Joachim Meyer

Hi @hiaselhans, the difference between those two commits is that the "bad" one uses the CPU acceleration compiler support by default. The warnings you see are kind of expected and...

Hi, thanks for the response. That indeed is troubling. Can you verify which hipSYCL commit you are on? It seems to me that you are at some intermediate state that...

Hi @hiaselhans, I can reproduce a linker segfault when using the CUDA target + LTO, when building the `sycl_tests` from hipSYCL's unit tests, with both BFD and LLD. Without LTO...

Alright.. I realized that LTO also worked up until recently, so I did a bisection myself and came to the conclusion that 7704095193a095fd33f2edc051b11ce23a0fe6cb is the offending commit, which I guess...

The initial compilation error is fixed in: #578. For a work-around just add the type to the functor: `maximum()`. I don't see a connection to #527, why the reference?

That's probably known as atomics on GPU are not implemented in master, see e.g. issue #213

Yeah.. just looked at the code and it looks like they do it _the right way_. In `hipStreamCreateWithFlags` a new stream is created: https://github.com/cpc/hipcl/blob/35cb616269cd81cd3652b19614e41e6f8d0c7219/lib/hipcl.cc#L438 which actually results in a new...

As you probably noticed, the SPIR-V backend is really experimental ;-) The CUDA path has to be specified as currently, the syclcc compilation driver reuses the CUDA command lines to...

Tested on a mixed GPU setup, seems good. Both HIP and CUDA group_reduce and group_inclusive_scan seem to be in the same performance ballpark as before (according to your sycl-bench fork...

You typically want to build inside a `build` directory. Following the [building guide](https://github.com/illuhad/hipSYCL/blob/develop/doc/installing.md#building-and-installing), you probably want to execute something like the following: ```bash $ cd hipSYCL # go to cloned...