Nol Moonen

Rebased the branch.

@stanleytsang-amd I can reproduce the bug reliably on `gfx906` and `gfx908` when letting `get_large_sizes` (not `get_sizes`) return size `30064767271` in `test/rocprim/test_device_partition.cpp`.

Rebased the PR.

The issue is also reproducible with ```c++ #include #include int main() { printf("starting..\n"); hipStream_t stream; hipStreamCreate(&stream); hipStreamDestroy(stream); hipStream_t stream2; hipStreamCreate(&stream2); hipStreamDestroy(stream2); printf("finished!\n"); } ``` so it is not specific to...

The issue is reproducible on a system with two `gfx1030` cards. It is not reproducible on a system with only one: if I create a `rocm/rocm-terminal:5.3` image and pass only...

@jrbyrnes `bool* val = static_cast(temp)` breaks strict aliasing rules and is undefined behavior. If PyTorch wants to invoke undefined behavior, that's up to PyTorch, but it isn't the responsibility of...

We are currently investigating the consequences of applying the PR. Aside from that: @bcahoon `bool* val = static_cast(temp)` breaks the strict aliasing rule, even if the underlying object type is...

Sorry, correction: the example I gave is also not allowed. While manipulating the data through the `char*` is allowed, it is not allowed to write a value outside of the...