use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::Result;
use crate::tensor::Tensor;
pub struct Tile {
input: Tensor,
repeats: Vec<usize>,
}
impl Tile {
pub fn new(input: Tensor, repeats: Vec<usize>) -> Result<Self> {
let num_dims = input.shape().len();
if repeats.len() != num_dims {
return Err(crate::error::BarracudaError::InvalidInput {
message: format!(
"Repeats length {} doesn't match tensor rank {}",
repeats.len(),
num_dims
),
});
}
if repeats.contains(&0) {
return Err(crate::error::BarracudaError::InvalidInput {
message: "Repeats must be positive".to_string(),
});
}
Ok(Self { input, repeats })
}
fn wgsl_shader() -> &'static str {
{
static S: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/tensor/tile_f64.wgsl").to_string()
});
&S
}
}
pub fn execute(self) -> Result<Tensor> {
let device = self.input.device();
let input_shape = self.input.shape();
let num_dims = input_shape.len();
let output_shape: Vec<usize> = input_shape
.iter()
.zip(self.repeats.iter())
.map(|(&s, &r)| s * r)
.collect();
let total_size: usize = output_shape.iter().product();
let mut input_strides = vec![1; num_dims];
for i in (0..num_dims - 1).rev() {
input_strides[i] = input_strides[i + 1] * input_shape[i + 1];
}
let mut output_strides = vec![1; num_dims];
for i in (0..num_dims - 1).rev() {
output_strides[i] = output_strides[i + 1] * output_shape[i + 1];
}
let input_buffer = self.input.buffer();
let input_shape_buffer =
device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Tile Input Shape"),
contents: bytemuck::cast_slice(
&input_shape.iter().map(|&x| x as u32).collect::<Vec<_>>(),
),
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
});
let output_shape_buffer =
device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Tile Output Shape"),
contents: bytemuck::cast_slice(
&output_shape.iter().map(|&x| x as u32).collect::<Vec<_>>(),
),
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
});
let input_strides_buffer =
device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Tile Input Strides"),
contents: bytemuck::cast_slice(
&input_strides.iter().map(|&x| x as u32).collect::<Vec<_>>(),
),
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
});
let output_strides_buffer =
device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Tile Output Strides"),
contents: bytemuck::cast_slice(
&output_strides.iter().map(|&x| x as u32).collect::<Vec<_>>(),
),
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
});
let output_buffer = device.create_buffer_f32(total_size)?;
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
total_size: u32,
num_dims: u32,
_pad1: u32,
_pad2: u32,
}
let params = Params {
total_size: total_size as u32,
num_dims: num_dims as u32,
_pad1: 0,
_pad2: 0,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Tile Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
let shader_module = device.compile_shader(Self::wgsl_shader(), Some("Tile Shader"));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("Tile 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: true },
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("Tile 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: input_shape_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: output_shape_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 4,
resource: input_strides_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 5,
resource: output_strides_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("Tile Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let compute_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("Tile 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("Tile Encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("Tile Pass"),
timestamp_writes: None,
});
compute_pass.set_pipeline(&compute_pipeline);
compute_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 = (total_size as u32).div_ceil(optimal_wg_size);
compute_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_tile_basic() {
let device = crate::device::test_pool::get_test_device().await;
let input = Tensor::from_data(&[1.0, 2.0, 3.0], vec![3], device).unwrap();
let tiled = Tile::new(input, vec![2]).unwrap().execute().unwrap();
assert_eq!(tiled.shape(), &vec![6]);
}
#[tokio::test]
async fn test_tile_2d() {
let device = crate::device::test_pool::get_test_device().await;
let data: Vec<f32> = (0..6).map(|i| i as f32).collect();
let input = Tensor::from_data(&data, vec![2, 3], device).unwrap();
let tiled = Tile::new(input, vec![2, 1]).unwrap().execute().unwrap();
assert_eq!(tiled.shape(), &vec![4, 3]);
}
#[tokio::test]
async fn test_tile_invalid_length() {
let device = crate::device::test_pool::get_test_device().await;
let input = Tensor::from_data(&[1.0, 2.0], vec![2], device).unwrap();
assert!(Tile::new(input, vec![2, 3]).is_err());
}
#[tokio::test]
async fn test_tile_zero_repeat() {
let device = crate::device::test_pool::get_test_device().await;
let input = Tensor::from_data(&[1.0, 2.0], vec![2], device).unwrap();
assert!(Tile::new(input, vec![0]).is_err());
}
#[tokio::test]
async fn test_tile_multiple_dims() {
let device = crate::device::test_pool::get_test_device().await;
let data: Vec<f32> = (0..12).map(|i| i as f32).collect();
let input = Tensor::from_data(&data, vec![2, 3, 2], device).unwrap();
let tiled = Tile::new(input, vec![2, 2, 2]).unwrap().execute().unwrap();
assert_eq!(tiled.shape(), &vec![4, 6, 4]);
}
}