tlinalg-competition on zen5-cpu

What was measured

Hardware

Every timed cell matches a successful verification census, including operation variants. Each cell is median of process medians ± sample SD of those medians; W is the median within-process sample SD. All values are milliseconds. These are descriptive spreads, not confidence intervals. The between-process spread includes the rotating-order schedule. Each process retains 5 chronological calls; 4 complete process sets are recorded. Spread guide (computed from this run's own rows): per-engine between-start range median 1.59%, worst 60.43% among cells of 1 us or more (a sub-microsecond row is timer and scheduler jitter, not kernel time); 53 of 396 engine cells exceed 5% and carry !; median within-process relative SD 0.34%. A cell whose ±, ! or p is as large as the difference you are reading is not evidence of a regression. The vs tlinalg-pool column is the rotation-averaged ratio exp(mean over starts of log(engine median / tlinalg-pool median)), with the per-start ratios beside it: R > 1 means the engine was slower than that first-declared row in that start. It removes a row-position effect only under one position effect shared by the engines and stable across starts, and only where every declared engine has a measured row (* marks a cell where one has no row, so no ratio there is position-balanced). ! marks a cell where some engine's own per-start median range exceeds 5% of its median, and p a balanced cell whose estimated position-effect ratio exceeds 1.05. Flagged cells are retained, not discarded; the between-start variation can exceed the difference being compared. The vs columns mark only differences larger than the spread this run recorded for those two rows (combined between-start SD, floored at 5%), always with the measured library as the subject: 🔵 <row> X faster than <row> / 🔴 <row> X slower than <row> (the subject is the first column of the table, the compared row is named in the cell, not only in the column header), · inside the band, - that row is not measured in this arm. notes carries the cell warnings (* a declared row has no measurement, ! between-start range over 5%, p position effect over 5%, dispatch-dominated means the row is under 20 us so 4T/8T pay thread-region start-up). Per-start ratios and every chronological sample are in the run directory named at the top of this page (run*.csv). These describe the recorded revisions; they are not a regression claim.

batch-small, f64, 8T (CPU 4-11)

