mirror of
https://github.com/huggingface/candle.git
synced 2025-06-17 11:08:52 +00:00
9cff7bc3f4
* Allow the use of tf32 accumulation in matmul. * Better timings. * Dummy versions for use when cuda is not enabled.
31 lines
867 B
Rust
31 lines
867 B
Rust
#[cfg(feature = "accelerate")]
|
|
extern crate accelerate_src;
|
|
|
|
#[cfg(feature = "mkl")]
|
|
extern crate intel_mkl_src;
|
|
|
|
use anyhow::Result;
|
|
use candle_core::{Device, Tensor};
|
|
|
|
fn main() -> Result<()> {
|
|
let device = Device::new_cuda(0)?;
|
|
let x = Tensor::randn(0f32, 1.0, (8 * 4096, 8 * 4096), &device)?;
|
|
candle_core::cuda::set_gemm_reduced_precision_f32(false);
|
|
let _x1 = x.matmul(&x)?;
|
|
drop(_x1);
|
|
let start_time = std::time::Instant::now();
|
|
let _x1 = x.matmul(&x)?;
|
|
device.synchronize()?;
|
|
println!("fp32: {:?}", start_time.elapsed());
|
|
drop(_x1);
|
|
candle_core::cuda::set_gemm_reduced_precision_f32(true);
|
|
let _x1 = x.matmul(&x)?;
|
|
drop(_x1);
|
|
let start_time = std::time::Instant::now();
|
|
let _x1 = x.matmul(&x)?;
|
|
device.synchronize()?;
|
|
println!("tf32: {:?}", start_time.elapsed());
|
|
drop(_x1);
|
|
Ok(())
|
|
}
|