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 1.70%, worst 106.67% among cells of 1 us or more (a sub-microsecond row is timer and scheduler jitter, not kernel time); 84 of 396 engine cells exceed 5% and carry !; median within-process relative SD 0.44%. 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 0.114013 ± 0.00137894 (W 0.00180289) 1.03086 ± 0.396971 (W 0.00210431) 🔵 tlinalg-pool 9.04× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
cholesky/n32/b1024 0.428276 ± 0.00296578 (W 0.00583484) 3.67068 ± 0.0295992 (W 0.00718008) 🔵 tlinalg-pool 8.57× faster than lapack-openblas
cholesky/n4/b1024 0.0299815 ± 0.0046906 (W 0.000992047) 0.197239 ± 0.00128434 (W 0.00151433) 🔵 tlinalg-pool 6.58× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
cholesky/n8/b1024 0.0527985 ± 0.000312541 (W 0.000424114) 0.390842 ± 0.00629782 (W 0.00239212) 🔵 tlinalg-pool 7.40× faster than lapack-openblas
eig/n16/b1024 11.8796 ± 0.02929 (W 0.0242826) 69.0092 ± 0.520932 (W 0.0793014) 🔵 tlinalg-pool 5.81× faster than lapack-openblas
eig/n32/b1024 53.3026 ± 1.97892 (W 0.633898) 324.264 ± 0.944573 (W 0.526396) 🔵 tlinalg-pool 6.08× faster than lapack-openblas ! between-start range over 5%
eig/n4/b1024 0.786482 ± 0.00577498 (W 0.00203535) 5.36275 ± 0.0363158 (W 0.00861698) 🔵 tlinalg-pool 6.82× faster than lapack-openblas
eig/n8/b1024 2.95819 ± 0.00381954 (W 0.00607215) 17.5112 ± 0.0388151 (W 0.0331013) 🔵 tlinalg-pool 5.92× faster than lapack-openblas
eigh/n16/b1024 3.75649 ± 0.00919627 (W 0.00594213) 40.4155 ± 0.336466 (W 0.0442196) 🔵 tlinalg-pool 10.76× faster than lapack-openblas
eigh/n32/b1024 14.4109 ± 0.299952 (W 0.0503679) 119.088 ± 0.730433 (W 0.297502) 🔵 tlinalg-pool 8.26× faster than lapack-openblas
eigh/n4/b1024 0.313926 ± 0.00281245 (W 0.00160513) 3.02856 ± 0.0166907 (W 0.00629503) 🔵 tlinalg-pool 9.65× faster than lapack-openblas
eigh/n8/b1024 1.07769 ± 0.00794658 (W 0.00308622) 10.8526 ± 0.023363 (W 0.0199779) 🔵 tlinalg-pool 10.07× faster than lapack-openblas
eigvals/n16/b1024 8.21531 ± 0.0341403 (W 0.0320197) 49.6174 ± 0.220314 (W 0.0470294) 🔵 tlinalg-pool 6.04× faster than lapack-openblas
eigvals/n32/b1024 36.2867 ± 0.0673943 (W 0.0412926) 218.246 ± 1.50518 (W 0.664291) 🔵 tlinalg-pool 6.01× faster than lapack-openblas
eigvals/n4/b1024 0.414184 ± 0.0101991 (W 0.00285983) 3.39492 ± 0.00638276 (W 0.0100707) 🔵 tlinalg-pool 8.20× faster than lapack-openblas ! between-start range over 5%
eigvals/n8/b1024 1.93992 ± 0.0195387 (W 0.0050936) 13.2285 ± 0.0205827 (W 0.0172476) 🔵 tlinalg-pool 6.82× faster than lapack-openblas
eigvalsh/n16/b1024 2.80272 ± 0.0113614 (W 0.00604842) 14.5783 ± 0.0211683 (W 0.0190993) 🔵 tlinalg-pool 5.20× faster than lapack-openblas
eigvalsh/n32/b1024 10.9364 ± 0.0139394 (W 0.0456235) 50.2066 ± 0.0569314 (W 0.0313319) 🔵 tlinalg-pool 4.59× faster than lapack-openblas
