sad4d neon: 4x4, 4x8

BUG=webm:1425

Change-Id: I5081b5ce131821d590c53ac1206a94f50cb8b468

diff --git a/test/sad_test.cc b/test/sad_test.cc
index ef8b01caeba924a9625bc983f641b0e78dbae23d..13eae70e422d44df1596fe0276f2e41c73070653 100644 (file)
--- a/test/sad_test.cc
+++ b/test/sad_test.cc
@@ -678,6 +678,8 @@ const SadMxNx4Param x4d_neon_tests[] = {
   SadMxNx4Param(64, 64, &vpx_sad64x64x4d_neon),
   SadMxNx4Param(32, 32, &vpx_sad32x32x4d_neon),
   SadMxNx4Param(16, 16, &vpx_sad16x16x4d_neon),
+  SadMxNx4Param(4, 8, &vpx_sad4x8x4d_neon),
+  SadMxNx4Param(4, 4, &vpx_sad4x4x4d_neon),
 };
 INSTANTIATE_TEST_CASE_P(NEON, SADx4Test, ::testing::ValuesIn(x4d_neon_tests));
 #endif  // HAVE_NEON

diff --git a/vpx_dsp/arm/sad4d_neon.c b/vpx_dsp/arm/sad4d_neon.c
index dc20398000aaa4daffb16ef65a9c0c7863dd7c78..9c45ae1ab02f8deea7e911f9839889eb25f3fe23 100644 (file)
--- a/vpx_dsp/arm/sad4d_neon.c
+++ b/vpx_dsp/arm/sad4d_neon.c
@@ -13,6 +13,39 @@
 #include "./vpx_config.h"
 #include "./vpx_dsp_rtcd.h"
 #include "vpx/vpx_integer.h"
+#include "vpx_dsp/arm/mem_neon.h"
+#include "vpx_dsp/arm/sum_neon.h"
+
+void vpx_sad4x4x4d_neon(const uint8_t *src, int src_stride,
+                        const uint8_t *const ref[4], int ref_stride,
+                        uint32_t *res) {
+  int i;
+  const uint8x16_t src_u8 = load_unaligned_u8q(src, src_stride);
+  for (i = 0; i < 4; ++i) {
+    const uint8x16_t ref_u8 = load_unaligned_u8q(ref[i], ref_stride);
+    uint16x8_t abs = vabdl_u8(vget_low_u8(src_u8), vget_low_u8(ref_u8));
+    abs = vabal_u8(abs, vget_high_u8(src_u8), vget_high_u8(ref_u8));
+    res[i] = vget_lane_u32(horizontal_add_uint16x8(abs), 0);
+  }
+}
+
+void vpx_sad4x8x4d_neon(const uint8_t *src, int src_stride,
+                        const uint8_t *const ref[4], int ref_stride,
+                        uint32_t *res) {
+  int i;
+  const uint8x16_t src_0 = load_unaligned_u8q(src, src_stride);
+  const uint8x16_t src_1 = load_unaligned_u8q(src + 4 * src_stride, src_stride);
+  for (i = 0; i < 4; ++i) {
+    const uint8x16_t ref_0 = load_unaligned_u8q(ref[i], ref_stride);
+    const uint8x16_t ref_1 =
+        load_unaligned_u8q(ref[i] + 4 * ref_stride, ref_stride);
+    uint16x8_t abs = vabdl_u8(vget_low_u8(src_0), vget_low_u8(ref_0));
+    abs = vabal_u8(abs, vget_high_u8(src_0), vget_high_u8(ref_0));
+    abs = vabal_u8(abs, vget_low_u8(src_1), vget_low_u8(ref_1));
+    abs = vabal_u8(abs, vget_high_u8(src_1), vget_high_u8(ref_1));
+    res[i] = vget_lane_u32(horizontal_add_uint16x8(abs), 0);
+  }
+}
 
 static INLINE unsigned int horizontal_long_add_16x8(const uint16x8_t vec_lo,
                                                     const uint16x8_t vec_hi) {

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 4a7859237945cd98f22090d9ef11a3102fede44d..57c6a1e358f09c698769f0e798389460e6cfaa53 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -890,10 +890,10 @@ add_proto qw/void vpx_sad8x4x4d/, "const uint8_t *src_ptr, int src_stride, const
 specialize qw/vpx_sad8x4x4d msa sse2/;
 
 add_proto qw/void vpx_sad4x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad4x8x4d msa sse2/;
+specialize qw/vpx_sad4x8x4d neon msa sse2/;
 
 add_proto qw/void vpx_sad4x4x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad4x4x4d msa sse2/;
+specialize qw/vpx_sad4x4x4d neon msa sse2/;
 
 add_proto qw/uint64_t vpx_sum_squares_2d_i16/, "const int16_t *src, int stride, int size";
 specialize qw/vpx_sum_squares_2d_i16 sse2 msa/;