const SHADER_F64: &str = include_str!("../shaders/gradient/clip_grad_norm_f64.wgsl");
use crate::device::DeviceCapabilities;
use crate::error::Result;
use crate::tensor::Tensor;
pub struct ClipGradNorm {
gradients: Tensor,
max_norm: f32,
}
impl ClipGradNorm {
pub fn new(gradients: Tensor, max_norm: f32) -> Result<Self> {
if max_norm < 0.0 {
return Err(crate::error::BarracudaError::InvalidInput {
message: "max_norm must be non-negative".to_string(),
});
}
Ok(Self {
gradients,
max_norm,
})
}
fn wgsl_shader() -> &'static str {
static SHADER: std::sync::LazyLock<String> =
std::sync::LazyLock::new(|| SHADER_F64.to_string());
SHADER.as_str()
}
pub fn execute(self) -> Result<Tensor> {
let device = self.gradients.device();
let size: usize = self.gradients.shape().iter().product();
let input_buffer = self.gradients.buffer();
let caps = DeviceCapabilities::from_device(device);
let num_workgroups = caps.dispatch_1d(size as u32);
let norm_buffer_size = num_workgroups.max(1) as usize;
let norm_buffer = device.create_buffer_f32(norm_buffer_size)?;
let output_buffer = device.create_buffer_f32(size)?;
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
size: u32,
max_norm: f32,
_pad1: u32,
_pad2: u32,
}
let params = Params {
size: size as u32,
max_norm: self.max_norm,
_pad1: 0,
_pad2: 0,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("ClipGradNorm Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
let shader_module = device.compile_shader(Self::wgsl_shader(), Some("ClipGradNorm Shader"));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("ClipGradNorm Bind Group Layout"),
entries: &[
wgpu::BindGroupLayoutEntry {
binding: 0,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Uniform,
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::Storage { read_only: false },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
],
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("ClipGradNorm Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: params_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: input_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: norm_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: output_buffer.as_entire_binding(),
},
],
});
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("ClipGradNorm Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let compute_pipeline_norm =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("ClipGradNorm Norm Pipeline"),
layout: Some(&pipeline_layout),
module: &shader_module,
entry_point: Some("compute_norm"),
cache: None,
compilation_options: Default::default(),
});
let compute_pipeline_norm_final =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("ClipGradNorm Norm Final Pipeline"),
layout: Some(&pipeline_layout),
module: &shader_module,
entry_point: Some("compute_norm_final"),
cache: None,
compilation_options: Default::default(),
});
let compute_pipeline_clip =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("ClipGradNorm Clip Pipeline"),
layout: Some(&pipeline_layout),
module: &shader_module,
entry_point: Some("clip_gradients"),
cache: None,
compilation_options: Default::default(),
});
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("ClipGradNorm Encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("ClipGradNorm Norm Pass"),
timestamp_writes: None,
});
compute_pass.set_pipeline(&compute_pipeline_norm);
compute_pass.set_bind_group(0, Some(&bind_group), &[]);
compute_pass.dispatch_workgroups(num_workgroups, 1, 1);
}
if num_workgroups > 1 {
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("ClipGradNorm Norm Final Pass"),
timestamp_writes: None,
});
compute_pass.set_pipeline(&compute_pipeline_norm_final);
compute_pass.set_bind_group(0, Some(&bind_group), &[]);
compute_pass.dispatch_workgroups(1, 1, 1);
}
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("ClipGradNorm Clip Pass"),
timestamp_writes: None,
});
compute_pass.set_pipeline(&compute_pipeline_clip);
compute_pass.set_bind_group(0, Some(&bind_group), &[]);
let workgroups = caps.dispatch_1d(size as u32);
compute_pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
Ok(Tensor::from_buffer(
output_buffer,
self.gradients.shape().to_vec(),
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_clip_grad_norm_basic() {
let device = crate::device::test_pool::get_test_device().await;
let gradients = Tensor::from_data(&[3.0, 4.0], vec![2], device).unwrap();
let clipped = ClipGradNorm::new(gradients, 1.0)
.unwrap()
.execute()
.unwrap();
let result = clipped.to_vec().unwrap();
let norm: f32 = result.iter().map(|&x| x * x).sum::<f32>().sqrt();
assert!((norm - 1.0).abs() < 0.1); }
#[tokio::test]
async fn test_clip_grad_norm_no_clip() {
let device = crate::device::test_pool::get_test_device().await;
let gradients = Tensor::from_data(&[0.1, 0.2, 0.3], vec![3], device).unwrap();
let clipped = ClipGradNorm::new(gradients, 1.0)
.unwrap()
.execute()
.unwrap();
let result = clipped.to_vec().unwrap();
let norm: f32 = result.iter().map(|&x| x * x).sum::<f32>().sqrt();
assert!(norm < 1.0);
}
#[tokio::test]
async fn test_clip_grad_norm_zero() {
let device = crate::device::test_pool::get_test_device().await;
let gradients = Tensor::from_data(&[0.0, 0.0, 0.0], vec![3], device).unwrap();
let clipped = ClipGradNorm::new(gradients, 1.0)
.unwrap()
.execute()
.unwrap();
let result = clipped.to_vec().unwrap();
assert_eq!(result, vec![0.0, 0.0, 0.0]);
}
#[tokio::test]
async fn test_clip_grad_norm_large() {
let device = crate::device::test_pool::get_test_device().await;
let data: Vec<f32> = (0..1000).map(|i| (i % 10) as f32).collect();
let gradients = Tensor::from_data(&data, vec![1000], device).unwrap();
let clipped = ClipGradNorm::new(gradients, 100.0)
.unwrap()
.execute()
.unwrap();
let result = clipped.to_vec().unwrap();
assert_eq!(result.len(), 1000);
let norm: f32 = result.iter().map(|&x| x * x).sum::<f32>().sqrt();
assert!(norm <= 100.0 + 1.0); }
#[tokio::test]
async fn test_clip_grad_norm_invalid() {
let device = crate::device::test_pool::get_test_device().await;
let gradients = Tensor::from_data(&[1.0, 2.0], vec![2], device).unwrap();
assert!(ClipGradNorm::new(gradients, -1.0).is_err());
}
}