Philip Turner

> Can you give a fuller description of what you're doing? Nanomechanical ALU with ~720 logic switches (mechanical transistors), each ~4000 atoms. Driven by a flywheel contained in an isolated...

> I think that's mostly building and compiling the kernels. The kernels by themselves take ~100 ms to compile with OpenCL, provided it hits the Metal shader cache. I think...

I could try a “nonbonded counter-force”. Include the 1,3 and maybe 1,2 pairs in the nonbonded force. Then add another bonded force between each pair with the inverse effect. It...

Is it possible to skip the detection of molecule groups entirely? I don't need this functionality. Atoms are sorted in Morton order beforehand. There is no solvent or identical water...

> Trying to disable molecule detection sounds dangerous to me. The potential benefit is small. The potential risk of messing something up is large. These are rigid bodies. They deform...

> Are parallel std::for_each loops from C++17 an option here? Is that like `DispatchQueue.concurrentPerform`? That's a convenient Swift function that performs a load-balanced execution of several function calls across all...

I do find, even in some compute-intensive code, parts that fail to parallelize. For example, a reduction operation across an array was always slower on multicore. Always make the code...

I was directed here to discuss my use case (https://github.com/openmm/openmm/issues/3942) I'm evaluating whether I can use OpenMM for a genetic algorithm. Simulate many small simulations, each possibly with a different...

Here's a great approach. Provided you aren't using PME, and the nonbonded forces are purely local. Make a massive system with many smaller systems, each separated farther than the nonbonded...

It depends on which application you're using. If the truncation of the mantissa is a problem, this might not work. You have to analyze exactly what parts of the code...