eigvalsh/n4/b1024 0.25408 ± 0.00233577 (W 0.00171765) 1.40364 ± 0.0100715 (W 0.00547328) 🔵 tlinalg-pool 5.52× faster than lapack-openblas
eigvalsh/n8/b1024 0.859099 ± 0.00438704 (W 0.00206387) 4.55921 ± 0.0156754 (W 0.0045994) 🔵 tlinalg-pool 5.31× faster than lapack-openblas
full_piv_lu/n16/b1024 0.458626 ± 0.00210503 (W 0.00367571) 4.70267 ± 0.109874 (W 0.0327222) 🔵 tlinalg-pool 10.25× faster than lapack-openblas
full_piv_lu/n32/b1024 2.02988 ± 0.386092 (W 0.0626466) 28.7661 ± 0.855986 (W 0.268266) 🔵 tlinalg-pool 14.17× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
full_piv_lu/n4/b1024 0.08562 ± 0.00176917 (W 0.00227666) 0.362813 ± 0.0209345 (W 0.002003) 🔵 tlinalg-pool 4.24× faster than lapack-openblas ! between-start range over 5%
full_piv_lu/n8/b1024 0.170755 ± 0.00191589 (W 0.00352671) 1.03613 ± 0.0288043 (W 0.00396397) 🔵 tlinalg-pool 6.07× faster than lapack-openblas ! between-start range over 5%
full_piv_lu_solve/n16/b1024 0.440162 ± 0.0048931 (W 0.00314707) 6.54438 ± 0.0766345 (W 0.0225205) 🔵 tlinalg-pool 14.87× faster than lapack-openblas
full_piv_lu_solve/n32/b1024 1.48402 ± 0.0234438 (W 0.01468) 31.3623 ± 0.113879 (W 0.217989) 🔵 tlinalg-pool 21.13× faster than lapack-openblas
full_piv_lu_solve/n4/b1024 0.0748345 ± 0.00342249 (W 0.00353063) 0.947234 ± 0.00283657 (W 0.00226179) 🔵 tlinalg-pool 12.66× faster than lapack-openblas ! between-start range over 5%
full_piv_lu_solve/n8/b1024 0.165053 ± 0.00362107 (W 0.00270673) 2.00375 ± 0.00431416 (W 0.0043334) 🔵 tlinalg-pool 12.14× faster than lapack-openblas
householder_apply/n16/b1024 0.603914 ± 0.0145007 (W 0.00206329) 4.53172 ± 0.0391948 (W 0.0225201) 🔵 tlinalg-pool 7.50× faster than lapack-openblas
householder_apply/n32/b1024 2.45922 ± 0.0654619 (W 0.00997565) 19.0205 ± 0.277388 (W 0.0569565) 🔵 tlinalg-pool 7.73× faster than lapack-openblas ! between-start range over 5%
householder_apply/n4/b1024 0.052809 ± 0.0238579 (W 0.000836941) 0.41799 ± 0.000968785 (W 0.00210265) 🔵 tlinalg-pool 7.92× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
householder_apply/n8/b1024 0.168339 ± 0.00353626 (W 0.00178244) 1.21052 ± 0.0037573 (W 0.00274804) 🔵 tlinalg-pool 7.19× faster than lapack-openblas
householder_factor/n16/b1024 0.492817 ± 0.0749994 (W 0.00451587) 4.37114 ± 0.0107916 (W 0.00782097) 🔵 tlinalg-pool 8.87× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
householder_factor/n32/b1024 1.6337 ± 0.00896215 (W 0.0111518) 14.7242 ± 0.402435 (W 0.037056) 🔵 tlinalg-pool 9.01× faster than lapack-openblas ! between-start range over 5%
householder_factor/n4/b1024 0.063929 ± 0.000441105 (W 0.000648369) 0.537723 ± 0.00143767 (W 0.00199268) 🔵 tlinalg-pool 8.41× faster than lapack-openblas
householder_factor/n8/b1024 0.164834 ± 0.0011015 (W 0.00166876) 1.50687 ± 0.00841654 (W 0.00403588) 🔵 tlinalg-pool 9.14× faster than lapack-openblas
