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
//! Cumprod - Cumulative product along a dimension - Pure WGSL
//!
//! Deep Debt Principles:
//! - Self-knowledge: Operation knows its dimension
//! - 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

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

/// Cumprod operation - Cumulative product along a dimension
pub struct Cumprod {
    input: Tensor,
    dim: usize,
}

impl Cumprod {
    /// Create a new cumprod operation
    #[must_use]
    pub fn new(input: Tensor, dim: usize) -> Self {
        Self { input, dim }
    }

    /// Get the WGSL shader source
    fn wgsl_shader() -> &'static str {
        include_str!("../shaders/reduce/cumprod.wgsl")
    }

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

        // Calculate dimension parameters
        let dim_size = shape[self.dim];
        let outer_size: usize = shape[..self.dim].iter().product();
        let inner_size: usize = shape[self.dim + 1..].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,
            dim_size: u32,
            outer_size: u32,
            inner_size: u32,
        }

        let params = Params {
            size: size as u32,
            dim_size: dim_size as u32,
            outer_size: outer_size as u32,
            inner_size: inner_size as u32,
        };

        let params_buffer = device
            .device
            .create_buffer_init(&wgpu::util::BufferInitDescriptor {
                label: Some("Cumprod 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("Cumprod 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("Cumprod 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("Cumprod Pipeline Layout"),
                    bind_group_layouts: &[&bind_group_layout],
                    immediate_size: 0,
                });

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

        {
            let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("Cumprod 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 = ((outer_size * inner_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, shape.to_vec(), device.clone()))
    }
}

impl Tensor {
    /// Compute cumulative product along a dimension
    /// # Arguments
    /// * `dim` - Dimension to accumulate along
    /// # Errors
    /// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
    /// readback fails (e.g. device lost or out of memory).
    pub fn cumprod_wgsl(self, dim: usize) -> Result<Self> {
        Cumprod::new(self, dim).execute()
    }
}

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

    #[tokio::test]
    async fn test_cumprod_1d() {
        let device = crate::device::test_pool::get_test_device().await;
        let data = vec![1.0, 2.0, 3.0, 4.0];
        let input = Tensor::new(data, vec![4], device);

        let output = input.cumprod_wgsl(0).unwrap();

        assert_eq!(output.shape(), &[4]);
        let result = output.to_vec().unwrap();
        assert_eq!(result[0], 1.0);
        assert_eq!(result[1], 2.0); // 1 * 2
        assert_eq!(result[2], 6.0); // 1 * 2 * 3
        assert_eq!(result[3], 24.0); // 1 * 2 * 3 * 4
    }

    #[tokio::test]
    async fn test_cumprod_2d_dim0() {
        let device = crate::device::test_pool::get_test_device().await;
        let data = vec![1.0, 2.0, 3.0, 4.0, 2.0, 2.0];
        let input = Tensor::new(data, vec![3, 2], device);

        let output = input.cumprod_wgsl(0).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); // 1 * 3
        assert_eq!(result[3], 8.0); // 2 * 4
        assert_eq!(result[4], 6.0); // 1 * 3 * 2
        assert_eq!(result[5], 16.0); // 2 * 4 * 2
    }
}