Implement Loop restoration multi-threading
Row-based multi-threading of loop-restoration has been added.
Change-Id: I26f96ded4b7b48d5513da1725828043703cd5c00
diff --git a/av1/common/restoration.c b/av1/common/restoration.c
index cd08995..be5825a 100644
--- a/av1/common/restoration.c
+++ b/av1/common/restoration.c
@@ -1241,14 +1241,12 @@
av1_loop_restoration_copy_planes(loop_rest_ctxt, cm, num_planes);
}
-void av1_foreach_rest_unit_in_row(RestorationTileLimits *limits,
- const AV1PixelRect *tile_rect,
- rest_unit_visitor_t on_rest_unit,
- int row_number, int unit_size, int unit_idx0,
- int hunits_per_tile, int plane, void *priv,
- int32_t *tmpbuf, RestorationLineBuffers *rlbs,
- sync_read_fn_t on_sync_read,
- sync_write_fn_t on_sync_write) {
+void av1_foreach_rest_unit_in_row(
+ RestorationTileLimits *limits, const AV1PixelRect *tile_rect,
+ rest_unit_visitor_t on_rest_unit, int row_number, int unit_size,
+ int unit_idx0, int hunits_per_tile, int plane, void *priv, int32_t *tmpbuf,
+ RestorationLineBuffers *rlbs, sync_read_fn_t on_sync_read,
+ sync_write_fn_t on_sync_write, struct AV1LrSyncData *const lr_sync) {
const int tile_w = tile_rect->right - tile_rect->left;
const int ext_size = unit_size * 3 / 2;
int x0 = 0, j = 0;
@@ -1262,11 +1260,11 @@
const int unit_idx = unit_idx0 + row_number * hunits_per_tile + j;
- on_sync_read(NULL, row_number, j, plane);
+ on_sync_read(lr_sync, row_number, j, plane);
on_rest_unit(limits, tile_rect, unit_idx, priv, tmpbuf, rlbs);
- on_sync_write(NULL, row_number, j, hunits_per_tile, plane);
+ on_sync_write(lr_sync, row_number, j, hunits_per_tile, plane);
x0 += w;
++j;
@@ -1319,7 +1317,7 @@
av1_foreach_rest_unit_in_row(&limits, tile_rect, on_rest_unit, i, unit_size,
unit_idx0, hunits_per_tile, plane, priv,
tmpbuf, rlbs, av1_lr_sync_read_dummy,
- av1_lr_sync_write_dummy);
+ av1_lr_sync_write_dummy, NULL);
y0 += h;
++i;
diff --git a/av1/common/restoration.h b/av1/common/restoration.h
index 81a03e3..b66b0b6 100644
--- a/av1/common/restoration.h
+++ b/av1/common/restoration.h
@@ -320,6 +320,7 @@
typedef void (*rest_tile_start_visitor_t)(int tile_row, int tile_col,
void *priv);
+struct AV1LrSyncData;
typedef void (*sync_read_fn_t)(void *const lr_sync, int r, int c, int plane);
@@ -356,14 +357,12 @@
int optimized_lr, int num_planes);
void av1_loop_restoration_copy_planes(AV1LrStruct *loop_rest_ctxt,
struct AV1Common *cm, int num_planes);
-void av1_foreach_rest_unit_in_row(RestorationTileLimits *limits,
- const AV1PixelRect *tile_rect,
- rest_unit_visitor_t on_rest_unit,
- int row_number, int unit_size, int unit_idx0,
- int hunits_per_tile, int plane, void *priv,
- int32_t *tmpbuf, RestorationLineBuffers *rlbs,
- sync_read_fn_t on_sync_read,
- sync_write_fn_t on_sync_write);
+void av1_foreach_rest_unit_in_row(
+ RestorationTileLimits *limits, const AV1PixelRect *tile_rect,
+ rest_unit_visitor_t on_rest_unit, int row_number, int unit_size,
+ int unit_idx0, int hunits_per_tile, int plane, void *priv, int32_t *tmpbuf,
+ RestorationLineBuffers *rlbs, sync_read_fn_t on_sync_read,
+ sync_write_fn_t on_sync_write, struct AV1LrSyncData *const lr_sync);
AV1PixelRect av1_whole_frame_rect(const struct AV1Common *cm, int is_uv);
int av1_lr_count_units_in_tile(int unit_size, int tile_size);
void av1_lr_sync_read_dummy(void *const lr_sync, int r, int c, int plane);
diff --git a/av1/common/thread_common.c b/av1/common/thread_common.c
index de97478..c365047 100644
--- a/av1/common/thread_common.c
+++ b/av1/common/thread_common.c
@@ -10,6 +10,7 @@
*/
#include "config/aom_config.h"
+#include "config/aom_scale_rtcd.h"
#include "aom_dsp/aom_dsp_common.h"
#include "aom_mem/aom_mem.h"
@@ -32,6 +33,24 @@
return 8;
}
+static INLINE int get_lr_sync_range(int width) {
+#if 0
+ // nsync numbers are picked by testing. For example, for 4k
+ // video, using 4 gives best performance.
