VoiceModelの実装を統合する
#746
Comments
Merged
|
パフォーマンス的に本当にasync_zipに統合してよいのかどうかを知るため、とりあえずmem版(ZIPファイル全体を一度読み込んでから解凍)をベンチしてみました。やはりzipよりは僅かに遅いように見えますが、実用上は誤差になると思います。
コード[package]
name = "bench-async-zip"
edition = "2021"
publish = false
[[bench]]
name = "bench"
harness = false
[dependencies]
async_zip = { version = "0.0.16", features = ["deflate"] }
criterion = { version = "0.5.1", features = ["html_reports"] }
futures-lite = "2.2.0"
zip = "0.6.6"use std::{
io::{self, Cursor},
time::Duration,
};
use criterion::{criterion_group, criterion_main, Criterion};
criterion_main!(benches);
criterion_group! {
name = benches;
config = Criterion::default()
.warm_up_time(Duration::from_secs(10))
.measurement_time(Duration::from_secs(20));
targets = bench
}
fn bench(criterion: &mut Criterion) {
criterion
.bench_function("zip-v0.6.6", |bencher| bencher.iter(bench_zip))
.bench_function("async_zip-v0.0.16", |bencher| bencher.iter(bench_async_zip));
}
fn bench_zip() -> impl Sized {
use zip::ZipArchive;
let mut zip = ZipArchive::new(Cursor::new(SAMPLE_VVM.to_owned())).unwrap();
let mut entry = zip.by_name("decode.onnx").unwrap();
let mut buf = Vec::with_capacity(entry.size() as _);
io::copy(&mut entry, &mut buf).unwrap();
buf
}
fn bench_async_zip() -> impl Sized {
futures_lite::future::block_on(async {
let zip = async_zip::base::read::mem::ZipFileReader::new(SAMPLE_VVM.to_owned())
.await
.unwrap();
let (idx, _) = zip
.file()
.entries()
.iter()
.enumerate()
.find(|(_, e)| e.filename().as_str().unwrap() == "decode.onnx")
.unwrap();
let mut rdr = zip.reader_with_entry(idx).await.unwrap();
let mut buf = Vec::with_capacity(rdr.entry().uncompressed_size() as _);
rdr.read_to_end_checked(&mut buf).await.unwrap();
buf
})
}
static SAMPLE_VVM: &[u8] = include_bytes!("../sample.vvm"); |
同感です！！ |
|
#828 についてを兼ねた検証を行ったので自分用のメモを残します。検証についての詳細は割愛します。 検証前はasync_zipはやめてzipにすべきかなと思ってましたが、async_zipの方で良さそうです。というかむしろ速度的にもこっちが良くなることがわかりました。 検証[package]
name = "bench-async-zip"
edition = "2021"
publish = false
[[bench]]
name = "bench"
harness = false
[dependencies]
async-fs = "2.1.2"
async_zip = { version = "=0.0.17", features = ["deflate", "tokio-fs"] }
#async_zip = { git = "https://github.com/Majored/rs-async-zip.git", rev = "2d841d600c6d509cb4ecc611001ec339876ca6c9", features = ["deflate", "tokio-fs"] }
criterion = { version = "0.5.1", features = ["html_reports"] }
futures-lite = "2.3.0"
futures-util = { version = "0.3.30", features = ["io"] }
pollster = { version = "0.3.0", features = ["macro"] }
tokio = { version = "1.40.0", features = ["macros", "rt-multi-thread"] }
zip = "2.2.0"use std::{
io::{self, Cursor},
sync::Arc,
};
use criterion::{criterion_group, criterion_main, Criterion};
use futures_lite::AsyncReadExt as _;
use tokio::task::JoinSet;
criterion_main!(benches);
criterion_group! {
name = benches;
//config = Criterion::default()
// .warm_up_time(Duration::from_secs(10))
// .measurement_time(Duration::from_secs(20));
config = Criterion::default();
targets = bench
}
fn bench(criterion: &mut Criterion) {
std::fs::read("./sample.vvm").unwrap();
criterion
.bench_function("zip-v2.2.0-mem-jsons", |bencher| {
bencher.iter(bench_zip_mem_jsons)
})
.bench_function("zip-v2.2.0-mem-onnxs", |bencher| {
bencher.iter(bench_zip_mem_onnxs)
})
.bench_function("zip-v2.2.0-seek-jsons", |bencher| {
bencher.iter(bench_zip_seek_jsons)
})
.bench_function("zip-v2.2.0-seek-onnxs", |bencher| {
bencher.iter(bench_zip_seek_onnxs)
})
.bench_function("async_zip-v0.0.17-mem-jsons", |bencher| {
bencher.iter(bench_async_zip_mem_jsons)
})
.bench_function("async_zip-v0.0.17-mem-onnxs", |bencher| {
bencher.iter(bench_async_zip_mem_rest_file)
})
.bench_function("async_zip-v0.0.17-without-tokio-seek-jsons", |bencher| {
bencher.iter(bench_async_zip_without_tokio_seek_jsons)
})
.bench_function("async_zip-v0.0.17-without-tokio-seek-onnxs", |bencher| {
bencher.iter(bench_async_zip_without_tokio_seek_onnxs)
})
.bench_function("async_zip-v0.0.17-with-tokio-seek-jsons", |bencher| {
bencher.iter(bench_async_zip_with_tokio_seek_jsons)
});
}
fn bench_zip_mem_jsons() -> impl Sized {
bench_zip_mem(&["manifest.json", "metas.json"])
}
fn bench_zip_mem_onnxs() -> impl Sized {
bench_zip_mem(&[
"predict_duration.onnx",
"predict_intonation.onnx",
"decode.onnx",
])
}
#[tokio::main(flavor = "multi_thread", worker_threads = 10)]
