use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
pub struct MessagePassing {
node_features: Tensor,
edge_index: Tensor,
edge_features: Option<Tensor>,
num_nodes: usize,
num_edges: usize,
node_feat_dim: usize,
edge_feat_dim: usize,
message_dim: usize,
aggr_type: u32, }
impl MessagePassing {
pub fn new(
node_features: Tensor,
edge_index: Tensor,
edge_features: Option<Tensor>,
aggr_type: u32,
) -> Result<Self> {
let node_shape = node_features.shape();
let num_nodes = node_shape[0];
let node_feat_dim = node_shape[1..].iter().product::<usize>();
let edge_shape = edge_index.shape();
if edge_shape.len() != 2 || edge_shape[1] != 2 {
return Err(BarracudaError::invalid_op(
"message_passing",
"edge_index must be [num_edges, 2]",
));
}
let num_edges = edge_shape[0];
let edge_feat_dim = if let Some(ref ef) = edge_features {
ef.shape()[1..].iter().product::<usize>()
} else {
0
};
Ok(Self {
node_features,
edge_index,
edge_features,
num_nodes,
num_edges,
node_feat_dim,
edge_feat_dim,
message_dim: node_feat_dim, aggr_type,
})
}
fn wgsl_shader() -> &'static str {
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/math/message_passing_f64.wgsl").to_string()
});
&SHADER
}
pub fn execute(self) -> Result<Tensor> {
let device = self.node_features.device();
let output_size = self.num_nodes * self.node_feat_dim;
let output_buffer = device.create_buffer_f32(output_size)?;
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
num_nodes: u32,
num_edges: u32,
node_feat_dim: u32,
edge_feat_dim: u32,
message_dim: u32,
aggr_type: u32,
}
let params = Params {
num_nodes: self.num_nodes as u32,
num_edges: self.num_edges as u32,
node_feat_dim: self.node_feat_dim as u32,
edge_feat_dim: self.edge_feat_dim as u32,
message_dim: self.message_dim as u32,
aggr_type: self.aggr_type,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("MessagePassing Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
let shader_module =
device.compile_shader(Self::wgsl_shader(), Some("MessagePassing Shader"));
let mut entries = vec![
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: true },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
];
if self.edge_features.is_some() {
entries.push(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,
});
}
entries.extend([
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::Uniform,
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
]);
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("MessagePassing Bind Group Layout"),
entries: &entries,
});
let mut bind_entries = vec![
wgpu::BindGroupEntry {
binding: 0,
resource: self.node_features.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.edge_index.buffer().as_entire_binding(),
},
];
if let Some(ref ef) = self.edge_features {
bind_entries.push(wgpu::BindGroupEntry {
binding: 2,
resource: ef.buffer().as_entire_binding(),
});
}
let placeholder = device.placeholder_buffer();
bind_entries.extend([
wgpu::BindGroupEntry {
binding: 3,
resource: placeholder.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 4,
resource: placeholder.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 5,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 6,
resource: params_buffer.as_entire_binding(),
},
]);
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("MessagePassing Bind Group"),
layout: &bind_group_layout,
entries: &bind_entries,
});
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("MessagePassing Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("MessagePassing 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("MessagePassing Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("MessagePassing 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::MatMul);
let workgroups = (self.num_nodes 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.node_feat_dim],
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_message_passing_basic() {
let device = crate::device::test_pool::get_test_device().await;
let num_nodes = 4;
let num_edges = 3;
let node_feat_dim = 8;
let node_features = Tensor::from_vec_on(
vec![1.0; num_nodes * node_feat_dim],
vec![num_nodes, node_feat_dim],
device.clone(),
)
.await
.unwrap();
let edge_index = Tensor::from_vec_on(
vec![0.0, 1.0, 1.0, 2.0, 2.0, 3.0],
vec![num_edges, 2],
device.clone(),
)
.await
.unwrap();
let output = MessagePassing::new(node_features, edge_index, None, 0)
.unwrap()
.execute()
.unwrap();
assert_eq!(output.shape(), &[num_nodes, node_feat_dim]);
}
}