Shaders
- ui-text and ui-fused turned a sample of the R8_UNORM atlas into a
distance with (128/255 - sdf) * 32. stb_truetype writes kOnEdgeValue +
kPixelDistScale * distance into each byte, so an atlas pixel is
kPixelDistScale / 255 of a sample; 32 made every distance 1024/255
(about 4) times too large, and the coverage ramp, meant to span one
screen pixel, spanned a quarter of one: nearly every edge pixel came out
fully on or fully off, and text looked pixelated at every size. On an
RTX 4090, UIFallbackParity's 22 px label had 98 partly covered pixels
where stb_truetype's own rasterizer has 415. Wrong since the v3 UI
(1f56973), and copied into the fused kernel (574242d).
- The decode now lives once, in ui-shared.glsl: UI_SDF_ON_EDGE and
UI_SDF_PIXEL_DIST_SCALE are FontAtlas's two constants in the encoder's
own units, uiSdfDistance decodes a sample, uiGlyphRamp works out a
glyph's ramp when it is loaded and uiGlyphCoverage the coverage per
pixel. Both kernels call them, and a copy of either should; the UI_
prefix keeps them clear of a copied kernel's own names.
- stb_truetype truncates, so a byte b stands for [b, b + 1) of the encoded
value and the decode takes its middle: decoding b pulled every outline
1/64 atlas pixel inwards.
- The ramp is capped at what a byte holds, about 4 atlas pixels each way
of the outline. With the corrected decode a wider one left every glyph
quad's padding translucent below 4 px.
- The UV span is taken with abs(): a glyph turned by swapping its UVs got
a negative scale, a ramp of 0.0001 and hard edges.
- A glyph scaled differently along x and y ramps along the distance's
gradient, from two one-texel differences. With the corrected decode, one
width for both axes, the more minified one's, would leave a row stretched
three times with stems three pixels soft. Glyphs scaled alike, as
ShapeText's are, keep one multiply-add per pixel, the ramp's reciprocal
worked out with the glyph.
- The atlas size is queried rather than FontAtlas::kAtlasSize hard-coded
as 1024.
WebGPU
- The WGSL text kernel decoded the distance right but ramped over one
atlas pixel whatever the size: half a screen pixel at 16 px, two at
64 px. It now repeats ui-shared.glsl's decode, cap, abs and gradient
path, and renders TextAntialiasing's scene as Vulkan does (run through
Deno's WebGPU: 1.02, 1.07 and 1.09 times the reference's edge pixels,
where the old kernel gave 0.55, 2.19 and 0.89).
FontAtlas
- kOnEdgeValue and kPixelDistScale say the kernels carry copies of them.
Tests
- TextAntialiasing draws one string at 14 px, 64 px and 14 px stretched
three times, a 'g' turned by its UVs and empty atlas texels shrunk to
12.8 atlas pixels per screen pixel, through DispatchText and again
through DispatchFused, and holds them to stb_truetype's antialiased
rasterizer, to the mirror image and to nothing. On master the two plain
rows have 0.31 and 0.21 times the reference's edge pixels; with the
decode fixed alone the stretched row had 1.62 times, the turned 'g' was
127 codes off its mirror and the empty texels drew 48. Now: 1.02, 1.07
and 1.09 times, 0 and 0, on the 4090, on llvmpipe and on descriptor
sets. It skips a swapchain that isn't 8-bit, which the capture can't
read, and captures the first frame.
- FontAtlasDecode reads ui-shared.glsl and dom-webgpu.js and holds their
copies of the two constants to FontAtlas's. No GPU.
Docs
- AGENTS.consumer.md has a custom text shader decode through the helpers
and gives the encoding; README says so in a line; AGENTS.md lists the
tests and where the decode lives.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
WebGPUBuffer::EnqueueReadback / PollReadback always copied the full
buffer `size` GPU→staging→wasm. Over-provisioned event-queue buffers
(e.g. Forts3D's GPU physics event drain) paid the full-capacity copy
every drain even when only a small live prefix held data.
Add an optional `byteCount` parameter (default 0 = whole buffer, so the
existing full-buffer callers are unchanged) bounding the readback to the
live prefix. Pass the same byteCount to the paired PollReadback so the
matching number of bytes lands in `.value`.
