From e40e78be246b2521749ce1fa6bdcd02e1f729a37 Mon Sep 17 00:00:00 2001
From: Johann <johannkoenig@google.com>
Date: Tue, 27 Jun 2017 13:02:28 -0700
Subject: [PATCH] sad neon: rewrite 8x8 and 8x16

BUG=webm:1425

Change-Id: I068f06c67b841f09ea07c04ada0c2f1706102138
---
 test/sad_test.cc             |  1 +
 vpx_dsp/arm/sad_neon.c       | 79 +++++++++++++++---------------------
 vpx_dsp/vpx_dsp_rtcd_defs.pl |  2 +-
 3 files changed, 35 insertions(+), 47 deletions(-)

diff --git a/test/sad_test.cc b/test/sad_test.cc
index 00a4bc084..045980642 100644
--- 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(8, 4, &vpx_sad8x4_neon),
   SadMxNParam(4, 8, &vpx_sad4x8_neon),
   SadMxNParam(4, 4, &vpx_sad4x4_neon),
 };
diff --git a/vpx_dsp/arm/sad_neon.c b/vpx_dsp/arm/sad_neon.c
index f5acb9149..cbc904feb 100644
--- a/vpx_dsp/arm/sad_neon.c
+++ b/vpx_dsp/arm/sad_neon.c
@@ -15,37 +15,6 @@
 #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) {
-  uint8x8_t d0, d8;
-  uint16x8_t q12;
-  uint32x4_t q1;
-  uint64x2_t q3;
-  uint32x2_t d5;
-  int i;
-
-  d0 = vld1_u8(src_ptr);
-  src_ptr += src_stride;
-  d8 = vld1_u8(ref_ptr);
-  ref_ptr += ref_stride;
-  q12 = vabdl_u8(d0, d8);
-
-  for (i = 0; i < 15; 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);
-  }
-
-  q1 = vpaddlq_u16(q12);
-  q3 = vpaddlq_u32(q1);
-  d5 = vadd_u32(vreinterpret_u32_u64(vget_low_u64(q3)),
-                vreinterpret_u32_u64(vget_high_u64(q3)));
-
-  return vget_lane_u32(d5, 0);
-}
-
 // 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);
@@ -80,6 +49,39 @@ uint32_t vpx_sad4x8_neon(const uint8_t *src_ptr, int src_stride,
   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;
+  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);
+    a += a_stride;
+    b += b_stride;
+    abs = vabal_u8(abs, a_u8, b_u8);
+  }
+  return abs;
+}
+
+uint32_t vpx_sad8x4_neon(const uint8_t *src, int src_stride, const uint8_t *ref,
+                         int ref_stride) {
+  const uint16x8_t abs = sad8x(src, src_stride, ref, ref_stride, 4);
+  return horizontal_add_16x8(abs);
+}
+
+uint32_t vpx_sad8x8_neon(const uint8_t *src, int src_stride, const uint8_t *ref,
+                         int ref_stride) {
+  const uint16x8_t abs = sad8x(src, src_stride, ref, ref_stride, 8);
+  return horizontal_add_16x8(abs);
+}
+
+uint32_t vpx_sad8x16_neon(const uint8_t *src, int src_stride,
+                          const uint8_t *ref, int ref_stride) {
+  const uint16x8_t abs = sad8x(src, src_stride, ref, ref_stride, 16);
+  return horizontal_add_16x8(abs);
+}
+
 unsigned int vpx_sad16x8_neon(unsigned char *src_ptr, int src_stride,
                               unsigned char *ref_ptr, int ref_stride) {
   uint8x16_t q0, q4;
@@ -206,18 +208,3 @@ unsigned int vpx_sad16x16_neon(const uint8_t *src, int src_stride,
   }
   return horizontal_add_16x8(vaddq_u16(vec_accum_lo, vec_accum_hi));
 }
-
-unsigned int vpx_sad8x8_neon(const uint8_t *src, int src_stride,
-                             const uint8_t *ref, int ref_stride) {
-  int i;
-  uint16x8_t vec_accum = vdupq_n_u16(0);
-
-  for (i = 0; i < 8; ++i) {
-    const uint8x8_t vec_src = vld1_u8(src);
-    const uint8x8_t vec_ref = vld1_u8(ref);
-    src += src_stride;
-    ref += ref_stride;
-    vec_accum = vabal_u8(vec_accum, vec_src, vec_ref);
-  }
-  return horizontal_add_16x8(vec_accum);
-}
diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 96e2bb7e4..88497b796 100644
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -726,7 +726,7 @@ add_proto qw/unsigned int vpx_sad8x8/, "const uint8_t *src_ptr, int src_stride,
 specialize qw/vpx_sad8x8 neon msa sse2/;
 
 add_proto qw/unsigned int vpx_sad8x4/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad8x4 msa sse2/;
+specialize qw/vpx_sad8x4 neon 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 neon msa sse2/;
-- 
2.49.0