use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
#[repr(C)]
#[derive(Copy, Clone, Debug, bytemuck::Pod, bytemuck::Zeroable)]
struct SpectralNormParams {
rows: u32,
cols: u32,
num_iterations: u32,
_padding: u32,
}
pub struct SpectralNorm {
weight: Tensor,
u: Tensor,
v: Tensor,
num_iterations: u32,
}
impl SpectralNorm {
pub fn new(weight: Tensor, u: Tensor, v: Tensor, num_iterations: u32) -> Result<Self> {
if num_iterations == 0 {
return Err(BarracudaError::invalid_op(
"SpectralNorm",
"num_iterations must be > 0",
));
}
Ok(Self {
weight,
u,
v,
num_iterations,
})
}
fn wgsl_shader() -> &'static str {
{
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/norm/spectral_norm_f64.wgsl").to_string()
});
std::sync::LazyLock::force(&SHADER).as_str()
}
}
pub fn execute(self) -> Result<Tensor> {
let device = self.weight.device();
let weight_shape = self.weight.shape();
if weight_shape.len() != 2 {
return Err(BarracudaError::invalid_op(
"SpectralNorm",
format!("weight must be 2D [rows, cols], got shape {weight_shape:?}"),
));
}
let rows = weight_shape[0];
let cols = weight_shape[1];
if self.u.shape() != [rows] {
return Err(BarracudaError::invalid_op(
"SpectralNorm",
format!("u must be 1D [rows], got shape {:?}", self.u.shape()),
));
}
if self.v.shape() != [cols] {
return Err(BarracudaError::invalid_op(
"SpectralNorm",
format!("v must be 1D [cols], got shape {:?}", self.v.shape()),
));
}
let params = SpectralNormParams {
rows: rows as u32,
cols: cols as u32,
num_iterations: self.num_iterations,
_padding: 0,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("SpectralNorm Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM,
});
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("SpectralNorm 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::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::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("SpectralNorm Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.weight.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.u.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: self.v.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: params_buffer.as_entire_binding(),
},
],
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("SpectralNorm"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("SpectralNorm Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("SpectralNorm 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("SpectralNorm Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("SpectralNorm Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&pipeline);
pass.set_bind_group(0, Some(&bind_group), &[]);
use crate::device::{DeviceCapabilities, WorkloadType};
let caps = DeviceCapabilities::from_device(device);
let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::ElementWise);
let max_dim = rows.max(cols);
let workgroups = (max_dim as u32).div_ceil(optimal_wg_size);
pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
Ok(self.weight)
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_spectral_norm_basic() {
let device = crate::device::test_pool::get_test_device().await;
let rows = 10;
let cols = 8;
let weight = Tensor::from_vec_on(vec![0.1; rows * cols], vec![rows, cols], device.clone())
.await
.unwrap();
let u = Tensor::from_vec_on(vec![1.0; rows], vec![rows], device.clone())
.await
.unwrap();
let v = Tensor::from_vec_on(vec![1.0; cols], vec![cols], device.clone())
.await
.unwrap();
let result = SpectralNorm::new(weight, u, v, 1)
.unwrap()
.execute()
.unwrap();
assert_eq!(result.shape(), &[rows, cols]);
}
}