JS bridge: the staging buffer is now sized to the full device-buffer
capacity (not the first call's byteSize) so a varying prefix length never
overflows it, while the copyBufferToBuffer + mapAsync/getMappedRange are
bounded to the requested prefix — that's where the copy saving lands.
Verified end-to-end in the browser via RayQueryPick: the full readback
still resolves correctly, and a follow-up 8-byte prefix read copies only
hit+customIndex while primitiveIndex keeps its poison sentinel, proving
the copy was bounded. Native `crafter-build test` suite: 24 passed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
uiResolveScreenPixel ran four float compares against hdr.clipRectPx for
every pixel of every UI dispatch, even though the clip rect is the
full-surface default {0,0,1e9,1e9} in the overwhelmingly common case.
Reserve the high bit of the header flags word (UI_FLAG_CLIP) and have
FillHeader set it only when the clip rect is narrower than the surface
(extracted into the unit-testable UIRenderer::ClipFlags). The shader
gates the compares on the bit; the branch is push-constant-uniform so it
never diverges. When the rect covers the surface the skipped compares
could never reject an in-surface pixel, so output is pixel-identical.
Mirrored on the WebGPU/WGSL path (dom-webgpu.js) for parity. Adds the
UIClipFlag CPU test covering the gate decision.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add zero-copy procedural BLAS overloads that take the AABB build input
straight from an existing device buffer instead of memcpy-ing a host span
through Mesh::aabbBuffer — the input-source companion to #36's refit work.
A GPU compute producer (e.g. a GPU-resident particle system) can now feed a
moving procedural BLAS that builds/refits each frame with no host round-trip.
Vulkan: new BuildProcedural(VkDeviceAddress, count, ...) and
RefitProcedural(VkDeviceAddress, count, ...) feed the device address directly
into VkAccelerationStructureGeometryAabbsDataKHR; the host-span path is
refactored to share RecordProceduralBuildFromAddress and is otherwise
unchanged.
WebGPU: device-buffer BuildProcedural/RefitProcedural copy the boxes GPU->GPU
into the mesh heap (new wgpuRegisterMeshBLASDeviceAabbs / wgpuRefitMeshBLAS-
DeviceAabbs bridge fns) and wrap them in a single root leaf bounded by a
caller-supplied worldBounds — no wasm round-trip, blasAddr stable across refit.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Window::onMouseScroll now fires on every backend and speaks one unit:
signed whole detents packed in the uint32 payload, +1 = wheel down
(DOM deltaY sign).
- Wayland: implement the previously-stubbed PointerListenerHandleAxis —
vertical axis only, wl_fixed_to_double / 15 (libinput's units-per-
detent), sub-detent remainder accumulated and reset on pointer leave.
- Win32: add the missing WM_MOUSEWHEEL case — -GET_WHEEL_DELTA_WPARAM /
WHEEL_DELTA with a remainder accumulator for free-spinning wheels.
- DOM: dom-env.js normalizes WheelEvent.deltaY per deltaMode (100px /
3 lines / 1 page per detent) with the same remainder scheme, so wasm
delivers the identical unit instead of raw browser-specific deltas.
- Document the contract on Window::onMouseScroll.
- tests/MouseScroll: compositor-free regression test driving the
Wayland axis handler directly.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
wfAccum becomes array<atomic<u32>> (4 slots/pixel, f32 bit patterns — same
16 B/pixel footprint) and rtAccumulate CASes each channel, so N rays for one
pixel may resolve in the same SHADE pass without the read-modify-write
racing. GENERATE clears with atomicStore (bitcast(0.0) == 0u); RESOLVE
atomicLoads + bitcasts the vec4 it hands runResolve.
Capacity half: RTPass::raysPerPixel (default 1) scales the wavefront
ray/hit/payload buffers to raysPerPixel·W·H rays per bounce so the
per-light emits actually fit instead of being dropped by rtEmitRay's
capacity guard. The accumulator stays per-pixel.
No flag-gating: uncontended the CAS succeeds first try — RTStress SHADE
stays at its documented ~1.0 ms and master-vs-branch renders are
byte-identical, so single-ray consumers pay nothing.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds the minimal WebGPU-backend surface an HDR bloom/post chain needs:
1. wgpuCreateStorageImage2D + StorageImage2D<>: runtime rgba16float (or
other format) textures with STORAGE_BINDING|TEXTURE_BINDING usage, no
CompressedTextureAsset upload — GPU-produced HDR targets.
