How to speed up the build with pyo3?

[antimora]

I have a simple use case. I call a python function from Rust. Building my rust code takes up to 30 seconds. Is there anything I could do to speed up the build? I even disabled all default features, but it does not help (default-features = false).

I noticed the initial build also includes pyo3-macros, which takes forever to compile. I certainly do not use any Py macros. Perhaps, it's causing the compilation slowness? Just an idea ...

[davidhewitt]

Disabling default features should mean that pyo3-macros does not get compiled. Can you provide a repro which demonstrates your problem?

[antimora]

OK. Let me provide with a code that can reproduce this.

[antimora]

Disabling the defaults still compiles pyo3-macros.

[adamreichold]

Could you share at least your Cargo.toml or the output of cargo tree --edges features to see if something else pulls in the feature (for example the numpy crate would do that).

[antimora]

@adamreichold, @davidhewitt provided the files and the build outputs down below. Thank you.

[birkenfeld]

Another question, is the 30 seconds for the initial build or later, after you change only your crate's code?

[antimora]

30 seconds are for later, after the initial build, triggered by changes to my code.

[antimora]

Here is my build time:

[pyo3-slowness]$ time cargo build # Clean project
   Compiling pyo3-slowness v0.1.0 (/workspaces/pyo3-slowness)
    Finished dev [unoptimized + debuginfo] target(s) in 6.34s

real    0m6.422s
user    0m0.360s
sys     0m0.788s



[pyo3-slowness]$ time cargo build # clean project with pyo3 dependency first time
   Compiling autocfg v1.1.0
   Compiling target-lexicon v0.12.4
   Compiling libc v0.2.135
   Compiling proc-macro2 v1.0.47
   Compiling once_cell v1.15.0
   Compiling quote v1.0.21
   Compiling unicode-ident v1.0.5
   Compiling syn v1.0.102
   Compiling cfg-if v1.0.0
   Compiling parking_lot_core v0.9.4
   Compiling scopeguard v1.1.0
   Compiling smallvec v1.10.0
   Compiling version_check v0.9.4
   Compiling rawpointer v0.2.1
   Compiling indoc v1.0.7
   Compiling unindent v0.1.10
   Compiling num-traits v0.2.15
   Compiling lock_api v0.4.9
   Compiling memoffset v0.6.5
   Compiling num-integer v0.1.45
   Compiling matrixmultiply v0.3.2
   Compiling ahash v0.7.6
   Compiling getrandom v0.2.7
   Compiling pyo3-build-config v0.17.2
   Compiling pyo3-macros-backend v0.17.2
   Compiling num-complex v0.4.2
   Compiling parking_lot v0.12.1
   Compiling ndarray v0.15.6
   Compiling pyo3-macros v0.17.2
   Compiling pyo3-ffi v0.17.2
   Compiling pyo3 v0.17.2
   Compiling numpy v0.17.2
   Compiling pyo3-slowness v0.1.0 (/workspaces/pyo3-slowness)
    Finished dev [unoptimized + debuginfo] target(s) in 3m 06s

real    3m6.448s
user    0m45.863s
sys     0m34.678s
[pyo3-slowness]$ 

[pyo3-slowness]$ time cargo build   #Build with code
   Compiling pyo3-slowness v0.1.0 (/workspaces/pyo3-slowness)
    Finished dev [unoptimized + debuginfo] target(s) in 22.55s

real    0m22.620s
user    0m0.801s
sys     0m2.569s

[pyo3-slowness]$ time cargo build # build without any modification
    Finished dev [unoptimized + debuginfo] target(s) in 0.37s

real    0m0.446s
user    0m0.065s
sys     0m0.064s


[pyo3-slowness]$ time cargo build # without some modification to trigger a new build
   Compiling pyo3-slowness v0.1.0 (/workspaces/pyo3-slowness)
    Finished dev [unoptimized + debuginfo] target(s) in 24.38s

real    0m24.435s
user    0m0.961s
sys     0m2.747s

Cargo.toml:

