Add scaled convolve functions for convolve_round
Change-Id: I343d94520955b6632c2a990af26454859b29a97e
diff --git a/av1/common/av1_rtcd_defs.pl b/av1/common/av1_rtcd_defs.pl
index 999d975..5c9836d 100755
--- a/av1/common/av1_rtcd_defs.pl
+++ b/av1/common/av1_rtcd_defs.pl
@@ -620,11 +620,14 @@
add_proto qw/void av1_convolve_rounding/, "const int32_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, int bits";
specialize qw/av1_convolve_rounding avx2/;
+ add_proto qw/void av1_convolve_2d_scale/, "const uint8_t *src, int src_stride, CONV_BUF_TYPE *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_q4, const int y_step_qn, ConvolveParams *conv_params";
+
if (aom_config("CONFIG_HIGHBITDEPTH") eq "yes") {
add_proto qw/void av1_highbd_convolve_2d/, "const uint16_t *src, int src_stride, CONV_BUF_TYPE *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd";
specialize qw/av1_highbd_convolve_2d ssse3/;
add_proto qw/void av1_highbd_convolve_rounding/, "const int32_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, int bits, int bd";
specialize qw/av1_highbd_convolve_rounding avx2/;
+ add_proto qw/void av1_highbd_convolve_2d_scale/, "const uint16_t *src, int src_stride, CONV_BUF_TYPE *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int x_step_qn, const int subpel_y_q4, const int y_step_qn, ConvolveParams *conv_params, int bd";
}
}
diff --git a/av1/common/convolve.c b/av1/common/convolve.c
index 66e6898..bc76a3c 100644
--- a/av1/common/convolve.c
+++ b/av1/common/convolve.c
@@ -63,7 +63,7 @@
ConvolveParams *conv_params) {
int x, y;
int filter_size = filter_params.taps;
- // assert(conv_params->round == CONVOLVE_OPT_ROUND);
+ assert(conv_params->round == CONVOLVE_OPT_ROUND);
src -= filter_size / 2 - 1;
for (y = 0; y < h; ++y) {
int x_qn = subpel_x_qn;
@@ -128,7 +128,7 @@
ConvolveParams *conv_params) {
int x, y;
int filter_size = filter_params.taps;
- // assert(conv_params->round == CONVOLVE_OPT_ROUND);
+ assert(conv_params->round == CONVOLVE_OPT_ROUND);
src -= src_stride * (filter_size / 2 - 1);
for (x = 0; x < w; ++x) {
int y_qn = subpel_y_qn;
@@ -226,7 +226,7 @@
const InterpFilterParams filter_params,
const int subpel_x_qn, int x_step_qn,
ConvolveParams *conv_params) {
- // assert(conv_params->round == CONVOLVE_OPT_ROUND);
+ assert(conv_params->round == CONVOLVE_OPT_ROUND);
if (filter_params.taps == SUBPEL_TAPS) {
const int16_t *filter_x = av1_get_interp_filter_subpel_kernel(
filter_params, subpel_x_qn >> SCALE_EXTRA_BITS);
@@ -292,7 +292,7 @@
const InterpFilterParams filter_params,
const int subpel_y_qn, int y_step_qn,
ConvolveParams *conv_params) {
- // assert(conv_params->round == CONVOLVE_OPT_ROUND);
+ assert(conv_params->round == CONVOLVE_OPT_ROUND);
if (filter_params.taps == SUBPEL_TAPS) {
const int16_t *filter_y = av1_get_interp_filter_subpel_kernel(
filter_params, subpel_y_qn >> SCALE_EXTRA_BITS);
@@ -365,7 +365,65 @@
}
}
}
+
+void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride,
+ CONV_BUF_TYPE *dst, int dst_stride, int w, int h,
+ InterpFilterParams *filter_params_x,
+ InterpFilterParams *filter_params_y,
+ const int subpel_x_qn, const int x_step_qn,
+ const int subpel_y_qn, const int y_step_qn,
+ ConvolveParams *conv_params) {
+ int x, y, k;
+ uint8_t im_block[(2 * MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE];
+ int im_h = (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) +
