sad neon: avg for 4x4 and 4x8

BUG=webm:1425

Change-Id: Ifc685a96cb34f7fd9243b4c674027480564b84fb

diff --git a/test/sad_test.cc b/test/sad_test.cc
index 8c7742717fb57870caed2f366ad89a6a4632c421..da15116399efaf2e683dacb730a0299446236ebe 100644 (file)
--- a/test/sad_test.cc
+++ b/test/sad_test.cc
@@ -657,6 +657,12 @@ const SadMxNParam neon_tests[] = {
 };
 INSTANTIATE_TEST_CASE_P(NEON, SADTest, ::testing::ValuesIn(neon_tests));
 
+const SadMxNAvgParam avg_neon_tests[] = {
+  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),

diff --git a/vpx_dsp/arm/sad_neon.c b/vpx_dsp/arm/sad_neon.c
index 41c764c6f6fa2deecfb04b510c8e30c7b150a796..c227ddc9d155857e62aff0342e7589e978b68531 100644 (file)
--- a/vpx_dsp/arm/sad_neon.c
+++ b/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;

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 869f29349d3074ab7862888903882769bfac2a9e..0dd496765ef8646b63c6dd6c3acf7fd8f8d90570 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -813,10 +813,10 @@ add_proto qw/unsigned int vpx_sad8x4_avg/, "const uint8_t *src_ptr, int src_stri
 specialize qw/vpx_sad8x4_avg 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