PDA-LSTM & baseline code

PDA-LSTM: Knowledge-driven page data arrangement based on LSTM for LCM supression in QLC 3D NAND flash memories

• common_lcm.py
  Random QLC data generator and the paper's LCM score, including the 16×16×16 score LUT, direct score reduction, and the N×N×N page-triple score cube.

• baseline_original.py
  Original data / no processing.

• baseline_random_bitline.py
  Random movement in the bitline direction. Each page independently permutes its cell positions, changing vertical neighboring-cell patterns.

• baseline_random_wordline.py
  Random wordline/page sorting. Thousands of random permutations can be scored in parallel from a precomputed N³ score cube. No N! enumeration.

• baseline_greedy.py
  Greedy wordline sequence generation. Because PDA's objective is based on three adjacent pages, the next page maximizes the score with the previous two. All ordered starting pairs are tried.

• baseline_tsp_sa.py
  TSP-style simulated annealing using parallel independent chains on GPU. The paper discusses SA as a TSP solver but does not state the exact solver configuration used in its experiment, so this is an explicit reproduction choice rather than a claim of hidden author hyperparameters.

• baseline_wbvm.py
  WBVM-style QLC proxy with CL64/CL128.

• baseline_dvds.py
  DVDS-style QLC proxy with CL64/CL128 and a PE-cycle parameter.

• run_all_baselines.py
  Runs all baselines on exactly the same random sample and prints a comparison table.

The paper explicitly states that it compares:

• WBVM and DVDS at different code lengths,
• random sorting in bitline and wordline directions,
• Greedy,
• TSP,
• Original/unprocessed data.

It also says code length is the number of cells handled by a WBVM/DVDS unit.

The paper's background describes:

• Greedy as nearest-neighbor sequence construction.
• TSP and simulated annealing as a sequence-search approach.
• TSP/Greedy repeatedly calculate LCM score during search.

The PDA-LSTM PDF does not reproduce the exact WBVM or DVDS mapping tables, BER-score tables, endurance equations, or their TLC-to-QLC conversion. The cited original works are TLC methods, whereas PDA-LSTM's model/data use QLC states 0..15.

Therefore the included WBVM/DVDS files are intentionally labeled QLC proxy implementations. They are useful for:

• testing the code-length mechanism,
• reproducing the comparison pipeline,
• measuring score/overhead trends,
• plugging in the exact mapping tables later.

They should not be presented as bit-exact implementations of the cited WBVM/DVDS papers.

The WBVM proxy uses the externally reported structural idea of chunk-level flipping with a flag bit and state-count/Vth-based decision.

pip install torch

CUDA is strongly recommended for full paper-size data.

The default is:

N = 16 pages
D = 18 * 1024 * 8 = 147456 cells/page
state = integer 0..15

Run all:

python run_all_baselines.py --device cuda

A faster smoke test:

python run_all_baselines.py \
  --page-cells 2048 \
  --random-wordline-trials 512 \
  --random-bitline-trials 2 \
  --sa-iterations 300 \
  --sa-chains 64

Individual examples:

python baseline_original.py --device cuda
python baseline_random_wordline.py --device cuda --trials 4096
python baseline_random_bitline.py --device cuda --trials 8
python baseline_greedy.py --device cuda
python baseline_tsp_sa.py --device cuda --iterations 5000 --chains 256
python baseline_wbvm.py --device cuda --cl 64
python baseline_wbvm.py --device cuda --cl 128
python baseline_dvds.py --device cuda --cl 64 --pe-cycles 500
python baseline_dvds.py --device cuda --cl 128 --pe-cycles 500

The wordline algorithms do not enumerate all permutations.

common_lcm.py first computes:

S_AC[N,N,N]

only 4096 ordered triples for N=16, using chunked broadcasting and LUT lookup.

Thousands of random full orders are then scored together using tensor advanced indexing. The SA implementation evolves many independent permutation chains in parallel.

DVDS's histogram-to-candidate-map cost uses Einstein summation:

costs = torch.einsum("ncs,ms->ncm", hist, mapped_penalty)

This keeps the implementation consistent with the tensor-first approach used for the PDA-LSTM reproduction.