case tlinalg-pool (ms) lapack-openblas (ms) vs lapack-openblas notes
cholesky/n16/b1024 0.115831 ± 0.000822909 (W 0.00177606) 1.02348 ± 0.0229 (W 0.00168164) 🔵 tlinalg-pool 8.84× faster than lapack-openblas
cholesky/n32/b1024 0.436202 ± 0.00688286 (W 0.00660227) 3.61067 ± 0.0579113 (W 0.00819164) 🔵 tlinalg-pool 8.28× faster than lapack-openblas
cholesky/n4/b1024 0.182434 ± 0.00103055 (W 0.00131396) 0.196366 ± 0.000277045 (W 0.00128066) 🔵 tlinalg-pool 1.08× faster than lapack-openblas
cholesky/n8/b1024 0.360556 ± 0.00325099 (W 0.00178496) 0.391438 ± 0.0069268 (W 0.00280302) 🔵 tlinalg-pool 1.09× faster than lapack-openblas
eig/n16/b1024 11.8888 ± 0.0427365 (W 0.0354367) 68.7331 ± 0.281975 (W 0.0728445) 🔵 tlinalg-pool 5.78× faster than lapack-openblas
eig/n32/b1024 53.4036 ± 0.0844592 (W 0.329476) 324.78 ± 1.8485 (W 0.679122) 🔵 tlinalg-pool 6.08× faster than lapack-openblas
eig/n4/b1024 6.19849 ± 0.0132486 (W 0.0126934) 5.32 ± 0.0169792 (W 0.00719111) 🔴 tlinalg-pool 1.17× slower than lapack-openblas
eig/n8/b1024 2.99139 ± 0.0228095 (W 0.0075552) 17.5878 ± 0.0260527 (W 0.0181406) 🔵 tlinalg-pool 5.88× faster than lapack-openblas
eigh/n16/b1024 3.80156 ± 0.109535 (W 0.0182831) 40.1353 ± 0.291824 (W 0.0949408) 🔵 tlinalg-pool 10.56× faster than lapack-openblas ! between-start range over 5%
eigh/n32/b1024 13.9991 ± 0.112837 (W 0.111004) 119.549 ± 0.490114 (W 0.204832) 🔵 tlinalg-pool 8.54× faster than lapack-openblas
eigh/n4/b1024 2.5466 ± 0.0123608 (W 0.00288794) 3.02976 ± 0.00810649 (W 0.0033896) 🔵 tlinalg-pool 1.19× faster than lapack-openblas
eigh/n8/b1024 1.11944 ± 0.0334521 (W 0.0274408) 10.856 ± 0.0668914 (W 0.0100859) 🔵 tlinalg-pool 9.70× faster than lapack-openblas ! between-start range over 5%
eigvals/n16/b1024 8.20468 ± 0.0498275 (W 0.0283853) 49.663 ± 0.258883 (W 0.046725) 🔵 tlinalg-pool 6.05× faster than lapack-openblas
eigvals/n32/b1024 36.4059 ± 0.0672407 (W 0.0880231) 218.11 ± 1.8602 (W 0.457356) 🔵 tlinalg-pool 5.99× faster than lapack-openblas
eigvals/n4/b1024 3.19327 ± 0.0163456 (W 0.0126401) 3.41338 ± 0.00561554 (W 0.00511686) 🔵 tlinalg-pool 1.07× faster than lapack-openblas
eigvals/n8/b1024 1.95964 ± 0.0559743 (W 0.01734) 13.2765 ± 0.0542738 (W 0.00917883) 🔵 tlinalg-pool 6.77× faster than lapack-openblas ! between-start range over 5%
eigvalsh/n16/b1024 2.81534 ± 0.0143008 (W 0.00574729) 14.5344 ± 0.0565757 (W 0.00699669) 🔵 tlinalg-pool 5.16× faster than lapack-openblas
eigvalsh/n32/b1024 10.9239 ± 0.0374498 (W 0.026731) 50.1168 ± 0.332884 (W 0.211124) 🔵 tlinalg-pool 4.59× faster than lapack-openblas
eigvalsh/n4/b1024 2.04936 ± 0.00234144 (W 0.00281848) 1.39635 ± 0.0129779 (W 0.0024437) 🔴 tlinalg-pool 1.47× slower than lapack-openblas
eigvalsh/n8/b1024 0.873355 ± 0.00405706 (W 0.00217548) 4.55874 ± 0.0119514 (W 0.00324856) 🔵 tlinalg-pool 5.22× faster than lapack-openblas
full_piv_lu/n16/b1024 0.513731 ± 0.00459812 (W 0.00711304) 4.74228 ± 0.118499 (W 0.0227884) 🔵 tlinalg-pool 9.23× faster than lapack-openblas ! between-start range over 5%
full_piv_lu/n32/b1024 2.17045 ± 0.0352224 (W 0.0126136) 28.9437 ± 1.1893 (W 0.110655) 🔵 tlinalg-pool 13.34× faster than lapack-openblas ! between-start range over 5%
full_piv_lu/n4/b1024 0.407417 ± 0.00305197 (W 0.00190131) 0.367258 ± 0.00156688 (W 0.00194645) 🔴 tlinalg-pool 1.11× slower than lapack-openblas
full_piv_lu/n8/b1024 1.14999 ± 0.0072062 (W 0.00418327) 1.04027 ± 0.0125749 (W 0.00260861) 🔴 tlinalg-pool 1.11× slower than lapack-openblas
full_piv_lu_solve/n16/b1024 0.466818 ± 0.00346781 (W 0.0023712) 6.49679 ± 0.100431 (W 0.0194758) 🔵 tlinalg-pool 13.92× faster than lapack-openblas
full_piv_lu_solve/n32/b1024 1.55165 ± 0.0192153 (W 0.0169212) 31.7534 ± 0.524983 (W 0.267357) 🔵 tlinalg-pool 20.46× faster than lapack-openblas
full_piv_lu_solve/n4/b1024 0.531243 ± 0.00271847 (W 0.00341891) 0.948203 ± 0.00417228 (W 0.00219706) 🔵 tlinalg-pool 1.78× faster than lapack-openblas
full_piv_lu_solve/n8/b1024 0.172431 ± 0.00393633 (W 0.0014649) 2.01761 ± 0.0138558 (W 0.00737823) 🔵 tlinalg-pool 11.70× faster than lapack-openblas
householder_apply/n16/b1024 0.633489 ± 0.0108165 (W 0.0116424) 4.46261 ± 0.0677051 (W 0.0096787) 🔵 tlinalg-pool 7.04× faster than lapack-openblas
householder_apply/n32/b1024 2.55651 ± 0.0361869 (W 0.030708) 19.0966 ± 0.375509 (W 0.0500507) 🔵 tlinalg-pool 7.47× faster than lapack-openblas
householder_apply/n4/b1024 0.385317 ± 0.00541028 (W 0.00163251) 0.418769 ± 0.00330711 (W 0.00191399) 🔵 tlinalg-pool 1.09× faster than lapack-openblas
householder_apply/n8/b1024 0.178131 ± 0.00361483 (W 0.00271061) 1.20379 ± 0.0133207 (W 0.00417576) 🔵 tlinalg-pool 6.76× faster than lapack-openblas
householder_factor/n16/b1024 0.50562 ± 0.0103861 (W 0.00196655) 4.32263 ± 0.00667262 (W 0.00688515) 🔵 tlinalg-pool 8.55× faster than lapack-openblas
householder_factor/n32/b1024 1.69067 ± 0.045729 (W 0.013099) 14.6004 ± 0.0403202 (W 0.0200138) 🔵 tlinalg-pool 8.64× faster than lapack-openblas ! between-start range over 5%
householder_factor/n4/b1024 0.4676 ± 0.00309011 (W 0.00229123) 0.535977 ± 0.00294267 (W 0.00231764) 🔵 tlinalg-pool 1.15× faster than lapack-openblas
householder_factor/n8/b1024 0.174174 ± 0.00271979 (W 0.00154645) 1.49543 ± 0.00400896 (W 0.00211746) 🔵 tlinalg-pool 8.59× faster than lapack-openblas
lu/n16/b1024 0.36277 ± 0.00653679 (W 0.00491152) 2.39565 ± 0.0329747 (W 0.0039499) 🔵 tlinalg-pool 6.60× faster than lapack-openblas
lu/n32/b1024 1.27034 ± 0.0107635 (W 0.00891927) 9.73588 ± 0.0568326 (W 0.0246988) 🔵 tlinalg-pool 7.66× faster than lapack-openblas
lu/n4/b1024 0.308009 ± 0.00658006 (W 0.00247144) 0.329839 ± 0.00533192 (W 0.00185811) 🔵 tlinalg-pool 1.07× faster than lapack-openblas
lu/n8/b1024 0.848789 ± 0.0114558 (W 0.00462561) 0.813463 ± 0.0121557 (W 0.00195453) ·
lu_factor/n16/b1024 0.267067 ± 0.00261926 (W 0.00246784) 1.80995 ± 0.0141442 (W 0.00413412) 🔵 tlinalg-pool 6.78× faster than lapack-openblas
