Add ScaleRowDown2Linear_NEON for ARM32/64

ARM32/64 NEON versions of ScaleRowDown2Linear_NEON are implemented.

BUG=319
TESTED=libyuvTest.ScaleDownBy2_Linear on ARM32/64 with Android
R=fbarchard@google.com

Change-Id: I2c7f43a0d56ed4dfded5bdbbb61765d87d65a2ba

Review URL: https://webrtc-codereview.appspot.com/43519005

git-svn-id: http://libyuv.googlecode.com/svn/trunk@1307 16f28f9a-4ce2-e073-06de-1de4eb20be90

include/libyuv/scale_row.h

@@ -268,6 +268,9 @@ void ScaleARGBRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
 void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                         uint8* dst, int dst_width);
 
+void ScaleRowDown2Linear_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                              uint8* dst, int dst_width);
+
 void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                            uint8* dst, int dst_width);
 

source/scale.cc

@@ -55,7 +55,9 @@ static void ScalePlaneDown2(int src_width, int src_height,
 
 #if defined(HAS_SCALEROWDOWN2_NEON)
   if (TestCpuFlag(kCpuHasNEON) && IS_ALIGNED(dst_width, 16)) {
-    ScaleRowDown2 = filtering ? ScaleRowDown2Box_NEON : ScaleRowDown2_NEON;
+    ScaleRowDown2 = filtering == kFilterNone ? ScaleRowDown2_NEON :
+        (filtering == kFilterLinear ? ScaleRowDown2Linear_NEON :
+        ScaleRowDown2Box_NEON);
   }
 #endif
 #if defined(HAS_SCALEROWDOWN2_SSE2)

source/scale_neon.cc

@@ -43,6 +43,30 @@ void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
   );
 }
 
+// Read 32x1 average down and write 16x1.
+void ScaleRowDown2Linear_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width) {
+  asm volatile (
+    ".p2align   2                              \n"
+  "1:                                          \n"
+    MEMACCESS(0)
+    "vld1.8     {q0, q1}, [%0]!                \n"  // load pixels and post inc
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop
+    "vpaddl.u8  q0, q0                         \n"  // add adjacent
+    "vpaddl.u8  q1, q1                         \n"
+    "vrshrn.u16 d0, q0, #1                     \n"  // downshift, round and pack
+    "vrshrn.u16 d1, q1, #1                     \n"
+    MEMACCESS(1)
+    "vst1.8     {q0}, [%1]!                    \n"
+    "bgt        1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "q0", "q1"     // Clobber List
+  );
+}
+
 // Read 32x2 average down and write 16x1.
 void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                            uint8* dst, int dst_width) {

source/scale_neon64.cc

@@ -40,6 +40,29 @@ void ScaleRowDown2_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
   );
 }
 
+// Read 32x1 average down and write 16x1.
+void ScaleRowDown2Linear_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                           uint8* dst, int dst_width) {
+  asm volatile (
+  "1:                                          \n"
+    MEMACCESS(0)
+    "ld1        {v0.16b,v1.16b}, [%0], #32     \n"  // load pixels and post inc
+    "subs       %2, %2, #16                    \n"  // 16 processed per loop
+    "uaddlp     v0.8h, v0.16b                  \n"  // add adjacent
+    "uaddlp     v1.8h, v1.16b                  \n"
+    "rshrn      v0.8b, v0.8h, #1               \n"  // downshift, round and pack
+    "rshrn2     v0.16b, v1.8h, #1              \n"
+    MEMACCESS(1)
+    "st1        {v0.16b}, [%1], #16            \n"
+    "b.gt       1b                             \n"
+  : "+r"(src_ptr),          // %0
+    "+r"(dst),              // %1
+    "+r"(dst_width)         // %2
+  :
+  : "v0", "v1"     // Clobber List
+  );
+}
+
 // Read 32x2 average down and write 16x1.
 void ScaleRowDown2Box_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                            uint8* dst, int dst_width) {