use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
use bytemuck::{Pod, Zeroable};
static SHADER_F64: &str = include_str!("../../shaders/interpolation/rbf_kernel_f64.wgsl");
static SHADER_F32: std::sync::LazyLock<String> =
std::sync::LazyLock::new(|| SHADER_F64.to_string());
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
#[repr(u32)]
#[derive(Default)]
pub enum RbfKernelType {
#[default]
ThinPlateSpline = 0,
Gaussian = 1,
Multiquadric = 2,
InverseMultiquadric = 3,
Cubic = 4,
Quintic = 5,
Linear = 6,
}
pub struct RbfKernel {
x: Tensor, y: Tensor, kernel_type: RbfKernelType,
epsilon: f32, }
impl RbfKernel {
#[must_use]
pub fn new(x: Tensor, y: Tensor, kernel_type: RbfKernelType, epsilon: f32) -> Self {
Self {
x,
y,
kernel_type,
epsilon,
}
}
fn wgsl_shader() -> &'static str {
&SHADER_F32
}
pub fn execute(self) -> Result<Tensor> {
let device = self.x.device();
let x_shape = self.x.shape();
let y_shape = self.y.shape();
if x_shape.len() != 2 || y_shape.len() != 2 {
return Err(BarracudaError::InvalidShape {
expected: vec![0, 0],
actual: x_shape.to_vec(),
});
}
let n_rows = x_shape[0]; let n_cols = y_shape[0]; let x_dims = x_shape[1]; let y_dims = y_shape[1];
if x_dims != y_dims {
return Err(BarracudaError::InvalidShape {
expected: vec![n_rows, x_dims],
actual: vec![n_cols, y_dims],
});
}
let n_dims = x_dims;
let output_size = n_rows * n_cols;
let output_buffer = device.create_buffer_f32(output_size)?;
#[repr(C)]
#[derive(Copy, Clone, Debug, Pod, Zeroable)]
struct RbfKernelParams {
n_rows: u32,
n_cols: u32,
n_dims: u32,
kernel_type: u32,
epsilon: f32,
_pad0: u32,
_pad1: u32,
_pad2: u32,
}
let params = RbfKernelParams {
n_rows: n_rows as u32,
n_cols: n_cols as u32,
n_dims: n_dims as u32,
kernel_type: self.kernel_type as u32,
epsilon: self.epsilon,
_pad0: 0,
_pad1: 0,
_pad2: 0,
};
let params_buffer = device.create_uniform_buffer("RbfKernel Params", ¶ms);
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("RbfKernel BGL"),
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: true },
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::Storage { read_only: false },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
wgpu::BindGroupLayoutEntry {
binding: 3,
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("RbfKernel BG"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.x.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.y.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: params_buffer.as_entire_binding(),
},
],
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("RbfKernel"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("RbfKernel PL"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("RbfKernel 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("RbfKernel Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("RbfKernel Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&pipeline);
pass.set_bind_group(0, Some(&bind_group), &[]);
let workgroup_size = 16;
let workgroups_x = (n_rows as u32).div_ceil(workgroup_size);
let workgroups_y = (n_cols as u32).div_ceil(workgroup_size);
pass.dispatch_workgroups(workgroups_x.max(1), workgroups_y.max(1), 1);
}
device.submit_commands(Some(encoder.finish()));
let output_size = n_rows * n_cols;
let output_data = crate::utils::read_buffer(device, &output_buffer, output_size)?;
Ok(Tensor::new(
output_data,
vec![n_rows, n_cols],
device.clone(),
))
}
}
impl Tensor {
pub fn rbf_kernel(
&self,
other: &Self,
kernel_type: RbfKernelType,
epsilon: f32,
) -> Result<Self> {
RbfKernel::new(self.clone(), other.clone(), kernel_type, epsilon).execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_rbf_kernel_same_points() {
let device = crate::device::test_pool::get_test_device().await;
let x_data = vec![1.0, 2.0, 3.0, 1.0, 2.0, 3.0];
let x = Tensor::from_vec_on(x_data, vec![2, 3], device)
.await
.unwrap();
let k = x
.rbf_kernel(&x, RbfKernelType::ThinPlateSpline, 1.0)
.unwrap();
let kernel = k.to_vec().unwrap();
assert_eq!(kernel.len(), 4);
assert!(
kernel[0].abs() < 1e-5,
"K[0,0] should be 0, got {}",
kernel[0]
);
assert!(
kernel[3].abs() < 1e-5,
"K[1,1] should be 0, got {}",
kernel[3]
);
}
#[tokio::test]
async fn test_rbf_kernel_gaussian() {
let device = crate::device::test_pool::get_test_device().await;
let x_data = vec![0.0, 0.0, 0.0];
let x = Tensor::from_vec_on(x_data, vec![1, 3], device)
.await
.unwrap();
let k = x.rbf_kernel(&x, RbfKernelType::Gaussian, 1.0).unwrap();
let kernel = k.to_vec().unwrap();
assert!(
(kernel[0] - 1.0).abs() < 1e-5,
"Gaussian at r=0 should be 1.0, got {}",
kernel[0]
);
}
#[tokio::test]
async fn test_rbf_kernel_dimensions() {
let device = crate::device::test_pool::get_test_device().await;
let x_data = vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0]; let y_data = vec![7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0];
let x = Tensor::from_vec_on(x_data, vec![2, 3], device.clone())
.await
.unwrap();
let y = Tensor::from_vec_on(y_data, vec![3, 3], device)
.await
.unwrap();
let k = x.rbf_kernel(&y, RbfKernelType::Cubic, 1.0).unwrap();
let shape = k.shape();
assert_eq!(shape, &[2, 3]);
let kernel = k.to_vec().unwrap();
assert_eq!(kernel.len(), 6);
}
}