Catcrafts/Crafter.Graphics
2
0
Fork 0
Code Issues 1 Pull requests Projects Releases Packages Wiki Activity Actions
Crafter.Graphics/shaders/ui-quads.comp.glsl

103 lines
3.8 KiB
Text
Raw Permalink Normal View History

UI rewrite 3rd attempt
2026-05-02 21:08:20 +02:00
#version 460
#extension GL_GOOGLE_include_directive : enable
#include "ui-shared.glsl"
perf(ui): cooperative shared-memory tile culling in UI compute shaders Each 8×8-tile workgroup previously had all 64 threads walk the entire item list, re-reading every item from the SSBO 64× and running the per-pixel reject for items that never touch the tile — O(tiles × N) item loads dominated by SSBO traffic. The workgroup now streams the item list in chunks of 64: each thread loads one item, tests its AABB against the whole tile, and the survivors are compacted — in original buffer order — into shared memory. Every thread then runs the unchanged per-pixel accumulate over only those survivors. This drops SSBO traffic ~64× (each item read once per workgroup instead of once per pixel) and shrinks the inner loop to the items that actually overlap the tile. Draw order is preserved exactly: chunks run in array order and the in-chunk compaction is a stable in-order scan, so later items still overdraw earlier ones (no atomic-append nondeterminism). The inner loop body is byte-for-byte the original per-pixel logic, so output is pixel-identical — the cull only decides which items reach it. Threads no longer early-return so every thread reaches the workgroup barriers. Applied to ui-quads, ui-circles, ui-images and ui-text; shared helpers (uiTileBounds / uiAabbOverlapsTile) live in ui-shared.glsl. Resolves #46 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-16 15:45:12 +00:00
// One workgroup per 8×8 screen tile. The workgroup cooperatively streams the
// QuadItem list in chunks of 64 (see ui-shared.glsl), culling each chunk
// against the tile and compacting survivors — in buffer order — into shared
// memory; every thread then runs the per-pixel accumulate over only those
// survivors. Item order in the buffer == draw order on screen.
UI rewrite 3rd attempt
2026-05-02 21:08:20 +02:00
layout(push_constant) uniform PC {
UIDispatchHeader hdr;
} pc;
layout(local_size_x = 8, local_size_y = 8, local_size_z = 1) in;
perf(ui): cooperative shared-memory tile culling in UI compute shaders Each 8×8-tile workgroup previously had all 64 threads walk the entire item list, re-reading every item from the SSBO 64× and running the per-pixel reject for items that never touch the tile — O(tiles × N) item loads dominated by SSBO traffic. The workgroup now streams the item list in chunks of 64: each thread loads one item, tests its AABB against the whole tile, and the survivors are compacted — in original buffer order — into shared memory. Every thread then runs the unchanged per-pixel accumulate over only those survivors. This drops SSBO traffic ~64× (each item read once per workgroup instead of once per pixel) and shrinks the inner loop to the items that actually overlap the tile. Draw order is preserved exactly: chunks run in array order and the in-chunk compaction is a stable in-order scan, so later items still overdraw earlier ones (no atomic-append nondeterminism). The inner loop body is byte-for-byte the original per-pixel logic, so output is pixel-identical — the cull only decides which items reach it. Threads no longer early-return so every thread reaches the workgroup barriers. Applied to ui-quads, ui-circles, ui-images and ui-text; shared helpers (uiTileBounds / uiAabbOverlapsTile) live in ui-shared.glsl. Resolves #46 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-16 15:45:12 +00:00
// Per-chunk cooperative-cull scratch (one slot per workgroup thread).
