generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Add high-precision sub-pixel search as a speed feature

Browse Source 
Using the up-sampled reference frames in sub-pixel motion search is
enabled as a speed feature for good-quality mode speed 0 and speed 1.

Change-Id: Ieb454bf8c646ddb99e87bd64c8e74dbd78d84a50
This commit is contained in:
Yunqing Wang
2016-03-10 11:07:50 -08:00
parent e38e2ad86e
commit e6e2d886d3
14 changed files with 245 additions and 290 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
configure vendored
View File

@@ -284,7 +284,6 @@ EXPERIMENT_LIST="
ext_partition
ext_tile
obmc
affine_motion
"
CONFIG_LIST="
dependency_tracking

198
vp10/encoder/encoder.c
View File

@@ -358,9 +358,7 @@ void vp10_initialize_enc(void) {
static void dealloc_compressor_data(VP10_COMP *cpi) {
VP10_COMMON *const cm = &cpi->common;
#if CONFIG_REF_MV
int i;
#endif
vpx_free(cpi->mbmi_ext_base);
cpi->mbmi_ext_base = NULL;
@@ -413,14 +411,9 @@ static void dealloc_compressor_data(VP10_COMP *cpi) {
vpx_free(cpi->active_map.map);
cpi->active_map.map = NULL;
#if CONFIG_AFFINE_MOTION
{
// Free up-sampled reference buffers.
int i;
for (i = 0; i < MAX_REF_FRAMES; i++)
vpx_free_frame_buffer(&cpi->upsampled_ref_bufs[i].buf);
}
#endif
// Free up-sampled reference buffers.
for (i = 0; i < MAX_REF_FRAMES; i++)
vpx_free_frame_buffer(&cpi->upsampled_ref_bufs[i].buf);
vp10_free_ref_frame_buffers(cm->buffer_pool);
#if CONFIG_VP9_POSTPROC
@@ -756,26 +749,6 @@ static void alloc_util_frame_buffers(VP10_COMP *cpi) {
NULL, NULL, NULL))
vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
"Failed to allocate scaled last source buffer");
#if CONFIG_AFFINE_MOTION
{
// Allocate up-sampled reference buffers.
int i;
for (i = 0; i < MAX_REF_FRAMES; i++)
if (vpx_realloc_frame_buffer(&cpi->upsampled_ref_bufs[i].buf,
(cm->width << 3), (cm->height << 3),
cm->subsampling_x, cm->subsampling_y,
#if CONFIG_VP9_HIGHBITDEPTH
cm->use_highbitdepth,
#endif
(VP9_ENC_BORDER_IN_PIXELS << 3),
cm->byte_alignment,
NULL, NULL, NULL))
vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
"Failed to allocate up-sampled reference frame buffer");
}
#endif
}
@@ -2069,6 +2042,14 @@ static void cal_nmvsadcosts_hp(int *mvsadcost[2]) {
} while (++i <= MV_MAX);
}
static INLINE void init_upsampled_ref_frame_bufs(VP10_COMP *cpi) {
int i;
for (i = 0; i < MAX_REF_FRAMES; ++i) {
cpi->upsampled_ref_bufs[i].ref_count = 0;
cpi->upsampled_ref_idx[i] = INVALID_IDX;
}
}
VP10_COMP *vp10_create_compressor(VP10EncoderConfig *oxcf,
BufferPool *const pool) {
@@ -2267,6 +2248,8 @@ VP10_COMP *vp10_create_compressor(VP10EncoderConfig *oxcf,
vp10_init_second_pass(cpi);
}
init_upsampled_ref_frame_bufs(cpi);
vp10_set_speed_features_framesize_independent(cpi);
vp10_set_speed_features_framesize_dependent(cpi);
@@ -2929,7 +2912,6 @@ static int recode_loop_test(VP10_COMP *cpi,
return force_recode;
}
#if CONFIG_AFFINE_MOTION
static INLINE int get_free_upsampled_ref_buf(EncRefCntBuffer *ubufs) {
int i;
@@ -2941,50 +2923,59 @@ static INLINE int get_free_upsampled_ref_buf(EncRefCntBuffer *ubufs) {
return INVALID_IDX;
}
// Up-sample reference frames.
static INLINE int upsample_ref_frame(RefCntBuffer *bufs,
#if CONFIG_VP9_HIGHBITDEPTH
EncRefCntBuffer *ubufs, int new_idx,
int bit_depth) {
#else
EncRefCntBuffer *ubufs, int new_idx) {
#endif
// Up-sample 1 reference frame.
static INLINE int upsample_ref_frame(VP10_COMP *cpi,
const YV12_BUFFER_CONFIG *const ref) {
VP10_COMMON * const cm = &cpi->common;
EncRefCntBuffer *ubufs = cpi->upsampled_ref_bufs;
int new_uidx = get_free_upsampled_ref_buf(ubufs);
if (new_uidx == INVALID_IDX) {
return INVALID_IDX;
} else {
const YV12_BUFFER_CONFIG *const ref = &bufs[new_idx].buf;
YV12_BUFFER_CONFIG *upsampled_ref = &ubufs[new_uidx].buf;
// Can allocate buffer for Y plane only.
if (upsampled_ref->buffer_alloc_sz < (ref->buffer_alloc_sz << 6))
if (vpx_realloc_frame_buffer(upsampled_ref,
(cm->width << 3), (cm->height << 3),
cm->subsampling_x, cm->subsampling_y,
#if CONFIG_VP9_HIGHBITDEPTH
cm->use_highbitdepth,
#endif
(VP9_ENC_BORDER_IN_PIXELS << 3),
