unified-memory

# unified-memory You are **unified-memory** - a specialized skill for CUDA Unified Memory and memory prefetching optimization. This skill provides expert capabilities for simplifying GPU memory management while maintaining high performance. ## Overview This skill enables AI-powered Unified Memory operations including: - Configuring managed memory allocations - Implementing memory prefetch strategies - Handling page fault analysis - Configuring memory hints and advise - Profiling unified memory migration - Optimizing for oversubscription scenarios - Handling multi-GPU unified memory - Comparing managed vs explicit memory performance ## Prerequisites - NVIDIA CUDA Toolkit 8.0+ (Unified Memory) - CUDA 9.0+ for hardware page faulting on Pascal+ - CUDA 11.0+ for advanced prefetching - GPU with compute capability 6.0+ for full UM features - nvidia-smi for migration monitoring - Nsight Systems for migration profiling ## Capabilities ### 1. Basic Unified Memory Allocation Allocate memory accessible from both CPU and GPU: ```cuda #include <cuda_runtime.h> // Allocate managed memory float* data; size_t size = N * sizeof(float); cudaMallocManaged(&data, size); // Initialize on CPU for (int i = 0; i < N; i++) { data[i] = (float)i; } // Use on GPU - data automatically migrates myKernel<<<blocks, threads>>>(data, N); cudaDeviceSynchronize(); // Access on CPU again - data migrates back printf("Result: %f\n", data[0]); // Free managed memory cudaFree(data); ``` ### 2. Memo...