Two distinct validation errors the native frame loop emitted, both
originating in Crafter.Graphics with no consumer-side influence.
Problem 1 — per-frame acquire-barrier access/stage mismatch. The
acquire->GENERAL barrier hardcoded dstAccessMask =
SHADER_WRITE|TRANSFER_WRITE but used the per-pass stage union as its dst
stage mask. For an all-compute frame the union narrows to COMPUTE_SHADER,
which does not support TRANSFER_WRITE, so VUID-02820 fired every frame.
Derive the access mask from the same stage union via a new
SwapchainWriterAccess() helper (mirroring SwapchainStageUnion), and apply
it to both the acquire dst and present src masks for symmetry.
Problem 2 — mid-session StartInit/FinishInit (and GetCmd/EndCmd) reuse the
shared drawCmdBuffers[currentBuffer]. With no steady-state wait-idle the
loop's last submission of that buffer may still be in flight when scene
setup runs (building map meshes / acceleration structures), so the old
code re-began (VUID-00049) and re-submitted (VUID-00071) a pending buffer,
and resources freed in the StartInit..FinishInit bracket could still be
referenced by it. Drain the queue at the start of StartInit/GetCmd before
re-recording; setup is rare, so a wait-idle is fine (FinishInit/EndCmd
already wait-idle at the end).
Tests: extend SwapchainBarrierScope with SwapchainWriterAccess coverage
(pure CPU), and add SetupCmdBufferReuse — a real-frame-loop regression
test driving a compute pass plus interleaved mid-session StartInit rounds,
asserting the validation layer stays silent. Verified both halves fail
(reproducing the exact VUIDs) when their respective fix is reverted.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
When `updating` is false (paused / minimized / stopped), the FIFO
present no longer paces StartSync()'s message loop, so the bare
PeekMessage(PM_REMOVE) spin pinned a core at 100% while uselessly
re-running Gamepad::Tick() (mutex + per-pad COM reads) and
onBeforeUpdate every iteration.
Block on MsgWaitForMultipleObjectsEx with a short timeout while
!updating: the loop now wakes on input or every ~16ms, eliminating the
busy-spin while preserving gamepad polling / onBeforeUpdate at a sane
rate (e.g. to detect a controller button asking to resume). The
updating path is unchanged — present remains vsync-gated.
Co-Authored-By: Claude Opus 4.8 <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>
BuildTLASUpload pushed GPU-driven (transformOwnedByGpu) runs one element
at a time — a separate FlushDeviceRange of the 16 strided metadata bytes
per element — each paying WebGPU validation / encode / JS-boundary cost,
while the CPU-driven arm already batched contiguously.
Upload the whole active instance range in a single writeBuffer instead.
Pushing the (stale) transform bytes for GPU-driven slots is harmless: the
only supported way to drive a transform from the GPU is the manual
Upload -> physics compute pass -> Build sequence, and that compute pass
runs after this upload and rewrites the transform on the GPU before the
TLAS build reads it. For the in-repo (all-CPU) usage the bytes uploaded
are identical to before — with one CPU-driven run the old loop already
emitted exactly this single FlushDeviceRange(0, 0, primitiveCount*64).
Verified: native suite (24 passed) and the RTStress wasm example (512
ray-traced instances) render correctly through the new path.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The metadata mirror is padded to kNPadded (65536) but only primitiveCount
slots are live. metadataBuffer.FlushDevice() wgpuWriteBuffer'd all 256 KB
through the WASM->JS staging path every frame (~100-250x waste for a
few-hundred-instance scene). Switch to the existing FlushDeviceRange
overload — the same one the instanceBuffer loop directly above uses —
sized to primitiveCount * sizeof(uint32_t). The Vulkan parallel already
sizes its flush to primitiveCount, so this was WebGPU-specific.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
UIRenderer::Dispatch built a fresh std::vector<uint32_t> with reserve on
every call, paying one malloc+free per dispatch. Custom compute dispatches
run per-frame, so make the vector thread_local and clear() it each call:
capacity grows to the high-water mark once and is reused thereafter,
eliminating the per-frame allocation churn. Behavior is identical.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
FontAtlas::Update used a single union bounding box for the dirty rect.
Scattered new glyphs landing on different shelves produced a tall union
box that re-uploaded mostly-unchanged texels between the shelves.
