Repository navigation
Conversation
…ruct, Reorth flag; validateattributes in +randlapack wrapper
…plicit-matrix escape hatch); memcpy marshal
…nse-pending) not vendored; THIRD_PARTY_LICENSES; gated behind RANDLAPACK_BINDINGS_BENCHMARKS
…hContent benchmark baselines (Epperly MIT, Persson acknowledged); test_fun_nystrom_pp trace-vs-exact
…icit Omega1 removed, sketch_type deprecated); MEX rebuilt + tests pass both precisions
…ized-Gaussian Omega2
…aptive Lanczos-QFA depth)
…ruct (adaptive depth for block_qfa, fixed d otherwise) for benchmark cost accounting
…crash (build blaspp as MKL ILP64 to match MATLAB's runtime BLAS)
…and OpenBLAS on macOS (avoid RandLAPACK's Accelerate BQRRP guard); guard GCC-only MEX link flags
…was scalar-only); document the times-struct lfa_us slots (block_qfa apply = compute_M incl. adaptive certificate)
… = library default) for the block_qfa certificate window
…es += auto_k/auto_s/probe_mv); AdaptiveDelay/AdaptiveMin knobs (inputs 15-16)
…const-marshal fix to avoid n^2 COW copy of A; budget-closure doc q*auto_k
…lly; widen lfa_us to six slots adaptive_delay / adaptive_min were assigned only when the caller passed a nonzero value, relying on the oracle being freshly constructed each call to supply the library defaults. That assumption fails as soon as the oracle is cached (the persistent-handle path), where the previous call's convergence window leaks forward and changes d_used, the estimate and the matvec count. Assign both unconditionally, restoring BlockLanczosQFA<T>::default_adaptive_delay / _min by name when the caller passes 0. The oracle is non-copyable and non-assignable, so "reset by reconstruction" is not available. lfa_us is now six wide (matvec, run_lanczos, apply, rest, total, reorth) to carry the reorthogonalization timer the Lanczos classes now record, and the size gate is >= 5 rather than == 5 so a wider vector is not silently rejected. Requires RandLAPACK with default_adaptive_delay / _min and the six-slot times vector.
…time)
Two fixes, both Windows-only and inert elsewhere:
1. Call randlapack_stage_runtime_dlls() on each MEX (guarded on the helper
existing) so blaspp/lapackpp/backend DLLs land next to the .mexw64;
MATLAB loads a MEX with its own directory on the DLL search path.
Note: consumers must pass -DRANDLAPACK_RUNTIME_DLL_DIRS=<backend bin>
at configure until the RandLAPACK package export bug is fixed (the
installed config's default is masked by the helper's empty cache
declaration).
2. Link MATLAB's own libiomp5md and suppress MSVC's libomp140 default-lib:
MATLAB already hosts Intel's OpenMP runtime, and loading a second
(LLVM) runtime aborts with OMP Error #15; the two share the kmpc ABI.
Validated on Windows: both MEX build (Ninja + MSVC), test_fun_nystrom_pp
and test_bqrrp pass in R2026a, harness validate('full') passes end to end.
Linux MEX rebuilt and re-tested unchanged.
Verified 2026-08-19 on a real Windows machine: without -G Ninja, CMake selects the multi-config Visual Studio generator even inside the x64 developer shell, and the MEX lands in matlab/+randlapack/private/Debug/ where MATLAB's package loader never looks (also as an unoptimized Debug binary). Without the DLL-dirs flag the MEX builds but fails to load in MATLAB (missing BLAS runtime DLLs).
Executing the documented build verbatim on Windows produced a Debug MEX (Ninja+MSVC initializes an unset build type to Debug: /Od /RTC1); on Linux an unset build type compiles without optimization. Correct either way, but timing-useless. Measured 2026-08-19.
