Home - News ( United States | United Kingdom | Italy | Germany ) - Football scores

Showing content from https://github.com/JuliaPackaging/BinaryBuilder.jl/issues/700 below:

Conflict between MKL_jll and libgomp in CompilerSupportLibraries_jll · Issue #700 · JuliaPackaging/BinaryBuilder.jl · GitHub

Today I was tracking down the test failure observed in JuliaMath/IntelVectorMath.jl#39 when moving to MKL_jll that occurred on a Mac.

After a while, with some help from @giordano, an MWE is:

# dlopen the libgomp inside CompilerSupportLibraries_jll (libgfortran 5). 
using Libdl
dlopen("/Users/kristoffercarlsson/.julia/artifacts/411c17f069dd12b3f79b19e4647949050a58223c/lib/libgomp.1.dylib")

using MKL_jll

gamma1(x) = (x = copy(x);    ccall((:vdTGamma, libmkl_rt), Cvoid, (Int32, Ptr{Float64}, Ptr{Float64}), length(x), x, x); x);
gamma2(x) = (y = similar(x); ccall((:vdTGamma, libmkl_rt), Cvoid, (Int32, Ptr{Float64}, Ptr{Float64}), length(x), x, y); y);
INPUT_2  = rand(1000);

gamma1(INPUT_2)[1:3]
gamma2(INPUT_2)[1:3]

the result from gamma1 and gamma2 should give the same answers, but they don't, as an example

julia> gamma1(INPUT_2)[1:3]
3-element Array{Float64,1}:
 Inf
 Inf
 Inf

julia> gamma2(INPUT_2)[1:3]
3-element Array{Float64,1}:
 16.209046961478283
  6.570129954697695
 41.08264424860374

(This doesn't seem to happen 100% of the time but the majority of the times).

This only happens if the length of the input vector is a certain size, cf

julia> gamma1(INPUT_2[1:602])[1:3] # ok
3-element Array{Float64,1}:
 1.0650782774656853
 1.4718294693439533
 1.1884483595544906

julia> gamma2(INPUT_2[1:602])[1:3] # ok
3-element Array{Float64,1}:
 1.0650782774656853
 1.4718294693439533
 1.1884483595544906

julia> gamma1(INPUT_2[1:603])[1:3] # bad
3-element Array{Float64,1}:
 0.9663910468243027
 0.8856476139287297
 0.9213177809196278

julia> gamma2(INPUT_2[1:603])[1:3] # ok
3-element Array{Float64,1}:
 1.0650782774656853
 1.4718294693439533
 1.1884483595544906

The cutoff seems to be when the length of the vector is 603. My guess is that MKL thinks this is a long enough vector to perhaps enable threading? Also, it is important that libgomp is loaded before MKL_jll for the bug to occur.

The versions used are:

  [e66e0078] CompilerSupportLibraries_jll v0.2.0+1
  [856f044c] MKL_jll v2020.0.166+0

and

julia> versioninfo()
Julia Version 1.4.0-rc2.0
Commit b99ed72c95 (2020-02-24 16:51 UTC)
Platform Info:
  OS: macOS (x86_64-apple-darwin19.0.0)
  CPU: Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-8.0.1 (ORCJIT, skylake)

RetroSearch is an open source project built by @garambo | Open a GitHub Issue

Search and Browse the WWW like it's 1997 | Search results from DuckDuckGo

HTML: 3.2 | Encoding: UTF-8 | Version: 0.7.4