diff --git a/.buildkite/pipeline.yml b/.buildkite/pipeline.yml index fa341ea459..94714b3a31 100755 --- a/.buildkite/pipeline.yml +++ b/.buildkite/pipeline.yml @@ -1242,6 +1242,16 @@ steps: agents: slurm_gpus: 1 + - label: "Perf: benchmark scripts benchmark_offset" + key: benchmark_offset + command: + - "julia --project=.buildkite -e 'using CUDA; CUDA.versioninfo()'" + - "julia --color=yes --project=.buildkite benchmarks/scripts/benchmark_offset.jl" + env: + CLIMACOMMS_DEVICE: "CUDA" + agents: + slurm_gpus: 1 + - group: "Perf: Operators" steps: diff --git a/benchmarks/scripts/benchmark_offset.jl b/benchmarks/scripts/benchmark_offset.jl new file mode 100644 index 0000000000..5304c01007 --- /dev/null +++ b/benchmarks/scripts/benchmark_offset.jl @@ -0,0 +1,309 @@ +#= +julia --project=.buildkite +using Revise; include(joinpath("benchmarks", "scripts", "benchmark_offset.jl")) + +# Info + + - This benchmark demos the performance for different offset styles: + - Array of structs with Cartesian offsets + - Array of structs with Linear offsets + - Struct of arrays with no offsets + +# Benchmark results: + +Clima A100: +``` +[ Info: ArrayType = CuArray +Problem size: (63, 4, 4, 1, 5400), float_type = Float32, device_bandwidth_GBs=2039 +┌────────────────────────────────────────────────────────────────────┬──────────────────────────────────┬─────────┬─────────────┬────────────────┬────────┐ +│ funcs │ time per call │ bw % │ achieved bw │ n-reads/writes │ n-reps │ +├────────────────────────────────────────────────────────────────────┼──────────────────────────────────┼─────────┼─────────────┼────────────────┼────────┤ +│ BO.aos_cart_offset!(X_aos_ref, Y_aos_ref, us; bm, nreps = 100) │ 68 microseconds, 834 nanoseconds │ 57.7908 │ 1178.35 │ 4 │ 100 │ +│ BO.aos_lin_offset!(X_aos, Y_aos, us; bm, nreps = 100) │ 58 microseconds, 153 nanoseconds │ 68.4046 │ 1394.77 │ 4 │ 100 │ +│ BO.soa_linear_index!(X_soa, Y_soa, us; bm, nreps = 100) │ 56 microseconds, 576 nanoseconds │ 70.3113 │ 1433.65 │ 4 │ 100 │ +│ BO.soa_cart_index!(X_soa, Y_soa, us; bm, nreps = 100) │ 67 microseconds, 185 nanoseconds │ 59.2089 │ 1207.27 │ 4 │ 100 │ +└────────────────────────────────────────────────────────────────────┴──────────────────────────────────┴─────────┴─────────────┴────────────────┴────────┘ + +[ Info: ArrayType = CuArray +Problem size: (63, 4, 4, 1, 5400), float_type = Float32, device_bandwidth_GBs=2039 +┌────────────────────────────────────────────────────────────────────┬──────────────────────────────────┬─────────┬─────────────┬────────────────┬────────┐ +│ funcs │ time per call │ bw % │ achieved bw │ n-reads/writes │ n-reps │ +├────────────────────────────────────────────────────────────────────┼──────────────────────────────────┼─────────┼─────────────┼────────────────┼────────┤ +│ BO.aos_cart_offset!