Proprietary LLVM-IR Vectorizer
NEC has released its proprietary vectorizer as a black box binary pluggable into LLVM. With some limitations, though, but following almost all compiler options and compiler directoves described in the ncc/nc++ compier manual.
The vectorizer is called llvm-vec and is supposed to be called on LLVM-IR. The NEC compiler people have worked on this since quite a while and have released several versions. The best one so far is somewhat hidden in the official yum repositories for the SX-Aurora Tsubasa: https://sxauroratsubasa.sakura.ne.jp/repos/additional/llvm-vec/llvm-vec_el/2.3.0/
The announcement for llvm-vec was posted in the NEC forum here.
llvm-vec needs a very heavily patched version of clang 12.0.0 which knows about all the new compiler options and directives. I have prepared a llvm package based on llvm-12.0.0 with the necessary patches here.
The combination of the two packages llvm-vec-2.3.0 (the patched LLVM 12.0.0) and nec-llvm-vec-2.3 (the proprietary vectorizer module) leads to a highly efficient freely usable vectorizing LLVM C/C++ compiler which is pretty much equivalent to teh proprietary NEC Compiler Collection ncc/nc++.
Compiling C or C++ code with
clang requires adding some options:
-target ve-linuxspecifies that the compiled code is for the vectore engine VE,
-fnvecenables the NEC VE IR vectorizer,
-fdiag-vector=2outupts more details on the vectorization process.
Very many of the ncc compiler options have been added for supporting ‘llvm-vec’, which can vectorize LLVM-IR when called separately: https://sxauroratsubasa.sakura.ne.jp/documents/llvm-vec/en/LLVMVecCommand.html
A large number of vectorization related pragmas are supported in exactly the same format
as used in the proprietary compiler
ncc/nc++ documented here.
The pragmas are written in the format
#pragma _NEC .... Supported pragmas are:
- move / move_unsafe / nomove
In order to simplify using the llvm-vec compiler I created a docker image including all necessary packages, the 12.0.0 LLVM as well as the nec-llev-vec vectorizer module. The project is available at https://github.com/efocht/llvm-vec-docker .
The docker image has been pushed to dockerhub under the name/tag
You can build the docker image on your own as follows:
git clone https://github.com/efocht/llvm-vec-docker.git docker build --tag efocht/llvm-vec:2.3.0 .
I had to add
--network host when building, but this should normally
not be necessary.
Run the docker image for example by invoking the script included into
It will mount the current user’s home directory into the image, as well as the /etc/passwd and /etc/group files (in order to provide the proper user IDs for the mounted home file system). This might lead to UID mismatches inside the container, but we’re not expecting the user to run any system services inside the container.