Specialize Neon OBMC subpel variance by filter value
Use the same specialization for Neon OBMC subpel variance functions as
used for the other Neon subpel variants. The rationale for the
specialization is as follows:
The optimal implementation of the bilinear interpolation depends on
the filter values being used. For both horizontal and vertical
interpolation this can simplify to just taking the source values, or
averaging the source and reference values - which can be computed
more easily than a bilinear interpolation with arbitrary filter
values.
This patch introduces tests to find the most optimal bilinear
interpolation implementation based on the filter values being used.
This new specialization is only used for larger block sizes
(>= 16x16) as we need to be doing enough work to make the cost of
finding the optimal implementation worth it.
Change-Id: If90f3b6120fe2a2a1a7a834dba96a73224224f30
diff --git a/aom_dsp/arm/subpel_variance_neon.c b/aom_dsp/arm/subpel_variance_neon.c
index 7946439..07f5801 100644
--- a/aom_dsp/arm/subpel_variance_neon.c
+++ b/aom_dsp/arm/subpel_variance_neon.c
@@ -564,6 +564,61 @@
return aom_obmc_variance##w##x##h(tmp1, w, wsrc, mask, sse); \
}
+#define SPECIALIZED_OBMC_SUBPEL_VARIANCE_WXH_NEON(w, h, padding) \
+ unsigned int aom_obmc_sub_pixel_variance##w##x##h##_neon( \
+ const uint8_t *pre, int pre_stride, int xoffset, int yoffset, \
+ const int32_t *wsrc, const int32_t *mask, unsigned int *sse) { \
+ if (xoffset == 0) { \
+ if (yoffset == 0) { \
+ return aom_obmc_variance##w##x##h##_neon(pre, pre_stride, wsrc, mask, \
+ sse); \
+ } else if (yoffset == 4) { \
+ uint8_t tmp[w * h]; \
+ var_filter_block2d_avg(pre, tmp, pre_stride, pre_stride, w, h); \
+ return aom_obmc_variance##w##x##h##_neon(tmp, w, wsrc, mask, sse); \
+ } else { \
+ uint8_t tmp[w * h]; \
+ var_filter_block2d_bil_w##w(pre, tmp, pre_stride, pre_stride, h, \
+ yoffset); \
+ return aom_obmc_variance##w##x##h##_neon(tmp, w, wsrc, mask, sse); \
+ } \
+ } else if (xoffset == 4) { \
+ uint8_t tmp0[w * (h + padding)]; \
+ if (yoffset == 0) { \
+ var_filter_block2d_avg(pre, tmp0, pre_stride, 1, w, h); \
+ return aom_obmc_variance##w##x##h##_neon(tmp0, w, wsrc, mask, sse); \
+ } else if (yoffset == 4) { \
+ uint8_t tmp1[w * (h + padding)]; \
+ var_filter_block2d_avg(pre, tmp0, pre_stride, 1, w, h + padding); \
+ var_filter_block2d_avg(tmp0, tmp1, w, w, w, h); \
+ return aom_obmc_variance##w##x##h##_neon(tmp1, w, wsrc, mask, sse); \
+ } else { \
+ uint8_t tmp1[w * (h + padding)]; \