shared vec4 s_rect[UI_CHUNK];
shared vec4 s_color[UI_CHUNK];
shared vec4 s_corners[UI_CHUNK];
shared vec4 s_outline[UI_CHUNK];
shared uint s_keep[UI_CHUNK];
shared uint s_order[UI_CHUNK];
shared uint s_count;
UI rewrite 3rd attempt
2026-05-02 21:08:20 +02:00
void main() {
perf(ui): cooperative shared-memory tile culling in UI compute shaders Each 8×8-tile workgroup previously had all 64 threads walk the entire item list, re-reading every item from the SSBO 64× and running the per-pixel reject for items that never touch the tile — O(tiles × N) item loads dominated by SSBO traffic. The workgroup now streams the item list in chunks of 64: each thread loads one item, tests its AABB against the whole tile, and the survivors are compacted — in original buffer order — into shared memory. Every thread then runs the unchanged per-pixel accumulate over only those survivors. This drops SSBO traffic ~64× (each item read once per workgroup instead of once per pixel) and shrinks the inner loop to the items that actually overlap the tile. Draw order is preserved exactly: chunks run in array order and the in-chunk compaction is a stable in-order scan, so later items still overdraw earlier ones (no atomic-append nondeterminism). The inner loop body is byte-for-byte the original per-pixel logic, so output is pixel-identical — the cull only decides which items reach it. Threads no longer early-return so every thread reaches the workgroup barriers. Applied to ui-quads, ui-circles, ui-images and ui-text; shared helpers (uiTileBounds / uiAabbOverlapsTile) live in ui-shared.glsl. Resolves #46 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-16 15:45:12 +00:00
// NOTE: do not early-return — every thread must reach the barriers below.
UI rewrite 3rd attempt
2026-05-02 21:08:20 +02:00
ivec2 screenPx;
perf(ui): cooperative shared-memory tile culling in UI compute shaders Each 8×8-tile workgroup previously had all 64 threads walk the entire item list, re-reading every item from the SSBO 64× and running the per-pixel reject for items that never touch the tile — O(tiles × N) item loads dominated by SSBO traffic. The workgroup now streams the item list in chunks of 64: each thread loads one item, tests its AABB against the whole tile, and the survivors are compacted — in original buffer order — into shared memory. Every thread then runs the unchanged per-pixel accumulate over only those survivors. This drops SSBO traffic ~64× (each item read once per workgroup instead of once per pixel) and shrinks the inner loop to the items that actually overlap the tile. Draw order is preserved exactly: chunks run in array order and the in-chunk compaction is a stable in-order scan, so later items still overdraw earlier ones (no atomic-append nondeterminism). The inner loop body is byte-for-byte the original per-pixel logic, so output is pixel-identical — the cull only decides which items reach it. Threads no longer early-return so every thread reaches the workgroup barriers. Applied to ui-quads, ui-circles, ui-images and ui-text; shared helpers (uiTileBounds / uiAabbOverlapsTile) live in ui-shared.glsl. Resolves #46 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-16 15:45:12 +00:00
bool valid = uiResolveScreenPixel(pc.hdr, screenPx);
UI rewrite 3rd attempt
2026-05-02 21:08:20 +02:00
perf(ui): cooperative shared-memory tile culling in UI compute shaders Each 8×8-tile workgroup previously had all 64 threads walk the entire item list, re-reading every item from the SSBO 64× and running the per-pixel reject for items that never touch the tile — O(tiles × N) item loads dominated by SSBO traffic. The workgroup now streams the item list in chunks of 64: each thread loads one item, tests its AABB against the whole tile, and the survivors are compacted — in original buffer order — into shared memory. Every thread then runs the unchanged per-pixel accumulate over only those survivors. This drops SSBO traffic ~64× (each item read once per workgroup instead of once per pixel) and shrinks the inner loop to the items that actually overlap the tile. Draw order is preserved exactly: chunks run in array order and the in-chunk compaction is a stable in-order scan, so later items still overdraw earlier ones (no atomic-append nondeterminism). The inner loop body is byte-for-byte the original per-pixel logic, so output is pixel-identical — the cull only decides which items reach it. Threads no longer early-return so every thread reaches the workgroup barriers. Applied to ui-quads, ui-circles, ui-images and ui-text; shared helpers (uiTileBounds / uiAabbOverlapsTile) live in ui-shared.glsl. Resolves #46 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-16 15:45:12 +00:00
