Document CUDA programming issues in GPU_Microbenchmark folder

[Copilot]

Overview

This PR adds comprehensive documentation of potential CUDA programming issues found in the src/cuda/GPU_Microbenchmark folder. The analysis scanned all 62 CUDA source files and identified 14 distinct categories of issues ranging from critical bugs to code maintainability concerns.

What's Changed

Added CUDA_ISSUES.md to the repository root containing:

• Detailed analysis of 14 categories of CUDA programming issues
• Severity classifications (High, Medium, Low priority)
• List of affected files for each issue
• Code examples demonstrating the problems
• Recommended fixes with example implementations
• Testing recommendations and priority-based remediation plan

Key Issues Documented

Critical Issues (High Priority)

1. Memory Leaks - 55+ files allocate memory with malloc() and cudaMalloc() but never free it with free() or cudaFree(), causing resource leaks in long-running or repeated executions.

2. Uninitialized Variables - Critical undefined behavior in files like ubench/atomics/Atomic_add_bw/atomic_add_bw.cu where accumulator variables are used without initialization:

   int32_t sum;  // Not initialized
   for (uint32_t i = 0; i < REPEAT_TIMES; i++) {
       sum = sum + atomicAdd(...);  // Undefined behavior
   }

3. Type Mismatches - Wrong sizeof used in cudaMemcpy operations, e.g., copying uint64_t arrays using sizeof(uint32_t), resulting in only half the data being transferred.

4. Undefined warpSize - 9 files use the variable warpSize in kernel code without declaring it, relying on CUDA's built-in constant without documentation.

5. Spinlock Deadlock Risk - ubench/atomics/Spinlock_simple/spinlock_simple.cu uses __syncthreads() inside a spinlock loop, which can cause deadlock when some threads hold the lock while others wait.

Medium Priority Issues

• Missing explicit device synchronization after kernel launches
• Missing error checking for CUDA event API calls
• Incorrect return values (56 files return 1 instead of 0 for success)

Low Priority Issues

• Undocumented bank conflicts in shared memory access patterns
• Magic numbers without named constants
• Inconsistent naming conventions
• Typo in function name: intilizeDeviceProp should be initializeDeviceProp

Statistics

• Total files scanned: 62 CUDA source files
• Files with memory leaks: 55+
• Files with incorrect return values: 56
• Critical issues identified: 5 categories
• Total issue categories: 14

Purpose

This documentation serves as:

• A reference for understanding existing code issues
• A guide for prioritizing fixes in future PRs
• A foundation for establishing CUDA coding standards
• A resource for code reviews and quality improvements

Next Steps

The issues documented here can be addressed in follow-up PRs based on priority:

1. Fix critical issues (uninitialized variables, type mismatches)
2. Add memory cleanup and proper error handling
3. Address code quality and maintainability issues

No code changes are made in this PR - this is purely documentation to inform future improvements.

Original prompt

Scan the src/cuda/GPU_Microbenchmark folder and identify all the potential CUDA programming issues and list them out in the PR description

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.