How do I use dynamic shared memory

[krzikalla]

The documentation mentions dynamic shared memory, but I can't find an example of how to do it.
At the moment we have

auto& shared_data = alpaka::declareSharedVar<T[NumberOfThreads], 0>(accelerator);

in our reduce code. Actually, the value behind NumberOfThreads is a run-time variable. NumberOfThreads is just a configurable constant, which equals to a reasonable high value. However some types T are so large, that sizeof(T) * NumberOfThreads exceed 0xc000. On the other hand, I don't want to restrict NumberOfThreads to a very low value just because some Ts are too large. So what I actually need is

auto& shared_data = alpaka::allocateSharedVar<T, 0>(accelerator, numberOfElements);

where numberOfElements is the actual number of threads (a run-time value).

How do I do this?

PS: We never paid attention to the ID template parameter. Since we only declare one shared variable inside a kernel invocation, should this matter?

[sbaldu]

You can find an example here: https://github.com/alpaka-group/alpaka/blob/develop/example/convolution2D/src/convolution2D.cpp#L86-L204.
You specialize the BlockSharedMemDynSizeBytes struct for your kernel, so that it knows how much shared memory is needed, and then you can access it from inside the kernel with alpaka::getDynSharedMem<T>(acc);

[krzikalla]

Thanks for the quick reply. Just to be sure: I guess, alpaka::getDynSharedMem and alpaka::declareSharedVar don't interoperate well. It's manageable for me, I just want to be sure before I am going to do some refactoring.

[psychocoderHPC]

Dynamic shared memory follows the cuda principles of dynamic shared memory. You need to define the amount at runtime outside of the kernel and within a kernel you are responsible to manage the memory, e.g. not accessing adresses behind the allocated size.
With getDynSharedMem you get the adress to the first value.
You can mix dynamic memory and static shared memory within a kernel but the total size of both combined is limited to around 32kiB.

[krzikalla]

OK, specializing the BlockSharedMemDynSizeBytes trait works. However I've now got another question: why is getBlockSharedMemDynSizeBytesmarked with ALPAKA_FN_HOST_ACC. It is never called in device context, so it should be ALPAKA_FN_HOST, shouldn't it.

[psychocoderHPC]

That is a very good question, from a short look it looked wrong and ALPAKA_FN_HOST should be fine.