cm->byte_alignment,
NULL, NULL, NULL))
vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
"Failed to allocate up-sampled frame buffer");
// Currently, only Y plane is up-sampled, U, V are not used.
#if CONFIG_VP9_HIGHBITDEPTH
scale_and_extend_frame(ref, upsampled_ref, 1, bit_depth);
scale_and_extend_frame(ref, upsampled_ref, 1, (int)cm->bit_depth);
#else
scale_and_extend_frame(ref, upsampled_ref, 1);
#endif
return new_uidx;
}
}
#endif
void vp10_update_reference_frames(VP10_COMP *cpi) {
VP10_COMMON * const cm = &cpi->common;
BufferPool *const pool = cm->buffer_pool;
const int use_upsampled_ref = cpi->sf.use_upsampled_references;
int new_uidx = 0;
#if CONFIG_EXT_REFS
int ref_frame;
#endif // CONFIG_EXT_REFS
#if CONFIG_AFFINE_MOTION
// Always up-sample the current encoded frame.
#if CONFIG_VP9_HIGHBITDEPTH
int new_uidx = upsample_ref_frame(pool->frame_bufs, cpi->upsampled_ref_bufs,
cm->new_fb_idx, (int)cm->bit_depth);
#else
int new_uidx = upsample_ref_frame(pool->frame_bufs, cpi->upsampled_ref_bufs,
cm->new_fb_idx);
#endif
#endif
if (use_upsampled_ref) {
// Up-sample the current encoded frame.
RefCntBuffer *bufs = pool->frame_bufs;
const YV12_BUFFER_CONFIG *const ref = &bufs[cm->new_fb_idx].buf;
new_uidx = upsample_ref_frame(cpi, ref);
}
// At this point the new frame has been encoded.
// If any buffer copy / swapping is signaled it should be done here.
@@ -2994,12 +2985,12 @@ void vp10_update_reference_frames(VP10_COMP *cpi) {
ref_cnt_fb(pool->frame_bufs,
&cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx);
#if CONFIG_AFFINE_MOTION
uref_cnt_fb(cpi->upsampled_ref_bufs,
&cpi->upsampled_ref_idx[cpi->gld_fb_idx], new_uidx);
uref_cnt_fb(cpi->upsampled_ref_bufs,
&cpi->upsampled_ref_idx[cpi->alt_fb_idx], new_uidx);
#endif
if (use_upsampled_ref) {
uref_cnt_fb(cpi->upsampled_ref_bufs,
&cpi->upsampled_ref_idx[cpi->gld_fb_idx], new_uidx);
uref_cnt_fb(cpi->upsampled_ref_bufs,
&cpi->upsampled_ref_idx[cpi->alt_fb_idx], new_uidx);
}
} else if (vp10_preserve_existing_gf(cpi)) {
// We have decided to preserve the previously existing golden frame as our
// new ARF frame. However, in the short term in function
@@ -3013,10 +3004,10 @@ void vp10_update_reference_frames(VP10_COMP *cpi) {
ref_cnt_fb(pool->frame_bufs,
&cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx);
#if CONFIG_AFFINE_MOTION
uref_cnt_fb(cpi->upsampled_ref_bufs,
&cpi->upsampled_ref_idx[cpi->alt_fb_idx], new_uidx);
#endif
if (use_upsampled_ref)
uref_cnt_fb(cpi->upsampled_ref_bufs,
&cpi->upsampled_ref_idx[cpi->alt_fb_idx], new_uidx);
tmp = cpi->alt_fb_idx;
cpi->alt_fb_idx = cpi->gld_fb_idx;
cpi->gld_fb_idx = tmp;
@@ -3030,10 +3021,10 @@ void vp10_update_reference_frames(VP10_COMP *cpi) {
ref_cnt_fb(pool->frame_bufs,
&cm->ref_frame_map[arf_idx], cm->new_fb_idx);
#if CONFIG_AFFINE_MOTION
uref_cnt_fb(cpi->upsampled_ref_bufs,
&cpi->upsampled_ref_idx[cpi->alt_fb_idx], new_uidx);
#endif
if (use_upsampled_ref)
uref_cnt_fb(cpi->upsampled_ref_bufs,
&cpi->upsampled_ref_idx[cpi->alt_fb_idx], new_uidx);
memcpy(cpi->interp_filter_selected[ALTREF_FRAME],
cpi->interp_filter_selected[0],
sizeof(cpi->interp_filter_selected[0]));
@@ -3042,10 +3033,10 @@ void vp10_update_reference_frames(VP10_COMP *cpi) {
if (cpi->refresh_golden_frame) {
ref_cnt_fb(pool->frame_bufs,
&cm->ref_frame_map[cpi->gld_fb_idx], cm->new_fb_idx);
#if CONFIG_AFFINE_MOTION
uref_cnt_fb(cpi->upsampled_ref_bufs,
&cpi->upsampled_ref_idx[cpi->gld_fb_idx], new_uidx);
#endif
if (use_upsampled_ref)
uref_cnt_fb(cpi->upsampled_ref_bufs,
&cpi->upsampled_ref_idx[cpi->gld_fb_idx], new_uidx);
if (!cpi->rc.is_src_frame_alt_ref)
memcpy(cpi->interp_filter_selected[GOLDEN_FRAME],
cpi->interp_filter_selected[0],
@@ -3080,10 +3071,10 @@ void vp10_update_reference_frames(VP10_COMP *cpi) {
if (cpi->refresh_last_frame) {
ref_cnt_fb(pool->frame_bufs,
&cm->ref_frame_map[cpi->lst_fb_idx], cm->new_fb_idx);
#if CONFIG_AFFINE_MOTION
uref_cnt_fb(cpi->upsampled_ref_bufs,
&cpi->upsampled_ref_idx[cpi->lst_fb_idx], new_uidx);
#endif
if (use_upsampled_ref)
uref_cnt_fb(cpi->upsampled_ref_bufs,
