GPU allocations

[aulwes]

We would like to track GPU allocations via hipMalloc and cudaMalloc calls as well. Would it be relatively easy for us to modify Malt to track these calls? Could you point us to where in the code that we would need to make the changes?

[svalat]

Is it just for the raw basic hipMalloc() & cudaMalloc() ? Not the cuda/hipMalloc3D() ? If it is the case, I can just shortly patch my GPU branch, it is mostly ready for those basic calls and put it somewher in the coming days.

[aulwes]

Just for the basic hipMalloc calls (also includes hipMallocAsync). Thank you!!

[svalat]

Hi,
I don't know if I will have time to touch it this week, but in case you can already give a look to the branch https://github.com/memtt/malt/tree/feature/gpu.

There is currently support for :

• cudaMalloc / cudaFree
• PGI OpenACC alloc interception via nvcc

Note: you need to compile your program with the shared cudart not the static one (-cudart shared on NVCC). You can have a look in the examples/cuda and examples/openacc.

I updated the full stack, from the profiling to the WebUI with dedicated pages + source annotations.

What is missing is mostly intercepting all the functions. If ever you have time to try to add some (mostly hip for what you asked) I will be pleased.

For this, look here to extend to other hip/cuda functions :

• Search for CUDA_FOUND in the CMakeLists.txt scripts, need to do similar for HIP.
• The injetor which replace the official function :

  malt/src/libinstrum/injectors/InjectGpuCudaAlloc.cpp

  Line 23 in f016754

  DLL_PUBLIC cudaError_enum cudaMalloc (void ** devPtr, size_t size)

• The wrapper which make the wrapping ops (you can copy-paste the fonction, the content is similar in all wrappers) :

  malt/src/libinstrum/wrappers/WrapperGpuCudaAlloc.cpp

  Line 37 in f016754

  cudaError_enum MALT::malt_wrap_cuda_malloc(void ** devPtr, size_t size, CudaMallocPtr real_cuda_malloc, void * retaddr)

The rest inside the core part is common and might not required to be touched.

I will need some work to get support of HIP/CUDA with a single build of MALT as I'm not totally sure we can mix both libs at the same time in the tool. Naively as it is we might build eighter for one or the other.

[aulwes]

Excellent, thank you! I'll start looking at modifying for HIP.