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
//! Gt operation - Greater than comparison
//! Pure WGSL implementation

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

/// f64 is the canonical source — f32 derived via `downcast_f64_to_f32` when needed.
const SHADER_F64: &str = include_str!("../shaders/misc/gt_f64.wgsl");

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

/// Element-wise greater-than comparison: output = lhs > rhs.
pub struct Gt {
    lhs: Tensor,
    rhs: Tensor,
}

impl Gt {
    /// Create a greater-than comparison operation.
    #[must_use]
    pub fn new(lhs: Tensor, rhs: Tensor) -> Self {
        Self { lhs, rhs }
    }
    fn wgsl_shader() -> &'static str {
        &SHADER_F32
    }

    /// Execute greater-than comparison on GPU.
    ///
    /// # 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.lhs.device();
        let size = self.lhs.len();
        let output_buffer = device.create_buffer_f32(size)?;

        let bind_group_layout =
            device
                .device
                .create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
                    label: Some("Gt BGL"),
                    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,
                        },
                    ],
                });

        let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
            label: Some("Gt BG"),
            layout: &bind_group_layout,
            entries: &[
                wgpu::BindGroupEntry {
                    binding: 0,
                    resource: self.lhs.buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 1,
                    resource: self.rhs.buffer().as_entire_binding(),
                },
                wgpu::BindGroupEntry {
                    binding: 2,
                    resource: output_buffer.as_entire_binding(),
                },
            ],
        });

        let shader = device.compile_shader(Self::wgsl_shader(), Some("Gt"));
        let pipeline_layout =
            device
                .device
                .create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
                    label: Some("Gt PL"),
                    bind_group_layouts: &[&bind_group_layout],
                    immediate_size: 0,
                });

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

        let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
            label: Some("Gt Encoder"),
        });
        {
            let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
                label: Some("Gt 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);
            pass.dispatch_workgroups(workgroups, 1, 1);
        }

        device.submit_commands(Some(encoder.finish()));
        Ok(Tensor::from_buffer(
            output_buffer,
            self.lhs.shape().to_vec(),
            device.clone(),
        ))
    }
}

impl Tensor {
    /// Element-wise greater-than comparison. Returns 1.0 where true, 0.0 where false.
    ///
    /// # Errors
    ///
    /// Returns [`Err`] if buffer allocation, GPU dispatch, or buffer
    /// readback fails (e.g. device lost or out of memory).
    pub fn gt(self, other: &Self) -> Result<Self> {
        Gt::new(self, other.clone()).execute()
    }
}

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

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

        let a = Tensor::from_vec_on(vec![1.0, 3.0, 2.0], vec![3], device.clone())
            .await
            .unwrap();
        let b = Tensor::from_vec_on(vec![2.0, 2.0, 2.0], vec![3], device)
            .await
            .unwrap();
        let result = a.gt(&b).unwrap().to_vec().unwrap();
        assert!((result[0] - 0.0).abs() < 1e-5); // 1 > 2? no
        assert!((result[1] - 1.0).abs() < 1e-5); // 3 > 2? yes
        assert!((result[2] - 0.0).abs() < 1e-5); // 2 > 2? no
    }
}