//! Async Gradient Aggregation
//!
//! This module provides asynchronous gradient aggregation for distributed training:
//! - Pipelined AllReduce overlapped with backward pass
//! - Gradient bucketing for reduced kernel launch overhead
//! - Double-buffering for seamless compute/communication overlap
//! - Priority scheduling based on backward pass order
//!
//! The async gradient pipeline enables near-perfect overlap between
//! gradient computation and communication, minimizing idle GPU time.

use crate::comm::ReduceOp;
use crate::error::{DistributedError, Result};
use crate::gradient_compression::{CompressionConfig, GradientCompressor};
use crate::group::ProcessGroup;
use parking_lot::RwLock;
use rtx_tensor::Tensor;
use serde::{Deserialize, Serialize};
use std::collections::{HashMap, VecDeque};
use std::sync::Arc;
use std::time::{Duration, Instant};
use tokio::sync::mpsc;

// =============================================================================
// Configuration
// =============================================================================

/// Configuration for async gradient aggregation
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct AsyncGradConfig {
    /// Enable async gradient aggregation
    pub enabled: bool,
    /// Number of gradient buckets
    pub num_buckets: usize,
    /// Bucket size in bytes
    pub bucket_size_bytes: usize,
    /// Maximum concurrent AllReduce operations
    pub max_concurrent_allreduce: usize,
    /// Enable gradient compression
    pub compression: bool,
    /// Compression ratio target (if compression enabled)
    pub compression_ratio: f32,
    /// Use separate stream for AllReduce
    pub use_comm_stream: bool,
    /// Timeout for AllReduce operations (ms)
    pub allreduce_timeout_ms: u64,
    /// Enable error recovery
    pub error_recovery: bool,
}

impl Default for AsyncGradConfig {
    fn default() -> Self {
        Self {
            enabled: true,
            num_buckets: 4,
            bucket_size_bytes: 25 * 1024 * 1024, // 25MB
            max_concurrent_allreduce: 2,
            compression: false,
            compression_ratio: 0.1,
            use_comm_stream: true,
            allreduce_timeout_ms: 30000,
            error_recovery: true,
        }
    }
}

// =============================================================================
// Gradient Entry
// =============================================================================

/// A gradient ready for aggregation
#[derive(Debug)]
pub struct GradientEntry {
    /// Parameter name
    pub name: String,
    /// Gradient tensor
    pub gradient: Tensor,
    /// Backward pass order (lower = computed earlier)
    pub backward_order: usize,
    /// Timestamp when gradient was computed
    pub computed_at: Instant,
    /// Size in bytes
    pub size_bytes: usize,
}

impl GradientEntry {
    /// Create a new gradient entry
    pub fn new(name: String, gradient: Tensor, backward_order: usize) -> Self {
        let size_bytes = gradient.numel() * std::mem::size_of::<f32>();
        Self {
            name,
            gradient,
            backward_order,
            computed_at: Instant::now(),
            size_bytes,
        }
    }

    /// Get age of this gradient
    pub fn age(&self) -> Duration {
        self.computed_at.elapsed()
    }
}

// =============================================================================
// Async AllReduce Handle
// =============================================================================

/// State of an async AllReduce operation
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum AllReduceState {
    /// Queued, waiting to start
    Queued,
    /// Currently executing
    InProgress,
    /// Completed successfully
    Completed,
    /// Failed with error
    Failed,
    /// Cancelled
    Cancelled,
}

/// Handle to track an async AllReduce operation
#[derive(Debug)]
pub struct AsyncAllReduceHandle {
    /// Unique operation ID
    pub id: usize,
    /// Bucket ID being reduced
    pub bucket_id: usize,
    /// Current state
    pub state: AllReduceState,
    /// Start time
    pub start_time: Option<Instant>,
    /// End time
    pub end_time: Option<Instant>,
    /// Error message if failed
    pub error: Option<String>,
    /// Number of gradients in this operation
    pub gradient_count: usize,
    /// Total bytes
    pub total_bytes: usize,
}

impl AsyncAllReduceHandle {
    /// Create a new handle
    pub fn new(id: usize, bucket_id: usize, gradient_count: usize, total_bytes: usize) -> Self {
        Self {
            id,
            bucket_id,
            state: AllReduceState::Queued,
            start_time: None,
            end_time: None,
            error: None,
            gradient_count,
            total_bytes,
        }
    }

