Tests for host interface functions

[christiangnrd]

supports_atomics in Metal.jl is broken and I realized we should probably test the interface definitions

All Claude and I reviewed

[github-actions]

Benchmark Results

Show table

	main	e540c24...	main / e540c24...
saxpy/default/Float32/1024	29.6 ± 20 μs	0.0327 ± 0.022 ms	0.904 ± 0.87
saxpy/default/Float32/1048576	0.34 ± 0.035 ms	0.349 ± 0.042 ms	0.976 ± 0.15
saxpy/default/Float32/16384	0.0435 ± 0.022 ms	0.0434 ± 0.021 ms	1 ± 0.7
saxpy/default/Float32/2048	0.0449 ± 0.02 ms	0.0485 ± 0.021 ms	0.927 ± 0.58
saxpy/default/Float32/256	29.4 ± 18 μs	26.6 ± 9 μs	1.1 ± 0.78
saxpy/default/Float32/262144	0.112 ± 0.017 ms	0.11 ± 0.016 ms	1.02 ± 0.22
saxpy/default/Float32/32768	0.0459 ± 0.022 ms	0.0476 ± 0.022 ms	0.963 ± 0.64
saxpy/default/Float32/4096	0.0534 ± 0.024 ms	0.052 ± 0.023 ms	1.03 ± 0.64
saxpy/default/Float32/512	26.3 ± 11 μs	30.4 ± 22 μs	0.867 ± 0.73
saxpy/default/Float32/64	25.6 ± 5.9 μs	26.4 ± 6.9 μs	0.97 ± 0.34
saxpy/default/Float32/65536	0.0567 ± 0.022 ms	0.06 ± 0.023 ms	0.944 ± 0.51
saxpy/default/Float64/1024	0.0363 ± 0.022 ms	0.035 ± 0.024 ms	1.04 ± 0.95
saxpy/default/Float64/1048576	0.615 ± 0.069 ms	0.635 ± 0.073 ms	0.969 ± 0.16
saxpy/default/Float64/16384	0.0435 ± 0.022 ms	0.0458 ± 0.023 ms	0.95 ± 0.68
saxpy/default/Float64/2048	0.0478 ± 0.022 ms	0.0472 ± 0.023 ms	1.01 ± 0.68
saxpy/default/Float64/256	25.4 ± 7.5 μs	26.1 ± 8.7 μs	0.972 ± 0.43
saxpy/default/Float64/262144	0.182 ± 0.025 ms	0.185 ± 0.028 ms	0.985 ± 0.2
saxpy/default/Float64/32768	0.0513 ± 0.023 ms	0.0527 ± 0.023 ms	0.974 ± 0.62
saxpy/default/Float64/4096	0.056 ± 0.022 ms	0.0528 ± 0.021 ms	1.06 ± 0.6
saxpy/default/Float64/512	26 ± 10 μs	26.6 ± 12 μs	0.977 ± 0.58
saxpy/default/Float64/64	26 ± 7.5 μs	26.1 ± 5.9 μs	0.997 ± 0.37
saxpy/default/Float64/65536	0.0713 ± 0.025 ms	0.0699 ± 0.023 ms	1.02 ± 0.49
saxpy/static workgroup=(1024,)/Float32/1024	31.6 ± 17 μs	0.032 ± 0.019 ms	0.985 ± 0.78
saxpy/static workgroup=(1024,)/Float32/1048576	0.431 ± 0.073 ms	0.403 ± 0.071 ms	1.07 ± 0.26
saxpy/static workgroup=(1024,)/Float32/16384	0.0442 ± 0.021 ms	0.0453 ± 0.021 ms	0.976 ± 0.64
saxpy/static workgroup=(1024,)/Float32/2048	0.0429 ± 0.019 ms	0.0442 ± 0.019 ms	0.971 ± 0.6
saxpy/static workgroup=(1024,)/Float32/256	31.4 ± 15 μs	30.8 ± 18 μs	1.02 ± 0.76
saxpy/static workgroup=(1024,)/Float32/262144	0.139 ± 0.016 ms	0.152 ± 0.015 ms	0.916 ± 0.14
saxpy/static workgroup=(1024,)/Float32/32768	0.0505 ± 0.02 ms	0.0505 ± 0.02 ms	0.999 ± 0.56
saxpy/static workgroup=(1024,)/Float32/4096	0.0505 ± 0.017 ms	0.0504 ± 0.017 ms	1 ± 0.48
saxpy/static workgroup=(1024,)/Float32/512	0.0329 ± 0.018 ms	31.6 ± 16 μs	1.04 ± 0.77
saxpy/static workgroup=(1024,)/Float32/64	30.8 ± 14 μs	29.9 ± 15 μs	1.03 ± 0.69
saxpy/static workgroup=(1024,)/Float32/65536	0.062 ± 0.02 ms	0.0638 ± 0.019 ms	0.973 ± 0.43
saxpy/static workgroup=(1024,)/Float64/1024	0.0342 ± 0.02 ms	0.0349 ± 0.021 ms	0.978 ± 0.81
saxpy/static workgroup=(1024,)/Float64/1048576	0.653 ± 0.091 ms	0.651 ± 0.1 ms	1 ± 0.21
saxpy/static workgroup=(1024,)/Float64/16384	0.044 ± 0.02 ms	0.0466 ± 0.022 ms	0.943 ± 0.63
saxpy/static workgroup=(1024,)/Float64/2048	0.049 ± 0.022 ms	0.0465 ± 0.022 ms	1.05 ± 0.7
saxpy/static workgroup=(1024,)/Float64/256	30.7 ± 9.9 μs	31.1 ± 14 μs	0.988 ± 0.54
saxpy/static workgroup=(1024,)/Float64/262144	0.198 ± 0.021 ms	0.189 ± 0.02 ms	1.05 ± 0.16
saxpy/static workgroup=(1024,)/Float64/32768	0.0528 ± 0.021 ms	0.0551 ± 0.021 ms	0.959 ± 0.53
saxpy/static workgroup=(1024,)/Float64/4096	0.0504 ± 0.018 ms	0.052 ± 0.018 ms	0.969 ± 0.49
saxpy/static workgroup=(1024,)/Float64/512	30.7 ± 13 μs	30.8 ± 17 μs	0.996 ± 0.69
saxpy/static workgroup=(1024,)/Float64/64	30.8 ± 13 μs	30.1 ± 15 μs	1.02 ± 0.66
saxpy/static workgroup=(1024,)/Float64/65536	0.0741 ± 0.019 ms	0.0749 ± 0.019 ms	0.989 ± 0.36
time_to_load	0.751 ± 0.024 s	0.758 ± 0.0094 s	0.991 ± 0.034

Benchmark Plots

A plot of the benchmark results have been uploaded as an artifact to the workflow run for this PR.
Go to "Actions"->"Benchmark a pull request"->[the most recent run]->"Artifacts" (at the bottom).

[christiangnrd]

Metal failure is fixed by JuliaGPU/Metal.jl#894.

oneAPI and CUDA fail because they return a device instead of nothing but that's not specified in the interface so probably just remove the tests?

[christiangnrd]

Similarly, priority! in AMDGPU returns the symbol instead of nothing, but that's not specified in the interface