On Wed, 3 Jan 2024 21:30:02 +0900 Takashi Yano wrote: > On Wed, 3 Jan 2024 21:00:07 +0900 > Takashi Yano wrote: > > On Wed, 3 Jan 2024 04:38:02 -0700 > > Brian Inglis wrote: > > > On 2024-01-03 02:29, Takashi Yano via Cygwin-apps wrote: > > > > On Wed, 3 Jan 2024 08:54:17 +0100 > > > > Marco Atzeri wrote: > > > >> On 03/01/2024 06:25, Takashi Yano via Cygwin-apps wrote: > > > >>> On Wed, 3 Jan 2024 14:14:12 +0900 > > > >>> Takashi Yano via Cygwin-apps <cygwin-apps@cygwin.com> wrote: > > > >>>> I'd like to adopt the pocl package. > > > >>>> - Update to latest upstream release. > > > > > > >> $ git diff |grep "^+" > > > >> +++ b/cygwin-pkg-maint > > > >> +pocl Takashi Yano > > > > > > >>> Sorry, the latest upstream release is 5.0 however, 4.0 and later > > > >>> cannot be built in current cygwin because LLVM package is old. > > > >>> This update is up to 3.1. > > > > > > >>>> - Enable CUDA support. > > > > > > >> Curiosity, how do we support CUDA on Cygwin ? > > > > > > > nvidia cuda toolkit is used in build stage of user programs. > > > > Although this is not very desirable for cygwin package, I thought > > > > that the advantage of being able to use the GPU was greater than > > > > the disadvantage. > > > > However, on the second thought, cuda support should be a separeted > > > > package from the base package, and suggest installing cuda toolkit > > > > in the install stage of of that package. > > > > Let me consider a bit. If you have any idea, please let me know. > > > > > > Please note CUDA is Nvidia proprietary closed source - I do not think we > > > can or > > > should touch it when OpenCL 3+ supports Nvidia devices. > > > Fedora does not support CUDA although others do in their non-free > > > "sources". > > > > We do not touch CUDA itself or distribute its binaries, but just use > > binaries > > distributed by NVIDIA. Source code in pocl is not NVIDIA proprietary. In > > that > > sense, cygwin itself uses microsoft proprietary closed source modules. > > > > So, enabling CUDA support in pocl should also be allowed, I think. > > D'oh! I missed that libpocl-devices-cuda.so is linked with cuda.lib > from CUDA Toolkit. > > I'll give up enabling CUDA support for pocl.
Ah, cuda.lib is just a import library for nvcuda.dll, so we can easily create non-proprietary import library using dlltool. Then, libpocl-devices-cuda.so will be free from NVIDIA binaries just like libkernel32.a of w32api-runtime. -- Takashi Yano <takashi.y...@nifty.ne.jp>