use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
use bytemuck::{Pod, Zeroable};
#[repr(C)]
#[derive(Copy, Clone, Debug, Pod, Zeroable)]
struct CosineEmbeddingLossParams {
size: u32,
margin: f32,
_padding: [u32; 2],
}
pub struct CosineEmbeddingLoss {
input1: Tensor,
input2: Tensor,
label: Tensor,
margin: f32,
}
impl CosineEmbeddingLoss {
pub fn new(input1: Tensor, input2: Tensor, label: Tensor, margin: f32) -> Result<Self> {
if input1.shape() != input2.shape() {
return Err(BarracudaError::invalid_op(
"cosine_embedding_loss",
format!(
"input1 shape {:?} must match input2 shape {:?}",
input1.shape(),
input2.shape()
),
));
}
if label.shape() != [1] {
return Err(BarracudaError::invalid_op(
"cosine_embedding_loss",
format!("label must be scalar [1], got shape {:?}", label.shape()),
));
}
Ok(Self {
input1,
input2,
label,
margin,
})
}
fn wgsl_shader() -> &'static str {
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/loss/cosine_embedding_loss_f64.wgsl").to_string()
});
&SHADER
}
pub fn execute(self) -> Result<Tensor> {
let device = self.input1.device();
let size = self.input1.len();
let output_buffer = device.create_buffer_f32(1)?;
let params = CosineEmbeddingLossParams {
size: size as u32,
margin: self.margin,
_padding: [0; 2],
};
let params_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
label: Some("CosineEmbeddingLoss Params"),
size: std::mem::size_of::<CosineEmbeddingLossParams>() 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 bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("CosineEmbeddingLoss 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 bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("CosineEmbeddingLoss Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.input1.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.input2.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: self.label.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 shader = device.compile_shader(Self::wgsl_shader(), Some("CosineEmbeddingLoss"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("CosineEmbeddingLoss Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("CosineEmbeddingLoss Pipeline"),
layout: Some(&pipeline_layout),
module: &shader,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("CosineEmbeddingLoss Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("CosineEmbeddingLoss 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::ElementWise);
let workgroups = (size as u32).div_ceil(optimal_wg_size);
pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
Ok(Tensor::from_buffer(output_buffer, vec![1], device.clone()))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_cosine_embedding_loss_basic() {
let device = crate::device::test_pool::get_test_device().await;
let input1 = Tensor::from_vec_on(vec![1.0, 2.0, 3.0], vec![3], device.clone())
.await
.unwrap();
let input2 = Tensor::from_vec_on(vec![1.0, 2.0, 3.0], vec![3], device.clone())
.await
.unwrap();
let label = Tensor::from_vec_on(vec![1.0], vec![1], device)
.await
.unwrap();
let output = CosineEmbeddingLoss::new(input1, input2, label, 0.5)
.unwrap()
.execute()
.unwrap();
let result = output.to_vec().unwrap();
assert_eq!(result.len(), 1);
assert!(result[0] >= 0.0);
}
}