use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::Result;
use crate::tensor::Tensor;
pub struct Stack {
tensors: Vec<Tensor>,
dim: usize,
}
impl Stack {
pub fn new(tensors: Vec<Tensor>, dim: usize) -> Result<Self> {
if tensors.is_empty() {
return Err(crate::error::BarracudaError::InvalidInput {
message: "Cannot stack empty tensor list".to_string(),
});
}
let first_shape = tensors[0].shape();
for (i, tensor) in tensors.iter().enumerate().skip(1) {
if tensor.shape() != first_shape {
return Err(crate::error::BarracudaError::InvalidInput {
message: format!(
"All tensors must have same shape. Tensor 0: {:?}, Tensor {}: {:?}",
first_shape,
i,
tensor.shape()
),
});
}
}
if dim > first_shape.len() {
return Err(crate::error::BarracudaError::InvalidInput {
message: format!("dim {} exceeds tensor rank {}", dim, first_shape.len()),
});
}
Ok(Self { tensors, dim })
}
fn wgsl_shader() -> &'static str {
{
static S: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/tensor/stack_f64.wgsl").to_string()
});
&S
}
}
pub fn execute(self) -> Result<Tensor> {
let device = self.tensors[0].device();
let num_tensors = self.tensors.len();
let tensor_size: usize = self.tensors[0].shape().iter().product();
let mut output_shape = self.tensors[0].shape().to_vec();
output_shape.insert(self.dim, num_tensors);
let output_size: usize = output_shape.iter().product();
let input_size = num_tensors * tensor_size;
let input_buffer = device.create_buffer_f32(input_size)?;
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("Stack Copy Encoder"),
});
for (i, tensor) in self.tensors.iter().enumerate() {
let offset = i * tensor_size * std::mem::size_of::<f32>();
encoder.copy_buffer_to_buffer(
tensor.buffer(),
0,
&input_buffer,
offset as u64,
(tensor_size * std::mem::size_of::<f32>()) as u64,
);
}
device.submit_commands(Some(encoder.finish()));
let output_buffer = device.create_buffer_f32(output_size)?;
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct StackParams {
num_tensors: u32,
tensor_size: u32,
output_size: u32,
stack_dim: u32,
}
let params = StackParams {
num_tensors: num_tensors as u32,
tensor_size: tensor_size as u32,
output_size: output_size as u32,
stack_dim: self.dim as u32,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Stack Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
let shader_module = device.compile_shader(Self::wgsl_shader(), Some("Stack Shader"));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("Stack 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("Stack Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: params_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: input_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("Stack Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("Stack 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("Stack Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("Stack 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::ElementWise);
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,
output_shape,
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_stack_basic() {
let device = crate::device::test_pool::get_test_device().await;
let t1 = Tensor::from_data(&[1.0, 2.0], vec![2], device.clone()).unwrap();
let t2 = Tensor::from_data(&[3.0, 4.0], vec![2], device).unwrap();
let stacked = Stack::new(vec![t1, t2], 0).unwrap().execute().unwrap();
assert_eq!(stacked.shape(), &vec![2, 2]);
}
#[tokio::test]
async fn test_stack_multiple() {
let device = crate::device::test_pool::get_test_device().await;
let tensors: Vec<Tensor> = (0..5)
.map(|i| Tensor::from_data(&[i as f32; 4], vec![2, 2], device.clone()).unwrap())
.collect();
let stacked = Stack::new(tensors, 0).unwrap().execute().unwrap();
assert_eq!(stacked.shape(), &vec![5, 2, 2]);
}
#[tokio::test]
async fn test_stack_empty() {
let _device = crate::device::test_pool::get_test_device().await;
assert!(Stack::new(vec![], 0).is_err());
}
#[tokio::test]
async fn test_stack_shape_mismatch() {
let device = crate::device::test_pool::get_test_device().await;
let t1 = Tensor::from_data(&[1.0, 2.0], vec![2], device.clone()).unwrap();
let t2 = Tensor::from_data(&[3.0, 4.0, 5.0], vec![3], device).unwrap();
assert!(Stack::new(vec![t1, t2], 0).is_err());
}
#[tokio::test]
async fn test_stack_dim_invalid() {
let device = crate::device::test_pool::get_test_device().await;
let t1 = Tensor::from_data(&[1.0, 2.0], vec![2], device.clone()).unwrap();
let t2 = Tensor::from_data(&[3.0, 4.0], vec![2], device).unwrap();
assert!(Stack::new(vec![t1, t2], 10).is_err());
}
}