NanoVDB: type the CUDA scratch buffers as cuda::Buffer<T> (CUDA) - #2322
Open
harrism wants to merge 2 commits into
Open
NanoVDB: type the CUDA scratch buffers as cuda::Buffer<T> (CUDA)#2322harrism wants to merge 2 commits into
harrism wants to merge 2 commits into
Conversation
Mask became trivially copyable in AcademySoftwareFoundation#2310, so the scratch that exists only to hold arrays of masks no longer needs to be a byte buffer reinterpreted at every use. TopologyBuilder's upper and lower mask members become Buffer<Mask<5>> and Buffer<Mask<4>>, and deviceUpperMasks/deviceLowerMasks return typed pointers, which the seven morphology functors and MeshToGrid now take directly - the only cast left is the array-shape view of the flat lower masks as one Mask<5>::SIZE row per upper node. MeshToGrid's two retain-mask buffers and its per-pass hit-mask buffer, the dilation example's leaf masks, and the dilation test's mask buffer (previously on the deprecated dual DeviceBuffer) are Buffer<Mask<3>>. Allocation sizes are element counts and memsets use size_bytes(). Closes AcademySoftwareFoundation#2312. Signed-off-by: Mark Harris <mharris@nvidia.com>
…oGrid and SignedFloodFill (CUDA) The same pattern as the mask buffers, applied to every other scratch allocation that holds a single element type for its whole life: TopologyBuilder's upper/lower/leaf/voxel offsets, lower/leaf parents, the three count arrays of its enumerate pass and its device Data staging; MeshToGrid's transformed triangles, box-triangle pairs, unique root origins, and its nine local key/count/offset/pair buffers; SignedFloodFill's root-child array. Allocations are sized in elements, the reinterpret_casts at every use are gone, and the accessors return typed pointers. The lower and leaf offsets, which the node-processing kernels read as one row of Mask<5>::SIZE per upper node, expose that row shape from accessors beside the allocation, as deviceLowerMasks does. TopologyBuilder's root staging stays a byte buffer (its size is RootT::memUsage(tileCount), not an element count), which is now the only reinterpretation its alignment static_assert guards. Addresses the sweep portion of AcademySoftwareFoundation#2327. Signed-off-by: Mark Harris <mharris@nvidia.com>
harrism
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #2312. Also lands the sweep portion of #2327 (everything but VoxelBlockManagerHandle, which stays open there pending coordination with #2291).
What
Maskbecame trivially copyable in #2310, which removed the last reason for the CUDA tools to hold typed arrays in byte buffers reinterpreted at every use. Two commits:1. The Mask sidecars (#2312) —
TopologyBuilder's upper/lower masks becomecuda::Buffer<Mask<5>>/cuda::Buffer<Mask<4>>;deviceUpperMasks()returnsMask<5>*anddeviceLowerMasks()returns the row-shapedMask<4> (*)[Mask<5>::SIZE], so the stride is fixed once beside the allocation and the seven morphology functors and every caller take typed pointers (eight casts gone). MeshToGrid's two retain-mask buffers and hit-mask buffer, the dilation example, and the dilation test (off the deprecated dualDeviceBuffer) becomecuda::Buffer<Mask<3>>.2. Every remaining single-type scratch buffer — TopologyBuilder's upper/lower/leaf/voxel offsets, lower/leaf parents, the three enumerate-pass count arrays and its device
Datastaging; MeshToGrid's transformed triangles, box-triangle pairs, unique root origins and nine local key/count/offset/pair buffers (its byteScratchTalias is gone); SignedFloodFill's root-child array. The lower/leaf offsets expose their per-upper-node row shape vialowerOffsetRows()/leafOffsetRows(), same pattern as the masks. TopologyBuilder's variable-length root staging stays a byte buffer (its size isRootT::memUsage(tileCount), not an element count) and is now the only thing its alignmentstatic_assertguards — strengthened toNANOVDB_DATA_ALIGNMENT, which is what the root reinterpret actually needs. MeshToGrid gains an equivalent assert for its 16-byte-aligned pair struct, which previously had none.Allocations are sized in elements and memsets use
size_bytes(), so the byte arithmetic disappears andBuffer's count×sizeof overflow check applies where hand-multiplied sizes had none.Verification
Full build clean under
NANOVDB_CUDA_WERROR; 156 CPU / 65 CUDA / 48 buffer / 19 memory-resource tests pass. Explore self-review and Codex review passes on both commits; every element type was checked againstBuffer'sis_trivially_copyablerequirement (includingRootChild, whose converting constructor leaves its copy operations trivial).