//! TFT Inference Latency Benchmark for HFT Production
//!
//! **Objective**: Measure TFT inference latency and ensure P95 <5ms for production HFT.
//!
//! **Performance Targets**:
//! - P95 latency: <5ms (5000μs)
//! - Mean latency: <2ms (2000μs)
//! - Consistent performance: P99/P50 ratio <2.0
//!
//! **Comparison with Other Models**:
//! - DQN: P95 2.1ms ✅
//! - PPO: P95 3.2ms ✅
//! - MAMBA-2: P95 1.8ms ✅
//! - TFT: P95 target <5ms
//!
//! **Optimization Strategies** (if >5ms):
//! 1. CUDA Kernel Fusion: Reduce kernel launch overhead
//! 2. Batch Size = 1: Single-sample inference (lowest latency)
//! 3. Mixed Precision: FP16 inference (2x faster)
//! 4. Model Quantization: INT8 inference (4x faster)
//! 5. Attention Optimization: Flash Attention (2-4x faster)

#![allow(unused_crate_dependencies)]

use candle_core::{Device, Tensor};
use ml::tft::{TFTConfig, TemporalFusionTransformer};
use ml::MLError;
use std::time::Instant;

/// Helper: Create realistic TFT input tensors for benchmarking
fn create_tft_inputs(
    config: &TFTConfig,
    device: &Device,
) -> Result<(Tensor, Tensor, Tensor), MLError> {
    // Static features: [batch=1, num_static_features]
    let static_data = vec![0.5f32; config.num_static_features];
    let static_features = Tensor::from_slice(&static_data, (1, config.num_static_features), device)?;

    // Historical features: [batch=1, seq_len, num_unknown_features]
    let hist_len = config.sequence_length;
    let hist_dim = config.num_unknown_features;
    let hist_data = vec![0.5f32; hist_len * hist_dim];
    let historical_features = Tensor::from_slice(&hist_data, (1, hist_len, hist_dim), device)?;

    // Future features: [batch=1, prediction_horizon, num_known_features]
    let fut_len = config.prediction_horizon;
    let fut_dim = config.num_known_features;
    let fut_data = vec![0.5f32; fut_len * fut_dim];
    let future_features = Tensor::from_slice(&fut_data, (1, fut_len, fut_dim), device)?;

    Ok((static_features, historical_features, future_features))
}

/// Primary benchmark: TFT inference latency with statistical analysis
#[test]
fn test_tft_inference_latency_p95_target() -> Result<(), MLError> {
    println!("\n=== TFT Inference Latency Benchmark ===");
    println!("Target: P95 <5ms (5000μs) for production HFT\n");

    let device = Device::cuda_if_available(0)?;
    println!("Device: {:?}", device);

    // Production-realistic TFT configuration
    let config = TFTConfig {
        input_dim: 64,
        hidden_dim: 128,
        num_heads: 8,
        num_layers: 3,
        prediction_horizon: 10,
        sequence_length: 50,
        num_quantiles: 9,
        num_static_features: 5,
        num_known_features: 10,
        num_unknown_features: 20,
        learning_rate: 1e-3,
        batch_size: 1, // HFT: Single-sample inference for lowest latency
        dropout_rate: 0.0, // Inference mode: No dropout
        l2_regularization: 1e-4,
        use_flash_attention: true,
        mixed_precision: false, // Test FP32 baseline first
        memory_efficient: true,
        max_inference_latency_us: 5000,
        target_throughput_pps: 100_000,
    };

    let mut tft = TemporalFusionTransformer::new(config.clone())?;

    // Prepare inputs
    let (static_features, historical_features, future_features) = create_tft_inputs(&config, &device)?;

    // ==================== WARMUP PHASE ====================
    // Critical for CUDA: Ensure kernels are compiled and caches are warm
    println!("Warmup: 5 iterations (CUDA kernel compilation)");
    for _ in 0..5 {
        let _ = tft.forward(&static_features, &historical_features, &future_features)?;
    }

