ref: 2346a6da4a3703eb2cb346f3b4a8e6d8a25c70f6
parent: 19e1ec8359069dd94ee06b9f2df63fd1c40497a1
author: Johann <johannkoenig@google.com>
date: Wed May 3 08:06:29 EDT 2017
subpel variance neon: add mixed sizes Add support for everything except block sizes of 4. Performance is better but numbers will improve again when the variance optimizations land. BUG=webm:1423 Change-Id: I92eb4312b20be423fa2fe6fdb18167a604ff4d80
--- a/test/variance_test.cc
+++ b/test/variance_test.cc
@@ -1227,9 +1227,16 @@
INSTANTIATE_TEST_CASE_P(
NEON, VpxSubpelVarianceTest,
::testing::Values(make_tuple(6, 6, &vpx_sub_pixel_variance64x64_neon, 0),
+ make_tuple(6, 5, &vpx_sub_pixel_variance64x32_neon, 0),
+ make_tuple(5, 6, &vpx_sub_pixel_variance32x64_neon, 0),
make_tuple(5, 5, &vpx_sub_pixel_variance32x32_neon, 0),
+ make_tuple(5, 4, &vpx_sub_pixel_variance32x16_neon, 0),
+ make_tuple(4, 5, &vpx_sub_pixel_variance16x32_neon, 0),
make_tuple(4, 4, &vpx_sub_pixel_variance16x16_neon, 0),
- make_tuple(3, 3, &vpx_sub_pixel_variance8x8_neon, 0)));
+ make_tuple(4, 3, &vpx_sub_pixel_variance16x8_neon, 0),
+ make_tuple(3, 4, &vpx_sub_pixel_variance8x16_neon, 0),
+ make_tuple(3, 3, &vpx_sub_pixel_variance8x8_neon, 0),
+ make_tuple(3, 2, &vpx_sub_pixel_variance8x4_neon, 0)));
#endif // HAVE_NEON
#if HAVE_MSA
--- a/vpx_dsp/arm/subpel_variance_neon.c
+++ b/vpx_dsp/arm/subpel_variance_neon.c
@@ -28,7 +28,6 @@
unsigned int src_pixels_per_line,
int pixel_step,
unsigned int output_height,
- unsigned int output_width,
const uint8_t *filter) {const uint8x8_t f0 = vmov_n_u8(filter[0]);
const uint8x8_t f1 = vmov_n_u8(filter[1]);
@@ -42,7 +41,7 @@
vst1_u8(&output_ptr[0], out);
// Next row...
src_ptr += src_pixels_per_line;
- output_ptr += output_width;
+ output_ptr += 8;
}
}
@@ -75,61 +74,36 @@
}
}
-unsigned int vpx_sub_pixel_variance8x8_neon(const uint8_t *src, int src_stride,
- int xoffset, int yoffset,
- const uint8_t *dst, int dst_stride,
- unsigned int *sse) {- DECLARE_ALIGNED(16, uint8_t, temp2[8 * 8]);
- DECLARE_ALIGNED(16, uint8_t, fdata3[9 * 8]);
+// TODO(johannkoenig): support 4xM block sizes.
+#define sub_pixel_varianceNxM(n, m) \
+ unsigned int vpx_sub_pixel_variance##n##x##m##_neon( \
+ const uint8_t *src, int src_stride, int xoffset, int yoffset, \
+ const uint8_t *dst, int dst_stride, unsigned int *sse) { \+ DECLARE_ALIGNED(16, uint8_t, fdata3[n * (m + 1)]); \
+ DECLARE_ALIGNED(16, uint8_t, temp2[n * m]); \
+ \
+ if (n == 8) { \+ var_filter_block2d_bil_w8(src, fdata3, src_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(src, fdata3, src_stride, 1, (m + 1), n, \
+ bilinear_filters[xoffset]); \
+ var_filter_block2d_bil_w16(fdata3, temp2, n, n, m, n, \
+ bilinear_filters[yoffset]); \
+ } \
+ return vpx_variance##n##x##m(temp2, n, dst, dst_stride, sse); \
+ }
- var_filter_block2d_bil_w8(src, fdata3, src_stride, 1, 9, 8,
- bilinear_filters[xoffset]);
- var_filter_block2d_bil_w8(fdata3, temp2, 8, 8, 8, 8,
- bilinear_filters[yoffset]);
- return vpx_variance8x8_neon(temp2, 8, dst, dst_stride, sse);
-}
-
-unsigned int vpx_sub_pixel_variance16x16_neon(const uint8_t *src,
- int src_stride, int xoffset,
- int yoffset, const uint8_t *dst,
- int dst_stride,
- unsigned int *sse) {- DECLARE_ALIGNED(16, uint8_t, temp2[16 * 16]);
- DECLARE_ALIGNED(16, uint8_t, fdata3[17 * 16]);
-
- var_filter_block2d_bil_w16(src, fdata3, src_stride, 1, 17, 16,
- bilinear_filters[xoffset]);
- var_filter_block2d_bil_w16(fdata3, temp2, 16, 16, 16, 16,
- bilinear_filters[yoffset]);
- return vpx_variance16x16(temp2, 16, dst, dst_stride, sse);
-}
-
-unsigned int vpx_sub_pixel_variance32x32_neon(const uint8_t *src,