lu/n16/b1024 0.36612 ± 0.0115324 (W 0.00356228) 2.36837 ± 0.0494122 (W 0.0104245) 🔵 tlinalg-pool 6.47× faster than lapack-openblas ! between-start range over 5%
lu/n32/b1024 1.31827 ± 0.02926 (W 0.0275567) 9.64593 ± 0.034258 (W 0.0632078) 🔵 tlinalg-pool 7.32× faster than lapack-openblas ! between-start range over 5%
lu/n4/b1024 0.0716485 ± 0.00185446 (W 0.00194333) 0.329105 ± 0.00259409 (W 0.00192127) 🔵 tlinalg-pool 4.59× faster than lapack-openblas ! between-start range over 5%
lu/n8/b1024 0.143363 ± 0.00272641 (W 0.00248077) 0.802662 ± 0.00408225 (W 0.00315308) 🔵 tlinalg-pool 5.60× faster than lapack-openblas
lu_factor/n16/b1024 0.267977 ± 0.1178 (W 0.00329673) 1.8119 ± 0.0286584 (W 0.00257533) 🔵 tlinalg-pool 6.76× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
lu_factor/n32/b1024 0.756586 ± 0.00617782 (W 0.00197991) 6.7539 ± 0.0435793 (W 0.0153262) 🔵 tlinalg-pool 8.93× faster than lapack-openblas
lu_factor/n4/b1024 0.0638695 ± 0.000699583 (W 0.0019649) 0.222246 ± 0.0184214 (W 0.00228406) 🔵 tlinalg-pool 3.48× faster than lapack-openblas ! between-start range over 5%
lu_factor/n8/b1024 0.112947 ± 0.000220322 (W 0.00186791) 0.569299 ± 0.00102102 (W 0.00302417) 🔵 tlinalg-pool 5.04× faster than lapack-openblas
lu_factor_solve/n16/b1024 0.343666 ± 0.00990042 (W 0.00477453) 3.57961 ± 0.206506 (W 0.00865259) 🔵 tlinalg-pool 10.42× faster than lapack-openblas ! between-start range over 5%
lu_factor_solve/n32/b1024 0.910732 ± 0.0109809 (W 0.00559059) 9.72477 ± 0.150397 (W 0.0123577) 🔵 tlinalg-pool 10.68× faster than lapack-openblas
lu_factor_solve/n4/b1024 0.064315 ± 0.00205333 (W 0.00145915) 1.16699 ± 0.0380646 (W 0.00463665) 🔵 tlinalg-pool 18.14× faster than lapack-openblas ! between-start range over 5%
lu_factor_solve/n8/b1024 0.130799 ± 0.00470993 (W 0.00161361) 1.74831 ± 0.26849 (W 0.010855) 🔵 tlinalg-pool 13.37× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
lu_solve_prepared/n16/b1024 0.105933 ± 0.000257152 (W 0.000881017) 1.36753 ± 0.233537 (W 0.00790662) 🔵 tlinalg-pool 12.91× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
lu_solve_prepared/n32/b1024 0.311361 ± 0.00860341 (W 0.0177477) 2.21633 ± 0.0816894 (W 0.00515781) 🔵 tlinalg-pool 7.12× faster than lapack-openblas ! between-start range over 5%
lu_solve_prepared/n4/b1024 0.033337 ± 0.000773019 (W 0.000793613) 0.839396 ± 0.0297006 (W 0.00734296) 🔵 tlinalg-pool 25.18× faster than lapack-openblas ! between-start range over 5%
lu_solve_prepared/n8/b1024 0.0556135 ± 0.00155589 (W 0.000881917) 1.02744 ± 0.273124 (W 0.00925467) 🔵 tlinalg-pool 18.47× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
qr/n16/b1024 1.12821 ± 0.0125196 (W 0.0024335) 7.86965 ± 0.062679 (W 0.0167366) 🔵 tlinalg-pool 6.98× faster than lapack-openblas
qr/n32/b1024 4.5408 ± 0.00759228 (W 0.0222397) 27.3997 ± 0.173552 (W 0.0475342) 🔵 tlinalg-pool 6.03× faster than lapack-openblas