lu_factor/n32/b1024 0.739486 ± 0.0111277 (W 0.00992806) 6.80178 ± 0.0892271 (W 0.00875737) 🔵 tlinalg-pool 9.20× faster than lapack-openblas
lu_factor/n4/b1024 0.256257 ± 0.0019459 (W 0.00163778) 0.218451 ± 0.000824783 (W 0.00101106) 🔴 tlinalg-pool 1.17× slower than lapack-openblas
lu_factor/n8/b1024 0.624927 ± 0.0141651 (W 0.00475539) 0.572445 ± 0.00148068 (W 0.00336155) 🔴 tlinalg-pool 1.09× slower than lapack-openblas
lu_factor_solve/n16/b1024 0.340734 ± 0.0244192 (W 0.00275711) 3.73993 ± 0.0293265 (W 0.00714342) 🔵 tlinalg-pool 10.98× faster than lapack-openblas ! between-start range over 5%
lu_factor_solve/n32/b1024 0.898277 ± 0.0109225 (W 0.00539666) 9.85064 ± 0.0570497 (W 0.0214597) 🔵 tlinalg-pool 10.97× faster than lapack-openblas
lu_factor_solve/n4/b1024 0.318176 ± 0.00600234 (W 0.00167628) 1.31778 ± 0.0256123 (W 0.00344876) 🔵 tlinalg-pool 4.14× faster than lapack-openblas
lu_factor_solve/n8/b1024 0.131269 ± 0.000870778 (W 0.00203876) 1.95336 ± 0.0198828 (W 0.00353414) 🔵 tlinalg-pool 14.88× faster than lapack-openblas
lu_solve_prepared/n16/b1024 0.108091 ± 0.00130538 (W 0.00305807) 1.56374 ± 0.0377564 (W 0.00420955) 🔵 tlinalg-pool 14.47× faster than lapack-openblas ! between-start range over 5%
lu_solve_prepared/n32/b1024 0.323277 ± 0.00957474 (W 0.0161677) 2.31328 ± 0.0101608 (W 0.00656351) 🔵 tlinalg-pool 7.16× faster than lapack-openblas ! between-start range over 5%
lu_solve_prepared/n4/b1024 0.0712075 ± 0.000213253 (W 0.000805738) 0.985288 ± 0.0387587 (W 0.00342259) 🔵 tlinalg-pool 13.84× faster than lapack-openblas ! between-start range over 5%
lu_solve_prepared/n8/b1024 0.233329 ± 0.000579608 (W 0.00203614) 1.25781 ± 0.0216744 (W 0.00590235) 🔵 tlinalg-pool 5.39× faster than lapack-openblas
qr/n16/b1024 1.1341 ± 0.0111089 (W 0.00585662) 7.79074 ± 0.0773897 (W 0.0100286) 🔵 tlinalg-pool 6.87× faster than lapack-openblas
qr/n32/b1024 4.57626 ± 0.0209588 (W 0.0151796) 27.1742 ± 0.237606 (W 0.0355756) 🔵 tlinalg-pool 5.94× faster than lapack-openblas
qr/n4/b1024 0.764017 ± 0.00703963 (W 0.00136798) 0.957666 ± 0.00472188 (W 0.00180489) 🔵 tlinalg-pool 1.25× faster than lapack-openblas
qr/n8/b1024 2.36233 ± 0.0350214 (W 0.00291797) 2.61491 ± 0.00292615 (W 0.00284527) 🔵 tlinalg-pool 1.11× faster than lapack-openblas
rank_revealing_qr/n16/b1024 1.34745 ± 0.0169477 (W 0.00347547) 9.41738 ± 0.0708146 (W 0.0372737) 🔵 tlinalg-pool 6.99× faster than lapack-openblas
rank_revealing_qr/n32/b1024 4.61331 ± 0.0484498 (W 0.0353822) 33.6207 ± 0.0462524 (W 0.102835) 🔵 tlinalg-pool 7.29× faster than lapack-openblas
rank_revealing_qr/n4/b1024 1.2613 ± 0.0247767 (W 0.00620184) 1.14852 ± 0.00462593 (W 0.00287006) 🔴 tlinalg-pool 1.10× slower than lapack-openblas
rank_revealing_qr/n8/b1024 3.40138 ± 0.038629 (W 0.00568472) 3.02578 ± 0.0222472 (W 0.0089922) 🔴 tlinalg-pool 1.12× slower than lapack-openblas
solve/n16/b1024 0.359449 ± 0.00547876 (W 0.00227685) 3.7492 ± 0.0140708 (W 0.00836091) 🔵 tlinalg-pool 10.43× faster than lapack-openblas
solve/n32/b1024 0.980148 ± 0.0125778 (W 0.0238265) 9.9059 ± 0.0509006 (W 0.0504676) 🔵 tlinalg-pool 10.11× faster than lapack-openblas
solve/n4/b1024 0.386254 ± 0.00380939 (W 0.00301288) 1.26375 ± 0.0607452 (W 0.00940769) 🔵 tlinalg-pool 3.27× faster than lapack-openblas ! between-start range over 5%
solve/n8/b1024 0.135386 ± 0.00260973 (W 0.00256063) 2.04149 ± 0.0259247 (W 0.00706886) 🔵 tlinalg-pool 15.08× faster than lapack-openblas
svd_full/n16/b1024 7.23096 ± 0.0697134 (W 0.112671) 80.4687 ± 0.642304 (W 0.24861) 🔵 tlinalg-pool 11.13× faster than lapack-openblas
svd_full/n32/b1024 26.237 ± 1.1376 (W 0.205987) 238.787 ± 0.413731 (W 0.098751) 🔵 tlinalg-pool 9.10× faster than lapack-openblas ! between-start range over 5%
svd_full/n4/b1024 4.34651 ± 0.0260325 (W 0.00830655) 6.30693 ± 0.0100899 (W 0.00319679) 🔵 tlinalg-pool 1.45× faster than lapack-openblas
svd_full/n8/b1024 1.97388 ± 0.0620062 (W 0.0211791) 21.2462 ± 0.0169103 (W 0.0158883) 🔵 tlinalg-pool 10.76× faster than lapack-openblas ! between-start range over 5%
svd_thin/n16/b1024 7.42111 ± 0.231423 (W 0.0985893) 80.9252 ± 0.555225 (W 0.24138) 🔵 tlinalg-pool 10.90× faster than lapack-openblas ! between-start range over 5%
svd_thin/n32/b1024 26.0906 ± 0.279536 (W 0.0593399) 238.864 ± 1.42338 (W 0.393455) 🔵 tlinalg-pool 9.16× faster than lapack-openblas
svd_thin/n4/b1024 4.36278 ± 0.023722 (W 0.00513781) 6.31736 ± 0.0272301 (W 0.00741009) 🔵 tlinalg-pool 1.45× faster than lapack-openblas
svd_thin/n8/b1024 1.9534 ± 0.0387594 (W 0.0125211) 21.2034 ± 0.203035 (W 0.0222155) 🔵 tlinalg-pool 10.85× faster than lapack-openblas
svd_values/n16/b1024 5.12872 ± 0.0101151 (W 0.0116694) 25.0499 ± 0.280895 (W 0.0202599) 🔵 tlinalg-pool 4.88× faster than lapack-openblas
svd_values/n32/b1024 20.0955 ± 0.0430097 (W 0.20246) 79.4625 ± 0.290644 (W 0.0866422) 🔵 tlinalg-pool 3.95× faster than lapack-openblas
svd_values/n4/b1024 3.18416 ± 0.0182354 (W 0.00574599) 3.46262 ± 0.0174327 (W 0.00480187) 🔵 tlinalg-pool 1.09× faster than lapack-openblas
svd_values/n8/b1024 1.44403 ± 0.0102175 (W 0.00269347) 8.92303 ± 0.0313565 (W 0.012487) 🔵 tlinalg-pool 6.18× faster than lapack-openblas
triangular_solve/n16/b1024 0.0785715 ± 0.0040262 (W 0.00193069) 1.04135 ± 0.108276 (W 0.00867309) 🔵 tlinalg-pool 13.25× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
triangular_solve/n32/b1024 0.207862 ± 0.015201 (W 0.00753265) 1.55002 ± 0.0421587 (W 0.0182887) 🔵 tlinalg-pool 7.46× faster than lapack-openblas ! between-start range over 5%
triangular_solve/n4/b1024 0.09137 ± 0.000248556 (W 0.000688929) 0.766261 ± 0.0151998 (W 0.00487141) 🔵 tlinalg-pool 8.39× faster than lapack-openblas
triangular_solve/n8/b1024 0.220525 ± 0.00148701 (W 0.00163755) 0.913679 ± 0.0411722 (W 0.00422571) 🔵 tlinalg-pool 4.14× faster than lapack-openblas ! between-start range over 5%

