Improve benchmarks layout

candle-core/Cargo.toml

@@ -46,6 +46,6 @@ accelerate = ["dep:libc", "dep:accelerate-src"]
 metal = ["dep:metal", "dep:candle-metal-kernels"]
 
 [[bench]]
-name = "matmul"
+name = "bench_main"
 harness = false
 

candle-core/benches/bench_main.rs

@@ -0,0 +1,4 @@
+mod benchmarks;
+
+use criterion::criterion_main;
+criterion_main!(benchmarks::matmul::benches);

candle-core/benches/benchmarks/matmul.rs

@@ -1,9 +1,7 @@
-mod utils;
-
+use crate::benchmarks::{bench_name, device, BenchDevice};
 use candle_core::{DType, Tensor};
-use criterion::{black_box, criterion_group, criterion_main, Criterion, Throughput};
+use criterion::{black_box, criterion_group, Criterion, Throughput};
 use std::time::Instant;
-use utils::{bench_name, device, BenchDevice};
 
 fn run(a: &Tensor, b: &Tensor) {
     a.matmul(&b.t().unwrap()).unwrap();
@@ -38,4 +36,3 @@ fn criterion_benchmark(c: &mut Criterion) {
 }
 
 criterion_group!(benches, criterion_benchmark);
-criterion_main!(benches);

candle-core/benches/benchmarks/mod.rs

@@ -1,3 +1,5 @@
+pub(crate) mod matmul;
+
 use candle_core::{Device, Result};
 
 pub(crate) trait BenchDevice {