aprender-db 0.38.0

//! Compute backend dispatcher
//!
//! Toyota Way Principles:
//! - Genchi Genbutsu: Physics-based cost model (`PCIe` Gen4 x16 = 32 GB/s)
//! - Muda elimination: GPU only if compute > 5x transfer time

/// Cost-based backend selection
///
/// References:
/// - Gregg & Hazelwood (2011): `PCIe` bus bottleneck analysis
/// - Breß et al. (2014): Operator variant selection on heterogeneous hardware
pub struct BackendDispatcher {
    _private: (),
}

/// `PCIe` Gen4 x16 bandwidth: 32 GB/s
const PCIE_BANDWIDTH_GBPS: f64 = 32.0;

/// Minimum data size for GPU consideration (10 MB)
const MIN_GPU_DATA_SIZE_BYTES: usize = 10_000_000;

/// GPU compute throughput (conservative estimate: 100 GFLOP/s)
/// Modern GPUs can achieve 1-10 TFLOP/s, but we use conservative estimate
const GPU_THROUGHPUT_GFLOPS: f64 = 100.0;

/// Transfer overhead multiplier (5x rule from spec)
/// GPU compute must be > 5x transfer time to be worthwhile
const TRANSFER_OVERHEAD_MULTIPLIER: f64 = 5.0;

impl BackendDispatcher {
    /// Select backend based on arithmetic intensity (FLOPs/Byte)
    ///
    /// # Arguments
    /// * `total_bytes` - Total data size in bytes
    /// * `estimated_flops` - Estimated floating point operations
    ///
    /// # Returns
    /// Backend selection (GPU, SIMD, or Scalar)
    ///
    /// # Algorithm
    /// 1. Check minimum data size threshold (10 MB)
    /// 2. Calculate `PCIe` transfer time: bytes / 32 GB/s
    /// 3. Estimate GPU compute time: FLOPs / 100 GFLOP/s
    /// 4. Apply 5x rule: GPU only if compute > 5x transfer
    #[must_use]
    #[allow(clippy::cast_precision_loss)]
    pub fn select(total_bytes: usize, estimated_flops: f64) -> super::Backend {
        // Rule 1: Minimum data size threshold (10 MB)
        if total_bytes < MIN_GPU_DATA_SIZE_BYTES {
            return super::Backend::Simd;
        }

        // Rule 2: Calculate transfer time (PCIe Gen4 x16 = 32 GB/s)
        let pcie_transfer_time_ms =
            (total_bytes as f64 / (PCIE_BANDWIDTH_GBPS * 1_000_000_000.0)) * 1000.0;

        // Rule 3: Estimate GPU compute time
        let estimated_gpu_compute_ms =
            (estimated_flops / (GPU_THROUGHPUT_GFLOPS * 1_000_000_000.0)) * 1000.0;

        // Rule 4: Apply 5x rule (Toyota Way: Genchi Genbutsu - physics-based decision)
        if estimated_gpu_compute_ms > pcie_transfer_time_ms * TRANSFER_OVERHEAD_MULTIPLIER {
            super::Backend::Gpu
        } else {
            super::Backend::Simd
        }
    }

    /// Calculate arithmetic intensity (FLOPs per byte)
    ///
    /// Higher arithmetic intensity means more compute per data transfer,
    /// making GPU acceleration more beneficial.
    ///
    /// # Arguments
    /// * `total_flops` - Total floating point operations
    /// * `total_bytes` - Total data size in bytes
    ///
    /// # Returns
    /// Arithmetic intensity ratio (FLOPs/Byte)
    ///
    /// # Example
    /// ```
    /// use trueno_db::backend::BackendDispatcher;
    ///
    /// // Matrix multiply: N^3 FLOPs for N^2 elements = N FLOPs/element
    /// let intensity = BackendDispatcher::arithmetic_intensity(1_000_000_000.0, 100_000_000);
    /// assert_eq!(intensity, 10.0); // 10 FLOPs per byte
    /// ```
    #[must_use]
    #[allow(clippy::cast_precision_loss)]
    pub const fn arithmetic_intensity(total_flops: f64, total_bytes: usize) -> f64 {
        total_flops / total_bytes as f64
    }

    /// Estimate FLOPs for simple aggregation (SUM, AVG, COUNT, MIN, MAX)
    ///
    /// Simple aggregations perform ~1 FLOP per element (single pass)
    ///
    /// # Arguments
    /// * `num_elements` - Number of elements to aggregate
    ///
    /// # Returns
    /// Estimated FLOPs
    ///
    /// # Example
    /// ```
    /// use trueno_db::backend::BackendDispatcher;
    ///
    /// // SUM over 100M elements = 100M FLOPs
    /// let flops = BackendDispatcher::estimate_simple_aggregation_flops(100_000_000);
    /// assert_eq!(flops, 100_000_000.0);
    /// ```
    #[must_use]
    #[allow(clippy::cast_precision_loss)]
    pub const fn estimate_simple_aggregation_flops(num_elements: usize) -> f64 {
        num_elements as f64
    }

    /// Estimate FLOPs for GROUP BY aggregation
    ///
    /// GROUP BY requires hashing (5 FLOPs/element) + aggregation (1 FLOP/element)
    ///
    /// # Arguments
    /// * `num_elements` - Number of elements to process
    ///
    /// # Returns
    /// Estimated FLOPs
    ///
    /// # Example
    /// ```
    /// use trueno_db::backend::BackendDispatcher;
    ///
    /// // GROUP BY over 100M elements = 600M FLOPs
    /// let flops = BackendDispatcher::estimate_group_by_flops(100_000_000);
    /// assert_eq!(flops, 600_000_000.0);
    /// ```
    #[must_use]
    #[allow(clippy::cast_precision_loss)]
    pub const fn estimate_group_by_flops(num_elements: usize) -> f64 {
        // Hashing: 5 FLOPs/element, Aggregation: 1 FLOP/element
        (num_elements as f64) * 6.0
    }

    /// Estimate FLOPs for WHERE filter
    ///
    /// Filters require predicate evaluation (~2 FLOPs per element)
    ///
    /// # Arguments
    /// * `num_elements` - Number of elements to filter
    ///
    /// # Returns
    /// Estimated FLOPs
    #[must_use]
    #[allow(clippy::cast_precision_loss)]
    pub const fn estimate_filter_flops(num_elements: usize) -> f64 {
        (num_elements as f64) * 2.0
    }

    /// Estimate FLOPs for JOIN operation
    ///
    /// Hash join: Build hash table (5 FLOPs/elem) + Probe (5 FLOPs/elem)
    ///
    /// # Arguments
    /// * `left_size` - Number of elements in left table
    /// * `right_size` - Number of elements in right table
    ///
    /// # Returns
    /// Estimated FLOPs
    #[must_use]
    #[allow(clippy::cast_precision_loss)]
    pub const fn estimate_join_flops(left_size: usize, right_size: usize) -> f64 {
        // Build phase: 5 FLOPs per left element
        // Probe phase: 5 FLOPs per right element
        ((left_size + right_size) as f64) * 5.0
    }
}