[maleadt]

Remove hipfunction's Function constraint so callable objects (structs with call methods) can be launched as kernels. Previously, this failed with a MethodError:

struct StoreValue
    val::Float64
end
(f::StoreValue)(a) = (a[1] = f.val; return)

@roc StoreValue(2.0)(a)  # MethodError: no method matching hipfunction(::StoreValue, ...)

JuliaGPU/KernelAbstractions.jl#931 launches a callable object in its "partition for the kernel function" test, which fails on AMDGPU. Nothing depends on the constraint, and CUDA.jl's cufunction has none.

@roc converts callable objects like closures, passing their fields as the kernel's first argument. Compiled kernels are cached by type; the added test checks that a new instance reuses the kernel with its own state.

Tests: device/launch (fails without the change, passes with it); kernelabstractions_tests, kernelinterface_tests and device/launch against KernelAbstractions#931 (a2dd21b3) on gfx1036.

[luraess]

cscs-ci run MI250,MI300

[github-actions]

Benchmark results

Compared with 3e7016582, the merge base.
The environment differs from the baseline's: AMDGPUExt 0.5.9 → 0.10.3, AcceleratedKernels 0.5.9 → 0.5.10, Atomix 1.6.0 → 1.7.0, BFloat16sExt 10.4.1 → 10.4.2, JSON 1.12.2 → 1.13.0, LLVM 10.4.1 → 10.4.2, Parsers 3.0.1 → 3.0.2, StructUtils 2.11.1 → 2.11.2, StructUtilsStaticArraysCoreExt 2.11.1 → 2.11.2, StructUtilsTablesExt 2.11.1 → 2.11.2, pocl_standalone_jll 7.2.3+1 → 7.2.3+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	−0.4%	6	9	69
Latency	−3.5%	0	1	3
Applications	−5.7%	0	2	14

Benchmark	Baseline	This PR	Change
Performance
array/copyto!/gpu_to_gpu	35.1 µs	57.2 µs	+62.8%	🔴 slower
kernel/rand	48.3 µs	72.1 µs	+49.2%	🔴 slower
array/broadcast	50.6 µs	41.1 µs	−18.8%	🟢 faster
array/reductions/mapreduce/Float32/dims=1	55.5 µs	45.5 µs	−18.1%	🟢 faster
array/reductions/mapreduce/Float32/dims=1L	59.7 µs	49.1 µs	−17.9%	🟢 faster
array/reductions/reduce/Float32/dims=2	59.4 µs	49.0 µs	−17.5%	🟢 faster
array/accumulate/Int64/dims=2	91.0 µs	109 µs	+19.6%	🔴 slower
array/reductions/mapreduce/Float32/1d	71.0 µs	60.7 µs	−14.5%	🟢 faster
array/reductions/reduce/Float32/dims=1L	50.4 µs	58.3 µs	+15.6%	🔴 slower
array/sorting/1d	343 µs	297 µs	−13.4%	🟢 faster
array/reverse/1d	37.3 µs	42.6 µs	+14.1%	🔴 slower
array/accumulate/Float32/dims=2L	124 µs	111 µs	−10.5%	🟢 faster
array/accumulate/Float32/dims=1	64.2 µs	57.8 µs	−10.0%	🟢 faster
array/reverse/2dL	80.5 µs	73.8 µs	−8.3%	🟢 faster
array/iteration/findfirst/bool	54.9 µs	58.4 µs	+6.4%	🔴 slower
Latency
import	1.37 s	1.30 s	−5.0%	🟢 faster
Applications
gemm/tiled	2.75 ms	1.86 ms	−32.3%	🟢 faster
gemm/tiled_unbounded	2.79 ms	1.89 ms	−32.2%	🟢 faster

Benchmark	Baseline allocations	This PR
Applications
prefix_sum	none	2 (84 B)	🔴 allocates