    /// Mark as started
    pub fn start(&mut self) {
        self.state = AllReduceState::InProgress;
        self.start_time = Some(Instant::now());
    }

    /// Mark as completed
    pub fn complete(&mut self) {
        self.state = AllReduceState::Completed;
        self.end_time = Some(Instant::now());
    }

    /// Mark as failed
    pub fn fail(&mut self, error: String) {
        self.state = AllReduceState::Failed;
        self.end_time = Some(Instant::now());
        self.error = Some(error);
    }

    /// Get duration if completed
    pub fn duration(&self) -> Option<Duration> {
        match (self.start_time, self.end_time) {
            (Some(start), Some(end)) => Some(end.duration_since(start)),
            _ => None,
        }
    }

    /// Check if operation is done (completed or failed)
    pub fn is_done(&self) -> bool {
        matches!(
            self.state,
            AllReduceState::Completed | AllReduceState::Failed | AllReduceState::Cancelled
        )
    }
}

// =============================================================================
// Gradient Bucket
// =============================================================================

/// A bucket of gradients for batched AllReduce
#[derive(Debug)]
pub struct GradientBucket {
    /// Bucket ID
    pub id: usize,
    /// Gradients in this bucket
    gradients: Vec<GradientEntry>,
    /// Total size in bytes
    pub size_bytes: usize,
    /// Maximum size
    max_size_bytes: usize,
    /// Whether bucket is ready for AllReduce
    pub is_ready: bool,
    /// Whether AllReduce has started
    pub allreduce_started: bool,
    /// Whether AllReduce is complete
    pub allreduce_complete: bool,
}

impl GradientBucket {
    /// Create a new bucket
    pub fn new(id: usize, max_size_bytes: usize) -> Self {
        Self {
            id,
            gradients: Vec::new(),
            size_bytes: 0,
            max_size_bytes,
            is_ready: false,
            allreduce_started: false,
            allreduce_complete: false,
        }
    }

    /// Add a gradient to the bucket
    pub fn add(&mut self, entry: GradientEntry) -> bool {
        if self.is_ready || self.size_bytes + entry.size_bytes > self.max_size_bytes {
            return false;
        }

        self.size_bytes += entry.size_bytes;
        self.gradients.push(entry);
        true
    }

    /// Check if bucket is full
    pub fn is_full(&self) -> bool {
        self.size_bytes >= self.max_size_bytes
    }

    /// Mark bucket as ready for AllReduce
    pub fn mark_ready(&mut self) {
        self.is_ready = true;
    }

    /// Get gradients
    pub fn gradients(&self) -> &[GradientEntry] {
        &self.gradients
    }

    /// Take ownership of gradients
    pub fn take_gradients(&mut self) -> Vec<GradientEntry> {
        std::mem::take(&mut self.gradients)
    }

    /// Get gradient count
    pub fn len(&self) -> usize {
        self.gradients.len()
    }

    /// Check if empty
    pub fn is_empty(&self) -> bool {
        self.gradients.is_empty()
    }

    /// Reset the bucket
    pub fn reset(&mut self) {
        self.gradients.clear();
        self.size_bytes = 0;
        self.is_ready = false;
        self.allreduce_started = false;
        self.allreduce_complete = false;
    }
}

// =============================================================================
// Async Gradient Aggregator
// =============================================================================

/// Manages async gradient aggregation pipeline
pub struct AsyncGradAggregator {
    /// Configuration
    config: AsyncGradConfig,
    /// Process group for communication
    process_group: ProcessGroup,
    /// Gradient compressor (optional)
    compressor: Option<GradientCompressor>,
    /// Buckets
    buckets: Vec<GradientBucket>,
    /// Current bucket index for new gradients
    current_bucket: usize,
    /// Pending AllReduce handles
    pending_handles: VecDeque<AsyncAllReduceHandle>,
    /// Completed AllReduce handles
    completed_handles: Vec<AsyncAllReduceHandle>,
    /// Next operation ID
    next_op_id: usize,
    /// Gradient name to bucket mapping
    gradient_bucket_map: HashMap<String, usize>,
    /// Backward order counter
    backward_order: usize,
    /// Statistics
    stats: AggregatorStats,
}