Track a dirty span per shelf instead and issue one tight UpdateRegion /
wgpuWriteAtlasRegion copy per armed span. A shelf packs glyphs
left-to-right at a fixed top, so each span stays a contiguous X run
capped by the shelf height. The whole-atlas zero-clear in Initialize
keeps using the existing dirtyRect span. `dirty` is now the OR of every
span and is cleared together with them in Update.
Verified: full test suite green (23 passed); HelloUI text renders
crisply on the WebGPU backend through the new multi-span upload path.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
DrawInputField re-measured the cursor prefix via Font::GetLineWidth on
every frame of a focused field, even though only the blink (caretVisible)
changes frame-to-frame — value, cursorPos and fontSize are stable across
the vast majority of frames.
Cache the measured prefix WIDTH on the InputField, keyed on (prefix bytes,
fontSize). A cheap byte compare of the cached prefix guards the expensive
per-glyph UTF-8 decode + advance accumulation. The cache is mutable so the
const-ref draw fn can refresh it.
The absolute caretX is deliberately NOT cached: it adds the layout-
dependent textX (rect.x + paddingX) each frame, so caching it would give a
relocated field a stale caret. Caching the width keeps that correct.
Adds tests/InputFieldCaretCache pinning: the cached caret matches the
uncached GetLineWidth oracle across states, a redraw (cache hit) is
identical to the miss, the cache invalidates on value/cursorPos/fontSize
changes, the blink never moves the caret, and a relocated field shifts the
caret by exactly the rect delta.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
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>
The standalone per-category UI shaders (quads/circles/images/text)
unconditionally imageLoad the destination pixel up front and imageStore
it at the end, paying a full read-modify-write even on tiles no item
touches — the common case for a sparse UI. The fused uber-kernel already
amortizes a single load/store across all four categories; the standalone
Dispatch* path has no such umbrella.
Defer the imageLoad to just before the first surviving blend (`loaded`
flag) and skip the imageStore unless something was actually blended.
Output is bit-identical: before the first blend `dst` is unused, and a
skipped store leaves the pixel exactly as a no-op load+store would have
rewritten it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
At kMaxShapedRuns (8192) UIRenderer::ShapeText did a full shapedRuns_.clear().
A stream of unique strings (FPS counters, timers) alongside many stable labels
periodically nuked every stable entry, forcing a full-UI reshape the next frame.
Track recency with an auxiliary std::list<const ShapedRunKey*> (front = most
recently used) holding pointers to the keys owned by the node-based map. A hit
splices its entry to the front; an overflow pops the least-recently-used entry
from the back and erases just that one — both O(1). The hot set of labels,
reshaped every frame, stays at the front and survives, so the churn of unique
strings only recycles the cold tail. Output is byte-identical either way.
InvalidateFont now drops matching entries from both the map and the LRU list.
Adds ShapedRunCacheSize()/IsShapedRunCached() introspection hooks (the policy
isn't observable through output or atlas.dirty) and a ShapeTextCache test that
asserts the cache plateaus at a fixed cap (evict-one, not clear-all), the hot
label survives a churn past the cap, and an untouched cold string is evicted.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The cooperative-load section already streams each item into shared memory
once per workgroup, but the per-pixel inner loop then recomputed values
that are constant for the whole item/glyph. The compiler can't hoist them
itself — they read shared memory at a varying index.
Precompute them once per item at load time into new shared slots:
- ui-images / ui-text / ui-fused (images+text phases): invRectSize =
1.0/rect.zw, turning the per-pixel `(sp-rect.xy)/rect.zw` vec2 divide
into a multiply.
- ui-text / ui-fused (text phase): the SDF AA `band` (a vec2 divide +
two maxes depending only on the glyph's uv span and rect size).
In ui-fused the new slots (s_inv, s_band) are reused across phases like
the existing s_v* scratch, so the LDS bump is per-category, not summed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
ShapeText built ShapedRunKey{..., std::string(utf8)} before find()
unconditionally — even on cache hits, on the per-frame onBuild path.
N labels meant N string copies + hashes (+ frees for non-SSO strings)
every frame, partially defeating the run cache.
Add is_transparent hash + equality functors on shapedRuns_ and probe
with a borrowing ShapedRunViewKey {const Font*, float, array<float,4>,
string_view}. The owning std::string is now materialised only on a
miss, for emplace. Cache hits copy and hash no string.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
BuildTLAS rebuilt the host instance+metadata buffers with an O(n) copy of
every 64 B VkAccelerationStructureInstanceKHR + metadata entry every frame,
unconditionally, then flushed the whole high-water capacity (VK_WHOLE_SIZE)
on both buffers. At the millions-of-instances target that copy dominates the
CPU frame, and the whole-buffer flush costs on non-coherent BAR/VRAM.
