huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-17 19:18:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

Updated feature separated benchmarks

Browse Source
This commit is contained in:
Ivar Flakstad
2024-01-09 19:04:31 +01:00
parent ad181f9cdc
commit 87efb5d8eb
4 changed files with 14 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
candle-core/Cargo.toml
View File

@ -48,8 +48,3 @@ metal = ["dep:metal", "dep:candle-metal-kernels"]
[[bench]] [[bench]]
name = "bench_main" name = "bench_main"
harness = false harness = false
[[bench]]
name = "random"
harness = false

2
candle-core/benches/bench_main.rs
View File

@ -1,4 +1,4 @@
mod benchmarks; mod benchmarks;
use criterion::criterion_main; use criterion::criterion_main;
criterion_main!(benchmarks::matmul::benches); criterion_main!(benchmarks::matmul::benches, benchmarks::random::benches);

1
candle-core/benches/benchmarks/mod.rs
View File

@ -1,4 +1,5 @@
pub(crate) mod matmul; pub(crate) mod matmul;
pub(crate) mod random;
use candle_core::{Device, Result}; use candle_core::{Device, Result};

32
candle-core/benches/random.rs → candle-core/benches/benchmarks/random.rs
View File

@ -1,9 +1,10 @@
use crate::benchmarks::{bench_name, device, BenchDevice};
use candle_core::{DType, Device, Tensor}; use candle_core::{DType, Device, Tensor};
use criterion::{black_box, criterion_group, criterion_main, Criterion, Throughput}; use criterion::{black_box, criterion_group, Criterion, Throughput};
use std::time::Instant; use std::time::Instant;
fn rand_uniform(a: &Tensor) { fn rand_uniform(a: &Tensor) {
a.rand_like(0.0, 1.0).unwrap(); a.rand_like(-1.0, 123.0).unwrap();
} }
fn rand_normal(a: &Tensor) { fn rand_normal(a: &Tensor) {
@ -16,14 +17,13 @@ fn criterion_benchmark(c: &mut Criterion) {
let rows = 2048; let rows = 2048;
let cols = 2048; let cols = 2048;
let device = Device::new_metal(0).unwrap(); let d = device().unwrap();
let device2 = device.clone();
let dtype = DType::F32; let dtype = DType::F32;
let tensor = Tensor::zeros((b, rows, cols), dtype, &device).unwrap(); let tensor = Tensor::zeros((b, rows, cols), dtype, &d).unwrap();
let flops = b * rows * cols; let flops = b * rows * cols * dtype.size_in_bytes();
let mut group = c.benchmark_group("metal_random_uniform"); let mut group = c.benchmark_group(bench_name("random_uniform"));
group.throughput(Throughput::Bytes(flops as u64)); group.throughput(Throughput::Bytes(flops as u64));
group.bench_function("iter", move |benches| { group.bench_function("iter", move |benches| {
benches.iter_custom(|iters| { benches.iter_custom(|iters| {
@ -31,19 +31,16 @@ fn criterion_benchmark(c: &mut Criterion) {
for _i in 0..iters { for _i in 0..iters {
rand_uniform(black_box(&tensor)); rand_uniform(black_box(&tensor));
} }
if let Device::Metal(device) = &device { d.sync().unwrap();
device.wait_until_completed().unwrap();
} else {
panic!("Expected metal device");
}
start.elapsed() start.elapsed()
}) })
}); });
group.finish(); group.finish();
let tensor = Tensor::zeros((b, rows, cols), dtype, &device2).unwrap(); let d = device().unwrap();
let tensor = Tensor::zeros((b, rows, cols), dtype, &d).unwrap();
let mut group = c.benchmark_group("metal_random_normal"); let mut group = c.benchmark_group(bench_name("random_normal"));
group.throughput(Throughput::Bytes(flops as u64)); group.throughput(Throughput::Bytes(flops as u64));
group.bench_function("iter", move |benches| { group.bench_function("iter", move |benches| {
benches.iter_custom(|iters| { benches.iter_custom(|iters| {
@ -51,11 +48,7 @@ fn criterion_benchmark(c: &mut Criterion) {
for _i in 0..iters { for _i in 0..iters {
rand_normal(black_box(&tensor)); rand_normal(black_box(&tensor));
} }
if let Device::Metal(device) = &device2 { d.sync().unwrap();
device.wait_until_completed().unwrap();
} else {
panic!("Expected metal device");
}
start.elapsed() start.elapsed()
}) })
}); });
@ -63,4 +56,3 @@ fn criterion_benchmark(c: &mut Criterion) {
} }
criterion_group!(benches, criterion_benchmark); criterion_group!(benches, criterion_benchmark);
criterion_main!(benches);