use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
pub struct GraphBatchNorm {
input: Tensor,
gamma: Tensor,
beta: Tensor,
num_nodes: usize,
num_features: usize,
epsilon: f32,
}
impl GraphBatchNorm {
pub fn new(input: Tensor, gamma: Tensor, beta: Tensor, epsilon: f32) -> Result<Self> {
let input_shape = input.shape();
if input_shape.len() != 2 {
return Err(BarracudaError::invalid_op(
"graph_batch_norm",
"input must be 2D [num_nodes, num_features]",
));
}
let num_nodes = input_shape[0];
let num_features = input_shape[1];
let gamma_shape = gamma.shape();
let gamma_size = gamma_shape.iter().product::<usize>();
if gamma_size != num_features {
return Err(BarracudaError::invalid_op(
"graph_batch_norm",
"gamma must have num_features elements",
));
}
let beta_shape = beta.shape();
let beta_size = beta_shape.iter().product::<usize>();
if beta_size != num_features {
return Err(BarracudaError::invalid_op(
"graph_batch_norm",
"beta must have num_features elements",
));
}
if epsilon <= 0.0 {
return Err(BarracudaError::invalid_op(
"graph_batch_norm",
"epsilon must be positive",
));
}
Ok(Self {
input,
gamma,
beta,
num_nodes,
num_features,
epsilon,
})
}
fn wgsl_shader() -> &'static str {
{
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/norm/graph_batch_norm_f64.wgsl").to_string()
});
std::sync::LazyLock::force(&SHADER).as_str()
}
}
pub fn execute(self) -> Result<Tensor> {
let device = self.input.device();
let mean_buffer = device.create_buffer_f32(self.num_features)?;
let variance_buffer = device.create_buffer_f32(self.num_features)?;
let output_size = self.num_nodes * self.num_features;
let output_buffer = device.create_buffer_f32(output_size)?;
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
num_nodes: u32,
num_features: u32,
epsilon: f32,
_pad1: u32,
}
let params = Params {
num_nodes: self.num_nodes as u32,
num_features: self.num_features as u32,
epsilon: self.epsilon,
_pad1: 0,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("GraphBatchNorm Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
let shader_module =
device.compile_shader(Self::wgsl_shader(), Some("GraphBatchNorm Shader"));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("GraphBatchNorm 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: true },
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: true },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
wgpu::BindGroupLayoutEntry {
binding: 4,
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: 5,
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: 6,
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("GraphBatchNorm Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: params_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.input.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: self.gamma.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: self.beta.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 4,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 5,
resource: mean_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 6,
resource: variance_buffer.as_entire_binding(),
},
],
});
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("GraphBatchNorm Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("GraphBatchNorm Encoder"),
});
let mean_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("GraphBatchNorm Mean Pipeline"),
layout: Some(&pipeline_layout),
module: &shader_module,
entry_point: Some("compute_mean"),
cache: None,
compilation_options: Default::default(),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("GraphBatchNorm Mean Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&mean_pipeline);
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 = (self.num_features as u32).div_ceil(optimal_wg_size);
pass.dispatch_workgroups(workgroups, 1, 1);
}
let variance_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("GraphBatchNorm Variance Pipeline"),
layout: Some(&pipeline_layout),
module: &shader_module,
entry_point: Some("compute_variance"),
cache: None,
compilation_options: Default::default(),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("GraphBatchNorm Variance Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&variance_pipeline);
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 = (self.num_features as u32).div_ceil(optimal_wg_size);
pass.dispatch_workgroups(workgroups, 1, 1);
}
let normalize_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("GraphBatchNorm Normalize Pipeline"),
layout: Some(&pipeline_layout),
module: &shader_module,
entry_point: Some("normalize"),
cache: None,
compilation_options: Default::default(),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("GraphBatchNorm Normalize Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&normalize_pipeline);
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 = (output_size 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,
vec![self.num_nodes, self.num_features],
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_graph_batch_norm_basic() {
let device = crate::device::test_pool::get_test_device().await;
let num_nodes = 4;
let num_features = 8;
let input = Tensor::from_vec_on(
vec![1.0; num_nodes * num_features],
vec![num_nodes, num_features],
device.clone(),
)
.await
.unwrap();
let gamma =
Tensor::from_vec_on(vec![1.0; num_features], vec![num_features], device.clone())
.await
.unwrap();
let beta = Tensor::from_vec_on(vec![0.0; num_features], vec![num_features], device.clone())
.await
.unwrap();
let batch_norm = GraphBatchNorm::new(input, gamma, beta, 1e-5).unwrap();
let output = batch_norm.execute().unwrap();
assert_eq!(output.shape(), &[num_nodes, num_features]);
}
#[tokio::test]
async fn test_graph_batch_norm_large_batch() {
let device = crate::device::test_pool::get_test_device().await;
let num_nodes = 100;
let num_features = 128;
let input = Tensor::from_vec_on(
vec![1.0; num_nodes * num_features],
vec![num_nodes, num_features],
device.clone(),
)
.await
.unwrap();
let gamma =
Tensor::from_vec_on(vec![1.0; num_features], vec![num_features], device.clone())
.await
.unwrap();
let beta = Tensor::from_vec_on(vec![0.0; num_features], vec![num_features], device.clone())
.await
.unwrap();
let batch_norm = GraphBatchNorm::new(input, gamma, beta, 1e-5).unwrap();
let output = batch_norm.execute().unwrap();
assert_eq!(output.shape(), &[num_nodes, num_features]);
}
}