[Question] Can one allocate the shared memory using CUDArt.jl? about cudart.jl

[timholy]

I'm not sure what you mean. If you use the commands in this section, you're explicitly allocating memory on the device. What kind of "sharing" did you have in mind? Do you know about SharedArrays? (They're on the CPU.)

[joaquimg]

I wanted to mean the memory dynamic shared memory between threads of some block. Like in the section dynamic shared memory here
The third field in:
yourFunction<<<1, n, n*sizeof(int)>>>(d_d, n);

used for sum reductions for instance.

I am aware of SharedArrays in CPU case.

[timholy]

Oh, right, now I remember this kind of thing. Honestly, I don't remember (it's been at least a year since I put a lot of effort into this package), this is probably the kind of thing you'll have to dig into the source code for. If it's not supported, a PR implementing it would be great!

[adambrewster]

Yes.

When calling launch (function launch(f::CuFunction, grid::CudaDims, block::CudaDims, args::Tuple; shmem_bytes::Int=4, stream=null_stream), set shmem_bytes to the desired size.

[maleadt]

This functionality is covered by CUDAdrv's cudacall nowadays, which accepts a keyword argument shmem to specify the bytes of shared memory to dynamically allocate.

I'll update the README and examples, which apparently are still using launch.