perf(ui): subgroup prefix-sum stream compaction in the UI compute shaders (#124) #145

Merged

catbot merged 1 commit from claude/issue-124 into master

2026-06-18 15:27:30 +02:00

catbot commented

2026-06-18 15:27:18 +02:00

Member

Summary

uiCompactChunk (and the same pattern inlined in the four standalone UI shaders) ran the per-chunk survivor scan entirely on gl_LocalInvocationIndex == 0 — ~64 serial iterations with 63/64 lanes idle, between two barriers, 4× per chunk in the fused kernel. It has to produce a stable in-order (buffer-order) exclusive prefix of the survivors into s_order[] so the per-pixel inner loop still sees items in draw order.

This replaces the lane-0 scan with a subgroup exclusive prefix-sum + cross-subgroup carry:

Each subgroup computes subgroupExclusiveAdd(keep) (rank within the subgroup) and subgroupAdd(keep) (its survivor total) at register speed.
One lane per subgroup publishes the total to shared s_subTotals[].
After one barrier, every lane sums the totals of all lower-id subgroups for its base offset.
A survivor's slot = base + subPrefix = the number of survivors with a smaller local index → buffer (draw) order preserved exactly. (An atomicAdd would scramble order — deliberately avoided, per the issue.)

The carry makes it correct for any subgroup width (2 subgroups on the 32-wide descriptor_heap target, up to 8/16 on narrower parts), relying only on the gl_LocalInvocationIndex ↔ (gl_SubgroupID, gl_SubgroupInvocationID) linear mapping every Vulkan compute implementation provides. The feature is free at the device baseline (apiVersion 1.4 + VK_EXT_descriptor_heap implies Vulkan 1.1 subgroup arithmetic), so no correctness fallback is needed. WebGPU is unaffected (it runs separate embedded WGSL).

Applied to ui-fused.comp.glsl and the inlined pattern in ui-quads/circles/images/text.

Verification

crafter-build test — all 23 pass, including UIFusedShader (recompiles ui-fused.comp.glsl with glslang, runs spirv-val, and pins the push-constant member offsets — unchanged).
140k-trial CPU simulation of the exact algorithm matches the serial reference for subgroup sizes {1,2,4,8,16,32,64}.
Runtime on real hardware: HelloUI (which uses DispatchFused with quads + circles + text) renders correctly on an RTX 4090 — the 2-subgroups-per-workgroup carry path. Text composites on top of the button quad and the circle knob on top of the slider track, confirming draw order is intact.

Screenshots

Resolves #124

## Summary `uiCompactChunk` (and the same pattern inlined in the four standalone UI shaders) ran the per-chunk survivor scan entirely on `gl_LocalInvocationIndex == 0` — ~64 serial iterations with 63/64 lanes idle, between two barriers, **4× per chunk** in the fused kernel. It has to produce a **stable in-order** (buffer-order) exclusive prefix of the survivors into `s_order[]` so the per-pixel inner loop still sees items in draw order. This replaces the lane-0 scan with a **subgroup exclusive prefix-sum + cross-subgroup carry**: - Each subgroup computes `subgroupExclusiveAdd(keep)` (rank within the subgroup) and `subgroupAdd(keep)` (its survivor total) at register speed. - One lane per subgroup publishes the total to shared `s_subTotals[]`. - After one barrier, every lane sums the totals of all **lower-id** subgroups for its base offset. - A survivor's slot = `base + subPrefix` = the number of survivors with a smaller local index → **buffer (draw) order preserved exactly**. (An `atomicAdd` would scramble order — deliberately avoided, per the issue.) The carry makes it correct for any subgroup width (2 subgroups on the 32-wide `descriptor_heap` target, up to 8/16 on narrower parts), relying only on the `gl_LocalInvocationIndex` ↔ `(gl_SubgroupID, gl_SubgroupInvocationID)` linear mapping every Vulkan compute implementation provides. The feature is free at the device baseline (`apiVersion 1.4` + `VK_EXT_descriptor_heap` implies Vulkan 1.1 subgroup arithmetic), so no correctness fallback is needed. WebGPU is unaffected (it runs separate embedded WGSL). Applied to `ui-fused.comp.glsl` and the inlined pattern in `ui-quads`/`circles`/`images`/`text`. ## Verification - **`crafter-build test` — all 23 pass**, including `UIFusedShader` (recompiles `ui-fused.comp.glsl` with glslang, runs `spirv-val`, and pins the push-constant member offsets — unchanged). - **140k-trial CPU simulation** of the exact algorithm matches the serial reference for subgroup sizes {1,2,4,8,16,32,64}. - **Runtime on real hardware**: `HelloUI` (which uses `DispatchFused` with quads + circles + text) renders correctly on an RTX 4090 — the 2-subgroups-per-workgroup carry path. Text composites on top of the button quad and the circle knob on top of the slider track, confirming draw order is intact. ## Screenshots ![HelloUI on RTX 4090 — DispatchFused quads+circles+text, draw order intact](https://forgejo.catcrafts.net/attachments/3c40c5a7-54b8-473d-9cef-cb6dc9cddc0f) Resolves #124

result.png

11 KiB

catbot added 1 commit

2026-06-18 15:27:18 +02:00

perf(ui): subgroup prefix-sum stream compaction, not lane-0 serial scan (#124 ) da9f2a89e8

uiCompactChunk ran the survivor scan entirely on gl_LocalInvocationIndex
== 0: ~64 serial iterations with 63/64 lanes idle, between two barriers,
4x per chunk in the fused kernel. It must emit a stable in-order
(buffer-order) exclusive prefix of the survivors so the per-pixel inner
loop still sees items in draw order.

Replace it with a per-subgroup subgroupExclusiveAdd of the keep bits plus
a carry across subgroups: each subgroup publishes its survivor total to
shared memory, then every lane sums the totals of all lower-id subgroups
for its base. A survivor's slot in s_order[] therefore equals the number
of survivors with a smaller local index — buffer (draw) order preserved
exactly, unlike an atomicAdd which would scramble it.

The carry makes the result correct for any subgroup width (2 subgroups on
the 32-wide descriptor_heap target, up to 8/16 on narrower parts), relying
only on the gl_LocalInvocationIndex <-> (gl_SubgroupID,
gl_SubgroupInvocationID) linear mapping every Vulkan compute
implementation provides. The feature is free at the device baseline
(apiVersion 1.4 + VK_EXT_descriptor_heap implies Vulkan 1.1 subgroup
arithmetic), so no correctness fallback is needed; WebGPU is unaffected
(separate embedded WGSL).

Applied to ui-fused.comp.glsl and the same pattern inlined in
ui-quads/circles/images/text.

Verified: all 23 tests pass (UIFusedShader recompiles + spirv-val +
pins push-constant offsets); a 140k-trial CPU simulation of the algorithm
matches the serial reference for subgroup sizes 1..64; HelloUI renders
correctly on an RTX 4090 (DispatchFused quads+circles+text) with draw
order intact.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

catbot merged commit 9046562a90 into master

2026-06-18 15:27:30 +02:00

catbot deleted branch claude/issue-124

2026-06-18 15:27:30 +02:00

catbot referenced this pull request from a commit

2026-06-18 15:27:31 +02:00

Merge pull request 'perf(ui): subgroup prefix-sum stream compaction in the UI compute shaders (#124)' (#145) from claude/issue-124 into master