burn/backend-comparison

Burn Benchmark

This crate allows to compare backend computation times, from tensor operations to complex models.

burnbench CLI

This crate comes with a CLI binary called burnbench which can be executed via cargo run --release --bin burnbench.

Note that you need to run the release target of burnbench otherwise you won't be able to share your benchmark results.

The end of options argument -- is used to pass arguments to the burnbench application. For instance cargo run --bin burnbench -- list passes the list argument to burnbench effectively calling burnbench list.

Commands

List benches and backends

To list all the available benches and backends use the list command:

> cargo run --release --bin burnbench -- list
    Finished dev [unoptimized] target(s) in 0.10s
     Running `target/debug/burnbench list`
Available Backends:
- candle-cpu
- candle-cuda
- candle-metal
- ndarray
- ndarray-blas-accelerate
- ndarray-blas-netlib
- ndarray-blas-openblas
- tch-cpu
- tch-gpu
- wgpu
- wgpu-fusion

Available Benchmarks:
- binary
- custom-gelu
- data
- matmul
- unary

Run benchmarks

To run a given benchmark against a specific backend we use the run command with the arguments --benches and --backends respectively. In the following example we execute the unary benchmark against the wgpu-fusion backend:

> cargo run --release --bin burnbench -- run --benches unary --backends wgpu-fusion

Shorthands can be used, the following command line is the same:

> cargo run --release --bin burnbench -- run -b unary -B wgpu-fusion

Multiple benchmarks and backends can be passed on the same command line. In this case, all the combinations of benchmarks with backends will be executed.

> cargo run --bin burnbench -- run --benches unary binary --backends wgpu-fusion tch-gpu
     Running `target/release/burnbench run --benches unary binary --backends wgpu-fusion wgpu`
Executing the following benchmark and backend combinations (Total: 4):
- Benchmark: unary, Backend: wgpu-fusion
- Benchmark: binary, Backend: wgpu-fusion
- Benchmark: unary, Backend: tch-gpu
- Benchmark: binary, Backend: tch-gpu
Running benchmarks...

By default burnbench uses a compact output with a progress bar which hides the compilation logs and benchmarks results as they are executed. If a benchmark failed to run, the --verbose flag can be used to investigate the error.

Authentication and benchmarks sharing

Burnbench can upload benchmark results to our servers so that users can share their results with the community and we can use this information to drive the development of Burn. The results can be explored on Burn website.

Sharing results is opt-in and it is enabled with the --share arguments passed to the run command:

> cargo run --release --bin burnbench -- run --share --benches unary --backends wgpu-fusion

To be able to upload results you must be authenticated. We only support GitHub authentication. To authenticate run the auth command, then follow the URL to enter your device code and authorize the Burnbench application:

> cargo run --release --bin burnbench -- auth

If everything is fine you should get a confirmation in the terminal that your token has been saved to the burn cache directory.

We don't store any of your personal information. An anonymized user name will be attributed to you and displayed in the terminal once you are authenticated. For instance:

🔑 Your username is: CuteFlame

You can now use the --share argument to upload and share your benchmarks. A URL to the results will displayed at the end of the report table.

Note that your access token will be refreshed automatically so you should not need to reauthorize the application again except if your refresh token itself becomes invalid.

Execute benchmarks with cargo

To execute a benchmark against a given backend using only cargo is done with the bench command. In this case the backend is a feature of this crate.

> cargo bench --features wgpu-fusion

Add a new benchmark

To add a new benchmark it must be first declared in the Cargo.toml file of this crate:

[[bench]]
name = "mybench"
harness = false

Then it must be registered in the BenchmarkValues enumeration:

#[derive(Debug, Clone, PartialEq, Eq, ValueEnum, Display, EnumIter)]
pub(crate) enum BackendValues {
    // ...
    #[strum(to_string = "mybench")]
    MyBench,
    // ...
}

Create a new file mybench.rs in the benches directory and implement the Benchmark trait over your benchmark structure. Then implement the bench function. At last call the macro backend_comparison::bench_on_backend!() in the main function.

Add a new backend

You can easily register and new backend in the BackendValues enumeration:

#[derive(Debug, Clone, PartialEq, Eq, ValueEnum, Display, EnumIter)]
pub(crate) enum BackendValues {
    // ...
    #[strum(to_string = "mybackend")]
    MyBackend,
    // ...
}

Then update the macro bench_on_backend to support the newly registered backend.