batch-small, c64, 8T (CPU 4-11)

case tlinalg-pool (ms) lapack-openblas (ms) vs lapack-openblas notes
cholesky/n16/b1024 0.168247 ± 0.00127109 (W 0.00193349) 1.5449 ± 0.0261815 (W 0.00827893) 🔵 tlinalg-pool 9.18× faster than lapack-openblas
cholesky/n32/b1024 0.815357 ± 0.0076747 (W 0.00906212) 6.28275 ± 0.0288274 (W 0.00921707) 🔵 tlinalg-pool 7.71× faster than lapack-openblas
cholesky/n4/b1024 0.20155 ± 0.00361586 (W 0.00112187) 0.229928 ± 0.00201718 (W 0.00139305) 🔵 tlinalg-pool 1.14× faster than lapack-openblas
cholesky/n8/b1024 0.420908 ± 0.000463632 (W 0.00274279) 0.488183 ± 0.00297212 (W 0.00208827) 🔵 tlinalg-pool 1.16× faster than lapack-openblas
eig/n16/b1024 16.353 ± 0.297862 (W 0.0532519) 198.83 ± 0.279421 (W 0.433126) 🔵 tlinalg-pool 12.16× faster than lapack-openblas
eig/n32/b1024 67.2542 ± 0.0447112 (W 0.0811742) 885.643 ± 23.407 (W 0.792737) 🔵 tlinalg-pool 13.17× faster than lapack-openblas ! between-start range over 5%
eig/n4/b1024 10.2527 ± 0.0176541 (W 0.0123505) 13.4065 ± 0.025134 (W 0.0159456) 🔵 tlinalg-pool 1.31× faster than lapack-openblas
eig/n8/b1024 4.43329 ± 0.0173455 (W 0.00953156) 46.6787 ± 0.143099 (W 0.223821) 🔵 tlinalg-pool 10.53× faster than lapack-openblas
eigh/n16/b1024 4.14293 ± 0.0837391 (W 0.0282233) 45.1621 ± 0.21983 (W 0.0356135) 🔵 tlinalg-pool 10.90× faster than lapack-openblas
eigh/n32/b1024 17.4172 ± 0.603188 (W 0.176511) 187.987 ± 0.624269 (W 0.118652) 🔵 tlinalg-pool 10.79× faster than lapack-openblas ! between-start range over 5%
eigh/n4/b1024 2.92837 ± 0.00827253 (W 0.00772235) 3.63664 ± 0.00847265 (W 0.00381445) 🔵 tlinalg-pool 1.24× faster than lapack-openblas
eigh/n8/b1024 1.23775 ± 0.0462286 (W 0.0208823) 12.4034 ± 0.103411 (W 0.0103568) 🔵 tlinalg-pool 10.02× faster than lapack-openblas ! between-start range over 5%
eigvals/n16/b1024 9.71274 ± 0.0205403 (W 0.0542253) 148.679 ± 0.216166 (W 0.444748) 🔵 tlinalg-pool 15.31× faster than lapack-openblas
eigvals/n32/b1024 42.4618 ± 0.0286741 (W 0.0462638) 597.723 ± 0.493331 (W 0.891888) 🔵 tlinalg-pool 14.08× faster than lapack-openblas
eigvals/n4/b1024 5.02102 ± 0.0192011 (W 0.0105101) 10.034 ± 0.0288924 (W 0.00461887) 🔵 tlinalg-pool 2.00× faster than lapack-openblas
eigvals/n8/b1024 2.53493 ± 0.00206675 (W 0.00871485) 36.948 ± 0.135348 (W 0.0216792) 🔵 tlinalg-pool 14.58× faster than lapack-openblas
eigvalsh/n16/b1024 3.01833 ± 0.00680026 (W 0.00987925) 19.3767 ± 0.0597495 (W 0.022882) 🔵 tlinalg-pool 6.42× faster than lapack-openblas
eigvalsh/n32/b1024 11.7765 ± 0.0225242 (W 0.0252773) 73.2416 ± 0.306103 (W 0.13452) 🔵 tlinalg-pool 6.22× faster than lapack-openblas
eigvalsh/n4/b1024 2.4269 ± 0.0115083 (W 0.00589869) 2.13264 ± 0.00847243 (W 0.00363091) 🔴 tlinalg-pool 1.14× slower than lapack-openblas
eigvalsh/n8/b1024 0.933999 ± 0.0039741 (W 0.00341139) 6.16366 ± 0.0311285 (W 0.00581578) 🔵 tlinalg-pool 6.60× faster than lapack-openblas
full_piv_lu/n16/b1024 0.654875 ± 0.00537657 (W 0.00962371) 33.2708 ± 0.14039 (W 0.0534897) 🔵 tlinalg-pool 50.80× faster than lapack-openblas
full_piv_lu/n32/b1024 3.07942 ± 0.0300656 (W 0.0132808) 199.78 ± 2.37727 (W 0.353201) 🔵 tlinalg-pool 64.88× faster than lapack-openblas
full_piv_lu/n4/b1024 0.462395 ± 0.00615094 (W 0.00454162) 0.901808 ± 0.00452438 (W 0.00199569) 🔵 tlinalg-pool 1.95× faster than lapack-openblas
full_piv_lu/n8/b1024 0.195121 ± 0.00354099 (W 0.00331943) 3.87555 ± 0.0312997 (W 0.00469206) 🔵 tlinalg-pool 19.86× faster than lapack-openblas
full_piv_lu_solve/n16/b1024 0.711339 ± 0.0124379 (W 0.0147774) 37.7339 ± 0.165327 (W 0.0340171) 🔵 tlinalg-pool 53.05× faster than lapack-openblas
full_piv_lu_solve/n32/b1024 2.65967 ± 0.0140767 (W 0.00961588) 211.664 ± 1.6357 (W 0.256585) 🔵 tlinalg-pool 79.58× faster than lapack-openblas
full_piv_lu_solve/n4/b1024 0.762271 ± 0.00269772 (W 0.00310748) 2.12774 ± 0.0190713 (W 0.00353211) 🔵 tlinalg-pool 2.79× faster than lapack-openblas
full_piv_lu_solve/n8/b1024 0.280707 ± 0.00526909 (W 0.00392276) 6.13932 ± 0.0586772 (W 0.00752616) 🔵 tlinalg-pool 21.87× faster than lapack-openblas
householder_apply/n16/b1024 0.642292 ± 0.0156504 (W 0.00913629) 7.19737 ± 0.153977 (W 0.0182638) 🔵 tlinalg-pool 11.21× faster than lapack-openblas ! between-start range over 5%
householder_apply/n32/b1024 3.56412 ± 0.078881 (W 0.0332999) 35.393 ± 0.669378 (W 0.0399594) 🔵 tlinalg-pool 9.93× faster than lapack-openblas
householder_apply/n4/b1024 0.36286 ± 0.00432104 (W 0.0019142) 0.438631 ± 0.000450266 (W 0.00193288) 🔵 tlinalg-pool 1.21× faster than lapack-openblas
householder_apply/n8/b1024 0.156626 ± 0.00401426 (W 0.00146448) 1.48481 ± 0.0136839 (W 0.00325063) 🔵 tlinalg-pool 9.48× faster than lapack-openblas ! between-start range over 5%
householder_factor/n16/b1024 0.599471 ± 0.0181174 (W 0.00672999) 7.32993 ± 0.0258324 (W 0.0089297) 🔵 tlinalg-pool 12.23× faster than lapack-openblas ! between-start range over 5%
householder_factor/n32/b1024 2.59585 ± 0.0204463 (W 0.0175617) 27.2451 ± 0.128436 (W 0.0197896) 🔵 tlinalg-pool 10.50× faster than lapack-openblas
householder_factor/n4/b1024 0.573148 ± 0.00363015 (W 0.00233865) 0.925486 ± 0.00339089 (W 0.00169663) 🔵 tlinalg-pool 1.61× faster than lapack-openblas
householder_factor/n8/b1024 0.200919 ± 0.000265695 (W 0.00294178) 2.36849 ± 0.00380375 (W 0.00335941) 🔵 tlinalg-pool 11.79× faster than lapack-openblas
lu/n16/b1024 0.44584 ± 0.00198473 (W 0.0100505) 3.82384 ± 0.128884 (W 0.0153596) 🔵 tlinalg-pool 8.58× faster than lapack-openblas ! between-start range over 5%
lu/n32/b1024 2.09234 ± 0.0125589 (W 0.0119431) 15.4322 ± 0.20034 (W 0.0189064) 🔵 tlinalg-pool 7.38× faster than lapack-openblas
lu/n4/b1024 0.378584 ± 0.00882074 (W 0.00249016) 0.376842 ± 0.0496584 (W 0.00250672) · ! between-start range over 5%
lu/n8/b1024 0.154432 ± 0.00221751 (W 0.00396397) 1.02056 ± 0.0920491 (W 0.00183171) 🔵 tlinalg-pool 6.61× faster than lapack-openblas ! between-start range over 5%
lu_factor/n16/b1024 0.350152 ± 0.094582 (W 0.0204632) 2.69263 ± 0.00924127 (W 0.00514524) 🔵 tlinalg-pool 7.69× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