&cpi->upsampled_ref_idx[cpi->lst_fb_idx], new_uidx);
if (!cpi->rc.is_src_frame_alt_ref) {
memcpy(cpi->interp_filter_selected[LAST_FRAME],
cpi->interp_filter_selected[0],
@@ -3249,8 +3240,9 @@ void vp10_scale_references(VP10_COMP *cpi) {
}
#endif // CONFIG_VP9_HIGHBITDEPTH
#if CONFIG_AFFINE_MOTION
{
if (cpi->sf.use_upsampled_references && (force_scaling ||
new_fb_ptr->buf.y_crop_width != cm->width ||
new_fb_ptr->buf.y_crop_height != cm->height)) {
const int map_idx = get_ref_frame_map_idx(cpi, ref_frame);
EncRefCntBuffer *ubuf =
&cpi->upsampled_ref_bufs[cpi->upsampled_ref_idx[map_idx]];
@@ -3267,15 +3259,12 @@ void vp10_scale_references(VP10_COMP *cpi) {
vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
"Failed to allocate up-sampled frame buffer");
#if CONFIG_VP9_HIGHBITDEPTH
scale_and_extend_frame(&new_fb_ptr->buf, &ubuf->buf, MAX_MB_PLANE,
scale_and_extend_frame(&new_fb_ptr->buf, &ubuf->buf, 1,
(int)cm->bit_depth);
#else
scale_and_extend_frame(&new_fb_ptr->buf, &ubuf->buf, MAX_MB_PLANE);
scale_and_extend_frame(&new_fb_ptr->buf, &ubuf->buf, 1);
#endif
cpi->scaled_ref_idx[ref_frame - LAST_FRAME] = new_fb;
alloc_frame_mvs(cm, new_fb);
}
#endif
} else {
const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
RefCntBuffer *const buf = &pool->frame_bufs[buf_idx];
@@ -3610,9 +3599,28 @@ static void set_frame_size(VP10_COMP *cpi) {
set_ref_ptrs(cm, xd, LAST_FRAME, LAST_FRAME);
}
static void reset_use_upsampled_references(VP10_COMP *cpi) {
MV_REFERENCE_FRAME ref_frame;
// reset up-sampled reference buffer structure.
init_upsampled_ref_frame_bufs(cpi);
for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
const YV12_BUFFER_CONFIG *const ref = get_ref_frame_buffer(cpi,
ref_frame);
int new_uidx = upsample_ref_frame(cpi, ref);
// Update the up-sampled reference index.
cpi->upsampled_ref_idx[get_ref_frame_map_idx(cpi, ref_frame)] =
new_uidx;
cpi->upsampled_ref_bufs[new_uidx].ref_count++;
}
}
static void encode_without_recode_loop(VP10_COMP *cpi) {
VP10_COMMON *const cm = &cpi->common;
int q = 0, bottom_index = 0, top_index = 0; // Dummy variables.
const int use_upsampled_ref = cpi->sf.use_upsampled_references;
vpx_clear_system_state();
@@ -3647,6 +3655,12 @@ static void encode_without_recode_loop(VP10_COMP *cpi) {
set_size_independent_vars(cpi);
set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
// cpi->sf.use_upsampled_references can be different from frame to frame.
// Every time when cpi->sf.use_upsampled_references is changed from 0 to 1.
// The reference frames for this frame have to be up-sampled before encoding.
if (!use_upsampled_ref && cpi->sf.use_upsampled_references)
reset_use_upsampled_references(cpi);
vp10_set_quantizer(cm, q);
vp10_set_variance_partition_thresholds(cpi, q);
@@ -3694,9 +3708,16 @@ static void encode_with_recode_loop(VP10_COMP *cpi,
int frame_over_shoot_limit;
int frame_under_shoot_limit;
int q = 0, q_low = 0, q_high = 0;
const int use_upsampled_ref = cpi->sf.use_upsampled_references;
set_size_independent_vars(cpi);
// cpi->sf.use_upsampled_references can be different from frame to frame.
// Every time when cpi->sf.use_upsampled_references is changed from 0 to 1.
// The reference frames for this frame have to be up-sampled before encoding.
if (!use_upsampled_ref && cpi->sf.use_upsampled_references)
reset_use_upsampled_references(cpi);
do {
vpx_clear_system_state();
@@ -4355,17 +4376,6 @@ static void init_ref_frame_bufs(VP10_COMMON *cm) {
}
}
#if CONFIG_AFFINE_MOTION
static INLINE void init_upsampled_ref_frame_bufs(VP10_COMP *cpi) {
int i;
for (i = 0; i < MAX_REF_FRAMES; ++i) {
cpi->upsampled_ref_bufs[i].ref_count = 0;
cpi->upsampled_ref_idx[i] = INVALID_IDX;
}
}
#endif
static void check_initial_width(VP10_COMP *cpi,
#if CONFIG_VP9_HIGHBITDEPTH
int use_highbitdepth,
@@ -4388,9 +4398,7 @@ static void check_initial_width(VP10_COMP *cpi,
alloc_raw_frame_buffers(cpi);
init_ref_frame_bufs(cm);
alloc_util_frame_buffers(cpi);
#if CONFIG_AFFINE_MOTION
init_upsampled_ref_frame_bufs(cpi);
#endif
init_motion_estimation(cpi); // TODO(agrange) This can be removed.
cpi->initial_width = cm->width;

