cuda-omp-translatorlisted

# CUDA ↔ OpenMP Translation Guide Reference for evaluating LLM-generated translations between CUDA and OpenMP. Organized by pattern category — each section describes the source construct, the correct target construct, and the failure modes LLMs commonly produce. **Trigger:** `/cuda-omp-translator` or when reviewing CUDA↔OMP eval results. ## When to use - Reviewing why an LLM translation got BUILD_FAIL or VERIFY_FAIL - Writing paper sections about translation difficulty or pattern analysis - Comparing translation quality across models for specific construct types - Creating new benchmark specs for CUDA/OpenMP kernel pairs ## Core Translation Patterns ### 1. Kernel Launch → Parallel Region **CUDA:** ```c __global__ void kernel(float *data, int n) { int tid = blockIdx.x * blockDim.x + threadIdx.x; if (tid < n) data[tid] = data[tid] * 2.0f; } // Launch: kernel<<<grid, block>>>(d_data, n); ``` **OpenMP equivalent:** ```c #pragma omp parallel for for (int tid = 0; tid < n; tid++) { data[tid] = data[tid] * 2.0f; } ``` **Common LLM failures:** - Preserving `blockIdx`/`threadIdx` arithmetic instead of using loop index - Missing `parallel for` — writing just `#pragma omp parallel` without work distribution - Adding unnecessary `num_threads()` clause that limits parallelism ### 2. Device Memory → Stack/Heap Allocation **CUDA:** `cudaMalloc`, `cudaMemcpy` (host↔device transfers) **OpenMP:** Direct pointer use (shared memory space) **Common LLM failures:** - Leavin