lu_factor/n32/b1024 1.29691 ± 0.038324 (W 0.0443762) 11.2921 ± 0.134472 (W 0.0157959) 🔵 tlinalg-pool 8.71× faster than lapack-openblas ! between-start range over 5%
lu_factor/n4/b1024 0.30354 ± 0.00715188 (W 0.00258671) 0.271375 ± 0.00171203 (W 0.00166938) 🔴 tlinalg-pool 1.12× slower than lapack-openblas ! between-start range over 5%
lu_factor/n8/b1024 0.114237 ± 0.00177571 (W 0.00193684) 0.754328 ± 0.00851392 (W 0.00282038) 🔵 tlinalg-pool 6.60× faster than lapack-openblas
lu_factor_solve/n16/b1024 0.464474 ± 0.00543965 (W 0.00424362) 5.07023 ± 0.108764 (W 0.0101428) 🔵 tlinalg-pool 10.92× faster than lapack-openblas
lu_factor_solve/n32/b1024 1.79378 ± 0.0435937 (W 0.0169653) 15.4837 ± 0.254168 (W 0.0161713) 🔵 tlinalg-pool 8.63× faster than lapack-openblas ! between-start range over 5%
lu_factor_solve/n4/b1024 0.457967 ± 0.0100565 (W 0.0029922) 1.10948 ± 0.163213 (W 0.00229131) 🔵 tlinalg-pool 2.42× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
lu_factor_solve/n8/b1024 0.174134 ± 0.01387 (W 0.00657279) 2.38791 ± 0.0304927 (W 0.00418359) 🔵 tlinalg-pool 13.71× faster than lapack-openblas ! between-start range over 5%
lu_solve_prepared/n16/b1024 0.197132 ± 0.0462067 (W 0.00376695) 1.93 ± 0.157526 (W 0.00585239) 🔵 tlinalg-pool 9.79× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
lu_solve_prepared/n32/b1024 0.89241 ± 0.033715 (W 0.0135466) 3.94291 ± 0.163753 (W 0.0126407) 🔵 tlinalg-pool 4.42× faster than lapack-openblas ! between-start range over 5%
lu_solve_prepared/n4/b1024 0.165468 ± 0.000333067 (W 0.00169133) 0.775949 ± 0.14247 (W 0.00940855) 🔵 tlinalg-pool 4.69× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
lu_solve_prepared/n8/b1024 0.0884495 ± 0.00181075 (W 0.00412834) 1.39704 ± 0.0239078 (W 0.00450177) 🔵 tlinalg-pool 15.79× faster than lapack-openblas
qr/n16/b1024 1.43366 ± 0.0130763 (W 0.0102023) 12.5423 ± 0.0216253 (W 0.0250228) 🔵 tlinalg-pool 8.75× faster than lapack-openblas
qr/n32/b1024 9.37455 ± 0.078953 (W 0.0264915) 49.5074 ± 0.149891 (W 0.209576) 🔵 tlinalg-pool 5.28× faster than lapack-openblas
qr/n4/b1024 0.91871 ± 0.00560864 (W 0.00204609) 1.42946 ± 0.0157339 (W 0.00372875) 🔵 tlinalg-pool 1.56× faster than lapack-openblas
qr/n8/b1024 0.351664 ± 0.00342454 (W 0.00284694) 3.77246 ± 0.0161733 (W 0.00258636) 🔵 tlinalg-pool 10.73× faster than lapack-openblas
rank_revealing_qr/n16/b1024 1.57377 ± 0.0129001 (W 0.00974695) 16.7227 ± 0.101798 (W 0.0297145) 🔵 tlinalg-pool 10.63× faster than lapack-openblas
rank_revealing_qr/n32/b1024 8.58819 ± 0.060905 (W 0.0349042) 65.6058 ± 0.347961 (W 0.355948) 🔵 tlinalg-pool 7.64× faster than lapack-openblas
rank_revealing_qr/n4/b1024 1.53686 ± 0.01878 (W 0.00321282) 1.80352 ± 0.012944 (W 0.00599221) 🔵 tlinalg-pool 1.17× faster than lapack-openblas
rank_revealing_qr/n8/b1024 0.496209 ± 0.00777758 (W 0.00638893) 5.04359 ± 0.0206025 (W 0.00391782) 🔵 tlinalg-pool 10.16× faster than lapack-openblas
solve/n16/b1024 0.513385 ± 0.0116489 (W 0.00864559) 5.16157 ± 0.183578 (W 0.0187679) 🔵 tlinalg-pool 10.05× faster than lapack-openblas ! between-start range over 5%
solve/n32/b1024 1.86693 ± 0.017851 (W 0.0163157) 16.2394 ± 0.0409975 (W 0.0157513) 🔵 tlinalg-pool 8.70× faster than lapack-openblas
solve/n4/b1024 0.574003 ± 0.00467282 (W 0.00328337) 1.51455 ± 0.046113 (W 0.00256951) 🔵 tlinalg-pool 2.64× faster than lapack-openblas ! between-start range over 5%
solve/n8/b1024 0.192383 ± 0.00168387 (W 0.00185518) 2.20817 ± 0.159966 (W 0.00909961) 🔵 tlinalg-pool 11.48× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
svd_full/n16/b1024 7.77616 ± 0.0771587 (W 0.03979) 93.4677 ± 0.284051 (W 0.261942) 🔵 tlinalg-pool 12.02× faster than lapack-openblas
svd_full/n32/b1024 33.721 ± 1.02104 (W 0.0904805) 294.879 ± 1.03672 (W 0.451659) 🔵 tlinalg-pool 8.74× faster than lapack-openblas ! between-start range over 5%
svd_full/n4/b1024 5.24338 ± 0.0141527 (W 0.00432623) 7.82344 ± 0.049825 (W 0.0225109) 🔵 tlinalg-pool 1.49× faster than lapack-openblas
svd_full/n8/b1024 2.20454 ± 0.0358642 (W 0.0594616) 24.8906 ± 0.0308784 (W 0.0176416) 🔵 tlinalg-pool 11.29× faster than lapack-openblas
svd_thin/n16/b1024 7.78827 ± 0.109083 (W 0.0492392) 93.5542 ± 0.302686 (W 0.0485448) 🔵 tlinalg-pool 12.01× faster than lapack-openblas
svd_thin/n32/b1024 33.4038 ± 0.220881 (W 0.0746673) 295.416 ± 0.603145 (W 0.567471) 🔵 tlinalg-pool 8.84× faster than lapack-openblas
svd_thin/n4/b1024 5.22826 ± 0.0122426 (W 0.00319908) 7.82854 ± 0.0414143 (W 0.00776528) 🔵 tlinalg-pool 1.50× faster than lapack-openblas
svd_thin/n8/b1024 2.18035 ± 0.08133 (W 0.0596297) 24.8259 ± 0.0480017 (W 0.0190909) 🔵 tlinalg-pool 11.39× faster than lapack-openblas ! between-start range over 5%
svd_values/n16/b1024 5.943 ± 0.102397 (W 0.0234551) 33.3801 ± 0.104509 (W 0.0138868) 🔵 tlinalg-pool 5.62× faster than lapack-openblas
svd_values/n32/b1024 22.6985 ± 0.107268 (W 0.0561621) 109.875 ± 0.388918 (W 0.276549) 🔵 tlinalg-pool 4.84× faster than lapack-openblas
svd_values/n4/b1024 4.01723 ± 0.0233529 (W 0.00336322) 4.56732 ± 0.0496255 (W 0.00361919) 🔵 tlinalg-pool 1.14× faster than lapack-openblas
svd_values/n8/b1024 1.67349 ± 0.00401093 (W 0.00908118) 11.1841 ± 0.0334278 (W 0.0155873) 🔵 tlinalg-pool 6.68× faster than lapack-openblas
triangular_solve/n16/b1024 0.137882 ± 0.00259031 (W 0.00213711) 1.25651 ± 0.0587948 (W 0.00533689) 🔵 tlinalg-pool 9.11× faster than lapack-openblas ! between-start range over 5%
triangular_solve/n32/b1024 0.636164 ± 0.0202597 (W 0.0110732) 2.9501 ± 0.0585679 (W 0.0102641) 🔵 tlinalg-pool 4.64× faster than lapack-openblas ! between-start range over 5%
triangular_solve/n4/b1024 0.164791 ± 0.0017035 (W 0.00196342) 0.746134 ± 0.0586288 (W 0.00381375) 🔵 tlinalg-pool 4.53× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
triangular_solve/n8/b1024 0.0671645 ± 0.00435348 (W 0.000638852) 1.02235 ± 0.0430839 (W 0.00438948) 🔵 tlinalg-pool 15.22× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%

