use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
pub struct GatConv {
node_features: Tensor,
edge_index: Vec<(usize, usize)>,
weight: Tensor,
attention: Tensor,
num_nodes: usize,
num_edges: usize,
in_features: usize,
out_features: usize,
leaky_slope: f32,
}
impl GatConv {
pub fn new(
node_features: Tensor,
edge_index: Vec<(usize, usize)>,
weight: Tensor,
attention: Tensor,
leaky_slope: f32,
) -> Result<Self> {
let node_shape = node_features.shape();
if node_shape.len() != 2 {
return Err(BarracudaError::invalid_op(
"gat_conv",
"node_features must be 2D [num_nodes, in_features]",
));
}
let num_nodes = node_shape[0];
let in_features = node_shape[1];
let weight_shape = weight.shape();
if weight_shape.len() != 2 || weight_shape[0] != in_features {
return Err(BarracudaError::invalid_op(
"gat_conv",
"weight must be [in_features, out_features]",
));
}
let out_features = weight_shape[1];
let attention_shape = attention.shape();
let attention_size = attention_shape.iter().product::<usize>();
if attention_size != 2 * out_features {
return Err(BarracudaError::invalid_op(
"gat_conv",
"attention must have 2 * out_features elements",
));
}
let num_edges = edge_index.len();
if num_edges == 0 {
return Err(BarracudaError::invalid_op(
"gat_conv",
"edge_index cannot be empty",
));
}
Ok(Self {
node_features,
edge_index,
weight,
attention,
num_nodes,
num_edges,
in_features,
out_features,
leaky_slope,
})
}
fn wgsl_shader() -> &'static str {
{
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/gnn/gat_conv_f64.wgsl").to_string()
});
SHADER.as_str()
}
}
pub fn execute(self) -> Result<Tensor> {
let device = self.node_features.device();
let edge_data: Vec<u32> = self
.edge_index
.iter()
.flat_map(|(src, dst)| vec![*src as u32, *dst as u32])
.collect();
let edge_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("GATConv Edge Index"),
contents: bytemuck::cast_slice(&edge_data),
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
});
let transformed_size = self.num_nodes * self.out_features;
let transformed_buffer = device.create_buffer_f32(transformed_size)?;
let output_buffer = device.create_buffer_f32(transformed_size)?;
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
num_nodes: u32,
num_edges: u32,
in_features: u32,
out_features: u32,
leaky_slope: f32,
_pad1: u32,
_pad2: u32,
_pad3: u32,
}
let params = Params {
num_nodes: self.num_nodes as u32,
num_edges: self.num_edges as u32,
in_features: self.in_features as u32,
out_features: self.out_features as u32,
leaky_slope: self.leaky_slope,
_pad1: 0,
_pad2: 0,
_pad3: 0,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("GATConv Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
let shader_module = device.compile_shader(Self::wgsl_shader(), Some("GATConv Shader"));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("GATConv 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: true },
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("GATConv 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: edge_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: self.weight.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 4,
resource: self.attention.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 5,
resource: transformed_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 6,
resource: output_buffer.as_entire_binding(),
},
],
});
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("GATConv Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("GATConv Encoder"),
});
let transform_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("GATConv Transform Pipeline"),
layout: Some(&pipeline_layout),
module: &shader_module,
entry_point: Some("transform_features"),
cache: None,
compilation_options: Default::default(),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("GATConv Transform Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&transform_pipeline);
pass.set_bind_group(0, Some(&bind_group), &[]);
let caps = DeviceCapabilities::from_device(device);
let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::MatMul);
let workgroups = (self.num_nodes as u32).div_ceil(optimal_wg_size);
pass.dispatch_workgroups(workgroups, 1, 1);
}
let aggregate_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("GATConv Aggregate Pipeline"),
layout: Some(&pipeline_layout),
module: &shader_module,
entry_point: Some("aggregate"),
cache: None,
compilation_options: Default::default(),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("GATConv Aggregate Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&aggregate_pipeline);
pass.set_bind_group(0, Some(&bind_group), &[]);
let caps = DeviceCapabilities::from_device(device);
let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::MatMul);
let workgroups = (self.num_edges 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.out_features],
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_gat_conv_basic() {
let device = crate::device::test_pool::get_test_device().await;
let num_nodes = 4;
let in_features = 8;
let out_features = 16;
let node_features = Tensor::from_vec_on(
vec![1.0; num_nodes * in_features],
vec![num_nodes, in_features],
device.clone(),
)
.await
.unwrap();
let edge_index = vec![(0, 1), (1, 2), (2, 3), (3, 0)];
let weight = Tensor::from_vec_on(
vec![0.1; in_features * out_features],
vec![in_features, out_features],
device.clone(),
)
.await
.unwrap();
let attention = Tensor::from_vec_on(
vec![0.5; 2 * out_features],
vec![2 * out_features],
device.clone(),
)
.await
.unwrap();
let gat = GatConv::new(node_features, edge_index, weight, attention, 0.01).unwrap();
let output = gat.execute().unwrap();
assert_eq!(output.shape(), &[num_nodes, out_features]);
}
#[tokio::test]
async fn test_gat_conv_edge_cases() {
let device = crate::device::test_pool::get_test_device().await;
let num_nodes = 2;
let in_features = 4;
let out_features = 8;
let node_features = Tensor::from_vec_on(
vec![1.0; num_nodes * in_features],
vec![num_nodes, in_features],
device.clone(),
)
.await
.unwrap();
let edge_index = vec![(0, 1)];
let weight = Tensor::from_vec_on(
vec![0.1; in_features * out_features],
vec![in_features, out_features],
device.clone(),
)
.await
.unwrap();
let attention = Tensor::from_vec_on(
vec![0.5; 2 * out_features],
vec![2 * out_features],
device.clone(),
)
.await
.unwrap();
let gat = GatConv::new(node_features, edge_index, weight, attention, 0.01).unwrap();
let output = gat.execute().unwrap();
assert_eq!(output.shape(), &[num_nodes, out_features]);
}
#[tokio::test]
async fn test_gat_conv_large_batch() {
let device = crate::device::test_pool::get_test_device().await;
let num_nodes = 100;
let in_features = 64;
let out_features = 128;
let node_features = Tensor::from_vec_on(
vec![1.0; num_nodes * in_features],
vec![num_nodes, in_features],
device.clone(),
)
.await
.unwrap();
let mut edge_index = Vec::new();
for i in 0..num_nodes {
edge_index.push((i, (i + 1) % num_nodes));
}
let weight = Tensor::from_vec_on(
vec![0.1; in_features * out_features],
vec![in_features, out_features],
device.clone(),
)
.await
.unwrap();
let attention = Tensor::from_vec_on(
vec![0.5; 2 * out_features],
vec![2 * out_features],
device.clone(),
)
.await
.unwrap();
let gat = GatConv::new(node_features, edge_index, weight, attention, 0.01).unwrap();
let output = gat.execute().unwrap();
assert_eq!(output.shape(), &[num_nodes, out_features]);
}
}