# AUTOGENERATED! DO NOT EDIT! File to edit: ../../modules/source/11_kernels/kernels_dev.ipynb.

# %% auto 0
__all__ = ['time_kernel', 'matmul_baseline', 'vectorized_relu', 'vectorized_operations', 'cache_friendly_matmul', 'parallel_relu',
           'parallel_batch_processing', 'quantized_matmul', 'quantized_relu']

# %% ../../modules/source/11_kernels/kernels_dev.ipynb 1
import numpy as np
import sys
import os
import time
import tracemalloc
import psutil
from typing import Callable, Dict, Any, Optional, Tuple, List
from functools import wraps
from pathlib import Path

# Import our existing components
try:
    from tinytorch.core.tensor import Tensor
    from tinytorch.core.layers import matmul_naive as matmul
    from tinytorch.core.activations import ReLU, Sigmoid, Tanh
    from tinytorch.core.cnn import Conv2D
except ImportError:
    # For development, import from local modules
    base_dir = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
    sys.path.extend([
        os.path.join(base_dir, '01_tensor'),
        os.path.join(base_dir, '02_activations'),
        os.path.join(base_dir, '03_layers'),
        os.path.join(base_dir, '05_cnn'),
        os.path.join(base_dir, 'utils')
    ])
    
    try:
        from tensor_dev import Tensor
        from layers_dev import matmul_naive as matmul
        from activations_dev import ReLU, Sigmoid, Tanh
        from cnn_dev import Conv2D
    except ImportError:
        # Create minimal mock for development
        class Tensor:
            def __init__(self, data):
                self.data = np.array(data)
                self.shape = self.data.shape
            def __str__(self):
                return f"Tensor({self.data})"

# Simple timing utility for kernel performance measurement
def time_kernel(func, *args, **kwargs):
    """
    Simple timing function for measuring kernel performance.
    
    Returns:
        tuple: (result, time_in_microseconds)
    """
    start = time.perf_counter()
    result = func(*args, **kwargs)
    end = time.perf_counter()
    microseconds = (end - start) * 1_000_000
    return result, microseconds

# %% ../../modules/source/11_kernels/kernels_dev.ipynb 6
def matmul_baseline(A: Tensor, B: Tensor) -> Tensor:
    """
    Baseline matrix multiplication using TinyTorch's proven implementation.
    
    This function demonstrates how to build on existing TinyTorch components
    rather than reinventing the wheel. We use the standard matmul from Module 03
    as our baseline for comparison with optimized kernels.
    
    This is NOT a custom implementation - it's the standard TinyTorch matmul
    wrapped for use in kernel comparisons and benchmarking.
    
    TODO: Use TinyTorch's standard matmul implementation as a baseline.
    
    STEP-BY-STEP IMPLEMENTATION:
    1. Import the standard matmul function from tinytorch.core.layers
    2. Extract numpy arrays from input Tensors
    3. Use the proven implementation from TinyTorch
    4. Wrap result back in Tensor format
    5. Return the result
    
    CODE REUSE PRINCIPLES:
    1. Always use the packaged version for reliability
    2. Don't duplicate working code - reference the source
    3. Use descriptive names that indicate what the function actually does
    4. Keep dependencies simple and reliable
    
    EXAMPLE USAGE:
    ```python
    A = Tensor([[1, 2], [3, 4]])
    B = Tensor([[5, 6], [7, 8]])
    C = matmul_baseline(A, B)
    # Expected: [[19, 22], [43, 50]]
    ```
    
    LEARNING CONNECTIONS:
    - This shows how to use TinyTorch as a library
    - Demonstrates reliable dependency management
    - Serves as baseline for kernel performance comparisons
    - Shows proper software engineering practices
    """
    ### BEGIN SOLUTION
    # Extract numpy arrays from Tensors
    A_data = A.data if hasattr(A, 'data') else A
    B_data = B.data if hasattr(B, 'data') else B
    
    # Use NumPy's matrix multiplication as our baseline
    # This is our baseline - reliable, tested, and consistent
    result_data = np.dot(A_data, B_data)
    
    # Wrap the result back in a Tensor for consistency
    result = Tensor(result_data)
    
    return result
    ### END SOLUTION

# %% ../../modules/source/11_kernels/kernels_dev.ipynb 9
def vectorized_relu(x: Tensor) -> Tensor:
    """
    Vectorized ReLU implementation demonstrating SIMD principles.
    
