currentNe_GPU

Modified GPU-accelerated currentNe(https://github.com/esrud/currentNe) with PED/MAP, and VCF input support, plus complete Ne estimation & confidence intervals. GPU-accelerated fork of currentNe adding PED/MAP and VCF input, and providing end-to-end Nₑ estimation with confidence intervals. The GPU path computes weighted LD (d²) in FP64 using atomicAdd(double*), while Nₑ and CIs follow the original integration and neural-network variance model.

Requires: NVIDIA GPU ≥ Pascal (SM ≥ 6.0), NVIDIA driver + CUDA Toolkit (12+), gcc/g++ & make, and ~1 GB free GPU memory (more for large datasets).

Cooling note: Not recommended to run on passively cooled (fanless) Tesla GPUs without server-grade, front-to-back airflow. The FP64 path saturates the FP units for extended periods, creating stress-test-level thermal load (stress FPU). Inadequate airflow will cause throttling or faults.

Citations: Santiago, E., Caballero, A., Köpke, C., & Novo, I. (2024). Estimation of the contemporary effective population size from SNP data while accounting for mating structure. Molecular Ecology Resources, 24, e13890. https://doi.org/10.1111/1755-0998.13890

Santiago, E., Köpke, C. & Caballero, A. Accounting for population structure and data quality in demographic inference with linkage disequilibrium methods. Nat Commun 16, 6054 (2025). https://doi.org/10.1038/s41467-025-61378-w##

CUDA build (recommended)

unzip currentNe_gpu_full.zip
cd currentNe_gpu_full
make ARCH=sm_89        # choose your GPU's SM arch (sm_70, sm_80, sm_86, sm_89 ...) also should be set `ARCH ?=sm_89` in Makefile accordingly.

This creates ./currentNe_gpu.

CPU fallback

make cpu

This creates ./currentNe_gpu_cpu (OpenMP).

Run

General form:

ulimit -s unlimited    #default Maxloci setting to 20 million, can increase in the cpp file.
./currentNe_gpu <datafile> <num_chromosomes> [options]

• <datafile>: one of
  • prefix.vcf
  • prefix.ped (requires prefix.map in the same folder)
  • prefix.tped (with individuals as columns following the first 4 fields)
• <num_chromosomes>: required (e.g., 22 for human-like autosomes, or the true count for your organism's autosomes).

Common options:

• -s <N> Number of SNPs to use (default: all segregating)
• -t <T> CPU threads (for non-GPU parts; default: OpenMP auto)
• -o <file> Output filename (default: <prefix>_currentNe_OUTPUT.txt)
• -k <int> Important, please see original description in currentNe
• -q Quiet: only print Ne (and with -v also 50% & 90% CI)
• -v With -q, also print CIs
• -p Print full analysis to stdout instead of file

Examples:

# TPED
./currentNe_gpu mydata.tped 19 -t 8

# PED/MAP
./currentNe_gpu mypop.ped 19 -t 8

# VCF
./currentNe_gpu cohort.vcf 19 -t 8
./currentNe_gpu cohort.vcf 19 -t 8 -k 1 

-t 8 is enough

Output

• Full report file (unless -p): <prefix>_currentNe_OUTPUT.txt
  Includes: input stats, d², expected/observed het, Ne point estimate, 50%/90% CI.

Notes

• Double atomicAdd requires GPU architecture sm_60+; set ARCH accordingly.
• For very large SNP counts, memory = L × N bytes (char). Consider filtering -s or thinning SNPs.