use crate::device::DeviceCapabilities;
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
#[repr(C)]
#[derive(Copy, Clone, Debug, bytemuck::Pod, bytemuck::Zeroable)]
struct SeparableConv2DParams {
batch_size: u32,
in_channels: u32,
out_channels: u32,
in_height: u32,
in_width: u32,
out_height: u32,
out_width: u32,
kernel_size: u32,
stride: u32,
padding: u32,
mode: u32, }
pub struct SeparableConv2D {
input: Tensor,
weight: Tensor,
bias: Tensor,
kernel_size: usize,
stride: usize,
padding: usize,
mode: SeparableConvMode,
}
#[derive(Clone, Copy)]
pub enum SeparableConvMode {
Depthwise,
Pointwise,
}
impl SeparableConv2D {
pub fn new(
input: Tensor,
weight: Tensor,
bias: Tensor,
kernel_size: usize,
stride: usize,
padding: usize,
mode: SeparableConvMode,
) -> Result<Self> {
let input_shape = input.shape();
if input_shape.len() != 4 {
return Err(BarracudaError::invalid_op(
"separable_conv2d",
"input must be 4D tensor [B, C, H, W]",
));
}
if kernel_size == 0 || stride == 0 {
return Err(BarracudaError::invalid_op(
"separable_conv2d",
"kernel_size and stride must be positive",
));
}
Ok(Self {
input,
weight,
bias,
kernel_size,
stride,
padding,
mode,
})
}
fn wgsl_shader() -> &'static str {
{
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/conv/separable_conv2d_f64.wgsl").to_string()
});
SHADER.as_str()
}
}
pub fn execute(self) -> Result<Tensor> {
let device = self.input.device();
let input_shape = self.input.shape();
let batch_size = input_shape[0];
let in_channels = input_shape[1];
let in_height = input_shape[2];
let in_width = input_shape[3];
let out_height = ((in_height + 2 * self.padding - self.kernel_size) / self.stride) + 1;
let out_width = ((in_width + 2 * self.padding - self.kernel_size) / self.stride) + 1;
let (out_channels, output_size) = match self.mode {
SeparableConvMode::Depthwise => (
in_channels,
batch_size * in_channels * out_height * out_width,
),
SeparableConvMode::Pointwise => {
let out_ch = self.weight.shape()[0];
(out_ch, batch_size * out_ch * out_height * out_width)
}
};
let output_buffer = device.create_buffer_f32(output_size)?;
let params = SeparableConv2DParams {
batch_size: batch_size as u32,
in_channels: in_channels as u32,
out_channels: out_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,
kernel_size: self.kernel_size as u32,
stride: self.stride as u32,
padding: self.padding as u32,
mode: match self.mode {
SeparableConvMode::Depthwise => 0,
SeparableConvMode::Pointwise => 1,
},
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("separable_conv2d_params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM,
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("separable_conv2d_shader"));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("separable_conv2d_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: 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: false },
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::Uniform,
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
],
});
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("separable_conv2d_pipeline_layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("separable_conv2d_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("separable_conv2d_bind_group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.input.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.weight.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: self.bias.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 4,
resource: params_buffer.as_entire_binding(),
},
],
});
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("separable_conv2d_encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("separable_conv2d_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 (workgroups_x, workgroups_y) =
caps.dispatch_2d(out_width as u32, out_height as u32);
let workgroups_z = batch_size * out_channels;
compute_pass.dispatch_workgroups(workgroups_x, workgroups_y, workgroups_z as u32);
}
device.submit_commands(Some(encoder.finish()));
let output_data = crate::utils::read_buffer(device, &output_buffer, output_size)?;
Ok(Tensor::new(
output_data,
vec![batch_size, out_channels, out_height, out_width],
device.clone(),
))
}
}
impl Tensor {
pub fn separable_conv2d(
self,
weight: Self,
bias: Self,
kernel_size: usize,
stride: usize,
padding: usize,
mode: SeparableConvMode,
) -> Result<Self> {
SeparableConv2D::new(self, weight, bias, kernel_size, stride, padding, mode)?.execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_separable_conv2d_basic() {
let device = crate::device::test_pool::get_test_device().await;
let input = Tensor::from_vec_on(vec![1.0; 3 * 4 * 4], vec![1, 3, 4, 4], device.clone())
.await
.unwrap();
let weight = Tensor::from_vec_on(vec![0.1; 3 * 3 * 3], vec![3, 1, 3, 3], device.clone())
.await
.unwrap();
let bias = Tensor::from_vec_on(vec![0.0; 3], vec![3], device.clone())
.await
.unwrap();
let output = input
.separable_conv2d(weight, bias, 3, 1, 1, SeparableConvMode::Depthwise)
.unwrap();
let result = output.to_vec().unwrap();
assert_eq!(output.shape()[0], 1);
assert_eq!(output.shape()[1], 3);
assert!(result.iter().all(|&x| x.is_finite()));
}
}