qr/n4/b1024 0.099676 ± 0.00023083 (W 0.000844724) 0.964805 ± 0.00144089 (W 0.00202545) 🔵 tlinalg-pool 9.68× faster than lapack-openblas
qr/n8/b1024 0.312143 ± 0.163618 (W 0.00270113) 2.66858 ± 0.0349439 (W 0.00870985) 🔵 tlinalg-pool 8.55× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
rank_revealing_qr/n16/b1024 1.32198 ± 0.0888906 (W 0.0028143) 9.42815 ± 0.1857 (W 0.0280232) 🔵 tlinalg-pool 7.13× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
rank_revealing_qr/n32/b1024 4.5677 ± 0.0252182 (W 0.0369771) 33.9171 ± 0.16101 (W 0.0880156) 🔵 tlinalg-pool 7.43× faster than lapack-openblas
rank_revealing_qr/n4/b1024 0.153932 ± 0.000936744 (W 0.00135299) 1.16522 ± 0.00285309 (W 0.0033963) 🔵 tlinalg-pool 7.57× faster than lapack-openblas
rank_revealing_qr/n8/b1024 0.4117 ± 0.00753713 (W 0.00207449) 3.06162 ± 0.0181944 (W 0.00529716) 🔵 tlinalg-pool 7.44× faster than lapack-openblas
solve/n16/b1024 0.352934 ± 0.00466674 (W 0.00208123) 3.71063 ± 0.245713 (W 0.0110674) 🔵 tlinalg-pool 10.51× faster than lapack-openblas ! between-start range over 5%
solve/n32/b1024 0.953971 ± 0.0197124 (W 0.0117132) 9.85251 ± 0.130554 (W 0.0151855) 🔵 tlinalg-pool 10.33× faster than lapack-openblas
solve/n4/b1024 0.0542265 ± 0.000753694 (W 0.0010221) 1.35134 ± 0.0188282 (W 0.00433961) 🔵 tlinalg-pool 24.92× faster than lapack-openblas
solve/n8/b1024 0.129142 ± 0.000823399 (W 0.0060238) 2.0483 ± 0.0430417 (W 0.00370634) 🔵 tlinalg-pool 15.86× faster than lapack-openblas ! between-start range over 5%
svd_full/n16/b1024 7.42299 ± 0.407255 (W 0.0152445) 81.0115 ± 0.189634 (W 0.252922) 🔵 tlinalg-pool 10.91× faster than lapack-openblas ! between-start range over 5%
svd_full/n32/b1024 25.9403 ± 0.407044 (W 0.0630652) 238.306 ± 1.41992 (W 0.354419) 🔵 tlinalg-pool 9.19× faster than lapack-openblas
svd_full/n4/b1024 0.544152 ± 0.0026785 (W 0.0044516) 6.32192 ± 0.0205514 (W 0.0119216) 🔵 tlinalg-pool 11.62× faster than lapack-openblas
svd_full/n8/b1024 1.93987 ± 0.0862479 (W 0.0340198) 21.2531 ± 0.0430521 (W 0.0406609) 🔵 tlinalg-pool 10.96× faster than lapack-openblas ! between-start range over 5%
svd_thin/n16/b1024 7.17199 ± 0.106423 (W 0.130466) 80.7471 ± 0.439779 (W 0.33974) 🔵 tlinalg-pool 11.26× faster than lapack-openblas
svd_thin/n32/b1024 27.9037 ± 1.20523 (W 0.146132) 238.811 ± 0.909317 (W 0.9686) 🔵 tlinalg-pool 8.56× faster than lapack-openblas ! between-start range over 5%
svd_thin/n4/b1024 0.546057 ± 0.00179162 (W 0.00701734) 6.34506 ± 0.0346932 (W 0.0157163) 🔵 tlinalg-pool 11.62× faster than lapack-openblas
svd_thin/n8/b1024 1.91697 ± 0.0508149 (W 0.0212547) 21.3123 ± 0.120726 (W 0.0307283) 🔵 tlinalg-pool 11.12× faster than lapack-openblas ! between-start range over 5%
svd_values/n16/b1024 5.13764 ± 0.0156723 (W 0.00612517) 25.1313 ± 0.168746 (W 0.0416756) 🔵 tlinalg-pool 4.89× faster than lapack-openblas
svd_values/n32/b1024 20.1428 ± 0.0992101 (W 0.0358681) 79.6873 ± 0.64862 (W 0.0775857) 🔵 tlinalg-pool 3.96× faster than lapack-openblas