- int src_stride, int xoffset,
- int yoffset, const uint8_t *dst,
- int dst_stride,
- unsigned int *sse) {- DECLARE_ALIGNED(16, uint8_t, temp2[32 * 32]);
- DECLARE_ALIGNED(16, uint8_t, fdata3[33 * 32]);
-
- var_filter_block2d_bil_w16(src, fdata3, src_stride, 1, 33, 32,
- bilinear_filters[xoffset]);
- var_filter_block2d_bil_w16(fdata3, temp2, 32, 32, 32, 32,
- bilinear_filters[yoffset]);
- return vpx_variance32x32(temp2, 32, dst, dst_stride, sse);
-}
-
-unsigned int vpx_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(16, uint8_t, temp2[64 * 64]);
- DECLARE_ALIGNED(16, uint8_t, fdata3[65 * 64]);
-
- var_filter_block2d_bil_w16(src, fdata3, src_stride, 1, 65, 64,
- bilinear_filters[xoffset]);
- var_filter_block2d_bil_w16(fdata3, temp2, 64, 64, 64, 64,
- bilinear_filters[yoffset]);
- return vpx_variance64x64(temp2, 64, dst, dst_stride, sse);
-}
+sub_pixel_varianceNxM(8, 4);
+sub_pixel_varianceNxM(8, 8);
+sub_pixel_varianceNxM(8, 16);
+sub_pixel_varianceNxM(16, 8);
+sub_pixel_varianceNxM(16, 16);
+sub_pixel_varianceNxM(16, 32);
+sub_pixel_varianceNxM(32, 16);
+sub_pixel_varianceNxM(32, 32);
+sub_pixel_varianceNxM(32, 64);
+sub_pixel_varianceNxM(64, 32);
+sub_pixel_varianceNxM(64, 64);
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -1240,34 +1240,34 @@
specialize qw/vpx_sub_pixel_variance64x64 avx2 neon msa sse2 ssse3/;
add_proto qw/uint32_t vpx_sub_pixel_variance64x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
- specialize qw/vpx_sub_pixel_variance64x32 msa sse2 ssse3/;
+ specialize qw/vpx_sub_pixel_variance64x32 neon msa sse2 ssse3/;
add_proto qw/uint32_t vpx_sub_pixel_variance32x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
- specialize qw/vpx_sub_pixel_variance32x64 msa sse2 ssse3/;
+ specialize qw/vpx_sub_pixel_variance32x64 neon msa sse2 ssse3/;
add_proto qw/uint32_t vpx_sub_pixel_variance32x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
specialize qw/vpx_sub_pixel_variance32x32 avx2 neon msa sse2 ssse3/;
add_proto qw/uint32_t vpx_sub_pixel_variance32x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
- specialize qw/vpx_sub_pixel_variance32x16 msa sse2 ssse3/;
+ specialize qw/vpx_sub_pixel_variance32x16 neon msa sse2 ssse3/;
add_proto qw/uint32_t vpx_sub_pixel_variance16x32/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
- specialize qw/vpx_sub_pixel_variance16x32 msa sse2 ssse3/;
+ specialize qw/vpx_sub_pixel_variance16x32 neon msa sse2 ssse3/;
add_proto qw/uint32_t vpx_sub_pixel_variance16x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
specialize qw/vpx_sub_pixel_variance16x16 neon msa sse2 ssse3/;
add_proto qw/uint32_t vpx_sub_pixel_variance16x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
- specialize qw/vpx_sub_pixel_variance16x8 msa sse2 ssse3/;
+ specialize qw/vpx_sub_pixel_variance16x8 neon msa sse2 ssse3/;
add_proto qw/uint32_t vpx_sub_pixel_variance8x16/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
- specialize qw/vpx_sub_pixel_variance8x16 msa sse2 ssse3/;
+ specialize qw/vpx_sub_pixel_variance8x16 neon msa sse2 ssse3/;
add_proto qw/uint32_t vpx_sub_pixel_variance8x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
specialize qw/vpx_sub_pixel_variance8x8 neon msa sse2 ssse3/;
add_proto qw/uint32_t vpx_sub_pixel_variance8x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
- specialize qw/vpx_sub_pixel_variance8x4 msa sse2 ssse3/;
+ specialize qw/vpx_sub_pixel_variance8x4 neon msa sse2 ssse3/;
add_proto qw/uint32_t vpx_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
specialize qw/vpx_sub_pixel_variance4x8 msa sse2 ssse3/;