Parfor on GPU

I want to run two functions in parallel on a GPU. For this i want to use pafor(eg: for ii=1 fun1 and ii=2 fun2).

Can variables on GPU be used for parfor operations on GPU?

Answer 1

[Edric Ellis]

Yes, you can do this. Whether you get much benefit depends on whether you have multiple GPUs in your system (under some circumstances, a single GPU might actually suffice if you have enough CPU work to keep things busy).

You might wish to do something like

spmd
  gpuDevice( 1 + mod( labindex - 1, gpuDeviceCount ) )
end

before you go any further (if you have multiple supported GPUs)

After that, gpuArrays can be passed into and out from PARFOR loops with no further modification. The following example shows this - but note that this is a proof of concept - it performs very badly because you're operating on scalar elements of the gpuArray.

g = gpuArray(1:10);
parfor ii=1:numel(g)
  x(ii) = 1/g(ii);
end

Answer 2

[Titus Edelhofer]

Hi Sampath,

probably not. I guess it would make not much sense anyway, because the two functions would share the same computational power of the GPU (like running parfor on a single core machine).

If you happen to have to GPUs you could use parfor/spmd to split the functions onto the two GPUs ...

Titus

[Pavel Sinha]

But the GPUs are milti-core computation engine. If a GPU has enough resources, can Matlab run two functions in parallel on the same GPU?

[Walter Roberson]

Nvidia gpu cores are restricted to running the same instruction as the other cores in the same SM. My reading of the linked article is that different SM could be running unrelated tasks efficiently. However, the end of task processing of bringing back results and status looks like it would potentially be inefficient.

https://yosefk.com/blog/simd-simt-smt-parallelism-in-nvidia-gpus.html