All benchmarks

Benchmark	Baseline	This PR	Change
Performance
amdgpu/synchronization/context/device	1.79 µs	1.74 µs	−3.1%
amdgpu/synchronization/stream/blocking	228 ns	230 ns	+1.1%
amdgpu/synchronization/stream/nonblocking	350 ns	340 ns	−2.9%
array/accumulate/Float32/1d	82.6 µs	82.7 µs	+0.1%
array/accumulate/Float32/dims=1	64.2 µs	57.8 µs	−10.0%	🟢 faster
array/accumulate/Float32/dims=1L	105 µs	107 µs	+1.9%
array/accumulate/Float32/dims=2	105 µs	105 µs	+0.5%
array/accumulate/Float32/dims=2L	124 µs	111 µs	−10.5%	🟢 faster
array/accumulate/Int64/1d	104 µs	104 µs	+0.1%
array/accumulate/Int64/dims=1	75.1 µs	76.6 µs	+2.1%
array/accumulate/Int64/dims=1L	111 µs	117 µs	+6.1%
array/accumulate/Int64/dims=2	91.0 µs	109 µs	+19.6%	🔴 slower
array/accumulate/Int64/dims=2L	157 µs	164 µs	+4.5%
array/broadcast	50.6 µs	41.1 µs	−18.8%	🟢 faster
array/construct	2.29 µs	2.33 µs	+1.6%
array/copy	40.3 µs	39.8 µs	−1.2%
array/copyto!/cpu_to_gpu	85.9 µs	85.9 µs	+0.1%
array/copyto!/gpu_to_cpu	85.8 µs	86.0 µs	+0.2%
array/copyto!/gpu_to_gpu	35.1 µs	57.2 µs	+62.8%	🔴 slower
array/iteration/findall/bool	94.7 µs	89.9 µs	−5.1%
array/iteration/findall/int	107 µs	103 µs	−3.1%
array/iteration/findfirst/bool	54.9 µs	58.4 µs	+6.4%	🔴 slower
array/iteration/findfirst/int	56.6 µs	56.4 µs	−0.3%
array/iteration/findmin/1d	95.0 µs	97.3 µs	+2.4%
array/iteration/findmin/2d	79.3 µs	79.4 µs	+0.2%
array/iteration/logical	91.0 µs	85.8 µs	−5.8%
array/iteration/scalar	307 µs	310 µs	+0.7%
array/permutedims/2d	62.1 µs	62.3 µs	+0.3%
array/permutedims/3d	49.4 µs	48.0 µs	−2.9%
array/permutedims/4d	54.1 µs	54.2 µs	+0.0%
array/random/rand!/Float32	50.9 µs	58.8 µs	+15.6%
array/random/rand!/Int64	59.8 µs	49.2 µs	−17.8%
array/random/rand/Float32	46.9 µs	46.8 µs	−0.2%
array/random/rand/Int64	46.4 µs	46.1 µs	−0.5%
array/random/randn!/Float32	82.6 µs	60.1 µs	−27.2%
array/random/randn/Float32	70.5 µs	70.8 µs	+0.5%
array/reductions/mapreduce/Float32/1d	71.0 µs	60.7 µs	−14.5%	🟢 faster
array/reductions/mapreduce/Float32/dims=1	55.5 µs	45.5 µs	−18.1%	🟢 faster
array/reductions/mapreduce/Float32/dims=1L	59.7 µs	49.1 µs	−17.9%	🟢 faster
array/reductions/mapreduce/Float32/dims=2	59.6 µs	57.6 µs	−3.3%
array/reductions/mapreduce/Float32/dims=2L	60.7 µs	57.3 µs	−5.5%
array/reductions/mapreduce/Int64/1d	71.7 µs	72.1 µs	+0.6%
array/reductions/mapreduce/Int64/dims=1	51.6 µs	50.6 µs	−2.0%
array/reductions/mapreduce/Int64/dims=1L	70.9 µs	70.1 µs	−1.2%
array/reductions/mapreduce/Int64/dims=2	60.7 µs	58.5 µs	−3.7%
array/reductions/mapreduce/Int64/dims=2L	78.4 µs	76.2 µs	−2.8%
array/reductions/reduce/Float32/1d	70.9 µs	71.2 µs	+0.5%
array/reductions/reduce/Float32/dims=1	55.7 µs	54.6 µs	−1.9%
array/reductions/reduce/Float32/dims=1L	50.4 µs	58.3 µs	+15.6%	🔴 slower
array/reductions/reduce/Float32/dims=2	59.4 µs	49.0 µs	−17.5%	🟢 faster
array/reductions/reduce/Float32/dims=2L	72.5 µs	70.1 µs	−3.3%
array/reductions/reduce/Int64/1d	71.4 µs	71.7 µs	+0.4%
array/reductions/reduce/Int64/dims=1	51.6 µs	50.5 µs	−2.2%
array/reductions/reduce/Int64/dims=1L	70.9 µs	69.3 µs	−2.2%
array/reductions/reduce/Int64/dims=2	60.6 µs	58.2 µs	−3.9%
array/reductions/reduce/Int64/dims=2L	78.2 µs	76.0 µs	−2.8%
array/reverse/1d	37.3 µs	42.6 µs	+14.1%	🔴 slower
array/reverse/1d_inplace	40.3 µs	50.6 µs	+25.8%
array/reverse/1dL	60.5 µs	68.5 µs	+13.3%
array/reverse/1dL_inplace	78.8 µs	79.4 µs	+0.8%
array/reverse/2d	47.8 µs	47.5 µs	−0.6%
array/reverse/2d_inplace	56.1 µs	65.5 µs	+16.6%
array/reverse/2dL	80.5 µs	73.8 µs	−8.3%	🟢 faster
array/reverse/2dL_inplace	90.5 µs	76.6 µs	−15.4%
array/sorting/1d	343 µs	297 µs	−13.4%	🟢 faster
kernel/indexing	32.1 µs	30.3 µs	−5.4%
kernel/indexing_checked	37.0 µs	44.7 µs	+20.7%
kernel/launch	900 ns	903 ns	+0.3%
kernel/rand	48.3 µs	72.1 µs	+49.2%	🔴 slower
Latency
import	1.37 s	1.30 s	−5.0%	🟢 faster
precompile	29.0 s	27.8 s	−4.1%
ttfp	1.77 s	1.74 s	−1.5%
Applications
bitonic_sort	1.11 ms	1.09 ms	−2.4%
byval/reference	39.2 µs	40.8 µs	+4.0%
byval/slices=1	41.1 µs	41.1 µs	−0.0%
byval/slices=2	158 µs	145 µs	−8.2%
byval/slices=3	233 µs	236 µs	+1.2%
convolution	76.0 µs	76.3 µs	+0.4%
floyd_warshall	9.06 ms	9.00 ms	−0.7%
gemm/tiled	2.75 ms	1.86 ms	−32.3%	🟢 faster
gemm/tiled_unbounded	2.79 ms	1.89 ms	−32.2%	🟢 faster
histogram	782 µs	795 µs	+1.7%
prefix_sum	194 µs	194 µs	+0.3%
stencil/diffusion3d	1.57 ms	1.57 ms	+0.2%
stencil/diffusion3d_checked	1.61 ms	1.61 ms	−0.0%
volumerhs	4.86 ms	4.85 ms	−0.2%

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