.16b          \n"
+  "2:                                          \n"
+    // load 16 pixels into q0
+    MEMACCESS(0)
+    "ld1       {v0.16b}, [%0], %3              \n"
+    "uaddw2    v3.8h, v3.8h, v0.16b            \n"
+    "uaddw     v2.8h, v2.8h, v0.8b             \n"
+    "subs      x12, x12, #1                    \n"
+    "b.gt      2b                              \n"
+    MEMACCESS(2)
+    "st1      {v2.8h, v3.8h}, [%2], #32        \n"  // store pixels
+    "add      %1, %1, #16                      \n"
+    "subs     %4, %4, #16                      \n"  // 16 processed per loop
+    "b.gt     1b                               \n"
+  : "+r"(src_tmp),          // %0
+    "+r"(src_ptr),          // %1
+    "+r"(dst_ptr),          // %2
+    "+r"(src_stride),       // %3
+    "+r"(src_width),        // %4
+    "+r"(src_height)        // %5
+  :
+  : "memory", "cc", "x12", "v0", "v1", "v2", "v3"  // Clobber List
+  );
+}
+
 // 16x2 -> 16x1
 void ScaleFilterRows_NEON(uint8* dst_ptr,
                           const uint8* src_ptr, ptrdiff_t src_stride,