canary-factor, f64, 8T (CPU 4-11)

case tlinalg-pool (ms) lapack-openblas (ms) vs lapack-openblas notes
cholesky/n128/b1 0.064355 ± 0.00133556 (W 0.00158861) 0.078426 ± 0.00133018 (W 0.000764699) 🔵 tlinalg-pool 1.22× faster than lapack-openblas
lu/n128/b1 0.175997 ± 0.00474629 (W 0.00402909) 0.145873 ± 0.000607088 (W 0.00160282) 🔴 tlinalg-pool 1.21× slower than lapack-openblas ! between-start range over 5%
lu_factor/n128/b1 0.118411 ± 0.00372906 (W 0.00525971) 0.130478 ± 9.71494e-05 (W 0.00107223) 🔵 tlinalg-pool 1.10× faster than lapack-openblas ! between-start range over 5%
lu_factor_solve/n128/b1 0.124997 ± 0.00160512 (W 0.00294206) 0.147056 ± 0.000290747 (W 0.000813873) 🔵 tlinalg-pool 1.18× faster than lapack-openblas
lu_solve_prepared/n128/b1 0.0102895 ± 0.000135377 (W 8.28236e-05) 0.012278 ± 5.058e-05 (W 0.000116698) 🔵 tlinalg-pool 1.19× faster than lapack-openblas dispatch-dominated (row under 20 us)
solve/n128/b1 0.138524 ± 0.0015091 (W 0.00196264) 0.154244 ± 0.000925326 (W 0.00102653) 🔵 tlinalg-pool 1.11× faster than lapack-openblas
svd_values/n128/b1 2.37734 ± 0.00253974 (W 0.00387491) 1.37925 ± 0.00208043 (W 0.00260626) 🔴 tlinalg-pool 1.72× slower than lapack-openblas
triangular_solve/n128/b1 0.0070835 ± 2.97811e-05 (W 4.04742e-05) 0.006812 ± 8.53913e-05 (W 0.000138746) · dispatch-dominated (row under 20 us)

