primal-barracuda 0.4.0

barraCuda — BARrier-free Rust Abstracted Cross-platform Unified Dimensional Algebra. Vendor-agnostic GPU compute via WGSL.
Documentation
// SPDX-License-Identifier: AGPL-3.0-or-later
//! L1 Loss - Mean Absolute Error - Pure WGSL
//!
//! **Deep Debt Principles**:
//! - ✅ Pure WGSL implementation (GPU-optimized)
//! - ✅ Safe Rust wrapper (no unsafe code)
//! - ✅ Hardware-agnostic via WebGPU
//! - ✅ Complete implementation (production-ready)
//!
//! ## Algorithm
//!
//! Computes Mean Absolute Error loss:
//! ```text
//! L1(pred, target) = mean(|pred - target|)
//!
//! Element-wise: loss[i] = |pred[i] - target[i]|
//! ```

use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::Result;
use crate::tensor::Tensor;

/// f64 is the canonical source — math is universal, precision is silicon.
const SHADER_F64: &str = include_str!("../shaders/loss/l1_loss_f64.wgsl");

/// f32 variant derived from f64 via precision downcast.
static SHADER_F32: std::sync::LazyLock<String> =
    std::sync::LazyLock::new(|| SHADER_F64.to_string());

/// Mean Absolute Error (L1) loss.
pub struct L1Loss {
    predictions: Tensor,
    targets: Tensor,
}

impl L1Loss {
    /// Create an L1 loss operation between predictions and targets.
    #[must_use]
    pub fn new(predictions: Tensor, targets: Tensor) -> Self {
        Self {
            predictions,
            targets,
        }
    }

    fn wgsl_shader() -> &'static str {
        &SHADER_F32
    }

    /// Execute L1 loss computation on GPU.
    ///
    /// # Errors
    ///
    /// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
    /// readback fails (e.g. device lost or out of memory).
    pub fn execute(self) -> Result<Tensor> {
        let device = self.predictions.device();
        let size = self.predictions.len();

        if size != self.targets.len() {
            return Err(crate::error::BarracudaError::InvalidShape {
                expected: self.predictions.shape().to_vec(),
                actual: self.targets.shape().to_vec(),
            });
        }

        let output_buffer = device.create_buffer_f32(size)?;

        // Create params buffer (size and reduction mode)
        let params_data = [
            size as u32,
            0u32, // reduction: 0=none (element-wise)
        ];
        let params_buffer = device.create_uniform_buffer("Params", &params_data);

        let bind_group_layout =
            device
                .device
                .create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
                    label: Some("L1Loss BGL"),
                    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("L1Loss BG"),
            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 shader = device.compile_shader(Self::wgsl_shader(), Some("L1Loss"));
        let pipeline_layout =
            device
                .device
                .create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
                    label: Some("L1Loss PL"),
                    bind_group_layouts: &[&bind_group_layout],
                    immediate_size: 0,
                });

        let pipeline = device
            .device
            .create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
                label: Some("L1Loss 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("L1Loss Encoder"),
        });

        {
            let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("L1Loss Pass"),
                timestamp_writes: None,
            });

            pass.set_pipeline(&pipeline);
            pass.set_bind_group(0, Some(&bind_group), &[]);

            // Deep Debt Evolution: Capability-based dispatch
            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,
            self.predictions.shape().to_vec(),
            device.clone(),
        ))
    }
}

impl Tensor {
    /// Compute element-wise L1 loss (absolute error) against targets.
    ///
    /// # Errors
    ///
    /// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
    /// readback fails (e.g. device lost or out of memory).
    pub fn l1_loss_wgsl(self, targets: Self) -> Result<Self> {
        L1Loss::new(self, targets).execute()
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[tokio::test]
    async fn test_l1_loss() {
        let device = crate::device::test_pool::get_test_device().await;
        let pred_data = vec![1.0, 2.0, 3.0, 4.0];
        let target_data = vec![1.5, 2.5, 2.0, 5.0];

        let predictions = Tensor::from_vec_on(pred_data, vec![4], device.clone())
            .await
            .unwrap();
        let targets = Tensor::from_vec_on(target_data, vec![4], device)
            .await
            .unwrap();

        let result = predictions.l1_loss_wgsl(targets).unwrap();
        let output = result.to_vec().unwrap();

        // |1.0 - 1.5| = 0.5
        assert!((output[0] - 0.5).abs() < 1e-5);
        // |2.0 - 2.5| = 0.5
        assert!((output[1] - 0.5).abs() < 1e-5);
        // |3.0 - 2.0| = 1.0
        assert!((output[2] - 1.0).abs() < 1e-5);
        // |4.0 - 5.0| = 1.0
        assert!((output[3] - 1.0).abs() < 1e-5);
    }
}