I've got a large codebase written in CUDA. I ported it to HIP, tested it on Nvidia platform. The computation results are correct, and cuda-memcheck finds no memory-related errors.
But I have encountered a lot of problems using ROCm on AMD platform. Without crashing at runtime, my code run on a Vega 56 GPU yields wrong results. I'm wondering if there is any memory checker available for ROCm that seeks invalid memory access and uninitialized variable use, similar to cuda-memcheck for Nvidia GPUs and valgrind for CPUs. Thanks for any advice.