svd_values/n4/b1024 0.410388 ± 0.00206496 (W 0.00200236) 3.44662 ± 0.0120687 (W 0.00414201) 🔵 tlinalg-pool 8.40× faster than lapack-openblas
svd_values/n8/b1024 1.43741 ± 0.0191923 (W 0.00389797) 8.93888 ± 0.0279815 (W 0.015362) 🔵 tlinalg-pool 6.22× faster than lapack-openblas
triangular_solve/n16/b1024 0.084317 ± 0.000841055 (W 0.00316438) 1.08434 ± 0.109825 (W 0.00582423) 🔵 tlinalg-pool 12.86× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
triangular_solve/n32/b1024 0.205809 ± 0.0169421 (W 0.00653264) 1.53271 ± 0.0483292 (W 0.00728908) 🔵 tlinalg-pool 7.45× faster than lapack-openblas ! between-start range over 5%
triangular_solve/n4/b1024 0.020784 ± 0.00275148 (W 0.000751476) 0.749399 ± 0.0626967 (W 0.00384918) 🔵 tlinalg-pool 36.06× faster than lapack-openblas ! between-start range over 5%
triangular_solve/n8/b1024 0.0349705 ± 0.000438961 (W 0.000540285) 0.882232 ± 0.0425801 (W 0.00376005) 🔵 tlinalg-pool 25.23× 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 0.172613 ± 0.00163386 (W 0.00271808) 1.61566 ± 0.0437038 (W 0.00678215) 🔵 tlinalg-pool 9.36× faster than lapack-openblas ! between-start range over 5%
cholesky/n32/b1024 0.822817 ± 0.0388665 (W 0.00801672) 6.37083 ± 0.00859705 (W 0.0071406) 🔵 tlinalg-pool 7.74× faster than lapack-openblas ! between-start range over 5%
cholesky/n4/b1024 0.032235 ± 0.000539727 (W 0.000529758) 0.232459 ± 0.00162113 (W 0.00140669) 🔵 tlinalg-pool 7.21× faster than lapack-openblas
cholesky/n8/b1024 0.061019 ± 0.00150283 (W 0.000386217) 0.498322 ± 0.01504 (W 0.00233742) 🔵 tlinalg-pool 8.17× faster than lapack-openblas ! between-start range over 5%
eig/n16/b1024 16.4234 ± 0.0427607 (W 0.0399982) 198.988 ± 0.650485 (W 0.364182) 🔵 tlinalg-pool 12.12× faster than lapack-openblas
eig/n32/b1024 67.2382 ± 0.0687037 (W 0.0303204) 886.777 ± 1.00318 (W 0.393702) 🔵 tlinalg-pool 13.19× faster than lapack-openblas
eig/n4/b1024 1.29051 ± 0.0118615 (W 0.0108906) 13.2532 ± 0.0377922 (W 0.0385004) 🔵 tlinalg-pool 10.27× faster than lapack-openblas
eig/n8/b1024 4.46854 ± 0.0183192 (W 0.0173177) 46.6002 ± 0.359382 (W 0.0771498) 🔵 tlinalg-pool 10.43× faster than lapack-openblas
eigh/n16/b1024 4.01947 ± 0.00858745 (W 0.00495748) 45.2356 ± 0.163908 (W 0.0926548) 🔵 tlinalg-pool 11.25× faster than lapack-openblas
eigh/n32/b1024 17.9039 ± 0.031762 (W 0.0274534) 187.539 ± 0.615919 (W 0.242937) 🔵 tlinalg-pool 10.47× faster than lapack-openblas
eigh/n4/b1024 0.380301 ± 0.0570137 (W 0.00157046) 3.58742 ± 0.0106668 (W 0.00434774) 🔵 tlinalg-pool 9.43× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
eigh/n8/b1024 1.15338 ± 0.00375158 (W 0.00197712) 12.3931 ± 0.105284 (W 0.0165128) 🔵 tlinalg-pool 10.75× faster than lapack-openblas
eigvals/n16/b1024 9.68117 ± 0.031304 (W 0.0105272) 149.162 ± 0.620377 (W 0.208299) 🔵 tlinalg-pool 15.41× faster than lapack-openblas
