Catcrafts/Crafter.Graphics
Watch
2
0
Fork
You've already forked Crafter.Graphics
 0
Code Issues 8 Pull requests Projects Releases Packages Wiki Activity Actions

refactor(ui): remove the unused per-member item loaders

Browse source 
LoadImageItem, LoadGlpyhtem, LoadCircleItem and LoadQuadItem in
ui-shared.glsl worked around NVIDIA losing the device when a whole
struct is loaded from a descriptor-heap storage buffer. No kernel here
calls them any more (they read members straight into shared memory), and
NVIDIA 615.78.08 no longer has the fault: every UI kernel rewritten to
whole-struct loads renders the same frame. Shaders that called them read
the item directly instead.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Jorijn van der Graaf 2026-10-19 03:03:16 +02:00
commit f7cfd53562
Signed by: jorijnvdgraaf
GPG key ID: 2937E59CDCC1BCFB
1 changed files with 0 additions and 43 deletions
Download patch file Download diff file Expand all files Collapse all files

43
shaders/ui-shared.glsl
View file

@ -99,49 +99,6 @@ layout(UI_HEAP_BUFFERS, std430) readonly buffer UICircleBuf { CircleItem items[]
layout(UI_HEAP_BUFFERS, std430) readonly buffer UIImageBuf { ImageItem items[]; } uiImageHeap[];
layout(UI_HEAP_BUFFERS, std430) readonly buffer UIGlyphBuf { GlyphItem items[]; } uiGlyphHeap[];
// ──── Driver workaround: per-member SSBO load ────────────────────────────
// `UIItem it = itemHeap[idx].items[i]` emits an OpLoad of a composite type
// from a descriptor-heap'd SSBO, which crashes the GPU on the NVIDIA
// VK_EXT_descriptor_heap path (verified with a 1-float struct repro).
// Reading individual members works (each becomes OpAccessChain + scalar
// OpLoad). LoadItem reassembles the struct member-by-member into a local;
// the rest of the shader then operates on a regular local var.
ImageItem LoadImageItem(uint heap, uint i) {
ImageItem it;
it.rect = uiImageHeap[heap].items[i].rect;
it.uv = uiImageHeap[heap].items[i].uv;
it.tint = uiImageHeap[heap].items[i].tint;
it.slots = uiImageHeap[heap].items[i].slots;
return it;
}
GlyphItem LoadGlpyhtem(uint heap, uint i) {
GlyphItem it;
it.rect = uiGlyphHeap[heap].items[i].rect;
it.uv = uiGlyphHeap[heap].items[i].uv;
it.color = uiGlyphHeap[heap].items[i].color;
return it;
}
CircleItem LoadCircleItem(uint heap, uint i) {
CircleItem it;
it.centerRadius = uiCircleHeap[heap].items[i].centerRadius;
it.color = uiCircleHeap[heap].items[i].color;
it.outline = uiCircleHeap[heap].items[i].outline;
return it;
}
QuadItem LoadQuadItem(uint heap, uint i) {
QuadItem it;
it.rect = uiQuadHeap[heap].items[i].rect;
it.color = uiQuadHeap[heap].items[i].color;
it.corners = uiQuadHeap[heap].items[i].corners;
it.outline = uiQuadHeap[heap].items[i].outline;
return it;
}
// ─── pixel-tile dispatch model ─────────────────────────────────────────
// Standard shaders dispatch one workgroup per 8×8 screen tile. Each thread
// owns ONE pixel and iterates ALL items in order, accumulating the result