trueno 0.17.4

//! Basic tuner coverage tests: features, errors, regressor, builder.

use crate::tuner::*;

#[test]
fn test_tuner_features_validate() {
    let features = TunerFeatures::builder().build();
    assert!(features.validate().is_ok());

    // Test with NaN
    let mut bad_features = features.clone();
    bad_features.model_params_b = f32::NAN;
    assert!(bad_features.validate().is_err());
}

#[test]
fn test_tuner_error_display() {
    let err = TunerError::InvalidFeature("test".to_string());
    assert!(format!("{}", err).contains("Invalid feature"));

    let err = TunerError::InsufficientData(5);
    assert!(format!("{}", err).contains("Insufficient"));

    let err = TunerError::Serialization("test".to_string());
    assert!(format!("{}", err).contains("Serialization"));

    let err = TunerError::ModelNotFound;
    assert!(format!("{}", err).contains("not found"));

    let err = TunerError::PredictionFailed("test".to_string());
    assert!(format!("{}", err).contains("Prediction failed"));
}

#[test]
fn test_throughput_regressor_predict_raw() {
    let regressor = ThroughputRegressor::new();
    let features = TunerFeatures::builder().batch_size(4).build();
    let vec = features.to_vector();
    let raw = regressor.predict_raw(&vec);
    assert!(raw > 0.0);
}

#[test]
fn test_brick_tuner_recommend() {
    let tuner = BrickTuner::new();
    let features = TunerFeatures::builder().model_params_b(1.5).batch_size(4).build();
    let rec = tuner.recommend(&features);

    assert!(rec.throughput.predicted_tps > 0.0);
    assert!(!rec.suggested_experiments.is_empty());
}

#[test]
fn test_experiment_suggestion_display() {
    let exp = ExperimentSuggestion::IncreaseBatchSize { from: 1, to: 4 };
    assert!(format!("{}", exp).contains("Increase batch size"));

    let exp = ExperimentSuggestion::EnableCudaGraphs;
    assert!(format!("{}", exp).contains("CUDA graphs"));

    let exp = ExperimentSuggestion::TryKernel { kernel: KernelType::BatchedQ4K };
    assert!(format!("{}", exp).contains("kernel"));

    let exp = ExperimentSuggestion::ReduceSequenceLength { factor: 0.5 };
    assert!(format!("{}", exp).contains("sequence"));

    let exp = ExperimentSuggestion::EnableMultiKvCache { count: 4 };
    assert!(format!("{}", exp).contains("KV"));
}

#[test]
fn test_tuner_data_collector() {
    let collector = TunerDataCollector::new();
    assert!(collector.is_empty());
    assert_eq!(collector.len(), 0);
    assert!(collector.samples().is_empty());
}

#[test]
fn test_feature_extractor_default() {
    let extractor = FeatureExtractor::new();
    assert!(extractor.hardware.is_none());
}

#[test]
fn test_feature_extractor_debug() {
    let extractor = FeatureExtractor::new();
    let debug_str = format!("{:?}", extractor);
    assert!(debug_str.contains("FeatureExtractor"));
}

#[test]
fn test_chrono_lite_now() {
    let timestamp = crate::tuner::chrono_lite_now();
    let parsed: u64 = timestamp.parse().expect("Should be a number");
    assert!(parsed > 0);
}

#[test]
fn test_pad_right() {
    assert_eq!(crate::tuner::pad_right("test", 10), "test      ");
    assert_eq!(crate::tuner::pad_right("longstring", 5), "longs");
}

#[test]
fn test_validation_infinite_features() {
    let features = TunerFeatures { model_params_b: f32::INFINITY, ..Default::default() };
    let result = features.validate();
    assert!(result.is_err());
    assert!(result.unwrap_err().to_string().contains("Infinite"));
}

#[test]
fn test_validation_out_of_range() {
    let features = TunerFeatures {
        batch_size_norm: 2.0, // Out of [0, 1]
        ..Default::default()
    };
    let result = features.validate();
    assert!(result.is_err());
    assert!(result.unwrap_err().to_string().contains("outside [0, 1]"));
}

