Don't use MKL except on x86_64, fixes #73. (#75)

Author: chriselrod

src/BLASBenchmarksCPU.jl

@@ -1,7 +1,10 @@
 module BLASBenchmarksCPU
 
 # BLAS libs (& Libdl)
-using MKL_jll, OpenBLAS_jll, blis_jll#, Libdl
+@static if Sys.ARCH === :x86_64
+using MKL_jll
+end
+using OpenBLAS_jll, blis_jll#, Libdl
 # Julia BLAS
 using Tullio, Octavian, Gaius
 
@@ -45,10 +48,12 @@ export gemmmkl!, gemmmkl_direct!
 export mkl_set_num_threads
 
 # set threads
+@static if Sys.ARCH === :x86_64
 const libMKL = MKL_jll.libmkl_rt # more convenient name
 function mkl_set_num_threads(N::Integer)
     ccall((:MKL_Set_Num_Threads,libMKL), Cvoid, (Int32,), N % Int32)
 end
+end
 const libOPENBLAS = OpenBLAS_jll.libopenblas # more convenient name
 function openblas_set_num_threads(N::Integer)
     ccall((:openblas_set_num_threads64_,libOPENBLAS), Cvoid, (Int64,), N)
@@ -69,7 +74,7 @@ include("runbenchmark.jl")
 include("plotting.jl")
 
 function __init__()
-    mkl_set_num_threads(num_cores())
+    Sys.ARCH === :x86_64 && mkl_set_num_threads(num_cores())
     openblas_set_num_threads(num_cores())
     blis_set_num_threads(num_cores())
 end

src/ccallblas.jl

@@ -54,6 +54,7 @@ for (name,BlasInt,suff) ∈ [
   ("openblas", :Int64, "_64_"),
   ("blis", :Int64, "_64_")
 ]
+  (Sys.ARCH !== :x86_64 && name === "mkl") && continue
   uname = uppercase(name)
   lib = Symbol("lib", uname)
   fgemm = Symbol("gemm", name, '!')
@@ -151,6 +152,7 @@ for (name,BlasInt,suff) ∈ [
 
   end
 end
+@static if Sys.ARCH === :x86_64
 let BlasInt = :Int32
   for (T,prefix) ∈ [(:Float32,'s'),(:Float64,'d')]
     f = Symbol(prefix, "gemm_direct")
@@ -176,6 +178,7 @@ let BlasInt = :Int32
     end
   end
 end
+end
 
 
 

src/runbenchmark.jl

@@ -201,12 +201,12 @@ function all_libs()
     libs = Symbol[
         :BLIS,
         :Gaius,
-        :MKL,
         :Octavian,
         :OpenBLAS,
         :Tullio,
         :LoopVectorization
     ]
+    Sys.ARCH === :x86_64 && push!(libs, :MKL)
     return libs
 end
 
@@ -276,11 +276,11 @@ function runbench(
   sleep_time = 0.0
 ) where {T}
   if threaded
-    mkl_set_num_threads(num_cores())
+    Sys.ARCH === :x86_64 && mkl_set_num_threads(num_cores())
     openblas_set_num_threads(num_cores())
     blis_set_num_threads(num_cores())
   else
-    mkl_set_num_threads(1)
+    Sys.ARCH === :x86_64 && mkl_set_num_threads(1)
     openblas_set_num_threads(1)
     blis_set_num_threads(1)
   end
@@ -355,7 +355,11 @@ function runbench(
   BenchmarkResult{T}(libs, sizes, gflop, times, threaded)
 end
 
+@static if Sys.ARCH === :x86_64
 const LUFUNCS = Dict(:RecursiveFactorization => RecursiveFactorization.lu!, :MKL => lumkl!, :OpenBLAS => luopenblas!)
+else
+const LUFUNCS = Dict(:RecursiveFactorization => RecursiveFactorization.lu!, :OpenBLAS => luopenblas)
+end
 struct LUWrapperFunc{F}; f::F; end
 (lu::LUWrapperFunc)(A,B,C) = lu.f(copyto!(A,B))
 function runlubench(