[maleadt]

Every launch of a kernel with work-group collectives (@groupreduce etc.) asks the back-end whether it can shuffle Float16 and Float64 values, also for kernels on other types. On PoCL, supports_shuffle (and supports_float64) answered that by querying the device's extensions and splitting them into strings each time: about 2.2 µs and 9 KB per query.

This checks the extensions once, when creating the session, next to the other device limits that launches need.

Launching a 64-element Float64 @groupreduce on an M3 Pro goes from 9 µs and 18 KB to 3.1 µs and 64 bytes (the latency/pocl/parallel/launch/groupreduce-f64 benchmark). The remaining 64 bytes are from the dynamic dispatch on the sub-group capabilities in launch_kernel.

[github-actions]

Benchmark results

Compared with d2ada37c6, the latest measured ancestor of the merge base (1 commit before it).
The environment differs from the baseline's: Parsers 3.0.1 → 3.0.2.

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	−0.7%	0	0	32
Latency	−7.8%	0	0	13
Applications	+5.7%	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	284 µs	284 µs	−0.0%	noise unknown
julia/serial/narrow/masked-reference	702 µs	700 µs	−0.3%	noise unknown
pocl/parallel/atomics/histogram	5.16 ms	5.15 ms	−0.1%	noise unknown
pocl/parallel/groupreduce/add-f64	1.66 ms	1.64 ms	−1.3%	noise unknown
pocl/parallel/groupreduce/fallback	18.1 ms	18.1 ms	−0.1%	noise unknown
pocl/parallel/groupscan/add-i32	2.53 ms	2.52 ms	−0.5%	noise unknown
pocl/parallel/indexing/cartesian3d	84.8 µs	79.3 µs	−6.5%	noise unknown
pocl/parallel/indexing/linear3d	84.4 µs	77.8 µs	−7.8%	noise unknown
pocl/parallel/indexing/ntuple3d	81.4 µs	80.1 µs	−1.6%	noise unknown
pocl/parallel/indexing/ntuple4d	164 µs	163 µs	−0.6%	noise unknown
pocl/parallel/localmem/transpose-naive	459 µs	454 µs	−1.1%	noise unknown
pocl/parallel/localmem/transpose-tiled	101 µs	104 µs	+3.3%	noise unknown
pocl/parallel/private/across-barrier	164 µs	167 µs	+2.2%	noise unknown
pocl/parallel/scheduling/default	181 µs	181 µs	−0.0%	noise unknown
pocl/parallel/scheduling/explicit	181 µs	181 µs	−0.0%	noise unknown
pocl/parallel/stream/copy/cache	6.06 µs	5.68 µs	−6.3%	noise unknown
pocl/parallel/stream/copy/dram	6.32 ms	6.55 ms	+3.6%	noise unknown
pocl/parallel/stream/triad/cache	6.90 µs	6.60 µs	−4.4%	noise unknown
pocl/parallel/stream/triad/dram	8.94 ms	9.56 ms	+7.0%	noise unknown
pocl/parallel/util/copyto-small	4.59 µs	4.60 µs	+0.4%	noise unknown
pocl/parallel/util/zeros-large	998 µs	1.01 ms	+1.4%	noise unknown
pocl/parallel/util/zeros-small	665 ns	633 ns	−4.8%	noise unknown
pocl/serial/arithmetic/contract	1.91 ms	1.91 ms	−0.1%	noise unknown
pocl/serial/arithmetic/muladd	759 µs	766 µs	+1.0%	noise unknown
pocl/serial/const/marked	728 µs	730 µs	+0.3%	noise unknown
pocl/serial/const/unmarked	915 µs	918 µs	+0.4%	noise unknown
pocl/serial/math/exp-f64	147 µs	146 µs	−0.6%	noise unknown
pocl/serial/math/sin-f64	202 µs	193 µs	−4.4%	noise unknown
pocl/serial/narrow/masked	854 µs	829 µs	−2.9%	noise unknown
pocl/serial/narrow/unmasked	765 µs	782 µs	+2.3%	noise unknown
pocl/serial/stream/triad/dram	8.63 ms	8.67 ms	+0.4%	noise unknown
Latency
host/import/ka	506 ms	503 ms	−0.6%	noise unknown
host/precompile/ka	44.9 s	44.8 s	−0.2%	noise unknown
host/startup	133 ms	131 ms	−2.0%	noise unknown
pocl/parallel/launch/dynamic-default	4.39 µs	4.39 µs	+0.0%	noise unknown
pocl/parallel/launch/groupreduce-f64	18.5 µs	6.72 µs	−63.6%	noise unknown
pocl/parallel/launch/many-args	12.0 µs	12.4 µs	+2.8%	noise unknown
pocl/parallel/launch/partial-workgroup	4.18 µs	4.03 µs	−3.6%	noise unknown
pocl/serial/compile/groupreduce	2.06 s	2.02 s	−2.0%	noise unknown
pocl/serial/compile/localmem-barrier	808 ms	805 ms	−0.5%	noise unknown
pocl/serial/compile/new-kernel	118 ms	118 ms	−0.5%	noise unknown
pocl/serial/compile/new-workgroupsize	13.6 ms	13.5 ms	−0.5%	noise unknown
pocl/serial/first-kernel/cold-cache	2.02 s	2.03 s	+0.4%	noise unknown
pocl/serial/first-kernel/warm-cache	360 ms	366 ms	+1.8%	noise unknown
Applications
pocl/parallel/dgsem/fluxdiff-node	291 µs	293 µs	+0.7%	noise unknown
pocl/parallel/dgsem/weak-element	160 µs	212 µs	+32.9%	noise unknown
pocl/parallel/dgsem/weak-node-localmem	118 µs	118 µs	−0.1%	noise unknown
pocl/parallel/stencil/timestep	7.91 ms	7.74 ms	−2.1%	noise unknown
pocl/serial/stencil/timestep	11.3 ms	11.3 ms	+0.6%	noise unknown

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

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 83.18%. Comparing base (d2ada37) to head (13660f6).
⚠️ Report is 2 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #934      +/-   ##
==========================================
+ Coverage   82.87%   83.18%   +0.30%     
==========================================
  Files          31       31              
  Lines        2564     2569       +5     
==========================================
+ Hits         2125     2137      +12     
+ Misses        439      432       -7     

☔ 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.