use gpgpru::config::{DeviceConfig, DeviceSelector};
use gpgpru::utils::GpuTimer;
use gpgpru::{Device, GpuSlice, Result, create_buffer, kernel, read_buffer};
#[kernel("examples/gemm.slang", "gemm", profiles = [Spirv1_6])]
fn gemm(
A: &GpuSlice<f32>,
B: &GpuSlice<f32>,
C: &mut GpuSlice<f32>,
M: u32,
N: u32,
K: u32,
alpha: f32,
beta: f32,
) {
}
fn main() -> Result<()> {
let config = DeviceConfig::default().with_selector(DeviceSelector::Default);
let device = Device::with(config)?;
println!("Vulkan device loaded successfully.");
let (m, n, k) = (2048, 2048, 2048);
let (alpha, beta) = (1.0f32, 0.0f32);
let host_a = vec![1.0f32; (m * k) as usize];
let host_b = vec![2.0f32; (k * n) as usize];
let mut host_c = vec![0.0f32; (m * n) as usize];
let buf_a = create_buffer(&host_a).with(&device)?;
let buf_b = create_buffer(&host_b).with(&device)?;
let mut buf_c = create_buffer(&host_c).with(&device)?;
let (bm, bn) = (16, 16);
let groups = ((n + bn - 1) / bn, (m + bm - 1) / bm, 1);
for _ in 0..10 {
gemm(&buf_a, &buf_b, &mut buf_c, m, n, k, alpha, beta).queue(&device, groups)?;
device.wait_timeline(device.dispatch()?)?;
}
let timer = GpuTimer::new(&device)?;
let s0 = timer.stamp(&device);
gemm(&buf_a, &buf_b, &mut buf_c, m, n, k, alpha, beta).queue(&device, groups)?;
let s1 = timer.stamp(&device);
device.wait_timeline(device.dispatch()?)?;
let t0 = timer.fetch(&device, s0)?;
let t1 = timer.fetch(&device, s1)?;
let avg_dur = t1 - t0;
let ops_per_gemm = 2.0 * (m as f64) * (n as f64) * (k as f64);
let gflops = (ops_per_gemm / avg_dur.as_secs_f64()) / 1e9;
println!("Execution time (avg): {:?}", avg_dur);
println!("Performance: {:.2} GFLOPS", gflops);
read_buffer(&buf_c).into(&mut host_c).with(&device)?;
let expected = (k as f32) * 2.0;
println!("Validation C[0..5]: {:?}", &host_c[0..5]);
assert_eq!(host_c[0], expected, "GEMM calculation mismatch!");
Ok(())
}