    // ==================== BENCHMARK PHASE ====================
    // 100 iterations for stable statistics
    println!("Benchmark: 100 iterations for stable statistics\n");
    let num_iterations = 100;
    let mut latencies_us = Vec::with_capacity(num_iterations);

    for _ in 0..num_iterations {
        let start = Instant::now();
        let _output = tft.forward(&static_features, &historical_features, &future_features)?;
        let elapsed_us = start.elapsed().as_micros() as u64;
        latencies_us.push(elapsed_us);
    }

    // ==================== STATISTICAL ANALYSIS ====================
    latencies_us.sort_unstable();

    let mean = latencies_us.iter().sum::<u64>() as f64 / latencies_us.len() as f64;
    let p50 = latencies_us[latencies_us.len() / 2];
    let p95 = latencies_us[latencies_us.len() * 95 / 100];
    let p99 = latencies_us[latencies_us.len() * 99 / 100];
    let min = latencies_us[0];
    let max = latencies_us[latencies_us.len() - 1];

    // Consistency metric: P99/P50 ratio (lower is better)
    let consistency_ratio = p99 as f64 / p50 as f64;

    // ==================== RESULTS ====================
    println!("📊 TFT Inference Latency Statistics:");
    println!("  Mean:   {:>6}μs ({:.2}ms)", mean as u64, mean / 1000.0);
    println!("  P50:    {:>6}μs ({:.2}ms)", p50, p50 as f64 / 1000.0);
    println!("  P95:    {:>6}μs ({:.2}ms) ← TARGET <5ms", p95, p95 as f64 / 1000.0);
    println!("  P99:    {:>6}μs ({:.2}ms)", p99, p99 as f64 / 1000.0);
    println!("  Min:    {:>6}μs ({:.2}ms)", min, min as f64 / 1000.0);
    println!("  Max:    {:>6}μs ({:.2}ms)", max, max as f64 / 1000.0);
    println!("  Consistency (P99/P50): {:.2}x", consistency_ratio);
    println!();

    // ==================== VALIDATION ====================
    // Primary target: P95 <5ms (5000μs)
    if p95 < 5000 {
        println!("✅ PASS: P95 latency {}μs is <5ms target", p95);
    } else {
        println!("⚠️  WARNING: P95 latency {}μs exceeds 5ms target", p95);
        println!("\n🔧 Optimization Strategies:");
        println!("  1. Enable Flash Attention (2-4x speedup)");
        println!("  2. Mixed Precision FP16 (2x speedup)");
        println!("  3. Model Quantization INT8 (4x speedup)");
        println!("  4. Reduce hidden_dim or num_layers");
        println!("  5. CUDA kernel fusion");
    }

    // Secondary target: Mean <2ms
    if mean < 2000.0 {
        println!("✅ PASS: Mean latency {:.0}μs is <2ms", mean);
    } else {
        println!("⚠️  INFO: Mean latency {:.0}μs exceeds 2ms (non-critical)", mean);
    }

    // Consistency check: P99/P50 ratio <2.0
    if consistency_ratio < 2.0 {
        println!("✅ PASS: Consistency ratio {:.2}x is <2.0 (stable performance)", consistency_ratio);
    } else {
        println!("⚠️  WARNING: Consistency ratio {:.2}x exceeds 2.0 (high variance)", consistency_ratio);
    }

    println!();

    // Test assertion: P95 must be <5ms for production readiness
    assert!(
        p95 < 5000,
        "FAIL: TFT P95 latency {}μs exceeds 5ms target ({}ms)",
        p95,
        p95 as f64 / 1000.0
    );

    Ok(())
}

/// Comparison benchmark: TFT vs other models (DQN, PPO, MAMBA-2)
#[test]
fn test_tft_latency_comparison_with_other_models() -> Result<(), MLError> {
    println!("\n=== Model Inference Latency Comparison ===\n");

    let device = Device::cuda_if_available(0)?;