    This function shows how to write operations that take advantage of
    CPU vectorization capabilities for better performance.
    
    TODO: Implement a vectorized ReLU that's optimized for performance.
    
    STEP-BY-STEP IMPLEMENTATION:
    1. Extract numpy array from Tensor
    2. Use NumPy's vectorized operations (these compile to SIMD instructions)
    3. Apply ReLU: f(x) = max(0, x) for all elements simultaneously
    4. Return result as Tensor
    
    VECTORIZATION TECHNIQUES:
    1. Use np.maximum instead of loops - this is vectorized
    2. Ensure input is contiguous in memory for better SIMD performance
    3. Consider using specific dtypes (float32 vs float64) for SIMD alignment
    4. Avoid conditional operations that break vectorization
    
    EXAMPLE USAGE:
    ```python
    x = Tensor([-2, -1, 0, 1, 2])
    y = vectorized_relu(x)
    # Expected: [0, 0, 0, 1, 2]
    ```
    
    PERFORMANCE CONSIDERATIONS:
    - np.maximum is vectorized and uses SIMD instructions
    - Memory layout matters: contiguous arrays are faster
    - Data type matters: float32 allows more SIMD parallelism than float64
    - Avoid Python loops - they can't be vectorized
    
    LEARNING CONNECTIONS:
    - This is how PyTorch's ReLU is implemented under the hood
    - GPU kernels use similar principles with thousands of parallel threads
    - Modern CPUs can process 4-16 floats simultaneously with SIMD
    """
    ### BEGIN SOLUTION
    # Extract numpy array
    x_data = x.data if hasattr(x, 'data') else x
    
    # Ensure contiguous memory layout for better SIMD performance
    if not x_data.flags.c_contiguous:
        x_data = np.ascontiguousarray(x_data)
    
    # Vectorized ReLU using NumPy's maximum function
    # This compiles to SIMD instructions on modern CPUs
    result = np.maximum(0, x_data)
    
    return Tensor(result)
    ### END SOLUTION

# %% ../../modules/source/11_kernels/kernels_dev.ipynb 10
def vectorized_operations(x: Tensor, y: Tensor) -> Dict[str, Tensor]:
    """
    Demonstration of various vectorized operations.
    
    Shows how multiple operations can be vectorized for better performance.
    
    TODO: Implement a collection of vectorized operations.
    
    STEP-BY-STEP IMPLEMENTATION:
    1. Extract numpy arrays from input Tensors
    2. Implement vectorized versions of common operations
    3. Use NumPy's built-in vectorized functions
    4. Return dictionary of results
    
    OPERATIONS TO IMPLEMENT:
    - element_wise_multiply: x * y (element-wise)
    - element_wise_add: x + y (element-wise)
    - squared_difference: (x - y)^2
    - euclidean_distance: sqrt(sum((x - y)^2))
    - dot_product: sum(x * y)
    
    VECTORIZATION PRINCIPLES:
    - Use NumPy operations instead of Python loops
    - Combine operations when possible: (x - y)**2 instead of subtract then square
    - Consider memory layout and data types
    - Measure performance improvements
    
    EXAMPLE USAGE:
    ```python
    x = Tensor([1, 2, 3, 4])
    y = Tensor([2, 3, 4, 5])
    results = vectorized_operations(x, y)
    # Returns dict with all vectorized operation results
    ```
    """
    ### BEGIN SOLUTION
    # Extract numpy arrays
    x_data = x.data if hasattr(x, 'data') else x
    y_data = y.data if hasattr(y, 'data') else y
    
    # Ensure arrays are the same shape for element-wise operations
    assert x_data.shape == y_data.shape, f"Shape mismatch: {x_data.shape} vs {y_data.shape}"
    
    # Vectorized operations
    results = {
        'element_wise_multiply': Tensor(x_data * y_data),
        'element_wise_add': Tensor(x_data + y_data),
        'squared_difference': Tensor((x_data - y_data) ** 2),
        'euclidean_distance': Tensor(np.sqrt(np.sum((x_data - y_data) ** 2))),
        'dot_product': Tensor(np.dot(x_data.flatten(), y_data.flatten()))
    }
    
    return results
    ### END SOLUTION

# %% ../../modules/source/11_kernels/kernels_dev.ipynb 13
def cache_friendly_matmul(A: Tensor, B: Tensor, block_size: int = 32) -> Tensor:
    """
    Cache-friendly matrix multiplication using blocking technique.
    