vec4 dst = vec4(0.0);
vec2 sp = vec2(0.0);
if (valid) {
dst = imageLoad(uiImages[pc.hdr.outImage], screenPx);
sp = vec2(screenPx) + 0.5;
UI rewrite 3rd attempt
2026-05-02 21:08:20 +02:00
}
perf(ui): cooperative shared-memory tile culling in UI compute shaders Each 8×8-tile workgroup previously had all 64 threads walk the entire item list, re-reading every item from the SSBO 64× and running the per-pixel reject for items that never touch the tile — O(tiles × N) item loads dominated by SSBO traffic. The workgroup now streams the item list in chunks of 64: each thread loads one item, tests its AABB against the whole tile, and the survivors are compacted — in original buffer order — into shared memory. Every thread then runs the unchanged per-pixel accumulate over only those survivors. This drops SSBO traffic ~64× (each item read once per workgroup instead of once per pixel) and shrinks the inner loop to the items that actually overlap the tile. Draw order is preserved exactly: chunks run in array order and the in-chunk compaction is a stable in-order scan, so later items still overdraw earlier ones (no atomic-append nondeterminism). The inner loop body is byte-for-byte the original per-pixel logic, so output is pixel-identical — the cull only decides which items reach it. Threads no longer early-return so every thread reaches the workgroup barriers. Applied to ui-quads, ui-circles, ui-images and ui-text; shared helpers (uiTileBounds / uiAabbOverlapsTile) live in ui-shared.glsl. Resolves #46 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-16 15:45:12 +00:00
vec2 tileMin, tileMax;
uiTileBounds(tileMin, tileMax);
uint lid = gl_LocalInvocationIndex;
for (uint base = 0u; base < pc.hdr.itemCount; base += UI_CHUNK) {
// Cooperative load: each thread reads one item member-by-member
// (the NVIDIA descriptor-heap workaround) into shared and culls it.
uint idx = base + lid;
bool keep = false;
if (idx < pc.hdr.itemCount) {
s_rect[lid] = uiQuadHeap[pc.hdr.itemBuffer].items[idx].rect;
s_color[lid] = uiQuadHeap[pc.hdr.itemBuffer].items[idx].color;
s_corners[lid] = uiQuadHeap[pc.hdr.itemBuffer].items[idx].corners;
s_outline[lid] = uiQuadHeap[pc.hdr.itemBuffer].items[idx].outline;
keep = uiAabbOverlapsTile(s_rect[lid].xy, s_rect[lid].xy + s_rect[lid].zw,
tileMin, tileMax);
}
s_keep[lid] = keep ? 1u : 0u;
barrier();
// Stable in-order compaction of this chunk's survivors.
if (lid == 0u) {
uint n = 0u;
uint lim = min(UI_CHUNK, pc.hdr.itemCount - base);
for (uint k = 0u; k < lim; ++k)
if (s_keep[k] != 0u) s_order[n++] = k;
s_count = n;
}
barrier();
if (valid) {
for (uint j = 0u; j < s_count; ++j) {
uint c = s_order[j];
// Cheap pre-test against the item's axis-aligned rect.
vec2 lo = s_rect[c].xy;
vec2 hi = s_rect[c].xy + s_rect[c].zw;
if (sp.x < lo.x || sp.y < lo.y) continue;
if (sp.x >= hi.x || sp.y >= hi.y) continue;
vec2 halfSize = s_rect[c].zw * 0.5;
vec2 p = sp - (s_rect[c].xy + halfSize);
float d = uiSdRoundRect(p, halfSize, s_corners[c]);
vec4 outline = s_outline[c];
float bodyA = clamp(0.5 - d, 0.0, 1.0);
if (bodyA <= 0.0 && outline.x <= 0.0) continue;
vec4 col = s_color[c];
vec4 src = vec4(col.rgb, col.a * bodyA);
if (outline.x > 0.0) {
float t = abs(d + outline.x * 0.5) - outline.x * 0.5;
float outlineA = clamp(0.5 - t, 0.0, 1.0);
src.rgb = mix(src.rgb, outline.yzw, outlineA);
src.a = max(src.a, outlineA);
}
if (src.a <= 0.0) continue;
dst = uiBlendOver(dst, src);
}
}
barrier(); // done reading shared for this chunk before it's overwritten
}
if (valid) imageStore(uiImages[pc.hdr.outImage], screenPx, dst);
UI rewrite 3rd attempt
2026-05-02 21:08:20 +02:00
}
Reference in a new issue Copy permalink