Accelerator emulation on PGVF90 10.3?

Accelerated Computing HPC Compilers Legacy PGI Compilers
#1

Hi ‘Accelerators’!

It is possible to compile and run F90 Accelerated code (based on directives) on my laptop, which has no GPU attached?

I assume that the -Mcuda=emu gives the option for CUDA Fortran code to execute on non-GPU machines in emulator mode, but what about simple accelerator directive based code? – The reason; I would like to develop locally and then port to a large system where production should take place.

Just made a clean install of PGVF (v.10.3, on WinXPpro), the entire bundle with VS etc. But depending on which flags I set in under Project properties -> Configuration Properties -> Fortran -> … etc I get either error at runtime telling med that “nvcuda.dll” is not found, or at linktime that linker cannot find “libacc1mprt.lib”. Do I need additional libraries/dlls, or is it just not possible to run directive based accelerator code in non-GPU emulator mode?

Best regards,
Jacob

#2

Jacob,

Our -Mcuda=emu switch generates an all x86 version of the code.
If you have cuda libs that you are calling from this emulation code,
they will fail since they would require a GPU to be present to run.

We are not running a simulator, we are replacing cuda runtime libs
with x86 equivalents, and we are compiling the program for an x86.
You can develop code and libs, but all the code and libs
need to be compiled -Mcuda=emu.

regards,
dave

#3

My mistake, I misread your mail and did not answer the question.
No, we do not emulate Accelerator directives, though technically we should be able
to.

Sorry about the last mail.

regards,
dave

#4

FYI, our next target for the Accelerator model will be multi-core x86 which will be similar to CUDA Fortran emulation.

  • Mat