I have a AMD Firepro v4800 Discrete graphics card and using AMD 2.6 SDK on a linux box running 3.0.0-16-generic linux kernel.
I want to overlap some kernel execution with some data transfer to the GPU.
I see that the SDK samples all use clEnqueueMapBuffer() in the TransferOverlap example.
I tried to do the same with clEnqueueRead/Write with out-of-order queue. But this is not working.
So is it even possible to do this with clEnqueueRead/WriteBuffer() ?
[ I couldnt find anything in the opencl 1.1 spec that
says it should not. I am using CL_FALSE option and synchronized the data movement with events. When I profile and see
the order of command execution I dont see any out-of-orderness even though I see the ENQUEUED event from profiling showing
time before the kernel completion time.]