6
vp10/encoder/encoder.h
View File

@@ -286,12 +286,10 @@ typedef struct IMAGE_STAT {
double worst;
} ImageStat;
#if CONFIG_AFFINE_MOTION
typedef struct {
int ref_count;
YV12_BUFFER_CONFIG buf;
} EncRefCntBuffer;
#endif
typedef struct VP10_COMP {
QUANTS quants;
@@ -311,11 +309,9 @@ typedef struct VP10_COMP {
YV12_BUFFER_CONFIG *unscaled_last_source;
YV12_BUFFER_CONFIG scaled_last_source;
#if CONFIG_AFFINE_MOTION
// Up-sampled reference buffers
EncRefCntBuffer upsampled_ref_bufs[MAX_REF_FRAMES];
int upsampled_ref_idx[MAX_REF_FRAMES];
#endif
TileDataEnc *tile_data;
int allocated_tiles; // Keep track of memory allocated for tiles.
@@ -702,7 +698,6 @@ void vp10_new_framerate(VP10_COMP *cpi, double framerate);
#define LAYER_IDS_TO_IDX(sl, tl, num_tl) ((sl) * (num_tl) + (tl))
#if CONFIG_AFFINE_MOTION
// Update up-sampled reference frame index.
static INLINE void uref_cnt_fb(EncRefCntBuffer *ubufs, int *uidx,
int new_uidx) {
@@ -714,7 +709,6 @@ static INLINE void uref_cnt_fb(EncRefCntBuffer *ubufs, int *uidx,
*uidx = new_uidx;
ubufs[new_uidx].ref_count++;
}
#endif
#ifdef __cplusplus
} // extern "C"

4
vp10/encoder/mbgraph.c
View File

@@ -64,11 +64,7 @@ static unsigned int do_16x16_motion_iteration(VP10_COMP *cpi,
&v_fn_ptr, 0, mv_sf->subpel_iters_per_step,
cond_cost_list(cpi, cost_list),
NULL, NULL,
#if CONFIG_AFFINE_MOTION
&distortion, &sse, NULL, 0, 0, 0);
#else
&distortion, &sse, NULL, 0, 0);
#endif
}
#if CONFIG_EXT_INTER

