bdenhollander

I ran into this as well in a fresh conda environment on Windows. Applying the changes in [`wrappers\python\openmm\__init__.py`](https://github.com/openmm/openmm/blob/master/wrappers/python/openmm/__init__.py) from 4142e19827fd6c9c98b4369e9908c78926f59929 and 952d3402d12dda9931a0af2dce24fd9875e9b730 to `lib\site-packages\openmm\__init.py__` in my conda env allowed OpenMM...

ROCm 5.7.0 includes a fallback to OpenCL-based printf that may resolve this issue. https://rocm.docs.amd.com/en/docs-5.7.0/release.html#non-hostcall-hip-printf

@jatinx Adding a note about PCI Gen3 Atomics to `printf` documentation would be helpful. It states very matter-of-factly that [printf works](https://rocm.docs.amd.com/projects/HIP/en/latest/reference/kernel_language.html#printf). Similar to #2266, `printf` hangs on Windows on gfx1032...

HIP 5.7.0 includes a fallback to OpenCL-based printf to allow use without PCIe atomics. https://rocm.docs.amd.com/en/docs-5.7.0/release.html#non-hostcall-hip-printf

I've been experimenting with the SDK on an RX 6600 (gfx1032) since it was released. Give it a try and see if it works. There's no mention of GCN4 but...

Are parallel `std::for_each` loops from C++17 an option here?

Does adding braces to wrap the `if` statement make any difference? If it helps then I would recommend the same for lines 534-535. ```C++ out

What about hoisting the variables outside of the `for` loop so that they're not re-declared when `nodes.size() > 1`? Re-declaration in the same scope normally works but most compilers will...

I generated .s files with and without `#pragma unroll 1` on Windows. Diffing ignoring whitespace and numbers shows differences in lines 3561-4090. [unroll-windows-hip-amdgcn-amd-amdhsa-gfx1032.s.txt](https://github.com/openmm/openmm/files/12383288/unroll-windows-hip-amdgcn-amd-amdhsa-gfx1032.s.txt) [hang-windows-hip-amdgcn-amd-amdhsa-gfx1032.s.txt](https://github.com/openmm/openmm/files/12383289/hang-windows-hip-amdgcn-amd-amdhsa-gfx1032.s.txt) hipcc.bin.exe prints `clang version 17.0.0 ([email protected]:Compute-Mirrors/llvm-project...

Retested this with Windows [HIP SDK 5.7.1](https://www.amd.com/en/developer/resources/rocm-hub/hip-sdk.html) and `#pragma unroll 1` is still needed. The clang hash reported by hipcc.bin.exe has changed since 5.5.1. ``` C:\AMD\ROCm\5.7\bin\hipcc.bin.exe --version HIP version: 5.7.32000-193a0b56e...