• Nicholas Wilson (12/12) Jul 17 2016 Available here[1][2], it can now build a simple add kernel for
• cym13 (5/17) Jul 17 2016 What did you have to modify in LDC? Was the change important? I'm
• Nicholas Wilson (9/30) Jul 17 2016 I think the only changes (as opposed to additions) were to

Nicholas Wilson <iamthewilsonator hotmail.com> writes:

Available here[1][2], it can now build a simple add kernel for 
both CUDA (.ptx) and OpenCL (.spv) (and of course the host code) 
in one compiler invocation.

A lot of things are still hardcoded and the optimisations need to 
tuned a bit more but it works.

Any and all testing, feedback,bug reports and pull requests 
welcome!

I'll be working on it the next two weeks, after which I go back 
to uni and will have less time to work on it, so the more feed 
back the better!

[1]https://github.com/thewilsonator/dcompute
[2]https://github.com/thewilsonator/ldc/tree/dcompute

cym13 <cpicard openmailbox.org> writes:

On Sunday, 17 July 2016 at 07:45:45 UTC, Nicholas Wilson wrote:
 Available here[1][2], it can now build a simple add kernel for 
 both CUDA (.ptx) and OpenCL (.spv) (and of course the host 
 code) in one compiler invocation.

 A lot of things are still hardcoded and the optimisations need 
 to tuned a bit more but it works.

 Any and all testing, feedback,bug reports and pull requests 
 welcome!

 I'll be working on it the next two weeks, after which I go back 
 to uni and will have less time to work on it, so the more feed 
 back the better!

 [1]https://github.com/thewilsonator/dcompute
 [2]https://github.com/thewilsonator/ldc/tree/dcompute

What did you have to modify in LDC? Was the change important? I'm 
asking because the lesser the change the more likely it is to be 
maintanable in the future and/or added to LDC as a proper 
extension.

Nicholas Wilson <iamthewilsonator hotmail.com> writes:

On Sunday, 17 July 2016 at 19:23:19 UTC, cym13 wrote:
 On Sunday, 17 July 2016 at 07:45:45 UTC, Nicholas Wilson wrote:
 Available here[1][2], it can now build a simple add kernel for 
 both CUDA (.ptx) and OpenCL (.spv) (and of course the host 
 code) in one compiler invocation.

 A lot of things are still hardcoded and the optimisations need 
 to tuned a bit more but it works.

 Any and all testing, feedback,bug reports and pull requests 
 welcome!

 I'll be working on it the next two weeks, after which I go 
 back to uni and will have less time to work on it, so the more 
 feed back the better!

 [1]https://github.com/thewilsonator/dcompute
 [2]https://github.com/thewilsonator/ldc/tree/dcompute

 What did you have to modify in LDC? Was the change important? 
 I'm asking because the lesser the change the more likely it is 
 to be maintanable in the future and/or added to LDC as a proper 
 extension.

I think the only changes (as opposed to additions) were to 
main.cpp and to how it translates D type to llvm types (making it 
somewhat aware of addrspaces) and possibly some logging suff to 
figure out what it does. I'll have to run clang-format on it at 
some point though.

There's a lot of semi-duplicate code though, I copied most of the 
declaration and statement codegen stuff, and got rid of all the 
stuff I wouldn't use, e.g. classes exceptions and whatnot.