aprender-db 0.38.0

//! GPU-Accelerated Database Aggregations
//!
//! This example demonstrates real GPU execution for database aggregations.
//! It shows the GPU being utilized for SUM, MIN, MAX, COUNT, and fused filter+sum operations.
//!
//! Run with: cargo run --example gpu_aggregations --features gpu
//!
//! Requirements:
//! - GPU hardware (Vulkan/Metal/DX12 compatible)
//! - Build with --features gpu flag

use arrow::array::Int32Array;
use std::time::Instant;
use trueno_db::gpu::GpuEngine;

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    println!("=== Trueno-DB GPU-Accelerated Database Aggregations ===\n");

    // Initialize GPU engine
    println!("🔧 Initializing GPU engine...");
    let start = Instant::now();

    let gpu = match GpuEngine::new().await {
        Ok(engine) => {
            println!("✅ GPU engine initialized in {:?}", start.elapsed());
            println!("   Device features: {:?}", engine.device.features());
            println!();
            engine
        }
        Err(e) => {
            eprintln!("❌ GPU not available: {}", e);
            eprintln!("   Hint: Ensure you have Vulkan/Metal/DX12 drivers installed");
            eprintln!("   Hint: Build with: cargo run --example gpu_aggregations --features gpu");
            return Ok(());
        }
    };

    // Test Case 1: SUM Aggregation (100K rows)
    println!("=== Test Case 1: GPU SUM (100,000 rows) ===");
    let data: Vec<i32> = (1..=100_000).collect();
    let array = Int32Array::from(data.clone());
    // Expected: sum of [1, 2, ..., 100000] = 100000 * 100001 / 2 = 5,000,050,000
    // But this overflows i32, so we'll use a smaller range

    let start = Instant::now();
    let gpu_sum = gpu.sum_i32(&array).await?;
    let gpu_time = start.elapsed();

    let expected_sum: i64 = data.iter().map(|&x| x as i64).sum();

    println!("  Data: [1, 2, 3, ..., 100000]");
    println!("  GPU Result: {} (i32)", gpu_sum);
    println!("  Expected: {} (i64)", expected_sum);
    println!("  GPU Time: {:?}", gpu_time);
    println!("  ✅ Correct: {}", gpu_sum as i64 == expected_sum);
    println!();

    // Test Case 2: MIN Aggregation
    println!("=== Test Case 2: GPU MIN (100,000 rows) ===");
    let data: Vec<i32> = vec![999; 50_000]
        .into_iter()
        .chain(vec![42].into_iter())
        .chain(vec![999; 49_999].into_iter())
        .collect();
    let array = Int32Array::from(data);

    let start = Instant::now();
    let gpu_min = gpu.min_i32(&array).await?;
    let gpu_time = start.elapsed();

    println!("  Data: [999 (50K times), 42, 999 (49,999 times)]");
    println!("  GPU Result: {}", gpu_min);
    println!("  Expected: 42");
    println!("  GPU Time: {:?}", gpu_time);
    println!("  ✅ Correct: {}", gpu_min == 42);
    println!();

    // Test Case 3: MAX Aggregation
    println!("=== Test Case 3: GPU MAX (100,000 rows) ===");
    let data: Vec<i32> = vec![1; 50_000]
        .into_iter()
        .chain(vec![9999].into_iter())
        .chain(vec![1; 49_999].into_iter())
        .collect();
    let array = Int32Array::from(data);

    let start = Instant::now();
    let gpu_max = gpu.max_i32(&array).await?;
    let gpu_time = start.elapsed();

    println!("  Data: [1 (50K times), 9999, 1 (49,999 times)]");
    println!("  GPU Result: {}", gpu_max);
    println!("  Expected: 9999");
    println!("  GPU Time: {:?}", gpu_time);
    println!("  ✅ Correct: {}", gpu_max == 9999);
    println!();

    // Test Case 4: COUNT Aggregation
    println!("=== Test Case 4: GPU COUNT (1,000,000 rows) ===");
    let data: Vec<i32> = (1..=1_000_000).collect();
    let array = Int32Array::from(data);

    let start = Instant::now();
    let gpu_count = gpu.count(&array).await?;
    let gpu_time = start.elapsed();

    println!("  Data: 1,000,000 integers");
    println!("  GPU Result: {}", gpu_count);
    println!("  Expected: 1000000");
    println!("  GPU Time: {:?}", gpu_time);
    println!("  ✅ Correct: {}", gpu_count == 1_000_000);
    println!();

    // Test Case 5: Fused Filter+Sum (Kernel Fusion - Muda Elimination)
    println!("=== Test Case 5: GPU Fused Filter+Sum (100,000 rows) ===");
    println!("  Operation: SELECT SUM(value) WHERE value > 50000");
    let data: Vec<i32> = (1..=100_000).collect();
    let array = Int32Array::from(data);

    let start = Instant::now();
    let gpu_result = gpu.fused_filter_sum(&array, 50_000, "gt").await?;
    let gpu_time = start.elapsed();

    // Expected: sum of [50001, 50002, ..., 100000]
    let expected: i64 = (50_001..=100_000).sum();

    println!("  Data: [1, 2, 3, ..., 100000]");
    println!("  Filter: value > 50000");
    println!("  GPU Result: {} (i32)", gpu_result);
    println!("  Expected: {} (i64)", expected);
    println!("  GPU Time: {:?}", gpu_time);
    println!("  ✅ Correct: {}", gpu_result as i64 == expected);
    println!("  🎯 Toyota Way: Muda elimination (single-pass, no intermediate buffer)");
    println!();

    // Test Case 6: Large-Scale Aggregation (10M rows)
    println!("=== Test Case 6: Large-Scale GPU SUM (10,000,000 rows) ===");
    let chunk_size = 1_000_000;
    let mut total_sum: i64 = 0;
    let start = Instant::now();

    for chunk_idx in 0..10 {
        let data: Vec<i32> = ((chunk_idx * chunk_size + 1)..=((chunk_idx + 1) * chunk_size))
            .map(|x| x % 1_000_000)
            .collect();
        let array = Int32Array::from(data);
        let chunk_sum = gpu.sum_i32(&array).await?;
        total_sum += chunk_sum as i64;
    }

    let gpu_time = start.elapsed();

    println!("  Data: 10M integers (processed in 1M chunks)");
    println!("  GPU Result: {}", total_sum);
    println!("  GPU Time: {:?}", gpu_time);
    println!(
        "  Throughput: {:.2} GB/s",
        (10_000_000 * 4) as f64 / gpu_time.as_secs_f64() / 1_000_000_000.0
    );
    println!();

    // GPU Device Information
    println!("=== GPU Device Information ===");
    println!("  Compute Backend: wgpu (Vulkan/Metal/DX12)");
    println!("  Workgroup Size: 256 threads (8 GPU warps)");
    println!("  Parallel Reduction: Harris 2007 algorithm");
    println!("  Kernel Fusion: Single-pass filter+aggregation");
    println!();

    println!("=== Performance Notes ===");
    println!("  ✅ Zero-copy transfers via Arrow columnar format");
    println!("  ✅ Parallel reduction for O(log N) aggregations");
    println!("  ✅ Kernel fusion eliminates intermediate buffers (Muda)");
    println!("  ✅ Morsel-based paging prevents VRAM OOM (Poka-Yoke)");
    println!();

    println!("🎉 GPU aggregations complete!");

    Ok(())
}