+ if (width < 640)
+ return 1;
+ else if (width <= 1280)
+ return 2;
+ else if (width <= 4096)
+ return 4;
+ else
+ return 8;
+#else
+ (void)width;
+ return 1;
+#endif
+}
+
// Allocate memory for lf row synchronization
static void loop_filter_alloc(AV1LfSync *lf_sync, AV1_COMMON *cm, int rows,
int width, int num_workers) {
@@ -363,3 +382,352 @@
loop_filter_rows_mt(frame, cm, xd, start_mi_row, end_mi_row, plane_start,
plane_end, workers, num_workers, lf_sync);
}
+
+static INLINE void lr_sync_read(void *const lr_sync, int r, int c, int plane) {
+#if CONFIG_MULTITHREAD
+ AV1LrSync *const loop_res_sync = (AV1LrSync *)lr_sync;
+ const int nsync = loop_res_sync->sync_range;
+
+ if (r && !(c & (nsync - 1))) {
+ pthread_mutex_t *const mutex = &loop_res_sync->mutex_[plane][r - 1];
+ pthread_mutex_lock(mutex);
+
+ while (c > loop_res_sync->cur_sb_col[plane][r - 1] - nsync) {
+ pthread_cond_wait(&loop_res_sync->cond_[plane][r - 1], mutex);
+ }
+ pthread_mutex_unlock(mutex);
+ }
+#else
+ (void)lr_sync;
+ (void)r;
+ (void)c;
+ (void)plane;
+#endif // CONFIG_MULTITHREAD
+}
+
+static INLINE void lr_sync_write(void *const lr_sync, int r, int c,
+ const int sb_cols, int plane) {
+#if CONFIG_MULTITHREAD
+ AV1LrSync *const loop_res_sync = (AV1LrSync *)lr_sync;
+ const int nsync = loop_res_sync->sync_range;
+ int cur;
+ // Only signal when there are enough filtered SB for next row to run.
