use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::Result;
use crate::tensor::Tensor;
pub struct Laguerre {
input: Tensor,
n: u32,
alpha: f32,
}
impl Laguerre {
#[must_use]
pub fn new(input: Tensor, n: u32, alpha: f32) -> Self {
Self { input, n, alpha }
}
#[must_use]
pub fn simple(input: Tensor, n: u32) -> Self {
Self::new(input, n, 0.0)
}
fn wgsl_shader() -> &'static str {
include_str!("../shaders/special/laguerre.wgsl")
}
pub fn execute(self) -> Result<Tensor> {
let device = self.input.device();
let size: usize = self.input.shape().iter().product();
let output_buffer = device.create_buffer_f32(size)?;
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
size: u32,
n: u32,
alpha: f32,
}
let params = Params {
size: size as u32,
n: self.n,
alpha: self.alpha,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Laguerre Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM,
});
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("Laguerre Bind Group Layout"),
entries: &[
wgpu::BindGroupLayoutEntry {
binding: 0,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Storage { read_only: true },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
wgpu::BindGroupLayoutEntry {
binding: 1,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Storage { read_only: false },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
wgpu::BindGroupLayoutEntry {
binding: 2,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Uniform,
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
],
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("Laguerre Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.input.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: params_buffer.as_entire_binding(),
},
],
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("Laguerre"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("Laguerre Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("Laguerre Pipeline"),
layout: Some(&pipeline_layout),
module: &shader,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("Laguerre Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("Laguerre Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&pipeline);
pass.set_bind_group(0, Some(&bind_group), &[]);
let caps = DeviceCapabilities::from_device(device);
let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::ElementWise);
let workgroups = (size as u32).div_ceil(optimal_wg_size);
pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
Ok(Tensor::from_buffer(
output_buffer,
self.input.shape().to_vec(),
device.clone(),
))
}
}
impl Tensor {
pub fn laguerre(self, n: u32, alpha: f32) -> Result<Self> {
Laguerre::new(self, n, alpha).execute()
}
pub fn laguerre_simple(self, n: u32) -> Result<Self> {
Laguerre::simple(self, n).execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_laguerre_l0() {
let device = crate::device::test_pool::get_test_device().await;
let data = vec![0.0, 1.0, 2.0, 5.0];
let input = Tensor::new(data, vec![4], device);
let output = input.laguerre_simple(0).unwrap();
let result = output.to_vec().unwrap();
for &v in &result {
assert!((v - 1.0).abs() < 1e-5, "L_0 should be 1, got {v}");
}
}
#[tokio::test]
async fn test_laguerre_l1() {
let device = crate::device::test_pool::get_test_device().await;
let data = vec![0.0, 1.0, 3.0];
let input = Tensor::new(data.clone(), vec![3], device);
let output = input.laguerre_simple(1).unwrap();
let result = output.to_vec().unwrap();
let expected = [1.0, 0.0, -2.0];
for (i, &v) in result.iter().enumerate() {
assert!(
(v - expected[i]).abs() < 1e-5,
"L_1({}) = {}, expected {}",
data[i],
v,
expected[i]
);
}
}
#[tokio::test]
async fn test_laguerre_l2() {
let device = crate::device::test_pool::get_test_device().await;
let data = vec![0.0, 1.0, 2.0];
let input = Tensor::new(data.clone(), vec![3], device);
let output = input.laguerre_simple(2).unwrap();
let result = output.to_vec().unwrap();
for (i, &v) in result.iter().enumerate() {
let x = data[i];
let expected = f32::midpoint(4.0f32.mul_add(-x, x * x), 2.0);
assert!(
(v - expected).abs() < 1e-4,
"L_2({x}) = {v}, expected {expected}"
);
}
}
#[tokio::test]
async fn test_laguerre_generalized() {
let device = crate::device::test_pool::get_test_device().await;
let data = vec![0.0, 2.0];
let input = Tensor::new(data, vec![2], device);
let output = input.laguerre(1, 1.0).unwrap();
let result = output.to_vec().unwrap();
assert!(
(result[0] - 2.0).abs() < 1e-5,
"L_1^(1)(0) = {}, expected 2",
result[0]
);
assert!(
(result[1] - 0.0).abs() < 1e-5,
"L_1^(1)(2) = {}, expected 0",
result[1]
);
}
}