Reran benchmarks after switching to use OpenBLAS_jll and MKL_jll.

Author: chriselrod

benchmark/Manifest.toml

@@ -440,9 +440,7 @@ uuid = "56ddb016-857b-54e1-b83d-db4d58db5568"
 
 [[LoopVectorization]]
 deps = ["DocStringExtensions", "LinearAlgebra", "OffsetArrays", "SIMDPirates", "SLEEFPirates", "UnPack", "VectorizationBase"]
-git-tree-sha1 = "1a9a648f0aa612fdc36acb784ba47d77e8707383"
-repo-rev = "master"
-repo-url = "https://github.com/chriselrod/LoopVectorization.jl.git"
+path = "/home/chriselrod/.julia/dev/LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 version = "0.8.2"
 

benchmark/benchmarkflops.jl

@@ -157,10 +157,10 @@ function dot_bench!(br, s, i)
     @assert edot(a,b) ≈ dotblas "eigen dot wrong?"
     br[8,i] = n_gflop / @belapsed iedot($a, $b)
     @assert iedot(a,b) ≈ dotblas "i-eigen dot wrong?"
-    br[9,i] = n_gflop / @belapsed dot($a, $b)
+    # br[9,i] = n_gflop / @belapsed dot($a, $b)
 end
 function benchmark_dot(sizes)
-    tests = ["LoopVectorization", "Julia", "Clang", "GFortran", "icc", "ifort", "g++ & Eigen-3", "clang++ & Eigen-3", "OpenBLAS"]
+    tests = ["LoopVectorization", "Julia", "Clang", "GFortran", "icc", "ifort", "g++ & Eigen-3", "clang++ & Eigen-3"]#, "OpenBLAS"]
     br = BenchmarkResult(tests, sizes)
     sm = br.sizedresults.results
     pmap(is -> dot_bench!(sm, is[2], is[1]), enumerate(sizes))
@@ -186,10 +186,10 @@ function selfdot_bench!(br, s, i)
     @assert eselfdot(a) ≈ dotblas "eigen dot wrong?"
     br[8,i] = n_gflop / @belapsed ieselfdot($a)
     @assert ieselfdot(a) ≈ dotblas "i-eigen dot wrong?"
-    br[9,i] = n_gflop / @belapsed dot($a, $a)
+    # br[9,i] = n_gflop / @belapsed dot($a, $a)
 end
 function benchmark_selfdot(sizes)
-    tests = ["LoopVectorization", "Julia", "Clang", "GFortran", "icc", "ifort", "g++ & Eigen-3", "clang++ & Eigen-3", "OpenBLAS"]
+    tests = ["LoopVectorization", "Julia", "Clang", "GFortran", "icc", "ifort", "g++ & Eigen-3", "clang++ & Eigen-3"]#, "OpenBLAS"]
     br = BenchmarkResult(tests, sizes)
     sm = br.sizedresults.results
     pmap(is -> selfdot_bench!(sm, is[2], is[1]), enumerate(sizes))
@@ -284,6 +284,7 @@ function benchmark_dot3(sizes)
     pmap(is -> dot3_bench!(sm, is[2], is[1]), enumerate(sizes))
     br
 end
+BLAS.set_num_threads(1)
 function sse!(Xβ, y, X, β)
     mul!(copyto!(Xβ, y), X, β, 1.0, -1.0)
     dot(Xβ, Xβ)
@@ -317,7 +318,7 @@ function sse_bench!(br, s, i)
     br[9,i] = n_gflop / @belapsed sse!($Xβ, $y, $X, $β)
 end
 function benchmark_sse(sizes)
-    tests = ["LoopVectorization", "Julia", "Clang", "GFortran", "icc", "ifort", "g++ & Eigen-3", "clang++ & Eigen-3", "OpenBLAS"]
+    tests = ["LoopVectorization", "Julia", "Clang", "GFortran", "icc", "ifort", "g++ & Eigen-3", "clang++ & Eigen-3", BLAS.vendor() === :mkl ? "MKL" : "OpenBLAS"]
     br = BenchmarkResult(tests, sizes)
     sm = br.sizedresults.results
     pmap(is -> sse_bench!(sm, is[2], is[1]), enumerate(sizes))

benchmark/loadsharedlibs.jl

@@ -4,6 +4,13 @@ using LoopVectorization.VectorizationBase: REGISTER_SIZE
 # const LOOPVECBENCHDIR = joinpath(pkgdir(LoopVectorization), "benchmark")
 include(joinpath(LOOPVECBENCHDIR, "looptests.jl"))
 
+const LIBCTEST = joinpath(LOOPVECBENCHDIR, "libctests.so")
+const LIBFTEST = joinpath(LOOPVECBENCHDIR, "libftests.so")
+const LIBICTEST = joinpath(LOOPVECBENCHDIR, "libictests.so")
+const LIBIFTEST = joinpath(LOOPVECBENCHDIR, "libiftests.so")
+const LIBEIGENTEST = joinpath(LOOPVECBENCHDIR, "libetest.so")
+const LIBIEIGENTEST = joinpath(LOOPVECBENCHDIR, "libietest.so")
+
 
