use crate::device::compute_pipeline::ComputeDispatch;
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
use bytemuck::{Pod, Zeroable};
const SHADER_F64: &str = include_str!("../shaders/augmentation/elastic_transform_f64.wgsl");
static SHADER_F32: std::sync::LazyLock<String> =
std::sync::LazyLock::new(|| SHADER_F64.to_string());
#[repr(C)]
#[derive(Copy, Clone, Debug, Pod, Zeroable)]
struct ElasticTransformParams {
batch_size: u32,
channels: u32,
height: u32,
width: u32,
alpha: f32,
sigma: f32,
_padding: [u32; 2],
}
pub struct ElasticTransform {
input: Tensor,
displacement_x: Tensor,
displacement_y: Tensor,
alpha: f32,
sigma: f32,
}
impl ElasticTransform {
pub fn new(
input: Tensor,
displacement_x: Tensor,
displacement_y: Tensor,
alpha: f32,
sigma: f32,
) -> Result<Self> {
let shape = input.shape();
if shape.len() != 4 {
return Err(BarracudaError::invalid_op(
"elastic_transform",
format!("input must be 4D [B, C, H, W], got shape {shape:?}"),
));
}
let height = shape[2];
let width = shape[3];
let displacement_size = height * width;
if displacement_x.shape() != [displacement_size] {
return Err(BarracudaError::invalid_op(
"elastic_transform",
format!(
"displacement_x shape {:?} must be [{:?}]",
displacement_x.shape(),
displacement_size
),
));
}
if displacement_y.shape() != [displacement_size] {
return Err(BarracudaError::invalid_op(
"elastic_transform",
format!(
"displacement_y shape {:?} must be [{:?}]",
displacement_y.shape(),
displacement_size
),
));
}
if alpha < 0.0 {
return Err(BarracudaError::invalid_op(
"elastic_transform",
format!("alpha must be non-negative, got {alpha}"),
));
}
if sigma <= 0.0 {
return Err(BarracudaError::invalid_op(
"elastic_transform",
format!("sigma must be positive, got {sigma}"),
));
}
Ok(Self {
input,
displacement_x,
displacement_y,
alpha,
sigma,
})
}
fn wgsl_shader() -> &'static str {
&SHADER_F32
}
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 height = shape[2];
let width = shape[3];
let size = self.input.len();
let output_buffer = device.create_buffer_f32(size)?;
let params = ElasticTransformParams {
batch_size: batch_size as u32,
channels: channels as u32,
height: height as u32,
width: width as u32,
alpha: self.alpha,
sigma: self.sigma,
_padding: [0; 2],
};
let params_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
label: Some("ElasticTransform Params"),
size: std::mem::size_of::<ElasticTransformParams>() 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 workgroups_x = (width as u32).div_ceil(8);
let workgroups_y = (height as u32).div_ceil(8);
let workgroups_z = (batch_size * channels) as u32;
ComputeDispatch::new(device, "elastic_transform")
.shader(Self::wgsl_shader(), "main")
.storage_read(0, self.input.buffer())
.storage_read(1, self.displacement_x.buffer())
.storage_read(2, self.displacement_y.buffer())
.storage_rw(3, &output_buffer)
.uniform(4, ¶ms_buffer)
.dispatch(workgroups_x, workgroups_y, workgroups_z)
.submit()?;
Ok(Tensor::from_buffer(
output_buffer,
shape.to_vec(),
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_elastic_transform_basic() {
let device = crate::device::test_pool::get_test_device().await;
if device.is_lost() {
return;
}
let input = Tensor::from_vec_on(vec![1.0; 2 * 3 * 4 * 4], vec![2, 3, 4, 4], device.clone())
.await
.unwrap();
let disp_size = 4 * 4;
let displacement_x =
Tensor::from_vec_on(vec![0.1; disp_size], vec![disp_size], device.clone())
.await
.unwrap();
let displacement_y = Tensor::from_vec_on(vec![0.1; disp_size], vec![disp_size], device)
.await
.unwrap();
let result = ElasticTransform::new(input, displacement_x, displacement_y, 1.0, 1.0)
.and_then(super::ElasticTransform::execute)
.and_then(|t| t.to_vec());
let Ok(result) = result else { return };
assert_eq!(result.len(), 2 * 3 * 4 * 4);
}
}