Boost backdrop parallelism for the prefix sums

piet-gpu/shader/backdrop.comp

@@ -2,9 +2,10 @@
 
 // Propagation of tile backdrop for filling.
 //
-// Each thread reads one path element and calculates the number of spanned tiles
+// Each thread reads one path element and calculates the row and column counts of spanned tiles
 // based on the bounding box.
-// In a further compaction step, the workgroup loops over the corresponding tile rows per element in parallel.
+// The row count then goes through a prefix sum to redistribute and load-balance the work across the workgroup.
+// In the following step, the workgroup loops over the corresponding tile rows per element in parallel.
 // For each row the per tile backdrop will be read, as calculated in the previous coarse path segment kernel,
 // and propagated from the left to the right (prefix summed).
 //
@@ -20,8 +21,13 @@
 
 #define LG_BACKDROP_WG (7 + LG_WG_FACTOR)
 #define BACKDROP_WG (1 << LG_BACKDROP_WG)
+// Some paths (those covering a large area) can generate a lot of backdrop tiles; BACKDROP_DIST_FACTOR defines how much
+// additional threads should we spawn for parallel row processing. The additional threads does not participate in the
+// earlier stages (calculating the tile counts) but does work in the final prefix sum stage which has a lot more
+// parallelism.
+#define BACKDROP_DIST_FACTOR 4
 
-layout(local_size_x = BACKDROP_WG, local_size_y = 1) in;
+layout(local_size_x = BACKDROP_WG, local_size_y = BACKDROP_DIST_FACTOR) in;
 
 layout(set = 0, binding = 1) readonly buffer ConfigBuf {
     Config conf;
@@ -35,54 +41,58 @@ shared Alloc sh_row_alloc[BACKDROP_WG];
 shared uint sh_row_width[BACKDROP_WG];
 
 void main() {
-    uint th_ix = gl_LocalInvocationID.x;
+    uint th_ix = gl_LocalInvocationIndex;
     uint element_ix = gl_GlobalInvocationID.x;
     AnnotatedRef ref = AnnotatedRef(conf.anno_alloc.offset + element_ix * Annotated_size);
 
     // Work assignment: 1 thread : 1 path element
     uint row_count = 0;
     bool mem_ok = mem_error == NO_ERROR;
-    if (element_ix < conf.n_elements) {
+    if (gl_LocalInvocationID.y == 0) {
-        AnnotatedTag tag = Annotated_tag(conf.anno_alloc, ref);
+        if (element_ix < conf.n_elements) {
-        switch (tag.tag) {
+            AnnotatedTag tag = Annotated_tag(conf.anno_alloc, ref);
-        case Annotated_Image:
+            switch (tag.tag) {
-        case Annotated_BeginClip:
+                case Annotated_Image:
-        case Annotated_Color:
+                case Annotated_BeginClip:
-            if (fill_mode_from_flags(tag.flags) != MODE_NONZERO) {
+                case Annotated_Color:
-                break;
+                if (fill_mode_from_flags(tag.flags) != MODE_NONZERO) {
+                    break;
+                }
+                // Fall through.
+                PathRef path_ref = PathRef(conf.tile_alloc.offset + element_ix * Path_size);
+                Path path = Path_read(conf.tile_alloc, path_ref);
+                sh_row_width[th_ix] = path.bbox.z - path.bbox.x;
+                row_count = path.bbox.w - path.bbox.y;
+                // Paths that don't cross tile top edges don't have backdrops.
+                // Don't apply the optimization to paths that may cross the y = 0
+                // top edge, but clipped to 1 row.
+                if (row_count == 1 && path.bbox.y > 0) {
+                    // Note: this can probably be expanded to width = 2 as
+                    // long as it doesn't cross the left edge.
+                    row_count = 0;
+                }
+                Alloc path_alloc = new_alloc(path.tiles.offset, (path.bbox.z - path.bbox.x) * (path.bbox.w - path.bbox.y) * Tile_size, mem_ok);
+                sh_row_alloc[th_ix] = path_alloc;
             }
-            // Fall through.
-            PathRef path_ref = PathRef(conf.tile_alloc.offset + element_ix * Path_size);
-            Path path = Path_read(conf.tile_alloc, path_ref);
-            sh_row_width[th_ix] = path.bbox.z - path.bbox.x;
-            row_count = path.bbox.w - path.bbox.y;
-            // Paths that don't cross tile top edges don't have backdrops.
-            // Don't apply the optimization to paths that may cross the y = 0
-            // top edge, but clipped to 1 row.
-            if (row_count == 1 && path.bbox.y > 0) {
-                // Note: this can probably be expanded to width = 2 as
-                // long as it doesn't cross the left edge.
-                row_count = 0;
-            }
-            Alloc path_alloc = new_alloc(path.tiles.offset, (path.bbox.z - path.bbox.x) * (path.bbox.w - path.bbox.y) * Tile_size, mem_ok);
-            sh_row_alloc[th_ix] = path_alloc;
         }
+        sh_row_count[th_ix] = row_count;
     }
 
-    sh_row_count[th_ix] = row_count;
     // Prefix sum of sh_row_count
     for (uint i = 0; i < LG_BACKDROP_WG; i++) {
         barrier();
-        if (th_ix >= (1 << i)) {
+        if (gl_LocalInvocationID.y == 0 && th_ix >= (1 << i)) {
             row_count += sh_row_count[th_ix - (1 << i)];
         }
         barrier();
-        sh_row_count[th_ix] = row_count;
+        if (gl_LocalInvocationID.y == 0) {
+            sh_row_count[th_ix] = row_count;
+        }
     }
     barrier();
     // Work assignment: 1 thread : 1 path element row
     uint total_rows = sh_row_count[BACKDROP_WG - 1];
-    for (uint row = th_ix; row < total_rows; row += BACKDROP_WG) {
+    for (uint row = th_ix; row < total_rows; row += BACKDROP_WG * BACKDROP_DIST_FACTOR) {
         // Binary search to find element
         uint el_ix = 0;
         for (uint i = 0; i < LG_BACKDROP_WG; i++) {