A few benchmark bug fixes.

Author: chriselrod

benchmarks/driver.jl

@@ -129,7 +129,7 @@ function sse!(Xβ, y, X, β)
     dot(Xβ, Xβ)
 end
 function benchmark_sse(sizes)
-    tests = [BLAS.vendor() === :mkl ? "IntelMKL" : "OpenBLAS", "Julia", "Clang-Polly", "GFort-loops", "GFort-intrinsic", "LoopVectorization"]
+    tests = [BLAS.vendor() === :mkl ? "IntelMKL" : "OpenBLAS", "Julia", "Clang-Polly", "GFort-loops", "LoopVectorization"]
     br = BenchmarkResult(tests, sizes)
     for (i,s) ∈ enumerate(sizes)
         N, P = totwotuple(s)
@@ -156,7 +156,7 @@ function benchmark_exp(sizes)
     br = BenchmarkResult(tests, sizes)
     for (i,s) ∈ enumerate(sizes)
         a = rand(s); b = similar(a)
-        n_gflop = s # not really gflops
+        n_gflop = 1e-9*s # not really gflops
         br[1,i] = n_gflop / @belapsed @. $b = exp($a)
         baseb = copy(b)
         br[2,i] = n_gflop / @belapsed fvexp!($b, $a)

benchmarks/loadsharedlibs.jl

@@ -113,16 +113,16 @@ function caplusBc!(D, a, B, c)
     M, K = size(B)
     ccall(
         (:aplusBc, LIBCTEST), Cvoid,
-        (Ptr{Float64}, Ptr{Float64}, Ptr{Float64}, Clong, Clong),
-        y, A, x, M, K
+        (Ptr{Float64}, Ptr{Float64}, Ptr{Float64}, Ptr{Float64}, Clong, Clong),
+        D, a, B, c, M, K
     )
 end
 function faplusBc!(D, a, B, c)
     M, K = size(B)
     ccall(
         (:aplusBc, LIBFTEST), Cvoid,
-        (Ptr{Float64}, Ptr{Float64}, Ptr{Float64}, Ref{Clong}, Ref{Clong}),
-        y, A, x, Ref(M), Ref(K)
+        (Ptr{Float64}, Ptr{Float64}, Ptr{Float64}, Ptr{Float64}, Ref{Clong}, Ref{Clong}),
+        D, a, B, c, Ref(M), Ref(K)
     )
 end
 function cOLSlp(y, X, β)

benchmarks/looptests.c

@@ -118,7 +118,7 @@ void unscaledvar(double* restrict s, double* restrict A, double* restrict xb, lo
   return;
 }
 
-void aplucBc(double* restrict D, double* restrict a, double* restrict B, double* restrict c, long M, long N){
+void aplusBc(double* restrict D, double* restrict a, double* restrict B, double* restrict c, long M, long N){
   for (long n = 0; n < N; n++){
     for (long m = 0; m < M; m++){
       D[m + n*M] = a[m] + B[m + n*M] * c[n];

benchmarks/looptests.f90

@@ -176,7 +176,7 @@ subroutine unscaledvar(s, A, x, M, N) BIND(C, name="unscaledvar")
          end do
       end do
     end subroutine unscaledvar
-    subroutine aplusBc(D, a, B, c, M, N) BIND(C, name="aplucBc")
+    subroutine aplusBc(D, a, B, c, M, N) BIND(C, name="aplusBc")
       integer(C_long), intent(in) :: M, N
       real(C_double), intent(in) :: a(M), B(M,N), c(N)
       real(C_double), dimension(M,N), intent(out) :: D

benchmarks/looptests.jl

@@ -113,9 +113,9 @@ japlucBcavx!(d, a, B, c) = @avx @. d = a + B * c';
 
 function jOLSlp(y, X, β)
     lp = 0.0
-    @inbounds for i ∈ eachindex(y)
+    @inbounds @fastmath for i ∈ eachindex(y)
         δ = y[i]
-        @simd for j ∈ eachindex(x)
+        @simd for j ∈ eachindex(β)
             δ -= X[i,j] * β[j]
         end
         lp += δ * δ