Skip to content

GitLab

Commit 81f6ecd1 authored by Sarah Parker's avatar Sarah Parker
Browse files

Early termination for warp error computation 

This terminates the computation for the warp error once
the frame error exceeds the best frame error found
so far to avoid unneccessary computation.

Change-Id: I094a0b3e13f8b91610e051cb91d20a815879dd80
parent 28c779be
Hide whitespace changes
Inline Side-by-side
Showing with 80 additions and 55 deletions
av1/common/warped_motion.c
View file @ 81f6ecd1
......@@ -18,6 +18,8 @@
#include "./av1_rtcd.h"
#include "av1/common/warped_motion.h"
#define WARP_ERROR_BLOCK 32
/* clang-format off */
static const int error_measure_lut[512] = {
// pow 0.7
......@@ -1080,14 +1082,13 @@ static void highbd_warp_plane(WarpedMotionParams *wm, const uint8_t *const ref8,
}
static int64_t highbd_frame_error(const uint16_t *const ref, int stride,
const uint16_t *const dst, int p_col,
int p_row, int p_width, int p_height,
int p_stride, int bd) {
const uint16_t *const dst, int p_width,
int p_height, int p_stride, int bd) {
int64_t sum_error = 0;
for (int i = 0; i < p_height; ++i) {
for (int j = 0; j < p_width; ++j) {
sum_error += highbd_error_measure(
dst[j + i * p_stride] - ref[(j + p_col) + (i + p_row) * stride], bd);
sum_error +=
highbd_error_measure(dst[j + i * p_stride] - ref[j + i * stride], bd);
}
}
return sum_error;
......@@ -1097,19 +1098,30 @@ static int64_t highbd_warp_error(
WarpedMotionParams *wm, const uint8_t *const ref8, int width, int height,
int stride, const uint8_t *const dst8, int p_col, int p_row, int p_width,
int p_height, int p_stride, int subsampling_x, int subsampling_y,
int x_scale, int y_scale, int bd) {
int x_scale, int y_scale, int bd, int64_t best_error) {
int64_t gm_sumerr = 0;
uint16_t *tmp = aom_malloc(p_width * p_height * sizeof(*tmp));
if (!tmp) return INT64_MAX;
highbd_warp_plane(wm, ref8, width, height, stride, CONVERT_TO_BYTEPTR(tmp),
p_col, p_row, p_width, p_height, p_width, subsampling_x,
subsampling_y, x_scale, y_scale, bd, 0);
gm_sumerr = highbd_frame_error(tmp, p_width, CONVERT_TO_SHORTPTR(dst8), p_col,
p_row, p_width, p_height, p_stride, bd);
aom_free(tmp);
int warp_w, warp_h;
int error_bsize_w = AOMMIN(p_width, WARP_ERROR_BLOCK);
int error_bsize_h = AOMMIN(p_height, WARP_ERROR_BLOCK);
uint16_t tmp[WARP_ERROR_BLOCK * WARP_ERROR_BLOCK];
for (int i = p_row; i < p_row + p_height; i += WARP_ERROR_BLOCK) {
for (int j = p_col; j < p_col + p_width; j += WARP_ERROR_BLOCK) {
// avoid warping extra 8x8 blocks in the padded region of the frame
// when p_width and p_height are not multiples of WARP_ERROR_BLOCK
warp_w = AOMMIN(error_bsize_w, p_col + p_width - j);
warp_h = AOMMIN(error_bsize_h, p_row + p_height - i);
highbd_warp_plane(wm, ref8, width, height, stride,
CONVERT_TO_BYTEPTR(tmp), j, i, warp_w, warp_h,
WARP_ERROR_BLOCK, subsampling_x, subsampling_y, x_scale,
y_scale, bd, 0);
gm_sumerr += highbd_frame_error(
tmp, WARP_ERROR_BLOCK, CONVERT_TO_SHORTPTR(dst8) + j + i * p_stride,
warp_w, warp_h, p_stride, bd);
if (gm_sumerr > best_error) return gm_sumerr;
}
}
return gm_sumerr;
}
#endif // CONFIG_HIGHBITDEPTH
......@@ -1368,13 +1380,13 @@ static void warp_plane(WarpedMotionParams *wm, const uint8_t *const ref,
}
static int64_t frame_error(const uint8_t *const ref, int stride,
const uint8_t *const dst, int p_col, int p_row,
int p_width, int p_height, int p_stride) {
const uint8_t *const dst, int p_width, int p_height,
int p_stride) {
int64_t sum_error = 0;
for (int i = 0; i < p_height; ++i) {
for (int j = 0; j < p_width; ++j) {
sum_error += (int64_t)error_measure(
dst[j + i * p_stride] - ref[(j + p_col) + (i + p_row) * stride]);
sum_error +=
(int64_t)error_measure(dst[j + i * p_stride] - ref[j + i * stride]);
}
}
return sum_error;