42
vp10/encoder/mcomp.c
View File

@@ -210,7 +210,6 @@ static INLINE const uint8_t *pre(const uint8_t *buf, int stride, int r, int c) {
#define CHECK_BETTER0(v, r, c) CHECK_BETTER(v, r, c)
#if CONFIG_AFFINE_MOTION
static INLINE const uint8_t *upre(const uint8_t *buf, int stride,
int r, int c) {
return &buf[(r) * stride + (c)];
@@ -232,7 +231,6 @@ static INLINE const uint8_t *upre(const uint8_t *buf, int stride,
} else { \
v = INT_MAX; \
}
#endif
#define FIRST_LEVEL_CHECKS \
{ \
@@ -438,11 +436,7 @@ int vp10_find_best_sub_pixel_tree_pruned_evenmore(
int *distortion,
unsigned int *sse1,
const uint8_t *second_pred,
#if CONFIG_AFFINE_MOTION
int w, int h, int use_upsampled_ref) {
#else
int w, int h) {
#endif
SETUP_SUBPEL_SEARCH;
besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp,
z, src_stride, y, y_stride, second_pred,
@@ -455,9 +449,7 @@ int vp10_find_best_sub_pixel_tree_pruned_evenmore(
(void) allow_hp;
(void) forced_stop;
(void) hstep;
#if CONFIG_AFFINE_MOTION
(void) use_upsampled_ref;
#endif
if (cost_list &&
cost_list[0] != INT_MAX && cost_list[1] != INT_MAX &&
@@ -524,16 +516,10 @@ int vp10_find_best_sub_pixel_tree_pruned_more(const MACROBLOCK *x,
int *distortion,
unsigned int *sse1,
const uint8_t *second_pred,
#if CONFIG_AFFINE_MOTION
int w, int h,
int use_upsampled_ref) {
#else
int w, int h) {
#endif
SETUP_SUBPEL_SEARCH;
#if CONFIG_AFFINE_MOTION
(void) use_upsampled_ref;
#endif
besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp,
z, src_stride, y, y_stride, second_pred,
@@ -607,15 +593,9 @@ int vp10_find_best_sub_pixel_tree_pruned(const MACROBLOCK *x,
int *distortion,
unsigned int *sse1,
const uint8_t *second_pred,
#if CONFIG_AFFINE_MOTION
int w, int h, int use_upsampled_ref) {
#else
int w, int h) {
#endif
SETUP_SUBPEL_SEARCH;
#if CONFIG_AFFINE_MOTION
(void) use_upsampled_ref;
#endif
besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp,
z, src_stride, y, y_stride, second_pred,
@@ -705,9 +685,8 @@ static const MV search_step_table[12] = {
{0, -1}, {0, 1}, {-1, 0}, {1, 0}
};
#if CONFIG_AFFINE_MOTION
#if CONFIG_VP9_HIGHBITDEPTH
// TODO(yunqing): Optimize the following 2 functions.
static void highbd_comp_avg_upsampled_pred(uint16_t *comp_pred,
const uint8_t *pred8,
int width, int height,
@@ -798,7 +777,6 @@ static unsigned int upsampled_setup_center_error(
besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
return besterr;
}
#endif
int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
MV *bestmv, const MV *ref_mv,
@@ -812,11 +790,7 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
int *distortion,
unsigned int *sse1,
const uint8_t *second_pred,
#if CONFIG_AFFINE_MOTION
int w, int h, int use_upsampled_ref) {
#else
int w, int h) {
#endif
const uint8_t *const z = x->plane[0].src.buf;
const uint8_t *const src_address = z;
const int src_stride = x->plane[0].src.stride;
@@ -852,7 +826,6 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
bestmv->row *= 8;
bestmv->col *= 8;
#if CONFIG_AFFINE_MOTION
// use_upsampled_ref can be 0 or 1
if (use_upsampled_ref)
besterr = upsampled_setup_center_error(xd, bestmv, ref_mv, error_per_bit,
@@ -860,7 +833,6 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
second_pred, w, h, (offset << 3),
mvjcost, mvcost, sse1, distortion);
else
#endif
besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp,
z, src_stride, y, y_stride, second_pred,
w, h, offset, mvjcost, mvcost,
@@ -876,7 +848,6 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
if (tc >= minc && tc <= maxc && tr >= minr && tr <= maxr) {
MV this_mv = {tr, tc};
#if CONFIG_AFFINE_MOTION
if (use_upsampled_ref) {
const uint8_t *const pre_address = y + tr * y_stride + tc;
@@ -884,7 +855,6 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
pre_address, y_stride, second_pred,
w, h, &sse);
} else {
#endif
const uint8_t *const pre_address = y + (tr >> 3) * y_stride +
(tc >> 3);
if (second_pred == NULL)
@@ -893,9 +863,7 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
else
thismse = vfp->svaf(pre_address, y_stride, sp(tc), sp(tr),
src_address, src_stride, &sse, second_pred);
#if CONFIG_AFFINE_MOTION
}
#endif
cost_array[idx] = thismse +
mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit);
@@ -920,7 +888,6 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
if (tc >= minc && tc <= maxc && tr >= minr && tr <= maxr) {
MV this_mv = {tr, tc};
#if CONFIG_AFFINE_MOTION
if (use_upsampled_ref) {
const uint8_t *const pre_address = y + tr * y_stride + tc;
@@ -928,7 +895,6 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
pre_address, y_stride, second_pred,
w, h, &sse);
} else {
#endif
const uint8_t *const pre_address = y + (tr >> 3) * y_stride + (tc >> 3);
if (second_pred == NULL)
@@ -937,9 +903,7 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
else
thismse = vfp->svaf(pre_address, y_stride, sp(tc), sp(tr),
src_address, src_stride, &sse, second_pred);
#if CONFIG_AFFINE_MOTION
}
#endif
cost_array[4] = thismse +
mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit);
@@ -963,15 +927,11 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
}
if (iters_per_step > 1 && best_idx != -1) {
#if CONFIG_AFFINE_MOTION
if (use_upsampled_ref) {
SECOND_LEVEL_CHECKS_BEST(1);
} else {
#endif
SECOND_LEVEL_CHECKS_BEST(0);
#if CONFIG_AFFINE_MOTION
}
#endif
}
tr = br;

4
vp10/encoder/mcomp.h
View File

@@ -116,11 +116,7 @@ typedef int (fractional_mv_step_fp) (
int *mvjcost, int *mvcost[2],
int *distortion, unsigned int *sse1,
const uint8_t *second_pred,
#if CONFIG_AFFINE_MOTION
int w, int h, int use_upsampled_ref);
#else
int w, int h);
#endif
extern fractional_mv_step_fp vp10_find_best_sub_pixel_tree;
extern fractional_mv_step_fp vp10_find_best_sub_pixel_tree_pruned;

