Add 64x variance Neon functions

Add optimized Neon functions of:
vp9_variance32x64
vp9_variance64x32
vp9_variance64x64

On Nexus 7 speed -5 and -6 saw about a 4% increase in perf.
Speeds -7 and -8 saw about a 6% increase in perf.
Tested on Nexus 7, built with ndk r10d, gcc 4.9.

Change-Id: I5a81f13c9897eb927fa39662530f5524a0f768fa

2 files changed, 30 insertions, 3 deletions

diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index 88f85a86d..b59e6ebe7 100644
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -798,16 +798,16 @@ add_proto qw/unsigned int vp9_variance16x32/, "const uint8_t *src_ptr, int sourc
 specialize qw/vp9_variance16x32/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance64x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance64x32 avx2/, "$sse2_x86inc";
+specialize qw/vp9_variance64x32 avx2 neon/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance32x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance32x64/, "$sse2_x86inc";
+specialize qw/vp9_variance32x64 neon/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance32x32/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
 specialize qw/vp9_variance32x32 avx2 neon/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_variance64x64 avx2/, "$sse2_x86inc";
+specialize qw/vp9_variance64x64 avx2 neon/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_variance16x16/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
 specialize qw/vp9_variance16x16 avx2 neon/, "$sse2_x86inc";
diff --git a/vp9/encoder/arm/neon/vp9_variance_neon.c b/vp9/encoder/arm/neon/vp9_variance_neon.c
index 816fbda1f..567b7deb1 100644
--- a/vp9/encoder/arm/neon/vp9_variance_neon.c
+++ b/vp9/encoder/arm/neon/vp9_variance_neon.c
@@ -10,6 +10,7 @@
 
 #include <arm_neon.h>
 #include "./vp9_rtcd.h"
+#include "./vpx_config.h"
 
 #include "vpx_ports/mem.h"
 #include "vpx/vpx_integer.h"
@@ -28,6 +29,8 @@ enum { kHeight16PlusOne = 17 };
 enum { kWidth32 = 32 };
 enum { kHeight32 = 32 };
 enum { kHeight32PlusOne = 33 };
+enum { kWidth64 = 64 };
+enum { kHeight64 = 64 };
 enum { kPixelStepOne = 1 };
 enum { kAlign16 = 16 };
 
@@ -208,6 +211,30 @@ unsigned int vp9_variance32x32_neon(const uint8_t *a, int a_stride,
   return *sse - (((int64_t)sum * sum) / (kWidth32 * kHeight32));
 }
 
+unsigned int vp9_variance32x64_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_neon_w8(a, a_stride, b, b_stride, kWidth32, kHeight64, sse, &sum);
+  return *sse - (((int64_t)sum * sum) >> 11);  // >> 11 = / 32 * 64
+}
+
+unsigned int vp9_variance64x32_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_neon_w8(a, a_stride, b, b_stride, kWidth64, kHeight32, sse, &sum);
+  return *sse - (((int64_t)sum * sum) >> 11);  // >> 11 = / 64 * 32
+}
+
+unsigned int vp9_variance64x64_neon(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    unsigned int *sse) {
+  int sum;
+  variance_neon_w8(a, a_stride, b, b_stride, kWidth64, kHeight64, sse, &sum);
+  return *sse - (((int64_t)sum * sum) >> 12);  // >> 12 = / 64 * 64
+}
+
 unsigned int vp9_sub_pixel_variance32x32_neon(const uint8_t *src,
                                               int src_stride,
                                               int xoffset,