Add a generation counter so only changed host-authored fields are copied, and
feed the same dirty span into the ranged FlushDevice(offset, bytes) overload:
- RenderingElement3D::hostDataVersion + MarkHostDataDirty() (bumps a global
monotonic counter). TlasWithBuffer::uploadedVersion records, per frame, the
version last copied into each slot. A slot is copied only when its element
advanced past the recorded version; version 0 ("untracked") reads dirty every
frame, so callers that don't opt in keep the prior copy-every-frame behaviour.
Globally-unique versions make this correct under relocation (Remove's
swap-and-pop, and remove+add that nets the same count on the refit path)
without tracking element identity. The reset on every topology change covers
buffer reallocation and the reshuffled element->slot mapping.
- The dirty [first, last] envelope drives both the copy and the flush: a new
VulkanBuffer::FlushDevice(cmd, access, stage, offset, bytes) overload flushes
+ barriers just that span for instanceBuffer, and the ranged
FlushDevice(offset, bytes) for metadataBuffer. When nothing is dirty both are
skipped — the skipped HOST->build barrier only ever ordered host writes, never
the application's compute-written GPU-owned transform (that compute->build
ordering is the caller's, and is unchanged).
Constraint honoured: transformOwnedByGpu transforms are still never host-copied.
The API field/method are mirrored on the WebGPU class for source portability
(the WebGPU build re-uploads its small mirror wholesale and ignores the version).
New test TLASInstanceDirtyTracking drives the real RT device and reads back the
host-mapped buffers to assert: tracked elements upload once then skip until
re-marked, untracked elements always upload, and relocation on the refit path
re-uploads exactly the moved slots — with zero validation-layer errors over the
ranged flush.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The staged branch of VulkanBuffer::UploadDeviceLocal did a full
Create (create+getreq+alloc+map) + DeferredClear (later free) per call.
Mesh::Refit / RecordProceduralBuild hit this every frame for deforming
meshes on no-/small-BAR hardware — a device-memory alloc/free cycle per
mesh per frame, on exactly the hardware the staged path targets.
Replace it with a persistent per-buffer staging ring sized to a
high-water mark, reallocated (via Resize, which defers the outgrown
allocation) only on growth — mirroring the TLAS instance/metadata reuse.
The ring is a per-frame-in-flight ring, not a single shared buffer: the
vkCmdCopyBuffer still reads staging after the call returns, and with
frames pipelined framesInFlight deep, overwriting one shared buffer next
frame would clobber data the previous frame's copy is still reading.
Indexing by frameCounter % framesInFlight gives each in-flight frame its
own slot, reused only after framesInFlight frames elapse — the same
window the #101 deletion queue relies on for GPU completion.
The ring is grown lazily, on first entry into the staged branch, so
ReBAR/UMA hardware (which always takes the direct-write branch) never
constructs a staging allocation it does not use — keeping this a runtime
no-op there and a straight upgrade on non-resizable-BAR machines.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Mesh::Refit re-uploaded the entire vertex array every frame on the
in-place UPDATE path (full host write + flush + barrier on the direct
path; full re-stage + copy on the staged path), even when only a handful
of vertices moved.
Add VulkanBuffer::UploadDeviceLocalRange — a dirty-range counterpart to
UploadDeviceLocal that writes/flushes/stages/copies and barriers only the
half-open element range [offset, offset+count) of an already-allocated
device-local buffer. It picks direct-map vs staged-copy from the memory
type the buffer was actually allocated with (not the sub-range size, which
PreferDirectDeviceWrite would mis-route), and the direct path's ranged
flush is rounded to nonCoherentAtomSize and clamped to the allocation size
(mappedSize is now recorded for every buffer, not just mapped ones).
Add a dirty-range Refit overload taking the full vertex/index arrays plus
a (dirtyVertexOffset, dirtyVertexCount) window. The full-span Refit now
delegates to it with the whole array as the window. On the in-place UPDATE
path only the declared window is uploaded — the rest of the device buffer
retains last refit's positions; when an UPDATE isn't possible it falls
back to the full-span rebuild, which is why the full arrays are still
passed. The WebGPU/DOM backend keeps API symmetry: it has no hardware AS,
so it ignores the window and rebuilds the host BVH from the full geometry.
BLASBuildOptions exercises the dirty-range refit on the direct path, the
staged path, and the count-change rebuild fallback, asserting AS-handle /
blasAddr stability and zero Vulkan validation errors.