+ int sig = 1;
+
+ if (c < sb_cols - 1) {
+ cur = c;
+ if (c % nsync) sig = 0;
+ } else {
+ cur = sb_cols + nsync;
+ }
+
+ if (sig) {
+ pthread_mutex_lock(&loop_res_sync->mutex_[plane][r]);
+
+ loop_res_sync->cur_sb_col[plane][r] = cur;
+
+ pthread_cond_broadcast(&loop_res_sync->cond_[plane][r]);
+ pthread_mutex_unlock(&loop_res_sync->mutex_[plane][r]);
+ }
+#else
+ (void)lr_sync;
+ (void)r;
+ (void)c;
+ (void)sb_cols;
+ (void)plane;
+#endif // CONFIG_MULTITHREAD
+}
+
+// Allocate memory for loop restoration row synchronization
+static void loop_restoration_alloc(AV1LrSync *lr_sync, AV1_COMMON *cm,
+ int num_workers, int num_rows_lr,
+ int num_planes, int width) {
+ lr_sync->rows = num_rows_lr;
+ lr_sync->num_planes = num_planes;
+#if CONFIG_MULTITHREAD
+ {
+ int i, j;
+
+ for (j = 0; j < num_planes; j++) {
+ CHECK_MEM_ERROR(cm, lr_sync->mutex_[j],
+ aom_malloc(sizeof(*(lr_sync->mutex_[j])) * num_rows_lr));
+ if (lr_sync->mutex_[j]) {
+ for (i = 0; i < num_rows_lr; ++i) {
+ pthread_mutex_init(&lr_sync->mutex_[j][i], NULL);
+ }
+ }
+
+ CHECK_MEM_ERROR(cm, lr_sync->cond_[j],
+ aom_malloc(sizeof(*(lr_sync->cond_[j])) * num_rows_lr));
+ if (lr_sync->cond_[j]) {
+ for (i = 0; i < num_rows_lr; ++i) {
+ pthread_cond_init(&lr_sync->cond_[j][i], NULL);
+ }
+ }
+ }
+
+ CHECK_MEM_ERROR(cm, lr_sync->job_mutex,
+ aom_malloc(sizeof(*(lr_sync->job_mutex))));
+ if (lr_sync->job_mutex) {
+ pthread_mutex_init(lr_sync->job_mutex, NULL);
+ }
+ }
+#endif // CONFIG_MULTITHREAD
+ CHECK_MEM_ERROR(cm, lr_sync->lrworkerdata,
+ aom_malloc(num_workers * sizeof(*(lr_sync->lrworkerdata))));
+
+ for (int worker_idx = 0; worker_idx < num_workers; ++worker_idx) {
+ if (worker_idx < num_workers - 1) {
+ CHECK_MEM_ERROR(cm, lr_sync->lrworkerdata[worker_idx].rst_tmpbuf,
+ (int32_t *)aom_memalign(16, RESTORATION_TMPBUF_SIZE));
+ CHECK_MEM_ERROR(cm, lr_sync->lrworkerdata[worker_idx].rlbs,
+ aom_malloc(sizeof(RestorationLineBuffers)));
+
+ } else {
+ lr_sync->lrworkerdata[worker_idx].rst_tmpbuf = cm->rst_tmpbuf;
+ lr_sync->lrworkerdata[worker_idx].rlbs = cm->rlbs;
+ }
+ }
+
+ lr_sync->num_workers = num_workers;
+
+ for (int j = 0; j < num_planes; j++) {
+ CHECK_MEM_ERROR(
+ cm, lr_sync->cur_sb_col[j],
+ aom_malloc(sizeof(*(lr_sync->cur_sb_col[j])) * num_rows_lr));
+ }
+ CHECK_MEM_ERROR(
+ cm, lr_sync->job_queue,
+ aom_malloc(sizeof(*(lr_sync->job_queue)) * num_rows_lr * num_planes));
+ // Set up nsync.
+ lr_sync->sync_range = get_lr_sync_range(width);
+}
+
+// Deallocate loop restoration synchronization related mutex and data
+void av1_loop_restoration_dealloc(AV1LrSync *lr_sync, int num_workers) {
+ if (lr_sync != NULL) {
+ int j;
+#if CONFIG_MULTITHREAD
+ int i;
+ for (j = 0; j < MAX_MB_PLANE; j++) {
+ if (lr_sync->mutex_[j] != NULL) {
+ for (i = 0; i < lr_sync->rows; ++i) {
+ pthread_mutex_destroy(&lr_sync->mutex_[j][i]);
+ }
+ aom_free(lr_sync->mutex_[j]);
+ }
+ if (lr_sync->cond_[j] != NULL) {
+ for (i = 0; i < lr_sync->rows; ++i) {
+ pthread_cond_destroy(&lr_sync->cond_[j][i]);
+ }
+ aom_free(lr_sync->cond_[j]);
+ }
+ }
+ if (lr_sync->job_mutex != NULL) {
+ pthread_mutex_destroy(lr_sync->job_mutex);
+ aom_free(lr_sync->job_mutex);
+ }
+#endif // CONFIG_MULTITHREAD
+ for (j = 0; j < MAX_MB_PLANE; j++) {
+ aom_free(lr_sync->cur_sb_col[j]);
+ }
+
+ aom_free(lr_sync->job_queue);
+
+ if (lr_sync->lrworkerdata) {
+ for (int worker_idx = 0; worker_idx < num_workers - 1; worker_idx++) {
+ LRWorkerData *const workerdata_data =
+ lr_sync->lrworkerdata + worker_idx;
+
+ aom_free(workerdata_data->rst_tmpbuf);
+ aom_free(workerdata_data->rlbs);
+ }
+ aom_free(lr_sync->lrworkerdata);
+ }
+
+ // clear the structure as the source of this call may be a resize in which
+ // case this call will be followed by an _alloc() which may fail.
