Merge changes I7b36a57e,If2ab51e3,Ifc685a96

* changes:
  sad neon: macroize 8xN definitions
  sad neon: avg for 8x[4,8,16]
  sad neon: avg for 4x4 and 4x8

test/sad_test.cc

@@ -657,6 +657,15 @@ const SadMxNParam neon_tests[] = {
 };
 INSTANTIATE_TEST_CASE_P(NEON, SADTest, ::testing::ValuesIn(neon_tests));
 
+const SadMxNAvgParam avg_neon_tests[] = {
+  SadMxNAvgParam(8, 16, &vpx_sad8x16_avg_neon),
+  SadMxNAvgParam(8, 8, &vpx_sad8x8_avg_neon),
+  SadMxNAvgParam(8, 4, &vpx_sad8x4_avg_neon),
+  SadMxNAvgParam(4, 8, &vpx_sad4x8_avg_neon),
+  SadMxNAvgParam(4, 4, &vpx_sad4x4_avg_neon),
+};
+INSTANTIATE_TEST_CASE_P(NEON, SADavgTest, ::testing::ValuesIn(avg_neon_tests));
+
 const SadMxNx4Param x4d_neon_tests[] = {
   SadMxNx4Param(64, 64, &vpx_sad64x64x4d_neon),
   SadMxNx4Param(32, 32, &vpx_sad32x32x4d_neon),

vpx_dsp/arm/sad_neon.c

@@ -33,6 +33,18 @@ uint32_t vpx_sad4x4_neon(const uint8_t *src_ptr, int src_stride,
   return horizontal_add_16x8(abs);
 }
 
+uint32_t vpx_sad4x4_avg_neon(const uint8_t *src_ptr, int src_stride,
+                             const uint8_t *ref_ptr, int ref_stride,
+                             const uint8_t *second_pred) {
+  const uint8x16_t src_u8 = load_unaligned_u8q(src_ptr, src_stride);
+  const uint8x16_t ref_u8 = load_unaligned_u8q(ref_ptr, ref_stride);
+  const uint8x16_t second_pred_u8 = vld1q_u8(second_pred);
+  const uint8x16_t avg = vrhaddq_u8(ref_u8, second_pred_u8);
+  uint16x8_t abs = vabdl_u8(vget_low_u8(src_u8), vget_low_u8(avg));
+  abs = vabal_u8(abs, vget_high_u8(src_u8), vget_high_u8(avg));
+  return horizontal_add_16x8(abs);
+}
+
 uint32_t vpx_sad4x8_neon(const uint8_t *src_ptr, int src_stride,
                          const uint8_t *ref_ptr, int ref_stride) {
   int i;
@@ -49,6 +61,26 @@ uint32_t vpx_sad4x8_neon(const uint8_t *src_ptr, int src_stride,
   return horizontal_add_16x8(abs);
 }
 
+uint32_t vpx_sad4x8_avg_neon(const uint8_t *src_ptr, int src_stride,
+                             const uint8_t *ref_ptr, int ref_stride,
+                             const uint8_t *second_pred) {
+  int i;
+  uint16x8_t abs = vdupq_n_u16(0);
+  for (i = 0; i < 8; i += 4) {
+    const uint8x16_t src_u8 = load_unaligned_u8q(src_ptr, src_stride);
+    const uint8x16_t ref_u8 = load_unaligned_u8q(ref_ptr, ref_stride);
+    const uint8x16_t second_pred_u8 = vld1q_u8(second_pred);
+    const uint8x16_t avg = vrhaddq_u8(ref_u8, second_pred_u8);
+    src_ptr += 4 * src_stride;
+    ref_ptr += 4 * ref_stride;
+    second_pred += 16;
+    abs = vabal_u8(abs, vget_low_u8(src_u8), vget_low_u8(avg));
+    abs = vabal_u8(abs, vget_high_u8(src_u8), vget_high_u8(avg));
+  }
+
+  return horizontal_add_16x8(abs);
+}
+
 static INLINE uint16x8_t sad8x(const uint8_t *a, int a_stride, const uint8_t *b,
                                int b_stride, const int height) {
   int i;
@@ -64,23 +96,43 @@ static INLINE uint16x8_t sad8x(const uint8_t *a, int a_stride, const uint8_t *b,
   return abs;
 }
 
