Linking Intel's Math Kernel Library (MKL) to R

问题:

Using an alternative BLAS for R has several advantages, see e.g. https://cran.r-project.org/web/packages/gcbd/vignettes/gcbd.pdf.

Microsoft R Open https://mran.revolutionanalytics.com/documents/rro/installation/#sysreq is using Intel's MKL instead of the default Reference BLAS to speed up calculations.

My question is:

What would be the exact steps to link Intel's MKL library **manually to R**'s most recent version on Windows (https://cran.r-project.org/bin/windows/base/)?

---

UPDATE 20-07-2016: Here is very detailed description on how to build a OpenBLAS-based Rblas.dll for 64-bit R for Windows for R ≥ 3.3.0: http://www.avrahamadler.com/r-tips/build-openblas-for-windows-r64/

回答1:

I was able to link R 3.6.0 with custom dlls you create using the builder. Basically you have to export the same symbols Rblas.dll and Rlapack.dll do. Start the Compiler 19.0 Update 4 for Intel 64 Visual Studio 2017 environment command prompt:

Get the symbols:

dumpbin /exports Rblas.dll > Rblas_list
dumpbin /exports Rlapack.dll > Rlapack_list_R

Edit both files deleting the "header" and "footer" and have all the lines with the symbol names (ex.: 248 F7 00138CE0 dgeevx_) be like dgeevx_ (only with the names). Copy the builder directory to somewhere in your pc and inside it run:

# blas links fine
nmake libintel64 export=..path..\Rblas_list name=Rblas 
# save lapack errors in another list
nmake libintel64 export=..path..\Rlapack_list_R name=Rlapack 1> undefined_symbols_list

Edit undefined_symbols_list keep only the names in each line and create a new list with the difference

findstr /v /g:undefined_symbols_list Rlapack_list_R > Rlapack_list
nmake libintel64 export=..path..\Rlapack_list name=Rlapack

With dumpbin /dependents Rlapack.dll, you can see that they depend on libiomp5md.dll, which you can find inside the redist folder in mkl installation.

---

Method 2

This method uses more disk space, but it's simpler. Copy all the contents from inside these folders

C:\Program Files (x86)\IntelSWTools\compilers_and_libraries\windows\redist\intel64\mkl
C:\Program Files (x86)\IntelSWTools\compilers_and_libraries\windows\redist\intel64\compiler

to

C:\Program Files\R\R-3.6.1\bin\x64

Inside the destination folder, create 2 copies of mkl_rt.dll and rename one of them Rblas.dll and the other Rlapack.dll replacing the originals and also keeping mkl_rt.dll.

回答2:

Easier solution than having to recompile R against the Intel MKL libraries on Windows is just to

1. Install Microsoft R Open from https://mran.microsoft.com/download, which comes with the outdated R version 3.5.3 but also with the Intel MKL multithreaded BLAS libraries
2. Install the latest version of R from https://cran.r-project.org/bin/windows/base/, i.e. currently R 3.6.2
3. copy files libiomp5md.dll, Rblas.dll and Rlapack.dll from C:\Program Files\Microsoft\R Open\R-3.5.3\bin\x64 to C:\Program Files\R\R-3.6.2\bin\x64 (you can back up your existing default non-hyperthreaded Rblas.dll and Rlapack.dll files first if you like)
4. copy Microsoft R Open libraries/packages MicrosoftR, RevoIOQ, RevoMods, RevoUtils, RevoUtilsMath and doParallel from C:\Program Files\Microsoft\R Open\R-3.5.5\library to your default package directory, e.g. C:\Documents\R\win-library\3.6
5. copy files Rprofile.site and Renviron.site from directory C:\Program Files\Microsoft\R Open\R-3.5.5\etc to C:\Progral Files\R\R-3.6.2\etc
6. replace line 24 in file Rprofile.site options(repos=r) with options(repos="https://cran.rstudio.com") (or your favourite CRAN repository - you can also use "https://cran.revolutionanalytics.com", the MRO repository that has the latest daily builds of all packages) to make sure that it will install the latest CRAN packages as opposed to the outdated mran.microsoft.com mirror that has outdated package versions, frozen at the 15th of April 2019. Also comment out lines 153, 154 and 155 with a #

Then restart RStudio to check that it works, with small SVD benchmark on my Intel Core i7-4700HQ 2.4GHz 4 core/8 thread laptop:

getMKLthreads()
4

# Singular Value Decomposition
m <- 10000
n <- 2000
A <- matrix (runif (m*n),m,n)
system.time (S <- svd (A,nu=0,nv=0))
   user  system elapsed 
  15.20    0.64    4.17

That same benchmark without Intel MKL installed ran at

   user  system elapsed 
  35.11    0.10   35.21 

so we get a >8 fold speed increase here!

Screenshot of Microsoft R Open 6.2 with Intel MKL up and running:

Alternatively, if you don't like copying files from MRO to your latest R installation, you can also copy the files from the free Intel MKL installation to your R installation to get multithreaded operation (as outlined in the other answer below):

1. Install Intel MKL from https://software.intel.com/en-us/mkl/choose-download (free)

Copy all the contents from inside these folders

C:\Program Files (x86)\IntelSWTools\compilers_and_libraries\windows\redist\intel64\mkl
C:\Program Files (x86)\IntelSWTools\compilers_and_libraries\windows\redist\intel64\compiler

to

C:\Program Files\R\R-3.6.1\bin\x64

2. Inside the destination folder, create 2 copies of mkl_rt.dll and rename one of them Rblas.dll and the other Rlapack.dll replacing the originals and also keeping mkl_rt.dll.
3. This will not provide you with the function setMKLthreads() and getMKLthreads() functions though to set the nr of intel MKL threads, as these come with the MRO package RevoUtilsMath. But for most people the default nr of threads set equal to the nr of physical cores will be OK though...

Not sure what's up with Microsoft, and why they are no longer updatig MRO... And why they also dropped Mac OS X support...

I hope that, given that Intel MKL is free now, the R core people will sooner or later provide a precompiled R version that is compiled to use the Intel MKL libs, or possibly detect at run time if Intel MKL is installed, and if it is, use it. I think this is important, especially since the easy availability of a good multithreaded BLAS also determines how one would develop packages - e.g. if a good multithreaded BLAS would be available for all OSes one would veer towards using RcppArmadillo, which falls back on whatever BLAS one has installed (but on Windows would give drastically worse timings if Intel MKL is not installed), and if not RcppEigen would be the best option, as that has its own multithreaded matrix algebra, irrespective of the BLAS against which R is compiled......

On Ubuntu btw it's very easy to make R use Intel MKL, without having to recompile R, as outlined here: https://github.com/eddelbuettel/mkl4deb

PS Slight problem is that running setMKLthreads(4) will crash RStudio (that was already a problem in the official MRO 3.5.3 though) but it does work OK in the R console...

回答3:

just tried for R 3.5.1 installation. I installed Microsoft R Open alongside with the CRAN R and copy libiomp5md.dll and overwrite Rblas.dll, Rlapack.dll from MRO MKL counterparts to link to CRAN R on Windows (similar to another answer above but need to copy the file libiomp5md.dll as well). This worked out fine and the CRAN R runs as fast as MRO according to the version.compare package on Github (https://github.com/andrie/version.compare)