+ filter_params_y->taps;
+ int im_stride = w;
+ const int fo_vert = filter_params_y->taps / 2 - 1;
+ const int fo_horiz = filter_params_x->taps / 2 - 1;
+
+ // horizontal filter
+ const uint8_t *src_horiz = src - fo_vert * src_stride;
+ for (y = 0; y < im_h; ++y) {
+ int x_qn = subpel_x_qn;
+ for (x = 0; x < w; ++x, x_qn += x_step_qn) {
+ const uint8_t *const src_x = &src_horiz[(x_qn >> SCALE_SUBPEL_BITS)];
+ const int x_filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
+ assert(x_filter_idx < SUBPEL_SHIFTS);
+ const int16_t *x_filter =
+ av1_get_interp_filter_subpel_kernel(*filter_params_x, x_filter_idx);
+ int sum = 0;
+ for (k = 0; k < filter_params_x->taps; ++k)
+ sum += x_filter[k] * src_x[k - fo_horiz];
+ im_block[y * im_stride + x] =
+ clip_pixel(ROUND_POWER_OF_TWO(sum, conv_params->round_0));
+ }
+ src_horiz += src_stride;
+ }
+
+ // vertical filter
+ const uint8_t *src_vert = im_block + fo_vert * im_stride;
+ for (x = 0; x < w; ++x) {
+ int y_qn = subpel_y_qn;
+ for (y = 0; y < h; ++y, y_qn += y_step_qn) {
+ const uint8_t *const src_y =
+ &src_vert[(y_qn >> SCALE_SUBPEL_BITS) * im_stride];
+ const int y_filter_idx = (y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
+ assert(y_filter_idx < SUBPEL_SHIFTS);
+ const int16_t *y_filter =
+ av1_get_interp_filter_subpel_kernel(*filter_params_y, y_filter_idx);
+ CONV_BUF_TYPE sum = 0;
+ for (k = 0; k < filter_params_y->taps; ++k) {
+ sum += y_filter[k] * src_y[(k - fo_vert) * im_stride];
+ }
+ CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1);
+ dst[y * dst_stride + x] += res;
+ }
+ src_vert++;
+ }
+}
+
#else
+
/* When convolve-round is enabled and compound-round is disabled, we use a
high-precision convolve filter.
Note: For notes on hardware implementations, including the required
@@ -421,13 +479,75 @@
}
}
}
-#endif
+
+void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride,
+ CONV_BUF_TYPE *dst, int dst_stride, int w, int h,
+ InterpFilterParams *filter_params_x,
+ InterpFilterParams *filter_params_y,
+ const int subpel_x_qn, const int x_step_qn,
+ const int subpel_y_qn, const int y_step_qn,
+ ConvolveParams *conv_params) {
+ int x, y, k;
+ int32_t im_block[(2 * MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE];
+ int im_h = (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) +
+ filter_params_y->taps;
+ int im_stride = w;
+ const int fo_vert = filter_params_y->taps / 2 - 1;
+ const int fo_horiz = filter_params_x->taps / 2 - 1;
+ const int bd = 8;
+
+ // horizontal filter
+ const uint8_t *src_horiz = src - fo_vert * src_stride;
+ for (y = 0; y < im_h; ++y) {
+ int x_qn = subpel_x_qn;
+ for (x = 0; x < w; ++x, x_qn += x_step_qn) {
+ const uint8_t *const src_x = &src_horiz[(x_qn >> SCALE_SUBPEL_BITS)];
+ const int x_filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
+ assert(x_filter_idx < SUBPEL_SHIFTS);
+ const int16_t *x_filter =
+ av1_get_interp_filter_subpel_kernel(*filter_params_x, x_filter_idx);
+ int32_t sum = (1 << (bd + FILTER_BITS - 1));
+ for (k = 0; k < filter_params_x->taps; ++k) {
+ sum += x_filter[k] * src_x[k - fo_horiz];
+ }
+ assert(0 <= sum && sum < (1 << (bd + FILTER_BITS + 1)));
+ im_block[y * im_stride + x] =
+ ROUND_POWER_OF_TWO(sum, conv_params->round_0);
+ }
+ src_horiz += src_stride;
+ }
+
+ // vertical filter
