Getting started: Setup LLVM-VE

LLVM-VE is a community supported open source compiler and not included in the official SDK for SX-Aurora TSUBASA. First of all you have to install LLVM-VE. The LLVM-VE rpm packages are released at our branch of llvm. Note that the official llvm repository also has VE backend, but it does not yet support VE intrinsics at the time of writing. If you prefer a yum repository, see https://www.sx-aurora.com/repos/llvm/x86_64/.

Install LLVM-VE.

% yum install llmv-ve-<version>.rpm llvm-ve-link-<version>.rpm

LLVM-VE is installed into /opt/nec/nosupport/llvm-ve-<version> and latest version is linked from /opt/nec/nosupport/llvm-ve. The compiler can be invoked as below.

% /opt/nec/nosupport/llvm-ve/clang -target ve-linux -O3 -c func.c

Because current LLVM-VE does not support automatic vectorization, we recommend to use llvm-ve with ncc/nc++/nfort, default compilers for VE. You typically create the dedicated source files for llvm-ve that includes minimum code with intrinsics. Other source files should be compiled by ncc, then linked by ncc. For example,

% /opt/nec/nosupport/llvm-ve/clang -target ve-linux -O3 -c intrinsics.c
% /opt/nec/ve/bin/ncc -c main.c
% /opt/nec/ve/bin/ncc -o a.out intrinsics.o main.o

OpenMP

LLVE-VE supports OpenMP but its runtime is different from ncc's runtime. We strongly recommend to use ncc's OpenMP. In other word, source code with intrinsics should not include OpenMP.