eigvals/n32/b1024 42.4545 ± 0.164689 (W 0.142495) 599.572 ± 0.960014 (W 0.682618) 🔵 tlinalg-pool 14.12× faster than lapack-openblas
eigvals/n4/b1024 0.626835 ± 0.00354899 (W 0.0042935) 10.0295 ± 0.0188524 (W 0.0208333) 🔵 tlinalg-pool 16.00× faster than lapack-openblas
eigvals/n8/b1024 2.51063 ± 0.0021953 (W 0.00455518) 36.9907 ± 0.0244126 (W 0.0558515) 🔵 tlinalg-pool 14.73× faster than lapack-openblas
eigvalsh/n16/b1024 3.00717 ± 0.00849306 (W 0.0071106) 19.3107 ± 0.0318797 (W 0.0500713) 🔵 tlinalg-pool 6.42× faster than lapack-openblas
eigvalsh/n32/b1024 11.8551 ± 0.0129255 (W 0.0459601) 73.0087 ± 0.167492 (W 0.0936851) 🔵 tlinalg-pool 6.16× faster than lapack-openblas
eigvalsh/n4/b1024 0.309313 ± 0.000507454 (W 0.00214657) 2.1057 ± 0.00581206 (W 0.00370838) 🔵 tlinalg-pool 6.81× faster than lapack-openblas
eigvalsh/n8/b1024 0.925642 ± 0.00460348 (W 0.00227389) 6.17183 ± 0.0192864 (W 0.0145016) 🔵 tlinalg-pool 6.67× faster than lapack-openblas
full_piv_lu/n16/b1024 0.628623 ± 0.00774172 (W 0.00825791) 33.3899 ± 0.0685432 (W 0.0873108) 🔵 tlinalg-pool 53.12× faster than lapack-openblas
full_piv_lu/n32/b1024 3.0551 ± 0.0312934 (W 0.0235906) 201.759 ± 1.24603 (W 1.4497) 🔵 tlinalg-pool 66.04× faster than lapack-openblas
full_piv_lu/n4/b1024 0.08825 ± 0.00264845 (W 0.00184992) 0.906719 ± 0.0286644 (W 0.00289831) 🔵 tlinalg-pool 10.27× faster than lapack-openblas ! between-start range over 5%
full_piv_lu/n8/b1024 0.200043 ± 0.00772681 (W 0.00405419) 3.92352 ± 0.0843634 (W 0.0100794) 🔵 tlinalg-pool 19.61× faster than lapack-openblas ! between-start range over 5%
full_piv_lu_solve/n16/b1024 0.706839 ± 0.00449727 (W 0.00563542) 37.8766 ± 0.167387 (W 0.0420825) 🔵 tlinalg-pool 53.59× faster than lapack-openblas
full_piv_lu_solve/n32/b1024 2.67704 ± 0.0176514 (W 0.012831) 211.557 ± 1.27952 (W 0.319527) 🔵 tlinalg-pool 79.03× faster than lapack-openblas
full_piv_lu_solve/n4/b1024 0.116398 ± 0.00528053 (W 0.0021543) 2.14216 ± 0.0115132 (W 0.00694872) 🔵 tlinalg-pool 18.40× faster than lapack-openblas ! between-start range over 5%
full_piv_lu_solve/n8/b1024 0.279006 ± 0.0569182 (W 0.00247743) 6.13752 ± 0.0486728 (W 0.0118665) 🔵 tlinalg-pool 22.00× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
householder_apply/n16/b1024 0.621229 ± 0.0142742 (W 0.00968236) 7.09994 ± 0.184628 (W 0.0269534) 🔵 tlinalg-pool 11.43× faster than lapack-openblas ! between-start range over 5%
householder_apply/n32/b1024 3.56848 ± 0.0779788 (W 0.0178369) 36.0122 ± 0.381408 (W 0.0721041) 🔵 tlinalg-pool 10.09× faster than lapack-openblas
householder_apply/n4/b1024 0.049653 ± 0.0031643 (W 0.00098818) 0.441435 ± 0.00116111 (W 0.00280425) 🔵 tlinalg-pool 8.89× faster than lapack-openblas ! between-start range over 5%
householder_apply/n8/b1024 0.161061 ± 0.0117862 (W 0.00532574) 1.49047 ± 0.00730228 (W 0.00336697) 🔵 tlinalg-pool 9.25× faster than lapack-openblas ! between-start range over 5%
