Build device code in advance with static library

[Luigi-Crisci]

Hello,
I have a cmake project with this structure:

• A static library lib.a with multiple kernel in different source files
• An executable a.exe linkinng to lib.a

When targeting the NVIDIA backend, I want to set different ptxas options for each kernel, but when building the project the assembler is called during a.exe linking phase, thus the options are ignored as the executable is a single translation unit.
Is there a way to tell DPC++ to build the device code in advance?

I was trying the same thing with OpenSYCL and it call ptxas while building the kernels during lib.a building phase, so it works smoothly.

[Luigi-Crisci]

I forgot to say that, to set the options for each kernel, I was using the cmake option set_source_files_properties("KERNEL.cpp" [COMPILE_FLAGS|LINK_FLAGS] "-Xcuda-ptxas OPTION"). But it's completely ignored as the assembler is not called wen building the static library target

[AlexeySachkov]

Hi @Luigi-Crisci,

Unfortunately, at the moment this functionality is not available neither for CUDA backend, nor for any other targets. The problem is in fundamental design of the toolchain: device compiler is invoked at link stage, when all device code from all translation units is visible. This was done to support SYCL_EXTERNAL.

I'm aware of other requests of the same functionality, i.e. a way to schedule device compiler pass early: either at static lib level, or even at a translation unit level. Motivation is the same as yours: have ability to specify per-translation unit backend-specific compilation flags.

AFAIK, someone is already working on that, but can't provide any timelines. Tagging @asudarsa and @mdtoguchi for awareness here.