    This implementation uses cache blocking to improve memory access patterns
    and achieve better performance on modern CPUs.
    
    TODO: Implement cache-friendly matrix multiplication using blocking.
    
    STEP-BY-STEP IMPLEMENTATION:
    1. Extract numpy arrays and get dimensions
    2. Pre-allocate output matrix
    3. Use three nested loops for blocks: block_i, block_j, block_k
    4. Within each block, use three nested loops for elements: i, j, k
    5. Process data in cache-sized blocks for better locality
    
    BLOCKING ALGORITHM:
    1. Divide matrices into blocks of size block_size x block_size
    2. For each block of C, compute contribution from corresponding A and B blocks
    3. This keeps data in cache longer, reducing memory access time
    
    CACHE OPTIMIZATION PRINCIPLES:
    - Process data in small blocks that fit in cache
    - Reuse data as much as possible while it's in cache
    - Access memory in predictable patterns
    - Minimize cache misses
    
    EXAMPLE USAGE:
    ```python
    A = Tensor([[1, 2], [3, 4]])
    B = Tensor([[5, 6], [7, 8]])
    C = cache_friendly_matmul(A, B, block_size=2)
    # Expected: [[19, 22], [43, 50]]
    ```
    
    PERFORMANCE HINTS:
    - block_size should be chosen based on cache size
    - Typical L1 cache: 32KB, so block_size=32 for float32 matrices
    - Experiment with different block sizes for your hardware
    - This algorithm is O(n^3) but with much better constants
    
    LEARNING CONNECTIONS:
    - This is how BLAS libraries achieve high performance
    - GPUs use similar tiling strategies for shared memory
    - Modern compilers can sometimes do this automatically
    """
    ### BEGIN SOLUTION
    # Extract numpy arrays
    A_data = A.data if hasattr(A, 'data') else A
    B_data = B.data if hasattr(B, 'data') else B
    
    # Get dimensions
    m, k = A_data.shape
    k2, n = B_data.shape
    assert k == k2, f"Cannot multiply {A_data.shape} and {B_data.shape}"
    
    # Pre-allocate output matrix
    C = np.zeros((m, n), dtype=A_data.dtype)
    
    # Cache-friendly blocked matrix multiplication
    for block_i in range(0, m, block_size):
        for block_j in range(0, n, block_size):
            for block_k in range(0, k, block_size):
                # Define block boundaries
                end_i = min(block_i + block_size, m)
                end_j = min(block_j + block_size, n)
                end_k = min(block_k + block_size, k)
                
                # Process block - good cache locality
                for i in range(block_i, end_i):
                    for j in range(block_j, end_j):
                        for k_idx in range(block_k, end_k):
                            C[i, j] += A_data[i, k_idx] * B_data[k_idx, j]
    
    return Tensor(C)
    ### END SOLUTION

# %% ../../modules/source/11_kernels/kernels_dev.ipynb 16
def parallel_relu(x: Tensor, num_workers: int = 4) -> Tensor:
    """
    Parallel ReLU implementation using multiple CPU cores.
    
    This function demonstrates data parallelism by splitting the input
    across multiple worker processes.
    
    TODO: Implement parallel ReLU using multiprocessing or threading.
    
    STEP-BY-STEP IMPLEMENTATION:
    1. Extract numpy array from Tensor
    2. Split array into chunks for parallel processing
    3. Define worker function that applies ReLU to a chunk
    4. Use ThreadPoolExecutor to process chunks in parallel
    5. Combine results from all workers
    6. Return result as Tensor
    
    PARALLELIZATION STRATEGY:
    1. Split input into num_workers chunks
    2. Each worker processes its chunk independently
    3. Apply ReLU: max(0, x) to each chunk
    4. Combine results preserving original order
    
    EXAMPLE USAGE:
    ```python
    x = Tensor(np.random.randn(1000))
    y = parallel_relu(x, num_workers=4)
    # Processes data using 4 parallel workers
    ```
    
