pub fn cpu_scope<R>(f: impl FnOnce() -> R) -> RExpand description
Run f with the GPU gates off on this thread (pure-CPU arm).
Examples found in repository?
examples/qwen_image_denoiser_metal_fixture.rs (lines 361-369)
354fn run(
355 transformer: &QwenImageTransformer,
356 fixture: &Fixture,
357 cpu: bool,
358) -> Result<(Vec<f32>, Duration), String> {
359 let start = Instant::now();
360 let output = if cpu {
361 gpu::cpu_scope(|| {
362 transformer.forward(
363 &fixture.image,
364 &fixture.text,
365 &fixture.shapes,
366 fixture.text_len,
367 fixture.timestep,
368 )
369 })?
370 } else {
371 transformer.forward(
372 &fixture.image,
373 &fixture.text,
374 &fixture.shapes,
375 fixture.text_len,
376 fixture.timestep,
377 )?
378 };
379 Ok((output, start.elapsed()))
380}More examples
examples/gemmbench.rs (lines 34-36)
12fn main() {
13 unsafe { std::env::set_var("CMF_GPU", "1") };
14 let cases = [
15 (384usize, 4096usize, 128usize),
16 (384, 128, 4096),
17 (384, 16384, 128),
18 (384, 128, 16384),
19 ];
20 for (n, k, m) in cases {
21 let x = vec![0.01f32; n * k];
22 let w = vec![0.02f32; m * k];
23 let mut y = vec![0f32; n * m];
24 // warm
25 cortiq_engine::fcd_ops::gemm_nt(&x, &w, &mut y, n, k, m, None);
26 let t = std::time::Instant::now();
27 for _ in 0..10 {
28 cortiq_engine::fcd_ops::gemm_nt(&x, &w, &mut y, n, k, m, None);
29 }
30 let d = t.elapsed().as_secs_f64() / 10.0;
31 let g = 2.0 * n as f64 * k as f64 * m as f64 / d / 1e9;
32 let t2 = std::time::Instant::now();
33 for _ in 0..10 {
34 cortiq_engine::gpu::cpu_scope(|| {
35 cortiq_engine::fcd_ops::gemm_nt(&x, &w, &mut y, n, k, m, None)
36 });
37 }
38 let d2 = t2.elapsed().as_secs_f64() / 10.0;
39 let g2 = 2.0 * n as f64 * k as f64 * m as f64 / d2 / 1e9;
40 println!(
41 "{n}x{k}x{m}: default {:.2} ms ({g:.0} GF/s) cpu_scope {:.2} ms ({g2:.0} GF/s)",
42 d * 1e3,
43 d2 * 1e3
44 );
45 }
46}