Added benchmarks for static arrays (#51)

Co-authored-by: Dilum Aluthge <[email protected]>

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "Octavian"
 uuid = "6fd5a793-0b7e-452c-907f-f8bfe9c57db4"
 authors = ["Mason Protter", "Chris Elrod", "Dilum Aluthge", "contributors"]
-version = "0.2.0"
+version = "0.2.1"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"

benchmark/staticarraybench.jl

@@ -0,0 +1,86 @@
+
+using Octavian, StaticArrays, LinearAlgebra, BenchmarkTools, ProgressMeter
+
+# BLAS.set_num_threads(1)
+
+# For laptops that thermally throttle, you can set the `JULIA_SLEEP_BENCH` environment variable for #seconds to sleep before each `@belapsed`
+const SLEEPTIME = parse(Float64, get(ENV, "JULIA_SLEEP_BENCH", "0"))
+maybe_sleep() = iszero(SLEEPTIME) || sleep(SLEEPTIME)
+# BenchmarkTools.DEFAULT_PARAMETERS.samples = 1_000_000
+# BenchmarkTools.DEFAULT_PARAMETERS.seconds = 10
+
+matrix_sizes(x::Integer) = (x,x,x)
+matrix_sizes(x::NTuple{3}) = x
+
+const matmulmethodnames = [:SMatrix, :MMatrix, :OctavianStatic, :OctavianDynamic];
+function fill_bench_results!(br, lp, (M,K,N), t, i, j)
+    name = matmulmethodnames[j]
+    br[i,j,1] = t
+    gflops = 2e-9M*K*N / t
+    br[i,j,2] = gflops
+    lp[j+1] = (name, gflops)
+    nothing
+end
+
+function runbenches(sr, ::Type{T} = Float64) where {T}
+    bench_results = Array{Float64}(undef, length(sr), 4, 2)
+    p = Progress(length(sr))
+    last_perfs = Vector{Tuple{Symbol,Union{Float64,NTuple{3,Int}}}}(undef, size(bench_results,2) + 1)
+    for (i,s) ∈ enumerate(sr)
+        M, K, N = matrix_sizes(s)
+        last_perfs[1] = (:Size, (M,K,N))
+        Astatic = @SMatrix rand(T, M, K);
+        Bstatic = @SMatrix rand(T, K, N);
+        maybe_sleep()
+        t = @belapsed $(Ref(Astatic))[] * $(Ref(Bstatic))[]
+        fill_bench_results!(bench_results, last_perfs, (M,K,N), t, i, 1)
+        Amutable = MArray(Astatic);
+        Bmutable = MArray(Bstatic);
+        Cmutable = MMatrix{M,N,T}(undef);
+        maybe_sleep()
+        t = @belapsed mul!($Cmutable, $Amutable, $Bmutable)
+        fill_bench_results!(bench_results, last_perfs, (M,K,N), t, i, 2)
+        Cmutable0 = copy(Cmutable); Cmutable .= NaN;
+        maybe_sleep()
+        t = @belapsed matmul!($Cmutable, $Amutable, $Bmutable)
+        fill_bench_results!(bench_results, last_perfs, (M,K,N), t, i, 3)
+        A = Array(Amutable); B = Array(Bmutable); C = Array(Cmutable);
+        maybe_sleep()
+        t = @belapsed matmul!($C, $A, $B)
+        fill_bench_results!(bench_results, last_perfs, (M,K,N), t, i, 4)
+        @assert Array(Cmutable) ≈ Array(Cmutable0) ≈ C
+        ProgressMeter.next!(p; showvalues = last_perfs)
+    end
+    bench_results
+end
+
+sizerange = 2:48
+br = runbenches(sizerange);
+using DataFrames, VegaLite
+
+df = DataFrame(@view(br[:,:,2]));
+rename!(df, matmulmethodnames);
+df.Size = sizerange
+
+function pick_suffix(desc = "")
+    suffix = if Octavian.VectorizationBase.AVX512F
+        "AVX512"
+    elseif Octavian.VectorizationBase.AVX2
+        "AVX2"
+    elseif Octavian.VectorizationBase.REGISTER_SIZE == 32
+        "AVX"
+    else
+        "REGSUZE$(Octavian.VectorizationBase.REGISTER_SIZE)"
+    end
+    if desc != ""
+        suffix *= '_' * desc
+    end
+    "$(Sys.CPU_NAME)_$suffix"
+end
+
+dfs = stack(df, matmulmethodnames, variable_name = :MatMulType, value_name = :GFLOPS);
+p = dfs |> @vlplot(:line, x = :Size, y = :GFLOPS, width = 900, height = 600, color = {:MatMulType});
+save(joinpath(pkgdir(Octavian), "docs/src/assets/sizedarraybenchmarks_$(pick_suffix()).svg"), p)
+
+
+