use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::Result;
use crate::tensor::Tensor;
pub struct Fold {
input: Tensor,
output_size: (usize, usize),
kernel_size: (usize, usize),
stride: usize,
padding: usize,
dilation: usize,
}
impl Fold {
pub fn new(
input: Tensor,
output_size: (usize, usize),
kernel_size: (usize, usize),
stride: usize,
padding: usize,
dilation: usize,
) -> Result<Self> {
let shape = input.shape();
if shape.len() != 3 {
return Err(crate::error::BarracudaError::InvalidInput {
message: format!("Fold expects 3D tensor [B, C*K*K, L], got shape {shape:?}"),
});
}
Ok(Self {
input,
output_size,
kernel_size,
stride,
padding,
dilation,
})
}
fn wgsl_shader() -> &'static str {
{
static S: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/tensor/fold_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_times_kernel = shape[1];
let kernel_elements = self.kernel_size.0 * self.kernel_size.1;
let channels = channels_times_kernel / kernel_elements;
if !channels_times_kernel.is_multiple_of(kernel_elements) {
return Err(crate::error::BarracudaError::InvalidInput {
message: format!(
"Input channels*kernel ({channels_times_kernel}) must be divisible by kernel elements ({kernel_elements})"
),
});
}
let out_height = self.output_size.0;
let out_width = self.output_size.1;
let output_size = batch_size * channels * out_height * out_width;
let num_blocks_h =
((out_height + 2 * self.padding - self.dilation * (self.kernel_size.0 - 1) - 1)
/ self.stride)
+ 1;
let num_blocks_w =
((out_width + 2 * self.padding - self.dilation * (self.kernel_size.1 - 1) - 1)
/ self.stride)
+ 1;
let input_buffer = self.input.buffer();
let output_buffer = device.create_buffer_f32(output_size)?;
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
batch_size: u32,
channels: u32,
out_height: u32,
out_width: u32,
kernel_height: u32,
kernel_width: u32,
stride: u32,
padding: u32,
dilation: u32,
num_blocks_h: u32,
num_blocks_w: u32,
_pad1: u32,
}
let params = Params {
batch_size: batch_size as u32,
channels: channels as u32,
out_height: out_height as u32,
out_width: out_width as u32,
kernel_height: self.kernel_size.0 as u32,
kernel_width: self.kernel_size.1 as u32,
stride: self.stride as u32,
padding: self.padding as u32,
dilation: self.dilation as u32,
num_blocks_h: num_blocks_h as u32,
num_blocks_w: num_blocks_w as u32,
_pad1: 0,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Fold Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
let shader_module = device.compile_shader(Self::wgsl_shader(), Some("Fold Shader"));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("Fold 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("Fold 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("Fold Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let compute_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("Fold 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("Fold Encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("Fold 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::Convolution);
let workgroups_x = (out_width as u32).div_ceil(optimal_wg_size);
let workgroups_y = (out_height as u32).div_ceil(optimal_wg_size);
let workgroups_z = (batch_size * channels) as u32;
compute_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_fold_basic() {
let device = crate::device::test_pool::get_test_device().await;
let data: Vec<f32> = (0..324).map(|i| i as f32).collect();
let input = Tensor::from_data(
&data,
vec![1, 9, 36], device,
)
.unwrap();
let folded = Fold::new(input, (6, 6), (3, 3), 1, 0, 1)
.unwrap()
.execute()
.unwrap();
assert_eq!(folded.shape(), &vec![1, 1, 6, 6]);
}
#[tokio::test]
async fn test_fold_invalid_shape() {
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();
assert!(Fold::new(input, (4, 4), (3, 3), 1, 0, 1).is_err());
}
#[tokio::test]
async fn test_fold_with_padding() {
let device = crate::device::test_pool::get_test_device().await;
let data: Vec<f32> = (0..576).map(|i| i as f32).collect();
let input = Tensor::from_data(&data, vec![1, 9, 64], device).unwrap();
let folded = Fold::new(input, (8, 8), (3, 3), 1, 1, 1)
.unwrap()
.execute()
.unwrap();
assert_eq!(folded.shape(), &vec![1, 1, 8, 8]);
}
}