use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::Result;
use crate::tensor::Tensor;
use std::sync::Arc;
pub struct RandomUniformGpu {
device: Arc<crate::device::WgpuDevice>,
n_samples: u32,
bounds: Vec<f32>, seed: u32,
}
impl RandomUniformGpu {
#[must_use]
pub fn new(
device: Arc<crate::device::WgpuDevice>,
n_samples: u32,
bounds: &[(f32, f32)],
seed: u32,
) -> Self {
let bounds_flat: Vec<f32> = bounds.iter().flat_map(|b| <[f32; 2]>::from(*b)).collect();
Self {
device,
n_samples,
bounds: bounds_flat,
seed,
}
}
fn wgsl_shader() -> &'static str {
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/sample/random_uniform_f64.wgsl").to_string()
});
&SHADER
}
pub fn generate(self) -> Result<Tensor> {
let device = &self.device;
let n_dims = self.bounds.len() / 2;
let output_size = (self.n_samples as usize) * n_dims;
let output_buffer = device.create_buffer_f32(output_size)?;
let bounds_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("RandomUniform Bounds"),
contents: bytemuck::cast_slice(&self.bounds),
usage: wgpu::BufferUsages::STORAGE,
});
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
n_samples: u32,
n_dims: u32,
seed: u32,
_pad: u32,
}
let params = Params {
n_samples: self.n_samples,
n_dims: n_dims as u32,
seed: self.seed,
_pad: 0,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("RandomUniform Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM,
});
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("RandomUniform 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("RandomUniform Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: bounds_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("RandomUniform"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("RandomUniform Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("RandomUniform 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("RandomUniform Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("RandomUniform 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 = self.n_samples.div_ceil(optimal_wg_size);
pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
Ok(Tensor::from_buffer(
output_buffer,
vec![self.n_samples as usize, n_dims],
self.device.clone(),
))
}
}
pub fn random_uniform_gpu(
device: Arc<crate::device::WgpuDevice>,
n_samples: u32,
bounds: &[(f32, f32)],
seed: u32,
) -> Result<Tensor> {
RandomUniformGpu::new(device, n_samples, bounds, seed).generate()
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_random_uniform_basic() {
let device = crate::device::test_pool::get_test_device().await;
let bounds = vec![(0.0, 1.0), (-1.0, 1.0)];
let result = random_uniform_gpu(device, 100, &bounds, 42).unwrap();
let data = result.to_vec().unwrap();
assert_eq!(data.len(), 200);
for i in 0..100 {
let x = data[i * 2];
let y = data[i * 2 + 1];
assert!((0.0..=1.0).contains(&x), "x={x} out of [0,1]");
assert!((-1.0..=1.0).contains(&y), "y={y} out of [-1,1]");
}
}
#[tokio::test]
async fn test_random_uniform_different_seeds() {
let device = crate::device::test_pool::get_test_device().await;
let bounds = vec![(0.0, 1.0)];
let r1 = random_uniform_gpu(device.clone(), 10, &bounds, 42).unwrap();
let r2 = random_uniform_gpu(device, 10, &bounds, 99).unwrap();
let d1 = r1.to_vec().unwrap();
let d2 = r2.to_vec().unwrap();
let different = d1.iter().zip(d2.iter()).any(|(a, b)| (a - b).abs() > 1e-6);
assert!(different, "Different seeds should give different results");
}
}