[package]
name = "pyo3-slowness"
version = "0.1.0"
edition = "2021"

# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html

[dependencies]
numpy = "0.17.2"
pyo3 = { version = "0.17.2",  default-features = false, features = ["auto-initialize"] }

Code main.rs:

use numpy::PyArray1;
use pyo3::{types::PyDict, PyResult, Python};

fn main() {
    let a = Python::with_gil(|py| -> PyResult<Vec<f64>> {
        let signal_module = py.import("scipy.signal")?;
        let locals = PyDict::new(py);
        locals.set_item("signal_module", signal_module)?;
        locals.set_item("window_name", "hann")?;
        locals.set_item("window_size", 5)?;
        let pyarray: &PyArray1<f64> = py
            .eval(
                "signal_module.get_window(window_name, window_size, True)",
                Some(locals),
                None,
            )?
            .extract()?;
        let readonly = pyarray.readonly();
        let items = readonly.as_slice()?;
        let mut result: Vec<f64> = vec![];
        result.extend(items);
        Ok(result)
    });
    println!("{:?}", a);
}

[birkenfeld]

With this main.rs and Cargo.toml, I can't reproduce slowness when triggering a rebuild; initial build time is 15 seconds, but rebuild with only main.rs changed is in the subsecond range for me.

(Toolchain is 1.64 stable.)

[antimora]

