[R-pkg-devel] using portable simd instructions

[Tomas Kalibera]

On 3/26/24 10:53, jesse koops wrote:
> Hello R-package-devel,
>
> I recently got inspired by the rcppsimdjson package to try out simd
> registers. It works fantastic on my computer but I struggle to find
> information on how to make it portable. It doesn't help in this case
> that R and Rcpp make including Cpp code so easy that I have never had
> to learn about cmake and compiler flags. I would appreciate any help,
> including of the type: "go read instructions at ...".
>
> I use RcppArmadillo and Rcpp. I currenlty include the following header:
>
> #include <immintrin.h>
>
> The functions in immintrin that I use are:
>
> _mm256_loadu_pd
> _mm256_set1_pd
> _mm256_mul_pd
> _mm256_fmadd_pd
> _mm256_storeu_pd
>
> and I define up to four __m256d registers. From information found
> online (not sure where anymore) I constructed the following makevars
> file:
>
> CXX_STD = CXX14
>
> PKG_CPPFLAGS = -I../inst/include -mfma -msse4.2 -mavx
>
> PKG_CXXFLAGS = $(SHLIB_OPENMP_CXXFLAGS)
> PKG_LIBS = $(SHLIB_OPENMP_CXXFLAGS) $(LAPACK_LIBS) $(BLAS_LIBS) $(FLIBS)
>
> (I also use openmp, that has always worked fine, I just included all
> lines for completeness) Rcheck gives me two notes:
>
> ─  using R version 4.3.2 (2023-10-31 ucrt)
> ─  using platform: x86_64-w64-mingw32 (64-bit)
> ─  R was compiled by
>         gcc.exe (GCC) 12.3.0
>         GNU Fortran (GCC) 12.3.0
>
> ❯ checking compilation flags used ... NOTE
>    Compilation used the following non-portable flag(s):
>      '-mavx' '-mfma' '-msse4.2'
>
> ❯ checking C++ specification ... NOTE
>      Specified C++14: please drop specification unless essential
>
> But as far as I understand, the flags are necessary, at least in GCC.
> How can I make this portable and CRAN-acceptable?

I think it the best way for portability is to use a higher-level library 
that already has done the low-level business of maintaining multiple 
versions of the code (with multiple instruction sets) and choosing one 
appropriate for the current CPU. It could be say LAPACK, BLAS, openmp, 
depending of the problem at hand. In some cases, code can be rewritten 
so that the compiler can vectorize it better, using the level of 
vectorized instructions that have been enabled.

Unconditionally using GCC-specific or architecture-specific options in 
packages would certainly not be portable. Even on Windows, R is now used 
also with clang and on aarch64, so one should not assume a concrete 
compiler and architecture.

Please note also that GCC on Windows has a bug due to which AVX2 
instructions cannot be used reliably - the compiler doesn't always 
properly align local variables on the stack when emitting these. See 
[1,2] for more information.

Best
Tomas

[1] https://stat.ethz.ch/pipermail/r-sig-windows/2024q1/000113.html
[2] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412

>
> kind regards,
> Jesse
>
> ______________________________________________
> R-package-devel using r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-package-devel