(X_aos_ref, Y_aos_ref, us; bm, nreps = 100) │ 68 microseconds, 967 nanoseconds │ 57.6793 │ 1176.08 │ 4 │ 100 │ +│ BO.aos_lin_offset!(X_aos, Y_aos, us; bm, nreps = 100) │ 58 microseconds, 82 nanoseconds │ 68.489 │ 1396.49 │ 4 │ 100 │ +│ BO.soa_linear_index!(X_soa, Y_soa, us; bm, nreps = 100) │ 56 microseconds, 597 nanoseconds │ 70.2858 │ 1433.13 │ 4 │ 100 │ +│ BO.soa_cart_index!(X_soa, Y_soa, us; bm, nreps = 100) │ 67 microseconds, 288 nanoseconds │ 59.1188 │ 1205.43 │ 4 │ 100 │ +└────────────────────────────────────────────────────────────────────┴──────────────────────────────────┴─────────┴─────────────┴────────────────┴────────┘ +``` +=# + +#! format: off +module BenchmarkOffset + +include("benchmark_utils.jl") + +add3(x1, x2, x3) = x1 + x2 + x3 + +function aos_cart_offset!(X, Y, us; nreps = 1, bm=nothing, n_trials = 30) + if Y isa Array + e = Inf + CI = CartesianIndices((get_Nv(us), get_Nij(us), get_Nij(us), 1, get_Nh(us))) + for t in 1:n_trials + et = Base.@elapsed begin + for i in 1:nreps + @inbounds @simd for I in 1:get_N(us) + CI1 = CI[I] + CI2 = CI1 + CartesianIndex((0, 0, 0, 1, 0)) + CI3 = CI1 + CartesianIndex((0, 0, 0, 2, 0)) + Y[CI1] = add3(X[CI1], X[CI2], X[CI3]) + end + end + end + e = min(e, et) + end + else + e = Inf + kernel = CUDA.@cuda always_inline = true launch = false aos_cart_offset_kernel!(X,Y,us) + config = CUDA.launch_configuration(kernel.fun) + threads = min(get_N(us), config.threads) + blocks = cld(get_N(us), threads) + for t in 1:n_trials + et = CUDA.@elapsed begin + for i in 1:nreps # reduce variance / impact of launch latency + kernel(X,Y,us; threads, blocks) + end + end + e = min(e, et) + end + end + push_info(bm; e, nreps, caller = @caller_name(@__FILE__),n_reads_writes=4) + return nothing +end; +function aos_cart_offset_kernel!(X, Y, us) + @inbounds begin + I = (CUDA.blockIdx().x - Int32(1)) * CUDA.blockDim().x + CUDA.threadIdx().x + if I ≤ get_N(us) + n = (get_Nv(us), get_Nij(us), get_Nij(us), 1, get_Nh(us)) + CI1 = CartesianIndices(map(x -> Base.OneTo(x), n))[I] + CI2 = CI1 + CartesianIndex((0, 0, 0, 1, 0)) + CI3 = CI1 + CartesianIndex((0, 0, 0, 2, 0)) + Y[CI1] = add3(X[CI1], X[CI2], X[CI3]) + end + end + return nothing +end; + +function aos_lin_offset!(X, Y, us; nreps = 1, bm=nothing, n_trials = 30) + if Y isa Array + e = Inf + for t in 1:n_trials + et = Base.@elapsed begin + for i in 1:nreps + @inbounds @simd for I in 1:get_N(us) + CI = CartesianIndices((get_Nv(us), get_Nij(us), get_Nij(us), 1, get_Nh(us))) + LI1 = LinearIndices((get_Nv(us), get_Nij(us), get_Nij(us), 1, get_Nh(us))) + LI3 = LinearIndices((get_Nv(us), get_Nij(us), get_Nij(us), 3, get_Nh(us))) + CI1 = CI[I] + CI2 = CI1 + CartesianIndex((0, 0, 0, 1, 0)) + CI3 = CI1 + CartesianIndex((0, 0, 0, 2, 0)) + IY1 = LI1[CI1] + IX1 = LI3[CI1] + IX2 = LI3[CI2] + IX3 = LI3[CI3] + Y[IY1] = add3(X[IX1], X[IX2], X[IX3]) + end + end + end + e = min(e, et) + end + else + e = Inf + kernel = CUDA.