canary-factor, c64, 8T (CPU 4-11)

case tlinalg-pool (ms) lapack-openblas (ms) vs lapack-openblas notes
cholesky/n128/b1 0.185707 ± 0.00121489 (W 0.0038702) 0.149728 ± 0.00119777 (W 0.00164674) 🔴 tlinalg-pool 1.24× slower than lapack-openblas
lu/n128/b1 0.340106 ± 0.000655558 (W 0.00520187) 0.234576 ± 0.00217091 (W 0.00158875) 🔴 tlinalg-pool 1.45× slower than lapack-openblas
lu_factor/n128/b1 0.265015 ± 0.00144075 (W 0.00260541) 0.191962 ± 0.00107912 (W 0.00196132) 🔴 tlinalg-pool 1.38× slower than lapack-openblas
lu_factor_solve/n128/b1 0.300068 ± 0.00322166 (W 0.00415643) 0.22598 ± 0.000869681 (W 0.00181596) 🔴 tlinalg-pool 1.33× slower than lapack-openblas
lu_solve_prepared/n128/b1 0.031774 ± 0.000112132 (W 8.05419e-05) 0.029896 ± 0.000547334 (W 0.000157844) 🔴 tlinalg-pool 1.06× slower than lapack-openblas
solve/n128/b1 0.318307 ± 0.00122954 (W 0.00348759) 0.234399 ± 0.00105116 (W 0.00139477) 🔴 tlinalg-pool 1.36× slower than lapack-openblas
svd_values/n128/b1 3.24942 ± 0.00994004 (W 0.00921002) 2.57648 ± 0.00754251 (W 0.00352385) 🔴 tlinalg-pool 1.26× slower than lapack-openblas
triangular_solve/n128/b1 0.0186745 ± 5.58592e-05 (W 7.49325e-05) 0.013335 ± 0.000418809 (W 7.57525e-05) 🔴 tlinalg-pool 1.40× slower than lapack-openblas ! between-start range over 5%; dispatch-dominated (row under 20 us)

