We're getting closer by the week!
Single-threaded solves are now within 1% of v2.9. We have ground to gain on multi-threaded operation. On my 18-core Mac, multi-threaded solves are roughly 7.5x faster than single-threaded solves. The ideal would be 18x. That's never achievable, but we can surely narrow the gap.
Issue #139 suggests integrating changes from DDSS. My experiments with it have not yet born fruit.
Switching compiler flags from -mtune=generic to -mcpu=native improved both single-threaded and multi-threaded performance by about 1% on my Mac, but there's a price to pay. The binary must be compiled on the machine it will run on. If we plan to host downloadable binaries this wouldn't fly. Also, the flag makes no difference on Windows or Linux. I've gated it behind a bazel flag in a local branch. I won't submit a PR unless there's a demand for this.
We're getting closer by the week!
Single-threaded solves are now within 1% of v2.9. We have ground to gain on multi-threaded operation. On my 18-core Mac, multi-threaded solves are roughly 7.5x faster than single-threaded solves. The ideal would be 18x. That's never achievable, but we can surely narrow the gap.
Issue #139 suggests integrating changes from DDSS. My experiments with it have not yet born fruit.
Switching compiler flags from -mtune=generic to -mcpu=native improved both single-threaded and multi-threaded performance by about 1% on my Mac, but there's a price to pay. The binary must be compiled on the machine it will run on. If we plan to host downloadable binaries this wouldn't fly. Also, the flag makes no difference on Windows or Linux. I've gated it behind a bazel flag in a local branch. I won't submit a PR unless there's a demand for this.