[maleadt]

The benchmark harness runs every process with one Julia thread, and only sets PoCL's worker count per mode. A backend that runs work-groups on Julia's threads therefore runs the parallel benchmarks on one thread. That is why #906's benchmark comment shows 3–8× slowdowns: the CI machine has 8 cores, and main used 8 PoCL workers.

This gives each process as many Julia threads as it has workers, with no interactive thread (JULIA_NUM_THREADS="$threads,0"). Serial and latency processes still get one thread, as before.

Minimum times on a 9950X (8 cores via taskset, JULIA_KA_CPU_THREADS=8):

	main -t1	main -t8,0	#906 -t1	#906 -t8,0
groupreduce/add-f64	642 µs	656 µs	5.07 ms	666 µs
groupscan/add-i32	1.18 ms	811 µs	6.13 ms	795 µs
indexing/ntuple4d	134 µs	119 µs	653 µs	84 µs
transpose-naive	370 µs	338 µs	2.55 ms	331 µs
private/across-barrier	78 µs	53 µs	373 µs	51 µs
scheduling/default	112 µs	110 µs	404 µs	53 µs

Main also changes with the extra threads (groupscan and across-barrier get faster), so this should merge before #906 is compared again.

[github-actions]

Benchmark results

Compared with 14002d752, the merge base.
Times are the minimum time of every benchmark. A change is unusual when it exceeds both three times the benchmark's typical change between consecutive runs on the default branch and a minimum per category (performance 3%, latency 5%, applications 3%, other 3%). A benchmark that starts or stops allocating is listed however few the allocations. A single run can still be a fluke, so rerun before drawing conclusions.

	Geometric mean	Unusual: slower	Unusual: faster	Compared
Performance	−0.1%	0	0	32
Latency	−1.1%	0	0	13
Applications	−14.3%	0	0	5

No unusual changes.

All benchmarks

Benchmark	Baseline	This PR	Change
Performance
julia/serial/arithmetic/muladd-reference	655 µs	655 µs	+0.0%	noise unknown
julia/serial/math/exp-f64-reference	383 µs	291 µs	−23.9%	noise unknown
julia/serial/narrow/masked-reference	678 µs	702 µs	+3.6%	noise unknown
pocl/parallel/atomics/histogram	5.14 ms	5.13 ms	−0.2%	noise unknown
pocl/parallel/groupreduce/add-f64	1.65 ms	1.65 ms	−0.1%	noise unknown
pocl/parallel/groupreduce/fallback	18.1 ms	18.9 ms	+4.0%	noise unknown
pocl/parallel/groupscan/add-i32	2.53 ms	2.52 ms	−0.1%	noise unknown
pocl/parallel/indexing/cartesian3d	84.3 µs	80.5 µs	−4.5%	noise unknown
pocl/parallel/indexing/linear3d	80.8 µs	77.9 µs	−3.5%	noise unknown
pocl/parallel/indexing/ntuple3d	84.2 µs	80.4 µs	−4.5%	noise unknown
pocl/parallel/indexing/ntuple4d	161 µs	164 µs	+1.6%	noise unknown
pocl/parallel/localmem/transpose-naive	456 µs	460 µs	+0.9%	noise unknown
pocl/parallel/localmem/transpose-tiled	108 µs	105 µs	−2.7%	noise unknown
pocl/parallel/private/across-barrier	165 µs	164 µs	−0.2%	noise unknown
pocl/parallel/scheduling/default	182 µs	182 µs	+0.1%	noise unknown
pocl/parallel/scheduling/explicit	182 µs	182 µs	−0.0%	noise unknown
pocl/parallel/stream/copy/cache	682 ns	688 ns	+0.9%	noise unknown
pocl/parallel/stream/copy/dram	6.00 ms	6.32 ms	+5.4%	noise unknown
pocl/parallel/stream/triad/cache	6.99 µs	7.19 µs	+2.9%	noise unknown
pocl/parallel/stream/triad/dram	9.37 ms	9.31 ms	−0.5%	noise unknown
pocl/parallel/util/copyto-small	23.0 ns	26.0 ns	+13.0%	noise unknown
pocl/parallel/util/zeros-large	999 µs	1.00 ms	+0.2%	noise unknown
pocl/parallel/util/zeros-small	606 ns	626 ns	+3.3%	noise unknown
pocl/serial/arithmetic/contract	1.90 ms	1.91 ms	+0.5%	noise unknown
pocl/serial/arithmetic/muladd	768 µs	770 µs	+0.2%	noise unknown
pocl/serial/const/marked	732 µs	712 µs	−2.7%	noise unknown
pocl/serial/const/unmarked	914 µs	924 µs	+1.1%	noise unknown
pocl/serial/math/exp-f64	138 µs	138 µs	−0.0%	noise unknown
pocl/serial/math/sin-f64	190 µs	192 µs	+1.0%	noise unknown
pocl/serial/narrow/masked	828 µs	846 µs	+2.2%	noise unknown
pocl/serial/narrow/unmasked	812 µs	825 µs	+1.6%	noise unknown
pocl/serial/stream/triad/dram	8.59 ms	8.66 ms	+0.8%	noise unknown
Latency
host/import/ka	504 ms	510 ms	+1.1%	noise unknown
host/precompile/ka	44.7 s	44.9 s	+0.5%	noise unknown
host/startup	131 ms	129 ms	−1.1%	noise unknown
pocl/parallel/launch/dynamic-default	4.53 µs	4.41 µs	−2.6%	noise unknown
pocl/parallel/launch/groupreduce-f64	6.72 µs	6.70 µs	−0.3%	noise unknown
pocl/parallel/launch/many-args	12.7 µs	11.1 µs	−12.9%	noise unknown
pocl/parallel/launch/partial-workgroup	4.22 µs	4.13 µs	−2.1%	noise unknown
pocl/serial/compile/groupreduce	1.85 s	1.87 s	+0.7%	noise unknown
pocl/serial/compile/localmem-barrier	720 ms	724 ms	+0.7%	noise unknown
pocl/serial/compile/new-kernel	109 ms	111 ms	+1.4%	noise unknown
pocl/serial/compile/new-workgroupsize	13.7 ms	13.9 ms	+1.2%	noise unknown
pocl/serial/first-kernel/cold-cache	2.05 s	2.06 s	+0.4%	noise unknown
pocl/serial/first-kernel/warm-cache	386 ms	385 ms	−0.2%	noise unknown
Applications
pocl/parallel/dgsem/fluxdiff-node	447 µs	287 µs	−35.9%	noise unknown
pocl/parallel/dgsem/weak-element	212 µs	159 µs	−24.9%	noise unknown
pocl/parallel/dgsem/weak-node-localmem	147 µs	146 µs	−0.8%	noise unknown
pocl/parallel/stencil/timestep	7.91 ms	7.78 ms	−1.6%	noise unknown
pocl/serial/stencil/timestep	11.4 ms	11.2 ms	−1.6%	noise unknown

History: https://juliagpu.org/perf/?pkg=kernelabstractions

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 82.46%. Comparing base (14002d7) to head (d663fa1).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #942      +/-   ##
==========================================
- Coverage   83.17%   82.46%   -0.72%     
==========================================
  Files          31       31              
  Lines        2567     2520      -47     
==========================================
- Hits         2135     2078      -57     
- Misses        432      442      +10     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.