Actually benchmark g++ and clang++-compiled Eigen.

Author: chriselrod

benchmark/benchmarkflops.jl

@@ -284,10 +284,10 @@ function benchmark_dot3(sizes)
     pmap(is -> dot3_bench!(sm, is[2], is[1]), enumerate(sizes))
     br
 end
-BLAS.set_num_threads(1)
+# BLAS.set_num_threads(1)
 function sse!(Xβ, y, X, β)
-    mul!(copyto!(Xβ, y), X, β, 1.0, -1.0)
-    dot(Xβ, Xβ)
+    dgemvmkl!(copyto!(Xβ, y), X, β, 1.0, -1.0)
+    jdot(Xβ, Xβ)
 end
 sse_totwotuple(s::NTuple{2}) = s
 sse_totwotuple(s::Integer) = ((3s) >> 1, s >> 1)
@@ -318,7 +318,7 @@ function sse_bench!(br, s, i)
     br[9,i] = n_gflop / @belapsed sse!($Xβ, $y, $X, $β)
 end
 function benchmark_sse(sizes)
-    tests = ["LoopVectorization", "Julia", "Clang", "GFortran", "icc", "ifort", "g++ & Eigen-3", "clang++ & Eigen-3", BLAS.vendor() === :mkl ? "MKL" : "OpenBLAS"]
+    tests = ["LoopVectorization", "Julia", "Clang", "GFortran", "icc", "ifort", "g++ & Eigen-3", "clang++ & Eigen-3", "MKL"]
     br = BenchmarkResult(tests, sizes)
     sm = br.sizedresults.results
     pmap(is -> sse_bench!(sm, is[2], is[1]), enumerate(sizes))

benchmark/driver.jl

@@ -34,15 +34,16 @@ AtmulB_bench = benchmark_AtmulB(sizes); println("A' * B benchmark results:"); pr
 Amulvb_bench = benchmark_Amulvb(sizes); println("A * b benchmark results:"); println(Amulvb_bench)
 Atmulvb_bench = benchmark_Atmulvb(sizes); println("A' * b benchmark results:"); println(Atmulvb_bench)
 
-filter2d_dynamic_bench = benchmark_filter2ddynamic(sizes); println("Benchmark results for dynamically sized 3x3 convolution:"); println(filter2d_dynamic_bench)
-filter2d_3x3_bench = benchmark_filter2d3x3(sizes); println("Benchmark results for statically sized 3x3 convolution:"); println(filter2d_3x3_bench)
-filter2d_unrolled_bench = benchmark_filter2dunrolled(sizes); println("Benchmark results for unrolled 3x3 convolution:"); println(filter2d_unrolled_bench)
-
 dot_bench = benchmark_dot(longsizes); println("a' * b benchmark results:"); println(dot_bench)
 selfdot_bench = benchmark_selfdot(longsizes); println("a' * a benchmark results:"); println(selfdot_bench)
 sse_bench = benchmark_sse(sizes); println("Benchmark resutls of summing squared error:"); println(sse_bench)
 aplusBc_bench = benchmark_aplusBc(sizes); println("Benchmark results of a .+ B .* c':"); println(aplusBc_bench)
 AplusAt_bench = benchmark_AplusAt(sizes); println("Benchmark results of A * A':"); println(AplusAt_bench)
+
+filter2d_dynamic_bench = benchmark_filter2ddynamic(sizes); println("Benchmark results for dynamically sized 3x3 convolution:"); println(filter2d_dynamic_bench)
+filter2d_3x3_bench = benchmark_filter2d3x3(sizes); println("Benchmark results for statically sized 3x3 convolution:"); println(filter2d_3x3_bench)
+filter2d_unrolled_bench = benchmark_filter2dunrolled(sizes); println("Benchmark results for unrolled 3x3 convolution:"); println(filter2d_unrolled_bench)
+
 vexp_bench = benchmark_exp(sizes); println("Benchmark results of exponentiating a vector:"); println(vexp_bench)
 randomaccess_bench = benchmark_random_access(sizes); println("Benchmark results from using a vector of indices:"); println(randomaccess_bench)
 