+ int32_t *src_vert = im_block + fo_vert * im_stride;
+ const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+ for (x = 0; x < w; ++x) {
+ int y_qn = subpel_y_qn;
+ for (y = 0; y < h; ++y, y_qn += y_step_qn) {
+ const int32_t *src_y = &src_vert[(y_qn >> SCALE_SUBPEL_BITS) * im_stride];
+ const int y_filter_idx = (y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
+ assert(y_filter_idx < SUBPEL_SHIFTS);
+ const int16_t *y_filter =
+ av1_get_interp_filter_subpel_kernel(*filter_params_y, y_filter_idx);
+ CONV_BUF_TYPE sum = 1 << offset_bits;
+ for (k = 0; k < filter_params_y->taps; ++k) {
+ sum += y_filter[k] * src_y[(k - fo_vert) * im_stride];
+ }
+ assert(0 <= sum && sum < (1 << (offset_bits + 2)));
+ CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1) -
+ ((1 << (offset_bits - conv_params->round_1)) +
+ (1 << (offset_bits - conv_params->round_1 - 1)));
+ dst[y * dst_stride + x] += res;
+ }
+ src_vert++;
+ }
+}
+#endif // CONFIG_COMPOUND_ROUND
void av1_convolve_2d_facade(const uint8_t *src, int src_stride, uint8_t *dst,
int dst_stride, int w, int h,
const InterpFilter *interp_filter,
const int subpel_x_q4, int x_step_q4,
- const int subpel_y_q4, int y_step_q4,
+ const int subpel_y_q4, int y_step_q4, int scaled,
ConvolveParams *conv_params) {
(void)x_step_q4;
(void)y_step_q4;
@@ -470,15 +590,28 @@
conv_params->dst_stride, w, h);
// horizontal and vertical parameters are swapped because of the transpose
- av1_convolve_2d(tr_src + fo_horiz * tr_src_stride + fo_vert, tr_src_stride,
- tr_dst, tr_dst_stride, h, w, &filter_params_y,
- &filter_params_x, subpel_y_q4, subpel_x_q4, conv_params);
+ if (scaled)
+ av1_convolve_2d_scale(tr_src + fo_horiz * tr_src_stride + fo_vert,
+ tr_src_stride, tr_dst, tr_dst_stride, h, w,
+ &filter_params_y, &filter_params_x, subpel_y_q4,
+ y_step_q4, subpel_x_q4, x_step_q4, conv_params);
+ else
+ av1_convolve_2d(tr_src + fo_horiz * tr_src_stride + fo_vert,
+ tr_src_stride, tr_dst, tr_dst_stride, h, w,
+ &filter_params_y, &filter_params_x, subpel_y_q4,
+ subpel_x_q4, conv_params);
transpose_int32(conv_params->dst, conv_params->dst_stride, tr_dst,
tr_dst_stride, h, w);
} else {
- av1_convolve_2d(src, src_stride, conv_params->dst, conv_params->dst_stride,
- w, h, &filter_params_x, &filter_params_y, subpel_x_q4,
- subpel_y_q4, conv_params);
+ if (scaled)
+ av1_convolve_2d_scale(src, src_stride, conv_params->dst,
+ conv_params->dst_stride, w, h, &filter_params_x,
+ &filter_params_y, subpel_x_q4, x_step_q4,
+ subpel_y_q4, y_step_q4, conv_params);
+ else
+ av1_convolve_2d(src, src_stride, conv_params->dst,
+ conv_params->dst_stride, w, h, &filter_params_x,
+ &filter_params_y, subpel_x_q4, subpel_y_q4, conv_params);
}
}
@@ -540,7 +673,66 @@
}
}
}
+
+void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride,
+ CONV_BUF_TYPE *dst, int dst_stride, int w,
+ int h, InterpFilterParams *filter_params_x,
+ InterpFilterParams *filter_params_y,
+ const int subpel_x_qn, const int x_step_qn,
+ const int subpel_y_qn, const int y_step_qn,
+ ConvolveParams *conv_params, int bd) {
+ int x, y, k;
+ uint16_t im_block[(2 * MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE];
+ int im_h = (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) +
+ filter_params_y->taps;
+ int im_stride = w;
+ const int fo_vert = filter_params_y->taps / 2 - 1;
+ const int fo_horiz = filter_params_x->taps / 2 - 1;
