Michael Mara

Create automated script to report performance

1

This is important to avoid performance regressions.

Example in #126

Since Opt uses a typeless layer to communicate parameters, it should have a validation function that ensures there will not be a segfault in the main solver and pinpoints the...

Update terra and CUDA dependencies

3

CUDA 8 seems to play nicer with external profiling tools, and there have been some useful enhancements to terra since the last release.

See #102 for example.

See #102 for an example of why the current message is confusing.

See https://github.com/zdevito/terra/blob/2a2c2b614def5fa9b19e7c38a35116e087e0dabb/tests/cudaoffline.t for an example.

Such as when you have >128 registers/thread. This will help us detect use cases that need to be improved in future versions faster.

This involves solving lots of issues with deploying terra/CUDA remotely.

Inspired by #88.