 # requires Clang with polly to build
 cfile = joinpath(LOOPVECBENCHDIR, "looptests.c")
@@ -28,12 +35,19 @@ end
 eigenfile = joinpath(LOOPVECBENCHDIR, "looptestseigen.cpp")
 if !isfile(LIBEIGENTEST) || mtime(eigenfile) > mtime(LIBEIGENTEST)
     # Clang seems to have trouble finding includes
-    run(`g++ -O3 -march=native -mprefer-vector-width=$(8REGISTER_SIZE) -DEIGEN_VECTORIZE_AVX512 -I/usr/include/eigen3 -shared -fPIC $eigenfile -o $LIBEIGENTEST`)
-    
+    if LoopVectorization.VectorizationBase.AVX512F
+        run(`g++ -O3 -march=native -mprefer-vector-width=$(8REGISTER_SIZE) -DEIGEN_VECTORIZE_AVX512 -I/usr/include/eigen3 -shared -fPIC $eigenfile -o $LIBEIGENTEST`)
+    else
+        run(`g++ -O3 -march=native -mprefer-vector-width=$(8REGISTER_SIZE) -I/usr/include/eigen3 -shared -fPIC $eigenfile -o $LIBEIGENTEST`)
+    end
 end
 if !isfile(LIBIEIGENTEST) || mtime(eigenfile) > mtime(LIBIEIGENTEST)
     # run(`/usr/bin/clang++ -Ofast -march=native -mprefer-vector-width=$(8REGISTER_SIZE) -DEIGEN_VECTORIZE_AVX512 -I/usr/include/c++/9 -I/usr/include/c++/9/x86_64-generic-linux -I/usr/include/eigen3 -shared -fPIC $eigenfile -o $LIBEIGENTEST`)
-    run(`/usr/bin/clang++ -Ofast -march=native -mprefer-vector-width=$(8REGISTER_SIZE) -DEIGEN_VECTORIZE_AVX512 -I/usr/include/eigen3 -shared -fPIC $eigenfile -o $LIBEIGENTEST`)
+    if LoopVectorization.VectorizationBase.AVX512F
+        run(`/usr/bin/clang++ -Ofast -march=native -mprefer-vector-width=$(8REGISTER_SIZE) -DEIGEN_VECTORIZE_AVX512 -I/usr/include/eigen3 -shared -fPIC $eigenfile -o $LIBEIGENTEST`)
+    else
+        run(`/usr/bin/clang++ -Ofast -march=native -mprefer-vector-width=$(8REGISTER_SIZE) -I/usr/include/eigen3 -shared -fPIC $eigenfile -o $LIBEIGENTEST`)
+    end
     # run(`icpc -fast -qopt-zmm-usage=high -fargument-noalias-global -qoverride-limits -I/usr/include/eigen3 -shared -fPIC $eigenfile -o $LIBIEIGENTEST`)
 end
 
@@ -47,13 +61,6 @@ end
 # end
 
 
-const LIBCTEST = joinpath(LOOPVECBENCHDIR, "libctests.so")
-const LIBFTEST = joinpath(LOOPVECBENCHDIR, "libftests.so")
-const LIBICTEST = joinpath(LOOPVECBENCHDIR, "libictests.so")
-const LIBIFTEST = joinpath(LOOPVECBENCHDIR, "libiftests.so")
-const LIBEIGENTEST = joinpath(LOOPVECBENCHDIR, "libetest.so")
-const LIBIEIGENTEST = joinpath(LOOPVECBENCHDIR, "libietest.so")
-
 using MKL_jll, OpenBLAS_jll
 
 const libMKL = Libdl.dlopen(MKL_jll.libmkl_rt)
@@ -105,9 +112,9 @@ function dgemmopenblas!(C::AbstractMatrix{Float64}, A::AbstractMatrix{Float64},
         transA, transB, M, N, K, α, pA, ldA, pB, ldB, β, C, ldC
     )
 end
-mkl_set_num_threads(N::Integer) = ccall(MKL_SET_NUM_THREADS, Cvoid, (Ref{UInt32},), Ref(N % UInt32))
+mkl_set_num_threads(N::Integer) = ccall(MKL_SET_NUM_THREADS, Cvoid, (Int32,), N % Int32)
 mkl_set_num_threads(1)
-openblas_set_num_threads(N::Integer) = ccall(OPENBLAS_SET_NUM_THREADS, Cvoid, (Ref{Int64},), Ref(N))
+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})
     transA = istransposed(A)
@@ -123,7 +130,7 @@ function dgemvmkl!(y::AbstractVector{Float64}, A::AbstractMatrix{Float64}, x::Ab
     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}),
-        transA, M32, N32, α, A, ldA, x, incx, β, y, incy
+        transA, M32, N32, α, pA, ldA, x, incx, β, y, incy
     )
 end
 function dgemvopenblas!(y::AbstractVector{Float64}, A::AbstractMatrix{Float64}, x::AbstractVector{Float64})
@@ -138,7 +145,7 @@ function dgemvopenblas!(y::AbstractVector{Float64}, A::AbstractMatrix{Float64},
     ccall(
         DGEMV_OpenBLAS, Cvoid,
         (Ref{UInt8}, Ref{Int64}, Ref{Int64}, Ref{Float64}, Ref{Float64}, Ref{Int64}, Ref{Float64}, Ref{Int64}, Ref{Float64}, Ref{Float64}, Ref{Int64}),
-        transA, M, N, α, A, ldA, x, incx, β, y, incy
+        transA, M, N, α, pA, ldA, x, incx, β, y, incy
     )
 end
 