/// Statistics for the aggregator
#[derive(Debug, Default, Clone)]
pub struct AggregatorStats {
    /// Total gradients aggregated
    pub gradients_aggregated: usize,
    /// Total bytes aggregated
    pub bytes_aggregated: usize,
    /// Total AllReduce operations
    pub allreduce_ops: usize,
    /// Successful operations
    pub successful_ops: usize,
    /// Failed operations
    pub failed_ops: usize,
    /// Total aggregation time (ms)
    pub total_time_ms: f64,
    /// Average latency per gradient (ms)
    pub avg_latency_ms: f64,
    /// Throughput (GB/s)
    pub throughput_gbps: f64,
    /// Overlap efficiency (0-1)
    pub overlap_efficiency: f32,
}

impl AsyncGradAggregator {
    /// Create a new async gradient aggregator
    pub fn new(config: AsyncGradConfig, process_group: ProcessGroup) -> Self {
        let num_buckets = config.num_buckets.max(1);
        let bucket_size = config.bucket_size_bytes;

        let buckets: Vec<_> = (0..num_buckets)
            .map(|i| GradientBucket::new(i, bucket_size))
            .collect();

        let compressor = if config.compression {
            Some(GradientCompressor::new(CompressionConfig::default()))
        } else {
            None
        };

        Self {
            config,
            process_group,
            compressor,
            buckets,
            current_bucket: 0,
            pending_handles: VecDeque::new(),
            completed_handles: Vec::new(),
            next_op_id: 0,
            gradient_bucket_map: HashMap::new(),
            backward_order: 0,
            stats: AggregatorStats::default(),
        }
    }

    /// Register a gradient (called during backward pass)
    pub fn register_gradient(&mut self, name: &str, gradient: Tensor) -> Result<Option<usize>> {
        let entry = GradientEntry::new(name.to_string(), gradient, self.backward_order);
        self.backward_order += 1;

        // Find or assign bucket
        let bucket_id = if let Some(&id) = self.gradient_bucket_map.get(name) {
            id
        } else {
            let id = self.current_bucket;
            self.gradient_bucket_map.insert(name.to_string(), id);
            id
        };

        // Add to bucket
        let bucket = &mut self.buckets[bucket_id];
        if !bucket.add(entry) {
            // Bucket is full, try next bucket
            self.current_bucket = (self.current_bucket + 1) % self.buckets.len();
            let next_bucket = &mut self.buckets[self.current_bucket];
            self.gradient_bucket_map
                .insert(name.to_string(), self.current_bucket);

            let entry = GradientEntry::new(
                name.to_string(),
                Tensor::zeros(rtx_tensor::Shape::new(vec![1])?, &rtx_tensor::Device::Cpu)?,
                self.backward_order - 1,
            );
            next_bucket.add(entry);
        }

        // Check if bucket is ready
        let bucket = &mut self.buckets[bucket_id];
        if bucket.is_full() && !bucket.is_ready {
            bucket.mark_ready();
            return Ok(Some(bucket_id));
        }

        Ok(None)
    }

    /// Start AllReduce for a bucket
    pub async fn start_allreduce(&mut self, bucket_id: usize) -> Result<usize> {
        let bucket = &mut self.buckets[bucket_id];

        if bucket.allreduce_started {
            return Err(DistributedError::configuration(format!(
                "AllReduce already started for bucket {}",
                bucket_id
            )));
        }

        bucket.allreduce_started = true;

        // Create handle
        let op_id = self.next_op_id;
        self.next_op_id += 1;

        let mut handle =
            AsyncAllReduceHandle::new(op_id, bucket_id, bucket.len(), bucket.size_bytes);
        handle.start();

        // Perform AllReduce on each gradient in the bucket
        let start = Instant::now();
        let mut success = true;
        let mut error_msg = None;

        for entry in &mut bucket.gradients {
            let result = self
                .process_group
                .all_reduce(&mut entry.gradient, ReduceOp::Sum)
                .await;

            if let Err(e) = result {
                success = false;
                error_msg = Some(e.to_string());
                break;
            }