246
vp10/encoder/rdopt.c
View File

@@ -4664,52 +4664,52 @@ static void joint_motion_search(VP10_COMP *cpi, MACROBLOCK *x,
if (bestsme < INT_MAX) {
int dis; /* TODO: use dis in distortion calculation later. */
unsigned int sse;
#if CONFIG_AFFINE_MOTION
// Use up-sampled reference frames.
struct macroblockd_plane *const pd = &xd->plane[0];
struct buf_2d backup_pred = pd->pre[0];
const YV12_BUFFER_CONFIG *upsampled_ref =
get_upsampled_ref(cpi, refs[id]);
if (cpi->sf.use_upsampled_references) {
// Use up-sampled reference frames.
struct macroblockd_plane *const pd = &xd->plane[0];
struct buf_2d backup_pred = pd->pre[0];
const YV12_BUFFER_CONFIG *upsampled_ref =
get_upsampled_ref(cpi, refs[id]);
// Set pred for Y plane
setup_pred_plane(&pd->pre[0], upsampled_ref->y_buffer,
upsampled_ref->y_stride, (mi_row << 3), (mi_col << 3),
NULL, pd->subsampling_x, pd->subsampling_y);
// Set pred for Y plane
setup_pred_plane(&pd->pre[0], upsampled_ref->y_buffer,
upsampled_ref->y_stride, (mi_row << 3), (mi_col << 3),
NULL, pd->subsampling_x, pd->subsampling_y);
// If bsize < BLOCK_8X8, adjust pred pointer for this block
if (bsize < BLOCK_8X8)
pd->pre[0].buf =
&pd->pre[0].buf[(vp10_raster_block_offset(BLOCK_8X8, block,
pd->pre[0].stride)) << 3];
// If bsize < BLOCK_8X8, adjust pred pointer for this block
if (bsize < BLOCK_8X8)
pd->pre[0].buf =
&pd->pre[0].buf[(vp10_raster_block_offset(BLOCK_8X8, block,
pd->pre[0].stride)) << 3];
bestsme = cpi->find_fractional_mv_step(
x, &tmp_mv,
&ref_mv[id].as_mv,
cpi->common.allow_high_precision_mv,
x->errorperbit,
&cpi->fn_ptr[bsize],
0, cpi->sf.mv.subpel_iters_per_step,
NULL,
x->nmvjointcost, x->mvcost,
&dis, &sse, second_pred,
pw, ph, 1);
bestsme = cpi->find_fractional_mv_step(
x, &tmp_mv,
&ref_mv[id].as_mv,
cpi->common.allow_high_precision_mv,
x->errorperbit,
&cpi->fn_ptr[bsize],
0, cpi->sf.mv.subpel_iters_per_step,
NULL,
x->nmvjointcost, x->mvcost,
&dis, &sse, second_pred,
pw, ph, 1);
// Restore the reference frames.
pd->pre[0] = backup_pred;
#else
(void) block;
bestsme = cpi->find_fractional_mv_step(
x, &tmp_mv,
&ref_mv[id].as_mv,
cpi->common.allow_high_precision_mv,
x->errorperbit,
&cpi->fn_ptr[bsize],
0, cpi->sf.mv.subpel_iters_per_step,
NULL,
x->nmvjointcost, x->mvcost,
&dis, &sse, second_pred,
pw, ph);
#endif
// Restore the reference frames.
pd->pre[0] = backup_pred;
} else {
(void) block;
bestsme = cpi->find_fractional_mv_step(
x, &tmp_mv,
&ref_mv[id].as_mv,
cpi->common.allow_high_precision_mv,
x->errorperbit,
&cpi->fn_ptr[bsize],
0, cpi->sf.mv.subpel_iters_per_step,
NULL,
x->nmvjointcost, x->mvcost,
&dis, &sse, second_pred,
pw, ph, 0);
}
}
// Restore the pointer to the first (possibly scaled) prediction buffer.
@@ -4990,57 +4990,57 @@ static int64_t rd_pick_best_sub8x8_mode(VP10_COMP *cpi, MACROBLOCK *x,
if (bestsme < INT_MAX) {
int distortion;
#if CONFIG_AFFINE_MOTION
const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
// Use up-sampled reference frames.
struct macroblockd_plane *const pd = &xd->plane[0];
struct buf_2d backup_pred = pd->pre[0];
const YV12_BUFFER_CONFIG *upsampled_ref =
get_upsampled_ref(cpi, mbmi->ref_frame[0]);
if (cpi->sf.use_upsampled_references) {
const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
// Use up-sampled reference frames.
struct macroblockd_plane *const pd = &xd->plane[0];
struct buf_2d backup_pred = pd->pre[0];
const YV12_BUFFER_CONFIG *upsampled_ref =
get_upsampled_ref(cpi, mbmi->ref_frame[0]);
// Set pred for Y plane
setup_pred_plane(&pd->pre[0], upsampled_ref->y_buffer,
upsampled_ref->y_stride,
(mi_row << 3), (mi_col << 3),
NULL, pd->subsampling_x, pd->subsampling_y);
// Set pred for Y plane
setup_pred_plane(&pd->pre[0], upsampled_ref->y_buffer,
upsampled_ref->y_stride,
(mi_row << 3), (mi_col << 3),
NULL, pd->subsampling_x, pd->subsampling_y);
// adjust pred pointer for this block
pd->pre[0].buf =
&pd->pre[0].buf[(vp10_raster_block_offset(BLOCK_8X8, i,
pd->pre[0].stride)) << 3];
// adjust pred pointer for this block
pd->pre[0].buf =
&pd->pre[0].buf[(vp10_raster_block_offset(BLOCK_8X8, i,
pd->pre[0].stride)) << 3];