@cuda always_inline = true launch = false aos_lin_offset_kernel!(X,Y,us) + config = CUDA.launch_configuration(kernel.fun) + threads = min(get_N(us), config.threads) + blocks = cld(get_N(us), threads) + for t in 1:n_trials + et = CUDA.@elapsed begin + for i in 1:nreps + kernel(X,Y,us; threads, blocks) + end + end + e = min(e, et) + end + end + push_info(bm; e, nreps, caller = @caller_name(@__FILE__),n_reads_writes=4) + return nothing +end; +function aos_lin_offset_kernel!(X, Y, us) + @inbounds begin + I = (CUDA.blockIdx().x - Int32(1)) * CUDA.blockDim().x + CUDA.threadIdx().x + if I ≤ get_N(us) + CI = CartesianIndices((get_Nv(us), get_Nij(us), get_Nij(us), 1, get_Nh(us))) + LI1 = LinearIndices((get_Nv(us), get_Nij(us), get_Nij(us), 1, get_Nh(us))) + LI3 = LinearIndices((get_Nv(us), get_Nij(us), get_Nij(us), 3, get_Nh(us))) + CI1 = CI[I] + CI2 = CI1 + CartesianIndex((0, 0, 0, 1, 0)) + CI3 = CI1 + CartesianIndex((0, 0, 0, 2, 0)) + IY1 = LI1[CI1] + IX1 = LI3[CI1] + IX2 = LI3[CI2] + IX3 = LI3[CI3] + Y[IY1] = add3(X[IX1], X[IX2], X[IX3]) + end + end + return nothing +end; + +function soa_cart_index!(X, Y, us; nreps = 1, bm=nothing, n_trials = 30) + e = Inf + if first(Y) isa Array + CI = CartesianIndices((get_Nv(us), get_Nij(us), get_Nij(us), get_Nh(us))) + for t in 1:n_trials + et = Base.@elapsed begin + for i in 1:nreps + (y1,) = Y + (x1, x2, x3) = X + @inbounds @simd for I in 1:get_N(us) + y1[CI[I]] = add3(x1[CI[I]], x2[CI[I]], x3[CI[I]]) + end + end + end + e = min(e, et) + end + else + kernel = CUDA.@cuda always_inline = true launch = false soa_cart_index_kernel!(X,Y,us) + config = CUDA.launch_configuration(kernel.fun) + threads = min(get_N(us), config.threads) + blocks = cld(get_N(us), threads) + for t in 1:n_trials + et = CUDA.@elapsed begin + for i in 1:nreps # reduce variance / impact of launch latency + kernel(X,Y,us; threads, blocks) + end + end + e = min(e, et) + end + end + push_info(bm; e, nreps, caller = @caller_name(@__FILE__),n_reads_writes=4) + return nothing +end; +function soa_cart_index_kernel!(X, Y, us) + @inbounds begin + I = (CUDA.blockIdx().x - Int32(1)) * CUDA.blockDim().x + CUDA.threadIdx().x + if I ≤ get_N(us) + CI = CartesianIndices((get_Nv(us), get_Nij(us), get_Nij(us), get_Nh(us))) + (y1,) = Y + (x1, x2, x3) = X + y1[CI[I]] = add3(x1[CI[I]], x2[CI[I]], x3[CI[I]]) + end + end + return nothing +end; + +function soa_linear_index!(X, Y, us; nreps = 1, bm=nothing, n_trials = 30) + e = Inf + if first(Y) isa Array + for t in 1:n_trials + et = Base.@elapsed begin + for i in 1:nreps + (y1,) = Y + (x1, x2, x3) = X + @inbounds @simd for I in 1:get_N(us) + y1[I] = add3(x1[I], x2[I], x3[I]) + end + end + end + e = min(e, et) + end + else + kernel = CUDA.