+ av1_zero(*lr_sync);
+ }
+}
+
+static void enqueue_lr_jobs(AV1LrSync *lr_sync, AV1LrStruct *lr_ctxt,
+ AV1_COMMON *cm) {
+ FilterFrameCtxt *ctxt = lr_ctxt->ctxt;
+
+ const int num_planes = av1_num_planes(cm);
+ AV1LrMTInfo *lr_job_queue = lr_sync->job_queue;
+ lr_sync->jobs_enqueued = 0;
+ lr_sync->jobs_dequeued = 0;
+
+ for (int plane = 0; plane < num_planes; plane++) {
+ if (cm->rst_info[plane].frame_restoration_type == RESTORE_NONE) continue;
+ const int is_uv = plane > 0;
+ const int ss_y = is_uv && cm->subsampling_y;
+
+ AV1PixelRect tile_rect = ctxt[plane].tile_rect;
+ const int unit_size = ctxt[plane].rsi->restoration_unit_size;
+
+ const int tile_h = tile_rect.bottom - tile_rect.top;
+ const int ext_size = unit_size * 3 / 2;
+
+ int y0 = 0, i = 0;
+ while (y0 < tile_h) {
+ int remaining_h = tile_h - y0;
+ int h = (remaining_h < ext_size) ? remaining_h : unit_size;
+
+ RestorationTileLimits limits;
+ limits.v_start = tile_rect.top + y0;
+ limits.v_end = tile_rect.top + y0 + h;
+ assert(limits.v_end <= tile_rect.bottom);
+ // Offset the tile upwards to align with the restoration processing stripe
+ const int voffset = RESTORATION_UNIT_OFFSET >> ss_y;
+ limits.v_start = AOMMAX(tile_rect.top, limits.v_start - voffset);
+ if (limits.v_end < tile_rect.bottom) limits.v_end -= voffset;
+
+ lr_job_queue->lr_unit_row = i;
+ lr_job_queue->plane = plane;
+ lr_job_queue->v_start = limits.v_start;
+ lr_job_queue->v_end = limits.v_end;
+ lr_job_queue++;
+ lr_sync->jobs_enqueued++;
+
+ y0 += h;
+ ++i;
+ }
+ }
+}
+
+AV1LrMTInfo *get_lr_job_info(AV1LrSync *lr_sync) {
+ AV1LrMTInfo *cur_job_info = NULL;
+
+#if CONFIG_MULTITHREAD
+ pthread_mutex_lock(lr_sync->job_mutex);
+
+ if (lr_sync->jobs_dequeued < lr_sync->jobs_enqueued) {
+ cur_job_info = lr_sync->job_queue + lr_sync->jobs_dequeued;
+ lr_sync->jobs_dequeued++;
+ }
+
+ pthread_mutex_unlock(lr_sync->job_mutex);
+#else
+ (void)lr_sync;
+#endif
+
+ return cur_job_info;
+}
+
+// Implement row loop restoration for each thread.
