On 4/04/2020, at 1:00 AM, Patrick Schratz <patrick.schratz at gmail.com> wrote:
Simon,
I don?t understand fully. I am using llvm for all C variants (just now shown) in combination with the 10.13 SDK.
So far this combination works flawlessly for all ?problematic? packages like data.table, igraph or Rcpp.
I don?t have deeper knowledge about the ?iomp? setup but as of right now I don?t know what I am mixing up here. Can you shine more light on this?
If it is about llvm in general: data.table recommends to use llvm openly in their wiki due to the lack of openMP for the standard clang.
And while this could be handled as ?any other package? in R, we all know that it has quite some impact and probably devs who know what they do in terms of C configs.
Regarding the SDK issue: Both Rcpp v1.0.4 and igraph 1.2.5 can?t be installed from source with SDK 10.15 and standard apple clang (or any C compiler). Switching to SDK 10.13 solves the issue. You were even commenting on the Rcpp issue.
Since SDK 10.15 is the current release and many users have no choice on which OS version they are (for different reasons), this issue should imo be inspected in more detail. What do you think?
Best, Patrick
On 2. Apr 2020, 23:38 +0200, Simon Urbanek <simon.urbanek at r-project.org>, wrote:
Patrick,
you can't mix compilers - it really matters which iomp your'e using. llvm has its own modified version which may not be the same as the official Intel release. From your tests it looks like you're using the llvm one which will likely work only with that compiler. Since Apple doesn't have official omp support it's hard to say which versions work and which don't.
On 3/04/2020, at 10:20 AM, Patrick Schratz <patrick.schratz at gmail.com> wrote:
Thanks Kevin, interesting approach.
However, when testing it against a few packages I get symbol pointer issues (e.g. for data.table and xgboost).
Using llvm everything works.
Could llvm be the best middle way? Easy installation via brew and still clang compliant.
Currently my Makevars look as follows
CFLAGS=-isysroot /Library/Developer/CommandLineTools/SDKs/MacOSX10.13.sdk
[?]
CC = ccache /usr/local/opt/llvm/bin/clang
[?]
Where llvm was installed via `brew install llvm`.
SDK 10.13 because of {igraph} and {Rcpp} issues with SDK 10.14 and SDK 10.15
I mentioned that before, but I do not see issues with 10.14 SDK.
Cheers,
Simon
On 2. Apr 2020, 23:01 +0200, Simon Urbanek <simon.urbanek at r-project.org>, wrote:
On 3/04/2020, at 2:01 AM, BATES Timothy <tim.bates at ed.ac.uk> wrote:
Moving to a compiler that drops support for OpenMP seems a sad choice, especially now we?ve all climbed the learning curve of the non-Apple compiler (the real barrier was lack of a pkg installer and that?s done now).
It has caused a lot of issues, it trips people on a daily basis which is just not worth it. Also with Apple's increasingly strict rules about what can be distributed we don't want to be in the business in maintaining a separate toolchain. There have been numerous issues with C++ exceptions so the fall out was much bigger than originally expected and it would only get worse.
Losing OpenMP for the CRAN version of OpenMx/umx (our SEM packages) would be a big loss for users (for whom the CRAN version now supports OpenMP giving them a 2-12x speed up). In general, R on Mac is made more viable by having OpenMP
Re Brian?s points, I?d say that the distribution problem is crucial: Packages not on CRAN have dramatically diminished accessibility/useage.
The idea is that if a package deems this critical, it can enable this for the users. As it is now, packages can still supply iomp and use it.
That said, I would open for discussion the ability to distribute iomp with the R binary, but it would not be supported by R directly, i.e., it would be on the package author to make sure that the way the package operates is compatible with that binary. Let me know what you think.
Second, a great range of compute-intensive problems are amenable to division amongst cores, including nearly all models that take more than a nominal amount of time to run: So simulations, CIs, bootstrapping, nearly everything in genetics all speeds up.
Yes, but OpenMP is rarely used for those tasks. In R OpenMP can be only used for very small subset of such tasks - which is why alternative approaches are much more common.
Cheers,
Simon
I?d say especially on desktop/laptop. The big advantage of multi blade systems requires snowfall-type solutions, but desktops profit automatically from their multi-core structure and don;?t have multiple processors (except graphics, which no-one seems to be exploiting on CRAN-style R), so OpenMP is their one trick. I?d hope not to lose it.
Best, t
On 2 Apr 2020, at 05:18, Prof Brian Ripley <ripley at stats.ox.ac.uk> wrote:
On 01/04/2020 22:02, Simon Urbanek wrote:
JJB,
1. correct, there was too much trouble in this. But please feel free to start a new thread about this here if you have strong opinions.
Also note that it is possible (and not hard) to install packages from source with an OpenMP-supporting compiler, and how to do so is in the R-admin manual. The problems come in distributing them.
The benefits of OpenMP are often overestimated, especially on desktop/laptop level hardware. But it is available for the small (tiny?) proportion of users who need it.
The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.