use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
use bytemuck::{Pod, Zeroable};
#[repr(C)]
#[derive(Copy, Clone, Debug, Pod, Zeroable)]
struct CircularPad2dParams {
batch_size: u32,
channels: u32,
in_height: u32,
in_width: u32,
out_height: u32,
out_width: u32,
pad_top: u32,
pad_left: u32,
}
pub struct CircularPad2d {
input: Tensor,
pad_top: u32,
pad_bottom: u32,
pad_left: u32,
pad_right: u32,
}
impl CircularPad2d {
pub fn new(
input: Tensor,
pad_top: u32,
pad_bottom: u32,
pad_left: u32,
pad_right: u32,
) -> Result<Self> {
let shape = input.shape();
if shape.len() != 4 {
return Err(BarracudaError::invalid_op(
"circular_pad2d",
format!("input must be 4D [B, C, H, W], got shape {shape:?}"),
));
}
Ok(Self {
input,
pad_top,
pad_bottom,
pad_left,
pad_right,
})
}
fn wgsl_shader() -> &'static str {
{
static S: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/tensor/circular_pad2d_f64.wgsl").to_string()
});
&S
}
}
pub fn execute(self) -> Result<Tensor> {
let device = self.input.device();
let shape = self.input.shape();
let batch_size = shape[0];
let channels = shape[1];
let in_height = shape[2];
let in_width = shape[3];
let out_height = in_height + self.pad_top as usize + self.pad_bottom as usize;
let out_width = in_width + self.pad_left as usize + self.pad_right as usize;
let output_size = batch_size * channels * out_height * out_width;
let output_buffer = device.create_buffer_f32(output_size)?;
let params = CircularPad2dParams {
batch_size: batch_size as u32,
channels: channels as u32,
in_height: in_height as u32,
in_width: in_width as u32,
out_height: out_height as u32,
out_width: out_width as u32,
pad_top: self.pad_top,
pad_left: self.pad_left,
};
let params_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
label: Some("CircularPad2d Params"),
size: std::mem::size_of::<CircularPad2dParams>() as u64,
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
mapped_at_creation: false,
});
device
.queue
.write_buffer(¶ms_buffer, 0, bytemuck::bytes_of(¶ms));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("CircularPad2d Bind Group Layout"),
entries: &[
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: false },
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::Uniform,
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
],
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("CircularPad2d Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.input.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: params_buffer.as_entire_binding(),
},
],
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("CircularPad2d"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("CircularPad2d Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("CircularPad2d Pipeline"),
layout: Some(&pipeline_layout),
module: &shader,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("CircularPad2d Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("CircularPad2d Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&pipeline);
pass.set_bind_group(0, Some(&bind_group), &[]);
let workgroups_x = (out_width as u32).div_ceil(8);
let workgroups_y = (out_height as u32).div_ceil(8);
let workgroups_z = (batch_size * channels) as u32;
pass.dispatch_workgroups(workgroups_x, workgroups_y, workgroups_z);
}
device.submit_commands(Some(encoder.finish()));
Ok(Tensor::from_buffer(
output_buffer,
vec![batch_size, channels, out_height, out_width],
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_circular_pad2d_basic() {
let device = crate::device::test_pool::get_test_device().await;
let input = Tensor::from_vec_on(vec![1.0; 2 * 3 * 4 * 4], vec![2, 3, 4, 4], device)
.await
.unwrap();
let output = CircularPad2d::new(input, 1, 1, 1, 1)
.unwrap()
.execute()
.unwrap();
let result = output.to_vec().unwrap();
assert_eq!(result.len(), 2 * 3 * 6 * 6);
}
}