use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
pub struct FocalLossV2 {
predictions: Tensor,
targets: Tensor,
alpha: f32,
gamma: f32,
}
impl FocalLossV2 {
pub fn new(predictions: Tensor, targets: Tensor, alpha: f32, gamma: f32) -> Result<Self> {
if predictions.shape() != targets.shape() {
return Err(BarracudaError::shape_mismatch(
predictions.shape().to_vec(),
targets.shape().to_vec(),
));
}
if !(0.0..=1.0).contains(&alpha) {
return Err(BarracudaError::invalid_op(
"FocalLossV2",
format!("alpha must be in [0, 1], got {alpha}"),
));
}
if gamma < 0.0 {
return Err(BarracudaError::invalid_op(
"FocalLossV2",
format!("gamma must be non-negative, got {gamma}"),
));
}
Ok(Self {
predictions,
targets,
alpha,
gamma,
})
}
fn wgsl_shader() -> &'static str {
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/loss/focal_loss_v2_f64.wgsl").to_string()
});
&SHADER
}
pub fn execute(self) -> Result<Tensor> {
let device = self.predictions.device();
let size = self.predictions.len();
let output_buffer = device.create_buffer_f32(size)?;
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
alpha: f32,
gamma: f32,
epsilon: f32,
size: u32,
_pad1: u32,
_pad2: u32,
_pad3: u32,
}
let params = Params {
alpha: self.alpha,
gamma: self.gamma,
epsilon: 1e-7,
size: size as u32,
_pad1: 0,
_pad2: 0,
_pad3: 0,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("FocalLossV2 Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
let shader_module = device.compile_shader(Self::wgsl_shader(), Some("FocalLossV2 Shader"));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("FocalLossV2 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: false },
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::Uniform,
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
],
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("FocalLossV2 Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.predictions.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.targets.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: params_buffer.as_entire_binding(),
},
],
});
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("FocalLossV2 Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let compute_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("FocalLossV2 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("FocalLossV2 Encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("FocalLossV2 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::Reduction);
let workgroups = (size as u32).div_ceil(optimal_wg_size);
compute_pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
Ok(Tensor::from_buffer(
output_buffer,
self.predictions.shape().to_vec(),
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_focal_loss_v2_basic() {
let device = crate::device::test_pool::get_test_device().await;
let preds = Tensor::from_vec_on(vec![0.9, 0.1, 0.8], vec![3], device.clone())
.await
.unwrap();
let targets = Tensor::from_vec_on(vec![1.0, 0.0, 1.0], vec![3], device.clone())
.await
.unwrap();
let loss = FocalLossV2::new(preds, targets, 0.25, 2.0)
.unwrap()
.execute()
.unwrap();
let result = loss.to_vec().unwrap();
assert_eq!(result.len(), 3);
assert!(result.iter().all(|&x| x >= 0.0 && x.is_finite()));
}
#[tokio::test]
async fn test_focal_loss_v2_edge_cases() {
let device = crate::device::test_pool::get_test_device().await;
let preds = Tensor::from_vec_on(vec![1.0, 0.0, 1.0, 0.0], vec![4], device.clone())
.await
.unwrap();
let targets = Tensor::from_vec_on(vec![1.0, 0.0, 1.0, 0.0], vec![4], device.clone())
.await
.unwrap();
let loss = FocalLossV2::new(preds, targets, 0.25, 2.0)
.unwrap()
.execute()
.unwrap();
let result = loss.to_vec().unwrap();
assert!(result.iter().all(|&x| x < 0.1));
}
}