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
//! Weight Normalization - 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
//!
//! Reparameterizes weights as: w = g * (v / ||v||)
//! Decouples magnitude and direction of weight vectors.
//! Speeds up training convergence.
//!
//! Reference: "Weight Normalization" by Salimans & Kingma (2016)

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

/// Weight Normalization operation
pub struct WeightNormalization {
    v: Tensor,
    g: Tensor,
    dim: u32,
}

impl WeightNormalization {
    /// Create a new weight normalization operation
    /// # Arguments
    /// * `v` - Direction vectors (weights to normalize)
    /// * `g` - Magnitude scalars
    /// * `dim` - Dimension to normalize over (0 = all)
    /// # Errors
    /// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
    /// readback fails (e.g. device lost or out of memory).
    pub fn new(v: Tensor, g: Tensor, dim: u32) -> Result<Self> {
        let v_shape = v.shape();

        if v_shape.is_empty() {
            return Err(BarracudaError::invalid_op(
                "WeightNormalization",
                format!("v must have at least 1 dimension, got shape {v_shape:?}"),
            ));
        }

        Ok(Self { v, g, dim })
    }

    /// Get the WGSL shader source
    fn wgsl_shader() -> &'static str {
        {
            static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
                include_str!("../shaders/norm/weight_norm_f64.wgsl").to_string()
            });
            std::sync::LazyLock::force(&SHADER).as_str()
        }
    }

    /// Execute the weight normalization 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.v.device();
        let v_shape = self.v.shape();
        let num_weights = self.v.len();

        // Create output buffer: same shape as v
        let output_buffer = device.create_buffer_f32(num_weights)?;

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

        let params = Params {
            num_weights: num_weights as u32,
            dim: self.dim,
            _padding: [0; 2],
        };

        let params_buffer = device
            .device
            .create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some("WeightNormalization 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("WeightNormalization Shader"));

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

        // Create bind group
        let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("WeightNormalization Bind Group"),
            layout: &bind_group_layout,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: self.v.buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: self.g.buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 2,
                    resource: output_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 3,
                    resource: params_buffer.as_entire_binding(),
                },
            ],
        });

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

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

        {
            let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("WeightNormalization 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::Reduction);
            let workgroups = (num_weights as u32).div_ceil(optimal_wg_size);
            compute_pass.dispatch_workgroups(workgroups, 1, 1);
        }

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

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

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

    #[tokio::test]
    async fn test_weight_normalization() {
        let device = crate::device::test_pool::get_test_device().await;
        // 2 filters, 3 weights each
        let v = Tensor::from_vec_on(
            vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0],
            vec![2, 3],
            device.clone(),
        )
        .await
        .unwrap();

        let g = Tensor::from_vec_on(vec![2.0], vec![1], device.clone())
            .await
            .unwrap();

        let result = WeightNormalization::new(v, g, 0)
            .unwrap()
            .execute()
            .unwrap();

        assert_eq!(result.shape(), &[2, 3]);
        let data = result.to_vec().unwrap();
        assert!(data.iter().all(|&x| x.is_finite()));
    }
}