René Widera

https://github.com/psychocoderHPC/mallocMC/tree/topic-hip-port contains now support for all variations of HIP. ATTENTION: HIP-hcc uses a workaround where the mallocMC heap is not freed when the heap object is destructed. I added the...

Even if you execute twice the 2D or 3D case it will fail. The reason is that the distribution policy XMallocSIMD has a bug. Chnaging https://github.com/alpaka-group/mallocMC/blob/bffe2aa2da5e83d356ff8d32f392935b8f7a59fa/tests/dimensions.cpp#L146 to `mallocMC::DistributionPolicies::Noop,` will fix...

The problem in the XMallocSIMD policy is how we communicate between threads in a warp. https://github.com/alpaka-group/mallocMC/blob/bffe2aa2da5e83d356ff8d32f392935b8f7a59fa/src/include/mallocMC/distributionPolicies/XMallocSIMD.hpp#L135 Since Volta threads in a warp can diverge, we should use in warp communication...

I tried libMesh with a blockchain daemon and it reduces the memory footprint by 20%. Next test would be with PIConGPU.

Copy a mallocMC buffer to host and later on back is only valid if the start pointer of the heap before and after the copy is the same. The reason...

@afanfa I would say that `atomicMin(unsigned long long, unsigned long long)` is broken too https://github.com/ROCm-Developer-Tools/hipamd/blob/3ec1ccdbbbee7090ba854eddd1dee281973a4498/include/hip/amd_detail/amd_hip_atomic.h#L938-L951 If the data in memory tmp==32 and value==4) the return value is 4 but should...

~~I found out that I checked the wrong branch, the main branch looks like and left over from older development, I opened an issue to remove this dead branch https://github.com/ROCm-Developer-Tools/hipamd/issues/41~~...

`unsigned long long` `atomicMin` and `atomicMax` is only wrong for architectures without `__hip_atomic_compare_exchange_strong`

The list of supported target is taken from: https://github.com/ROCm-Developer-Tools/HIP/blob/main/docs/markdown/hip_porting_guide.md#compiler-options-supported-on-amd-platforms

@pfultz2 Is there anything to do before this can be merged?