cpi->find_fractional_mv_step(
x,
new_mv,
&bsi->ref_mv[0]->as_mv,
cm->allow_high_precision_mv,
x->errorperbit, &cpi->fn_ptr[bsize],
cpi->sf.mv.subpel_force_stop,
cpi->sf.mv.subpel_iters_per_step,
cond_cost_list(cpi, cost_list),
x->nmvjointcost, x->mvcost,
&distortion,
&x->pred_sse[mbmi->ref_frame[0]],
NULL, pw, ph, 1);
cpi->find_fractional_mv_step(
x,
new_mv,
&bsi->ref_mv[0]->as_mv,
cm->allow_high_precision_mv,
x->errorperbit, &cpi->fn_ptr[bsize],
cpi->sf.mv.subpel_force_stop,
cpi->sf.mv.subpel_iters_per_step,
cond_cost_list(cpi, cost_list),
x->nmvjointcost, x->mvcost,
&distortion,
&x->pred_sse[mbmi->ref_frame[0]],
NULL, pw, ph, 1);
// Restore the reference frames.
pd->pre[0] = backup_pred;
#else
cpi->find_fractional_mv_step(
x,
new_mv,
&bsi->ref_mv[0]->as_mv,
cm->allow_high_precision_mv,
x->errorperbit, &cpi->fn_ptr[bsize],
cpi->sf.mv.subpel_force_stop,
cpi->sf.mv.subpel_iters_per_step,
cond_cost_list(cpi, cost_list),
x->nmvjointcost, x->mvcost,
&distortion,
&x->pred_sse[mbmi->ref_frame[0]],
NULL, 0, 0);
#endif
// Restore the reference frames.
pd->pre[0] = backup_pred;
} else {
cpi->find_fractional_mv_step(
x,
new_mv,
&bsi->ref_mv[0]->as_mv,
cm->allow_high_precision_mv,
x->errorperbit, &cpi->fn_ptr[bsize],
cpi->sf.mv.subpel_force_stop,
cpi->sf.mv.subpel_iters_per_step,
cond_cost_list(cpi, cost_list),
x->nmvjointcost, x->mvcost,
&distortion,
&x->pred_sse[mbmi->ref_frame[0]],
NULL, 0, 0, 0);
}
// save motion search result for use in compound prediction
#if CONFIG_EXT_INTER
@@ -5637,43 +5637,43 @@ static void single_motion_search(VP10_COMP *cpi, MACROBLOCK *x,
if (bestsme < INT_MAX) {
int dis; /* TODO: use dis in distortion calculation later. */
#if CONFIG_AFFINE_MOTION
const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
// Use up-sampled reference frames.
struct macroblockd_plane *const pd = &xd->plane[0];
struct buf_2d backup_pred = pd->pre[ref_idx];
const YV12_BUFFER_CONFIG *upsampled_ref = get_upsampled_ref(cpi, ref);
if (cpi->sf.use_upsampled_references) {
const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
// Use up-sampled reference frames.
struct macroblockd_plane *const pd = &xd->plane[0];
struct buf_2d backup_pred = pd->pre[ref_idx];
const YV12_BUFFER_CONFIG *upsampled_ref = get_upsampled_ref(cpi, ref);
// Set pred for Y plane
setup_pred_plane(&pd->pre[ref_idx], upsampled_ref->y_buffer,
upsampled_ref->y_stride, (mi_row << 3), (mi_col << 3),
NULL, pd->subsampling_x, pd->subsampling_y);
// Set pred for Y plane
setup_pred_plane(&pd->pre[ref_idx], upsampled_ref->y_buffer,
upsampled_ref->y_stride, (mi_row << 3), (mi_col << 3),
NULL, pd->subsampling_x, pd->subsampling_y);
bestsme = cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
cm->allow_high_precision_mv,
x->errorperbit,
&cpi->fn_ptr[bsize],
cpi->sf.mv.subpel_force_stop,
cpi->sf.mv.subpel_iters_per_step,
cond_cost_list(cpi, cost_list),
x->nmvjointcost, x->mvcost,
&dis, &x->pred_sse[ref], NULL,
pw, ph, 1);
bestsme = cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
cm->allow_high_precision_mv,
x->errorperbit,
&cpi->fn_ptr[bsize],
cpi->sf.mv.subpel_force_stop,
cpi->sf.mv.subpel_iters_per_step,
cond_cost_list(cpi, cost_list),
x->nmvjointcost, x->mvcost,
&dis, &x->pred_sse[ref], NULL,
pw, ph, 1);
// Restore the reference frames.
pd->pre[ref_idx] = backup_pred;
#else
cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
cm->allow_high_precision_mv,
x->errorperbit,
&cpi->fn_ptr[bsize],
cpi->sf.mv.subpel_force_stop,
cpi->sf.mv.subpel_iters_per_step,
cond_cost_list(cpi, cost_list),
x->nmvjointcost, x->mvcost,
&dis, &x->pred_sse[ref], NULL, 0, 0);
#endif
// Restore the reference frames.
pd->pre[ref_idx] = backup_pred;
} else {
cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
cm->allow_high_precision_mv,
x->errorperbit,
&cpi->fn_ptr[bsize],
cpi->sf.mv.subpel_force_stop,
cpi->sf.mv.subpel_iters_per_step,
cond_cost_list(cpi, cost_list),
x->nmvjointcost, x->mvcost,
&dis, &x->pred_sse[ref], NULL, 0, 0, 0);
}
}
*rate_mv = vp10_mv_bit_cost(&tmp_mv->as_mv, &ref_mv,
x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);

