factor::table: Implement a batched version w/ improved performance #2143
Conversation
|
Rebased to deal with merge conflict, applied |
|
Fixed the benchmark not to rely on unstable features; surprisingly, this slightly improved the performance of the benchmark code, |
|
please create a BENCHMARKING.md file like src/uu/ls/BENCHMARKING.md to reproduce your benchmark :) |
OK? Though I'm not sure what I'd say beyond “run |
This is good enough :) |
|
|
||
| Instead, we could run the µbenchmarks in a simulated CPU with [`cachegrind`], | ||
| measure execution “time” in that model (in CI), and use it to detect and report | ||
| performance improvements and regressions. |
There was a problem hiding this comment.
@sylvestre Do you agree it would make sense to use something like iai in factor's CI, and automatically report significant performance changes on the PR's discussion?
There was a problem hiding this comment.
why not? Are you aware of any success story with iai ?
There was a problem hiding this comment.
Not yet, but the tool is very new. I seem to recall (non-Rust) projects having decent success using cachegrind in CI that way, though.
This will be easier to adapt to working with multiple numbers to process at once.
The factor_chunk implementation is a strawman, but getting it in place allows us to set up the microbenchmarking etc.
This keeps the traversal of `P_INVS_U64` (a large table) to a single pass in-order, rather than `CHUNK_SIZE` passes.
Previous version would perform an amount of work proportional to `CHUNK_SIZE`, so this wasn't a valid way to benchmark at multiple values of that constant. The `TryInto` implementation for `&mut [T]` to `&mut [T; N]` relies on `const` generics, and is available in (stable) Rust v1.51 and later.
Adding array-init v2.0.0 Updating cast v0.2.5 -> v0.2.6 Adding pest v2.1.3 Updating rustc_version v0.2.3 -> v0.3.3 Adding semver v0.11.0 Adding semver-parser v0.10.2 Updating serde v1.0.125 -> v1.0.126 Updating serde_derive v1.0.125 -> v1.0.126 Adding ucd-trie v0.1.3 Adding uu_factor_benches v0.0.0 (#tests/benches/factor)
|
Rebase to deal with merge conflict in |
|
Tests are blocked on BurntSushi/quickcheck#282, so let's move ahead for now. |
|
CI breakage is unrelated: |
factor::table::factor_chunktablebenchmark to work on stable RustAdd testsFuture work
factor_chunkIterator<u64>intoIterator<Factors>and automatically using the batched/vectorized implementation internally.
factor::table::factor_chunkin the CLI application (presumably via the high-level API)