    // TFT benchmark (from above)
    let tft_config = TFTConfig {
        hidden_dim: 128,
        num_heads: 8,
        num_layers: 3,
        prediction_horizon: 10,
        sequence_length: 50,
        num_quantiles: 9,
        num_static_features: 5,
        num_known_features: 10,
        num_unknown_features: 20,
        batch_size: 1,
        dropout_rate: 0.0,
        ..Default::default()
    };

    let mut tft = TemporalFusionTransformer::new(tft_config.clone())?;
    let (static_features, historical_features, future_features) = create_tft_inputs(&tft_config, &device)?;

    // Warmup
    for _ in 0..5 {
        let _ = tft.forward(&static_features, &historical_features, &future_features)?;
    }

    // Benchmark TFT
    let mut tft_latencies = Vec::new();
    for _ in 0..100 {
        let start = Instant::now();
        let _ = tft.forward(&static_features, &historical_features, &future_features)?;
        tft_latencies.push(start.elapsed().as_micros() as u64);
    }
    tft_latencies.sort_unstable();

    let tft_mean = tft_latencies.iter().sum::<u64>() as f64 / tft_latencies.len() as f64;
    let tft_p50 = tft_latencies[tft_latencies.len() / 2];
    let tft_p95 = tft_latencies[tft_latencies.len() * 95 / 100];
    let tft_p99 = tft_latencies[tft_latencies.len() * 99 / 100];

    // ==================== COMPARISON TABLE ====================
    println!("Model         Mean      P50       P95       P99     Target    Status");
    println!("─────────────────────────────────────────────────────────────────────");
    println!("DQN           150μs     200μs     2.1ms     3ms     <5ms      ✅");
    println!("PPO           280μs     324μs     3.2ms     4ms     <5ms      ✅");
    println!("MAMBA-2       400μs     500μs     1.8ms    2.5ms    <5ms      ✅");
    println!(
        "TFT          {: >4}μs    {: >4}μs    {: >4.1}ms   {: >4.1}ms    <5ms      {}",
        tft_mean as u64,
        tft_p50,
        tft_p95 as f64 / 1000.0,
        tft_p99 as f64 / 1000.0,
        if tft_p95 < 5000 { "✅" } else { "❌" }
    );
    println!();

    // Verify TFT meets target
    assert!(
        tft_p95 < 5000,
        "TFT P95 latency {}μs exceeds 5ms target",
        tft_p95
    );

    Ok(())
}

/// Test: TFT latency with different batch sizes (batch optimization)
#[test]
fn test_tft_batch_size_latency_tradeoff() -> Result<(), MLError> {
    println!("\n=== TFT Batch Size Latency Trade-off ===\n");

    let device = Device::cuda_if_available(0)?;

    let batch_sizes = vec![1, 2, 4, 8];

    println!("Batch    Total      Per-Sample    Throughput");
    println!("Size     Latency    Latency       (samples/sec)");
    println!("───────────────────────────────────────────────");

    for batch_size in batch_sizes {
        let config = TFTConfig {
            hidden_dim: 128,
            num_heads: 8,
            num_layers: 3,
            prediction_horizon: 10,
            sequence_length: 50,
            num_quantiles: 9,
            num_static_features: 5,
            num_known_features: 10,
            num_unknown_features: 20,
            batch_size,
            dropout_rate: 0.0,
            ..Default::default()
        };

        let mut tft = TemporalFusionTransformer::new(config.clone())?;

        // Create batched inputs
        let static_data = vec![0.5f32; batch_size * config.num_static_features];
        let static_features = Tensor::from_slice(
            &static_data,
            (batch_size, config.num_static_features),
            &device,
        )?;

        let hist_data = vec![0.5f32; batch_size * config.sequence_length * config.num_unknown_features];
        let historical_features = Tensor::from_slice(
            &hist_data,
            (batch_size, config.sequence_length, config.num_unknown_features),
            &device,
        )?;

        let fut_data = vec![0.5f32; batch_size * config.prediction_horizon * config.num_known_features];
        let future_features = Tensor::from_slice(
            &fut_data,
            (batch_size, config.prediction_horizon, config.num_known_features),
            &device,
        )?;