@@ -51,24 +52,27 @@ using Cairo, Fontconfig
 const PICTURES = joinpath(pkgdir(LoopVectorization), "docs", "src", "assets")
 saveplot(f, br) = draw(PNG(joinpath(PICTURES, f * "$v.png"), 12inch, 8inch), plot(br))
 
-saveplot("bench_logdettriangle_v", logdettriangle_bench);
-saveplot("bench_filter2d_dynamic_v", filter2d_dynamic_bench);
-saveplot("bench_filter2d_3x3_v", filter2d_3x3_bench);
-saveplot("bench_filter2d_unrolled_v", filter2d_unrolled_bench);
+saveplot("bench_dot3_v", dot3_bench);
 saveplot("bench_dot_v", dot_bench);
 saveplot("bench_selfdot_v", selfdot_bench);
-saveplot("bench_dot3_v", dot3_bench);
 saveplot("bench_sse_v", sse_bench);
 saveplot("bench_aplusBc_v", aplusBc_bench);
 saveplot("bench_AplusAt_v", AplusAt_bench);
-saveplot("bench_exp_v", vexp_bench);
-saveplot("bench_random_access_v", randomaccess_bench);
 saveplot("bench_AmulB_v", AmulB_bench);
 saveplot("bench_AmulBt_v", AmulBt_bench);
 saveplot("bench_AtmulB_v", AtmulB_bench);
 saveplot("bench_AtmulBt_v", AtmulBt_bench);
 saveplot("bench_Amulvb_v", Amulvb_bench);
 saveplot("bench_Atmulvb_v", Atmulvb_bench);
 
+saveplot("bench_logdettriangle_v", logdettriangle_bench);
+saveplot("bench_filter2d_dynamic_v", filter2d_dynamic_bench);
+saveplot("bench_filter2d_3x3_v", filter2d_3x3_bench);
+saveplot("bench_filter2d_unrolled_v", filter2d_unrolled_bench);
+saveplot("bench_exp_v", vexp_bench);
+saveplot("bench_random_access_v", randomaccess_bench);
+
+# @load "benchmarkresults.jld2" logdettriangle_bench filter2d_dynamic_bench filter2d_3x3_bench filter2d_unrolled_bench vexp_bench randomaccess_bench
+
 @save "benchmarkresults.jld2" logdettriangle_bench filter2d_dynamic_bench filter2d_3x3_bench filter2d_unrolled_bench dot_bench selfdot_bench dot3_bench sse_bench aplusBc_bench AplusAt_bench vexp_bench randomaccess_bench AmulB_bench AmulBt_bench AtmulB_bench AtmulBt_bench Amulvb_bench Atmulvb_bench
 

benchmark/loadsharedlibs.jl

@@ -116,7 +116,7 @@ mkl_set_num_threads(N::Integer) = ccall(MKL_SET_NUM_THREADS, Cvoid, (Int32,), N
 mkl_set_num_threads(1)
 openblas_set_num_threads(N::Integer) = ccall(OPENBLAS_SET_NUM_THREADS, Cvoid, (Int64,), N)
 openblas_set_num_threads(1)
-function dgemvmkl!(y::AbstractVector{Float64}, A::AbstractMatrix{Float64}, x::AbstractVector{Float64})
+function dgemvmkl!(y::AbstractVector{Float64}, A::AbstractMatrix{Float64}, x::AbstractVector{Float64}, α = 1.0, β = 0.0)
     transA = istransposed(A)
     pA = parent(A)
     M, N = size(pA)
@@ -125,8 +125,6 @@ function dgemvmkl!(y::AbstractVector{Float64}, A::AbstractMatrix{Float64}, x::Ab
     ldA = stride(pA, 2) % Int32
     incx = LinearAlgebra.stride1(x) % Int32
     incy = LinearAlgebra.stride1(y) % Int32
-    α = 1.0
-    β = 0.0
     ccall(
         DGEMV_MKL, Cvoid,
         (Ref{UInt8}, Ref{Int32}, Ref{Int32}, Ref{Float64}, Ref{Float64}, Ref{Int32}, Ref{Float64}, Ref{Int32}, Ref{Float64}, Ref{Float64}, Ref{Int32}),