+ var_filter_block2d_avg(pre, tmp0, pre_stride, 1, w, h + padding); \
+ var_filter_block2d_bil_w##w(tmp0, tmp1, w, w, h, yoffset); \
+ return aom_obmc_variance##w##x##h##_neon(tmp1, w, wsrc, mask, sse); \
+ } \
+ } else { \
+ uint8_t tmp0[w * (h + padding)]; \
+ if (yoffset == 0) { \
+ var_filter_block2d_bil_w##w(pre, tmp0, pre_stride, 1, h, xoffset); \
+ return aom_obmc_variance##w##x##h##_neon(tmp0, w, wsrc, mask, sse); \
+ } else if (yoffset == 4) { \
+ uint8_t tmp1[w * h]; \
+ var_filter_block2d_bil_w##w(pre, tmp0, pre_stride, 1, h + padding, \
+ xoffset); \
+ var_filter_block2d_avg(tmp0, tmp1, w, w, w, h); \
+ return aom_obmc_variance##w##x##h##_neon(tmp1, w, wsrc, mask, sse); \
+ } else { \
+ uint8_t tmp1[w * h]; \
+ var_filter_block2d_bil_w##w(pre, tmp0, pre_stride, 1, h + padding, \
+ xoffset); \
+ var_filter_block2d_bil_w##w(tmp0, tmp1, w, w, h, yoffset); \
+ return aom_obmc_variance##w##x##h##_neon(tmp1, w, wsrc, mask, sse); \
+ } \
+ } \
+ }
+
OBMC_SUBPEL_VARIANCE_WXH_NEON(4, 4, 2)
OBMC_SUBPEL_VARIANCE_WXH_NEON(4, 8, 2)
OBMC_SUBPEL_VARIANCE_WXH_NEON(4, 16, 2)
@@ -575,22 +630,23 @@
OBMC_SUBPEL_VARIANCE_WXH_NEON(16, 4, 1)
OBMC_SUBPEL_VARIANCE_WXH_NEON(16, 8, 1)
-OBMC_SUBPEL_VARIANCE_WXH_NEON(16, 16, 1)
-OBMC_SUBPEL_VARIANCE_WXH_NEON(16, 32, 1)
-OBMC_SUBPEL_VARIANCE_WXH_NEON(16, 64, 1)
+SPECIALIZED_OBMC_SUBPEL_VARIANCE_WXH_NEON(16, 16, 1)
+SPECIALIZED_OBMC_SUBPEL_VARIANCE_WXH_NEON(16, 32, 1)
+SPECIALIZED_OBMC_SUBPEL_VARIANCE_WXH_NEON(16, 64, 1)
-OBMC_SUBPEL_VARIANCE_WXH_NEON(32, 8, 1)
-OBMC_SUBPEL_VARIANCE_WXH_NEON(32, 16, 1)
-OBMC_SUBPEL_VARIANCE_WXH_NEON(32, 32, 1)
-OBMC_SUBPEL_VARIANCE_WXH_NEON(32, 64, 1)
+SPECIALIZED_OBMC_SUBPEL_VARIANCE_WXH_NEON(32, 8, 1)
+SPECIALIZED_OBMC_SUBPEL_VARIANCE_WXH_NEON(32, 16, 1)
+SPECIALIZED_OBMC_SUBPEL_VARIANCE_WXH_NEON(32, 32, 1)
+SPECIALIZED_OBMC_SUBPEL_VARIANCE_WXH_NEON(32, 64, 1)
-OBMC_SUBPEL_VARIANCE_WXH_NEON(64, 16, 1)
-OBMC_SUBPEL_VARIANCE_WXH_NEON(64, 32, 1)
-OBMC_SUBPEL_VARIANCE_WXH_NEON(64, 64, 1)
-OBMC_SUBPEL_VARIANCE_WXH_NEON(64, 128, 1)
+SPECIALIZED_OBMC_SUBPEL_VARIANCE_WXH_NEON(64, 16, 1)
+SPECIALIZED_OBMC_SUBPEL_VARIANCE_WXH_NEON(64, 32, 1)
+SPECIALIZED_OBMC_SUBPEL_VARIANCE_WXH_NEON(64, 64, 1)
+SPECIALIZED_OBMC_SUBPEL_VARIANCE_WXH_NEON(64, 128, 1)
-OBMC_SUBPEL_VARIANCE_WXH_NEON(128, 64, 1)
-OBMC_SUBPEL_VARIANCE_WXH_NEON(128, 128, 1)
+SPECIALIZED_OBMC_SUBPEL_VARIANCE_WXH_NEON(128, 64, 1)
+SPECIALIZED_OBMC_SUBPEL_VARIANCE_WXH_NEON(128, 128, 1)
#undef OBMC_SUBPEL_VARIANCE_WXH_NEON
+#undef SPECIALIZED_OBMC_SUBPEL_VARIANCE_WXH_NEON
#endif // !CONFIG_REALTIME_ONLY