Resolves#119
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The inter-pass and acquire/present barriers in the frame loop set both
stage masks to ALL_COMMANDS, and the inter-pass dependency used a
queue-wide VkMemoryBarrier — fully serialising against every pipeline
stage and flushing all caches every frame, when all the next pass needs
is the swapchain image the previous one wrote.
Replace the inter-pass global VkMemoryBarrier with an image memory
barrier scoped to the swapchain image's single colour subresource (as
the intra-pass UI barrier already does), and derive the barrier stage
masks per pass: RenderPass::SwapchainStage() is overridden by UIRenderer
(COMPUTE_SHADER) and RTPass (RAY_TRACING_SHADER), so a compute->compute
edge only serialises COMPUTE while an RT pass pulls in RAY_TRACING — the
acquire/present frame-edge masks use the real union of the frame's
passes (SwapchainStageUnion). The base default and the empty-passes
fallback are the conservative COMPUTE | RAY_TRACING | TRANSFER union, so
a polymorphic or un-overridden pass can only be over- not
under-synchronised.
Adds SwapchainBarrierScope (pure CPU) pinning the per-pass derivation,
the union narrowing, and the inter-pass barrier scope; FrameLoopSync
already drives the real GPU frame loop with validation enabled.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
ImageVulkan kept its host-visible staging `buffer` (sized w*h, persistently
mapped) alive for the whole life of the image and Destroy() never freed it, so
static textures (e.g. Sponza albedo) pinned HOST_VISIBLE / small-BAR memory
forever.
Update now releases the staging via VulkanBuffer::DeferredClear() right after
recording the buffer→image copy — the same fence-keyed deletion queue
(#101/#102) the compressed Mesh path already uses — so it is freed once the
upload submit's frame has cleared. A new `streamed` flag (set by FontAtlas)
keeps the persistent map for images re-uploaded from a CPU-side staging buffer
every frame; its uploads go through UpdateRegion, which never releases. Destroy
now also frees any staging the image still owns, gated on a live handle so a
released static texture can't double-free.
Adds tests/ImageStagingRelease driving the real upload + readback on a headless
device: asserts the staging is enqueued (not pinned), the image reads back
byte-equal (released staging outlived the submit), the entry retires only after
framesInFlight frames, and a streamed image keeps then frees its staging.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
ReclaimDeletions ran an O(n) walk + survivor compaction over the whole
deletion queue every frame. frameCounter is monotonic, so retireAfter is
non-decreasing down the queue and the entries ready to reclaim are always
a contiguous prefix. Switch deletionQueue from std::vector to std::deque
and pop the ready prefix off the front, stopping at the first entry still
in flight — O(ready) instead of O(queue).
Resolves#116
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Issue #110 reported `crafter-build test MeshDecompressStagingRelease` failing
deterministically on the GPU decompress path: the "exactly one allocation
handed to the deletion queue" assertion saw more than one entry.
Root cause (confirmed at Crafter.Graphics-Mesh.cpp:175-176): a static
(allowUpdate=false) Build cannot refit, so it DeferredClear()s its now-dead
per-mesh BLAS scratch in addition to the compressed staging (#67) — two
deferred allocations, not one. The original test built statically yet asserted
size==1, so the scratch was the unaccounted-for second entry. The behaviour
was already corrected on master by PR #109 (the #73 merge, ed9b3f6), which
switched the assertion's build to allowUpdate=true so the scratch is retained.
This hardens that fix: it adds a sibling case that Builds with allowUpdate=false
and asserts the queue holds exactly TWO entries (staging + dead scratch),
builds with zero validation errors, and that both retire on schedule. The count
the #67 assertion relies on is now a named, explicitly-tested quantity rather
than an implicit comment, so a future change to scratch deferral can't silently
shift it back.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Add TRANSFER_SRC to RT geometry usage so device-local geometry (now in
VRAM, no longer host-mappable) stays copyable/inspectable.
- MeshDecompressStagingRelease: read decompressed vertex/index back via a
device->host copy instead of the removed host-mapped .value/FlushHost, and
build the mesh with allowUpdate=true so the retained per-mesh scratch (#66)
doesn't also land in the deletion queue — isolating the assertion to the
released compressed staging (#67).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The TLAS metadata buffer is CPU-written every frame (one userMetadata word
per element in BuildTLAS's copy loop) and read by the ray shaders as a
STORAGE_BUFFER via device address every frame. Allocated
HOST_VISIBLE | HOST_COHERENT (system RAM), every per-frame shader read
traversed PCIe — the same cost the sibling instance buffer paid before #65.
