CODEWITHSUNDEEP
cuda
AmirC
AmirC

Reputation: 326

Memory Mapping in CUDA

I am trying to understand the data layout in GPU memory. For example, when we call cudaMalloc3D or cudaMalloc, how is the data laid out in memory in terms of banks/row/columns? Is this information available?

Upvotes: 0

Views: 579

Answers (1)

talonmies
talonmies

Reputation: 72349

None of what you are asking for is officially documented, as far as I am aware.

Simple linear memory allocations made with malloc area just that. Pitched linear memory is just the same, but with a pitch which is determined at runtime. The pitch heuristics are undocumented. CUDA arrays are an opaque data container only accessible via the texture or surface APIs. How any of this maps onto hardware is undocumented.

The memory layout of the device will vary from hardware version to hardware version. You might be able to find at least some hints in the various architecture white papers which NVIDIA normally release at the announcement of a new architecture. Beyond that, I don't believe there is more in the public domain.

Upvotes: 1

Related Questions