mirror of
https://github.com/italicsjenga/vello.git
synced 2025-01-10 20:51:29 +11:00
ac3ac3ddff
Both the Vulkan and OpenGL ES spec allow implementations to limit workgroups to 128 threads. Add a LG_WG_FACTOR setting for easy switching between 128 and 256 threads, with 256 being kept as the default setting. Manually tested that LG_WG_FACTOR = 0 (128 threads) works as expected. Signed-off-by: Elias Naur <mail@eliasnaur.com>
100 lines
2.8 KiB
Plaintext
100 lines
2.8 KiB
Plaintext
// Allocation and initialization of tiles for paths.
|
|
|
|
#version 450
|
|
#extension GL_GOOGLE_include_directive : enable
|
|
|
|
#include "setup.h"
|
|
|
|
#define LG_TILE_ALLOC_WG (7 + LG_WG_FACTOR)
|
|
#define TILE_ALLOC_WG (1 << LG_TILE_ALLOC_WG)
|
|
|
|
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))
|
|
|
|
shared uint sh_tile_count[TILE_ALLOC_WG];
|
|
shared uint sh_tile_alloc;
|
|
|
|
void main() {
|
|
uint th_ix = gl_LocalInvocationID.x;
|
|
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 tile_count = (x1 - x0) * (y1 - y0);
|
|
|
|
sh_tile_count[th_ix] = tile_count;
|
|
// Prefix sum of sh_tile_count
|
|
for (uint i = 0; i < LG_TILE_ALLOC_WG; i++) {
|
|
barrier();
|
|
if (th_ix >= (1 << i)) {
|
|
tile_count += sh_tile_count[th_ix - (1 << i)];
|
|
}
|
|
barrier();
|
|
sh_tile_count[th_ix] = tile_count;
|
|
}
|
|
if (th_ix == TILE_ALLOC_WG - 1) {
|
|
sh_tile_alloc = atomicAdd(alloc, tile_count * Tile_size);
|
|
}
|
|
barrier();
|
|
uint alloc_start = sh_tile_alloc;
|
|
|
|
if (element_ix < n_elements) {
|
|
uint tile_subix = th_ix > 0 ? sh_tile_count[th_ix - 1] : 0;
|
|
path.tiles = TileRef(alloc_start + Tile_size * tile_subix);
|
|
Path_write(path_ref, path);
|
|
}
|
|
|
|
// Zero out allocated tiles efficiently
|
|
uint total_count = sh_tile_count[TILE_ALLOC_WG - 1] * (Tile_size / 4);
|
|
uint start_ix = alloc_start >> 2;
|
|
for (uint i = th_ix; i < total_count; i += TILE_ALLOC_WG) {
|
|
// Note: this interleaving is faster than using Tile_write
|
|
// by a significant amount.
|
|
tile[start_ix + i] = 0;
|
|
}
|
|
}
|