        // Warmup
        for _ in 0..5 {
            let _ = tft.forward(&static_features, &historical_features, &future_features)?;
        }

        // Benchmark
        let num_iterations = 50;
        let mut latencies = Vec::new();
        for _ in 0..num_iterations {
            let start = Instant::now();
            let _ = tft.forward(&static_features, &historical_features, &future_features)?;
            latencies.push(start.elapsed().as_micros() as u64);
        }

        let avg_latency_us = latencies.iter().sum::<u64>() / num_iterations;
        let per_sample_us = avg_latency_us as f64 / batch_size as f64;
        let throughput = 1_000_000.0 / per_sample_us;

        println!(
            "{: >4}     {: >6}μs    {: >6.0}μs      {: >6.0}",
            batch_size, avg_latency_us, per_sample_us, throughput
        );
    }

    println!("\n💡 Insight: Larger batches amortize overhead, increasing throughput");
    println!("   HFT recommendation: batch_size=1 for lowest latency (<5ms)");
    println!();

    Ok(())
}

/// Test: TFT latency with Flash Attention enabled/disabled
#[test]
fn test_tft_flash_attention_speedup() -> Result<(), MLError> {
    println!("\n=== TFT Flash Attention Speedup ===\n");

    let device = Device::cuda_if_available(0)?;

    // Test both configurations
    let flash_configs = vec![
        ("Standard Attention", false),
        ("Flash Attention", true),
    ];

    println!("Configuration           P50       P95      Speedup");
    println!("────────────────────────────────────────────────────");

    let mut baseline_p95 = 0u64;

    for (name, use_flash) in flash_configs {
        let config = TFTConfig {
            hidden_dim: 128,
            num_heads: 8,
            num_layers: 3,
            prediction_horizon: 10,
            sequence_length: 50,
            num_quantiles: 9,
            num_static_features: 5,
            num_known_features: 10,
            num_unknown_features: 20,
            use_flash_attention: use_flash,
            batch_size: 1,
            dropout_rate: 0.0,
            ..Default::default()
        };

        let mut tft = TemporalFusionTransformer::new(config.clone())?;
        let (static_features, historical_features, future_features) = create_tft_inputs(&config, &device)?;

        // Warmup
        for _ in 0..5 {
            let _ = tft.forward(&static_features, &historical_features, &future_features)?;
        }

        // Benchmark
        let mut latencies = Vec::new();
        for _ in 0..100 {
            let start = Instant::now();
            let _ = tft.forward(&static_features, &historical_features, &future_features)?;
            latencies.push(start.elapsed().as_micros() as u64);
        }
        latencies.sort_unstable();

        let p50 = latencies[latencies.len() / 2];
        let p95 = latencies[latencies.len() * 95 / 100];

        let speedup = if baseline_p95 > 0 {
            baseline_p95 as f64 / p95 as f64
        } else {
            baseline_p95 = p95;
            1.0
        };

        println!(
            "{: <22}  {: >6}μs   {: >6}μs    {:.2}x",
            name,
            p50,
            p95,
            speedup
        );
    }

    println!("\n💡 Flash Attention expected speedup: 2-4x for long sequences");
    println!();

    Ok(())
}

/// Test: TFT latency with different model sizes (hidden_dim, num_layers)
#[test]
fn test_tft_model_size_latency_scaling() -> Result<(), MLError> {
    println!("\n=== TFT Model Size Latency Scaling ===\n");

    let device = Device::cuda_if_available(0)?;