Upgrade the request to HOST_VISIBLE with a best-effort DEVICE_LOCAL preference
(BAR/VRAM), keeping the single persistently-mapped buffer. The CPU's writes are
write-only, sequential, never-read-back — the ideal write-combined BAR load —
and the shader reads now hit local VRAM.
Unlike the instance buffer there is no GPU co-write, so the direct upgrade is
unconditional. Staging is still deliberately avoided: this buffer is rewritten
by the CPU every frame, so a per-frame stage+copy+barrier would defeat the
purpose (#75 caveat 2).
Correctness:
- DEVICE_LOCAL is a preference only (depends on #59): GetMemoryType falls back
to plain HOST_VISIBLE (the previous behaviour) on GPUs without resizable BAR.
- HOST_COHERENT is dropped from the required set — the combined type may not be
coherent. A FlushDevice() is added after the copy loop to make the host writes
available on a non-coherent type; it self-gates to a no-op on a coherent type
(#60), so the previous behaviour is preserved wherever the type ends up
coherent. The metadata buffer is not an AS build input and has no GPU
co-write, so it needs neither the build-stage barrier the instance buffer
takes nor a HOST->shader command barrier: the queue submit already orders host
writes made available before it against every command in the submission,
exactly as it did when this buffer was HOST_COHERENT.
Extends the TLASHighWaterMark hardware test with a metadata-buffer assertion
block mirroring #65's instance-buffer check: the chosen memory type must equal
what GetMemoryType resolves for the HOST_VISIBLE | prefer-DEVICE_LOCAL request,
and must be DEVICE_LOCAL wherever a host-visible device-local type is reachable.
The zero-validation-errors check covers the dropped HOST_COHERENT path.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
vertexBuffer/indexBuffer/aabbBuffer were HOST_VISIBLE (system RAM), so the
BLAS build, every refit, and any hit-shader geometry fetch read them over
PCIe. The usage flags (SHADER_DEVICE_ADDRESS, plus STORAGE on the index
buffer) exist precisely to expose geometry for hit-shader fetch, the dominant
RT shading pattern.
Add VulkanBuffer::UploadDeviceLocal, which places a CPU-written/GPU-read
buffer in device-local memory and picks the upload mechanism at runtime via
the #89 strategy (Device::PreferDirectDeviceWrite):
- ReBAR/UMA: allocate HOST_VISIBLE|DEVICE_LOCAL (DEVICE_LOCAL preferred),
map transiently, memcpy, flush-if-non-coherent. No staging buffer.
- no/small BAR: allocate pure DEVICE_LOCAL + TRANSFER_DST, fill a transient
HOST_VISIBLE staging buffer, vkCmdCopyBuffer, then DeferredClear the
staging buffer onto the fence-keyed deletion queue (#101/#102) so it
outlives the copy submit.
The geometry buffers become non-mapped so the destination is free to be
device-local-only. Same-size re-uploads reuse the allocation, so the device
address stays stable across an in-place AS UPDATE refit. The compressed Build
path now allocates vertex/index as pure DEVICE_LOCAL — the GPU decompressor
fills them directly, no host write.
Tests: BLASBuildOptions asserts device-local placement on the triangle,
procedural, and (budget-forced) staged paths, and exercises the staged copy +
deferred-deletion + in-place refit under GPU-assisted validation with zero
validation errors.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
SaveFrame's readback staging buffer was allocated HOST_VISIBLE |
HOST_COHERENT, whose first matching type is write-combined (uncached) on
most discrete GPUs. The CPU then reads the entire frame back from it —
and reads from write-combined memory are pathologically slow.
Select a HOST_CACHED type instead (preferring a cached-coherent type
when one exists), falling back to any HOST_VISIBLE type if no cached type
is advertised (#59 mechanism). When the chosen type is cached-but-not-
coherent, invalidate the mapped range before the CPU read — the read-path
counterpart of FlushDevice (#60) — gated on the chosen type's actual
coherency, not the requested flags. Whole-range invalidate sidesteps
nonCoherentAtomSize.
Deliberately not routed through the #89 upload-strategy helper, which is
upload-only and steers toward write-combined DEVICE_LOCAL|HOST_VISIBLE —
the worst type for readback.
Verified by driving SaveFrame from the VulkanTriangle example: the
barycentric RGB triangle reads back with correct pixels and channel
order on an RTX 4090.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
