|
| 1 | +using Test |
| 2 | +using BenchmarkTools |
| 3 | +import ClimaComms |
| 4 | +import ClimaCore: ClimaCore, DataLayouts |
| 5 | +@static pkgversion(ClimaComms) >= v"0.6" && ClimaComms.@import_required_backends |
| 6 | +if ClimaComms.device() isa ClimaComms.CUDADevice |
| 7 | + import CUDA |
| 8 | + device_name = CUDA.name(CUDA.device()) # Move to ClimaComms |
| 9 | +else |
| 10 | + device_name = "CPU" |
| 11 | +end |
| 12 | + |
| 13 | +include(joinpath(pkgdir(ClimaCore), "benchmarks/scripts/benchmark_utils.jl")) |
| 14 | + |
| 15 | +function benchmarkfill!(bm, device, data, val) |
| 16 | + caller = string(DataLayouts.layout_constructor(data)) |
| 17 | + @info "Benchmarking $caller..." |
| 18 | + trial = @benchmark ClimaComms.@cuda_sync $device fill!($data, $val) |
| 19 | + kernel_time_s = minimum(trial.times) * 1e-9 # to seconds |
| 20 | + nreps = length(trial.times) |
| 21 | + problem_size = size(data) |
| 22 | + n_reads_writes = DataLayouts.ncomponents(data) |
| 23 | + push_info(bm; kernel_time_s, nreps, caller, problem_size, n_reads_writes) |
| 24 | +end |
| 25 | + |
| 26 | +@testset "fill! with Nf = 1" begin |
| 27 | + device = ClimaComms.device() |
| 28 | + FT = Float64 |
| 29 | + A = ClimaComms.array_type(device){FT} |
| 30 | + bm = Benchmark(; float_type = FT, device_name) |
| 31 | + |
| 32 | + data = DataLayouts.DataF{FT}(A) |
| 33 | + benchmarkfill!(bm, device, data, 3) |
| 34 | + @test all(parent(data) .== 3) |
| 35 | + |
| 36 | + (Nv, Nij, Nh) = (63, 4, 30 * 30 * 6) |
| 37 | + for Nv in (1, Nv), (Ni, Nj) in ((1, 1), (Nij, 1), (Nij, Nij)), Nh in (1, Nh) |
| 38 | + for D in (DataLayouts.VIJFH, DataLayouts.VIJHF) |
| 39 | + data = D{FT, Nv, Ni, Nj, Nh == 1 ? 1 : nothing}(A, Nh) |
| 40 | + benchmarkfill!(bm, device, data, 3) |
| 41 | + @test all(parent(data) .== 3) |
| 42 | + end |
| 43 | + end |
| 44 | + for Nv in (1, Nv), Ni in (1, Nij), Nh in (1, Nh) |
| 45 | + data = DataLayouts.VIH1{FT, Nv, Ni, Nh == 1 ? 1 : nothing}(A, Nh) |
| 46 | + benchmarkfill!(bm, device, data, 3) |
| 47 | + @test all(parent(data) .== 3) |
| 48 | + end |
| 49 | + for (Ni, Nj) in ((1, 1), (Nij, 1), (Nij, Nij)), Nh in (1, Nh) |
| 50 | + data = DataLayouts.IH1JH2{FT, Ni, Nj, Nh == 1 ? 1 : nothing}(A, Nh) |
| 51 | + benchmarkfill!(bm, device, data, 3) |
| 52 | + @test all(parent(data) .== 3) |
| 53 | + end |
| 54 | + |
| 55 | + tabulate_benchmark(bm) |
| 56 | +end |
0 commit comments