+ (void)bd;
+
+ // horizontal filter
+ const uint16_t *src_horiz = src - fo_vert * src_stride;
+ for (y = 0; y < im_h; ++y) {
+ int x_qn = subpel_x_qn;
+ for (x = 0; x < w; ++x, x_qn += x_step_qn) {
+ const uint16_t *const src_x = &src_horiz[(x_qn >> SCALE_SUBPEL_BITS)];
+ const int x_filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
+ assert(x_filter_idx < SUBPEL_SHIFTS);
+ const int16_t *x_filter =
+ av1_get_interp_filter_subpel_kernel(*filter_params_x, x_filter_idx);
+ int sum = 0;
+ for (k = 0; k < filter_params_x->taps; ++k)
+ sum += x_filter[k] * src_x[k - fo_horiz];
+ im_block[y * im_stride + x] =
+ clip_pixel(ROUND_POWER_OF_TWO(sum, conv_params->round_0));
+ }
+ src_horiz += src_stride;
+ }
+
+ // vertical filter
+ uint16_t *src_vert = im_block + fo_vert * im_stride;
+ for (x = 0; x < w; ++x) {
+ int y_qn = subpel_y_qn;
+ for (y = 0; y < h; ++y, y_qn += y_step_qn) {
+ const uint16_t *const src_y =
+ &src_vert[(y_qn >> SCALE_SUBPEL_BITS) * im_stride];
+ const int y_filter_idx = (y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
+ assert(y_filter_idx < SUBPEL_SHIFTS);
+ const int16_t *y_filter =
+ av1_get_interp_filter_subpel_kernel(*filter_params_y, y_filter_idx);
+ CONV_BUF_TYPE sum = 0;
+ for (k = 0; k < filter_params_y->taps; ++k) {
+ sum += y_filter[k] * src_y[(k - fo_vert) * im_stride];
+ }
+ CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1);
+ dst[y * dst_stride + x] += res;
+ }
+ src_vert++;
+ }
+}
+
#else
+
void av1_highbd_convolve_2d_c(const uint16_t *src, int src_stride,
CONV_BUF_TYPE *dst, int dst_stride, int w, int h,
InterpFilterParams *filter_params_x,
@@ -590,14 +782,76 @@
}
}
}
-#endif
+
+void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride,
+ CONV_BUF_TYPE *dst, int dst_stride, int w,
+ int h, InterpFilterParams *filter_params_x,
+ InterpFilterParams *filter_params_y,
+ const int subpel_x_qn, const int x_step_qn,
+ const int subpel_y_qn, const int y_step_qn,
+ ConvolveParams *conv_params, int bd) {
+ int x, y, k;
+ int32_t im_block[(2 * MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE];
+ int im_h = (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) +
+ filter_params_y->taps;
+ int im_stride = w;
+ const int fo_vert = filter_params_y->taps / 2 - 1;
+ const int fo_horiz = filter_params_x->taps / 2 - 1;
+
+ // horizontal filter
+ const uint16_t *src_horiz = src - fo_vert * src_stride;
+ for (y = 0; y < im_h; ++y) {
+ int x_qn = subpel_x_qn;
+ for (x = 0; x < w; ++x, x_qn += x_step_qn) {
+ const uint16_t *const src_x = &src_horiz[(x_qn >> SCALE_SUBPEL_BITS)];
+ const int x_filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
+ assert(x_filter_idx < SUBPEL_SHIFTS);
+ const int16_t *x_filter =
+ av1_get_interp_filter_subpel_kernel(*filter_params_x, x_filter_idx);
+ int32_t sum = (1 << (bd + FILTER_BITS - 1));
+ for (k = 0; k < filter_params_x->taps; ++k) {
+ sum += x_filter[k] * src_x[k - fo_horiz];
+ }
+ assert(0 <= sum && sum < (1 << (bd + FILTER_BITS + 1)));
+ im_block[y * im_stride + x] =
+ ROUND_POWER_OF_TWO(sum, conv_params->round_0);
+ }
+ src_horiz += src_stride;
+ }
+
+ // vertical filter
+ int32_t *src_vert = im_block + fo_vert * im_stride;
+ const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+ for (x = 0; x < w; ++x) {
+ int y_qn = subpel_y_qn;
+ for (y = 0; y < h; ++y, y_qn += y_step_qn) {