+static int loop_restoration_row_worker(AV1LrSync *const lr_sync,
+ LRWorkerData *lrworkerdata) {
+ AV1LrStruct *lr_ctxt = (AV1LrStruct *)lrworkerdata->lr_ctxt;
+ FilterFrameCtxt *ctxt = lr_ctxt->ctxt;
+ int lr_unit_row;
+ int plane;
+ const int tile_row = LR_TILE_ROW;
+ const int tile_col = LR_TILE_COL;
+ const int tile_cols = LR_TILE_COLS;
+ const int tile_idx = tile_col + tile_row * tile_cols;
+
+ while (1) {
+ AV1LrMTInfo *cur_job_info = get_lr_job_info(lr_sync);
+ if (cur_job_info != NULL) {
+ RestorationTileLimits limits;
+ limits.v_start = cur_job_info->v_start;
+ limits.v_end = cur_job_info->v_end;
+ lr_unit_row = cur_job_info->lr_unit_row;
+ plane = cur_job_info->plane;
+ const int unit_idx0 = tile_idx * ctxt[plane].rsi->units_per_tile;
+
+ av1_foreach_rest_unit_in_row(
+ &limits, &(ctxt[plane].tile_rect), lr_ctxt->on_rest_unit, lr_unit_row,
+ ctxt[plane].rsi->restoration_unit_size, unit_idx0,
+ ctxt[plane].rsi->horz_units_per_tile, plane, &ctxt[plane],
+ lrworkerdata->rst_tmpbuf, lrworkerdata->rlbs, lr_sync_read,
+ lr_sync_write, lr_sync);
+ } else {
+ break;
+ }
+ }
+ return 1;
+}
+
+static void foreach_rest_unit_in_planes_mt(AV1LrStruct *lr_ctxt,
+ AVxWorker *workers, int nworkers,
+ AV1LrSync *lr_sync, AV1_COMMON *cm) {
+ FilterFrameCtxt *ctxt = lr_ctxt->ctxt;
+
+ const int num_planes = av1_num_planes(cm);
+
+ const AVxWorkerInterface *const winterface = aom_get_worker_interface();
+ int num_rows_lr = 0;
+
+ for (int plane = 0; plane < num_planes; plane++) {
+ const AV1PixelRect tile_rect = ctxt[plane].tile_rect;
+ const int max_tile_h = tile_rect.bottom - tile_rect.top;
+
+ const int unit_size = cm->seq_params.sb_size == BLOCK_128X128 ? 128 : 64;
+
+ num_rows_lr =
+ AOMMAX(num_rows_lr, av1_lr_count_units_in_tile(unit_size, max_tile_h));
+ }
+
+ const int num_workers = nworkers;
+ int i;
+ assert(MAX_MB_PLANE == 3);
+
+ if (!lr_sync->sync_range || num_rows_lr != lr_sync->rows ||
+ num_workers > lr_sync->num_workers || num_planes != lr_sync->num_planes) {
+ av1_loop_restoration_dealloc(lr_sync, num_workers);
+ loop_restoration_alloc(lr_sync, cm, num_workers, num_rows_lr, num_planes,
+ cm->width);
+ }
+
+ // Initialize cur_sb_col to -1 for all SB rows.
+ for (i = 0; i < num_planes; i++) {
+ memset(lr_sync->cur_sb_col[i], -1,
+ sizeof(*(lr_sync->cur_sb_col[i])) * num_rows_lr);
+ }
+
+ enqueue_lr_jobs(lr_sync, lr_ctxt, cm);
+
+ // Set up looprestoration thread data.
