[libre-riscv-dev] whole stack of vulkan llvm spirv stuff

[Jacob Lifshay]

On Wed, Sep 11, 2019, 18:38 Luke Kenneth Casson Leighton <lkcl at lkcl.net>
wrote:

> On Thu, Sep 12, 2019 at 12:30 AM Jacob Lifshay <programmerjake at gmail.com>
> wrote:
> >
> > On Wed, Sep 11, 2019, 13:38 Luke Kenneth Casson Leighton <lkcl at lkcl.net>
> > wrote:
> >
> > > https://github.com/KhronosGroup/SPIRV-LLVM
> >
> > the reason I didn't use SPIRV-LLVM is because it only handles the OpenCL
> > variant of SPIR-V (last I checked)
>
> the release date is what... four years?  still, i'm supposed to know
> these things, in order to be able to make informed decisions!
>
> > as well as being only for much older versions of LLVM.
>
> okaay.  that would explain why AMD forked the whole of LLVM as well:
> https://github.com/GPUOpen-Drivers/llvm
>
> although it looks like they've been mix-and-match merging.
>
>
> > >
> > > https://github.com/GPUOpen-Drivers/llpc
> > >
> > >
> https://hub.packtpub.com/amd-releases-amd-open-source-driver-for-vulkan-v-2019-q1-2/
> > >
> > > That SPIRV to LLVM compiler is apparently two-way!
> > >
> > > How did we manage to miss that? moo??
> > >
> >
> > we would still need the scalar to vector conversion code, which is almost
> > all of the spir-v to llvm converter I'm writing as part of Kazan
> > (converting scalar spir-v to scalar llvm is trivial by comparison).
>
> okaay, got it.  which, hm, makes me curious as to how the heck AMD did it.


they do it in their AMDGPU backend for LLVM, which works because AMDGPU is
a GPU-only ISA.

>


> l.
>
> _______________________________________________
> libre-riscv-dev mailing list
> libre-riscv-dev at lists.libre-riscv.org
> http://lists.libre-riscv.org/mailman/listinfo/libre-riscv-dev
>