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
//! `MarginRankingLoss` - Pure WGSL
//!
//! Deep Debt Principles:
//! - Self-knowledge: Operation knows its computation
//! - Zero hardcoding: Hardware-agnostic implementation
//! - Modern idiomatic Rust: Safe, zero unsafe code
//! - Complete implementation: Production-ready, no mocks
//! - Hardware-agnostic: Pure WGSL for universal compute

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

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

/// Margin Ranking Loss operation
pub struct MarginRankingLoss {
    input1: Tensor,
    input2: Tensor,
    target: Tensor,
    margin: f32,
}

impl MarginRankingLoss {
    /// Create a new margin ranking loss operation
    ///
    /// # Errors
    ///
    /// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
    /// readback fails (e.g. device lost or out of memory).
    pub fn new(input1: Tensor, input2: Tensor, target: Tensor, margin: f32) -> Result<Self> {
        if input1.shape() != input2.shape() {
            return Err(BarracudaError::invalid_op(
                "margin_ranking_loss",
                "input1 and input2 shapes must match",
            ));
        }

        let batch_size = input1.shape()[0];
        if target.shape()[0] != batch_size {
            return Err(BarracudaError::invalid_op(
                "margin_ranking_loss",
                "target batch size mismatch",
            ));
        }

        Ok(Self {
            input1,
            input2,
            target,
            margin,
        })
    }

    /// Get the WGSL shader source
    fn wgsl_shader() -> &'static str {
        &SHADER_F32
    }

    /// Execute the margin ranking loss operation
    ///
    /// # 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.input1.device();

        let size = self.input1.shape().iter().product::<usize>();
        let output_buffer = device.create_buffer_f32(size)?;

        // Create uniform buffer for parameters
        #[repr(C)]
        #[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
        struct Params {
            size: u32,
            margin: f32,
            _padding: [u32; 2],
        }

        let params = Params {
            size: size as u32,
            margin: self.margin,
            _padding: [0, 0],
        };

        let params_buffer = device
            .device
            .create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some("MarginRankingLoss Params"),
                contents: bytemuck::cast_slice(&[params]),
                usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
            });

        // Compile shader
        let shader_module =
            device.compile_shader(Self::wgsl_shader(), Some("MarginRankingLoss Shader"));

        // Create bind group layout
        let bind_group_layout =
            device
                .device
                .create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
                    label: Some("MarginRankingLoss 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,
                        },
                    ],
                });

        // Create bind group
        let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("MarginRankingLoss 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.target.buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 3,
                    resource: output_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 4,
                    resource: params_buffer.as_entire_binding(),
                },
            ],
        });

        // Create pipeline
        let pipeline_layout =
            device
                .device
                .create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
                    label: Some("MarginRankingLoss Pipeline Layout"),
                    bind_group_layouts: &[&bind_group_layout],
                    immediate_size: 0,
                });

        let pipeline = device
            .device
            .create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
                label: Some("MarginRankingLoss Pipeline"),
                layout: Some(&pipeline_layout),
                module: &shader_module,
                entry_point: Some("main"),
                cache: None,
                compilation_options: Default::default(),
            });

        // Encode and execute
        let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
            label: Some("MarginRankingLoss Encoder"),
        });

        {
            let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("MarginRankingLoss 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.as_ref());
            let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::Reduction);
            let workgroups = (size as u32).div_ceil(optimal_wg_size);
            pass.dispatch_workgroups(workgroups, 1, 1);
        }

        device.submit_commands(Some(encoder.finish()));

        // Create output tensor
        Ok(Tensor::from_buffer(
            output_buffer,
            self.input1.shape().to_vec(),
            device.clone(),
        ))
    }
}

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

    #[tokio::test]
    async fn test_margin_ranking_loss_basic() {
        let device = crate::device::test_pool::get_test_device().await;
        let size = 10;

        let input1 = Tensor::from_vec_on(vec![2.0; size], vec![size], device.clone())
            .await
            .unwrap();

        let input2 = Tensor::from_vec_on(vec![1.0; size], vec![size], device.clone())
            .await
            .unwrap();

        let target = Tensor::from_vec_on(vec![1.0; size], vec![size], device.clone())
            .await
            .unwrap();

        let output = MarginRankingLoss::new(input1, input2, target, 0.5)
            .unwrap()
            .execute()
            .unwrap();

        assert_eq!(output.shape(), &[size]);
    }
}