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 LocalResponseNormParams {
batch_size: u32,
channels: u32,
height: u32,
width: u32,
size: u32,
alpha: f32,
beta: f32,
k: f32,
}
pub struct LocalResponseNorm {
input: Tensor,
size: usize,
alpha: f32,
beta: f32,
k: f32,
}
impl LocalResponseNorm {
pub fn new(input: Tensor, size: usize, alpha: f32, beta: f32, k: f32) -> Result<Self> {
let shape = input.shape();
if shape.len() != 4 {
return Err(BarracudaError::invalid_op(
"local_response_norm",
"input must be 4D tensor [B, C, H, W]",
));
}
if size == 0 {
return Err(BarracudaError::invalid_op(
"local_response_norm",
"size must be positive",
));
}
Ok(Self {
input,
size,
alpha,
beta,
k,
})
}
fn wgsl_shader() -> &'static str {
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/norm/local_response_norm_f64.wgsl").to_string()
});
std::sync::LazyLock::force(&SHADER).as_str()
}
pub fn execute(self) -> Result<Tensor> {
let device = self.input.device();
let shape = self.input.shape();
let batch_size = shape[0];
let channels = shape[1];
let height = shape[2];
let width = shape[3];
let output_size = batch_size * channels * height * width;
let output_buffer = device.create_buffer_f32(output_size)?;
let params = LocalResponseNormParams {
batch_size: batch_size as u32,
channels: channels as u32,
height: height as u32,
width: width as u32,
size: self.size as u32,
alpha: self.alpha,
beta: self.beta,
k: self.k,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("local_response_norm_params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM,
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("local_response_norm_shader"));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("local_response_norm_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: 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::Uniform,
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
],
});
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("local_response_norm_pipeline_layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("local_response_norm_pipeline"),
layout: Some(&pipeline_layout),
module: &shader,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("local_response_norm_bind_group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.input.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: params_buffer.as_entire_binding(),
},
],
});
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("local_response_norm_encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("local_response_norm_pass"),
timestamp_writes: None,
});
compute_pass.set_pipeline(&pipeline);
compute_pass.set_bind_group(0, Some(&bind_group), &[]);
let caps = DeviceCapabilities::from_device(device);
let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::Reduction);
let workgroups_x = (width as u32).div_ceil(8);
let workgroups_y = (height as u32).div_ceil(8);
let workgroups_z = ((batch_size * channels) as u32).div_ceil(optimal_wg_size);
compute_pass.dispatch_workgroups(workgroups_x, workgroups_y, workgroups_z);
}
device.submit_commands(Some(encoder.finish()));
Ok(Tensor::from_buffer(
output_buffer,
vec![batch_size, channels, height, width],
device.clone(),
))
}
}
impl Tensor {
pub fn local_response_norm(self, size: usize, alpha: f32, beta: f32, k: f32) -> Result<Self> {
LocalResponseNorm::new(self, size, alpha, beta, k)?.execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_local_response_norm_basic() {
let device = crate::device::test_pool::get_test_device().await;
let input = Tensor::from_vec_on(vec![1.0; 3 * 4 * 4], vec![1, 3, 4, 4], device.clone())
.await
.unwrap();
let output = input.local_response_norm(5, 1e-4, 0.75, 1.0).unwrap();
let result = output.to_vec().unwrap();
assert_eq!(output.shape(), &[1, 3, 4, 4]);
assert_eq!(result.len(), 48);
assert!(result.iter().all(|&x| x.is_finite()));
}
}