sad neon: rewrite 4x4 and add 4x8

The previous implementation loaded 8 values (discarding half)

BUG=webm:1425

Change-Id: Icb72a94e2557a4ee2db7091266ab58fd92f72158

diff --git a/test/sad_test.cc b/test/sad_test.cc
index 7bf6e13856cebc84326b23cefa4dcb24dd60ae4a..00a4bc084a074822efd1c382225f966775605652 100644 (file)
--- a/test/sad_test.cc
+++ b/test/sad_test.cc
@@ -649,6 +649,7 @@ const SadMxNParam neon_tests[] = {
   SadMxNParam(16, 8, &vpx_sad16x8_neon),
   SadMxNParam(8, 16, &vpx_sad8x16_neon),
   SadMxNParam(8, 8, &vpx_sad8x8_neon),
+  SadMxNParam(4, 8, &vpx_sad4x8_neon),
   SadMxNParam(4, 4, &vpx_sad4x4_neon),
 };
 INSTANTIATE_TEST_CASE_P(NEON, SADTest, ::testing::ValuesIn(neon_tests));

diff --git a/vpx_dsp/arm/sad_neon.c b/vpx_dsp/arm/sad_neon.c
index ff3228768cefc8455f65267982b881c9c4ebf7d2..f5acb9149dc88ff67036f2562dc385436ecdd5d7 100644 (file)
--- a/vpx_dsp/arm/sad_neon.c
+++ b/vpx_dsp/arm/sad_neon.c
@@ -13,6 +13,7 @@
 #include "./vpx_config.h"
 
 #include "vpx/vpx_integer.h"
+#include "vpx_dsp/arm/mem_neon.h"
 
 unsigned int vpx_sad8x16_neon(unsigned char *src_ptr, int src_stride,
                               unsigned char *ref_ptr, int ref_stride) {
@@ -45,32 +46,38 @@ unsigned int vpx_sad8x16_neon(unsigned char *src_ptr, int src_stride,
   return vget_lane_u32(d5, 0);
 }
 
-unsigned int vpx_sad4x4_neon(unsigned char *src_ptr, int src_stride,
-                             unsigned char *ref_ptr, int ref_stride) {
-  uint8x8_t d0, d8;
-  uint16x8_t q12;
-  uint32x2_t d1;
-  uint64x1_t d3;
-  int i;
+// TODO(johannkoenig): combine with avg_neon.h version.
+static INLINE uint32_t horizontal_add_16x8(const uint16x8_t vec_16x8) {
+  const uint32x4_t a = vpaddlq_u16(vec_16x8);
+  const uint64x2_t b = vpaddlq_u32(a);
+  const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
+                                vreinterpret_u32_u64(vget_high_u64(b)));
+  return vget_lane_u32(c, 0);
+}
 
-  d0 = vld1_u8(src_ptr);
-  src_ptr += src_stride;
-  d8 = vld1_u8(ref_ptr);
-  ref_ptr += ref_stride;
-  q12 = vabdl_u8(d0, d8);
+uint32_t vpx_sad4x4_neon(const uint8_t *src_ptr, int src_stride,
+                         const uint8_t *ref_ptr, int ref_stride) {
+  const uint8x16_t src_u8 = load_unaligned_u8q(src_ptr, src_stride);
+  const uint8x16_t ref_u8 = load_unaligned_u8q(ref_ptr, 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));
+  return horizontal_add_16x8(abs);
+}
 
-  for (i = 0; i < 3; i++) {
-    d0 = vld1_u8(src_ptr);
-    src_ptr += src_stride;
-    d8 = vld1_u8(ref_ptr);
-    ref_ptr += ref_stride;
-    q12 = vabal_u8(q12, d0, d8);
+uint32_t vpx_sad4x8_neon(const uint8_t *src_ptr, int src_stride,
+                         const uint8_t *ref_ptr, int ref_stride) {
+  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);
+    src_ptr += 4 * src_stride;
+    ref_ptr += 4 * ref_stride;
+    abs = vabal_u8(abs, vget_low_u8(src_u8), vget_low_u8(ref_u8));
+    abs = vabal_u8(abs, vget_high_u8(src_u8), vget_high_u8(ref_u8));
   }
 
-  d1 = vpaddl_u16(vget_low_u16(q12));
-  d3 = vpaddl_u32(d1);
-
-  return vget_lane_u32(vreinterpret_u32_u64(d3), 0);
+  return horizontal_add_16x8(abs);
 }
 
 unsigned int vpx_sad16x8_neon(unsigned char *src_ptr, int src_stride,
@@ -119,13 +126,6 @@ static INLINE unsigned int horizontal_long_add_16x8(const uint16x8_t vec_lo,
                                 vreinterpret_u32_u64(vget_high_u64(b)));
   return vget_lane_u32(c, 0);
 }
-static INLINE unsigned int horizontal_add_16x8(const uint16x8_t vec_16x8) {
-  const uint32x4_t a = vpaddlq_u16(vec_16x8);
-  const uint64x2_t b = vpaddlq_u32(a);
-  const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
-                                vreinterpret_u32_u64(vget_high_u64(b)));
-  return vget_lane_u32(c, 0);
-}
 
 unsigned int vpx_sad64x64_neon(const uint8_t *src, int src_stride,
                                const uint8_t *ref, int ref_stride) {

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 51aa9f637ffe7d2547e90b82f3eaaf5ad52beebe..96e2bb7e4c52c567533fdd93d5938d7a04ea4be2 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -729,7 +729,7 @@ add_proto qw/unsigned int vpx_sad8x4/, "const uint8_t *src_ptr, int src_stride,
 specialize qw/vpx_sad8x4 msa sse2/;
 
 add_proto qw/unsigned int vpx_sad4x8/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad4x8 msa sse2/;
+specialize qw/vpx_sad4x8 neon msa sse2/;
 
 add_proto qw/unsigned int vpx_sad4x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 specialize qw/vpx_sad4x4 neon msa sse2/;