#!/usr/bin/env rust-script

//! Standalone test for magnitude pruning implementation
//! This allows us to test the functionality without dealing with broader compilation issues

use std::collections::HashMap;

/// Result type for this standalone test
type Result<T> = std::result::Result<T, Box<dyn std::error::Error>>;

/// Simple error type for testing
#[derive(Debug)]
struct TestError(String);

impl std::fmt::Display for TestError {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        write!(f, "Test error: {}", self.0)
    }
}

impl std::error::Error for TestError {}

// Include our implementation directly for standalone testing
// (This is normally not recommended, but useful for TDD when dependencies have issues)

/// Configuration for magnitude-based pruning
#[derive(Debug, Clone, PartialEq)]
pub struct PruningConfig {
    /// Pruning strategy to apply
    pub strategy: PruningStrategy,
    /// Target sparsity ratio (0.0 to 1.0)
    pub sparsity_ratio: f64,
    /// Pruning granularity
    pub granularity: PruningGranularity,
    /// Optional pruning schedule for gradual pruning
    pub schedule: Option<PruningSchedule>,
    /// Random seed for deterministic pruning
    pub seed: Option<u64>,
}

/// Strategy for selecting weights to prune
#[derive(Debug, Clone, PartialEq)]
pub enum PruningStrategy {
    /// Global magnitude pruning across all parameters
    Global,
    /// Layer-wise magnitude pruning within each layer
    LayerWise,
    /// Structured pruning (channels, filters)
    Structured {
        /// Type of structured pattern
        pattern_type: StructuredPattern,
    },
}

/// Granularity of pruning operations
#[derive(Debug, Clone, PartialEq)]
pub enum PruningGranularity {
    /// Element-wise pruning (unstructured)
    Element,
    /// Channel-wise pruning
    Channel,
    /// Filter-wise pruning
    Filter,
    /// Block-wise pruning
    Block { height: usize, width: usize },
}

/// Structured pruning patterns
#[derive(Debug, Clone, PartialEq)]
pub enum StructuredPattern {
    /// Channel pruning
    Channel,
    /// Filter pruning
    Filter,
    /// N:M sparsity pattern
    NM { n: usize, m: usize },
}

/// Pruning schedule for gradual sparsification during training
#[derive(Debug, Clone, PartialEq)]
pub enum PruningSchedule {
    /// Linear schedule from initial to final sparsity
    Linear {
        /// Initial sparsity ratio
        initial_sparsity: f64,
        /// Final sparsity ratio  
        final_sparsity: f64,
        /// Duration in training steps
        duration_steps: usize,
    },
}

/// Binary mask for indicating which weights to keep/prune
#[derive(Debug, Clone)]
pub struct PruningMask {
    /// Binary mask values (true = keep, false = prune)
    mask: Vec<bool>,
    /// Shape of the mask
    shape: Vec<usize>,
    /// Number of parameters kept
    kept_parameters: usize,
    /// Total number of parameters
    total_parameters: usize,
}

/// Statistics about pruning results
#[derive(Debug, Clone, PartialEq)]
pub struct PruningStats {
    /// Original parameter count
    pub original_params: usize,
    /// Parameters after pruning
    pub pruned_params: usize,
    /// Achieved sparsity ratio
    pub sparsity_ratio: f64,
    /// Compression ratio
    pub compression_ratio: f64,
    /// Memory savings (in bytes, estimated)
    pub memory_savings: usize,
}

/// Main struct for magnitude-based pruning operations
#[derive(Debug)]
pub struct MagnitudePruning;

/// Simple tensor for testing
#[derive(Debug, Clone)]
pub struct DenseTensor {
    pub data: Vec<f32>,
    pub shape: Vec<usize>,
}

impl DenseTensor {
    pub fn new(data: Vec<f32>, shape: Vec<usize>) -> Self {
        assert_eq!(data.len(), shape.iter().product::<usize>());
        Self { data, shape }
    }
    
    pub fn shape(&self) -> &[usize] {
        &self.shape
    }
    
    pub fn data(&self) -> &[f32] {
        &self.data
    }
}

impl PruningConfig {
    /// Create global magnitude pruning configuration
    pub fn global(sparsity_ratio: f64) -> Self {
        Self {
            strategy: PruningStrategy::Global,
            sparsity_ratio,
            granularity: PruningGranularity::Element,
            schedule: None,
            seed: None,
        }
    }
    
