sbalint98

Hi @mmeterel thank you for reminding me. I'll have a look and prepare this PR until Monday. I'll let you know if it is ready for review

Sorry for the delay. I have rebased the branch and updated the readme. All tests are passing on my sid [cublas_tests_hipsycl.txt](https://github.com/oneapi-src/oneMKL/files/8846920/cublas_tests_hipsycl.txt) e

@nilsfriess and @normallytangent are working on the cuRAND support. I believe for them it might be easier to open a new PR. Would this be acceptable for you?

Pinging @nilsfriess and @normallytangent as they took this over from me.

I pushed a version that, in my opinion, is the most error safe for the user, and also provides the option of specifying the installation root dir without pointing to...

> Can you try extracting a pointer from the accessor? Unfortunately, this does not resolve this issue for me > What clang turns that into is not up to us....

Thank you, @al42and, for the link. I couldn't find your workaround on the official GROMACS repo. Could you point me to the right branch? When `atomicAddNoRet` was used with hipSYCL...

As additional information: [the simple hip reproducer](https://github.com/illuhad/hipSYCL/files/8451133/hip_reproducer.txt) Generates fast assembly without using `atomicAddNoRet` and additional flags on rocm 5.1, 5.0, 4.5.2.

Well, I had a look at this change, but unfortunately, it did not resolve the issue. The problem, in this case, seems to be that the memory is marked to...

Thank you @nilsfriess for looking into this. Can you share which version of ROCm and which target did you use for trying to reproduce the behaviour with USM?