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
//! GPU compute operations for `NAdam` Optimizer
//!
//! This module contains the single-pass GPU execution for `NAdam` optimizer
//! with Nesterov momentum.

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

impl Nadam {
    /// Execute `NAdam` optimizer step (GPU single-pass)
    ///
    /// **Deep Debt**: Efficient single-pass update with Nesterov momentum
    ///
    /// Returns: (`new_weights`, `new_m`, `new_v`)
    ///
    /// # 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, Tensor, Tensor)> {
        let device = self.weights().device();
        let size = self.weights().len();

        // Create parameters
        let params = NadamParams {
            learning_rate: self.learning_rate(),
            beta1: self.beta1(),
            beta2: self.beta2(),
            epsilon: self.epsilon(),
            weight_decay: self.weight_decay(),
            step: self.step(),
            _padding: [0, 0],
        };

        let params_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
            label: Some("NAdam Params"),
            size: std::mem::size_of::<NadamParams>() as u64,
            usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
            mapped_at_creation: false,
        });
        device
            .queue
            .write_buffer(&params_buffer, 0, bytemuck::bytes_of(&params));

        // Output buffers
        let weights_out_buffer = device.create_buffer_f32(size)?;
        let m_out_buffer = device.create_buffer_f32(size)?;
        let v_out_buffer = device.create_buffer_f32(size)?;

        // Compile shader
        let shader = device.compile_shader(Self::shader(), Some("NAdam"));

        // Create bind group layout (8 bindings)
        let bgl = device
            .device
            .create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
                label: Some("NAdam BGL"),
                entries: &[
                    // weights (read)
                    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,
                    },
                    // gradients (read)
                    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,
                    },
                    // m_in (read)
                    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,
                    },
                    // v_in (read)
                    wgpu::BindGroupLayoutEntry {
                        binding: 3,
                        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,
                    },
                    // weights_out (write)
                    wgpu::BindGroupLayoutEntry {
                        binding: 4,
                        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,
                    },
                    // m_out (write)
                    wgpu::BindGroupLayoutEntry {
                        binding: 5,
                        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,
                    },
                    // v_out (write)
                    wgpu::BindGroupLayoutEntry {
                        binding: 6,
                        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,
                    },
                    // params (uniform)
                    wgpu::BindGroupLayoutEntry {
                        binding: 7,
                        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("NAdam BG"),
            layout: &bgl,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: self.weights().buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: self.gradients().buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 2,
                    resource: self.m().buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 3,
                    resource: self.v().buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 4,
                    resource: weights_out_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 5,
                    resource: m_out_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 6,
                    resource: v_out_buffer.as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 7,
                    resource: params_buffer.as_entire_binding(),
                },
            ],
        });

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

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

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

        {
            let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("NAdam 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).max(1);
            pass.dispatch_workgroups(workgroups, 1, 1);
        }

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

        // Return all three outputs
        Ok((
            Tensor::from_buffer(
                weights_out_buffer,
                self.weights().shape().to_vec(),
                device.clone(),
            ),
            Tensor::from_buffer(m_out_buffer, self.m().shape().to_vec(), device.clone()),
            Tensor::from_buffer(v_out_buffer, self.v().shape().to_vec(), device.clone()),
        ))
    }
}