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
//! `PReLU` - Parametric Rectified Linear Unit - Pure WGSL
//!
//! Deep Debt Principles:
//! - Self-knowledge: Operation knows its alpha parameter
//! - Zero hardcoding: All parameters passed at runtime
//! - Modern idiomatic Rust: Safe, zero unsafe code
//! - Complete implementation: Production-ready, no mocks
//! - Hardware-agnostic: Pure WGSL for universal compute
//! - ✅ Capability-based dispatch (vendor-optimized workgroups)

/// f64 is the canonical source.
const SHADER_F64: &str = include_str!("../shaders/activation/prelu_f64.wgsl");

static SHADER_F32: std::sync::LazyLock<String> =
    std::sync::LazyLock::new(|| SHADER_F64.to_string());

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

/// `PReLU` operation - Parametric Rectified Linear Unit
pub struct PReLU {
    input: Tensor,
    alpha: f32,
}

impl PReLU {
    /// Create a new `PReLU` operation
    #[must_use]
    pub fn new(input: Tensor, alpha: f32) -> Self {
        Self { input, alpha }
    }

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

    /// Execute the `PReLU` 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.input.device();
        let size: usize = self.input.shape().iter().product();

        // Create buffers
        // Access input buffer directly (zero-copy)
        let input_buffer = self.input.buffer();

        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,
            alpha: f32,
        }

        let params = Params {
            size: size as u32,
            alpha: self.alpha,
        };

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

        // Create bind group layout
        let bind_group_layout =
            device
                .device
                .create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
                    label: Some("PReLU 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: false },
                                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::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("PReLU Bind Group"),
            layout: &bind_group_layout,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: input_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: output_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 2,
                    resource: params_buffer.as_entire_binding(),
                },
            ],
        });

        // Create compute pipeline
        let shader_module = device.compile_shader(Self::wgsl_shader(), Some("Shader"));

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

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

        // Execute compute shader
        let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
            label: Some("PReLU Encoder"),
        });

        {
            let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("PReLU Pass"),
                timestamp_writes: None,
            });
            compute_pass.set_pipeline(&compute_pipeline);
            compute_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);
            compute_pass.dispatch_workgroups(workgroups, 1, 1);
        }

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

        // Read back results
        let output_data = crate::utils::read_buffer(device, &output_buffer, size)?;

        Ok(Tensor::new(
            output_data,
            self.input.shape().to_vec(),
            device.clone(),
        ))
    }
}

impl Tensor {
    /// Apply Parametric `ReLU` activation
    ///
    /// # Arguments
    ///
    /// * `alpha` - Slope for negative values (typically 0.01 to 0.3)
    ///
    /// # Errors
    ///
    /// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
    /// readback fails (e.g. device lost or out of memory).
    pub fn prelu_wgsl(self, alpha: f32) -> Result<Self> {
        PReLU::new(self, alpha).execute()
    }
}

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

    #[tokio::test]
    async fn test_prelu_positive() {
        let device = crate::device::test_pool::get_test_device().await;
        let data = vec![1.0, 2.0, 3.0];
        let input = Tensor::new(data, vec![3], device);

        let output = input.prelu_wgsl(0.1).unwrap();

        let result = output.to_vec().unwrap();
        assert_eq!(result[0], 1.0);
        assert_eq!(result[1], 2.0);
        assert_eq!(result[2], 3.0);
    }

    #[tokio::test]
    async fn test_prelu_negative() {
        let device = crate::device::test_pool::get_test_device().await;
        let data = vec![-1.0, -2.0, -3.0];
        let input = Tensor::new(data, vec![3], device);

        let output = input.prelu_wgsl(0.25).unwrap();

        let result = output.to_vec().unwrap();
        assert_eq!(result[0], -0.25);
        assert_eq!(result[1], -0.5);
        assert_eq!(result[2], -0.75);
    }

    #[tokio::test]
    async fn test_prelu_mixed() {
        let device = crate::device::test_pool::get_test_device().await;
        let data = vec![-2.0, 0.0, 2.0];
        let input = Tensor::new(data, vec![3], device);

        let output = input.prelu_wgsl(0.1).unwrap();

        let result = output.to_vec().unwrap();
        assert_eq!(result[0], -0.2);
        assert_eq!(result[1], 0.0);
        assert_eq!(result[2], 2.0);
    }
}