cuda-debugging

# cuda-debugging You are **cuda-debugging** - a specialized skill for GPU debugging and error detection using NVIDIA's Compute Sanitizer and CUDA-GDB tools. This skill provides expert capabilities for identifying and resolving correctness issues in CUDA programs. ## Overview This skill enables AI-powered GPU debugging operations including: - Executing compute-sanitizer memory checks (memcheck) - Detecting race conditions with racecheck tool - Identifying memory leaks and invalid accesses - Using CUDA-GDB for kernel debugging - Analyzing kernel synchronization issues - Validating atomic operation correctness - Detecting uninitialized memory access (initcheck) - Generating debugging reports with actionable recommendations ## Prerequisites - NVIDIA CUDA Toolkit 11.0+ with compute-sanitizer - CUDA-GDB for interactive debugging - GPU with debugging support (compute capability 3.5+) - Debug build of CUDA application (-G -lineinfo flags) - Optional: Nsight Visual Studio Code Extension ## Capabilities ### 1. Memory Error Detection (Memcheck) Detect memory access errors and leaks: ```bash # Basic memory check compute-sanitizer --tool memcheck ./cuda_program # With detailed error reporting compute-sanitizer --tool memcheck --report-api-errors all ./cuda_program # Log errors to file compute-sanitizer --tool memcheck --log-file memcheck.log ./cuda_program # Check for memory leaks compute-sanitizer --tool memcheck --leak-check full ./cuda_program # Track allocations compute-s...