shithub: libvpx

Download patch

ref: f6fcd3410d7277e67b2a46cbc9f48436011be4db
parent: 188d58eaa99ed5b47af77b354e303a82614ce62a
author: Johann <johannkoenig@google.com>
date: Thu May 4 10:48:43 EDT 2017

sub pel avg variance neon: 4x block sizes

BUG=webm:1423

Change-Id: Iaab2b9a183fdb54aae5f717aba95d90dc36a9e3b

--- a/test/variance_test.cc
+++ b/test/variance_test.cc
@@ -1292,7 +1292,9 @@
         make_tuple(4, 3, &vpx_sub_pixel_avg_variance16x8_neon, 0),
         make_tuple(3, 4, &vpx_sub_pixel_avg_variance8x16_neon, 0),
         make_tuple(3, 3, &vpx_sub_pixel_avg_variance8x8_neon, 0),
-        make_tuple(3, 2, &vpx_sub_pixel_avg_variance8x4_neon, 0)));
+        make_tuple(3, 2, &vpx_sub_pixel_avg_variance8x4_neon, 0),
+        make_tuple(2, 3, &vpx_sub_pixel_avg_variance4x8_neon, 0),
+        make_tuple(2, 2, &vpx_sub_pixel_avg_variance4x4_neon, 0)));
 #endif  // HAVE_NEON
 
 #if HAVE_MSA
--- a/vpx_dsp/arm/subpel_variance_neon.c
+++ b/vpx_dsp/arm/subpel_variance_neon.c
@@ -141,33 +141,41 @@
 sub_pixel_varianceNxM(64, 32);
 sub_pixel_varianceNxM(64, 64);
 
-// TODO(johannkoenig): support 4xM block sizes.
-#define sub_pixel_avg_varianceNxM(n, m)                              \
-  uint32_t vpx_sub_pixel_avg_variance##n##x##m##_neon(               \
-      const uint8_t *a, int a_stride, int xoffset, int yoffset,      \
-      const uint8_t *b, int b_stride, uint32_t *sse,                 \
-      const uint8_t *second_pred) {                                  \
-    DECLARE_ALIGNED(16, uint8_t, fdata3[n * (m + 1)]);               \
-    DECLARE_ALIGNED(16, uint8_t, temp2[n * m]);                      \
-    DECLARE_ALIGNED(16, uint8_t, temp3[n * m]);                      \
-                                                                     \
-    if (n == 8) {                                                    \
-      var_filter_block2d_bil_w8(a, fdata3, a_stride, 1, (m + 1),     \
-                                bilinear_filters[xoffset]);          \
-      var_filter_block2d_bil_w8(fdata3, temp2, n, n, m,              \
-                                bilinear_filters[yoffset]);          \
-    } else {                                                         \
-      var_filter_block2d_bil_w16(a, fdata3, a_stride, 1, (m + 1), n, \
-                                 bilinear_filters[xoffset]);         \
-      var_filter_block2d_bil_w16(fdata3, temp2, n, n, m, n,          \
-                                 bilinear_filters[yoffset]);         \
-    }                                                                \
-                                                                     \
-    vpx_comp_avg_pred(temp3, second_pred, n, m, temp2, n);           \
-                                                                     \
-    return vpx_variance##n##x##m(temp3, n, b, b_stride, sse);        \
+// 4xM filter writes an extra row to fdata because it processes two rows at a
+// time.
+#define sub_pixel_avg_varianceNxM(n, m)                               \
+  uint32_t vpx_sub_pixel_avg_variance##n##x##m##_neon(                \
+      const uint8_t *a, int a_stride, int xoffset, int yoffset,       \
+      const uint8_t *b, int b_stride, uint32_t *sse,                  \
+      const uint8_t *second_pred) {                                   \
+    DECLARE_ALIGNED(16, uint8_t, fdata3[n * (m + (n == 4 ? 2 : 1))]); \
+    DECLARE_ALIGNED(16, uint8_t, temp2[n * m]);                       \
+    DECLARE_ALIGNED(16, uint8_t, temp3[n * m]);                       \
+                                                                      \
+    if (n == 4) {                                                     \
+      var_filter_block2d_bil_w4(a, fdata3, a_stride, 1, (m + 2),      \
+                                bilinear_filters[xoffset]);           \
+      var_filter_block2d_bil_w4(fdata3, temp2, n, n, m,               \
+                                bilinear_filters[yoffset]);           \
+    } else if (n == 8) {                                              \
+      var_filter_block2d_bil_w8(a, fdata3, a_stride, 1, (m + 1),      \
+                                bilinear_filters[xoffset]);           \
+      var_filter_block2d_bil_w8(fdata3, temp2, n, n, m,               \
+                                bilinear_filters[yoffset]);           \
+    } else {                                                          \
+      var_filter_block2d_bil_w16(a, fdata3, a_stride, 1, (m + 1), n,  \
+                                 bilinear_filters[xoffset]);          \
+      var_filter_block2d_bil_w16(fdata3, temp2, n, n, m, n,           \
+                                 bilinear_filters[yoffset]);          \
+    }                                                                 \
+                                                                      \
+    vpx_comp_avg_pred(temp3, second_pred, n, m, temp2, n);            \
+                                                                      \
+    return vpx_variance##n##x##m(temp3, n, b, b_stride, sse);         \
   }
 
+sub_pixel_avg_varianceNxM(4, 4);
+sub_pixel_avg_varianceNxM(4, 8);
 sub_pixel_avg_varianceNxM(8, 4);
 sub_pixel_avg_varianceNxM(8, 8);
 sub_pixel_avg_varianceNxM(8, 16);
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -1253,10 +1253,10 @@
   specialize qw/vpx_sub_pixel_avg_variance8x4 neon msa sse2 ssse3/;
 
 add_proto qw/uint32_t vpx_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
-  specialize qw/vpx_sub_pixel_avg_variance4x8 msa sse2 ssse3/;
+  specialize qw/vpx_sub_pixel_avg_variance4x8 neon msa sse2 ssse3/;
 
 add_proto qw/uint32_t vpx_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
-  specialize qw/vpx_sub_pixel_avg_variance4x4 msa sse2 ssse3/;
+  specialize qw/vpx_sub_pixel_avg_variance4x4 neon msa sse2 ssse3/;
 
 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/unsigned int vpx_highbd_12_variance64x64/, "const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse";