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 1.45%, worst 95.69% among cells of 1 us or more (a sub-microsecond row is timer and scheduler jitter, not kernel time); 70 of 396 engine cells exceed 5% and carry !; median within-process relative SD 0.47%. 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-mkl (ms) vs lapack-mkl notes
cholesky/n16/b1024 0.121606 ± 0.00237869 (W 0.00154702) 1.056 ± 0.00499906 (W 0.00262263) 🔵 tlinalg-pool 8.68× faster than lapack-mkl
cholesky/n32/b1024 0.462256 ± 0.213866 (W 0.0122161) 3.68226 ± 0.017886 (W 0.0182962) 🔵 tlinalg-pool 7.97× faster than lapack-mkl ! between-start range over 5%; p position effect over 5%
cholesky/n4/b1024 0.030006 ± 0.000192588 (W 0.000887028) 0.120394 ± 0.00426816 (W 0.00142392) 🔵 tlinalg-pool 4.01× faster than lapack-mkl ! between-start range over 5%
cholesky/n8/b1024 0.0536445 ± 0.000393714 (W 0.00103241) 0.322405 ± 0.00308103 (W 0.00208426) 🔵 tlinalg-pool 6.01× faster than lapack-mkl
eig/n16/b1024 11.9316 ± 0.0168178 (W 0.0272405) 52.7762 ± 0.0626004 (W 0.067591) 🔵 tlinalg-pool 4.42× faster than lapack-mkl
eig/n32/b1024 53.2071 ± 0.117634 (W 0.188094) 283.706 ± 0.949039 (W 0.188217) 🔵 tlinalg-pool 5.33× faster than lapack-mkl
eig/n4/b1024 0.785547 ± 0.0200052 (W 0.00521261) 2.83342 ± 0.00299599 (W 0.00551067) 🔵 tlinalg-pool 3.61× faster than lapack-mkl ! between-start range over 5%
eig/n8/b1024 2.94833 ± 0.0362855 (W 0.00737106) 11.675 ± 0.067638 (W 0.0301321) 🔵 tlinalg-pool 3.96× faster than lapack-mkl
eigh/n16/b1024 3.74781 ± 0.0152911 (W 0.00567988) 23.4827 ± 0.0893444 (W 0.0540381) 🔵 tlinalg-pool 6.27× faster than lapack-mkl
eigh/n32/b1024 14.3294 ± 0.0779087 (W 0.0212585) 100.663 ± 0.325364 (W 0.352732) 🔵 tlinalg-pool 7.02× faster than lapack-mkl
eigh/n4/b1024 0.318563 ± 0.0028685 (W 0.0019919) 2.82322 ± 0.0118562 (W 0.00700393) 🔵 tlinalg-pool 8.86× faster than lapack-mkl
eigh/n8/b1024 1.07701 ± 0.00623048 (W 0.00257358) 7.73299 ± 0.0295757 (W 0.0157157) 🔵 tlinalg-pool 7.18× faster than lapack-mkl
eigvals/n16/b1024 8.16946 ± 0.047101 (W 0.0165768) 36.5073 ± 0.0826389 (W 0.0472979) 🔵 tlinalg-pool 4.47× faster than lapack-mkl
eigvals/n32/b1024 36.3321 ± 0.0349917 (W 0.099252) 169.323 ± 0.205683 (W 0.0950808) 🔵 tlinalg-pool 4.66× faster than lapack-mkl
eigvals/n4/b1024 0.409017 ± 0.00415587 (W 0.00245429) 2.10652 ± 0.00777113 (W 0.00432676) 🔵 tlinalg-pool 5.15× faster than lapack-mkl
eigvals/n8/b1024 1.94011 ± 0.00572955 (W 0.00547737) 8.92663 ± 0.0588898 (W 0.0172059) 🔵 tlinalg-pool 4.60× faster than lapack-mkl
eigvalsh/n16/b1024 2.79746 ± 0.00933589 (W 0.00420352) 12.9566 ± 0.0118618 (W 0.0178961) 🔵 tlinalg-pool 4.63× faster than lapack-mkl
eigvalsh/n32/b1024 10.8985 ± 0.0421512 (W 0.0345427) 43.0437 ± 0.0572277 (W 0.0275762) 🔵 tlinalg-pool 3.95× faster than lapack-mkl
eigvalsh/n4/b1024 0.263185 ± 0.0120034 (W 0.00214293) 1.45673 ± 0.00545739 (W 0.00458291) 🔵 tlinalg-pool 5.54× faster than lapack-mkl ! between-start range over 5%
eigvalsh/n8/b1024 0.858198 ± 0.00135313 (W 0.00237775) 4.39312 ± 0.00777091 (W 0.00964448) 🔵 tlinalg-pool 5.12× faster than lapack-mkl
full_piv_lu/n16/b1024 0.483691 ± 0.00730585 (W 0.007675) 3.91803 ± 0.0507473 (W 0.00812744) 🔵 tlinalg-pool 8.10× faster than lapack-mkl
full_piv_lu/n32/b1024 2.00754 ± 0.0334886 (W 0.022313) 26.2629 ± 0.0668475 (W 0.0341475) 🔵 tlinalg-pool 13.08× faster than lapack-mkl
full_piv_lu/n4/b1024 0.0831855 ± 0.000814367 (W 0.00241145) 0.40023 ± 0.0019237 (W 0.0029399) 🔵 tlinalg-pool 4.81× faster than lapack-mkl
full_piv_lu/n8/b1024 0.162934 ± 0.00161981 (W 0.00352367) 0.984033 ± 0.00918252 (W 0.00253502) 🔵 tlinalg-pool 6.04× faster than lapack-mkl
full_piv_lu_solve/n16/b1024 0.45119 ± 0.00839727 (W 0.00437766) 5.77847 ± 0.028728 (W 0.0108206) 🔵 tlinalg-pool 12.81× faster than lapack-mkl
full_piv_lu_solve/n32/b1024 1.47911 ± 0.0196214 (W 0.0160905) 29.6368 ± 0.0897295 (W 0.0630729) 🔵 tlinalg-pool 20.04× faster than lapack-mkl
full_piv_lu_solve/n4/b1024 0.0748145 ± 0.00103573 (W 0.00294667) 1.11298 ± 0.00477068 (W 0.0062395) 🔵 tlinalg-pool 14.88× faster than lapack-mkl
full_piv_lu_solve/n8/b1024 0.172626 ± 0.00632894 (W 0.00389568) 1.93897 ± 0.0107034 (W 0.00453228) 🔵 tlinalg-pool 11.23× faster than lapack-mkl ! between-start range over 5%
householder_apply/n16/b1024 0.616158 ± 0.00625151 (W 0.00357615) 2.66743 ± 0.014905 (W 0.0184974) 🔵 tlinalg-pool 4.33× faster than lapack-mkl
householder_apply/n32/b1024 2.46837 ± 0.0548104 (W 0.00771775) 13.0572 ± 0.151938 (W 0.0751233) 🔵 tlinalg-pool 5.29× faster than lapack-mkl
householder_apply/n4/b1024 0.0544015 ± 0.00130761 (W 0.000547492) 0.323087 ± 0.00211473 (W 0.00268647) 🔵 tlinalg-pool 5.94× faster than lapack-mkl ! between-start range over 5%
householder_apply/n8/b1024 0.167156 ± 0.00166951 (W 0.00169292) 0.802088 ± 0.00313458 (W 0.00838076) 🔵 tlinalg-pool 4.80× faster than lapack-mkl
householder_factor/n16/b1024 0.483381 ± 0.00571188 (W 0.00282935) 2.95332 ± 0.0107464 (W 0.00609265) 🔵 tlinalg-pool 6.11× faster than lapack-mkl
householder_factor/n32/b1024 1.68564 ± 0.0151862 (W 0.010412) 12.3122 ± 0.0795421 (W 0.0377682) 🔵 tlinalg-pool 7.30× faster than lapack-mkl
householder_factor/n4/b1024 0.0646655 ± 0.0133385 (W 0.000502017) 0.306296 ± 0.000843377 (W 0.0019344) 🔵 tlinalg-pool 4.74× faster than lapack-mkl ! between-start range over 5%; p position effect over 5%
householder_factor/n8/b1024 0.16671 ± 0.00164425 (W 0.00170395) 0.844823 ± 0.00193488 (W 0.00165113) 🔵 tlinalg-pool 5.07× faster than lapack-mkl
lu/n16/b1024 0.360947 ± 0.00620923 (W 0.00791757) 1.74714 ± 0.054895 (W 0.0095497) 🔵 tlinalg-pool 4.84× faster than lapack-mkl ! between-start range over 5%
lu/n32/b1024 1.2934 ± 0.0476024 (W 0.0202452) 11.7919 ± 0.03303 (W 0.0252065) 🔵 tlinalg-pool 9.12× faster than lapack-mkl ! between-start range over 5%
lu/n4/b1024 0.072881 ± 0.00237852 (W 0.00279209) 0.167577 ± 0.000955277 (W 0.000798285) 🔵 tlinalg-pool 2.30× faster than lapack-mkl ! between-start range over 5%
lu/n8/b1024 0.143387 ± 0.00242898 (W 0.00249593) 0.499109 ± 0.00880948 (W 0.00261349) 🔵 tlinalg-pool 3.48× faster than lapack-mkl
lu_factor/n16/b1024 0.261622 ± 0.00142314 (W 0.00181453) 1.19426 ± 0.0321788 (W 0.00235867) 🔵 tlinalg-pool 4.56× faster than lapack-mkl ! between-start range over 5%
lu_factor/n32/b1024 0.754455 ± 0.00237821 (W 0.00461927) 8.97074 ± 0.0369859 (W 0.0243152) 🔵 tlinalg-pool 11.89× faster than lapack-mkl
lu_factor/n4/b1024 0.0616795 ± 0.00135633 (W 0.00192381) 0.0665495 ± 0.00332041 (W 0.00142339) 🔵 tlinalg-pool 1.08× faster than lapack-mkl ! between-start range over 5%
lu_factor/n8/b1024 0.108051 ± 0.000605489 (W 0.00147) 0.26352 ± 0.00205984 (W 0.00216778) 🔵 tlinalg-pool 2.44× faster than lapack-mkl
lu_factor_solve/n16/b1024 0.335093 ± 0.00297363 (W 0.00233134) 2.38985 ± 0.0301702 (W 0.0114445) 🔵 tlinalg-pool 7.13× faster than lapack-mkl
lu_factor_solve/n32/b1024 0.913478 ± 0.00754984 (W 0.00402883) 13.9445 ± 0.0495587 (W 0.0415731) 🔵 tlinalg-pool 15.27× faster than lapack-mkl
lu_factor_solve/n4/b1024 0.0653015 ± 0.00175172 (W 0.00263859) 0.182671 ± 0.00175467 (W 0.00148126) 🔵 tlinalg-pool 2.80× faster than lapack-mkl ! between-start range over 5%
lu_factor_solve/n8/b1024 0.130684 ± 0.00482292 (W 0.00142943) 0.828128 ± 0.00331211 (W 0.00223364) 🔵 tlinalg-pool 6.34× faster than lapack-mkl ! between-start range over 5%
lu_solve_prepared/n16/b1024 0.130608 ± 0.000986676 (W 0.000902833) 1.06414 ± 0.0220334 (W 0.00546557) 🔵 tlinalg-pool 8.15× faster than lapack-mkl
lu_solve_prepared/n32/b1024 0.314947 ± 0.00505309 (W 0.0102524) 4.56719 ± 0.0320861 (W 0.0273012) 🔵 tlinalg-pool 14.50× faster than lapack-mkl
lu_solve_prepared/n4/b1024 0.032581 ± 0.00154034 (W 0.00101987) 0.133524 ± 0.000917668 (W 0.000760623) 🔵 tlinalg-pool 4.10× faster than lapack-mkl ! between-start range over 5%
lu_solve_prepared/n8/b1024 0.055568 ± 0.000468844 (W 0.00136576) 0.586542 ± 0.00201042 (W 0.0026807) 🔵 tlinalg-pool 10.56× faster than lapack-mkl
qr/n16/b1024 1.14411 ± 0.0229985 (W 0.0047767) 5.26367 ± 0.0203896 (W 0.0138891) 🔵 tlinalg-pool 4.60× faster than lapack-mkl
qr/n32/b1024 4.52839 ± 0.0387685 (W 0.0281934) 21.0132 ± 0.0528052 (W 0.0365718) 🔵 tlinalg-pool 4.64× faster than lapack-mkl
qr/n4/b1024 0.105401 ± 0.000586171 (W 0.000607712) 0.724506 ± 0.0303314 (W 0.00325838) 🔵 tlinalg-pool 6.87× faster than lapack-mkl ! between-start range over 5%
qr/n8/b1024 0.309922 ± 0.016623 (W 0.00184263) 1.68542 ± 0.0108696 (W 0.0042348) 🔵 tlinalg-pool 5.44× faster than lapack-mkl ! between-start range over 5%
rank_revealing_qr/n16/b1024 1.30287 ± 0.0216488 (W 0.00418838) 8.3419 ± 0.0566898 (W 0.0185861) 🔵 tlinalg-pool 6.40× faster than lapack-mkl
rank_revealing_qr/n32/b1024 4.60245 ± 0.0765412 (W 0.0209092) 31.0211 ± 0.0431944 (W 0.0591707) 🔵 tlinalg-pool 6.74× faster than lapack-mkl
rank_revealing_qr/n4/b1024 0.160384 ± 0.00231577 (W 0.00177472) 1.22276 ± 0.0111805 (W 0.00330205) 🔵 tlinalg-pool 7.62× faster than lapack-mkl
rank_revealing_qr/n8/b1024 0.420799 ± 0.00417054 (W 0.00337789) 2.91399 ± 0.00612433 (W 0.00714062) 🔵 tlinalg-pool 6.92× faster than lapack-mkl
solve/n16/b1024 0.354386 ± 0.00362849 (W 0.00236981) 2.42507 ± 0.0211184 (W 0.00504438) 🔵 tlinalg-pool 6.84× faster than lapack-mkl
solve/n32/b1024 0.944541 ± 0.00723642 (W 0.0148452) 13.4746 ± 0.0472758 (W 0.0210078) 🔵 tlinalg-pool 14.27× faster than lapack-mkl
solve/n4/b1024 0.054887 ± 0.000824336 (W 0.0010276) 0.23129 ± 0.00364693 (W 0.00205204) 🔵 tlinalg-pool 4.21× faster than lapack-mkl
solve/n8/b1024 0.132221 ± 0.0020718 (W 0.00268931) 0.89272 ± 0.00278028 (W 0.00248544) 🔵 tlinalg-pool 6.75× faster than lapack-mkl
svd_full/n16/b1024 7.11742 ± 0.504694 (W 0.112523) 51.5429 ± 0.100078 (W 0.0559778) 🔵 tlinalg-pool 7.24× faster than lapack-mkl ! between-start range over 5%
svd_full/n32/b1024 26.8402 ± 1.04279 (W 0.329673) 194.433 ± 0.372513 (W 0.2389) 🔵 tlinalg-pool 7.24× faster than lapack-mkl ! between-start range over 5%
svd_full/n4/b1024 0.565664 ± 0.0194703 (W 0.0236725) 6.09128 ± 0.0189407 (W 0.0133952) 🔵 tlinalg-pool 10.77× faster than lapack-mkl ! between-start range over 5%
svd_full/n8/b1024 1.98392 ± 0.181571 (W 0.0126685) 18.5901 ± 0.0397399 (W 0.0315521) 🔵 tlinalg-pool 9.37× faster than lapack-mkl ! between-start range over 5%
svd_thin/n16/b1024 7.22998 ± 0.130381 (W 0.104104) 51.3581 ± 0.0924534 (W 0.0548919) 🔵 tlinalg-pool 7.10× faster than lapack-mkl
svd_thin/n32/b1024 25.8891 ± 0.212137 (W 0.0939896) 194.618 ± 0.555514 (W 0.202059) 🔵 tlinalg-pool 7.52× faster than lapack-mkl
svd_thin/n4/b1024 0.583554 ± 0.0427739 (W 0.00269515) 6.08683 ± 0.00901789 (W 0.00629693) 🔵 tlinalg-pool 10.43× faster than lapack-mkl ! between-start range over 5%
svd_thin/n8/b1024 1.9441 ± 0.183939 (W 0.00483177) 18.5729 ± 0.0566969 (W 0.0323642) 🔵 tlinalg-pool 9.55× faster than lapack-mkl ! between-start range over 5%
svd_values/n16/b1024 5.18946 ± 0.0467734 (W 0.0253505) 19.2555 ± 0.0360632 (W 0.0245227) 🔵 tlinalg-pool 3.71× faster than lapack-mkl
svd_values/n32/b1024 20.1271 ± 0.0597423 (W 0.0193201) 64.7517 ± 0.0929919 (W 0.227805) 🔵 tlinalg-pool 3.22× faster than lapack-mkl
svd_values/n4/b1024 0.412333 ± 0.00215029 (W 0.00532137) 2.67339 ± 0.0108872 (W 0.00613369) 🔵 tlinalg-pool 6.48× faster than lapack-mkl
svd_values/n8/b1024 1.45567 ± 0.0159755 (W 0.00597205) 6.68622 ± 0.0128886 (W 0.012011) 🔵 tlinalg-pool 4.59× faster than lapack-mkl
triangular_solve/n16/b1024 0.0854245 ± 0.00087811 (W 0.00254948) 0.594836 ± 0.00797623 (W 0.00376138) 🔵 tlinalg-pool 6.96× faster than lapack-mkl
triangular_solve/n32/b1024 0.189663 ± 0.00263774 (W 0.00563782) 2.44578 ± 0.0179001 (W 0.0130905) 🔵 tlinalg-pool 12.90× faster than lapack-mkl
triangular_solve/n4/b1024 0.0189505 ± 0.00709933 (W 0.000805262) 0.294478 ± 0.00747112 (W 0.00254375) 🔵 tlinalg-pool 15.54× faster than lapack-mkl ! between-start range over 5%; p position effect over 5%; dispatch-dominated (row under 20 us)
triangular_solve/n8/b1024 0.03496 ± 0.000188229 (W 0.000623236) 0.397275 ± 0.00448465 (W 0.00289698) 🔵 tlinalg-pool 11.36× faster than lapack-mkl

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

case tlinalg-pool (ms) lapack-mkl (ms) vs lapack-mkl notes
cholesky/n16/b1024 0.173364 ± 0.00322754 (W 0.00303411) 1.79841 ± 0.0172309 (W 0.0143502) 🔵 tlinalg-pool 10.37× faster than lapack-mkl
cholesky/n32/b1024 0.808486 ± 0.0162844 (W 0.0042831) 8.8493 ± 0.025694 (W 0.0252183) 🔵 tlinalg-pool 10.95× faster than lapack-mkl
cholesky/n4/b1024 0.032721 ± 0.000564251 (W 0.000453778) 0.178227 ± 0.000737568 (W 0.00165621) 🔵 tlinalg-pool 5.45× faster than lapack-mkl
cholesky/n8/b1024 0.0604275 ± 0.00157147 (W 0.000936475) 0.470487 ± 0.0119452 (W 0.00302609) 🔵 tlinalg-pool 7.79× faster than lapack-mkl ! between-start range over 5%
eig/n16/b1024 16.4433 ± 0.0333485 (W 0.0737915) 175.394 ± 0.193957 (W 0.311316) 🔵 tlinalg-pool 10.67× faster than lapack-mkl
eig/n32/b1024 67.3984 ± 0.33371 (W 0.181206) 793.593 ± 1.13117 (W 1.05096) 🔵 tlinalg-pool 11.77× faster than lapack-mkl
eig/n4/b1024 1.30493 ± 0.0100795 (W 0.0141901) 12.4539 ± 0.0223056 (W 0.0203771) 🔵 tlinalg-pool 9.54× faster than lapack-mkl
eig/n8/b1024 4.46184 ± 0.0311922 (W 0.0172622) 43.2435 ± 0.0674282 (W 0.0401267) 🔵 tlinalg-pool 9.69× faster than lapack-mkl
eigh/n16/b1024 4.06105 ± 0.477726 (W 0.0313327) 32.2381 ± 0.184573 (W 0.0279093) 🔵 tlinalg-pool 7.94× faster than lapack-mkl ! between-start range over 5%; p position effect over 5%
eigh/n32/b1024 17.9443 ± 0.0691245 (W 0.0270159) 152.062 ± 0.405791 (W 0.233277) 🔵 tlinalg-pool 8.47× faster than lapack-mkl
eigh/n4/b1024 0.378531 ± 0.004092 (W 0.00184677) 2.82244 ± 0.0347823 (W 0.00542107) 🔵 tlinalg-pool 7.46× faster than lapack-mkl
eigh/n8/b1024 1.15495 ± 0.00335156 (W 0.00213873) 8.91985 ± 0.00959031 (W 0.0137286) 🔵 tlinalg-pool 7.72× faster than lapack-mkl
eigvals/n16/b1024 9.69366 ± 0.0161279 (W 0.0276532) 124.922 ± 0.0316303 (W 0.0808156) 🔵 tlinalg-pool 12.89× faster than lapack-mkl
eigvals/n32/b1024 42.4887 ± 0.0728142 (W 0.183488) 523.51 ± 0.365211 (W 0.538437) 🔵 tlinalg-pool 12.32× faster than lapack-mkl
eigvals/n4/b1024 0.613599 ± 0.0114077 (W 0.010526) 9.56403 ± 0.00625228 (W 0.0156493) 🔵 tlinalg-pool 15.59× faster than lapack-mkl
eigvals/n8/b1024 2.54607 ± 0.237655 (W 0.00853226) 34.3614 ± 0.0444391 (W 0.0257885) 🔵 tlinalg-pool 13.50× faster than lapack-mkl ! between-start range over 5%; p position effect over 5%
eigvalsh/n16/b1024 3.03481 ± 0.0573121 (W 0.00798423) 18.3868 ± 0.00784189 (W 0.0193171) 🔵 tlinalg-pool 6.06× faster than lapack-mkl
eigvalsh/n32/b1024 11.917 ± 0.0806169 (W 0.0555404) 77.4789 ± 0.0173485 (W 0.0355516) 🔵 tlinalg-pool 6.50× faster than lapack-mkl
eigvalsh/n4/b1024 0.309251 ± 0.00172078 (W 0.00172928) 1.78252 ± 0.00409837 (W 0.00334899) 🔵 tlinalg-pool 5.76× faster than lapack-mkl
eigvalsh/n8/b1024 0.929762 ± 0.00357479 (W 0.00270853) 5.58392 ± 0.00774665 (W 0.00911304) 🔵 tlinalg-pool 6.01× faster than lapack-mkl
full_piv_lu/n16/b1024 0.621653 ± 0.00834721 (W 0.00762647) 5.57642 ± 0.117981 (W 0.0150546) 🔵 tlinalg-pool 8.97× faster than lapack-mkl
full_piv_lu/n32/b1024 3.04533 ± 0.0328277 (W 0.0378511) 27.0181 ± 0.140721 (W 0.047077) 🔵 tlinalg-pool 8.87× faster than lapack-mkl
full_piv_lu/n4/b1024 0.088395 ± 0.000742232 (W 0.00212617) 0.519534 ± 0.00426149 (W 0.00283793) 🔵 tlinalg-pool 5.88× faster than lapack-mkl
full_piv_lu/n8/b1024 0.196401 ± 0.002794 (W 0.00379327) 1.31036 ± 0.0233579 (W 0.00324217) 🔵 tlinalg-pool 6.67× faster than lapack-mkl
full_piv_lu_solve/n16/b1024 0.701342 ± 0.176973 (W 0.00814064) 8.44998 ± 0.0416518 (W 0.0205723) 🔵 tlinalg-pool 12.05× faster than lapack-mkl ! between-start range over 5%; p position effect over 5%
full_piv_lu_solve/n32/b1024 2.65749 ± 0.0112326 (W 0.00627269) 34.5518 ± 0.0364858 (W 0.0415497) 🔵 tlinalg-pool 13.00× faster than lapack-mkl
full_piv_lu_solve/n4/b1024 0.106182 ± 0.0025064 (W 0.00257367) 1.59675 ± 0.00427573 (W 0.00519253) 🔵 tlinalg-pool 15.04× faster than lapack-mkl
full_piv_lu_solve/n8/b1024 0.264273 ± 0.00136225 (W 0.00229533) 3.15822 ± 0.0202753 (W 0.00481074) 🔵 tlinalg-pool 11.95× faster than lapack-mkl
householder_apply/n16/b1024 0.636 ± 0.0192611 (W 0.0102557) 6.19001 ± 0.0893384 (W 0.0229932) 🔵 tlinalg-pool 9.73× faster than lapack-mkl ! between-start range over 5%
householder_apply/n32/b1024 3.52716 ± 0.030043 (W 0.0208449) 37.8521 ± 0.188014 (W 0.0757596) 🔵 tlinalg-pool 10.73× faster than lapack-mkl
householder_apply/n4/b1024 0.0503635 ± 0.00414148 (W 0.00152745) 0.416065 ± 0.00209416 (W 0.00543492) 🔵 tlinalg-pool 8.26× faster than lapack-mkl ! between-start range over 5%
householder_apply/n8/b1024 0.156632 ± 0.00337081 (W 0.00370772) 1.27306 ± 0.0137394 (W 0.00453788) 🔵 tlinalg-pool 8.13× faster than lapack-mkl ! between-start range over 5%
householder_factor/n16/b1024 0.580852 ± 0.00487671 (W 0.00416013) 6.29767 ± 0.0306936 (W 0.0103353) 🔵 tlinalg-pool 10.84× faster than lapack-mkl
householder_factor/n32/b1024 2.6052 ± 0.0334947 (W 0.0218756) 27.4504 ± 0.08103 (W 0.031061) 🔵 tlinalg-pool 10.54× faster than lapack-mkl
householder_factor/n4/b1024 0.075135 ± 0.000446439 (W 0.000542159) 0.516177 ± 0.00301964 (W 0.00236094) 🔵 tlinalg-pool 6.87× faster than lapack-mkl
householder_factor/n8/b1024 0.196511 ± 0.000678453 (W 0.00175495) 1.4871 ± 0.00611152 (W 0.00344517) 🔵 tlinalg-pool 7.57× faster than lapack-mkl
lu/n16/b1024 0.44003 ± 0.00484161 (W 0.00809407) 3.57708 ± 0.211586 (W 0.0175123) 🔵 tlinalg-pool 8.13× faster than lapack-mkl ! between-start range over 5%
lu/n32/b1024 2.15443 ± 0.0445072 (W 0.0183997) 19.7081 ± 0.125036 (W 0.0438559) 🔵 tlinalg-pool 9.15× faster than lapack-mkl
lu/n4/b1024 0.070622 ± 0.0010373 (W 0.00144865) 0.28456 ± 0.0433377 (W 0.00231896) 🔵 tlinalg-pool 4.03× faster than lapack-mkl ! between-start range over 5%; p position effect over 5%
lu/n8/b1024 0.155469 ± 0.000494925 (W 0.00390492) 0.893088 ± 0.0907577 (W 0.00195389) 🔵 tlinalg-pool 5.74× faster than lapack-mkl ! between-start range over 5%
lu_factor/n16/b1024 0.293417 ± 0.00674373 (W 0.0116637) 2.65038 ± 0.0092202 (W 0.0049128) 🔵 tlinalg-pool 9.03× faster than lapack-mkl ! between-start range over 5%
lu_factor/n32/b1024 1.27303 ± 0.0177549 (W 0.0212376) 15.5367 ± 0.0449932 (W 0.0310101) 🔵 tlinalg-pool 12.20× faster than lapack-mkl
lu_factor/n4/b1024 0.069199 ± 0.00233399 (W 0.000857185) 0.182981 ± 0.00053724 (W 0.00175373) 🔵 tlinalg-pool 2.64× faster than lapack-mkl ! between-start range over 5%
lu_factor/n8/b1024 0.116241 ± 0.00142189 (W 0.00252596) 0.621463 ± 0.00166102 (W 0.00267804) 🔵 tlinalg-pool 5.35× faster than lapack-mkl
lu_factor_solve/n16/b1024 0.476187 ± 0.00763988 (W 0.00877875) 5.2816 ± 0.0125948 (W 0.00806216) 🔵 tlinalg-pool 11.09× faster than lapack-mkl
lu_factor_solve/n32/b1024 1.84965 ± 0.0445352 (W 0.0265043) 21.1286 ± 0.0843237 (W 0.0736158) 🔵 tlinalg-pool 11.42× faster than lapack-mkl ! between-start range over 5%
lu_factor_solve/n4/b1024 0.080315 ± 0.00247729 (W 0.00109621) 0.415484 ± 0.00354636 (W 0.00256839) 🔵 tlinalg-pool 5.17× faster than lapack-mkl ! between-start range over 5%
lu_factor_solve/n8/b1024 0.172837 ± 0.0018201 (W 0.00143852) 1.66619 ± 0.00791159 (W 0.00250563) 🔵 tlinalg-pool 9.64× faster than lapack-mkl
lu_solve_prepared/n16/b1024 0.207797 ± 0.0171933 (W 0.00654444) 2.6259 ± 0.0135436 (W 0.010536) 🔵 tlinalg-pool 12.64× faster than lapack-mkl ! between-start range over 5%
lu_solve_prepared/n32/b1024 0.867197 ± 0.01092 (W 0.0172731) 5.66791 ± 0.125035 (W 0.0996096) 🔵 tlinalg-pool 6.54× faster than lapack-mkl ! between-start range over 5%
lu_solve_prepared/n4/b1024 0.0455955 ± 0.000540081 (W 0.00161651) 0.241335 ± 0.00113386 (W 0.00132768) 🔵 tlinalg-pool 5.29× faster than lapack-mkl
lu_solve_prepared/n8/b1024 0.0876935 ± 0.00148239 (W 0.000997389) 1.07515 ± 0.00232468 (W 0.00176766) 🔵 tlinalg-pool 12.26× faster than lapack-mkl
qr/n16/b1024 1.4044 ± 0.0122321 (W 0.00593186) 10.5538 ± 0.0564987 (W 0.0202349) 🔵 tlinalg-pool 7.51× faster than lapack-mkl
qr/n32/b1024 9.33379 ± 0.039464 (W 0.0193785) 49.5435 ± 0.14507 (W 0.0342353) 🔵 tlinalg-pool 5.31× faster than lapack-mkl
qr/n4/b1024 0.117885 ± 0.000417618 (W 0.0010384) 0.92321 ± 0.00364486 (W 0.00247312) 🔵 tlinalg-pool 7.83× faster than lapack-mkl
qr/n8/b1024 0.341927 ± 0.00588858 (W 0.00379584) 2.66796 ± 0.0514307 (W 0.00640667) 🔵 tlinalg-pool 7.80× faster than lapack-mkl
rank_revealing_qr/n16/b1024 1.58768 ± 0.0780705 (W 0.0522101) 15.8708 ± 0.0828307 (W 0.0238319) 🔵 tlinalg-pool 10.00× faster than lapack-mkl ! between-start range over 5%
rank_revealing_qr/n32/b1024 8.61367 ± 0.0102682 (W 0.0483492) 69.4617 ± 0.153003 (W 0.0629084) 🔵 tlinalg-pool 8.06× faster than lapack-mkl
rank_revealing_qr/n4/b1024 0.185044 ± 0.0019961 (W 0.00134591) 1.7521 ± 0.00401051 (W 0.00707548) 🔵 tlinalg-pool 9.47× faster than lapack-mkl
rank_revealing_qr/n8/b1024 0.483617 ± 0.00874206 (W 0.00251864) 4.73709 ± 0.0260353 (W 0.00983401) 🔵 tlinalg-pool 9.80× faster than lapack-mkl
solve/n16/b1024 0.518392 ± 0.00956028 (W 0.0161653) 5.58063 ± 0.0289174 (W 0.00865029) 🔵 tlinalg-pool 10.77× faster than lapack-mkl
solve/n32/b1024 1.81743 ± 0.0097747 (W 0.0101357) 21.4786 ± 0.025381 (W 0.0252269) 🔵 tlinalg-pool 11.82× faster than lapack-mkl
solve/n4/b1024 0.077113 ± 0.00272944 (W 0.0021663) 0.477595 ± 0.00211723 (W 0.00236312) 🔵 tlinalg-pool 6.19× faster than lapack-mkl ! between-start range over 5%
solve/n8/b1024 0.18553 ± 0.00102222 (W 0.00180009) 1.76741 ± 0.0190619 (W 0.00302924) 🔵 tlinalg-pool 9.53× faster than lapack-mkl
svd_full/n16/b1024 7.81511 ± 0.0717807 (W 0.0633612) 67.551 ± 0.0903696 (W 0.060876) 🔵 tlinalg-pool 8.64× faster than lapack-mkl
svd_full/n32/b1024 33.3445 ± 0.106819 (W 0.0274504) 285.392 ± 0.975622 (W 0.593116) 🔵 tlinalg-pool 8.56× faster than lapack-mkl
svd_full/n4/b1024 0.642908 ± 0.00984228 (W 0.00184133) 7.39152 ± 0.0152648 (W 0.0165202) 🔵 tlinalg-pool 11.50× faster than lapack-mkl
svd_full/n8/b1024 2.18547 ± 0.0258755 (W 0.0796521) 22.1487 ± 0.0264823 (W 0.0292237) 🔵 tlinalg-pool 10.13× faster than lapack-mkl
svd_thin/n16/b1024 8.12453 ± 0.596378 (W 0.113471) 67.3997 ± 0.100529 (W 0.0890074) 🔵 tlinalg-pool 8.30× faster than lapack-mkl ! between-start range over 5%; p position effect over 5%
svd_thin/n32/b1024 33.3332 ± 0.747337 (W 0.044308) 286.197 ± 1.04957 (W 0.288326) 🔵 tlinalg-pool 8.59× faster than lapack-mkl
svd_thin/n4/b1024 0.641545 ± 0.0103913 (W 0.00232642) 7.38461 ± 0.0145112 (W 0.0188571) 🔵 tlinalg-pool 11.51× faster than lapack-mkl
svd_thin/n8/b1024 2.16662 ± 0.0540666 (W 0.00577077) 22.1473 ± 0.0353436 (W 0.024648) 🔵 tlinalg-pool 10.22× faster than lapack-mkl ! between-start range over 5%
svd_values/n16/b1024 5.80778 ± 0.10072 (W 0.0172) 28.2591 ± 0.0803377 (W 0.0466292) 🔵 tlinalg-pool 4.87× faster than lapack-mkl
svd_values/n32/b1024 22.3648 ± 0.0194513 (W 0.0290461) 105.96 ± 0.249328 (W 0.0438984) 🔵 tlinalg-pool 4.74× faster than lapack-mkl
svd_values/n4/b1024 0.492733 ± 0.00233315 (W 0.00233874) 3.51227 ± 0.00825161 (W 0.00640414) 🔵 tlinalg-pool 7.13× faster than lapack-mkl
svd_values/n8/b1024 1.65043 ± 0.00453485 (W 0.00628035) 9.11357 ± 0.0273703 (W 0.0182357) 🔵 tlinalg-pool 5.52× faster than lapack-mkl
triangular_solve/n16/b1024 0.142841 ± 0.00384874 (W 0.00740819) 1.33724 ± 0.0384065 (W 0.00366983) 🔵 tlinalg-pool 9.36× faster than lapack-mkl ! between-start range over 5%
triangular_solve/n32/b1024 0.539736 ± 0.252901 (W 0.0113826) 3.56187 ± 0.030996 (W 0.0255159) 🔵 tlinalg-pool 6.60× faster than lapack-mkl ! between-start range over 5%; p position effect over 5%
triangular_solve/n4/b1024 0.0280625 ± 0.000659926 (W 0.000936743) 0.302073 ± 0.000953517 (W 0.00210788) 🔵 tlinalg-pool 10.76× faster than lapack-mkl ! between-start range over 5%
triangular_solve/n8/b1024 0.065492 ± 0.00046021 (W 0.00110029) 0.584052 ± 0.00559501 (W 0.00344394) 🔵 tlinalg-pool 8.92× faster than lapack-mkl

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

case tlinalg-pool (ms) lapack-mkl (ms) vs lapack-mkl notes
cholesky/n128/b1 0.066459 ± 0.000630893 (W 0.000138984) 0.0941155 ± 0.000481287 (W 0.0010573) 🔵 tlinalg-pool 1.42× faster than lapack-mkl
lu/n128/b1 0.185446 ± 0.00385348 (W 0.0062456) 0.136149 ± 0.000407392 (W 0.00112943) 🔴 tlinalg-pool 1.36× slower than lapack-mkl
lu_factor/n128/b1 0.126771 ± 0.00120733 (W 0.00284903) 0.12106 ± 0.000175549 (W 0.00095346) ·
lu_factor_solve/n128/b1 0.134817 ± 0.0011646 (W 0.00429039) 0.144019 ± 0.000851477 (W 0.00167171) 🔵 tlinalg-pool 1.07× faster than lapack-mkl
lu_solve_prepared/n128/b1 0.009928 ± 3.90758e-05 (W 0.000120485) 0.016556 ± 5.25198e-05 (W 0.000297585) 🔵 tlinalg-pool 1.67× faster than lapack-mkl dispatch-dominated (row under 20 us)
solve/n128/b1 0.149193 ± 0.00187656 (W 0.00452737) 0.150891 ± 0.000601184 (W 0.00119282) ·
svd_values/n128/b1 2.38738 ± 0.00651744 (W 0.00262694) 1.13233 ± 0.00443002 (W 0.00302846) 🔴 tlinalg-pool 2.11× slower than lapack-mkl
triangular_solve/n128/b1 0.0068575 ± 7.90907e-05 (W 7.45264e-05) 0.0104045 ± 0.000504383 (W 0.000240707) 🔵 tlinalg-pool 1.52× faster than lapack-mkl ! between-start range over 5%; dispatch-dominated (row under 20 us)

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

case tlinalg-pool (ms) lapack-mkl (ms) vs lapack-mkl notes
cholesky/n128/b1 0.187138 ± 0.000541455 (W 0.00239505) 0.409728 ± 0.00170022 (W 0.00333755) 🔵 tlinalg-pool 2.19× faster than lapack-mkl
lu/n128/b1 0.342293 ± 0.00224635 (W 0.00467664) 0.451326 ± 0.0014239 (W 0.00285133) 🔵 tlinalg-pool 1.32× faster than lapack-mkl
lu_factor/n128/b1 0.262019 ± 0.00149553 (W 0.00305705) 0.415199 ± 0.00127375 (W 0.00298318) 🔵 tlinalg-pool 1.58× faster than lapack-mkl
lu_factor_solve/n128/b1 0.28965 ± 0.00207095 (W 0.00283012) 0.446241 ± 0.00193535 (W 0.00205305) 🔵 tlinalg-pool 1.54× faster than lapack-mkl
lu_solve_prepared/n128/b1 0.0312485 ± 0.000162574 (W 0.00011645) 0.0251265 ± 0.000131226 (W 9.79448e-05) 🔴 tlinalg-pool 1.24× slower than lapack-mkl
solve/n128/b1 0.316584 ± 0.00320948 (W 0.00515612) 0.463139 ± 0.000578367 (W 0.00276384) 🔵 tlinalg-pool 1.46× faster than lapack-mkl
svd_values/n128/b1 3.15856 ± 0.00961659 (W 0.00540219) 1.92098 ± 0.00552202 (W 0.00562493) 🔴 tlinalg-pool 1.64× slower than lapack-mkl
triangular_solve/n128/b1 0.01864 ± 0.000108152 (W 0.000121584) 0.013009 ± 6.89704e-05 (W 9.18845e-05) 🔴 tlinalg-pool 1.43× slower than lapack-mkl dispatch-dominated (row under 20 us)

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

case tlinalg-pool (ms) lapack-mkl (ms) vs lapack-mkl notes
eigh/n512/b1 23.6612 ± 0.10555 (W 0.110916) 15.038 ± 0.0728209 (W 0.0355972) 🔴 tlinalg-pool 1.57× slower than lapack-mkl
eigvalsh/n512/b1 18.5617 ± 0.0338739 (W 0.0346088) 12.6663 ± 0.0200189 (W 0.0127108) 🔴 tlinalg-pool 1.47× slower than lapack-mkl
householder_apply/n512/b1 2.73573 ± 0.0321844 (W 0.0524102) 3.1582 ± 0.0766923 (W 0.0272867) 🔵 tlinalg-pool 1.15× faster than lapack-mkl ! between-start range over 5%
householder_factor/n512/b1 2.66062 ± 0.0242662 (W 0.0329361) 2.1333 ± 0.00375801 (W 0.00374449) 🔴 tlinalg-pool 1.25× slower than lapack-mkl
qr/n512/b1 7.84731 ± 0.0534165 (W 0.077981) 8.74615 ± 0.0145672 (W 0.0249704) 🔵 tlinalg-pool 1.11× faster than lapack-mkl
rank_revealing_qr/n512/b1 9.6927 ± 0.14648 (W 0.144917) 24.6562 ± 0.211599 (W 0.0746781) 🔵 tlinalg-pool 2.54× faster than lapack-mkl
svd_thin/n512/b1 60.7376 ± 0.197726 (W 0.135853) 49.131 ± 0.0549404 (W 0.170158) 🔴 tlinalg-pool 1.24× slower than lapack-mkl

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

case tlinalg-pool (ms) lapack-mkl (ms) vs lapack-mkl notes
eigh/n512/b1 39.1729 ± 0.531057 (W 0.490657) 65.8481 ± 0.086262 (W 0.118165) 🔵 tlinalg-pool 1.68× faster than lapack-mkl
eigvalsh/n512/b1 28.8647 ± 0.0685283 (W 0.0615579) 28.9568 ± 0.806039 (W 0.109505) · ! between-start range over 5%
householder_apply/n512/b1 7.06691 ± 0.033475 (W 0.103459) 8.12216 ± 0.0128529 (W 0.0148617) 🔵 tlinalg-pool 1.15× faster than lapack-mkl
householder_factor/n512/b1 6.94726 ± 0.0697756 (W 0.120228) 5.45011 ± 0.0443646 (W 0.00953294) 🔴 tlinalg-pool 1.27× slower than lapack-mkl
qr/n512/b1 16.7766 ± 0.0802107 (W 0.168827) 15.2289 ± 0.0573613 (W 0.0281039) 🔴 tlinalg-pool 1.10× slower than lapack-mkl
rank_revealing_qr/n512/b1 18.8542 ± 0.120696 (W 0.130801) 59.6579 ± 0.390596 (W 0.141225) 🔵 tlinalg-pool 3.16× faster than lapack-mkl
svd_thin/n512/b1 100.861 ± 0.432426 (W 0.600913) 86.5714 ± 0.1206 (W 0.216242) 🔴 tlinalg-pool 1.17× slower than lapack-mkl

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

case tlinalg-pool (ms) lapack-mkl (ms) vs lapack-mkl notes
householder_factor/4096x256/b1 7.84882 ± 0.982383 (W 0.0352985) 6.14661 ± 0.28496 (W 0.0351048) 🔴 tlinalg-pool 1.28× slower than lapack-mkl ! between-start range over 5%
qr/4096x256/b1 22.5045 ± 1.47175 (W 0.565956) 18.9593 ± 2.43203 (W 0.0872895) 🔴 tlinalg-pool 1.19× slower than lapack-mkl ! between-start range over 5%
rank_revealing_qr/4096x256/b1 24.3732 ± 3.98711 (W 0.467811) 63.204 ± 18.767 (W 1.12433) 🔵 tlinalg-pool 2.59× faster than lapack-mkl ! between-start range over 5%
svd_thin/4096x256/b1 39.6408 ± 1.9777 (W 0.399292) 34.6223 ± 3.91767 (W 0.153239) 🔴 tlinalg-pool 1.14× slower than lapack-mkl ! between-start range over 5%

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

case tlinalg-pool (ms) lapack-mkl (ms) vs lapack-mkl notes
householder_factor/4096x256/b1 19.9705 ± 1.13992 (W 0.0807203) 16.889 ± 2.42121 (W 0.0951264) 🔴 tlinalg-pool 1.18× slower than lapack-mkl ! between-start range over 5%
qr/4096x256/b1 53.5904 ± 1.62775 (W 0.122476) 40.8881 ± 4.13948 (W 0.120474) 🔴 tlinalg-pool 1.31× slower than lapack-mkl ! between-start range over 5%
rank_revealing_qr/4096x256/b1 60.7077 ± 14.6941 (W 0.3409) 162.302 ± 14.7213 (W 0.772922) 🔵 tlinalg-pool 2.67× faster than lapack-mkl ! between-start range over 5%; p position effect over 5%
svd_thin/4096x256/b1 115.649 ± 5.50524 (W 0.483385) 76.0307 ± 4.37265 (W 0.505635) 🔴 tlinalg-pool 1.52× slower than lapack-mkl ! between-start range over 5%

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.