General Performance Guidelines

Below are some general guidelines for achieving decent performance.

• One good measurement is worth more than a million expert opinions. We have a Benchmark executable that uses Google Benchmark so one can compare different implementations and see how they perform. This executable is only available in release builds.
• Reduce memory allocations. On all modern hardware (many core CPUs, GPUs, and FPGAs), memory is almost always the bottleneck. Memory allocations are especially expensive since this is a quasi-serial process: the OS has to manage memory allocations for all running threads and processes. SpECTRE has various classes to optimize this. For example, there are Variables, TempBuffer, and DynamicBuffer that allow making large contiguous memory allocations that are then used for individual tensor components.