    /// Create layer-wise magnitude pruning configuration
    pub fn layerwise(sparsity_ratio: f64) -> Self {
        Self {
            strategy: PruningStrategy::LayerWise,
            sparsity_ratio,
            granularity: PruningGranularity::Element,
            schedule: None,
            seed: None,
        }
    }
    
    /// Validate configuration parameters
    pub fn validate(&self) -> Result<()> {
        if !(0.0..=1.0).contains(&self.sparsity_ratio) {
            return Err(Box::new(TestError("Sparsity ratio must be between 0.0 and 1.0".into())));
        }
        Ok(())
    }
}

impl PruningMask {
    /// Create new pruning mask
    pub fn new(mask: Vec<bool>, shape: Vec<usize>) -> Self {
        let total_parameters = shape.iter().product::<usize>();
        let kept_parameters = mask.iter().filter(|&&x| x).count();
        
        Self {
            mask,
            shape,
            kept_parameters,
            total_parameters,
        }
    }
    
    /// Get mask values
    pub fn mask(&self) -> &[bool] {
        &self.mask
    }
    
    /// Get mask shape
    pub fn shape(&self) -> &[usize] {
        &self.shape
    }
    
    /// Get sparsity ratio
    pub fn sparsity_ratio(&self) -> f64 {
        1.0 - (self.kept_parameters as f64 / self.total_parameters as f64)
    }
    
    /// Apply mask to tensor (element-wise multiplication)
    pub fn apply(&self, tensor: &DenseTensor) -> Result<DenseTensor> {
        if tensor.shape != self.shape {
            return Err(Box::new(TestError("Tensor and mask shapes must match".into())));
        }
        
        let masked_data: Vec<f32> = tensor.data.iter()
            .zip(self.mask.iter())
            .map(|(&value, &keep)| if keep { value } else { 0.0 })
            .collect();
        
        Ok(DenseTensor::new(masked_data, tensor.shape.clone()))
    }
}

impl PruningStats {
    /// Create new pruning statistics
    pub fn new(original_params: usize, pruned_params: usize) -> Self {
        let sparsity_ratio = 1.0 - (pruned_params as f64 / original_params as f64);
        let compression_ratio = if pruned_params == 0 {
            f64::INFINITY
        } else {
            original_params as f64 / pruned_params as f64
        };
        
        // Estimate memory savings (assuming 4 bytes per f32)
        let memory_savings = (original_params - pruned_params) * 4;
        
        Self {
            original_params,
            pruned_params,
            sparsity_ratio,
            compression_ratio,
            memory_savings,
        }
    }
}

impl MagnitudePruning {
    /// Apply magnitude-based pruning to tensor
    pub fn prune(tensor: &DenseTensor, config: &PruningConfig) -> Result<DenseTensor> {
        config.validate()?;
        
        match config.strategy {
            PruningStrategy::Global => {
                Self::prune_global(tensor, config)
            }
            PruningStrategy::LayerWise => {
                Self::prune_layerwise(tensor, config)
            }
            PruningStrategy::Structured { .. } => {
                Err(Box::new(TestError("Structured pruning not implemented yet".into())))
            }
        }
    }
    
    /// Create pruning mask without modifying the original tensor
    pub fn create_mask(tensor: &DenseTensor, config: &PruningConfig) -> Result<PruningMask> {
        config.validate()?;
        
        match config.strategy {
            PruningStrategy::Global => {
                Self::create_global_mask(tensor, config)
            }
            PruningStrategy::LayerWise => {
                Self::create_layerwise_mask(tensor, config)
            }
            PruningStrategy::Structured { .. } => {
                Err(Box::new(TestError("Structured mask creation not implemented yet".into())))
            }
        }
    }
    
    /// Compute importance scores based on magnitude
    pub fn compute_importance_scores(tensor: &DenseTensor) -> Result<Vec<f32>> {
        let scores: Vec<f32> = tensor.data.iter().map(|&x| x.abs()).collect();
        Ok(scores)
    }
    
    /// Analyze pruning results
    pub fn analyze_pruning(original: &DenseTensor, pruned: &DenseTensor) -> Result<PruningStats> {
        if original.shape != pruned.shape {
            return Err(Box::new(TestError("Tensor shapes must match for analysis".into())));
        }
        
        let original_nonzeros = original.data.iter().filter(|&&x| x != 0.0).count();
        let pruned_nonzeros = pruned.data.iter().filter(|&&x| x != 0.0).count();
        