6
vp10/encoder/rdopt.h
View File

@@ -103,13 +103,16 @@ void vp10_build_prediction_by_left_preds(VP10_COMP *cpi,
int tmp_stride[MAX_MB_PLANE]);
#endif // CONFIG_OBMC
#if CONFIG_AFFINE_MOTION
static INLINE const YV12_BUFFER_CONFIG *get_upsampled_ref(VP10_COMP *cpi,
const int ref) {
// Use up-sampled reference frames.
int ref_idx = 0;
if (ref == LAST_FRAME)
#if CONFIG_EXT_REFS
ref_idx = cpi->lst_fb_idxes[ref - LAST_FRAME];
#else
ref_idx = cpi->lst_fb_idx;
#endif
else if (ref == GOLDEN_FRAME)
ref_idx = cpi->gld_fb_idx;
else if (ref == ALTREF_FRAME)
@@ -117,7 +120,6 @@ static INLINE const YV12_BUFFER_CONFIG *get_upsampled_ref(VP10_COMP *cpi,
return &cpi->upsampled_ref_bufs[cpi->upsampled_ref_idx[ref_idx]].buf;
}
#endif
#ifdef __cplusplus
} // extern "C"

7
vp10/encoder/speed_features.c
View File

@@ -177,6 +177,7 @@ static void set_good_speed_feature(VP10_COMP *cpi, VP10_COMMON *cm,
sf->comp_inter_joint_search_thresh = BLOCK_SIZES;
sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX;
sf->allow_partition_search_skip = 1;
sf->use_upsampled_references = 0;
#if CONFIG_EXT_TX
sf->tx_type_search = PRUNE_TWO;
#endif
@@ -279,6 +280,7 @@ static void set_rt_speed_feature(VP10_COMP *cpi, SPEED_FEATURES *sf,
sf->use_fast_coef_costing = 1;
sf->allow_exhaustive_searches = 0;
sf->exhaustive_searches_thresh = INT_MAX;
sf->use_upsampled_references = 0;
// Use transform domain distortion computation
// Note var-tx expt always uses pixel domain distortion.
@@ -495,6 +497,11 @@ void vp10_set_speed_features_framesize_independent(VP10_COMP *cpi) {
sf->disable_filter_search_var_thresh = 0;
sf->adaptive_interp_filter_search = 0;
sf->allow_partition_search_skip = 0;
#if CONFIG_EXT_REFS
sf->use_upsampled_references = 0;
#else
sf->use_upsampled_references = 1;
#endif
for (i = 0; i < TX_SIZES; i++) {
sf->intra_y_mode_mask[i] = INTRA_ALL;

3
vp10/encoder/speed_features.h
View File

@@ -479,6 +479,9 @@ typedef struct SPEED_FEATURES {
// Fast approximation of vp10_model_rd_from_var_lapndz
int simple_model_rd_from_var;
// Do sub-pixel search in up-sampled reference frames
int use_upsampled_references;
// Whether to compute distortion in the image domain (slower but
// more accurate), or in the transform domain (faster but less acurate).
int use_transform_domain_distortion;

4
vp10/encoder/temporal_filter.c
View File

@@ -320,11 +320,7 @@ static int temporal_filter_find_matching_mb_c(VP10_COMP *cpi,
0, mv_sf->subpel_iters_per_step,
cond_cost_list(cpi, cost_list),
NULL, NULL,
#if CONFIG_AFFINE_MOTION
&distortion, &sse, NULL, 0, 0, 0);
#else
&distortion, &sse, NULL, 0, 0);
#endif
// Restore input state
x->plane[0].src = src;

2
vpx_dsp/variance.c
View File

@@ -267,7 +267,6 @@ void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred,
}
}
#if CONFIG_AFFINE_MOTION
// Get pred block from up-sampled reference.
void vpx_upsampled_pred_c(uint8_t *comp_pred,
int width, int height,
@@ -300,7 +299,6 @@ void vpx_comp_avg_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred,
ref += stride;
}
}
#endif
#if CONFIG_VP9_HIGHBITDEPTH
static void highbd_variance64(const uint8_t *a8, int a_stride,

10
vpx_dsp/vpx_dsp_rtcd_defs.pl
View File

@@ -1233,12 +1233,10 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
#
# ...
#
if (vpx_config("CONFIG_AFFINE_MOTION") eq "yes") {
add_proto qw/void vpx_upsampled_pred/, "uint8_t *comp_pred, int width, int height, const uint8_t *ref, int ref_stride";
specialize qw/vpx_upsampled_pred sse2/;
add_proto qw/void vpx_comp_avg_upsampled_pred/, "uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride";
specialize qw/vpx_comp_avg_upsampled_pred sse2/;
}
add_proto qw/void vpx_upsampled_pred/, "uint8_t *comp_pred, int width, int height, const uint8_t *ref, int ref_stride";
specialize qw/vpx_upsampled_pred sse2/;
add_proto qw/void vpx_comp_avg_upsampled_pred/, "uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride";
specialize qw/vpx_comp_avg_upsampled_pred sse2/;
#
# ...

2
vpx_dsp/x86/variance_sse2.c
View File

@@ -476,7 +476,6 @@ FNS(ssse3, ssse3);
#undef FN
#endif // CONFIG_USE_X86INC
#if CONFIG_AFFINE_MOTION
void vpx_upsampled_pred_sse2(uint8_t *comp_pred,
int width, int height,
const uint8_t *ref, int ref_stride) {
@@ -703,4 +702,3 @@ void vpx_comp_avg_upsampled_pred_sse2(uint8_t *comp_pred, const uint8_t *pred,
}
}
}
#endif