use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
#[repr(C)]
#[derive(Copy, Clone, Debug, bytemuck::Pod, bytemuck::Zeroable)]
struct AffineGridParams {
batch_size: u32,
height: u32,
width: u32,
align_corners: u32,
}
pub struct AffineGrid {
theta: Tensor,
size: (usize, usize),
align_corners: bool,
}
impl AffineGrid {
pub fn new(theta: Tensor, size: (usize, usize), align_corners: bool) -> Result<Self> {
let shape = theta.shape();
if shape.len() != 3 || shape[1] != 2 || shape[2] != 3 {
return Err(BarracudaError::invalid_op(
"affine_grid",
"theta must be 3D tensor [B, 2, 3]",
));
}
let _ = shape[0];
let height = size.0;
let width = size.1;
if height == 0 || width == 0 {
return Err(BarracudaError::invalid_op(
"affine_grid",
"height and width must be positive",
));
}
Ok(Self {
theta,
size,
align_corners,
})
}
fn wgsl_shader() -> &'static str {
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/misc/affine_grid_f64.wgsl").to_string()
});
&SHADER
}
pub fn execute(self) -> Result<Tensor> {
let device = self.theta.device();
let shape = self.theta.shape();
let batch_size = shape[0];
let height = self.size.0;
let width = self.size.1;
let output_size = batch_size * height * width * 2;
let output_buffer = device.create_buffer_f32(output_size)?;
let params = AffineGridParams {
batch_size: batch_size as u32,
height: height as u32,
width: width as u32,
align_corners: u32::from(self.align_corners),
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("affine_grid_params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM,
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("affine_grid_shader"));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("affine_grid_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 pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("affine_grid_pipeline_layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("affine_grid_pipeline"),
layout: Some(&pipeline_layout),
module: &shader,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("affine_grid_bind_group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.theta.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 mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("affine_grid_encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("affine_grid_pass"),
timestamp_writes: None,
});
compute_pass.set_pipeline(&pipeline);
compute_pass.set_bind_group(0, Some(&bind_group), &[]);
let caps = DeviceCapabilities::from_device(device);
let (wg_x, wg_y, wg_z) = caps.optimal_workgroup_size_3d(WorkloadType::MatMul);
let workgroups_x = (width as u32).div_ceil(wg_x);
let workgroups_y = (height as u32).div_ceil(wg_y);
let workgroups_z = (batch_size as u32).div_ceil(wg_z);
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, height, width, 2],
device.clone(),
))
}
}
impl Tensor {
pub fn affine_grid(self, size: (usize, usize), align_corners: bool) -> Result<Self> {
AffineGrid::new(self, size, align_corners)?.execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_affine_grid_basic() {
let device = crate::device::test_pool::get_test_device().await;
let theta_data = vec![1.0, 0.0, 0.0, 0.0, 1.0, 0.0];
let theta = Tensor::from_vec_on(theta_data, vec![1, 2, 3], device.clone())
.await
.unwrap();
let output = theta.affine_grid((4, 4), false).unwrap();
let result = output.to_vec().unwrap();
assert_eq!(output.shape(), &[1, 4, 4, 2]);
assert_eq!(result.len(), 32);
assert!(result.iter().all(|&x| x.is_finite()));
}
}