update on compiling with big sure and clang11:

[dokipen3d]

I've managed to go from clang 10 to clang 11 now. There were a few API changes in clang 11.

• StringRefs used as return types instead of strings
• constant/callee extraction
• Load/Store instruction signatures changed
• links against libxml2 so needed a few cmake tweaks
• c++14 mode for cocl and clblast
• Sequence type deprecated. I replaced with arrayType as that is what llvm code seems to do, but I am ignoring vectorType.
• VectorType is now FixedVectorType and ScaleVectorType so had to add a couple of cases to a switch/case.

I know what llvm is and what it does but I've never dived into the API and nowhere near an expert (I made a toy interpreter based on craftinginterpreter.com book for fun but thats about it).

I've tested compiling a few simple example.cu files and they seem to work still.

Is anyone interested in me pushing this to a PR? I'm just messing around as I've been ill with a cold in bed and this looked like a fun project. It looks like it could still go on to do some cool things. OR we could start a new project and have it spit out spirv and use something like the vulkan kompute library instead of opencl and get this running even on macs with moltenvk. crazy I know, but would be fun!

[dokipen3d]

ah there are some issues with structs as the api for inserting instructions changed, and I can't get it quite right.

[dokipen3d]

actually, things aren't so successful. I'm getting nullptr returned from MDevice->getFunction(info->kernelName); in getLaunchTypes and it can't find a kernel. no idea why. Not even sure why some things were working before. I think I might have had clang 10 in my path still.

I'll pause for a while on this as I wan't to learn a bit more about llvm modules befire I hack any more. Also I want to do some cuda coding!

[dokipen3d]

the struct test fails to compile on llvm 9 and above. other ones seem to work but for now, the most stable seems to be 8 (so far from the tests ive been trying out).

[hughperkins]

"Is anyone interested in me pushing this to a PR?"

Yes please :)

(Erased my ridiculously over-long previous reply lol :) )

[dokipen3d]

Thank you very much for all of the advice Hugh. Some very good points there. I myself am quite fond of nvidia and have always used their cards ( I've owned GeForce 6600 GT, 8800 GS, GTX 460, GTX 770 and now 1080 in my lifetime). I'm pursuing this mainly as a way to be able to enable small educational projects on various hardware such as lower end laptops and mini pcs. Its probably something I'l dive in and out of over months/years.

Cheers!

[dokipen3d]

More thoughts on spirv. There are actually two 'dialiects' of spirv, one for opencl and one for vulkan. There is an unofficial google project clspv that translates opencl LLVM IR into vulkan SPIRV. So I'm currently looking at possible ways forward. The cool thing about vulkan spirv is that it supports subgroup operations which includes shuffle(), that you commented on in your very nice paper.

Amazingly, even the new opencl 3.0 spec does not have subgroup shuffle and is an intel extension. so if we can go from Nvidia Cuda kernel to OpenCL, we would have to delve into weird extension territory. But a route to vulkan spirv is certainly possible.

I'm currently learning about address spaces and IR which is fun. Last week I went down the Machine Learning rabbit hole and this week this. Fun times.

[hughperkins]

I've created a branch llvm-13. I just changed a few things, certainly not all, and it doesnt compile yet, and I havent tested it. Perhaps any work on making things compatible with recent-er versions of llvm can go into this branch for now?

[dokipen3d]

Hi Hugh, yeah sounds good. I haven't looked too much into this recently, but if get a chance, i'll take a look :)