+ for (i = 0; i < num_workers; ++i) {
+ AVxWorker *const worker = &workers[i];
+ lr_sync->lrworkerdata[i].lr_ctxt = (void *)lr_ctxt;
+ worker->hook = (AVxWorkerHook)loop_restoration_row_worker;
+ worker->data1 = lr_sync;
+ worker->data2 = &lr_sync->lrworkerdata[i];
+
+ // Start loopfiltering
+ if (i == num_workers - 1) {
+ winterface->execute(worker);
+ } else {
+ winterface->launch(worker);
+ }
+ }
+
+ // Wait till all rows are finished
+ for (i = 0; i < num_workers; ++i) {
+ winterface->sync(&workers[i]);
+ }
+}
+
+void av1_loop_restoration_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
+ AV1_COMMON *cm, int optimized_lr,
+ AVxWorker *workers, int num_workers,
+ AV1LrSync *lr_sync, void *lr_ctxt) {
+ assert(!cm->all_lossless);
+
+ const int num_planes = av1_num_planes(cm);
+
+ AV1LrStruct *loop_rest_ctxt = (AV1LrStruct *)lr_ctxt;
+
+ av1_loop_restoration_filter_frame_init(loop_rest_ctxt, frame, cm,
+ optimized_lr, num_planes);
+
+ foreach_rest_unit_in_planes_mt(loop_rest_ctxt, workers, num_workers, lr_sync,
+ cm);
+
+ av1_loop_restoration_copy_planes(loop_rest_ctxt, cm, num_planes);
+}
diff --git a/av1/common/thread_common.h b/av1/common/thread_common.h
index 1327251..cf47562 100644
--- a/av1/common/thread_common.h
+++ b/av1/common/thread_common.h
@@ -54,6 +54,46 @@
int jobs_dequeued;
} AV1LfSync;
+typedef struct AV1LrMTInfo {
+ int v_start;
+ int v_end;
+ int lr_unit_row;
+ int plane;
+} AV1LrMTInfo;
+
+typedef struct LoopRestorationWorkerData {
+ int32_t *rst_tmpbuf;
+ void *rlbs;
+ void *lr_ctxt;
+} LRWorkerData;
+
+// Looprestoration row synchronization
+typedef struct AV1LrSyncData {
+#if CONFIG_MULTITHREAD
+ pthread_mutex_t *mutex_[MAX_MB_PLANE];
+ pthread_cond_t *cond_[MAX_MB_PLANE];
+#endif
+ // Allocate memory to store the loop-restoration block index in each row.
+ int *cur_sb_col[MAX_MB_PLANE];
+ // The optimal sync_range for different resolution and platform should be
+ // determined by testing. Currently, it is chosen to be a power-of-2 number.
+ int sync_range;
+ int rows;
+ int num_planes;
+
+ int num_workers;
+
+#if CONFIG_MULTITHREAD
+ pthread_mutex_t *job_mutex;
+#endif
+ // Row-based parallel loopfilter data
+ LRWorkerData *lrworkerdata;
+
+ AV1LrMTInfo *job_queue;
+ int jobs_enqueued;
+ int jobs_dequeued;
+} AV1LrSync;
+
// Deallocate loopfilter synchronization related mutex and data.
void av1_loop_filter_dealloc(AV1LfSync *lf_sync);
@@ -62,6 +102,12 @@
int plane_end, int partial_frame,
AVxWorker *workers, int num_workers,
AV1LfSync *lf_sync);
+void av1_loop_restoration_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
+ struct AV1Common *cm,
+ int optimized_lr, AVxWorker *workers,
+ int num_workers, AV1LrSync *lr_sync,
+ void *lr_ctxt);
+void av1_loop_restoration_dealloc(AV1LrSync *lr_sync, int num_workers);
#ifdef __cplusplus
} // extern "C"
diff --git a/av1/decoder/decodeframe.c b/av1/decoder/decodeframe.c
index de96409..0f6685c 100644
--- a/av1/decoder/decodeframe.c
+++ b/av1/decoder/decodeframe.c
@@ -4029,17 +4029,32 @@
if (do_loop_restoration) {
av1_loop_restoration_save_boundary_lines(&pbi->cur_buf->buf, cm, 1);
- av1_loop_restoration_filter_frame((YV12_BUFFER_CONFIG *)xd->cur_buf, cm,
- optimized_loop_restoration,
- &pbi->lr_ctxt);
+ if (pbi->num_workers > 1) {
+ av1_loop_restoration_filter_frame_mt(
+ (YV12_BUFFER_CONFIG *)xd->cur_buf, cm, optimized_loop_restoration,
+ pbi->tile_workers, pbi->num_workers, &pbi->lr_row_sync,
+ &pbi->lr_ctxt);
+ } else {
+ av1_loop_restoration_filter_frame((YV12_BUFFER_CONFIG *)xd->cur_buf,
+ cm, optimized_loop_restoration,
+ &pbi->lr_ctxt);
+ }
}
} else {
// In no cdef and no superres case. Provide an optimized version of
// loop_restoration_filter.
