Prompt file imported from ravindu439/e20-4yp-onchip-offline-neuromorphic-computing (.github/prompts/plan-smemCsvExport.prompt.md). Copyright stays with the author.

Plan: Testbench CSV Export Refinement

Completed Steps

1. Deleted unnecessary proxy python scripts (export_unified_csv.py, read_smem_snapshot.py, compare_dumps.py, verify_spike_counts.py).
2. Modified mnist_infertest_tb.v to natively dump the required CSV format (smem_all_samples.csv).
3. Set export schema to: sample, ts, inp_0...inp_783, spike_h0...spike_h199, spike_o0...spike_o9, vmem_h0...vmem_h199, vmem_o0...vmem_o9.
4. Extracted the true hardware 8-bit truncated LUT index for vmem (uut.all_pre_spike[... + 23 -: 8]) rather than using debug Q28 values.
5. Preserved input spike inputs (ts_input_spikes) until the dump phase (S_WAIT_DONE) to accurately reflect inputs in the CSV.

Next Steps for Refinement

• Run full simulation with all 320 samples to ensure the correct output dimensions (rows = 320 * 16).
• Process the generated smem_all_samples.csv through the downstream ML backprop pipeline to test compatibility.
• Address any column alignment mismatches if the backprop training script expects a different column sequence.
• (Add any details to refine the testbench or data collection mechanism further)