2. UICustomBindingKind::StorageTexture (kind 5): write-only storage-texture
binding carrying a texel format (the freed UICustomBinding::format byte),
wired through all three binding paths (UI custom / RT / plain compute).
Float textures already sample via SampledTexture (sampleType float).
3. PipelineRTWebGPU::Init(..., hdrOutputFormat): RESOLVE writes the linear
accumulator into a user rgba16float texture (RTPass::outTexHandle)
instead of the rgba8unorm canvas, leaving the composite→swapchain pass to
the app. Default RGBA8Unorm keeps the canvas path byte-identical.
Existing RT examples updated for the _pad→format field rename; the default
paths are unchanged (verified RTStress builds + renders).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The software rayQuery shim's _rqTraverseTlas detected BVH leaves with a
compile-time constant TLAS_BVH_LEAVES_START = 16384 - 1, while the actual
TLAS sweep tree is built at depth log2(next_pow2(instanceCount)). For any
scene with fewer than 8193 instances the padded leaf count is far below
16384, so no node index ever reached 16383: every node looked internal,
the descent walked into zeroed out-of-tree AABBs, and the pick reported a
permanent miss. This broke every rayQuery=true compute shader (builder
picking, splash queries) on the WebGPU backend.
Pass the per-build padded leaf count to the shim the same way the
megakernel _rtwTraverseTlas reads wfParams.tlasNPadded: a small uniform
(RqTlasMeta.nPadded) at @group(1) @binding(10), written each wgpuBuildTLAS
from wfNextPow2(instanceCount), and bound by both rayQuery dispatch paths.
_rqTraverseTlas now computes leavesStart = nPadded - 1 dynamically.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The TRACE-stage BLAS descent now threads the payload through, runs the
any-hit shader for non-opaque candidates (DXR/VK opacity resolution:
ray FORCE flags > instance FORCE flags > geometry opaque bit), and
handles AABB leaves via the intersection shader. MeshRecord grows to 64
bytes with geomType + opaque. When any-hit/intersection are present the
TRACE pipeline takes the user bind-group layout so those shaders can
sample @group(3+) resources; otherwise TRACE keeps its zero-user-code
path unchanged. rayQuery stays triangle-only (skips AABB leaves).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A 1-D indirect dispatch of ceil(W*H/64) workgroups for the wavefront
TRACE/SHADE stages overflows maxComputeWorkgroupsPerDimension (65535 on
Dawn/Firefox) once the surface exceeds ~4.19M rays (~2560x1640). Per the
WebGPU spec such a dispatch is silently dropped — no validation error —
so at 4K the world is never traced and the accumulator stays black while
non-RT passes survive.
_wfPrep now spreads the workgroups across a 2-D grid (x clamped to 65535,
y = ceil(wg/65535)), and the wfTrace/wfShade entry points rebuild the
linear ray index from (global_invocation_id, num_workgroups). The existing
`i >= _wfCurCount()` guard absorbs the grid overshoot. GENERATE/RESOLVE
already use a 2-D tile dispatch and are unchanged.
Verified in Firefox/WebGPU with RTStress at a 3449x1739 surface (5.99M
rays, 93716 workgroups — well over the 65535 cap): renders the full cube
grid where master shows a black screen.
Resolves#11
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
dom-webgpu.js capped maxStorageBuffersPerShaderStage at 16 even when the
adapter reports far more (64 in our test env). The wavefront SHADE kernel
already binds ~16 storage buffers before any user binding, so any RT
pipeline declaring 2+ user storage buffers at @group(3) overflowed the
limit and failed to build with "Too many bindings of type StorageBuffers".
Request the adapter's reported maxStorageBuffersPerShaderStage /
maxStorageBuffersInPipelineLayout instead of a fixed 16. `clamp` already
mins against the adapter cap, so baseline-only devices still get a valid
request, and the `|| 16` fallback + the `typeof cap === "number"` guard
handle limit names a browser doesn't expose (Firefox returns null for
maxStorageBuffersInPipelineLayout).
