CUDA development tools additions What is missing from current CUDA debuggers?

Hi all,

I am working on adding correctness checks to GPU Ocelot’s instrumentations. This will implement checks for the GPU backend, which will cause some slow downs but it would be much faster than running the same checks through the emulator backend. The goal is to provide developers better ways to debug their programs and ensure program correctness.

My plan is to add some simple, feasible checks: alignment checking, shared memory access race detection, and out-of-bound checks.

My question is, what other checks are there that CUDA developers would benefit from? Are there any missing features from existing CUDA development tools (nSight, Ocelot debugger, etc) that you want to see implemented?

Thanks,
Si

It would be very helpful to have a tool that could discover expressions that use uninitialized values from memory.