+ const int32_t *src_y = &src_vert[(y_qn >> SCALE_SUBPEL_BITS) * im_stride];
+ const int y_filter_idx = (y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
+ assert(y_filter_idx < SUBPEL_SHIFTS);
+ const int16_t *y_filter =
+ av1_get_interp_filter_subpel_kernel(*filter_params_y, y_filter_idx);
+ CONV_BUF_TYPE sum = 1 << offset_bits;
+ for (k = 0; k < filter_params_y->taps; ++k) {
+ sum += y_filter[k] * src_y[(k - fo_vert) * im_stride];
+ }
+ assert(0 <= sum && sum < (1 << (offset_bits + 2)));
+ CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1) -
+ ((1 << (offset_bits - conv_params->round_1)) +
+ (1 << (offset_bits - conv_params->round_1 - 1)));
+ dst[y * dst_stride + x] += res;
+ }
+ src_vert++;
+ }
+}
+#endif // CONFIG_COMPOUND_ROUND
void av1_highbd_convolve_2d_facade(const uint8_t *src8, int src_stride,
uint8_t *dst, int dst_stride, int w, int h,
const InterpFilter *interp_filter,
const int subpel_x_q4, int x_step_q4,
const int subpel_y_q4, int y_step_q4,
- ConvolveParams *conv_params, int bd) {
+ int scaled, ConvolveParams *conv_params,
+ int bd) {
(void)x_step_q4;
(void)y_step_q4;
(void)dst;
@@ -639,17 +893,89 @@
conv_params->dst_stride, w, h);
// horizontal and vertical parameters are swapped because of the transpose
- av1_highbd_convolve_2d(tr_src + fo_horiz * tr_src_stride + fo_vert,
- tr_src_stride, tr_dst, tr_dst_stride, h, w,
- &filter_params_y, &filter_params_x, subpel_y_q4,
- subpel_x_q4, conv_params, bd);
+ if (scaled)
+ av1_highbd_convolve_2d_scale(
+ tr_src + fo_horiz * tr_src_stride + fo_vert, tr_src_stride, tr_dst,
+ tr_dst_stride, h, w, &filter_params_y, &filter_params_x, subpel_y_q4,
+ y_step_q4, subpel_x_q4, x_step_q4, conv_params, bd);
+ else
+ av1_highbd_convolve_2d(tr_src + fo_horiz * tr_src_stride + fo_vert,
+ tr_src_stride, tr_dst, tr_dst_stride, h, w,
+ &filter_params_y, &filter_params_x, subpel_y_q4,
+ subpel_x_q4, conv_params, bd);
transpose_int32(conv_params->dst, conv_params->dst_stride, tr_dst,
tr_dst_stride, h, w);
} else {
- av1_highbd_convolve_2d(src, src_stride, conv_params->dst,
- conv_params->dst_stride, w, h, &filter_params_x,
- &filter_params_y, subpel_x_q4, subpel_y_q4,
- conv_params, bd);
+ if (scaled)
+ av1_highbd_convolve_2d_scale(
+ src, src_stride, conv_params->dst, conv_params->dst_stride, w, h,
+ &filter_params_x, &filter_params_y, subpel_x_q4, x_step_q4,
+ subpel_y_q4, y_step_q4, conv_params, bd);
+ else
+ av1_highbd_convolve_2d(src, src_stride, conv_params->dst,
+ conv_params->dst_stride, w, h, &filter_params_x,
+ &filter_params_y, subpel_x_q4, subpel_y_q4,
+ conv_params, bd);
+ }
+}
+
+void av1_highbd_convolve_2d_facade_scale(const uint8_t *src8, int src_stride,
+ uint8_t *dst, int dst_stride, int w,
+ int h,
+ const InterpFilter *interp_filter,
+ const int subpel_x_qn, int x_step_qn,
+ const int subpel_y_qn, int y_step_qn,
+ ConvolveParams *conv_params, int bd) {
+ (void)dst;
+ (void)dst_stride;
+#if CONFIG_DUAL_FILTER
+ InterpFilterParams filter_params_x =
+ av1_get_interp_filter_params(interp_filter[1 + 2 * conv_params->ref]);
+ InterpFilterParams filter_params_y =
+ av1_get_interp_filter_params(interp_filter[0 + 2 * conv_params->ref]);
+
+#if USE_EXTRA_FILTER
+ if (filter_params_x.interp_filter == MULTITAP_SHARP &&
+ filter_params_y.interp_filter == MULTITAP_SHARP) {
+ // Avoid two directions both using 12-tap filter.
