Can we directly read/write a file from GPU instead of doing memcpy from CPU to GPU and GPU to CPU? What is the fastest way to read and write file to and from GPU?
The answer to the first question is “No”.
The answer to the second question is probably “it doesn’t really matter because disk I/O is always going to be the bottleneck and the device-host transfers will have negligible impact on overall throughput”. What will be optimal will depend a lot on how much data you are writing. For a small amount of data, zero copy is probably the best approach. For a lot of data, using several pinned allocations and overlapping I/O, device-host-device transfers, and kernel execution is probably the way to hide most of the latency. But ultimately the disk I/O is still going to kill performance compared to anything the GPU does.
Thank you for your suggestion… I am facing a similar problem…GPU is working fast but CPU is limiting the speed and the final result is slow…So I was just curious about how to make processing faster…