Cuda portability

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

I’m wondering how are your experience with portability (performance and source-code) for cuda?

  • Are there any issues porting code from compiler to compiler?
  • Does the performance show consistent performance (using comparable metric) over all Hardware?

Thanks in advance

2

It is not an issue because there is no portability for CUDA code. That is because there is only one compiler for it : Nvidia’s. Other compilers are used for the host-side code (gcc, clang, VS, etc.) but all device code is compiled by Nvidia’s compiler.

3

All device code is indeed compiled finally/ultimately for NVIDIA GPUs by the ptxas compiler (or the equivalent functionality in the GPU driver). There are a few user-created assemblers out there (e.g. maxas) but these aren’t that relevant to this discussion, I don’t think.

However, the CUDA device code compilation process doesn’t necessarily begin with ptxas, and the conversion of source code (in whatever form it may be) to PTX may follow a number of available paths, some of which are not wholly created by NVIDIA or part of the NVIDIA provided toolchain(s). I’ll mention 2 examples:

clang has the ability to compile CUDA C++ device code:

https://llvm.org/docs/CompileCudaWithLLVM.html

gnu tools have the ability compile OpenACC device source code:

https://gcc.gnu.org/wiki/OpenACC

As far as I know, both of these examples build fatbinaries with embedded PTX, so they are runnable directly as a “CUDA executable”. The conversion to CUDA machine code would be handled by the GPU driver, equivalently to a CUDA executable built with e.g. -gencode arch=compute_30,code=compute_30 using NVIDIA nvcc toolchain.

I’m not trying to provide any value judgments here, or any statements of suitability.

4

I appreciate the elaboration Robert. Thank you.