Ugh, NVidia's nvcc GPU kernel compiler has some very painful regressions for us going from v10 to v11.

Our heavily optimized kernels that do molecular mechanics calculations compile just fine on nvcc 10.2 and just barely fit into the 64 register limit imposed by the launch bounds.

But for some reason, the nvcc v11 blows waaaay past the register limits by like 70%.

It's not great.

· · Web · 0 · 0 · 1
Sign in to participate in the conversation
Mastodon for Tech Folks

This Mastodon instance is for people interested in technology. Discussions aren't limited to technology, because tech folks shouldn't be limited to technology either!