async fn bench_zip_mem(filenames: &[&'static str]) -> impl Sized {
let zip = Arc::new(std::fs::read("./sample.vvm").unwrap());
let mut tasks = JoinSet::new();
for &filename in filenames {
let zip = zip.clone();
tasks.spawn_blocking(move || {
let mut zip = zip::ZipArchive::new(Cursor::new(&**zip)).unwrap();
let mut entry = zip.by_name(filename).unwrap();
let mut buf = Vec::with_capacity(entry.size() as _);
io::copy(&mut entry, &mut buf).unwrap();
buf
});
}
tasks.join_all().await
}
fn bench_zip_seek_jsons() -> impl Sized {
bench_zip_seek(&["manifest.json", "metas.json"])
}
fn bench_zip_seek_onnxs() -> impl Sized {
bench_zip_seek(&[
"predict_duration.onnx",
"predict_intonation.onnx",
"decode.onnx",
])
}
fn bench_zip_seek(filenames: &[&str]) -> impl Sized {
let zip = std::fs::File::open("./sample.vvm").unwrap();
let mut zip = zip::ZipArchive::new(zip).unwrap();
filenames
.iter()
.map(|filename| {
let mut entry = zip.by_name(filename).unwrap();
let mut buf = Vec::with_capacity(entry.size() as _);
io::copy(&mut entry, &mut buf).unwrap();
buf
})
.collect::<Vec<_>>()
}
fn bench_async_zip_mem_jsons() -> impl Sized {
bench_async_zip_mem(&["manifest.json", "metas.json"])
}
fn bench_async_zip_mem_rest_file() -> impl Sized {
bench_async_zip_mem(&[
"predict_duration.onnx",
"predict_intonation.onnx",
"decode.onnx",
])
}
#[tokio::main(flavor = "multi_thread", worker_threads = 10)]
async fn bench_async_zip_mem(filenames: &[&'static str]) -> impl Sized {
let zip = tokio::fs::read("./sample.vvm").await.unwrap();
let zip = async_zip::base::read::mem::ZipFileReader::new(zip)
.await
.unwrap();
let zip = Arc::new(zip);
let mut tasks = JoinSet::new();
for &filename in filenames {
let zip = zip.clone();
tasks.spawn(async move {
let (idx, _) = zip
.file()
.entries()
.iter()
.enumerate()
.find(|(_, e)| e.filename().as_str().unwrap() == filename)
.unwrap();
let mut rdr = zip.reader_with_entry(idx).await.unwrap();
let mut buf = Vec::with_capacity(rdr.entry().uncompressed_size() as _);
rdr.read_to_end_checked(&mut buf).await.unwrap();
buf
});
}
tasks.join_all().await
}
fn bench_async_zip_without_tokio_seek_jsons() -> impl Sized {
bench_async_zip_without_tokio_seek(&["manifest.json", "metas.json"])
}
fn bench_async_zip_without_tokio_seek_onnxs() -> impl Sized {
bench_async_zip_without_tokio_seek(&[
"predict_duration.onnx",
"predict_intonation.onnx",
"decode.onnx",
])
}
#[pollster::main]
async fn bench_async_zip_without_tokio_seek(filenames: &[&str]) -> impl Sized {
let zip = futures_util::io::BufReader::new(async_fs::File::open("./sample.vvm").await.unwrap());
let mut zip = async_zip::base::read::seek::ZipFileReader::new(zip)
.await
.unwrap();
let mut ret = vec![];
for filename in filenames {
let (idx, _) = zip
.file()
.entries()
.iter()
.enumerate()
.find(|(_, e)| e.filename().as_str().ok() == Some(filename))
.unwrap();
let mut entry = zip.reader_without_entry(idx).await.unwrap();
let mut buf = vec![];
entry.read_to_end(&mut buf).await.unwrap();
ret.push(buf);
}
ret
}
#[tokio::main(flavor = "current_thread")]
async fn bench_async_zip_with_tokio_seek_jsons() -> impl Sized {
let zip = async_zip::tokio::read::fs::ZipFileReader::new("./sample.vvm")
.await
.unwrap();
let zip = Arc::new(zip);
let mut tasks = JoinSet::new();
for filename in ["manifest.json", "metas.json"] {
let zip = zip.clone();
tasks.spawn(async move {
let (idx, _) = zip
.file()
.entries()
.iter()
.enumerate()
.find(|(_, e)| e.filename().as_str().ok() == Some(filename))
.unwrap();
let mut entry = zip.reader_without_entry(idx).await.unwrap();
let mut buf = vec![];
entry.read_to_end(&mut buf).await.unwrap();
buf
});
}
tasks.join_all().await
}
|
|
async-fs使って思ったのですが他の 現在(#830 での) |
|
キャンセルの導入が円滑になるの良さそうですね！！ コードがどれくらい複雑になっちゃうかがちょっと気になってます。 |
This was referenced Sep 17, 2024
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
内容
現在
blocking::VoiceModelはzipクレート、tokio::VoiceModelはasync_zipクレートで別々に実装されています。blocking::VoiceModelもasync_zipを使うようにして、実装を統一します。Pros 良くなる点
実装がバラけているのを解消できる
Cons 悪くなる点
実現方法
async_zipはtokio無しでも動くことを利用する。Rust 1.75からtrait定義に
-> impl Traitを書けることを利用して上手く抽象化を行い、blocking::VoiceModelはasync_zip::base、tokio::VoiceModelはasync_zip::tokioで駆動するようにする（async_zip v0.0.16の場合）。VOICEVOXのバージョン
N/A
OSの種類/ディストリ/バージョン
その他
The text was updated successfully, but these errors were encountered: