use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
#[derive(Debug, Clone, Copy)]
pub enum AggregationType {
Sum,
Mean,
Max,
}
pub struct GlobalPooling {
node_features: Tensor,
num_nodes: usize,
num_features: usize,
aggregation_type: AggregationType,
}
impl GlobalPooling {
pub fn new(node_features: Tensor, aggregation_type: AggregationType) -> Result<Self> {
let node_shape = node_features.shape();
if node_shape.len() != 2 {
return Err(BarracudaError::invalid_op(
"global_pooling",
"node_features must be 2D [num_nodes, num_features]",
));
}
let num_nodes = node_shape[0];
let num_features = node_shape[1];
Ok(Self {
node_features,
num_nodes,
num_features,
aggregation_type,
})
}
fn wgsl_shader() -> &'static str {
{
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/pooling/global_pooling_f64.wgsl").to_string()
});
std::sync::LazyLock::force(&SHADER).as_str()
}
}
pub fn execute(self) -> Result<Tensor> {
let device = self.node_features.device();
let output_buffer = device.create_buffer_f32(self.num_features)?;
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
num_nodes: u32,
num_features: u32,
aggregation_type: u32,
_pad1: u32,
}
let aggregation_code = match self.aggregation_type {
AggregationType::Sum => 0u32,
AggregationType::Mean => 1u32,
AggregationType::Max => 2u32,
};
let params = Params {
num_nodes: self.num_nodes as u32,
num_features: self.num_features as u32,
aggregation_type: aggregation_code,
_pad1: 0,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("GlobalPooling Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
let shader_module =
device.compile_shader(Self::wgsl_shader(), Some("GlobalPooling Shader"));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("GlobalPooling 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,
},
],
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("GlobalPooling Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: params_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.node_features.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: output_buffer.as_entire_binding(),
},
],
});
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("GlobalPooling Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("GlobalPooling Pipeline"),
layout: Some(&pipeline_layout),
module: &shader_module,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("GlobalPooling Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("GlobalPooling 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::Reduction);
let workgroups = (self.num_features 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_features],
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_global_pooling_sum() {
let device = crate::device::test_pool::get_test_device().await;
let num_nodes = 4;
let num_features = 8;
let node_features = Tensor::from_vec_on(
vec![1.0; num_nodes * num_features],
vec![num_nodes, num_features],
device.clone(),
)
.await
.unwrap();
let pooling = GlobalPooling::new(node_features, AggregationType::Sum).unwrap();
let output = pooling.execute().unwrap();
assert_eq!(output.shape(), &[num_features]);
}
#[tokio::test]
async fn test_global_pooling_mean() {
let device = crate::device::test_pool::get_test_device().await;
let num_nodes = 3;
let num_features = 4;
let node_features = Tensor::from_vec_on(
vec![1.0; num_nodes * num_features],
vec![num_nodes, num_features],
device.clone(),
)
.await
.unwrap();
let pooling = GlobalPooling::new(node_features, AggregationType::Mean).unwrap();
let output = pooling.execute().unwrap();
assert_eq!(output.shape(), &[num_features]);
}
#[tokio::test]
async fn test_global_pooling_max() {
let device = crate::device::test_pool::get_test_device().await;
let num_nodes = 5;
let num_features = 16;
let node_features = Tensor::from_vec_on(
vec![1.0; num_nodes * num_features],
vec![num_nodes, num_features],
device.clone(),
)
.await
.unwrap();
let pooling = GlobalPooling::new(node_features, AggregationType::Max).unwrap();
let output = pooling.execute().unwrap();
assert_eq!(output.shape(), &[num_features]);
}
#[tokio::test]
async fn test_global_pooling_large_batch() {
let device = crate::device::test_pool::get_test_device().await;
let num_nodes = 100;
let num_features = 128;
let node_features = Tensor::from_vec_on(
vec![1.0; num_nodes * num_features],
vec![num_nodes, num_features],
device.clone(),
)
.await
.unwrap();
let pooling = GlobalPooling::new(node_features, AggregationType::Mean).unwrap();
let output = pooling.execute().unwrap();
assert_eq!(output.shape(), &[num_features]);
}
}