www.digitalmars.com         C & C++   DMDScript  

digitalmars.D.learn - D meets GPU: recommendations?

reply Bruce Carneal <bcarneal gmail.com> writes:
The project I've been working on for the last few months has a 
compute backend that is currently written MT+SIMD.  I would like 
to bring up a GPU variant.

If you have experience with this sort of thing, I'd love to hear 
from you, either within this forum or at beerconf.

In a past life I was a CUDA programmer, but I'd prefer to use 
something now that has at least some chance of running on most 
mobile devices in the future.

In addition to the CUDA fallback, I've looked at SycL, dcompute 
and, as suggested by Ethan a couple of beerconfs back, I looked 
in to the D/MLIR work.  That work is very nice but I didn't find 
the connection to GPUs.

Guidance from experience regarding any of the above, or other, 
GPU possibilities would be most welcome.
Jan 29 2021
next sibling parent reply Guillaume Piolat <first.last gmail.com> writes:
On Friday, 29 January 2021 at 16:34:25 UTC, Bruce Carneal wrote:

 The project I've been working on for the last few months has a 
 compute backend that is currently written MT+SIMD.  I would 
 like to bring up a GPU variant.


What you could do is ressurect DerelictCL, port it to BindBC, and 
write vanilla OpenCL 1.2 + OpenCL C.
Not up to date on both, but CUDA is messier than OpenCL.

I don't really know about the other possibilities, like OpenGL + 
compute shaders or Vulkan + compute shaders.
Jan 29 2021
parent reply Bruce Carneal <bcarneal gmail.com> writes:
On Friday, 29 January 2021 at 17:46:05 UTC, Guillaume Piolat 
wrote:

 On Friday, 29 January 2021 at 16:34:25 UTC, Bruce Carneal wrote:

 The project I've been working on for the last few months has a 
 compute backend that is currently written MT+SIMD.  I would 
 like to bring up a GPU variant.


 What you could do is ressurect DerelictCL, port it to BindBC, 
 and write vanilla OpenCL 1.2 + OpenCL C.
 Not up to date on both, but CUDA is messier than OpenCL.

 I don't really know about the other possibilities, like OpenGL 
 + compute shaders or Vulkan + compute shaders.


Thanks for the pointer to Mike's bindings.  If I run in to 
trouble plugging in with both dcompute and C++/SycL I'll take a 
look before falling back to CUDA.

I've not looked in to OpenGL compute shaders seriously either but 
I did look at Vulkan compute shaders.  They looked very strong 
with respect to future deployability but were weak in other ways 
so I kept them off the list.  I don't think I can warp my code to 
fit in to glsl or similar, at least not easily. Custom 
neighborhood indexing, group/sub-group scheduling control and 
ability to manage group local memory efficiently were the main 
concerns, IIRC.
Jan 29 2021
parent Bruce Carneal <bcarneal gmail.com> writes:
On Friday, 29 January 2021 at 20:01:17 UTC, Bruce Carneal wrote:

 On Friday, 29 January 2021 at 17:46:05 UTC, Guillaume Piolat 
 wrote:

 On Friday, 29 January 2021 at 16:34:25 UTC, Bruce Carneal 
 wrote:

 The project I've been working on for the last few months has 
 a compute backend that is currently written MT+SIMD.  I would 
 like to bring up a GPU variant.


 What you could do is ressurect DerelictCL, port it to BindBC, 
 and write vanilla OpenCL 1.2 + OpenCL C.
 Not up to date on both, but CUDA is messier than OpenCL.

 I don't really know about the other possibilities, like OpenGL 
 + compute shaders or Vulkan + compute shaders.




[...]

 I've not looked in to OpenGL compute shaders seriously either 
 but I did look at Vulkan compute shaders.  They looked very 
 strong with respect to future deployability but were weak in 
 other ways so I kept them off the list.  I don't think I can 
 warp my code to fit in to glsl or similar, at least not easily. 
 Custom neighborhood indexing, group/sub-group scheduling 
 control and ability to manage group local memory efficiently 
 were the main concerns, IIRC.


I took another look at Vulkan compute shaders.  They have come a 
long way.  Not nearly as nice as programming in D but good enough 
that I'll sketch out a few kernels if I can't make headway with 
dcompute.

Anybody with recent dcompute experience, please speak up.  The 
code in the repository is pretty nice but it looks unfinished or 
at least frozen in time.
Jan 29 2021
prev sibling parent reply mw <mingwu gmail.com> writes:
On Friday, 29 January 2021 at 16:34:25 UTC, Bruce Carneal wrote:

 Guidance from experience regarding any of the above, or other, 
 GPU possibilities would be most welcome.




https://dlang.org/blog/2017/10/30/d-compute-running-d-on-the-gpu/


or Google search:

https://www.google.com/search?&q=dlang+gpu
Jan 29 2021
parent Bruce Carneal <bcarneal gmail.com> writes:
On Friday, 29 January 2021 at 18:23:40 UTC, mw wrote:

 On Friday, 29 January 2021 at 16:34:25 UTC, Bruce Carneal wrote:

 Guidance from experience regarding any of the above, or other, 
 GPU possibilities would be most welcome.




 https://dlang.org/blog/2017/10/30/d-compute-running-d-on-the-gpu/


 or Google search:

 https://www.google.com/search?&q=dlang+gpu


Yes, there is a lot of readily available information on the 
possibilities that I listed.  And I have read/viewed much of 
that, including the text and video items prominent in a simple 
google search. (shame on me if I ever ask for something trivially 
available)

Current day "Guidance from *experience*..." is in another 
category, hence the request.
Jan 29 2021