        Ok(PruningStats::new(original_nonzeros, pruned_nonzeros))
    }
    
    fn prune_global(tensor: &DenseTensor, config: &PruningConfig) -> Result<DenseTensor> {
        let mask = Self::create_global_mask(tensor, config)?;
        mask.apply(tensor)
    }
    
    fn prune_layerwise(tensor: &DenseTensor, config: &PruningConfig) -> Result<DenseTensor> {
        let mask = Self::create_layerwise_mask(tensor, config)?;
        mask.apply(tensor)
    }
    
    fn create_global_mask(tensor: &DenseTensor, config: &PruningConfig) -> Result<PruningMask> {
        let total_elements = tensor.data.len();
        let elements_to_keep = ((1.0 - config.sparsity_ratio) * total_elements as f64).round() as usize;
        
        // Compute importance scores (absolute magnitudes)
        let importance_scores = Self::compute_importance_scores(tensor)?;
        
        // Create (index, score) pairs for sorting
        let mut indexed_scores: Vec<(usize, f32)> = importance_scores.iter()
            .enumerate()
            .map(|(i, &score)| (i, score))
            .collect();
        
        // Sort by importance score in descending order (keep highest magnitude values)
        indexed_scores.sort_by(|a, b| b.1.total_cmp(&a.1));
        
        // Create mask: true for kept elements, false for pruned elements
        let mut mask_data = vec![false; total_elements];
        for i in 0..elements_to_keep {
            if i < indexed_scores.len() {
                let (index, _) = indexed_scores[i];
                mask_data[index] = true;
            }
        }
        
        Ok(PruningMask::new(mask_data, tensor.shape.clone()))
    }
    
    fn create_layerwise_mask(tensor: &DenseTensor, config: &PruningConfig) -> Result<PruningMask> {
        // For layer-wise pruning, we currently support 2D tensors (treating each row as a layer)
        if tensor.shape.len() != 2 {
            return Err(Box::new(TestError("Layer-wise pruning currently only supports 2D tensors".into())));
        }
        
        let rows = tensor.shape[0];
        let cols = tensor.shape[1];
        let elements_per_row = cols;
        let elements_to_keep_per_row = ((1.0 - config.sparsity_ratio) * elements_per_row as f64).round() as usize;
        
        let mut mask_data = vec![false; tensor.data.len()];
        
        // Process each row independently
        for row in 0..rows {
            let row_start = row * cols;
            let row_end = row_start + cols;
            
            // Get magnitudes for this row
            let row_scores: Vec<(usize, f32)> = (row_start..row_end)
                .map(|i| (i, tensor.data[i].abs()))
                .collect();
            
            // Sort by magnitude in descending order
            let mut sorted_scores = row_scores;
            sorted_scores.sort_by(|a, b| b.1.total_cmp(&a.1));
            
            // Keep the top elements_to_keep_per_row elements in this row
            for i in 0..elements_to_keep_per_row.min(sorted_scores.len()) {
                let (original_index, _) = sorted_scores[i];
                mask_data[original_index] = true;
            }
        }
        
        Ok(PruningMask::new(mask_data, tensor.shape.clone()))
    }
}

/// Test utilities
mod test_utils {
    use super::*;
    
    /// Create a simple 2D test tensor with known values
    pub fn create_test_tensor_2d() -> DenseTensor {
        // 3x4 tensor with specific values for predictable pruning results
        let data = vec![
            0.1, 0.9, 0.2, 0.8,  // Row 0: magnitudes [0.1, 0.9, 0.2, 0.8]
            0.3, 0.7, 0.4, 0.6,  // Row 1: magnitudes [0.3, 0.7, 0.4, 0.6]
            0.05, 0.95, 0.15, 0.85  // Row 2: magnitudes [0.05, 0.95, 0.15, 0.85]
        ];
        DenseTensor::new(data, vec![3, 4])
    }
}

fn main() -> Result<()> {
    println!("Running magnitude pruning standalone tests...");
    
    // Test 1: Basic configuration
    println!("\n=== Test 1: Basic Configuration ===");
    let config = PruningConfig::global(0.5);
    println!("Global config created: {:?}", config);
    config.validate()?;
    println!("✓ Configuration validation passed");
    
