Managed to get an #Intel Arc A750 #gpu running on #risc_v using #OpenCL, #SYCL, and #AdaptiveCpp. Software PR's submitted for review.

#hpc #supercomputing

@risc_v

#oneAPI / #SYCL 2020 / #DPC++ / #AdaptiveCpp Is starting to look pretty nice. I’m not sure if there is a story for mobile devices and older devices. It seems that OpenCL 1.2 isn’t enough for SYCL 2020, but unsure. Also not sure how Vulcan fits in here. It’s also unclear how it scales to smaller kernels and data. The API looks like kernel/buffer/queue. Not sure if you can queue up many kernels or if they can be pipelined.

How does the new IR that llvm is using fit in here? So many questions.

Like C++ standard parallelism offload? #AdaptiveCpp substantially outperforms NVIDIA HPC SDK's nvc++ in C++ PSTL offloading for the LULESH app on A100. Perhaps this a good moment to reconsider your use of vendor compilers 😉 (Also, the same AdaptiveCpp binary runs runs on AMD/Intel hw too!)

For a while (and still AFAIK) you needed to build clang from source to get #OpenMP working on MacOS.

Remember running into that with #AdaptiveCPP+#Spack