    // Test configurations: (hidden_dim, num_layers, name)
    let model_configs = vec![
        (64, 2, "Small"),
        (128, 3, "Medium (Production)"),
        (256, 4, "Large"),
        (512, 6, "Extra Large"),
    ];

    println!("Model Size            Hidden  Layers   P95       Status");
    println!("─────────────────────────────────────────────────────────");

    for (hidden_dim, num_layers, name) in model_configs {
        let config = TFTConfig {
            hidden_dim,
            num_heads: 8,
            num_layers,
            prediction_horizon: 10,
            sequence_length: 50,
            num_quantiles: 9,
            num_static_features: 5,
            num_known_features: 10,
            num_unknown_features: 20,
            batch_size: 1,
            dropout_rate: 0.0,
            use_flash_attention: true,
            ..Default::default()
        };

        let mut tft = TemporalFusionTransformer::new(config.clone())?;
        let (static_features, historical_features, future_features) = create_tft_inputs(&config, &device)?;

        // Warmup
        for _ in 0..5 {
            let _ = tft.forward(&static_features, &historical_features, &future_features)?;
        }

        // Benchmark
        let mut latencies = Vec::new();
        for _ in 0..100 {
            let start = Instant::now();
            let _ = tft.forward(&static_features, &historical_features, &future_features)?;
            latencies.push(start.elapsed().as_micros() as u64);
        }
        latencies.sort_unstable();

        let p95 = latencies[latencies.len() * 95 / 100];
        let status = if p95 < 5000 { "✅" } else { "⚠️" };

        println!(
            "{: <22} {: >6}  {: >6}   {: >6}μs   {}",
            name, hidden_dim, num_layers, p95, status
        );
    }

    println!("\n💡 Latency scales with model size: Smaller models = lower latency");
    println!("   Production: Medium (128, 3 layers) balances accuracy and latency");
    println!();

    Ok(())
}

/// Test: TFT memory usage during inference
#[test]
fn test_tft_inference_memory_usage() -> Result<(), MLError> {
    println!("\n=== TFT Inference Memory Usage ===\n");

    let device = Device::cuda_if_available(0)?;

    let config = TFTConfig {
        hidden_dim: 128,
        num_heads: 8,
        num_layers: 3,
        prediction_horizon: 10,
        sequence_length: 50,
        num_quantiles: 9,
        num_static_features: 5,
        num_known_features: 10,
        num_unknown_features: 20,
        batch_size: 1,
        dropout_rate: 0.0,
        ..Default::default()
    };

    let mut tft = TemporalFusionTransformer::new(config.clone())?;
    let (static_features, historical_features, future_features) = create_tft_inputs(&config, &device)?;

    // Single inference
    let _ = tft.forward(&static_features, &historical_features, &future_features)?;

    // Estimate memory usage
    let static_mem = config.num_static_features * 4; // f32
    let hist_mem = config.sequence_length * config.num_unknown_features * 4;
    let fut_mem = config.prediction_horizon * config.num_known_features * 4;
    let output_mem = config.prediction_horizon * config.num_quantiles * 4;

    let total_input_mem = static_mem + hist_mem + fut_mem;
    let total_mem = total_input_mem + output_mem;

    println!("Memory Usage Breakdown:");
    println!("  Static features:     {} bytes ({:.2} KB)", static_mem, static_mem as f64 / 1024.0);
    println!("  Historical features: {} bytes ({:.2} KB)", hist_mem, hist_mem as f64 / 1024.0);
    println!("  Future features:     {} bytes ({:.2} KB)", fut_mem, fut_mem as f64 / 1024.0);
    println!("  Output (quantiles):  {} bytes ({:.2} KB)", output_mem, output_mem as f64 / 1024.0);
    println!("  ──────────────────────────────────────────");
    println!("  Total per inference: {} bytes ({:.2} KB)", total_mem, total_mem as f64 / 1024.0);
    println!();

    println!("💡 Target: <10MB per inference (TFT well below at ~{:.2} KB)", total_mem as f64 / 1024.0);
    println!();

    // Verify memory is reasonable
    assert!(total_mem < 10 * 1024 * 1024, "Memory usage exceeds 10MB target");

    Ok(())
}