Michel Migdal

Hello @slgogar, I've worked on porting the back-end to SYCL too. I'm getting 80% of the CUDA performance. Where are you at ? Porting the code using DPCT yields bad...

> @Michoumichmich I wasn't aware of this ongoing effort to enable SYCL backend. Can you please share a link to the referred branch? Is there a particular test you are...

Hello @lindstro, I will follow-up with the benchmarks. Regarding the tests, for the fixed-rate compression everything passes on the GPU as well as the new tests from `feature/cuda-variable-rate` branch. However,...

> @Michoumichmich could you please tell me the oneApi version that is used in your branch? Hi, I worked on the tip of the Intel/llvm as bugs got fixed (one...

@jeffhammond It would be really nice to support Apple Metal. I was able to run SPIR-V on Apple Metal using C++ So it should be achievable as hipSYCL already provides...

> @Michoumichmich Are we talking SPIR-V in OpenCL kernel model or shader model? One of the big issues of shader model is that it only has opaque pointers of undefined...

> hmm...I'm currently working towards a generic toolchain-independent LLVM backend, so maybe we could eventually leverage this to generate MSL.. Although this sounds like reinventing the LLVM C backend. Is...

It could be addressed using https://github.com/illuhad/hipSYCL/pull/578 ?

> The current specification has both the init and no-init version, but asks for the identity. > Currently `sycl::known_identity` is not yet implemented in hipSYCL, The version I am currently...

What if the accessor `auto gm_data = buf.get_access(cgh);`, line 26, is created in the in the command group scope, outside of the kernel, line 51 in your code before passing...