…I, numeric-scalar inputs Adaptive for block_qfa: 0 = fixed, 1 = Radau-certified (consistent with scalar_qfa), 2 = legacy window; new RadauReturn (midpoint), AutoDepthCap and AutoProbeFrac inputs; VecNnz = 0 selects the driver's auto density. Times struct gains tr_U/tr_L/certified (block_qfa), probe_ms/probe_converged/ phase2_certified (auto), and a real oracle_mv for block_qfa; NaN marks not-applicable fields. Marshalled buffers are unique_ptr-owned so no error path leaks; read_int/read_double accept any real numeric scalar class; an n-by-0 Omega2 raises instead of dividing by zero; infeasible auto budgets surface as randlapack:fun_nystrom_pp:infeasibleBudget. Fixed the error-id degradation: raise() threw through the engine and MATLAB's exception re-entered the catch chain across the static-libstdc++/RTTI split, so every identifier collapsed to StdError; raise() now throws a TU-local type and the engine error is emitted last. Tests round-trip all six LFATypes in double and single and assert the cost-accounting fields the campaign relies on.
RandLAPACK_DIR must name the directory holding RandLAPACKConfig.cmake, which varies by distro (lib/ vs lib64/) and is the step collaborators trip on. Search the known install layouts beneath the given directory so the RandNLA-project root works directly, and replace find_package's generic failure with an actionable message when nothing is found.
LFAType='adaptive' with AdaptiveMatvecCap; default-on symmetry check with SkipSymCheck opt-out; ignored-knob warnings moved to the wrapper on inputParser.UsingDefaults; Omega2 marshal skipped for auto/adaptive; certified flag marshalled for scalar_qfa; unified error ids; vec_nnz per-row docs; README example fixed. Tests green in double and single.
…, trace_estimate wrapper, degradation diagnostics The Gauss / Radau-at-0 bracket needs the derivatives of f from order two up to alternate in sign, not operator monotonicity; -x log x has the same pattern as log(1+x), so entropy is admitted to scalar_auto/block_auto (poly and square stay blocked). The scalar tiers now report tr_U and tr_L as sums of their per-column Gauss and Gauss-Radau values. Also: scalar_auto/block_auto as LFAType values, CapRankFraction and QuadratureFraction, FirstRowQL with its fallback reported, the symmetric-sketch build option, and randlapack.trace_estimate with its tests. The times struct grows from 22 to 27 fields to carry the kernel's degradation counters out: ritz_clamped, rank_deficient_steps, min_diag_ratio, nystrom_clamped and bracket_evaluated, summed over whichever Lanczos objects the call used. Without a field an event cannot reach a result file however carefully the kernel records it. The help text documents them, and no longer states tr_U >= tr_L as unconditional: the kernel itself notes tr_L can exceed tr_U outside the operator-monotone regime, which is the entropy pages.
Input 28 spend_cap, output phase2_checked; under SpendCap the unevaluated lower bracket value and certificates are NaN; trace_estimate reports the pilot's certificate. Tests check the MEX allocation against a copy of spend_cap_split on a case that separates it from the rejected rule.
…acement helpers - The MEX reads A, Omega2 and the sparse CSC arrays straight from MATLAB's buffers through const pointers, so a dense n x n problem holds one copy of A instead of two. The private copy stays available behind RANDLAPACK_FNPP_COPY_INPUT=1 (the paper's code path); builds without RANDLAPACK_SYMMETRIC_SKETCH always copy. times.input_copied reports which path ran. - RANDLAPACK_PERF_PARCOPY=1 now warns once that it has no effect without the copy. - fun_nystrom_pp.m: the symmetry check runs over column blocks of about 32 MB instead of forming A - A.' and abs(A(:)) (two n x n temporaries); the maxima are exact, so the result is bit-identical. - New numa_place_mex (Linux): randlapack.numa_spread moves an array's pages round-robin over the NUMA nodes in place (move_pages), randlapack.numa_query reports the placement, and randlapack.load_dense reads a raw dense file with threads under an interleave policy. Verified bit-identical to 2fa9639 in every load and placement mode (80/80 cases locally, 12 arms at n = 50,000 on ISAAC); peak memory at n = 50,000 drops from 41 to 22 GiB.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
WIP