......@@ -1385,19 +1397,28 @@ static int64_t warp_error(WarpedMotionParams *wm, const uint8_t *const ref,
const uint8_t *const dst, int p_col, int p_row,
int p_width, int p_height, int p_stride,
int subsampling_x, int subsampling_y, int x_scale,
int y_scale) {
int y_scale, int64_t best_error) {
int64_t gm_sumerr = 0;
uint8_t *tmp = aom_malloc(p_width * p_height);
if (!tmp) return INT64_MAX;
warp_plane(wm, ref, width, height, stride, tmp, p_col, p_row, p_width,
p_height, p_width, subsampling_x, subsampling_y, x_scale, y_scale,
0);
gm_sumerr =
frame_error(tmp, p_width, dst, p_col, p_row, p_width, p_height, p_stride);
aom_free(tmp);
int warp_w, warp_h;
int error_bsize_w = AOMMIN(p_width, WARP_ERROR_BLOCK);
int error_bsize_h = AOMMIN(p_height, WARP_ERROR_BLOCK);
uint8_t tmp[WARP_ERROR_BLOCK * WARP_ERROR_BLOCK];
for (int i = p_row; i < p_row + p_height; i += WARP_ERROR_BLOCK) {
for (int j = p_col; j < p_col + p_width; j += WARP_ERROR_BLOCK) {
// avoid warping extra 8x8 blocks in the padded region of the frame
// when p_width and p_height are not multiples of WARP_ERROR_BLOCK
warp_w = AOMMIN(error_bsize_w, p_col + p_width - j);
warp_h = AOMMIN(error_bsize_h, p_row + p_height - i);
warp_plane(wm, ref, width, height, stride, tmp, j, i, warp_w, warp_h,
WARP_ERROR_BLOCK, subsampling_x, subsampling_y, x_scale,
y_scale, 0);
gm_sumerr += frame_error(tmp, WARP_ERROR_BLOCK, dst + j + i * p_stride,
warp_w, warp_h, p_stride);
if (gm_sumerr > best_error) return gm_sumerr;
}
}
return gm_sumerr;
}
......@@ -1405,17 +1426,16 @@ int64_t av1_frame_error(
#if CONFIG_HIGHBITDEPTH
int use_hbd, int bd,
#endif // CONFIG_HIGHBITDEPTH
const uint8_t *ref, int stride, uint8_t *dst, int p_col, int p_row,
int p_width, int p_height, int p_stride) {
const uint8_t *ref, int stride, uint8_t *dst, int p_width, int p_height,
int p_stride) {
#if CONFIG_HIGHBITDEPTH
if (use_hbd) {
return highbd_frame_error(CONVERT_TO_SHORTPTR(ref), stride,
CONVERT_TO_SHORTPTR(dst), p_col, p_row, p_width,
p_height, p_stride, bd);
CONVERT_TO_SHORTPTR(dst), p_width, p_height,
p_stride, bd);
}
#endif // CONFIG_HIGHBITDEPTH
return frame_error(ref, stride, dst, p_col, p_row, p_width, p_height,
p_stride);
return frame_error(ref, stride, dst, p_width, p_height, p_stride);
}
int64_t av1_warp_error(WarpedMotionParams *wm,
......@@ -1425,18 +1445,19 @@ int64_t av1_warp_error(WarpedMotionParams *wm,
const uint8_t *ref, int width, int height, int stride,
uint8_t *dst, int p_col, int p_row, int p_width,
int p_height, int p_stride, int subsampling_x,
int subsampling_y, int x_scale, int y_scale) {
int subsampling_y, int x_scale, int y_scale,
int64_t best_error) {
if (wm->wmtype <= AFFINE)
if (!get_shear_params(wm)) return 1;
#if CONFIG_HIGHBITDEPTH
if (use_hbd)
return highbd_warp_error(wm, ref, width, height, stride, dst, p_col, p_row,
p_width, p_height, p_stride, subsampling_x,
subsampling_y, x_scale, y_scale, bd);
subsampling_y, x_scale, y_scale, bd, best_error);
#endif // CONFIG_HIGHBITDEPTH
return warp_error(wm, ref, width, height, stride, dst, p_col, p_row, p_width,
p_height, p_stride, subsampling_x, subsampling_y, x_scale,
y_scale);
y_scale, best_error);
}
void av1_warp_plane(WarpedMotionParams *wm,
......
av1/common/warped_motion.h
View file @ 81f6ecd1
......@@ -82,7 +82,8 @@ int64_t av1_warp_error(WarpedMotionParams *wm,
const uint8_t *ref, int width, int height, int stride,
uint8_t *dst, int p_col, int p_row, int p_width,
int p_height, int p_stride, int subsampling_x,
int subsampling_y, int x_scale, int y_scale);
int subsampling_y, int x_scale, int y_scale,
int64_t best_error);
// Returns the error between the frame described by 'ref' and the frame
// described by 'dst'.
......@@ -90,8 +91,8 @@ int64_t av1_frame_error(
#if CONFIG_HIGHBITDEPTH
int use_hbd, int bd,
#endif // CONFIG_HIGHBITDEPTH
