kernels need to be compiled with NEC’s proprietary compiler.

[freemin7]

Your publication neoSYCL: a SYCL implementation for SX-Aurora TSUBASA reads

neoSYCL needs to separate the host and kernel codes so that they can be compiled with different
compilers. In the case of SX-Aurora, kernels need to be compiled with NEC’s proprietary compiler.

With the efforts of an open source LLVM based compiler would it be possible to add support for a non NEC compiler (probably at worse performance)?
If this is not possible with your current staffing, what would be involved in getting that working?

[keichi]

Hi, NEC’s compiler is not required if you don’t run on the Vector Engine. Please use the CPU device or the host device instead of the VE device (this is how we run the tests on GitHub CI).

[freemin7]

I am sorry for causing a misunderstanding. I want to run on a VectorEngine but not use the NEC compiler and instead use the LLVM based compiler linked above.

[keichi]

I see. We have not tested with LLVM-VE, but neoSYCL does not rely on a specific feature of NEC's compiler.

In the case of SX-Aurora, kernels need to be compiled with NEC’s proprietary compiler.

What we meant by this sentence is simply that the compiler needs to emit VE binary. Please try setting the environment variable DEVICE_COMPILER to your installation of LLVM-VE clang++ when you invoke nsc++.

[freemin7]

I will get back to you and perform those steps.

[efocht]

When using LLVM (either llvm-vec or llvm-ve-rv) we need to add some flags. Like -target ve-linux or -rvec.
It doesn't look like I could add any compile options to nsc++

$ src/nsc++ -h
usage: nsc++ [-h] [--hcode filename] [--dcode filename] [--soname filename] [--device {ve,cpu}] [--debug] file

neoSYCL compilation helper

positional arguments:
  file               SYCL source filename

optional arguments:
  -h, --help         show this help message and exit
  --hcode filename   specify the output host code filename
  --dcode filename   specify the output device code filename
  --soname filename  specify the output shared object filename
  --device {ve,cpu}  specify the device to use
  --debug            enable debug options

Unrecognized options are simply passed to compilers.