Allow CDEF and LR to run sbrows in parallel

1 files changed, 13 insertions, 7 deletions

diff --git a/src/internal.h b/src/internal.h
index d389906..f3cba13 100644
--- a/src/internal.h
+++ b/src/internal.h
@@ -175,6 +175,7 @@ enum TaskType {
     DAV1D_TASK_TYPE_SUPER_RESOLUTION,
     DAV1D_TASK_TYPE_LOOP_RESTORATION,
     DAV1D_TASK_TYPE_ENTROPY_PROGRESS,
+    DAV1D_TASK_TYPE_RECONSTRUCTION_PROGRESS,
 };
 
 struct Dav1dTask {
@@ -183,7 +184,7 @@ struct Dav1dTask {
     int sby;                    // sbrow
 
     // task dependencies
-    int recon_progress, deblock_progress, cdef_progress, lr_progress;
+    int recon_progress, deblock_progress;
     int deps_skip;
     struct Dav1dTask *next; // only used in task queue
 };
@@ -226,7 +227,7 @@ struct Dav1dFrameContext {
         filter_sbrow_fn filter_sbrow;
         filter_sbrow_fn filter_sbrow_deblock_cols;
         filter_sbrow_fn filter_sbrow_deblock_rows;
-        filter_sbrow_fn filter_sbrow_cdef;
+        void (*filter_sbrow_cdef)(Dav1dTaskContext *tc, int sby);
         filter_sbrow_fn filter_sbrow_resize;
         filter_sbrow_fn filter_sbrow_lr;
         backup_ipred_edge_fn backup_ipred_edge;
@@ -248,7 +249,8 @@ struct Dav1dFrameContext {
     struct {
         int next_tile_row[2 /* 0: reconstruction, 1: entropy */];
         int entropy_progress;
-        atomic_int deblock_progress, cdef_progress, lr_progress; // in sby units
+        atomic_int deblock_progress; // in sby units
+        atomic_uint *frame_progress, *copy_lpf_progress;
         // indexed using t->by * f->b4_stride + t->bx
         Av1Block *b;
         struct CodedBlockInfo {
@@ -260,6 +262,7 @@ struct Dav1dFrameContext {
         // iterated over inside tile state
         uint8_t *pal_idx;
         coef *cf;
+        int prog_sz;
         int pal_sz, pal_idx_sz, cf_sz;
         // start offsets per tile
         int *tile_start_off;
@@ -270,21 +273,23 @@ struct Dav1dFrameContext {
         uint8_t (*level)[4];
         Av1Filter *mask;
         Av1Restoration *lr_mask;
-        int top_pre_cdef_toggle;
-        int mask_sz /* w*h */, lr_mask_sz, cdef_line_sz[2] /* stride */;
-        size_t lr_plane_sz; /* w*sbh*4*is_sb128 if n_tc > 1, else w*12 */
+        int mask_sz /* w*h */, lr_mask_sz;
+        int cdef_buf_plane_sz[2]; /* stride*sbh*4 */
+        int lr_buf_plane_sz[2]; /* (stride*sbh*4) << sb128 if n_tc > 1, else stride*4 */
         int re_sz /* h */;
         ALIGN(Av1FilterLUT lim_lut, 16);
         int last_sharpness;
         uint8_t lvl[8 /* seg_id */][4 /* dir */][8 /* ref */][2 /* is_gmv */];
         uint8_t *tx_lpf_right_edge[2];
-        uint8_t *cdef_line_buf;
+        uint8_t *cdef_line_buf, *lr_line_buf;
         pixel *cdef_line[2 /* pre, post */][3 /* plane */];
+        pixel *cdef_lpf_line[3 /* plane */];
         pixel *lr_lpf_line[3 /* plane */];
 
         // in-loop filter per-frame state keeping
         uint8_t *start_of_tile_row;
         int start_of_tile_row_sz;
+        int need_cdef_lpf_copy;
         pixel *p[3], *sr_p[3];
         Av1Filter *mask_ptr, *prev_mask_ptr;
         int restore_planes; // enum LrRestorePlanes
@@ -403,6 +408,7 @@ struct Dav1dTaskContext {
 
     Dav1dWarpedMotionParams warpmv;
     Av1Filter *lf_mask;
+    int top_pre_cdef_toggle;
     int8_t *cur_sb_cdef_idx_ptr;
     // for chroma sub8x8, we need to know the filter for all 4 subblocks in
     // a 4x4 area, but the top/left one can go out of cache already, so this