check the impact of different cudaMallocHost flags

[fwyzard]

According to the documentation, cudaMallocHost flags can affect the way the memory is allocated, pinned, and shared with the GPU.

We should check the impact these flags have on the time spent in memory copies, on different architectures.

[fwyzard]

Using cudaHostAllocWriteCombined for most of the CPU --> GPU buffers gave a negligible or negative impact; for the moment we do not plan to use it.

[cmsbot]

A new Issue was created by @fwyzard Andrea Bocci.

can you please review it and eventually sign/assign? Thanks.

cms-bot commands are listed here

[fwyzard]

As a follow up, here are some quasi-benchmark of the host to device memory transfer throughput for

• pageable memory;
• pinned memory;
• pinned, write-combined memory.

The measurements are taken with CUDA bandwitdhTest and show the average and standard deviation form 4 repeated measurements.

In all cases:

• pinned memory is much faster than pageable memory;
• write-combined memory does not show any gains, and is sometimes worse for small transfers.