Precompilation - #487
Conversation
ce58308 to
f6b5b49
Compare
Codecov Report❌ Patch coverage is
... and 7 files with indirect coverage changes 🚀 New features to boost your workflow:
|
| if p[1] === codomainind(tsrc) && p[2] === domainind(tsrc) | ||
| add!(tdst, tsrc, α, β) | ||
| else | ||
| p2 = (linearize(p), ()) |
There was a problem hiding this comment.
Sorry for my naive question here (I'm not too familiar with this part of the code), but I guess passing the same type of tuple, one containing the indices and the other always being empty, is beneficial for precompilation, and somehow doesn't matter within TensorOperations thatindices are grouped up like this now? So why would you not make the same changes to transpose! and braid in the cases that they don't pass add_transform! first?
There was a problem hiding this comment.
The main point is that for anything that is just raw data (i.e. stridedview), this does not matter and therefore it is beneficial to pick a single canonical option that is then only compiled once. For the tensormaps themselves, this is actually more difficult to get to because the spaces as well as the recoupling data need to be partitioned correctly
Add `src/precompile.jl`: a default-on `@compile_workload` that runs representative contractions, traces and permutations for `Trivial`, `Z2Irrep`, `SU2Irrep` and `FermionParity` over `Float64`/`ComplexF64` and arities [2,3,4]. The heavy kernels (tree transformers, `add_transform!`, per-block BLAS `mul!`) are sector-agnostic, so this also speeds up the first contraction of other symmetries -- including user-defined ones. The exported `precompile_contract(V; eltypes, ndims)` helper is the reusable unit; downstream packages can call it for their own symmetry inside their own `@compile_workload`. The workload is tunable/disable-able through Preferences (`precompile_workload`, `precompile_eltypes`, `precompile_sectors`, `precompile_ndims`), mirroring TensorOperations. Also move `_repartition_body` above the `@generated repartition` that calls it at generation time: the workload is the first code to trigger `repartition` at precompile time, which exposed a latent world-age ordering bug (`UndefVarError` that runtime masks via a later world age). Adds PrecompileTools and Preferences dependencies. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2 participants
This PR adds a precompilation suite to TensorKit.
It takes about 300 seconds on my machine, and is enabled by default.
Given the general migration of the ecosystem towards more heavy precompilation workflows, I think this is warranted, and I'm hoping this improves our CI times as well as the general user experiences.
I've also structured it in such a way that you could conceivably reuse this locally for projects with specific symmetry needs, to ensure faster startup times.
As in general TensorKit is not updated all that often, I hope this is not something that has to be retriggered too much, therefore I felt the ~5mins of precompilation time to be reasonable.