- if (do_loop_restoration)
- av1_loop_restoration_filter_frame((YV12_BUFFER_CONFIG *)xd->cur_buf, cm,
- optimized_loop_restoration,
- &pbi->lr_ctxt);
+ if (do_loop_restoration) {
+ if (pbi->num_workers > 1) {
+ av1_loop_restoration_filter_frame_mt(
+ (YV12_BUFFER_CONFIG *)xd->cur_buf, cm, optimized_loop_restoration,
+ pbi->tile_workers, pbi->num_workers, &pbi->lr_row_sync,
+ &pbi->lr_ctxt);
+ } else {
+ av1_loop_restoration_filter_frame((YV12_BUFFER_CONFIG *)xd->cur_buf,
+ cm, optimized_loop_restoration,
+ &pbi->lr_ctxt);
+ }
+ }
}
}
diff --git a/av1/decoder/decoder.c b/av1/decoder/decoder.c
index 6a4907d..066d416 100644
--- a/av1/decoder/decoder.c
+++ b/av1/decoder/decoder.c
@@ -156,6 +156,7 @@
if (pbi->num_workers > 0) {
av1_loop_filter_dealloc(&pbi->lf_row_sync);
+ av1_loop_restoration_dealloc(&pbi->lr_row_sync, pbi->num_workers);
}
#if CONFIG_ACCOUNTING
diff --git a/av1/decoder/decoder.h b/av1/decoder/decoder.h
index a854ccb..e6e61f7 100644
--- a/av1/decoder/decoder.h
+++ b/av1/decoder/decoder.h
@@ -68,6 +68,7 @@
AVxWorker *frame_worker_owner; // frame_worker that owns this pbi.
AVxWorker lf_worker;
AV1LfSync lf_row_sync;
+ AV1LrSync lr_row_sync;
AV1LrStruct lr_ctxt;
AVxWorker *tile_workers;
int num_workers;
diff --git a/av1/encoder/encoder.c b/av1/encoder/encoder.c
index efd8a39..e3ce750 100644
--- a/av1/encoder/encoder.c
+++ b/av1/encoder/encoder.c
@@ -2895,6 +2895,7 @@
if (cpi->num_workers > 1) {
av1_loop_filter_dealloc(&cpi->lf_row_sync);
+ av1_loop_restoration_dealloc(&cpi->lr_row_sync, cpi->num_workers);
}
dealloc_compressor_data(cpi);
@@ -4119,8 +4120,13 @@
if (cm->rst_info[0].frame_restoration_type != RESTORE_NONE ||
cm->rst_info[1].frame_restoration_type != RESTORE_NONE ||
cm->rst_info[2].frame_restoration_type != RESTORE_NONE) {
- av1_loop_restoration_filter_frame(cm->frame_to_show, cm, 0,
- &cpi->lr_ctxt);
+ if (cpi->num_workers > 1)
+ av1_loop_restoration_filter_frame_mt(cm->frame_to_show, cm, 0,
+ cpi->workers, cpi->num_workers,
+ &cpi->lr_row_sync, &cpi->lr_ctxt);
+ else
+ av1_loop_restoration_filter_frame(cm->frame_to_show, cm, 0,
+ &cpi->lr_ctxt);
}
}
}
diff --git a/av1/encoder/encoder.h b/av1/encoder/encoder.h
index 3673da7..507190f 100644
--- a/av1/encoder/encoder.h
+++ b/av1/encoder/encoder.h
@@ -686,6 +686,7 @@
int ref_conv[REF_FRAMES];
AV1LfSync lf_row_sync;
+ AV1LrSync lr_row_sync;
AV1LrStruct lr_ctxt;
} AV1_COMP;