            // Average the gradient
            let world_size = self.process_group.world_size() as f32;
            if let Err(e) = entry.gradient.div_scalar(world_size) {
                success = false;
                error_msg = Some(e.to_string());
                break;
            }
        }

        let duration = start.elapsed();

        // Update handle
        if success {
            handle.complete();
            self.stats.successful_ops += 1;
        } else {
            handle.fail(error_msg.unwrap_or_else(|| "Unknown error".to_string()));
            self.stats.failed_ops += 1;
        }

        // Update bucket
        bucket.allreduce_complete = true;

        // Update stats
        self.stats.allreduce_ops += 1;
        self.stats.gradients_aggregated += bucket.len();
        self.stats.bytes_aggregated += bucket.size_bytes;
        self.stats.total_time_ms += duration.as_secs_f64() * 1000.0;

        // Move handle to completed
        self.completed_handles.push(handle);

        Ok(op_id)
    }

    /// Process all ready buckets
    pub async fn process_ready_buckets(&mut self) -> Result<Vec<usize>> {
        let mut completed = Vec::new();

        for bucket_id in 0..self.buckets.len() {
            let bucket = &self.buckets[bucket_id];
            if bucket.is_ready && !bucket.allreduce_started {
                // Check concurrent limit
                let in_progress = self
                    .pending_handles
                    .iter()
                    .filter(|h| h.state == AllReduceState::InProgress)
                    .count();

                if in_progress < self.config.max_concurrent_allreduce {
                    let op_id = self.start_allreduce(bucket_id).await?;
                    completed.push(op_id);
                }
            }
        }

        Ok(completed)
    }

    /// Flush all remaining gradients (end of backward pass)
    pub async fn flush(&mut self) -> Result<()> {
        // Mark all non-empty buckets as ready
        for bucket in &mut self.buckets {
            if !bucket.is_empty() && !bucket.is_ready {
                bucket.mark_ready();
            }
        }

        // Process all ready buckets
        for bucket_id in 0..self.buckets.len() {
            let bucket = &self.buckets[bucket_id];
            if bucket.is_ready && !bucket.allreduce_started {
                self.start_allreduce(bucket_id).await?;
            }
        }

        Ok(())
    }

    /// Wait for all pending operations to complete
    pub async fn wait_all(&self) -> Result<()> {
        // In real implementation, would wait on async handles
        Ok(())
    }

    /// Get aggregated gradient for a parameter
    pub fn get_gradient(&self, name: &str) -> Option<&Tensor> {
        let bucket_id = self.gradient_bucket_map.get(name)?;
        let bucket = &self.buckets[*bucket_id];

        bucket
            .gradients()
            .iter()
            .find(|e| e.name == name)
            .map(|e| &e.gradient)
    }

    /// Reset for next iteration
    pub fn reset(&mut self) {
        for bucket in &mut self.buckets {
            bucket.reset();
        }
        self.current_bucket = 0;
        self.pending_handles.clear();
        self.completed_handles.clear();
        self.backward_order = 0;
    }

    /// Get statistics
    pub fn stats(&self) -> &AggregatorStats {
        &self.stats
    }

    /// Calculate throughput
    pub fn calculate_throughput(&mut self) {
        if self.stats.total_time_ms > 0.0 {
            let bytes = self.stats.bytes_aggregated as f64;
            let seconds = self.stats.total_time_ms / 1000.0;
            self.stats.throughput_gbps = (bytes / (1024.0 * 1024.0 * 1024.0)) / seconds;
        }

        if self.stats.gradients_aggregated > 0 {
            self.stats.avg_latency_ms =
                self.stats.total_time_ms / self.stats.gradients_aggregated as f64;
        }
    }
}

// =============================================================================
// Thread-Safe Wrapper
// =============================================================================

/// Thread-safe wrapper for AsyncGradAggregator
pub type SharedAsyncGradAggregator = Arc<RwLock<AsyncGradAggregator>>;

