[cfe-dev] Ask for CUDA supports in Clang

[lipengcheng]

By the way, this simple header file https://gist.github.com/eliben/b014ac17cbe5a452803f was pushed into Github
3 hours ago. Haha~ :)

On Oct 6, 2014, at 8:54 PM, lipengcheng <pli at cs.rochester.edu> wrote:

> Thanks, Eli and Jingyue. I just thought to fake them that way. Plus, consider corresponding 
> semantics of CUDA syntax when processing AST. 
> 
> Hope to see you guys at google next summer. :) 
> 
> Pengcheng
> 
> On Oct 6, 2014, at 6:46 PM, Eli Bendersky <eliben at google.com> wrote:
> 
>> 
>> 
>> On Mon, Oct 6, 2014 at 3:36 PM, Jingyue Wu <jingyue at google.com> wrote:
>> Eli, 
>> 
>> Lacking of real support for threadIdx, blockIdx, etc. may be fine for Pengcheng (?) because he's doing CUDA->AST->CUDA transformation. 
>> 
>> Yep, I think this is what I wrote below suggests. The linked header creates fake declarations of these globals, which is good enough for Clang's parsing.
>> 
>> Eli
>> 
>> 
>>  
>> Jingyue
>> 
>> On Mon Oct 06 2014 at 3:21:11 PM Eli Bendersky <eliben at google.com> wrote:
>> On Mon, Oct 6, 2014 at 11:25 AM, lipengcheng <pli at cs.rochester.edu> wrote:
>> Thanks, Eli.
>> 
>> May I know what your initial support is? Can you please open it to me?
>> I don’t need too strong CUDA support, say generating CUDA code to LLVM
>> IR. What I only need to do is to parse CUDA code to Clang AST, and then
>> translate it back .
>> 
>> 
>> So this part should already work, if you're careful. The main missing parts for just parsing the CUDA source into the AST are:
>> 
>> 1. Real support for the threadIdx, blockIdx, etc. special globals
>> 2. Various header definitions for __global__ & friends, CUDA builtins, etc.
>> 
>> If all you need is to *parse*, then a simple header like https://gist.github.com/eliben/b014ac17cbe5a452803f should do the trick.
>> 
>> Then you can do:
>> 
>> $ clang  -cc1 -ast-dump -x cuda  -include <path-to-header-file-in-gist> cuda-input-file.cu
>> 
>> And this should dump the AST. If you need to support more CUDA builtins (and/or runtime functions), you should add them to the header, *or* try to get Clang to parse the NVIDIA CUDA headers.
>> 
>> Eli
>> 
>> 
>> 
>> 
>> 
>> 
>> 
>> 
>>  
>> Thanks,
>> Pengcheng
>> 
>> On Oct 6, 2014, at 12:34 PM, Eli Bendersky <eliben at google.com> wrote:
>> 
>>> 
>>> 
>>> On Mon, Oct 6, 2014 at 6:40 AM, lipengcheng <pli at cs.rochester.edu> wrote:
>>> > Hi there,
>>> >
>>> > May I ask what’s the status of supporting CUDA on Clang?
>>> > Are there any incomplete patches to use for CUDA? I am doing
>>> > a source-to-source translation work for CUDA based on Clang.
>>> >
>>> > Thanks,
>>> > Pengcheng
>>> 
>>> 
>>> Hi Pengcheng,
>>> 
>>> CUDA is still only partially supported. This has been discussed a few times this year:
>>> 
>>> http://lists.cs.uiuc.edu/pipermail/cfe-dev/2014-September/038979.html
>>> http://lists.cs.uiuc.edu/pipermail/cfe-dev/2014-March/035782.html
>>> 
>>> Things haven't changed significantly since then. Patches welcome!
>>> Eli
>>> 
>>>  
>>> 
>> 
>> _______________________________________________
>> cfe-dev mailing list
>> cfe-dev at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/cfe-dev
>> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20141006/d732d7af/attachment.html>