Add Neon implementation of vpx_sad_skip_<w>x<h> functions

Add Neon implementations of standard bitdepth downsampling SAD
functions for all block sizes.

Also add corresponding unit tests.

Change-Id: Ibda734c270278d947673ffcc29ef17a2f4970b01

diff --git a/test/sad_test.cc b/test/sad_test.cc
index 561da5d..e43d9ac 100644 (file)
--- a/test/sad_test.cc
+++ b/test/sad_test.cc
@@ -1129,6 +1129,24 @@ const SadMxNParam neon_tests[] = {
 };
 INSTANTIATE_TEST_SUITE_P(NEON, SADTest, ::testing::ValuesIn(neon_tests));
 
+const SadSkipMxNParam skip_neon_tests[] = {
+  SadSkipMxNParam(64, 64, &vpx_sad_skip_64x64_neon),
+  SadSkipMxNParam(64, 32, &vpx_sad_skip_64x32_neon),
+  SadSkipMxNParam(32, 64, &vpx_sad_skip_32x64_neon),
+  SadSkipMxNParam(32, 32, &vpx_sad_skip_32x32_neon),
+  SadSkipMxNParam(32, 16, &vpx_sad_skip_32x16_neon),
+  SadSkipMxNParam(16, 32, &vpx_sad_skip_16x32_neon),
+  SadSkipMxNParam(16, 16, &vpx_sad_skip_16x16_neon),
+  SadSkipMxNParam(16, 8, &vpx_sad_skip_16x8_neon),
+  SadSkipMxNParam(8, 16, &vpx_sad_skip_8x16_neon),
+  SadSkipMxNParam(8, 8, &vpx_sad_skip_8x8_neon),
+  SadSkipMxNParam(8, 4, &vpx_sad_skip_8x4_neon),
+  SadSkipMxNParam(4, 8, &vpx_sad_skip_4x8_neon),
+  SadSkipMxNParam(4, 4, &vpx_sad_skip_4x4_neon)
+};
+INSTANTIATE_TEST_SUITE_P(NEON, SADSkipTest,
+                         ::testing::ValuesIn(skip_neon_tests));
+
 const SadMxNAvgParam avg_neon_tests[] = {
   SadMxNAvgParam(64, 64, &vpx_sad64x64_avg_neon),
   SadMxNAvgParam(64, 32, &vpx_sad64x32_avg_neon),

diff --git a/vpx_dsp/arm/sad_neon.c b/vpx_dsp/arm/sad_neon.c
index 9382b80..566a1f8 100644 (file)
--- a/vpx_dsp/arm/sad_neon.c
+++ b/vpx_dsp/arm/sad_neon.c
@@ -250,6 +250,36 @@ SAD_WXH_NEON(32, 64)
 SAD_WXH_NEON(64, 32)
 SAD_WXH_NEON(64, 64)
 
+#undef SAD_WXH_NEON
+
+#define SAD_SKIP_WXH_NEON(w, h)                                                \
+  unsigned int vpx_sad_skip_##w##x##h##_neon(                                  \
+      const uint8_t *src, int src_stride, const uint8_t *ref,                  \
+      int ref_stride) {                                                        \
+    return 2 *                                                                 \
+           sad##w##xh_neon(src, 2 * src_stride, ref, 2 * ref_stride, (h) / 2); \
+  }
+
+SAD_SKIP_WXH_NEON(4, 4)
+SAD_SKIP_WXH_NEON(4, 8)
+
+SAD_SKIP_WXH_NEON(8, 4)
+SAD_SKIP_WXH_NEON(8, 8)
+SAD_SKIP_WXH_NEON(8, 16)
+
+SAD_SKIP_WXH_NEON(16, 8)
+SAD_SKIP_WXH_NEON(16, 16)
+SAD_SKIP_WXH_NEON(16, 32)
+
+SAD_SKIP_WXH_NEON(32, 16)
+SAD_SKIP_WXH_NEON(32, 32)
+SAD_SKIP_WXH_NEON(32, 64)
+
+SAD_SKIP_WXH_NEON(64, 32)
+SAD_SKIP_WXH_NEON(64, 64)
+
+#undef SAD_SKIP_WXH_NEON
+
 #if defined(__ARM_FEATURE_DOTPROD)
 
 static INLINE unsigned int sadwxh_avg_neon(const uint8_t *src_ptr,

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index e3d48f4..05d031b 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -787,41 +787,43 @@ add_proto qw/unsigned int vpx_sad4x4/, "const uint8_t *src_ptr, int src_stride,
 specialize qw/vpx_sad4x4 neon msa sse2 mmi/;
 
 add_proto qw/unsigned int vpx_sad_skip_64x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad_skip_64x64 avx2 sse2/;
+specialize qw/vpx_sad_skip_64x64 neon avx2 sse2/;
 
 add_proto qw/unsigned int vpx_sad_skip_64x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad_skip_64x32 avx2 sse2/;
+specialize qw/vpx_sad_skip_64x32 neon avx2 sse2/;
 
 add_proto qw/unsigned int vpx_sad_skip_32x64/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad_skip_32x64 avx2 sse2/;
+specialize qw/vpx_sad_skip_32x64 neon avx2 sse2/;
 
 add_proto qw/unsigned int vpx_sad_skip_32x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad_skip_32x32 avx2 sse2/;
+specialize qw/vpx_sad_skip_32x32 neon avx2 sse2/;
 
 add_proto qw/unsigned int vpx_sad_skip_32x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad_skip_32x16 avx2 sse2/;
+specialize qw/vpx_sad_skip_32x16 neon avx2 sse2/;
 
 add_proto qw/unsigned int vpx_sad_skip_16x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad_skip_16x32 sse2/;
+specialize qw/vpx_sad_skip_16x32 neon sse2/;
 
 add_proto qw/unsigned int vpx_sad_skip_16x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad_skip_16x16 sse2/;
+specialize qw/vpx_sad_skip_16x16 neon sse2/;
 
 add_proto qw/unsigned int vpx_sad_skip_16x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad_skip_16x8 sse2/;
+specialize qw/vpx_sad_skip_16x8 neon sse2/;
 
 add_proto qw/unsigned int vpx_sad_skip_8x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad_skip_8x16 sse2/;
+specialize qw/vpx_sad_skip_8x16 neon sse2/;
 
 add_proto qw/unsigned int vpx_sad_skip_8x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad_skip_8x8 sse2/;
+specialize qw/vpx_sad_skip_8x8 neon sse2/;
 
 add_proto qw/unsigned int vpx_sad_skip_8x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+specialize qw/vpx_sad_skip_8x4 neon/;
 
 add_proto qw/unsigned int vpx_sad_skip_4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad_skip_4x8 sse2/;
+specialize qw/vpx_sad_skip_4x8 neon sse2/;
 
 add_proto qw/unsigned int vpx_sad_skip_4x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
+specialize qw/vpx_sad_skip_4x4 neon/;
 
 #
 # Avg