paracodex-ocl-cuda-step2

name	paracodex-ocl-cuda-step2
description	ParaCodex prompt for ocl cuda step2 step.

Performance Tuning - OpenCL to CUDA Migration

Directory: {kernel_dir}/
Files: {file_listing}
Profile: {profile_log_path} Reference: {kernel_dir}/cuda_migration_plan.md

Context

Migrated from OpenCL → CUDA. Target is >1.0x speedup (match or beat OpenCL).

Workflow

1. Verify Baseline Correctness

cd {kernel_dir}
{clean_cmd_str}
timeout 300 {correctness_run_cmd} > current_output.txt 2>&1
diff baseline_output.txt current_output.txt

2. Analyze Performance Profile

1. Read profile (cuda_gpu_kern_sum).
2. Compare with OpenCL baseline.
3. If slower: Major optimization needed.
4. Compare combined GPU kernel + memcpy + sync time to end-to-end runtime and host overhead.
5. If kernel time improves but combined cost regresses, structural work is still needed.
6. Recheck with at least one larger practical input when hardware and timeout budget allow; choose a size that materially exercises the GPU on the available hardware rather than one that remains launch-dominated, and reject changes that only win on the smallest/default run.
7. Use system_info_summary.txt to size that run against the actual device. Prefer the largest short-run, memory-safe input that should load the GPU meaningfully, not merely a modest increase over the default.

3. Create Optimization Plan

MANDATORY: Create optimization_plan.md in {kernel_dir} using template in references/output.md.

Fill Analysis:

• Block Size: Suboptimal?
• Transfer Overhead: >50%?
• Slow Math: Using sinf vs __sinf?
• Warp Divergence?
• Wrong kernel decomposition?

4. Execute Optimization Plan

Reference: references/examples.md (Optimization Examples).

• 4A. Block Size: Tune for occupancy (multiple of 32).
• 4S. Structural Rewrite: Fuse kernels, flatten data, or replace a weak step1 split before micro-tuning.
• 4B. Fast Math: Use intrinsics __sinf, __cosf.
• 4C. Warp Primitives: Use __shfl_down_sync for reductions.
• 4D. Transfers: Async, Pinned, or Unified Memory.
• 4E. Shared Mem: Pad to avoid bank conflicts.
• Keep combined GPU kernel + memcpy + sync cost improving; do not accept a kernel-only win that increases transfer cost.
• Reject any change that improves only the tiny/default run but scales worse at the larger practical input.
• Do not accept a "larger" input that still leaves the GPU mostly idle. If utilization is still weak, push higher within the short-run budget and record the hardware/memory reason if you stop.

5. Verify & Profile (Iterative)

{clean_cmd_str}
{build_cmd_str}
# Check correctness
timeout 300 {run_cmd_str} > optimized_output.txt 2>&1
diff baseline_output.txt optimized_output.txt
# Profile
{profile_cmd_str} > {profile_log_path}_optimized 2>&1

6. Final Summary

Update optimization_plan.md with final metrics and checklist.

Rules

• NO GIT COMMANDS
• DO NOT READ/WRITE OUTSIDE THE WORKING DIRECTORY
• ALWAYS CLEAN BEFORE BUILD
• PRESERVE CORRECTNESS - diff against baseline.
• VERIFY PERFORMANCE GAIN - revert if slower.