+ // This will reduce hardware implementation cost.
+ filter_params_y = av1_get_interp_filter_params(EIGHTTAP_SHARP);
+ }
+#endif
+#else
+ InterpFilterParams filter_params_x =
+ av1_get_interp_filter_params(*interp_filter);
+ InterpFilterParams filter_params_y =
+ av1_get_interp_filter_params(*interp_filter);
+#endif
+ const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+ if (filter_params_y.taps < filter_params_x.taps) {
+ uint16_t tr_src[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) *
+ (MAX_SB_SIZE + MAX_FILTER_TAP - 1)];
+ int tr_src_stride = MAX_SB_SIZE + MAX_FILTER_TAP - 1;
+ CONV_BUF_TYPE tr_dst[MAX_SB_SIZE * MAX_SB_SIZE];
+ int tr_dst_stride = MAX_SB_SIZE;
+ int fo_vert = filter_params_y.taps / 2 - 1;
+ int fo_horiz = filter_params_x.taps / 2 - 1;
+
+ transpose_uint16(
+ tr_src, tr_src_stride, src - fo_vert * src_stride - fo_horiz,
+ src_stride, w + filter_params_x.taps - 1, h + filter_params_y.taps - 1);
+ transpose_int32(tr_dst, tr_dst_stride, conv_params->dst,
+ conv_params->dst_stride, w, h);
+
+ // horizontal and vertical parameters are swapped because of the transpose
+ av1_highbd_convolve_2d_scale(
+ tr_src + fo_horiz * tr_src_stride + fo_vert, tr_src_stride, tr_dst,
+ tr_dst_stride, h, w, &filter_params_y, &filter_params_x, subpel_y_qn,
+ y_step_qn, subpel_x_qn, x_step_qn, conv_params, bd);
+ transpose_int32(conv_params->dst, conv_params->dst_stride, tr_dst,
+ tr_dst_stride, h, w);
+ } else {
+ av1_highbd_convolve_2d_scale(
+ src, src_stride, conv_params->dst, conv_params->dst_stride, w, h,
+ &filter_params_x, &filter_params_y, subpel_x_qn, x_step_qn, subpel_y_qn,
+ y_step_qn, conv_params, bd);
}
}
#endif // CONFIG_HIGHBITDEPTH
@@ -812,7 +1138,7 @@
InterpFilterParams filter_params =
av1_get_interp_filter_params(interp_filter);
#endif
- // assert(conv_params->round == CONVOLVE_OPT_ROUND);
+ assert(conv_params->round == CONVOLVE_OPT_ROUND);
assert(w <= MAX_BLOCK_WIDTH);
assert(h <= MAX_BLOCK_HEIGHT);
diff --git a/av1/common/convolve.h b/av1/common/convolve.h
index 4627339..99ef1aed 100644
--- a/av1/common/convolve.h
+++ b/av1/common/convolve.h
@@ -54,7 +54,7 @@
int dst_stride, int w, int h,
const InterpFilter *interp_filter,
const int subpel_x_q4, int x_step_q4,
- const int subpel_y_q4, int y_step_q4,
+ const int subpel_y_q4, int y_step_q4, int scaled,
ConvolveParams *conv_params);
static INLINE ConvolveParams get_conv_params_no_round(int ref, int do_average,
@@ -83,7 +83,8 @@
const InterpFilter *interp_filter,
const int subpel_x_q4, int x_step_q4,
const int subpel_y_q4, int y_step_q4,