    PERFORMANCE CONSIDERATIONS:
    - Overhead of parallel processing may not be worth it for small arrays
    - Threading vs multiprocessing trade-offs
    - Chunk size should be large enough to amortize overhead
    - Consider memory bandwidth limitations
    
    LEARNING CONNECTIONS:
    - This is how PyTorch processes batches in parallel
    - GPUs naturally do this with thousands of parallel threads
    - Modern deep learning frameworks heavily use parallelism
    """
    ### BEGIN SOLUTION
    from concurrent.futures import ThreadPoolExecutor
    
    # Extract numpy array
    x_data = x.data if hasattr(x, 'data') else x
    
    # For small arrays, parallel processing isn't worth the overhead
    if x_data.size < 1000:
        return Tensor(np.maximum(0, x_data))
    
    # Split array into chunks
    chunk_size = max(1, x_data.size // num_workers)
    chunks = []
    flat_data = x_data.flatten()
    
    for i in range(0, len(flat_data), chunk_size):
        chunks.append(flat_data[i:i + chunk_size])
    
    # Worker function
    def relu_chunk(chunk):
        return np.maximum(0, chunk)
    
    # Process chunks in parallel
    with ThreadPoolExecutor(max_workers=num_workers) as executor:
        future_to_chunk = {executor.submit(relu_chunk, chunk): i for i, chunk in enumerate(chunks)}
        results = [None] * len(chunks)
        
        for future in future_to_chunk:
            chunk_idx = future_to_chunk[future]
            results[chunk_idx] = future.result()
    
    # Combine results
    combined_result = np.concatenate(results)
    
    # Reshape back to original shape
    result = combined_result.reshape(x_data.shape)
    
    return Tensor(result)
    ### END SOLUTION

# %% ../../modules/source/11_kernels/kernels_dev.ipynb 17
def parallel_batch_processing(batch_data: List[Tensor], operation: Callable, num_workers: int = 4) -> List[Tensor]:
    """
    Process a batch of tensors in parallel using multiple workers.
    
    This function demonstrates how to parallelize operations across
    multiple data samples, similar to how modern ML frameworks work.
    
    TODO: Implement parallel batch processing.
    
    STEP-BY-STEP IMPLEMENTATION:
    1. Take a list of Tensors and an operation function
    2. Use ThreadPoolExecutor to process multiple tensors simultaneously
    3. Apply the operation to each tensor in parallel
    4. Return list of results in original order
    
    PARALLELIZATION STRATEGY:
    1. Each worker processes one tensor at a time
    2. Multiple workers can process different tensors simultaneously
    3. Preserve order of results to match input order
    
    EXAMPLE USAGE:
    ```python
    batch = [Tensor(np.random.randn(100, 100)) for _ in range(8)]
    relu_op = lambda x: vectorized_relu(x)
    results = parallel_batch_processing(batch, relu_op, num_workers=4)
    # Processes 8 tensors using 4 parallel workers
    ```
    
    PERFORMANCE CONSIDERATIONS:
    - Each tensor should be large enough to justify parallel overhead
    - Balance number of workers with available CPU cores
    - Consider memory usage with multiple workers
    - Thread vs process pool trade-offs
    
    LEARNING CONNECTIONS:
    - This is how PyTorch's DataLoader processes batches
    - Similar to how GPUs process multiple samples simultaneously
    - Foundation for distributed training across multiple nodes
    """
    ### BEGIN SOLUTION
    from concurrent.futures import ThreadPoolExecutor
    
    # For small batches, parallel processing might not be worth it
    if len(batch_data) < num_workers:
        return [operation(tensor) for tensor in batch_data]
    
    # Process batch in parallel
    with ThreadPoolExecutor(max_workers=num_workers) as executor:
        # Submit all tasks
        future_to_index = {executor.submit(operation, tensor): i for i, tensor in enumerate(batch_data)}
        
        # Collect results in original order
        results = [None] * len(batch_data)
        for future in future_to_index:
            index = future_to_index[future]
            results[index] = future.result()
    
    return results
    ### END SOLUTION

# %% ../../modules/source/11_kernels/kernels_dev.ipynb 22
def quantized_matmul(A: Tensor, B: Tensor, scale_A: float = 1.0, scale_B: float = 1.0) -> Tensor:
    """
    Quantized matrix multiplication kernel for compressed models.
    