const uint8_t *ref, int stride, uint8_t *dst, int p_col, int p_row,
int p_width, int p_height, int p_stride);
const uint8_t *ref, int stride, uint8_t *dst, int p_width, int p_height,
int p_stride);
void av1_warp_plane(WarpedMotionParams *wm,
#if CONFIG_HIGHBITDEPTH
......
av1/encoder/encodeframe.c
View file @ 81f6ecd1
......@@ -4716,8 +4716,8 @@ static void encode_frame_internal(AV1_COMP *cpi) {
xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH, xd->bd,
#endif // CONFIG_HIGHBITDEPTH
ref_buf[frame]->y_buffer, ref_buf[frame]->y_stride,
cpi->source->y_buffer, 0, 0, cpi->source->y_width,
cpi->source->y_height, cpi->source->y_stride);
cpi->source->y_buffer, cpi->source->y_width, cpi->source->y_height,
cpi->source->y_stride);
if (ref_frame_error == 0) continue;
......@@ -4752,7 +4752,8 @@ static void encode_frame_internal(AV1_COMP *cpi) {
ref_buf[frame]->y_buffer, ref_buf[frame]->y_width,
ref_buf[frame]->y_height, ref_buf[frame]->y_stride,
cpi->source->y_buffer, cpi->source->y_width,
cpi->source->y_height, cpi->source->y_stride, 3);
cpi->source->y_height, cpi->source->y_stride, 3,
best_warp_error);
if (warp_error < best_warp_error) {
best_warp_error = warp_error;
// Save the wm_params modified by refine_integerized_param()
......
av1/encoder/global_motion.c
View file @ 81f6ecd1
......@@ -131,8 +131,8 @@ int64_t refine_integerized_param(WarpedMotionParams *wm,
#endif // CONFIG_HIGHBITDEPTH
uint8_t *ref, int r_width, int r_height,
int r_stride, uint8_t *dst, int d_width,
int d_height, int d_stride,
int n_refinements) {
int d_height, int d_stride, int n_refinements,
int64_t best_frame_error) {
static const int max_trans_model_params[TRANS_TYPES] = {
0, 2, 4, 6, 8, 8, 8
};
......@@ -154,7 +154,8 @@ int64_t refine_integerized_param(WarpedMotionParams *wm,
ref, r_width, r_height, r_stride,
dst + border * d_stride + border, border, border,
d_width - 2 * border, d_height - 2 * border,
d_stride, 0, 0, 16, 16);
d_stride, 0, 0, 16, 16, best_frame_error);
best_error = AOMMIN(best_error, best_frame_error);
step = 1 << (n_refinements + 1);
for (i = 0; i < n_refinements; i++, step >>= 1) {
for (p = 0; p < n_params; ++p) {
......@@ -174,7 +175,7 @@ int64_t refine_integerized_param(WarpedMotionParams *wm,
#endif // CONFIG_HIGHBITDEPTH
ref, r_width, r_height, r_stride, dst + border * d_stride + border,
border, border, d_width - 2 * border, d_height - 2 * border, d_stride,
0, 0, 16, 16);
0, 0, 16, 16, best_error);
if (step_error < best_error) {
best_error = step_error;
best_param = *param;
......@@ -190,7 +191,7 @@ int64_t refine_integerized_param(WarpedMotionParams *wm,
#endif // CONFIG_HIGHBITDEPTH
ref, r_width, r_height, r_stride, dst + border * d_stride + border,
border, border, d_width - 2 * border, d_height - 2 * border, d_stride,
0, 0, 16, 16);
0, 0, 16, 16, best_error);
if (step_error < best_error) {
best_error = step_error;
best_param = *param;
......@@ -209,7 +210,7 @@ int64_t refine_integerized_param(WarpedMotionParams *wm,
#endif // CONFIG_HIGHBITDEPTH
ref, r_width, r_height, r_stride, dst + border * d_stride + border,
border, border, d_width - 2 * border, d_height - 2 * border,
d_stride, 0, 0, 16, 16);
d_stride, 0, 0, 16, 16, best_error);
if (step_error < best_error) {
best_error = step_error;
best_param = *param;
......
av1/encoder/global_motion.h
View file @ 81f6ecd1
......@@ -36,7 +36,8 @@ int64_t refine_integerized_param(WarpedMotionParams *wm,
#endif // CONFIG_HIGHBITDEPTH
uint8_t *ref, int r_width, int r_height,
int r_stride, uint8_t *dst, int d_width,
int d_height, int d_stride, int n_refinements);
int d_height, int d_stride, int n_refinements,
int64_t best_frame_error);
/*
Computes "num_motions" candidate global motion parameters between two frames.
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment