Optimize and cleanup supertx predictor.
Use vpx_blend_a64_hmask and vpx_blend_a64_vmask to speed up
computing the supertx predictor.
Decoder speedup of up to 4% has been observed.
Change-Id: I255a5ba4cc24f78dc905d25b6e2f7fbafac13253
diff --git a/vp10/common/reconinter.c b/vp10/common/reconinter.c
index 3a196a5..53fd1a6 100644
--- a/vp10/common/reconinter.c
+++ b/vp10/common/reconinter.c
@@ -1067,177 +1067,123 @@
28, 18, 10, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
};
-static void generate_1dmask(int length, uint8_t *mask, int plane) {
+static const uint8_t* get_supertx_mask(int length, int plane) {
switch (length) {
case 8:
- memcpy(mask, plane ? mask_8_uv : mask_8, length);
- break;
+ return plane ? mask_8_uv : mask_8;
case 16:
- memcpy(mask, plane ? mask_16_uv : mask_16, length);
- break;
+ return plane ? mask_16_uv : mask_16;
case 32:
- memcpy(mask, plane ? mask_32_uv : mask_32, length);
- break;
+ return plane ? mask_32_uv : mask_32;
default:
assert(0);
}
+ return NULL;
}
void vp10_build_masked_inter_predictor_complex(
MACROBLOCKD *xd,
- uint8_t *dst, int dst_stride, uint8_t *dst2, int dst2_stride,
+ uint8_t *dst, int dst_stride,
+ const uint8_t *pre, int pre_stride,
int mi_row, int mi_col,
int mi_row_ori, int mi_col_ori, BLOCK_SIZE bsize, BLOCK_SIZE top_bsize,
PARTITION_TYPE partition, int plane) {
- int i, j;
const struct macroblockd_plane *pd = &xd->plane[plane];
- uint8_t mask[MAX_TX_SIZE];
- int top_w = 4 << b_width_log2_lookup[top_bsize];
- int top_h = 4 << b_height_log2_lookup[top_bsize];
- int w = 4 << b_width_log2_lookup[bsize];
- int h = 4 << b_height_log2_lookup[bsize];
- int w_offset = (mi_col - mi_col_ori) * MI_SIZE;
- int h_offset = (mi_row - mi_row_ori) * MI_SIZE;
+ const int ssx = pd->subsampling_x;
+ const int ssy = pd->subsampling_y;
+ const int top_w = (4 << b_width_log2_lookup[top_bsize]) >> ssx;
+ const int top_h = (4 << b_height_log2_lookup[top_bsize]) >> ssy;
+ const int w = (4 << b_width_log2_lookup[bsize]) >> ssx;
+ const int h = (4 << b_height_log2_lookup[bsize]) >> ssy;
+ const int w_offset = ((mi_col - mi_col_ori) * MI_SIZE) >> ssx;
+ const int h_offset = ((mi_row - mi_row_ori) * MI_SIZE) >> ssy;
+
+ int w_remain, h_remain;
#if CONFIG_VP9_HIGHBITDEPTH
- uint16_t *dst16= CONVERT_TO_SHORTPTR(dst);
- uint16_t *dst216 = CONVERT_TO_SHORTPTR(dst2);
- int b_hdb = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? 1 : 0;
+ const int is_hdb = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? 1 : 0;
#endif // CONFIG_VP9_HIGHBITDEPTH
assert(bsize <= BLOCK_32X32);
-
- top_w >>= pd->subsampling_x;
- top_h >>= pd->subsampling_y;
- w >>= pd->subsampling_x;
- h >>= pd->subsampling_y;
- w_offset >>= pd->subsampling_x;
- h_offset >>= pd->subsampling_y;
+ assert(IMPLIES(plane == 0, ssx == 0));
+ assert(IMPLIES(plane == 0, ssy == 0));
switch (partition) {
- case PARTITION_HORZ:
- {
+ case PARTITION_HORZ: {
+ const uint8_t *const mask = get_supertx_mask(h, ssy);
+
+ w_remain = top_w;
+ h_remain = top_h - h_offset - h;
+ dst += h_offset * dst_stride;
+ pre += h_offset * pre_stride;
+
#if CONFIG_VP9_HIGHBITDEPTH
- if (b_hdb) {
- uint16_t *dst_tmp = dst16 + h_offset * dst_stride;
- uint16_t *dst2_tmp = dst216 + h_offset * dst2_stride;
- generate_1dmask(h, mask + h_offset,
- plane && xd->plane[plane].subsampling_y);
-
- for (i = h_offset; i < h_offset + h; i++) {
- for (j = 0; j < top_w; j++) {
- const int m = mask[i]; assert(m >= 0 && m <= 64);
- if (m == 64)
- continue;
-
- if (m == 0)
- dst_tmp[j] = dst2_tmp[j];
- else
- dst_tmp[j] = ROUND_POWER_OF_TWO(dst_tmp[j] * m +
- dst2_tmp[j] * (64 - m), 6);
- }
- dst_tmp += dst_stride;
- dst2_tmp += dst2_stride;
- }
-
- for (; i < top_h; i ++) {
- memcpy(dst_tmp, dst2_tmp, top_w * sizeof(uint16_t));
- dst_tmp += dst_stride;
- dst2_tmp += dst2_stride;
- }
- } else {
+ if (is_hdb)
+ vpx_highbd_blend_a64_vmask(dst, dst_stride,
+ dst, dst_stride,
+ pre, pre_stride,