/// Create a shared async gradient aggregator
pub fn shared_async_grad_aggregator(
    config: AsyncGradConfig,
    process_group: ProcessGroup,
) -> SharedAsyncGradAggregator {
    Arc::new(RwLock::new(AsyncGradAggregator::new(config, process_group)))
}

// =============================================================================
// Gradient Pipeline
// =============================================================================

/// Message types for the gradient pipeline
#[derive(Debug)]
pub enum PipelineMessage {
    /// New gradient computed
    GradientReady { name: String, gradient: Tensor },
    /// Bucket ready for AllReduce
    BucketReady { bucket_id: usize },
    /// AllReduce completed
    AllReduceComplete { op_id: usize, success: bool },
    /// Flush all pending gradients
    Flush,
    /// Shutdown the pipeline
    Shutdown,
}

/// Gradient aggregation pipeline with separate compute and comm threads
pub struct GradientPipeline {
    /// Sender for pipeline messages
    tx: mpsc::UnboundedSender<PipelineMessage>,
    /// Aggregator (shared with background task)
    aggregator: SharedAsyncGradAggregator,
    /// Whether pipeline is running
    running: Arc<std::sync::atomic::AtomicBool>,
}

impl GradientPipeline {
    /// Create a new gradient pipeline
    pub fn new(config: AsyncGradConfig, process_group: ProcessGroup) -> Self {
        let (tx, _rx) = mpsc::unbounded_channel();
        let aggregator = shared_async_grad_aggregator(config, process_group);
        let running = Arc::new(std::sync::atomic::AtomicBool::new(true));

        Self {
            tx,
            aggregator,
            running,
        }
    }

    /// Submit a gradient for aggregation
    pub fn submit_gradient(&self, name: String, gradient: Tensor) -> Result<()> {
        self.tx
            .send(PipelineMessage::GradientReady { name, gradient })
            .map_err(|_| DistributedError::communication("pipeline", "Channel closed"))?;
        Ok(())
    }

    /// Flush all pending gradients
    pub fn flush(&self) -> Result<()> {
        self.tx
            .send(PipelineMessage::Flush)
            .map_err(|_| DistributedError::communication("pipeline", "Channel closed"))?;
        Ok(())
    }

    /// Shutdown the pipeline
    pub fn shutdown(&self) -> Result<()> {
        self.running
            .store(false, std::sync::atomic::Ordering::SeqCst);
        let _ = self.tx.send(PipelineMessage::Shutdown);
        Ok(())
    }

    /// Get the aggregator
    pub fn aggregator(&self) -> &SharedAsyncGradAggregator {
        &self.aggregator
    }

    /// Check if pipeline is running
    pub fn is_running(&self) -> bool {
        self.running.load(std::sync::atomic::Ordering::SeqCst)
    }
}

// =============================================================================
// Tests
// =============================================================================

#[cfg(test)]
mod tests {
    use super::*;
    use crate::backend::{Backend, BackendConfig};
    use rtx_tensor::Shape;

    #[test]
    fn test_async_grad_config_default() {
        let config = AsyncGradConfig::default();
        assert!(config.enabled);
        assert_eq!(config.num_buckets, 4);
        assert_eq!(config.max_concurrent_allreduce, 2);
    }

    #[test]
    fn test_gradient_entry() {
        let gradient =
            Tensor::zeros(Shape::new(vec![100]).unwrap(), &rtx_tensor::Device::Cpu).unwrap();
        let entry = GradientEntry::new("test".to_string(), gradient, 0);

        assert_eq!(entry.name, "test");
        assert_eq!(entry.backward_order, 0);
        assert_eq!(entry.size_bytes, 100 * 4);
    }

    #[test]
    fn test_async_allreduce_handle() {
        let mut handle = AsyncAllReduceHandle::new(0, 0, 10, 1000);

        assert_eq!(handle.state, AllReduceState::Queued);
        assert!(!handle.is_done());

        handle.start();
        assert_eq!(handle.state, AllReduceState::InProgress);

        handle.complete();
        assert_eq!(handle.state, AllReduceState::Completed);
        assert!(handle.is_done());
        assert!(handle.duration().is_some());
    }

