Repository navigation
Conversation
`@kernel` tests `__validindex` in every work-item and runs the body under that mask, even in workgroups that lie entirely inside the ndrange. PoCL vectorizes over the work-items of a workgroup and keeps the mask in the vectorized loop, which made cheap kernel bodies about twice as slow as without it (#845). Emit each region of the body twice: without the mask when the whole workgroup lies inside the ndrange, which is the same for all of its work-items, and masked otherwise. `__fullgroup` decides, and only the PoCL backend enables it; elsewhere it is `false`, and the unmasked copy is removed when compiling. This is the single-launch form of #449. Assisted-by: Claude Code
vchuravy
commented
Oct 6, 2026
Comment on lines
+279
to
+286
| @inline full_group(ctx, launch::Launch, iterspace, ndrange) = | ||
| if builtin(iterspace) && ndrange isa CartesianIndices | ||
| T = index_type(launch) | ||
| groupsize = narrow(T, size(workitems(iterspace))) | ||
| reduce(&, map((g, w, n) -> g * w <= n, group_position(ctx, launch), groupsize, narrow(T, size(ndrange))); init = true) | ||
| else | ||
| false | ||
| end |
Member
Author
There was a problem hiding this comment.
Write this as function full_group
Contributor
Benchmark ResultsShow table
Benchmark PlotsA plot of the benchmark results have been uploaded as an artifact to the workflow run for this PR. |
Member
Author
|
This causes problems for kernels with Each region between Example: a neighbour sum through
Results are still correct (n = 4096 and 4099). I haven't measured what the error paths cost on PoCL. They do cost on GPUs. I checked whether enabling
Kernels without Emitting the whole body twice behind one uniform branch ( So for PoCL I'd limit the duplication to kernels without |
vchuravy
marked this pull request as draft
October 7, 2026 08:16
Member
Author
|
Testing by @maleadt
|
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
1 participant
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.
@kerneltests__validindexin every work-item and runs the body under that mask, also in workgroups that lie entirely inside thendrange. On the CPU backend that costs about 2x for cheap kernel bodies (#845): PoCL vectorizes over the work-items of a workgroup and keeps the mask inside the vectorized loop.This emits each region of the body twice: without the mask when the whole workgroup lies inside the
ndrange, and masked otherwise. The test is the same for all work-items of a workgroup, so PoCL's work-item loop only runs the unmasked copy for full workgroups. It is the single-launch form of #449: it covers padded ranges and tuned workgroup sizes, without a second launch for the remainder (which costs several µs on this backend).__fullgroup(ctx)decides. It isfalseby default, so the unmasked copy is removed when compiling and GPU back-ends are unaffected; the PoCL backend overrides it with__fullgroup_check(ctx). Other backends can opt in the same way if the mask is expensive for them.Alternatives I tried
__validindexreturningfull_group | validindex: depends on LLVM unswitching PoCL's work-item loop. It did for a tuned(128, 32, 1)workgroup (983 → 456 µs) but not for(64, 64, 1), which got 33% slower than main.64case most, but not tuned sizes or padded ranges.Measurements
7-point stencil (
r[I] = x[I±e₁] + x[I±e₂] + x[I±e₃] - 6x[I]) on a 128³ or 130³ interior, Julia 1.13, Ryzen 9 5950X, 1 thread pinned to fixed cores (other jobs were running on the machine), best of 3 interleaved runs of best-of-7, µs per launch:6464(64, 64)(64, 64)Results checked against a host reference for exact and padded ranges.
WaterLily v1.6.1 at 4 threads (Julia 1.13, seconds per 25 steps, measured with the same change before restricting it to PoCL, on top of #844, which doesn't change anything on Julia 1.13):
Compile time (WaterLily's constructor and first step) changed within noise, despite the duplicated body.
Tests
ndrangegives two copies of the body on PoCL, a static one (where every workgroup is full) one. It fails on main.@printcodegen check now uses a staticndrange, since a dynamic one has twoprintfcalls (one per copy). It still checks that a@printlowers to a single call.Closes #845. Related: #449.
🤖 Generated with Claude Code