use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
pub struct GroupedConv2D {
input: Tensor,
kernel: Tensor,
bias: Option<Tensor>,
stride: usize,
padding: usize,
groups: usize,
}
impl GroupedConv2D {
pub fn new(
input: Tensor,
kernel: Tensor,
stride: usize,
padding: usize,
groups: usize,
bias: Option<Tensor>,
) -> Result<Self> {
let input_shape = input.shape();
if input_shape.len() != 4 {
return Err(BarracudaError::invalid_op(
"grouped_conv2d",
"input must be 4D [batch, channels, height, width]",
));
}
let kernel_shape = kernel.shape();
if kernel_shape.len() != 4 {
return Err(BarracudaError::invalid_op(
"grouped_conv2d",
"kernel must be 4D [out_channels, in_channels_per_group, kernel_h, kernel_w]",
));
}
let _ = input_shape[0];
let in_channels = input_shape[1];
let in_height = input_shape[2];
let in_width = input_shape[3];
let out_channels = kernel_shape[0];
let in_per_group = kernel_shape[1];
let kernel_size = kernel_shape[2];
if !in_channels.is_multiple_of(groups) {
return Err(BarracudaError::invalid_op(
"grouped_conv2d",
"in_channels must be divisible by groups",
));
}
if !out_channels.is_multiple_of(groups) {
return Err(BarracudaError::invalid_op(
"grouped_conv2d",
"out_channels must be divisible by groups",
));
}
if in_channels / groups != in_per_group {
return Err(BarracudaError::invalid_op(
"grouped_conv2d",
"kernel in_channels_per_group must match in_channels / groups",
));
}
if let Some(ref bias_tensor) = bias {
let bias_size = bias_tensor.shape().iter().product::<usize>();
if bias_size != out_channels {
return Err(BarracudaError::invalid_op(
"grouped_conv2d",
"bias must have out_channels elements",
));
}
}
let _ = (
(in_height + 2 * padding - kernel_size) / stride + 1,
(in_width + 2 * padding - kernel_size) / stride + 1,
);
Ok(Self {
input,
kernel,
bias,
stride,
padding,
groups,
})
}
fn wgsl_shader() -> &'static str {
{
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/conv/grouped_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 kernel_shape = self.kernel.shape();
let out_channels = kernel_shape[0];
let in_per_group = kernel_shape[1];
let kernel_size = kernel_shape[2];
let out_height = (in_height + 2 * self.padding - kernel_size) / self.stride + 1;
let out_width = (in_width + 2 * self.padding - kernel_size) / self.stride + 1;
let out_per_group = out_channels / self.groups;
let output_size = batch_size * out_channels * out_height * out_width;
let output_buffer = device.create_buffer_f32(output_size)?;
let zero_bias;
let bias_buffer: &wgpu::Buffer = if let Some(ref bias_tensor) = self.bias {
bias_tensor.buffer()
} else {
let zeros = vec![0.0f32; out_channels];
zero_bias = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("GroupedConv2D Bias Zeros"),
contents: bytemuck::cast_slice(&zeros),
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
});
&zero_bias
};
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
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,
groups: u32,
in_per_group: u32,
out_per_group: u32,
_pad1: u32,
_pad2: u32,
_pad3: u32,
}
let params = Params {
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: kernel_size as u32,
stride: self.stride as u32,
padding: self.padding as u32,
groups: self.groups as u32,
in_per_group: in_per_group as u32,
out_per_group: out_per_group as u32,
_pad1: 0,
_pad2: 0,
_pad3: 0,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("GroupedConv2D Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
let shader_module =
device.compile_shader(Self::wgsl_shader(), Some("GroupedConv2D Shader"));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("GroupedConv2D 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: 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: true },
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::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("GroupedConv2D Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: params_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.input.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: self.kernel.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: bias_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 4,
resource: output_buffer.as_entire_binding(),
},
],
});
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("GroupedConv2D Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("GroupedConv2D 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("GroupedConv2D Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("GroupedConv2D Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&pipeline);
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 spatial_size = (out_height * out_width) as u32;
let workgroups_x = spatial_size.div_ceil(optimal_wg_size);
let workgroups_y = (out_channels as u32).div_ceil(optimal_wg_size);
let workgroups_z = batch_size 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, out_channels, out_height, out_width],
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_grouped_conv2d_basic() {
let device = crate::device::test_pool::get_test_device().await;
let batch_size = 1;
let in_channels = 8;
let in_height = 32;
let in_width = 32;
let out_channels = 16;
let kernel_size = 3;
let groups = 2;
let input = Tensor::from_vec_on(
vec![1.0; batch_size * in_channels * in_height * in_width],
vec![batch_size, in_channels, in_height, in_width],
device.clone(),
)
.await
.unwrap();
let in_per_group = in_channels / groups;
let kernel = Tensor::from_vec_on(
vec![0.1; out_channels * in_per_group * kernel_size * kernel_size],
vec![out_channels, in_per_group, kernel_size, kernel_size],
device.clone(),
)
.await
.unwrap();
let bias = Tensor::from_vec_on(vec![0.0; out_channels], vec![out_channels], device.clone())
.await
.unwrap();
let conv = GroupedConv2D::new(input, kernel, 1, 1, groups, Some(bias)).unwrap();
let output = conv.execute().unwrap();
assert_eq!(output.shape().len(), 4);
}
#[tokio::test]
async fn test_grouped_conv2d_no_bias() {
let device = crate::device::test_pool::get_test_device().await;
let batch_size = 1;
let in_channels = 4;
let in_height = 16;
let in_width = 16;
let out_channels = 8;
let kernel_size = 3;
let groups = 2;
let input = Tensor::from_vec_on(
vec![1.0; batch_size * in_channels * in_height * in_width],
vec![batch_size, in_channels, in_height, in_width],
device.clone(),
)
.await
.unwrap();
let in_per_group = in_channels / groups;
let kernel = Tensor::from_vec_on(
vec![0.1; out_channels * in_per_group * kernel_size * kernel_size],
vec![out_channels, in_per_group, kernel_size, kernel_size],
device.clone(),
)
.await
.unwrap();
let conv = GroupedConv2D::new(input, kernel, 1, 1, groups, None).unwrap();
let output = conv.execute().unwrap();
assert_eq!(output.shape().len(), 4);
}
}