-uint32_t vpx_sad8x4_neon(const uint8_t *src, int src_stride, const uint8_t *ref,
+static INLINE uint16x8_t sad8x_avg(const uint8_t *a, int a_stride,
-                         int ref_stride) {
+                                   const uint8_t *b, int b_stride,
-  const uint16x8_t abs = sad8x(src, src_stride, ref, ref_stride, 4);
+                                   const uint8_t *c, const int height) {
-  return horizontal_add_16x8(abs);
+  int i;
+  uint16x8_t abs = vdupq_n_u16(0);
+
+  for (i = 0; i < height; ++i) {
+    const uint8x8_t a_u8 = vld1_u8(a);
+    const uint8x8_t b_u8 = vld1_u8(b);
+    const uint8x8_t c_u8 = vld1_u8(c);
+    const uint8x8_t avg = vrhadd_u8(b_u8, c_u8);
+    a += a_stride;
+    b += b_stride;
+    c += 8;
+    abs = vabal_u8(abs, a_u8, avg);
+  }
+  return abs;
 }
 
-uint32_t vpx_sad8x8_neon(const uint8_t *src, int src_stride, const uint8_t *ref,
+#define sad8xN(n)                                                      \
-                         int ref_stride) {
+  uint32_t vpx_sad8x##n##_neon(const uint8_t *src, int src_stride,     \
-  const uint16x8_t abs = sad8x(src, src_stride, ref, ref_stride, 8);
+                               const uint8_t *ref, int ref_stride) {   \
-  return horizontal_add_16x8(abs);
+    const uint16x8_t abs = sad8x(src, src_stride, ref, ref_stride, n); \
-}
+    return horizontal_add_16x8(abs);                                   \
+  }                                                                    \
+                                                                       \
+  uint32_t vpx_sad8x##n##_avg_neon(const uint8_t *src, int src_stride, \
+                                   const uint8_t *ref, int ref_stride, \
+                                   const uint8_t *second_pred) {       \
+    const uint16x8_t abs =                                             \
+        sad8x_avg(src, src_stride, ref, ref_stride, second_pred, n);   \
+    return horizontal_add_16x8(abs);                                   \
+  }
 
-uint32_t vpx_sad8x16_neon(const uint8_t *src, int src_stride,
+sad8xN(4);
-                          const uint8_t *ref, int ref_stride) {
+sad8xN(8);
-  const uint16x8_t abs = sad8x(src, src_stride, ref, ref_stride, 16);
+sad8xN(16);
-  return horizontal_add_16x8(abs);
-}
 
 static INLINE uint16x8_t sad16x(const uint8_t *a, int a_stride,
                                 const uint8_t *b, int b_stride,

vpx_dsp/vpx_dsp_rtcd_defs.pl

@@ -804,19 +804,19 @@ add_proto qw/unsigned int vpx_sad16x8_avg/, "const uint8_t *src_ptr, int src_str
 specialize qw/vpx_sad16x8_avg msa sse2 vsx/;
 
 add_proto qw/unsigned int vpx_sad8x16_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vpx_sad8x16_avg msa sse2/;
+specialize qw/vpx_sad8x16_avg neon msa sse2/;
 
 add_proto qw/unsigned int vpx_sad8x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vpx_sad8x8_avg msa sse2/;
+specialize qw/vpx_sad8x8_avg neon msa sse2/;
 
 add_proto qw/unsigned int vpx_sad8x4_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vpx_sad8x4_avg msa sse2/;
+specialize qw/vpx_sad8x4_avg neon msa sse2/;
 
 add_proto qw/unsigned int vpx_sad4x8_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vpx_sad4x8_avg msa sse2/;
+specialize qw/vpx_sad4x8_avg neon msa sse2/;
 
 add_proto qw/unsigned int vpx_sad4x4_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
-specialize qw/vpx_sad4x4_avg msa sse2/;
+specialize qw/vpx_sad4x4_avg neon msa sse2/;
 
 #
 # Multi-block SAD, comparing a reference to N blocks 1 pixel apart horizontally