+ mask, h, top_w, xd->bd);
+ else
#endif // CONFIG_VP9_HIGHBITDEPTH
- uint8_t *dst_tmp = dst + h_offset * dst_stride;
- uint8_t *dst2_tmp = dst2 + h_offset * dst2_stride;
- generate_1dmask(h, mask + h_offset,
- plane && xd->plane[plane].subsampling_y);
+ vpx_blend_a64_vmask(dst, dst_stride,
+ dst, dst_stride,
+ pre, pre_stride,
+ mask, h, top_w);
- for (i = h_offset; i < h_offset + h; i++) {
- for (j = 0; j < top_w; j++) {
- const int m = mask[i]; assert(m >= 0 && m <= 64);
- if (m == 64)
- continue;
-
- if (m == 0)
- dst_tmp[j] = dst2_tmp[j];
- else
- dst_tmp[j] = ROUND_POWER_OF_TWO(dst_tmp[j] * m +
- dst2_tmp[j] * (64 - m), 6);
- }
- dst_tmp += dst_stride;
- dst2_tmp += dst2_stride;
- }
-
- for (; i < top_h; i ++) {
- memcpy(dst_tmp, dst2_tmp, top_w * sizeof(uint8_t));
- dst_tmp += dst_stride;
- dst2_tmp += dst2_stride;
- }
-#if CONFIG_VP9_HIGHBITDEPTH
- }
-#endif // CONFIG_VP9_HIGHBITDEPTH
- }
-
+ dst += h * dst_stride;
+ pre += h * pre_stride;
break;
- case PARTITION_VERT:
- {
-#if CONFIG_VP9_HIGHBITDEPTH
- if (b_hdb) {
- uint16_t *dst_tmp = dst16;
- uint16_t *dst2_tmp = dst216;
- generate_1dmask(w, mask + w_offset,
- plane && xd->plane[plane].subsampling_x);
-
- for (i = 0; i < top_h; i++) {
- for (j = w_offset; j < w_offset + w; j++) {
- const int m = mask[j]; assert(m >= 0 && m <= 64);
- if (m == 64)
- continue;
-
- if (m == 0)
- dst_tmp[j] = dst2_tmp[j];
- else
- dst_tmp[j] = ROUND_POWER_OF_TWO(dst_tmp[j] * m +
- dst2_tmp[j] * (64 - m), 6);
- }
- memcpy(dst_tmp + j, dst2_tmp + j,
- (top_w - w_offset - w) * sizeof(uint16_t));
- dst_tmp += dst_stride;
- dst2_tmp += dst2_stride;
- }
- } else {
-#endif // CONFIG_VP9_HIGHBITDEPTH
- uint8_t *dst_tmp = dst;
- uint8_t *dst2_tmp = dst2;
- generate_1dmask(w, mask + w_offset,
- plane && xd->plane[plane].subsampling_x);
-
- for (i = 0; i < top_h; i++) {
- for (j = w_offset; j < w_offset + w; j++) {
- const int m = mask[j]; assert(m >= 0 && m <= 64);
- if (m == 64)
- continue;
-
- if (m == 0)
- dst_tmp[j] = dst2_tmp[j];
- else
- dst_tmp[j] = ROUND_POWER_OF_TWO(dst_tmp[j] * m +
- dst2_tmp[j] * (64 - m), 6);
- }
- memcpy(dst_tmp + j, dst2_tmp + j,
- (top_w - w_offset - w) * sizeof(uint8_t));
- dst_tmp += dst_stride;
- dst2_tmp += dst2_stride;
- }
-#if CONFIG_VP9_HIGHBITDEPTH
- }
-#endif // CONFIG_VP9_HIGHBITDEPTH
}
+ case PARTITION_VERT: {
+ const uint8_t *const mask = get_supertx_mask(w, ssx);
+
+ w_remain = top_w - w_offset - w;
+ h_remain = top_h;
+ dst += w_offset;
+ pre += w_offset;
+
+#if CONFIG_VP9_HIGHBITDEPTH
+ if (is_hdb)
+ vpx_highbd_blend_a64_hmask(dst, dst_stride,
+ dst, dst_stride,
+ pre, pre_stride,
+ mask, top_h, w, xd->bd);
+ else
+#endif // CONFIG_VP9_HIGHBITDEPTH
+ vpx_blend_a64_hmask(dst, dst_stride,
+ dst, dst_stride,
+ pre, pre_stride,
+ mask, top_h, w);
+
+ dst += w;
+ pre += w;
break;
- default:
+ }
+ default: {
assert(0);
+ return;
+ }
}
- (void) xd;
+
+ if (w_remain == 0 || h_remain == 0) {
+ return;
+ }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+ if (is_hdb) {
+ dst = (uint8_t*)CONVERT_TO_SHORTPTR(dst);
+ pre = (const uint8_t*)CONVERT_TO_SHORTPTR(pre);
+ dst_stride *= 2;
+ pre_stride *= 2;
+ w_remain *= 2;
+ }
+#endif // CONFIG_VP9_HIGHBITDEPTH
+
+ do {
+ memcpy(dst, pre, w_remain * sizeof(uint8_t));
+ dst += dst_stride;
+ pre += pre_stride;
+ } while (--h_remain);
}
void vp10_build_inter_predictors_sb_sub8x8_extend(
diff --git a/vp10/common/reconinter.h b/vp10/common/reconinter.h
index 7321831..ac4a004 100644
--- a/vp10/common/reconinter.h
+++ b/vp10/common/reconinter.h
@@ -396,7 +396,8 @@
struct macroblockd_plane;
void vp10_build_masked_inter_predictor_complex(
MACROBLOCKD *xd,
- uint8_t *dst, int dst_stride, uint8_t *dst2, int dst2_stride,
+ uint8_t *dst, int dst_stride,
+ const uint8_t *pre, int pre_stride,
int mi_row, int mi_col, int mi_row_ori, int mi_col_ori,
BLOCK_SIZE bsize, BLOCK_SIZE top_bsize,
PARTITION_TYPE partition, int plane);