Verified in-browser: a 17-storage-buffer compute pipeline fails with the
exact reported error on a device clamped to 16, and builds cleanly on a
device requesting the adapter's 64. RTStress renders correctly.
Resolves#8
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The LBVH bitonic sort still runs over the full 16384 (sentinels sink to
the tail), but the sweep tree is now built and traced at depth
log2(next_pow2(nReal)) instead of a fixed 14. Add nPadded to LbvhPC; leaf
init + bottom-up refit use it; the host passes the same next_pow2 to the
trace via WfParams.tlasNPadded. Renders correctly at 512 instances
(depth 9). The fragile sort phases are untouched.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The RT pipeline now only builds the wavefront kernels, so the old
single-megakernel traversal/traceRay block (rtWgslMegakernelHelpers) and
the unused rtWgslPrelude alias are dead. Remove them. The rayQuery compute
path keeps rtWgslMegakernelBindings (its own _rq* traversal uses it).
RTStress still renders correctly with the trimmed prelude.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add _rtAabbT (AABB test returning entry-t); in both _rtwTraverseBlas and
_rtwTraverseTlas descend the nearer child first and push the farther only
when it hits, re-culling it against the (tightened) bestT when popped.
Render is identical (same closest hit) on VulkanTriangle, RTStress
(512/4096), and Sponza; cuts node visits on dense scenes.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Request the timestamp-query feature; write begin/end timestamps around
each wavefront pass via timestampWrites; resolve + read back (deferred to
after submit) and print a per-pass us breakdown ~1x/sec. RTStress @ 512
instances, 1920x995: TRACE dominates, total ~1.8-3.0ms/frame.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replace the megakernel @compute entry with five wavefront kernels sharing
one module, connected by GPU ray/hit/payload buffers and a GPU-driven
indirect bounce loop:
GENERATE -> (PREP -> TRACE -> SHADE) x maxDepth -> RESOLVE
- TRACE contains zero user code (pure _rtwTraverseTlas/Blas, opaque-only).
- PREP publishes dispatchWorkgroupsIndirect args from the live ray count;
the indirect-args buffer lives in its own bind group so it is never
bound read-write in the same dispatch that consumes it as INDIRECT.
- New emit/accumulate API: rtEmitPrimaryRay / rtEmitRay / rtAccumulate,
plus an optional user Resolve stage (tonemap hook; identity by default).
- Per-pass WfParams via a dynamic-offset uniform ring (curIsA/bounce vary
between passes within one submit).
- Payload-typed wfPayload binding emitted in the codegen region after the
user's struct Payload; payload travels with each ray (2*W*H slots).
- Request maxBufferSize / maxStorageBufferBindingSize / maxComputeWorkgroups
PerDimension so the W*H-sized work buffers fit past the 128MB baseline.
VulkanTriangle ported to the new API and renders bit-identical to the
megakernel baseline at maxDepth=1.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replace the disabled LSD radix sort in lbvhBuildMain with a data-oblivious
workgroup bitonic sorting network and enable it. The radix scatter was gated
behind `if (false)` because it produced count/distribution-dependent
corruption (TODO-lbvh-sort.md) — a memory-ordering bug in the Hillis-Steele
scan / parallel scatter that surfaced only for certain Morton distributions
(a small object beside a tight cluster), making geometry flicker.
A bitonic network's compare-exchange schedule depends only on N_PADDED, never
on key values, so it sidesteps that entire class of distribution-dependent
races (TODO strategy #5). 105 sub-stages over 2^14 keys, single workgroup of
1024 threads, 8 compare-exchanges/thread/sub-stage, operating in-place on
sortA with a storageBarrier between sub-stages. Sentinel keys (0xFFFFFFFF)
compare largest and settle at the tail, exactly where Phase 4 expects them.
Restores Morton (Z-order) spatial coherence to TLAS BVH leaves, which the
many-instance case needs. Removes the now-dead radix histogram/scan workgroup
memory and constants.
Verified on the Firefox/Dawn WebGPU stack: a GPU unit test diffs the kernel
output against a CPU oracle across all three required distributions
(all-uniform, all-one-bucket, small-object-next-to-cluster) plus random,
reverse, and empty inputs — all match bit-for-bit with a valid index
permutation. Sponza renders correctly with the sort live.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