    // Test 2: Importance score computation
    println!("\n=== Test 2: Importance Score Computation ===");
    let tensor = test_utils::create_test_tensor_2d();
    println!("Test tensor: {:?}", tensor);
    let scores = MagnitudePruning::compute_importance_scores(&tensor)?;
    println!("Importance scores: {:?}", scores);
    
    let expected_scores = vec![0.1, 0.9, 0.2, 0.8, 0.3, 0.7, 0.4, 0.6, 0.05, 0.95, 0.15, 0.85];
    assert_eq!(scores, expected_scores);
    println!("✓ Importance scores computation passed");
    
    // Test 3: Global magnitude pruning
    println!("\n=== Test 3: Global Magnitude Pruning ===");
    let config = PruningConfig::global(0.5); // 50% sparsity
    let mask = MagnitudePruning::create_mask(&tensor, &config)?;
    println!("Global mask sparsity ratio: {:.2}", mask.sparsity_ratio());
    println!("Kept parameters: {}/{}", mask.kept_parameters, mask.total_parameters);
    
    let pruned = MagnitudePruning::prune(&tensor, &config)?;
    println!("Pruned tensor data: {:?}", pruned.data);
    
    let nonzero_count = pruned.data.iter().filter(|&&x| x != 0.0).count();
    println!("Non-zero count after pruning: {}", nonzero_count);
    assert_eq!(nonzero_count, 6); // Should keep 6 out of 12 (50% sparsity)
    
    // Verify the largest magnitude values are kept: [0.95, 0.9, 0.85, 0.8, 0.7, 0.6]
    let mut nonzero_values: Vec<f32> = pruned.data.iter().filter(|&&x| x != 0.0).copied().collect();
    nonzero_values.sort_by(|a, b| b.abs().total_cmp(&a.abs()));
    let expected_kept = vec![0.95, 0.9, 0.85, 0.8, 0.7, 0.6];
    assert_eq!(nonzero_values, expected_kept);
    println!("✓ Global magnitude pruning passed");
    
    // Test 4: Layer-wise magnitude pruning
    println!("\n=== Test 4: Layer-wise Magnitude Pruning ===");
    let config = PruningConfig::layerwise(0.5);
    let pruned = MagnitudePruning::prune(&tensor, &config)?;
    println!("Layer-wise pruned tensor data: {:?}", pruned.data);
    
    // With layer-wise pruning on a 2D tensor, each row should have 50% sparsity
    // Row 0: keep 2 largest [0.9, 0.8], prune [0.1, 0.2]
    // Row 1: keep 2 largest [0.7, 0.6], prune [0.3, 0.4]
    // Row 2: keep 2 largest [0.95, 0.85], prune [0.05, 0.15]
    let expected_data = vec![
        0.0, 0.9, 0.0, 0.8,
        0.0, 0.7, 0.0, 0.6,
        0.0, 0.95, 0.0, 0.85
    ];
    assert_eq!(pruned.data, expected_data);
    println!("✓ Layer-wise magnitude pruning passed");
    
    // Test 5: Pruning analysis
    println!("\n=== Test 5: Pruning Analysis ===");
    let stats = MagnitudePruning::analyze_pruning(&tensor, &pruned)?;
    println!("Pruning statistics: {:?}", stats);
    assert_eq!(stats.original_params, 12);
    assert_eq!(stats.pruned_params, 6);
    assert!((stats.sparsity_ratio - 0.5).abs() < f64::EPSILON);
    println!("✓ Pruning analysis passed");
    
    // Test 6: Mask application
    println!("\n=== Test 6: Mask Application ===");
    let tensor_for_mask = test_utils::create_test_tensor_2d();
    let mask_data = vec![
        true, false, true, false,  // Keep [0.1, _, 0.2, _]
        false, true, false, true,  // Keep [_, 0.7, _, 0.6] 
        true, true, false, false   // Keep [0.05, 0.95, _, _]
    ];
    let mask = PruningMask::new(mask_data, vec![3, 4]);
    
    let masked_tensor = mask.apply(&tensor_for_mask)?;
    let expected_masked = vec![
        0.1, 0.0, 0.2, 0.0,
        0.0, 0.7, 0.0, 0.6,
        0.05, 0.95, 0.0, 0.0
    ];
    assert_eq!(masked_tensor.data, expected_masked);
    println!("✓ Mask application passed");
    
    println!("\n🎉 All tests passed! Magnitude pruning implementation is working correctly.");
    
    Ok(())
}