|
| 1 | +group = addgroup!(SUITE, "array") |
| 2 | + |
| 3 | +const m = 512 |
| 4 | +const n = 1000 |
| 5 | + |
| 6 | +# generate some arrays |
| 7 | +cpu_mat = rand(rng, Float32, m, n) |
| 8 | +gpu_mat = MtlArray{Float32}(undef, size(cpu_mat)) |
| 9 | +gpu_vec = reshape(gpu_mat, length(gpu_mat)) |
| 10 | +gpu_arr_3d = reshape(gpu_mat, (m, 40, 25)) |
| 11 | +gpu_arr_4d = reshape(gpu_mat, (m, 10, 10, 10)) |
| 12 | +gpu_mat_ints = MtlArray(rand(rng, Int, m, n)) |
| 13 | +gpu_vec_ints = reshape(gpu_mat_ints, length(gpu_mat_ints)) |
| 14 | +gpu_mat_bools = MtlArray(rand(rng, Bool, m, n)) |
| 15 | +gpu_vec_bools = reshape(gpu_mat_bools, length(gpu_mat_bools)) |
| 16 | + |
| 17 | +group["construct"] = @benchmarkable MtlArray{Int}(undef, 1) |
| 18 | + |
| 19 | +group["copy"] = @async_benchmarkable copy($gpu_mat) |
| 20 | + |
| 21 | +gpu_mat2 = copy(gpu_mat) |
| 22 | +let group = addgroup!(group, "copyto!") |
| 23 | + group["cpu_to_gpu"] = @async_benchmarkable copyto!($gpu_mat, $cpu_mat) |
| 24 | + group["gpu_to_cpu"] = @async_benchmarkable copyto!($cpu_mat, $gpu_mat) |
| 25 | + group["gpu_to_gpu"] = @async_benchmarkable copyto!($gpu_mat2, $gpu_mat) |
| 26 | +end |
| 27 | + |
| 28 | +let group = addgroup!(group, "iteration") |
| 29 | + group["scalar"] = @benchmarkable Metal.@allowscalar [$gpu_vec[i] for i in 1:10] |
| 30 | + |
| 31 | + group["logical"] = @benchmarkable $gpu_vec[$gpu_vec_bools] |
| 32 | + |
| 33 | + let group = addgroup!(group, "findall") |
| 34 | + group["bool"] = @benchmarkable findall($gpu_vec_bools) |
| 35 | + group["int"] = @benchmarkable findall(isodd, $gpu_vec_ints) |
| 36 | + end |
| 37 | + |
| 38 | + let group = addgroup!(group, "findfirst") |
| 39 | + group["bool"] = @benchmarkable findfirst($gpu_vec_bools) |
| 40 | + group["int"] = @benchmarkable findfirst(isodd, $gpu_vec_ints) |
| 41 | + end |
| 42 | + |
| 43 | + let group = addgroup!(group, "findmin") # findmax |
| 44 | + group["1d"] = @async_benchmarkable findmin($gpu_vec) |
| 45 | + group["2d"] = @async_benchmarkable findmin($gpu_mat; dims=1) |
| 46 | + end |
| 47 | +end |
| 48 | + |
| 49 | +# let group = addgroup!(group, "reverse") |
| 50 | +# group["1d"] = @async_benchmarkable reverse($gpu_vec) |
| 51 | +# group["2d"] = @async_benchmarkable reverse($gpu_mat; dims=1) |
| 52 | +# group["1d_inplace"] = @async_benchmarkable reverse!($gpu_vec) |
| 53 | +# group["2d_inplace"] = @async_benchmarkable reverse!($gpu_mat; dims=1) |
| 54 | +# end |
| 55 | + |
| 56 | +group["broadcast"] = @async_benchmarkable $gpu_mat .= 0f0 |
| 57 | + |
| 58 | +# no need to test inplace version, which performs the same operation (but with an alloc) |
| 59 | +let group = addgroup!(group, "accumulate") |
| 60 | + group["1d"] = @async_benchmarkable accumulate(+, $gpu_vec) |
| 61 | + group["2d"] = @async_benchmarkable accumulate(+, $gpu_mat; dims=1) |
| 62 | +end |
| 63 | + |
| 64 | +let group = addgroup!(group, "reductions") |
| 65 | + let group = addgroup!(group, "reduce") |
| 66 | + group["1d"] = @async_benchmarkable reduce(+, $gpu_vec) |
| 67 | + group["2d"] = @async_benchmarkable reduce(+, $gpu_mat; dims=1) |
| 68 | + end |
| 69 | + |
| 70 | + let group = addgroup!(group, "mapreduce") |
| 71 | + group["1d"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_vec) |
| 72 | + group["2d"] = @async_benchmarkable mapreduce(x->x+1, +, $gpu_mat; dims=1) |
| 73 | + end |
| 74 | + |
| 75 | + # used by sum, prod, minimum, maximum, all, any, count |
| 76 | +end |
| 77 | + |
| 78 | +let group = addgroup!(group, "random") |
| 79 | + let group = addgroup!(group, "rand") |
| 80 | + group["Float32"] = @async_benchmarkable Metal.rand(Float32, m*n) |
| 81 | + group["Int64"] = @async_benchmarkable Metal.rand(Int64, m*n) |
| 82 | + end |
| 83 | + |
| 84 | + let group = addgroup!(group, "rand!") |
| 85 | + group["Float32"] = @async_benchmarkable Metal.rand!($gpu_vec) |
| 86 | + group["Int64"] = @async_benchmarkable Metal.rand!($gpu_vec_ints) |
| 87 | + end |
| 88 | + |
| 89 | + let group = addgroup!(group, "randn") |
| 90 | + group["Float32"] = @async_benchmarkable Metal.randn(Float32, m*n) |
| 91 | + # group["Int64"] = @async_benchmarkable Metal.randn(Int64, m*n) |
| 92 | + end |
| 93 | + |
| 94 | + let group = addgroup!(group, "randn!") |
| 95 | + group["Float32"] = @async_benchmarkable Metal.randn!($gpu_vec) |
| 96 | + # group["Int64"] = @async_benchmarkable Metal.randn!($gpu_vec_ints) |
| 97 | + end |
| 98 | +end |
| 99 | + |
| 100 | +# let group = addgroup!(group, "sorting") |
| 101 | +# group["1d"] = @async_benchmarkable sort($gpu_vec) |
| 102 | +# group["2d"] = @async_benchmarkable sort($gpu_mat; dims=1) |
| 103 | +# group["by"] = @async_benchmarkable sort($gpu_vec; by=sin) |
| 104 | +# end |
| 105 | + |
| 106 | +let group = addgroup!(group, "permutedims") |
| 107 | + group["2d"] = @async_benchmarkable permutedims($gpu_mat, (2,1)) |
| 108 | + group["3d"] = @async_benchmarkable permutedims($gpu_arr_3d, (3,1,2)) |
| 109 | + group["4d"] = @async_benchmarkable permutedims($gpu_arr_4d, (2,1,4,3)) |
| 110 | +end |
0 commit comments