householder_factor/n16/b1024 0.575476 ± 0.00539725 (W 0.00385144) 7.38155 ± 0.0282787 (W 0.0176688) 🔵 tlinalg-pool 12.83× faster than lapack-openblas
householder_factor/n32/b1024 2.63168 ± 0.0388916 (W 0.0416234) 27.1298 ± 0.1888 (W 0.0318975) 🔵 tlinalg-pool 10.31× faster than lapack-openblas
householder_factor/n4/b1024 0.075456 ± 0.000501121 (W 0.000577728) 0.924272 ± 0.00593815 (W 0.0016966) 🔵 tlinalg-pool 12.25× faster than lapack-openblas
householder_factor/n8/b1024 0.192549 ± 0.00143758 (W 0.00181665) 2.3766 ± 0.0110249 (W 0.00670962) 🔵 tlinalg-pool 12.34× faster than lapack-openblas
lu/n16/b1024 0.432979 ± 0.00562674 (W 0.00685623) 3.98662 ± 0.0567853 (W 0.00983231) 🔵 tlinalg-pool 9.21× faster than lapack-openblas
lu/n32/b1024 2.12432 ± 0.0246833 (W 0.0177218) 15.4428 ± 0.145575 (W 0.0262942) 🔵 tlinalg-pool 7.27× faster than lapack-openblas
lu/n4/b1024 0.070557 ± 0.00121861 (W 0.00146337) 0.382379 ± 0.00260219 (W 0.00258645) 🔵 tlinalg-pool 5.42× faster than lapack-openblas
lu/n8/b1024 0.149985 ± 0.00154069 (W 0.00342833) 1.05255 ± 0.0134039 (W 0.00266987) 🔵 tlinalg-pool 7.02× faster than lapack-openblas
lu_factor/n16/b1024 0.295186 ± 0.0182973 (W 0.00759162) 2.71974 ± 0.0159372 (W 0.00581023) 🔵 tlinalg-pool 9.21× faster than lapack-openblas ! between-start range over 5%
lu_factor/n32/b1024 1.32229 ± 0.0258548 (W 0.0174975) 11.3386 ± 0.169502 (W 0.0223366) 🔵 tlinalg-pool 8.57× faster than lapack-openblas
lu_factor/n4/b1024 0.067411 ± 0.000383415 (W 0.00138786) 0.271418 ± 0.00181876 (W 0.00185564) 🔵 tlinalg-pool 4.03× faster than lapack-openblas
lu_factor/n8/b1024 0.117985 ± 0.0138991 (W 0.00161758) 0.759049 ± 0.00183941 (W 0.00303478) 🔵 tlinalg-pool 6.43× faster than lapack-openblas ! between-start range over 5%
lu_factor_solve/n16/b1024 0.456869 ± 0.00506595 (W 0.00378305) 5.11843 ± 0.0586761 (W 0.019864) 🔵 tlinalg-pool 11.20× faster than lapack-openblas
lu_factor_solve/n32/b1024 1.75546 ± 0.0200154 (W 0.0333742) 15.5752 ± 0.27609 (W 0.0158843) 🔵 tlinalg-pool 8.87× faster than lapack-openblas
lu_factor_solve/n4/b1024 0.080911 ± 0.000810398 (W 0.00110093) 1.40062 ± 0.0395304 (W 0.0049225) 🔵 tlinalg-pool 17.31× faster than lapack-openblas ! between-start range over 5%
lu_factor_solve/n8/b1024 0.16555 ± 0.0683428 (W 0.00227442) 2.20147 ± 0.164157 (W 0.00511664) 🔵 tlinalg-pool 13.30× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
lu_solve_prepared/n16/b1024 0.19572 ± 0.0189682 (W 0.00352107) 1.97275 ± 0.0165671 (W 0.00849174) 🔵 tlinalg-pool 10.08× faster than lapack-openblas ! between-start range over 5%
lu_solve_prepared/n32/b1024 0.884143 ± 0.0216936 (W 0.00869384) 3.90134 ± 0.0412984 (W 0.0111773) 🔵 tlinalg-pool 4.41× faster than lapack-openblas ! between-start range over 5%