canary-decompose, f64, 8T (CPU 4-11)

case tlinalg-pool (ms) lapack-openblas (ms) vs lapack-openblas notes
eigh/n512/b1 29.0606 ± 0.0476514 (W 0.0309587) 23.4437 ± 0.0966878 (W 0.0201093) 🔴 tlinalg-pool 1.24× slower than lapack-openblas
eigvalsh/n512/b1 17.6059 ± 0.0238696 (W 0.030695) 11.0839 ± 0.036557 (W 0.00404075) 🔴 tlinalg-pool 1.59× slower than lapack-openblas
householder_apply/n512/b1 1.77507 ± 0.00929193 (W 0.0081583) 9.85781 ± 0.0591743 (W 0.0133266) 🔵 tlinalg-pool 5.55× faster than lapack-openblas
householder_factor/n512/b1 2.06826 ± 0.012508 (W 0.0145947) 6.78018 ± 0.0410363 (W 0.0147275) 🔵 tlinalg-pool 3.28× faster than lapack-openblas
qr/n512/b1 6.14632 ± 0.026018 (W 0.0114485) 13.5395 ± 0.0764982 (W 0.0244065) 🔵 tlinalg-pool 2.20× faster than lapack-openblas
rank_revealing_qr/n512/b1 8.60556 ± 0.0557393 (W 0.0390378) 23.2618 ± 0.0763545 (W 0.0829385) 🔵 tlinalg-pool 2.70× faster than lapack-openblas
svd_thin/n512/b1 60.069 ± 0.368649 (W 0.0909446) 84.6495 ± 0.549728 (W 0.123854) 🔵 tlinalg-pool 1.41× faster than lapack-openblas

canary-decompose, c64, 8T (CPU 4-11)

case tlinalg-pool (ms) lapack-openblas (ms) vs lapack-openblas notes
eigh/n512/b1 39.9188 ± 0.36905 (W 0.363537) 205.463 ± 1.14552 (W 0.405502) 🔵 tlinalg-pool 5.15× faster than lapack-openblas
eigvalsh/n512/b1 23.6418 ± 0.0569227 (W 0.0479428) 29.8076 ± 0.0570086 (W 0.0173851) 🔵 tlinalg-pool 1.26× faster than lapack-openblas
householder_apply/n512/b1 5.84963 ± 0.0221755 (W 0.0164069) 14.4934 ± 0.173237 (W 0.0181259) 🔵 tlinalg-pool 2.48× faster than lapack-openblas
householder_factor/n512/b1 5.54873 ± 0.0227664 (W 0.00830392) 10.7518 ± 0.0219535 (W 0.0140543) 🔵 tlinalg-pool 1.94× faster than lapack-openblas
qr/n512/b1 14.983 ± 0.0233723 (W 0.0207769) 21.7938 ± 0.109075 (W 0.0248756) 🔵 tlinalg-pool 1.45× faster than lapack-openblas
rank_revealing_qr/n512/b1 16.8923 ± 0.0317839 (W 0.0170109) 24.4289 ± 0.0977114 (W 0.0479901) 🔵 tlinalg-pool 1.45× faster than lapack-openblas
svd_thin/n512/b1 95.7224 ± 0.430571 (W 0.263016) 97.7779 ± 0.712128 (W 0.0826973) ·

canary-rectangular, f64, 8T (CPU 4-11)

case tlinalg-pool (ms) lapack-openblas (ms) vs lapack-openblas notes
householder_factor/4096x256/b1 6.93783 ± 0.0214457 (W 0.0158027) 17.9967 ± 0.104293 (W 0.0232321) 🔵 tlinalg-pool 2.59× faster than lapack-openblas
qr/4096x256/b1 23.593 ± 0.0920066 (W 0.0622897) 35.5275 ± 0.413347 (W 0.03468) 🔵 tlinalg-pool 1.51× faster than lapack-openblas
rank_revealing_qr/4096x256/b1 24.5624 ± 0.0657247 (W 0.0362153) 38.8237 ± 0.218431 (W 0.0505704) 🔵 tlinalg-pool 1.58× faster than lapack-openblas
svd_thin/4096x256/b1 40.9522 ± 0.134037 (W 0.128143) 60.6598 ± 0.618124 (W 0.0548918) 🔵 tlinalg-pool 1.48× faster than lapack-openblas

canary-rectangular, c64, 8T (CPU 4-11)

case tlinalg-pool (ms) lapack-openblas (ms) vs lapack-openblas notes
householder_factor/4096x256/b1 18.8769 ± 0.0770746 (W 0.0237199) 20.1445 ± 0.0222079 (W 0.0229127) 🔵 tlinalg-pool 1.07× faster than lapack-openblas
qr/4096x256/b1 55.3953 ± 0.741779 (W 0.804664) 40.1408 ± 0.135086 (W 0.0468689) 🔴 tlinalg-pool 1.38× slower than lapack-openblas
rank_revealing_qr/4096x256/b1 61.7249 ± 0.708337 (W 0.928784) 51.3527 ± 0.116022 (W 0.113671) 🔴 tlinalg-pool 1.20× slower than lapack-openblas
svd_thin/4096x256/b1 111.587 ± 0.660808 (W 1.17114) 85.4027 ± 0.201127 (W 0.0476096) 🔴 tlinalg-pool 1.31× slower than lapack-openblas

Repeated process sets describe observed repeatability under the recorded order schedule; they are not a paired baseline/candidate experiment and do not establish equal performance between engines.