Currently, some of the benchmarks allocate memory that's accessed from the kernels using:
mallocaligned_alloc(2MiB, ...)- C++
new
- vendor specific apis (cudaMallocManaged, etc.)
How the memory is allocated does impact benchmark performance a bit.
I think we should extract these into their own file, to ensure all benchmarks may pick the exact same defaults.
Currently, some of the benchmarks allocate memory that's accessed from the kernels using:
mallocaligned_alloc(2MiB, ...)newHow the memory is allocated does impact benchmark performance a bit.
I think we should extract these into their own file, to ensure all benchmarks may pick the exact same defaults.