Update with feature separated benchmarks

candle-core/Cargo.toml

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

candle-core/benches/bench_main.rs

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

candle-core/benches/benchmarks/fill.rs

@@ -0,0 +1,43 @@
+use crate::benchmarks::{bench_name, device, BenchDevice};
+use candle_core::{DType, Device, Tensor};
+use criterion::{black_box, criterion_group, Criterion, Throughput};
+use std::time::Instant;
+
+fn run(shape: (usize, usize, usize), dtype: DType, device: &Device) {
+    Tensor::ones(shape, dtype, device).unwrap();
+}
+
+fn run_fill_benchmark(c: &mut Criterion, name: &str, dtype: DType) {
+    let b = 1;
+    let rows = 4096;
+    let columns = 4096;
+
+    let flops = b * rows * columns * dtype.size_in_bytes();
+
+    let device = device().unwrap();
+
+    let mut group = c.benchmark_group(bench_name(name));
+    group.throughput(Throughput::Bytes(flops as u64));
+    group.bench_function("iter", move |bencher| {
+        bencher.iter_custom(|iters| {
+            let start = Instant::now();
+            for _i in 0..iters {
+                run(
+                    black_box((b, rows, columns)),
+                    black_box(dtype),
+                    black_box(&device),
+                );
+            }
+            device.sync().unwrap();
+            start.elapsed()
+        })
+    });
+    group.finish();
+}
+
+fn criterion_benchmark(c: &mut Criterion) {
+    run_fill_benchmark(c, "fill_u8", DType::U8);
+    run_fill_benchmark(c, "fill_f32", DType::F32);
+}
+
+criterion_group!(benches, criterion_benchmark);

candle-core/benches/benchmarks/mod.rs

@@ -1,3 +1,4 @@
+pub(crate) mod fill;
 pub(crate) mod matmul;
 
 use candle_core::{Device, Result};

candle-core/benches/fill.rs

@@ -1,67 +0,0 @@
-use candle_core::{DType, Device, Tensor};
-use criterion::{black_box, criterion_group, criterion_main, Criterion, Throughput};
-use std::time::Instant;
-
-fn run(shape: (usize, usize, usize), dtype: DType, device: &Device) {
-    Tensor::ones(shape, dtype, device).unwrap();
-}
-
-fn criterion_benchmark(c: &mut Criterion) {
-    let b = 1;
-    let rows = 4096;
-    let columns = 4096;
-
-    let flops = b * rows * columns;
-
-    let device1 = Device::new_metal(0).unwrap();
-    let device2 = device1.clone();
-
-    let mut group = c.benchmark_group("fill_metal_u8");
-    group.throughput(Throughput::Bytes(flops as u64));
-    group.bench_function("iter", move |bencher| {
-        bencher.iter_custom(|iters| {
-            let start = Instant::now();
-            for _i in 0..iters {
-                run(
-                    black_box((b, rows, columns)),
-                    black_box(DType::U8),
-                    black_box(&device1),
-                );
-            }
-            if let Device::Metal(device) = &device1 {
-                device.wait_until_completed().unwrap();
-            } else {
-                panic!("Expected metal device");
-            }
-            start.elapsed()
-        })
-    });
-    group.finish();
-
-    let mut group = c.benchmark_group("fill_metal_f32");
-    group.throughput(Throughput::Bytes(
-        (flops * DType::F32.size_in_bytes()) as u64,
-    ));
-    group.bench_function("iter", move |bencher| {
-        bencher.iter_custom(|iters| {
-            let start = Instant::now();
-            for _i in 0..iters {
-                run(
-                    black_box((b, rows, columns)),
-                    black_box(DType::F32),
-                    black_box(&device2),
-                );
-            }
-            if let Device::Metal(device) = &device2 {
-                device.wait_until_completed().unwrap();
-            } else {
-                panic!("Expected metal device");
-            }
-            start.elapsed()
-        })
-    });
-    group.finish();
-}
-
-criterion_group!(benches, criterion_benchmark);
-criterion_main!(benches);