Enable GPU-accelerated nucleotide ungapped prefilter

[KimBioInfoStudio]

Summary

• Map nucleotide PSSM rows from NucleotideMatrix encoding (A=0, C=1, T=2, G=3, X=4) to ConvertAA_20 positions (A=0, C=1, G=5, T=16, X=20), so existing 21-row GPU kernels score nucleotide sequences correctly
• Always allocate 21-row PSSM for GPU path regardless of alphabet size
• Allow GPU nucleotide search in ungapped prefilter mode (reject only gapped rescore mode)

Approach

The GPU database (makepaddedseqdb) encodes sequences using SubstitutionMatrix::aa2num which follows ConvertAA_20 alphabetical order. The GPU PSSM kernels use hardcoded 21-row shared memory. Rather than modifying CUDA kernels or database encoding, we remap the nucleotide PSSM rows in ungappedprefilter.cpp to match ConvertAA_20 positions. This keeps full compatibility with getUnpadded(), gpuserver, and the existing protein GPU path.

2 files changed, 29 insertions, 11 deletions. Zero CUDA kernel or libmarv modifications.

Test plan

Tested on DGX Spark (Blackwell B200, CUDA 13.0, aarch64):

• Nucleotide GPU vs CPU: 30,000 scores identical (1000 queries × 10K targets, --comp-bias-corr 0 --mask 0)
• Nucleotide small dataset: perfect match scores verified (e.g. 16bp exact match = 32)
• Protein GPU regression: no change (patched vs unpatched GPU output identical)
• Performance: 6.1× speedup over 20-core CPU (1000 queries × 10K targets)

Benchmark (1000q × 10K targets)	GPU	CPU (20 cores)	Speedup
Nucleotide ungapped prefilter	940ms	5780ms	6.1×

🤖 Generated with Claude Code