paracodex-baseline

name	paracodex-baseline
description	ParaCodex baseline skill: translate serial / parallel code to GPU-offloaded target API, optimize, compile, and run — all in a single session.

GPU Offload — Baseline (Single Session)

Source: {source_dir}/ Target API: {target_api} Deliverable directory: {kernel_dir}/ Files to create: {file_listing} Executable: ./{executable}

---

Your Task

1. Translate the serial source code in {source_dir}/ to a GPU-offloaded {target_api} version.
2. Apply GPU offloading directives/pragmas/kernels as appropriate for {target_api}.
3. Optimize the code for GPU performance while preserving its original functionality.
4. Compile with {clean_cmd_str} then {build_cmd_str}, and fix any errors until it succeeds.
5. Run with {run_cmd} and confirm the output looks correct.
6. Deliver the complete, modified source code to {kernel_dir}/.

---

Workflow

1. Read the Source Code

Read all relevant files in {source_dir}/ to understand the data structures, compute kernels, and build requirements.

2. Translate to GPU Offload

Write the translated file(s) to {kernel_dir}/: {file_listing}

Offload compute-intensive work to the GPU using {target_api} constructs. Keep I/O and initialization on the CPU. Manage data movement explicitly.

3. Update Makefile if Needed

Check {kernel_dir}/Makefile.nvc and update if required — missing source files, include paths, or linker flags.

4. Compile and Fix

{clean_cmd_str}
{build_cmd_str}

Fix any compiler errors and repeat until the build succeeds.

5. Run and Verify

{run_cmd}

Confirm the output is correct.

---

Deliverable

The complete, modified source code for: {file_listing}

in {kernel_dir}/, that compiles with {build_cmd_str} and produces correct output when run with {run_cmd}.

---

Rules

• NO GIT COMMANDS.
• DO NOT READ/WRITE OUTSIDE THE WORKING DIRECTORY.
• ALWAYS CLEAN BEFORE BUILD ({clean_cmd_str} before every {build_cmd_str}).
• ALWAYS run with exactly {run_cmd} — do not substitute or simplify this command.
• The core compute loops MUST run on the GPU. Offloading only peripheral or setup code while leaving the main computation on the CPU is a failure — fix it.
• If {run_cmd} aborts with an offload error, the GPU pragmas are wrong or missing. Fix the source code. Do NOT remove or relax the run command.