Add 64x64 sub_pel_variance Neon function

On Nexus 7 speed -5, -6, -7, and -8 saw about a 15% increase
in perf for 480p. Speeds -5, -6, -7, and -8 saw about a 10%
increase in perf for 720p.

Tested on Nexus 7, built with ndk r10d, gcc 4.9.

Change-Id: I2fa5315845e3021c9a6e2ea47e52e68b398d8334

diff --git a/test/variance_test.cc b/test/variance_test.cc
index e7517a7d807cbf2bd801ce5916e69f367200fc87..a8dd7de139280b5720b79a0de99eccbde199842c 100644 (file)
--- a/test/variance_test.cc
+++ b/test/variance_test.cc
@@ -1932,11 +1932,14 @@ const vp9_subpixvariance_fn_t subpel_variance16x16_neon =
     vp9_sub_pixel_variance16x16_neon;
 const vp9_subpixvariance_fn_t subpel_variance32x32_neon =
     vp9_sub_pixel_variance32x32_neon;
+const vp9_subpixvariance_fn_t subpel_variance64x64_neon =
+    vp9_sub_pixel_variance64x64_neon;
 INSTANTIATE_TEST_CASE_P(
     NEON, VP9SubpelVarianceTest,
     ::testing::Values(make_tuple(3, 3, subpel_variance8x8_neon, 0),
                       make_tuple(4, 4, subpel_variance16x16_neon, 0),
-                      make_tuple(5, 5, subpel_variance32x32_neon, 0)));
+                      make_tuple(5, 5, subpel_variance32x32_neon, 0),
+                      make_tuple(6, 6, subpel_variance64x64_neon, 0)));
 #endif  // HAVE_NEON
 #endif  // CONFIG_VP9_ENCODER
 

diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index b59e6ebe794b146a4585bb6ff128f735b9a95d89..11df21f07e7dd58302df6ece10db6c9e42181f1c 100644 (file)
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -837,7 +837,7 @@ add_proto qw/unsigned int vp9_variance4x4/, "const uint8_t *src_ptr, int source_
 specialize qw/vp9_variance4x4/, "$sse2_x86inc";
 
 add_proto qw/unsigned int vp9_sub_pixel_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";
-specialize qw/vp9_sub_pixel_variance64x64 avx2/, "$sse2_x86inc", "$ssse3_x86inc";
+specialize qw/vp9_sub_pixel_variance64x64 avx2 neon/, "$sse2_x86inc", "$ssse3_x86inc";
 
 add_proto qw/unsigned int vp9_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred";
 specialize qw/vp9_sub_pixel_avg_variance64x64 avx2/, "$sse2_x86inc", "$ssse3_x86inc";

diff --git a/vp9/encoder/arm/neon/vp9_variance_neon.c b/vp9/encoder/arm/neon/vp9_variance_neon.c
index 567b7deb17d636e12149dff15e9df731d560bdaa..3e1b3de8cd0ef3778d871680bf23b6e1f53824e5 100644 (file)
--- a/vp9/encoder/arm/neon/vp9_variance_neon.c
+++ b/vp9/encoder/arm/neon/vp9_variance_neon.c
@@ -31,6 +31,7 @@ enum { kHeight32 = 32 };
 enum { kHeight32PlusOne = 33 };
 enum { kWidth64 = 64 };
 enum { kHeight64 = 64 };
+enum { kHeight64PlusOne = 65 };
 enum { kPixelStepOne = 1 };
 enum { kAlign16 = 16 };
 
@@ -252,3 +253,21 @@ unsigned int vp9_sub_pixel_variance32x32_neon(const uint8_t *src,
                              kWidth32, BILINEAR_FILTERS_2TAP(yoffset));
   return vp9_variance32x32_neon(temp2, kWidth32, dst, dst_stride, sse);
 }
+
+unsigned int vp9_sub_pixel_variance64x64_neon(const uint8_t *src,
+                                              int src_stride,
+                                              int xoffset,
+                                              int yoffset,
+                                              const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse) {
+  DECLARE_ALIGNED_ARRAY(kAlign16, uint8_t, temp2, kHeight64 * kWidth64);
+  DECLARE_ALIGNED_ARRAY(kAlign16, uint8_t, fdata3, kHeight64PlusOne * kWidth64);
+
+  var_filter_block2d_bil_w16(src, fdata3, src_stride, kPixelStepOne,
+                             kHeight64PlusOne, kWidth64,
+                             BILINEAR_FILTERS_2TAP(xoffset));
+  var_filter_block2d_bil_w16(fdata3, temp2, kWidth64, kWidth64, kHeight64,
+                             kWidth64, BILINEAR_FILTERS_2TAP(yoffset));
+  return vp9_variance64x64_neon(temp2, kWidth64, dst, dst_stride, sse);
+}