[pyo3-slowness]$ cargo tree --edges features
pyo3-slowness v0.1.0 (/workspaces/pyo3-slowness)
├── numpy feature "default"
│   └── numpy v0.17.2
│       ├── ahash feature "default"
│       │   ├── ahash v0.7.6
│       │   │   ├── getrandom feature "default"
│       │   │   │   └── getrandom v0.2.7
│       │   │   │       ├── libc v0.2.135
│       │   │   │       └── cfg-if feature "default"
│       │   │   │           └── cfg-if v1.0.0
│       │   │   └── once_cell feature "alloc"
│       │   │       ├── once_cell v1.15.0
│       │   │       └── once_cell feature "race"
│       │   │           └── once_cell v1.15.0
│       │   │   [build-dependencies]
│       │   │   └── version_check feature "default"
│       │   │       └── version_check v0.9.4
│       │   └── ahash feature "std"
│       │       └── ahash v0.7.6 (*)
│       ├── libc feature "default"
│       │   ├── libc v0.2.135
│       │   └── libc feature "std"
│       │       └── libc v0.2.135
│       ├── ndarray feature "default"
│       │   ├── ndarray v0.15.6
│       │   │   ├── num-complex v0.4.2
│       │   │   │   └── num-traits feature "i128"
│       │   │   │       └── num-traits v0.2.15
│       │   │   │           [build-dependencies]
│       │   │   │           └── autocfg feature "default"
│       │   │   │               └── autocfg v1.1.0
│       │   │   ├── num-integer v0.1.45
│       │   │   │   └── num-traits v0.2.15 (*)
│       │   │   │   [build-dependencies]
│       │   │   │   └── autocfg feature "default" (*)
│       │   │   ├── num-traits v0.2.15 (*)
│       │   │   ├── matrixmultiply feature "cgemm"
│       │   │   │   └── matrixmultiply v0.3.2
│       │   │   │       └── rawpointer feature "default"
│       │   │   │           └── rawpointer v0.2.1
│       │   │   └── rawpointer feature "default" (*)
│       │   └── ndarray feature "std"
│       │       ├── ndarray v0.15.6 (*)
│       │       ├── matrixmultiply feature "std"
│       │       │   └── matrixmultiply v0.3.2 (*)
│       │       └── num-traits feature "std"
│       │           └── num-traits v0.2.15 (*)
│       ├── num-complex feature "default"
│       │   ├── num-complex v0.4.2 (*)
│       │   └── num-complex feature "std"
│       │       ├── num-complex v0.4.2 (*)
│       │       └── num-traits feature "std" (*)
│       ├── num-traits feature "default"
│       │   ├── num-traits v0.2.15 (*)
│       │   └── num-traits feature "std" (*)
│       ├── num-integer feature "default"
│       │   ├── num-integer v0.1.45 (*)
│       │   └── num-integer feature "std"
│       │       ├── num-integer v0.1.45 (*)
│       │       └── num-traits feature "std" (*)
│       └── pyo3 feature "macros"
│           ├── pyo3 v0.17.2
│           │   ├── cfg-if feature "default" (*)
│           │   ├── libc feature "default" (*)
│           │   ├── indoc feature "default"
│           │   │   └── indoc v1.0.7 (proc-macro)
│           │   ├── memoffset feature "default"
│           │   │   └── memoffset v0.6.5
│           │   │       [build-dependencies]
│           │   │       └── autocfg feature "default" (*)
│           │   ├── parking_lot feature "default"
│           │   │   └── parking_lot v0.12.1
│           │   │       ├── lock_api feature "default"
│           │   │       │   └── lock_api v0.4.9
│           │   │       │       └── scopeguard v1.1.0
│           │   │       │       [build-dependencies]
│           │   │       │       └── autocfg feature "default" (*)
│           │   │       └── parking_lot_core feature "default"
│           │   │           └── parking_lot_core v0.9.4
│           │   │               ├── cfg-if feature "default" (*)
│           │   │               ├── libc feature "default" (*)
│           │   │               └── smallvec feature "default"
│           │   │                   └── smallvec v1.10.0
│           │   ├── pyo3-ffi feature "default"
│           │   │   └── pyo3-ffi v0.17.2
│           │   │       └── libc feature "default" (*)
│           │   │       [build-dependencies]
│           │   │       ├── pyo3-build-config feature "default"
│           │   │       │   └── pyo3-build-config v0.17.2
│           │   │       │       ├── once_cell feature "default"
│           │   │       │       │   ├── once_cell v1.15.0
│           │   │       │       │   └── once_cell feature "std"
│           │   │       │       │       ├── once_cell v1.15.0
│           │   │       │       │       └── once_cell feature "alloc"
│           │   │       │       │           ├── once_cell v1.15.0
│           │   │       │       │           └── once_cell feature "race"
│           │   │       │       │               └── once_cell v1.15.0
│           │   │       │       └── target-lexicon feature "default"
│           │   │       │           └── target-lexicon v0.12.4
│           │   │       │       [build-dependencies]
│           │   │       │       └── target-lexicon feature "default" (*)
│           │   │       └── pyo3-build-config feature "resolve-config"
│           │   │           └── pyo3-build-config v0.17.2 (*)
│           │   ├── pyo3-macros feature "default"
│           │   │   └── pyo3-macros v0.17.2 (proc-macro)
│           │   │       ├── proc-macro2 v1.0.47
│           │   │       │   └── unicode-ident feature "default"
│           │   │       │       └── unicode-ident v1.0.5
│           │   │       ├── pyo3-macros-backend feature "default"
│           │   │       │   └── pyo3-macros-backend v0.17.2
│           │   │       │       ├── proc-macro2 v1.0.47 (*)
│           │   │       │       ├── quote v1.0.21
│           │   │       │       │   └── proc-macro2 v1.0.47 (*)
│           │   │       │       ├── syn feature "clone-impls"
│           │   │       │       │   └── syn v1.0.102
│           │   │       │       │       ├── proc-macro2 v1.0.47 (*)
│           │   │       │       │       ├── quote v1.0.21 (*)
│           │   │       │       │       └── unicode-ident feature "default" (*)
│           │   │       │       ├── syn feature "derive"
│           │   │       │       │   └── syn v1.0.102 (*)
│           │   │       │       ├── syn feature "extra-traits"
│           │   │       │       │   └── syn v1.0.102 (*)
│           │   │       │       ├── syn feature "full"
│           │   │       │       │   └── syn v1.0.102 (*)
│           │   │       │       ├── syn feature "parsing"
│           │   │       │       │   └── syn v1.0.102 (*)
│           │   │       │       └── syn feature "printing"
│           │   │       │           ├── syn v1.0.102 (*)
│           │   │       │           └── syn feature "quote"
│           │   │       │               └── syn v1.0.102 (*)
│           │   │       ├── quote feature "default"
│           │   │       │   ├── quote v1.0.21 (*)
│           │   │       │   └── quote feature "proc-macro"
│           │   │       │       ├── quote v1.0.21 (*)
│           │   │       │       └── proc-macro2 feature "proc-macro"
│           │   │       │           └── proc-macro2 v1.0.47 (*)
│           │   │       ├── syn feature "default"
│           │   │       │   ├── syn v1.0.102 (*)
│           │   │       │   ├── syn feature "clone-impls" (*)
│           │   │       │   ├── syn feature "derive" (*)
│           │   │       │   ├── syn feature "parsing" (*)
│           │   │       │   ├── syn feature "printing" (*)
│           │   │       │   └── syn feature "proc-macro"
│           │   │       │       ├── syn v1.0.102 (*)
│           │   │       │       ├── proc-macro2 feature "proc-macro" (*)
│           │   │       │       ├── quote feature "proc-macro" (*)
│           │   │       │       └── syn feature "quote" (*)
│           │   │       ├── syn feature "extra-traits" (*)
│           │   │       └── syn feature "full" (*)
│           │   └── unindent feature "default"
│           │       └── unindent v0.1.10
│           │   [build-dependencies]
│           │   ├── pyo3-build-config feature "default" (*)
│           │   └── pyo3-build-config feature "resolve-config" (*)
│           ├── pyo3 feature "indoc"
│           │   └── pyo3 v0.17.2 (*)
│           ├── pyo3 feature "pyo3-macros"
│           │   └── pyo3 v0.17.2 (*)
│           └── pyo3 feature "unindent"
│               └── pyo3 v0.17.2 (*)
└── pyo3 feature "auto-initialize"
    └── pyo3 v0.17.2 (*)