@cuda always_inline = true launch = false soa_linear_index_kernel!(X,Y,us) + config = CUDA.launch_configuration(kernel.fun) + threads = min(get_N(us), config.threads) + blocks = cld(get_N(us), threads) + for t in 1:n_trials + et = CUDA.@elapsed begin + for i in 1:nreps # reduce variance / impact of launch latency + kernel(X,Y,us; threads, blocks) + end + end + e = min(e, et) + end + end + push_info(bm; e, nreps, caller = @caller_name(@__FILE__),n_reads_writes=4) + return nothing +end; +function soa_linear_index_kernel!(X, Y, us) + @inbounds begin + I = (CUDA.blockIdx().x - Int32(1)) * CUDA.blockDim().x + CUDA.threadIdx().x + if I ≤ get_N(us) + (y1,) = Y + (x1, x2, x3) = X + y1[I] = add3(x1[I], x2[I], x3[I]) + end + end + return nothing +end; + +end # module + +import .BenchmarkOffset as BO + +function fill_with_rand!(arr) + FT = eltype(arr) + T = typeof(arr) + s = size(arr) + arr .= T(rand(FT, s)) +end + +using CUDA +using Test +@testset "Offset benchmark" begin + bm = BO.Benchmark(;problem_size=(63,4,4,1,5400), float_type=Float32) # size(problem_size, 4) == 1 to avoid double counting reads/writes + ArrayType = CUDA.CuArray; + # ArrayType = Base.identity; + arr(float_type, problem_size, T) = T(zeros(float_type, problem_size...)) + + FT = Float64; + s = (63,4,4,3,5400); + sY = (63,4,4,1,5400); + st = (63,4,4,5400); + ndofs = prod(st); + us = BO.UniversalSizesStatic(s[1], s[2], s[end]); + + X_aos = arr(bm.float_type, s, ArrayType); + Y_aos = arr(bm.float_type, sY, ArrayType); + X_aos_ref = arr(bm.float_type, s, ArrayType); + Y_aos_ref = arr(bm.float_type, sY, ArrayType); + X_soa = ntuple(_ -> arr(bm.float_type, st, ArrayType), 3); + Y_soa = ntuple(_ -> arr(bm.float_type, st, ArrayType), 1); + fill_with_rand!(X_aos) + fill_with_rand!(Y_aos) + X_aos_ref .= X_aos + Y_aos_ref .= Y_aos + for i in 1:3; X_soa[i] .= X_aos[:,:,:,i,:]; end + for i in 1:1; Y_soa[i] .= Y_aos[:,:,:,i,:]; end + @info "ArrayType = $ArrayType" + + BO.aos_cart_offset!(X_aos_ref, Y_aos_ref, us; n_trials = 1, nreps = 1) + BO.aos_lin_offset!(X_aos, Y_aos, us; n_trials = 1, nreps = 1) + BO.soa_linear_index!(X_soa, Y_soa, us; n_trials = 1, nreps = 1) + + @test all(X_aos .== X_aos_ref) + @test all(Y_aos .== Y_aos_ref) + for i in 1:3; @test all(X_soa[i] .== X_aos_ref[:,:,:,i,:]); end + for i in 1:1; @test all(Y_soa[i] .== Y_aos_ref[:,:,:,i,:]); end + + BO.soa_cart_index!(X_soa, Y_soa, us; n_trials = 1, nreps = 1) + + for i in 1:3; @test all(X_soa[i] .== X_aos_ref[:,:,:,i,:]); end + for i in 1:1; @test all(Y_soa[i] .== Y_aos_ref[:,:,:,i,:]); end + + BO.aos_cart_offset!(X_aos_ref, Y_aos_ref, us; bm, nreps = 100) + BO.aos_lin_offset!(X_aos, Y_aos, us; bm, nreps = 100) + BO.soa_linear_index!(X_soa, Y_soa, us; bm, nreps = 100) + BO.soa_cart_index!(X_soa, Y_soa, us; bm, nreps = 100) + + BO.tabulate_benchmark(bm) +end + +# #! format: on