Skip to contents

Changelog

prego 0.0.12

  • Fix: PWM scoring no longer needs OMP_NUM_THREADS=1 in the environment to avoid the thread explosion. 0.0.11 pinned the OpenMP thread count from R, which helped but could not finish the job: omp_set_num_threads() writes the calling thread’s libgomp ICV, while the workers that actually run the BLAS call are separate threads that inherit the untouched global one. The count is now pinned inside each worker thread itself, in PWMWorker::operator(), resolved via dlsym so prego still does not need to be built with OpenMP (and is a no-op where no OpenMP runtime is loaded).

    Measured standalone against the OpenBLAS conda ships (which links libgomp despite the p in libopenblasp-*.so), 16 threads each running a dgemm: unpinned 54.91s at 2,034 OS threads; pinned per worker 0.08s at 1 thread - matching OMP_NUM_THREADS=1 set before the process starts (0.09s). End to end through extract_pwm() over 20,000 x 500bp sequences and 20 motifs at set_parallel(16), on an idle 128-core node: 8.28s at 2,049 threads before, 2.03s at 17 threads after. The gap is far larger on a loaded machine, which is where this bites.

    Results are unchanged apart from floating-point reassociation inside dgemm (max absolute difference 3.6e-15 over a 5,000 x 20 energy matrix, identical NA pattern), which is inherent to changing a BLAS thread count.

prego 0.0.11

  • Fix: the thread-explosion guard added in 0.0.10 did not actually work on an OpenMP-linked OpenBLAS - the libopenblas-*-openmp build conda ships, and therefore what most lab environments run. RhpcBLASctl::blas_set_num_threads(1) reports success there while changing nothing: openblas_set_num_threads() is a no-op for that build and omp_get_max_threads() is what sizes the thread team. local_serial_blas() now pins the OpenMP thread count as well. Measured on a 128-core node, extract_pwm() over 3,000 x 500bp sequences and 20 motifs at set_parallel(16): unguarded and blas-pin-only both exceeded 600s at 2,049 OS threads and ~11,000% CPU (~80% of it system time); with the OpenMP pin, 4.8s.

    This does not close the gap completely. omp_set_num_threads() writes the calling thread’s libgomp ICV, while the parallel workers that actually run the BLAS call are separate threads which still inherit the untouched global one. The same benchmark runs in 0.47s with 17 threads when OMP_NUM_THREADS=1 is set in the environment before R starts, so on a many-core machine that is still worth setting: OMP_NUM_THREADS=1 Rscript your_script.R. Note that Sys.setenv(OMP_NUM_THREADS = 1) inside the script cannot work - libgomp reads the variable in an ELF constructor at process load, long before R runs line 1. Pinning the count from inside the worker threads themselves would remove the need for the variable; that is worth doing but needs validating on a machine where the affected BLAS can be rebuilt against.

prego 0.0.10

  • Fix: compute_pwm() scored N (and the * wildcard) inconsistently between strands - the forward strand used the column’s average log-probability while the reverse strand used a flat log(0.25). With bidirect = TRUE this made a sequence and its reverse-complement score differently whenever an N fell on an informative position. Both strands now use the column average (get_avg_log_prob()), matching the other likelihood routines, so scoring is strand-symmetric again. Only affects sequences containing N/*.
  • Fix: calc_seq_pwm() / extract_pwm() now error clearly when given sequences of unequal length instead of silently recycling the shorter ones (via rbind) and returning wrong scores. These functions build a single rectangular one-hot matrix and require equal-length sequences; for variable lengths use compute_pwm().
  • Fix: PWM scoring no longer opens thousands of threads / fails on core-limited machines. compute_pwm(), compute_local_pwm() and calc_seq_pwm() (hence extract_pwm()) run a small per-sequence BLAS dgemm inside an RcppParallel/TBB parallelFor. With a multi-threaded BLAS (MKL, OpenBLAS) each TBB worker spawned its own BLAS thread team, so one call opened n_threads * n_blas_threads OS threads (thousands on a many-core node) - which oversubscribed the machine and could fail outright on a cluster job with a thread/process (cgroup pids) limit. The inner BLAS is now pinned to a single thread for the duration of these calls (the TBB loop over sequences is the parallelism layer), and the previous BLAS thread count is restored afterwards. Results are unchanged; throughput is the same or better. Adds a dependency on the RhpcBLASctl package.
  • Fix: compute_pwm() gave a sequence a different score depending on the other sequences in the batch. The motif-scan window was capped to the length of the first sequence for the whole batch, so a longer sequence sitting behind a shorter one was only scanned over its first nchar(sequences[1]) positions and missed motif hits further along. Each sequence is now scored on its own full length, independently of its batch companions.
  • Added return_all parameter to regress_pwm (multi-kmer path). When TRUE, returns every candidate-kmer regression (sorted by validation score) instead of just the best one - useful for getting N independent motifs without the residual-rounds approach used by motif_num > 1. When sample_for_kmers = TRUE, each candidate is refit on the full data.
  • Improved docs for regress_pwm (clarified the three operating modes, fixed n_motifs/comb_modle typos in the return-value section).

prego 0.0.9

  • Fix: Floating point discrepancies between predict and regress_pwm output.

prego 0.0.8

  • Fix: compute_local_pwm now returns NA when sequence length < motif length (issue #35).

prego 0.0.7

  • Fix: regress_multiple_motifs now works with multiple response variables.

prego 0.0.6

  • Added screen_local_pwm to find positions in sequences that match a PSSM.
  • Added return_list parameter to compute_local_pwm.

prego 0.0.5

  • Faster pssm correlation computation using RcppParallel.
  • pssm_match when best=FALSE now returns a score field instead of cor.
  • Implemented computation of KL divergence between two PSSMs. Note that spearman correlation is still the best way to match PSSMs.

prego 0.0.4

  • Added size paramter to intervals_to_seq.

prego 0.0.3

  • Added MotifDB object to store motif information.
  • Implmented a faster energy computation method, which is now used by default.

prego 0.0.0.9000

  • Added a NEWS.md file to track changes to the package.