[antimora]

@adamreichold Thanks for your comments. Sounds like there is no general solution. Is there any quick workaround for my code, perhaps? I can convert this question to an open issue to deal with compilation slowness in general. Honestly, it would be great if build time could be improved. It's very easy to lose focus if the compilation takes too long.

[adamreichold]

Is there any quick workaround for my code, perhaps?

It depends whether your sample above is really the extent of your interaction with Python. If that is the case, you could avoid the numpy crate altogether, evaluate list(signal_module.get_window(window_name, window_size, True)) and extract a Vec<f64> directly from that. But that would not cover more extensive interactions with ndarrays.

[antimora]

Yeah, I was thinking of not relying on ndarray for my use case. Your suggestion should work. Currently, I am trying to use list in the call but getting this error. Most likely because the built-in types are not imported. Do you know how to import the python builtins by default? I am not finding this in the documentation.

Err(PyErr { type: <class 'NameError'>, value: NameError("name 'list' is not defined"), traceback: Some(<traceback object at 0xffff6b31ac00>) })

[antimora]

Found a fix for missing list name. One must import "builtins" and prequalify list like this "builtins.list". Now I don't have to import numpy but this only helped me to shave off 6 seconds from the build. I'll run the compiler report suggested by @davidhewitt to see more details on the slowdown.

[adamreichold]

I am sorry, I think using list was actually bad advice or at least pointless: With the current PyO3 release, you should be able to extract a Vec<f64> directly from the one-dimensional NumPy array as it implements the relevant parts of the sequence protocol.

[davidhewitt]

I think there's also other questions here like what actually is taking the time to compile.

Consider running cargo +nightly rustc -- -Zself-profile and using the tools to report what's apparently consuming the compilation time? https://doc.rust-lang.org/nightly/unstable-book/compiler-flags/self-profile.html