shithub: libvpx

Info  •  Files  •  Log  •  Branches

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

--- a/test/variance_test.cc

+++ b/test/variance_test.cc

@@ -1914,11 +1914,17 @@

 const vp9_variance_fn_t variance8x8_neon = vp9_variance8x8_neon;

 const vp9_variance_fn_t variance16x16_neon = vp9_variance16x16_neon;

 const vp9_variance_fn_t variance32x32_neon = vp9_variance32x32_neon;

+const vp9_variance_fn_t variance32x64_neon = vp9_variance32x64_neon;

+const vp9_variance_fn_t variance64x32_neon = vp9_variance64x32_neon;

+const vp9_variance_fn_t variance64x64_neon = vp9_variance64x64_neon;

 INSTANTIATE_TEST_CASE_P(

     NEON, VP9VarianceTest,

     ::testing::Values(make_tuple(3, 3, variance8x8_neon, 0),

                       make_tuple(4, 4, variance16x16_neon, 0),

-                      make_tuple(5, 5, variance32x32_neon, 0)));

+                      make_tuple(5, 5, variance32x32_neon, 0),

+                      make_tuple(5, 6, variance32x64_neon, 0),

+                      make_tuple(6, 5, variance64x32_neon, 0),

+                      make_tuple(6, 6, variance64x64_neon, 0)));

 const vp9_subpixvariance_fn_t subpel_variance8x8_neon =

     vp9_sub_pixel_variance8x8_neon;

--- a/vp9/common/vp9_rtcd_defs.pl

+++ b/vp9/common/vp9_rtcd_defs.pl

@@ -798,16 +798,16 @@

 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";

--- 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 { kWidth32 = 32 };

 enum { kHeight32 = 32 };

 enum { kHeight32PlusOne = 33 };

+enum { kWidth64 = 64 };

+enum { kHeight64 = 64 };

 enum { kPixelStepOne = 1 };

 enum { kAlign16 = 16 };

@@ -206,6 +209,30 @@

   int sum;

   variance_neon_w8(a, a_stride, b, b_stride, kWidth32, kHeight32, sse, &sum);

   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,