- ConvolveParams *conv_params, int bd);
+ int scaled, ConvolveParams *conv_params,
+ int bd);
#endif
#endif // CONFIG_CONVOLVE_ROUND
diff --git a/av1/common/reconinter.h b/av1/common/reconinter.h
index 7d6badb..b452c61 100644
--- a/av1/common/reconinter.h
+++ b/av1/common/reconinter.h
@@ -66,8 +66,24 @@
if (has_scale(xs, ys)) {
// TODO(afergs, debargha): Use a different scale convolve function
// that uses higher precision for subpel_x, subpel_y, xs, ys
- av1_convolve_scale(src, src_stride, dst, dst_stride, w, h, interp_filter,
- subpel_x, xs, subpel_y, ys, conv_params);
+ if (conv_params->round == CONVOLVE_OPT_NO_ROUND) {
+#if CONFIG_CONVOLVE_ROUND
+ av1_convolve_2d_facade(src, src_stride, dst, dst_stride, w, h,
+#if CONFIG_DUAL_FILTER
+ interp_filter,
+#else // CONFIG_DUAL_FILTER
+ &interp_filter,
+#endif // CONFIG_DUAL_FILTER
+ subpel_x, xs, subpel_y, ys, 1, conv_params);
+ conv_params->do_post_rounding = 1;
+#else
+ assert(0);
+#endif // CONFIG_CONVOLVE_ROUND
+ } else {
+ assert(conv_params->round == CONVOLVE_OPT_ROUND);
+ av1_convolve_scale(src, src_stride, dst, dst_stride, w, h, interp_filter,
+ subpel_x, xs, subpel_y, ys, conv_params);
+ }
} else {
subpel_x >>= SCALE_EXTRA_BITS;
subpel_y >>= SCALE_EXTRA_BITS;
@@ -85,7 +101,7 @@
#else // CONFIG_DUAL_FILTER
&interp_filter,
#endif // CONFIG_DUAL_FILTER
- subpel_x, xs, subpel_y, ys, conv_params);
+ subpel_x, xs, subpel_y, ys, 0, conv_params);
conv_params->do_post_rounding = 1;
#else
assert(0);
@@ -138,10 +154,25 @@
#endif
if (has_scale(xs, ys)) {
- av1_highbd_convolve_scale(
- src, src_stride, dst, dst_stride, w, h, interp_filter,
- subpel_x >> SCALE_EXTRA_BITS, xs >> SCALE_EXTRA_BITS,
- subpel_y >> SCALE_EXTRA_BITS, ys >> SCALE_EXTRA_BITS, avg, bd);
+ if (conv_params->round == CONVOLVE_OPT_NO_ROUND) {
+#if CONFIG_CONVOLVE_ROUND
+ av1_highbd_convolve_2d_facade(src, src_stride, dst, dst_stride, w, h,
+#if CONFIG_DUAL_FILTER
+ interp_filter,
+#else // CONFIG_DUAL_FILTER
+ &interp_filter,
+#endif // CONFIG_DUAL_FILTER
+ subpel_x, xs, subpel_y, ys, 1, conv_params,
+ bd);
+ conv_params->do_post_rounding = 1;
+#else
+ assert(0);
+#endif // CONFIG_CONVOLVE_ROUND
+ } else {
+ av1_highbd_convolve_scale(src, src_stride, dst, dst_stride, w, h,
+ interp_filter, subpel_x, xs, subpel_y, ys, avg,
+ bd);
+ }
} else {
subpel_x >>= SCALE_EXTRA_BITS;
subpel_y >>= SCALE_EXTRA_BITS;
@@ -159,7 +190,7 @@
#else // CONFIG_DUAL_FILTER
&interp_filter,
#endif // CONFIG_DUAL_FILTER
- subpel_x, xs, subpel_y, ys, conv_params,
+ subpel_x, xs, subpel_y, ys, 0, conv_params,
bd);
conv_params->do_post_rounding = 1;
#else
