[PATCH] D47849: [OpenMP][Clang][NVPTX] Enable math functions called in an OpenMP NVPTX target device region to be resolved as device-native function calls

[Gheorghe-Teodor Bercea via Phabricator via cfe-commits]

gtbercea added a comment.

In https://reviews.llvm.org/D47849#1192245, @Hahnfeld wrote:

> In https://reviews.llvm.org/D47849#1192134, @gtbercea wrote:
>
> > This patch is concerned with calling device functions when you're on the device. The correctness issues you mention are orthogonal to this and should be handled by another patch. I don't think this patch should be held up any longer.
>
>
> I'm confused by now, could you please highlight the point that I'm missing?


You're bringing up the correctness of the header files which is a detail that is orthogonal to this patch. Even if the header files worked correctly I would still want to use the libdevice functions. Fixing the header files themselves should be therefore done in a separate patch.
Using the libdevice functions guarantees correctness (no weird assembly instructions that the device doesn't recognize etc.) and may improve performance (if for example the libdevice contained device specific assembly).

The purpose of this patch is to call NVIDIA's libdevice math functions which should in principle be more efficient in terms of runtime and register usage. Not all of them may be more effecient today (like @tra suggested) but some of them will be. Maybe others will be improved in the future, maybe not, again that's an orthogonal point. The benefit of using libdevice functions is that any improvements NVIDIA makes we will be there to use them in the OpenMP NVPTX toolchain. The premise of the OpenMP NVPTX toolchain is that it will leverage as much of the CUDA toolchain as possible.

Another point is that users specifically ask for NVIDIA math functions to be called on the device when using OpenMP NVPTX device offloading. The libdevice library offers __nv_fast_* variants of some math functions. Users want to have access to those functions and other functions that the libdevice library contains.


Repository:
  rC Clang

https://reviews.llvm.org/D47849