Make host Tiled broadcast shape-correct#264
Merged
Merged
Conversation
maleadt
force-pushed
the
ao/broadcast-shape-fix
branch
from
5daeebd to
0bd80d1
Compare
Host-level Tiled broadcasts previously loaded every array leaf with the
destination's block index and tile shape, silently producing garbage for
any shape-mismatched input, and never validated axes at all. Fix:
- Validate via Broadcast.check_broadcast_axes: Base-illegal shapes now
throw DimensionMismatch instead of computing garbage.
- Base-legal size-1 (and missing trailing) dims now work: such leaves are
wrapped in BroadcastLeaf{P}, load a 1-wide slice at index 1 in flagged
dims, and are expanded by the device-side tile broadcast.
- Scalar-only RHS (Tiled(C) .= 0) now fills the destination instead of
MethodError-ing on convert(Tile{T}, ::Number).
- Empty destinations are a no-op instead of launching a 0-block grid.
- ct.@. in-place assignment returns the original destination array.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
0-dim destinations previously died with an IR error; they now run as a 1-element 1-D broadcast (0-dim leaves are reshaped to 1-element vectors, since the kernel path assumes rank >= 1). CPU Array leaves previously compiled and crashed with an illegal memory access on the device (poisoning the CUDA context), and ranges failed with a cryptic pointer MethodError; both now throw a clear ArgumentError on the host before launch. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
maleadt
force-pushed
the
ao/broadcast-shape-fix
branch
from
0bd80d1 to
c336349
Compare
Align the kernel-wrapper machinery with Base.Broadcast's vocabulary and
trim the helper zoo introduced by the shape fixes:
- Replace BroadcastLeaf{P} with Extruded{K}, mirroring
Base.Broadcast.Extruded but with the keep-mask in the type domain
(it determines the static tile shape). Every array leaf is wrapped,
so the separate exact-shape load path disappears.
- Replace _to_tiled_bc with preprocess, the counterpart of
Base.Broadcast.preprocess.
- Drop _bc_result: scalars now convert through a new
convert(Tile{T}, ::Number) method, and the store site uniformly
expands with broadcast_to (a codegen no-op for full-shaped tiles).
- Drop _eval_bc_args in favor of a tuple map.
- Move host-memory rejection from a parent-walking blocklist into
TileArray construction itself: pointer(arr) returning a plain Ptr
means host memory. This also covers launch() arguments and
mapreduce, not just broadcast.
- Drop the macro return-value rewrite: materialize! already returns
the parent array since #258.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
maleadt
force-pushed
the
ao/broadcast-shape-fix
branch
from
c336349 to
976e8b9
Compare
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.
Host-level
Tiledbroadcasts loaded every array leaf with the destination's block index and tile shape, and never validated axes — so shape-mismatched inputs (both Base-legal(1,N)vs(M,N)expansion and outright wrong sizes) silently produced garbage.Changes
_tiled_broadcast!now callsBroadcast.check_broadcast_axes, so Base-illegal shapes throwDimensionMismatchinstead of computing garbage.BroadcastLeaf{P}(P = static mask of expanded dims), load a 1-wide slice at index 1 in those dims, and are expanded by the existing device-side tile broadcast (broadcast_shape/broadcast_to). Exact-shape leaves keep the previous fast path unchanged.Tiled(C) .= 0used toMethodErroronconvert(Tile{T}, ::Number); the kernel result is now routed through_bc_result, which splats scalars viabroadcast_to(Tile(T(x)), tile_size)(and also expands sub-shaped result tiles when every leaf is size-1 in a dim).ct.@.return value: in-place assignment (ct.@. C = A + B) now evaluates to the original destination arrayCrather than theTiledwrapper.Tests
9 new testsets in
test/host/broadcast.jl: mismatch throws (1D/2D, nested), row/column size-1 expansion, expansion-only RHS, rank expansion (vector + matrix), allocating form with expansion, scalar fill, empty no-op, and macro return identity. Follow-up commit: 0-dim broadcasts now work (delegated to a 1-element 1-D launch), and host-side leaves (CPUArrays — previously an illegal memory access that poisoned the CUDA context — and ranges) are rejected with a clearArgumentErrorbefore launch. Full host broadcast suite passes on GPU: 37/37.(Supersedes #261/#263, closed by GitHub during head-branch renames.)
🤖 Generated with Claude Code