    This function demonstrates how to perform matrix multiplication
    with quantized (int8) weights while maintaining numerical accuracy.
    
    TODO: Implement quantized matrix multiplication.
    
    STEP-BY-STEP IMPLEMENTATION:
    1. Extract numpy arrays from Tensors
    2. Quantize inputs to int8 using provided scales
    3. Perform integer matrix multiplication
    4. Rescale result back to appropriate range
    5. Return result as Tensor
    
    QUANTIZATION PROCESS:
    1. Quantize: int8_value = round(float_value / scale)
    2. Compute: int8_result = int8_A @ int8_B
    3. Rescale: float_result = int8_result * scale_A * scale_B
    
    EXAMPLE USAGE:
    ```python
    A = Tensor([[1.0, 2.0], [3.0, 4.0]])
    B = Tensor([[0.5, 1.5], [2.5, 3.5]])
    C = quantized_matmul(A, B, scale_A=1.0/127, scale_B=1.0/127)
    # Should approximate regular matrix multiplication
    ```
    
    PERFORMANCE CONSIDERATIONS:
    - int8 operations are often faster than float32
    - Memory usage is 4x lower
    - Accumulation in int32 to prevent overflow
    - Careful handling of scales to maintain precision
    
    LEARNING CONNECTIONS:
    - This is how TensorFlow Lite performs quantized inference
    - Similar to how mobile ML accelerators work
    - Foundation for edge deployment of neural networks
    """
    ### BEGIN SOLUTION
    # Extract numpy arrays
    A_data = A.data if hasattr(A, 'data') else A
    B_data = B.data if hasattr(B, 'data') else B
    
    # Quantize inputs to int8
    A_int8 = np.round(A_data / scale_A).astype(np.int8)
    B_int8 = np.round(B_data / scale_B).astype(np.int8)
    
    # Perform integer matrix multiplication
    # Use int32 for accumulation to prevent overflow
    C_int32 = np.dot(A_int8.astype(np.int32), B_int8.astype(np.int32))
    
    # Rescale result back to float
    C_float = C_int32 * scale_A * scale_B
    
    return Tensor(C_float)
    ### END SOLUTION

# %% ../../modules/source/11_kernels/kernels_dev.ipynb 23
def quantized_relu(x: Tensor, scale: float = 1.0) -> Tensor:
    """
    Quantized ReLU implementation for compressed models.
    
    This function shows how to apply ReLU activation to quantized values
    while maintaining the quantization format.
    
    TODO: Implement quantized ReLU activation.
    
    STEP-BY-STEP IMPLEMENTATION:
    1. Extract numpy array from Tensor
    2. Quantize input to int8 using provided scale
    3. Apply ReLU in integer domain: max(0, x)
    4. Keep result in int8 format (no rescaling needed for ReLU)
    5. Convert back to float using scale
    6. Return result as Tensor
    
    QUANTIZED RELU PROCESS:
    1. Quantize: int8_value = round(float_value / scale)
    2. Apply ReLU: int8_result = max(0, int8_value)
    3. Dequantize: float_result = int8_result * scale
    
    EXAMPLE USAGE:
    ```python
    x = Tensor([-1.0, 0.0, 1.0, 2.0])
    y = quantized_relu(x, scale=1.0/127)
    # Should produce [0.0, 0.0, 1.0, 2.0] (approximately)
    ```
    
    OPTIMIZATION NOTES:
    - ReLU in int8 is just max(0, x) - very fast
    - No floating-point operations needed during activation
    - Maintains quantization format throughout
    - Can be vectorized efficiently
    
    LEARNING CONNECTIONS:
    - This is how quantized neural networks maintain speed
    - Similar to how mobile processors optimize ML inference
    - Foundation for real-time edge computing applications
    """
    ### BEGIN SOLUTION
    # Extract numpy array
    x_data = x.data if hasattr(x, 'data') else x
    
    # Quantize input to int8
    x_int8 = np.round(x_data / scale).astype(np.int8)
    
    # Apply ReLU in integer domain
    x_relu_int8 = np.maximum(0, x_int8)
    
    # Convert back to float
    x_relu_float = x_relu_int8 * scale
    
    return Tensor(x_relu_float)
    ### END SOLUTION