benchmark/looptests.c

@@ -248,8 +248,8 @@ void filter2d(double* restrict B, double* restrict A, double* restrict K, long M
   for (long na = offset; na < N-offset; na++){
     for (long ma = offset; ma < M-offset; ma++){
       double tmp = 0.0;
-      for (long mk = -offset; mk < offset + 1; mk++){
-	for (long nk = -offset; nk < offset + 1; nk++){
+      for (long nk = -offset; nk < offset + 1; nk++){
+	for (long mk = -offset; mk < offset + 1; mk++){
 	  tmp += A[(ma+mk) + (na+nk)*M] * K[(mk+offset) + (nk+offset)*(2*offset+1)];
 	}
       }
@@ -262,8 +262,8 @@ void filter2d3x3(double* restrict B, double* restrict A, double* restrict K, lon
   for (long na = offset; na < N-offset; na++){
     for (long ma = offset; ma < M-offset; ma++){
       double tmp = 0.0;
-      for (long mk = -offset; mk < offset + 1; mk++){
-	for (long nk = -offset; nk < offset + 1; nk++){
+      for (long nk = -offset; nk < offset + 1; nk++){
+	for (long mk = -offset; mk < offset + 1; mk++){
 	  tmp += A[(ma+mk) + (na+nk)*M] * K[(mk+offset) + (nk+offset)*(2*offset+1)];
 	}
       }

benchmark/looptests.f90

@@ -324,10 +324,8 @@ subroutine filter2d(B, A, K, Ma, Na, offset) BIND(C, name="filter2d")
       real(C_double) :: tmp
       do concurrent(mma = 1+offset:Ma-offset, nna = 1+offset:Na-offset)
          tmp = 0
-         do mmk = -offset,offset
-            do nnk = -offset,offset
-               tmp = tmp + A(mma + mmk, nna + nnk) * K(mmk, nnk)
-            end do
+         do concurrent(mmk = -offset:offset, nnk = -offset:offset)
+            tmp = tmp + A(mma + mmk, nna + nnk) * K(mmk, nnk)
          end do
          B(mma,nna) = tmp
       end do
@@ -342,10 +340,8 @@ subroutine filter2d3x3(B, A, K, Ma, Na) BIND(C, name="filter2d3x3")
       real(C_double) :: tmp
       do concurrent(mma = 1+offset:Ma-offset, nna = 1+offset:Na-offset)
          tmp = 0
-         do mmk = -offset,offset
-            do nnk = -offset,offset
-               tmp = tmp + A(mma + mmk, nna + nnk) * K(mmk, nnk)
-            end do
+         do concurrent(mmk = -offset:offset, nnk = -offset:offset)
+            tmp = tmp + A(mma + mmk, nna + nnk) * K(mmk, nnk)
          end do
          B(mma,nna) = tmp
       end do

benchmark/plotbenchmarks.jl

@@ -35,8 +35,11 @@ function Gadfly.plot(br::BenchmarkResult)
     # sizes = Vector{eltype(brsizes)}(undef, length(res))
     tests = replace_and.(@view(br.tests[2:end]))
     colors = getcolor.(tests)
+    addlabel = false
 
-    xt = 0:20:260
+    maxxval, maxxind = findmax(sizes)
+    maxxtick = 10cld(maxxval,10) + (addlabel ? 20 : 0)
+    xt = 0:20:maxxtick
     maxres = maximum(res)
     maxtick = 10round(Int, 0.1maxres)
     yt = if iszero(maxtick)
@@ -59,6 +62,7 @@ function Gadfly.plot(br::BenchmarkResult)
     for i ∈ eachindex(tests)
         push!(p, layer(x = sizes, y = res[i,:], Geom.line, Theme(default_color=colors[i])))
     end
+    addlabel && push!(p, layer(x = fill(maxxtick - 10, length(tests)), y = res[:,maxxind], label=tests, Geom.label(position=:centered)))
     p
 end