use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::Result;
use crate::tensor::Tensor;
use std::sync::Arc;
pub struct SobolGpu {
device: Arc<crate::device::WgpuDevice>,
n_samples: u32,
n_dims: u32,
skip: u32,
}
impl SobolGpu {
#[must_use]
pub fn new(device: Arc<crate::device::WgpuDevice>, n_samples: u32, n_dims: u32) -> Self {
Self {
device,
n_samples,
n_dims,
skip: 0,
}
}
#[must_use]
pub fn with_skip(mut self, skip: u32) -> Self {
self.skip = skip;
self
}
fn wgsl_shader() -> &'static str {
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/sample/sobol_f64.wgsl").to_string()
});
&SHADER
}
pub fn generate(self) -> Result<Tensor> {
let device = &self.device;
let output_size = (self.n_samples * self.n_dims) as usize;
let output_buffer = device.create_buffer_f32(output_size)?;
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
n_samples: u32,
n_dims: u32,
skip: u32,
_pad: u32,
}
let params = Params {
n_samples: self.n_samples,
n_dims: self.n_dims,
skip: self.skip,
_pad: 0,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Sobol Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM,
});
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("Sobol Bind Group Layout"),
entries: &[
wgpu::BindGroupLayoutEntry {
binding: 0,
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: 1,
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("Sobol Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: params_buffer.as_entire_binding(),
},
],
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("Sobol"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("Sobol Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("Sobol 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("Sobol Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("Sobol 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, self.n_dims as usize],
self.device.clone(),
))
}
}
pub fn sobol_gpu(
device: Arc<crate::device::WgpuDevice>,
n_samples: u32,
n_dims: u32,
) -> Result<Tensor> {
SobolGpu::new(device, n_samples, n_dims).generate()
}
pub fn sobol_gpu_skip(
device: Arc<crate::device::WgpuDevice>,
n_samples: u32,
n_dims: u32,
skip: u32,
) -> Result<Tensor> {
SobolGpu::new(device, n_samples, n_dims)
.with_skip(skip)
.generate()
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_sobol_basic() {
let device = crate::device::test_pool::get_test_device().await;
let result = sobol_gpu(device, 16, 2).unwrap();
let data = result.to_vec().unwrap();
assert_eq!(data.len(), 32);
for &v in &data {
assert!((0.0..1.0).contains(&v), "Value {v} out of range");
}
assert!(data[0].abs() < 1e-6, "First point should be 0");
}
#[tokio::test]
async fn test_sobol_low_discrepancy() {
let device = crate::device::test_pool::get_test_device().await;
let result = sobol_gpu(device, 64, 1).unwrap();
let data = result.to_vec().unwrap();
let q1 = data.iter().filter(|&&x| x < 0.25).count();
let q2 = data.iter().filter(|&&x| (0.25..0.5).contains(&x)).count();
let q3 = data.iter().filter(|&&x| (0.5..0.75).contains(&x)).count();
let q4 = data.iter().filter(|&&x| x >= 0.75).count();
assert!((10..=22).contains(&q1), "Q1 has {q1} points");
assert!((10..=22).contains(&q2), "Q2 has {q2} points");
assert!((10..=22).contains(&q3), "Q3 has {q3} points");
assert!((10..=22).contains(&q4), "Q4 has {q4} points");
}
#[tokio::test]
async fn test_sobol_with_skip() {
let device = crate::device::test_pool::get_test_device().await;
let result = sobol_gpu_skip(device, 8, 1, 8).unwrap();
let data = result.to_vec().unwrap();
assert!(
data[0].abs() > 1e-6,
"First point after skip shouldn't be 0"
);
}
}