use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
#[repr(C)]
#[derive(Copy, Clone, Debug, bytemuck::Pod, bytemuck::Zeroable)]
struct WeightNormParams {
num_weights: u32,
dim: u32,
_padding: [u32; 2],
}
pub struct WeightNorm {
v: Tensor,
g: Tensor,
dim: u32,
}
impl WeightNorm {
pub fn new(v: Tensor, g: Tensor, dim: u32) -> Result<Self> {
Ok(Self { v, g, dim })
}
fn wgsl_shader() -> &'static str {
{
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/norm/weight_norm_f64.wgsl").to_string()
});
std::sync::LazyLock::force(&SHADER).as_str()
}
}
pub fn execute(self) -> Result<Tensor> {
let device = self.v.device();
let v_shape = self.v.shape();
if v_shape.is_empty() {
return Err(BarracudaError::invalid_op(
"WeightNorm",
format!("v must have at least 1 dimension, got shape {v_shape:?}"),
));
}
let num_weights = self.v.len();
let output_buffer = device.create_buffer_f32(num_weights)?;
let params = WeightNormParams {
num_weights: num_weights as u32,
dim: self.dim,
_padding: [0; 2],
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("WeightNorm Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM,
});
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("WeightNorm 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: 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("WeightNorm Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.v.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.g.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("WeightNorm"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("WeightNorm Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("WeightNorm 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("WeightNorm Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("WeightNorm 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 = (num_weights 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,
v_shape.to_vec(),
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_weight_norm_basic() {
let device = crate::device::test_pool::get_test_device().await;
let num_weights = 20;
let v = Tensor::from_vec_on(vec![1.0; num_weights], vec![num_weights], device.clone())
.await
.unwrap();
let g = Tensor::from_vec_on(vec![2.0], vec![1], device.clone())
.await
.unwrap();
let result = WeightNorm::new(v, g, 0).unwrap().execute().unwrap();
assert_eq!(result.shape(), &[num_weights]);
}
}