Added vp9_sad64x64_neon(), vp9_sad32x32_neon()

and vp9_sad16x16_neon()

On a Nexus 7, vpxenc (in realtime mode, speed -6)
reported a performance improvement of ~17%.

Change-Id: I91e070cde2973451083d3f3d63b49b7886de9a85

test/sad_test.cc

@@ -473,6 +473,17 @@ INSTANTIATE_TEST_CASE_P(NEON, SADTest, ::testing::Values(
                         make_tuple(8, 8, sad_8x8_neon),
                         make_tuple(4, 4, sad_4x4_neon)));
 #endif  // CONFIG_VP8_ENCODER
+#if CONFIG_VP9_ENCODER
+const sad_m_by_n_fn_vp9_t sad_64x64_neon_vp9 = vp9_sad64x64_neon;
+const sad_m_by_n_fn_vp9_t sad_32x32_neon_vp9 = vp9_sad32x32_neon;
+const sad_m_by_n_fn_vp9_t sad_16x16_neon_vp9 = vp9_sad16x16_neon;
+const sad_m_by_n_test_param_vp9_t neon_vp9_tests[] = {
+  make_tuple(64, 64, sad_64x64_neon_vp9),
+  make_tuple(32, 32, sad_32x32_neon_vp9),
+  make_tuple(16, 16, sad_16x16_neon_vp9),
+};
+INSTANTIATE_TEST_CASE_P(NEON, SADVP9Test, ::testing::ValuesIn(neon_vp9_tests));
+#endif  // CONFIG_VP9_ENCODER
 #endif  // HAVE_NEON
 
 //------------------------------------------------------------------------------

vp9/common/vp9_rtcd_defs.pl

@@ -527,7 +527,7 @@ add_proto qw/unsigned int vp9_sub_pixel_avg_variance4x4/, "const uint8_t *src_pt
 specialize qw/vp9_sub_pixel_avg_variance4x4/, "$sse_x86inc", "$ssse3_x86inc";
 
 add_proto qw/unsigned int vp9_sad64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
-specialize qw/vp9_sad64x64/, "$sse2_x86inc";
+specialize qw/vp9_sad64x64 neon/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_sad32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride";
 specialize qw/vp9_sad32x64/, "$sse2_x86inc";
@@ -542,10 +542,10 @@ add_proto qw/unsigned int vp9_sad16x32/, "const uint8_t *src_ptr, int source_str
 specialize qw/vp9_sad16x32/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_sad32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
-specialize qw/vp9_sad32x32/, "$sse2_x86inc";
+specialize qw/vp9_sad32x32 neon/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_sad16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
-specialize qw/vp9_sad16x16 mmx/, "$sse2_x86inc";
+specialize qw/vp9_sad16x16 mmx neon/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_sad16x8/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride";
 specialize qw/vp9_sad16x8 mmx/, "$sse2_x86inc";

vp9/encoder/arm/neon/vp9_sad_neon.c

@@ -0,0 +1,116 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+
+#include "vpx/vpx_integer.h"
+
+static INLINE unsigned int horizontal_long_add_16x8(const uint16x8_t vec_lo,
+                                                    const uint16x8_t vec_hi) {
+  const uint32x4_t vec_l_lo = vaddl_u16(vget_low_u16(vec_lo),
+                                        vget_high_u16(vec_lo));
+  const uint32x4_t vec_l_hi = vaddl_u16(vget_low_u16(vec_hi),
+                                        vget_high_u16(vec_hi));
+  const uint32x4_t a = vaddq_u32(vec_l_lo, vec_l_hi);
+  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);
+}
+static INLINE unsigned int horizontal_add_16x8(const uint16x8_t vec_lo,
+                                               const uint16x8_t vec_hi) {
+  const uint32x4_t a = vpaddlq_u16(vaddq_u16(vec_lo, vec_hi));
+  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 vp9_sad64x64_neon(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum_lo = vdupq_n_u16(0);
+  uint16x8_t vec_accum_hi = vdupq_n_u16(0);
+  for (i = 0; i < 64; ++i) {
+    const uint8x16_t vec_src_00 = vld1q_u8(src);
+    const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
+    const uint8x16_t vec_src_32 = vld1q_u8(src + 32);
+    const uint8x16_t vec_src_48 = vld1q_u8(src + 48);
+    const uint8x16_t vec_ref_00 = vld1q_u8(ref);
+    const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
+    const uint8x16_t vec_ref_32 = vld1q_u8(ref + 32);
+    const uint8x16_t vec_ref_48 = vld1q_u8(ref + 48);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_00),
+                            vget_low_u8(vec_ref_00));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_00),
+                            vget_high_u8(vec_ref_00));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_16),
+                            vget_low_u8(vec_ref_16));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_16),
+                            vget_high_u8(vec_ref_16));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_32),
+                            vget_low_u8(vec_ref_32));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_32),
+                            vget_high_u8(vec_ref_32));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_48),
+                            vget_low_u8(vec_ref_48));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_48),
+                            vget_high_u8(vec_ref_48));
+  }
+  return horizontal_long_add_16x8(vec_accum_lo, vec_accum_hi);
+}
+
+unsigned int vp9_sad32x32_neon(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum_lo = vdupq_n_u16(0);
+  uint16x8_t vec_accum_hi = vdupq_n_u16(0);
+
+  for (i = 0; i < 32; ++i) {
+    const uint8x16_t vec_src_00 = vld1q_u8(src);
+    const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
+    const uint8x16_t vec_ref_00 = vld1q_u8(ref);
+    const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_00),
+                            vget_low_u8(vec_ref_00));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_00),
+                            vget_high_u8(vec_ref_00));
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src_16),
+                            vget_low_u8(vec_ref_16));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_16),
+                            vget_high_u8(vec_ref_16));
+  }
+  return horizontal_add_16x8(vec_accum_lo, vec_accum_hi);
+}
+
+unsigned int vp9_sad16x16_neon(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride) {
+  int i;
+  uint16x8_t vec_accum_lo = vdupq_n_u16(0);
+  uint16x8_t vec_accum_hi = vdupq_n_u16(0);
+
+  for (i = 0; i < 16; ++i) {
+    const uint8x16_t vec_src = vld1q_u8(src);
+    const uint8x16_t vec_ref = vld1q_u8(ref);
+    src += src_stride;
+    ref += ref_stride;
+    vec_accum_lo = vabal_u8(vec_accum_lo, vget_low_u8(vec_src),
+                            vget_low_u8(vec_ref));
+    vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src),
+                            vget_high_u8(vec_ref));
+  }
+  return horizontal_add_16x8(vec_accum_lo, vec_accum_hi);
+}

vp9/vp9cx.mk

@@ -129,4 +129,6 @@ VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_dct_avx2.c
 VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_error_intrin_avx2.c
 VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_variance_avx2.c
 
+VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_sad_neon.c
+
 VP9_CX_SRCS-yes := $(filter-out $(VP9_CX_SRCS_REMOVE-yes),$(VP9_CX_SRCS-yes))