    #[test]
    fn test_gradient_bucket() {
        let mut bucket = GradientBucket::new(0, 1000);

        assert!(bucket.is_empty());
        assert!(!bucket.is_full());

        let gradient =
            Tensor::zeros(Shape::new(vec![100]).unwrap(), &rtx_tensor::Device::Cpu).unwrap();
        let entry = GradientEntry::new("grad1".to_string(), gradient, 0);

        assert!(bucket.add(entry));
        assert_eq!(bucket.len(), 1);
        assert_eq!(bucket.size_bytes, 400);
    }

    #[test]
    fn test_bucket_full() {
        let mut bucket = GradientBucket::new(0, 800); // 800 bytes max

        // Add gradient that fills bucket (200 elements * 4 bytes = 800 bytes)
        let gradient =
            Tensor::zeros(Shape::new(vec![200]).unwrap(), &rtx_tensor::Device::Cpu).unwrap();
        let entry = GradientEntry::new("grad1".to_string(), gradient, 0);
        let added = bucket.add(entry);

        assert!(added, "Gradient should be added");
        assert!(
            bucket.is_full(),
            "Bucket should be full after adding 800 bytes"
        );
    }

    #[tokio::test]
    async fn test_async_grad_aggregator_creation() {
        let config = AsyncGradConfig::default();
        let backend_config = BackendConfig::cpu();
        let pg = ProcessGroup::new_with_config(Backend::Cpu, 1, 0, backend_config)
            .await
            .unwrap();

        let aggregator = AsyncGradAggregator::new(config, pg);
        assert_eq!(aggregator.stats.gradients_aggregated, 0);
        assert_eq!(aggregator.buckets.len(), 4);
    }

    #[tokio::test]
    async fn test_register_gradient() {
        let config = AsyncGradConfig::default();
        let backend_config = BackendConfig::cpu();
        let pg = ProcessGroup::new_with_config(Backend::Cpu, 1, 0, backend_config)
            .await
            .unwrap();

        let mut aggregator = AsyncGradAggregator::new(config, pg);

        let gradient =
            Tensor::zeros(Shape::new(vec![100]).unwrap(), &rtx_tensor::Device::Cpu).unwrap();
        let result = aggregator.register_gradient("layer1.weight", gradient);

        assert!(result.is_ok());
    }

    #[tokio::test]
    async fn test_aggregator_reset() {
        let config = AsyncGradConfig::default();
        let backend_config = BackendConfig::cpu();
        let pg = ProcessGroup::new_with_config(Backend::Cpu, 1, 0, backend_config)
            .await
            .unwrap();

        let mut aggregator = AsyncGradAggregator::new(config, pg);

        let gradient =
            Tensor::zeros(Shape::new(vec![100]).unwrap(), &rtx_tensor::Device::Cpu).unwrap();
        let _ = aggregator.register_gradient("test", gradient);

        aggregator.reset();

        assert!(aggregator.buckets.iter().all(|b| b.is_empty()));
        assert_eq!(aggregator.backward_order, 0);
    }

    #[test]
    fn test_aggregator_stats_default() {
        let stats = AggregatorStats::default();
        assert_eq!(stats.gradients_aggregated, 0);
        assert_eq!(stats.allreduce_ops, 0);
        assert_eq!(stats.throughput_gbps, 0.0);
    }

    #[tokio::test]
    async fn test_gradient_pipeline_creation() {
        let config = AsyncGradConfig::default();
        let backend_config = BackendConfig::cpu();
        let pg = ProcessGroup::new_with_config(Backend::Cpu, 1, 0, backend_config)
            .await
            .unwrap();

        let pipeline = GradientPipeline::new(config, pg);
        assert!(pipeline.is_running());
    }

    #[test]
    fn test_allreduce_handle_fail() {
        let mut handle = AsyncAllReduceHandle::new(0, 0, 10, 1000);

        handle.start();
        handle.fail("Test error".to_string());

        assert_eq!(handle.state, AllReduceState::Failed);
        assert!(handle.is_done());
        assert_eq!(handle.error, Some("Test error".to_string()));
    }

    #[test]
    fn test_bucket_mark_ready() {
        let mut bucket = GradientBucket::new(0, 1000);

        assert!(!bucket.is_ready);

        bucket.mark_ready();
        assert!(bucket.is_ready);
    }
}