René Widera

I was not able to follow the full discussion but I try to read all soon. I would like to point all to the [Mephisto buffers](https://github.com/mephisto-hpc/testing/blob/master/mephisto/include/mephisto/buffer). The device-side buffer representation...

Thanks for starting the discussion about the memory topic! :+1: > 1. 1.3 `makeBuf(TDev const& dev, TElem* ptr, TExtent const& extent, bool useHostPtr)` - create a buffer from a host...

> > enum class viewTarget > > { > > HostMemory, > > GlobalMemory, > > ConstantMemory, > > SharedMemory > > }; > > Let's light up the bomb:...

> > Let's light up the bomb: how does unified memory fit into this picture? I think unified memory is getting increasingly relevant. Also because it is the default model...

> > ``` > > auto get_byte_distance(); // = pitch > > ``` > > > I do not like the name. This function also only makes sense for 2D...

> ```c++ > ```c++ > auto const taskKernel(alpaka::createTaskKernel( > workDiv, > kernel, > alpaka::require(bufAccA), > alpaka::require(bufAccB), > alpaka::require(bufAccC), > numElements)); > ``` > > > With my proposed implicit recognition...

> Let me clarify as well: > I wanted to implicitely convert an alpaka buffer passed to `alpaka::createTaskKernel` into a `T*` at the kernel entry function. So I would like...

Using cmath functions directly was not possible in the past. The attribute `__device__` is missing for those functions. I am not sure if CUDA is maybe shipping for some function...

> Seems to work in nvcc, HIP clang and normal clang for CUDA: https://cuda.godbolt.org/z/n48chKxGe > Maybe we have an old compiler somewhere that complains. But I took the oldest available...

> `boost::alignment::aligned_alloc` could be replaced > > * by C11 `aligned_alloc` but this [will most probably never be implemented by Visual Studio](https://devblogs.microsoft.com/cppblog/c17-features-and-stl-fixes-in-vs-2017-15-3/) > * `std::aligned_alloc` which is available starting from...