#[test]
fn test_validation_bad_quant_onehot() {
    let features = TunerFeatures {
        quant_type_onehot: [0.5, 0.5, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], // Sums to 1 but invalid one-hot
        ..Default::default()
    };
    // This should actually pass since sum is 1.0
    assert!(features.validate().is_ok());

    // Now test with sum != 1
    let features2 = TunerFeatures {
        quant_type_onehot: [0.5, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0], // Sums to 0.5
        ..Default::default()
    };
    let result = features2.validate();
    assert!(result.is_err());
}

#[test]
fn test_validation_bad_kernel_onehot() {
    let mut features = TunerFeatures {
        kernel_type_onehot: [0.0; 16], // All zeros, sum = 0
        ..Default::default()
    };
    // Zero sum is allowed (unspecified kernel)
    assert!(features.validate().is_ok());

    // Sum != 0 and != 1 should fail
    features.kernel_type_onehot[0] = 0.5;
    let result = features.validate();
    assert!(result.is_err());
}

#[test]
fn test_builder_gpu_l2_cache_mb() {
    let features = TunerFeatures::builder()
        .gpu_l2_cache_mb(96.0) // 96MB L2 cache
        .build();
    // Normalized: 96 / 128 = 0.75
    assert!((features.gpu_l2_cache_norm - 0.75).abs() < 0.01);
}

#[test]
fn test_builder_is_zero_copy() {
    let features_enabled = TunerFeatures::builder().is_zero_copy(true).build();
    assert_eq!(features_enabled.is_zero_copy, 1.0);

    let features_disabled = TunerFeatures::builder().is_zero_copy(false).build();
    assert_eq!(features_disabled.is_zero_copy, 0.0);
}

#[test]
fn test_builder_hardware() {
    use crate::hardware::{GpuBackend, GpuCapability};

    let gpu = GpuCapability {
        vendor: "NVIDIA".to_string(),
        model: "Test GPU".to_string(),
        backend: GpuBackend::Cuda,
        compute_capability: Some("8.9".to_string()),
        peak_tflops_fp32: 100.0,
        peak_tflops_tensor: Some(400.0),
        memory_bw_gbps: 1000.0,
        vram_gb: 24.0,
    };

    // Directly test the normalization without HardwareCapability
    let features = TunerFeatures::builder()
        .gpu_mem_bw_gbs(gpu.memory_bw_gbps as f32)
        .gpu_compute_tflops(gpu.peak_tflops_fp32 as f32)
        .build();

    // Memory BW: 1000 / 3000 approx 0.333
    assert!((features.gpu_mem_bw_norm - (1000.0 / 3000.0)).abs() < 0.01);
    // Compute: 100 / 500 = 0.2
    assert!((features.gpu_compute_norm - 0.2).abs() < 0.01);
}

#[test]
fn test_brick_tuner_train() {
    let mut tuner = BrickTuner::new();

    // Create minimal training data
    let data: Vec<(TunerFeatures, f32)> = (0..15)
        .map(|i| {
            let features = TunerFeatures::builder()
                .batch_size((i % 4) as u32 + 1)
                .model_params_b(1.5 + (i as f32) * 0.1)
                .build();
            (features, 100.0 + (i as f32) * 10.0)
        })
        .collect();

    let result = tuner.train(&data);
    assert!(result.is_ok());
    assert_eq!(tuner.sample_count, 15);
}

#[test]
fn test_brick_tuner_train_insufficient_data() {
    let mut tuner = BrickTuner::new();

    // Too few samples
    let data: Vec<(TunerFeatures, f32)> = (0..5)
        .map(|i| {
            let features = TunerFeatures::builder().batch_size(i as u32 + 1).build();
            (features, 100.0)
        })
        .collect();

    let result = tuner.train(&data);
    assert!(result.is_err());
    assert!(matches!(result.unwrap_err(), TunerError::InsufficientData(5)));
}