vello/piet-gpu/shader/tile_alloc.comp
Raph Levien 294f6fd1db Experiment with new sorting scheme
Path segments are unsorted, but other elements are using the same
sort-middle approach as before.

This is a checkpoint. At this point, there are unoptimized versions
of tile init and coarse path raster, but it isn't wired up into a
working pipeline. Also observing about a 3x performance regression in
element processing, which needs to be investigated.
2020-06-03 09:29:25 -07:00

74 lines
2 KiB
Plaintext

// Allocation and initialization of tiles for paths.
#version 450
#extension GL_GOOGLE_include_directive : enable
#include "setup.h"
#define TILE_ALLOC_WG 32
layout(local_size_x = TILE_ALLOC_WG, local_size_y = 1) in;
layout(set = 0, binding = 0) buffer AnnotatedBuf {
uint[] annotated;
};
layout(set = 0, binding = 1) buffer AllocBuf {
uint n_elements;
uint n_pathseg;
uint alloc;
};
layout(set = 0, binding = 2) buffer TileBuf {
uint[] tile;
};
#include "annotated.h"
#include "tile.h"
// scale factors useful for converting coordinates to tiles
#define SX (1.0 / float(TILE_WIDTH_PX))
#define SY (1.0 / float(TILE_HEIGHT_PX))
void main() {
uint element_ix = gl_GlobalInvocationID.x;
PathRef path_ref = PathRef(element_ix * Path_size);
AnnotatedRef ref = AnnotatedRef(element_ix * Annotated_size);
uint tag = Annotated_Nop;
if (element_ix < n_elements) {
tag = Annotated_tag(ref);
}
int x0 = 0, y0 = 0, x1 = 0, y1 = 0;
switch (tag) {
case Annotated_Fill:
case Annotated_Stroke:
// Note: we take advantage of the fact that fills and strokes
// have compatible layout.
AnnoFill fill = Annotated_Fill_read(ref);
x0 = int(floor(fill.bbox.x * SX));
y0 = int(floor(fill.bbox.y * SY));
x1 = int(ceil(fill.bbox.z * SX));
y1 = int(ceil(fill.bbox.w * SY));
break;
}
x0 = clamp(x0, 0, WIDTH_IN_TILES);
y0 = clamp(y0, 0, HEIGHT_IN_TILES);
x1 = clamp(x1, 0, WIDTH_IN_TILES);
y1 = clamp(y1, 0, HEIGHT_IN_TILES);
Path path;
path.bbox = uvec4(x0, y0, x1, y1);
uint n_tiles = (x1 - x0) * (y1 - y0);
path.tiles = TileRef(0);
if (n_tiles > 0) {
path.tiles.offset = atomicAdd(alloc, n_tiles * Tile_size);
Tile init_tile = Tile(TileSegRef(0), 0);
// TODO: improve load balancing
for (uint i = 0; i < n_tiles; i++) {
Tile_write(Tile_index(path.tiles, i), init_tile);
}
}
Path_write(path_ref, path);
}