DCompute backend

[ShigekiKarita]

After https://github.com/ShigekiKarita/grain/tree/v0.0.6, I removed the dcompute library and CUDA kernels written in D from grain. If these issues are fixed, I wanna go back

• call atomic ops in cuda kernel
• call sync threads in cuda kernel
• call fast math functions in cuda kernel
• load nvptx module written in D with the other modules written in C++

[thewilsonator]

sync threads is dcompute.std.sync.barrier()

for fast math functions LDC's @fastmath should do the trick, if it doesn't let me know.

I'm waiting on the LLVM7 release and then I'll get LDC into shape and get the atomics working.

Could you please point me to an example of loading multiple module so I can have an idea of what thats supposed to look like?

[ShigekiKarita]

Thank you for your kindness! It really helps me. I'm trying to go back dcompute soon.

Could you please point me to an example of loading multiple module so I can have an idea of what thats supposed to look like?

OK but this might come from my wrong usage of CUDA driver API. Anyway I will make a minimal example.