lu_solve_prepared/n4/b1024 0.0447585 ± 0.000346288 (W 0.00121656) 1.03755 ± 0.0374376 (W 0.00516968) 🔵 tlinalg-pool 23.18× faster than lapack-openblas ! between-start range over 5%
lu_solve_prepared/n8/b1024 0.087363 ± 0.00106372 (W 0.000902096) 1.2164 ± 0.152874 (W 0.00248359) 🔵 tlinalg-pool 13.92× faster than lapack-openblas ! between-start range over 5%
qr/n16/b1024 1.39947 ± 0.0033504 (W 0.00511432) 12.5603 ± 0.0431006 (W 0.0297604) 🔵 tlinalg-pool 8.98× faster than lapack-openblas
qr/n32/b1024 9.41301 ± 0.0234656 (W 0.0347283) 49.5516 ± 0.292016 (W 0.064714) 🔵 tlinalg-pool 5.26× faster than lapack-openblas
qr/n4/b1024 0.114249 ± 0.00126011 (W 0.00141198) 1.37963 ± 0.00344576 (W 0.00300589) 🔵 tlinalg-pool 12.08× faster than lapack-openblas
qr/n8/b1024 0.335957 ± 0.000620989 (W 0.00196992) 3.74028 ± 0.0112216 (W 0.006076) 🔵 tlinalg-pool 11.13× faster than lapack-openblas
rank_revealing_qr/n16/b1024 1.56728 ± 0.013733 (W 0.00587322) 16.6501 ± 0.044163 (W 0.0259538) 🔵 tlinalg-pool 10.62× faster than lapack-openblas
rank_revealing_qr/n32/b1024 8.72147 ± 1.31232 (W 0.0826752) 65.6304 ± 0.230549 (W 0.0871364) 🔵 tlinalg-pool 7.53× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
rank_revealing_qr/n4/b1024 0.183537 ± 0.000960837 (W 0.00172398) 1.77411 ± 0.0130734 (W 0.00438807) 🔵 tlinalg-pool 9.67× faster than lapack-openblas
rank_revealing_qr/n8/b1024 0.472859 ± 0.016882 (W 0.00208419) 4.91663 ± 0.0182032 (W 0.0100596) 🔵 tlinalg-pool 10.40× faster than lapack-openblas ! between-start range over 5%
solve/n16/b1024 0.508539 ± 0.00741731 (W 0.0123776) 5.15597 ± 0.197197 (W 0.0190523) 🔵 tlinalg-pool 10.14× faster than lapack-openblas ! between-start range over 5%
solve/n32/b1024 1.84572 ± 0.0203152 (W 0.00847055) 16.0648 ± 0.210122 (W 0.0144144) 🔵 tlinalg-pool 8.70× faster than lapack-openblas
solve/n4/b1024 0.0765125 ± 0.00110691 (W 0.00263016) 1.47565 ± 0.171204 (W 0.00588362) 🔵 tlinalg-pool 19.29× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
solve/n8/b1024 0.186287 ± 0.00370951 (W 0.00165544) 2.36366 ± 0.157112 (W 0.0032374) 🔵 tlinalg-pool 12.69× faster than lapack-openblas ! between-start range over 5%
svd_full/n16/b1024 7.75212 ± 0.275074 (W 0.0844928) 93.7184 ± 0.139146 (W 0.0647117) 🔵 tlinalg-pool 12.09× faster than lapack-openblas ! between-start range over 5%
svd_full/n32/b1024 34.5236 ± 1.5545 (W 0.788951) 296.788 ± 1.00723 (W 0.507046) 🔵 tlinalg-pool 8.60× faster than lapack-openblas ! between-start range over 5%
svd_full/n4/b1024 0.637073 ± 0.00127364 (W 0.00267627) 7.9199 ± 0.0500496 (W 0.0128364) 🔵 tlinalg-pool 12.43× faster than lapack-openblas
svd_full/n8/b1024 2.16674 ± 0.0757461 (W 0.0537403) 25.1488 ± 0.141991 (W 0.0231398) 🔵 tlinalg-pool 11.61× faster than lapack-openblas ! between-start range over 5%
svd_thin/n16/b1024 7.76759 ± 0.07083 (W 0.0488894) 93.9546 ± 0.354904 (W 0.296305) 🔵 tlinalg-pool 12.10× faster than lapack-openblas
svd_thin/n32/b1024 34.6575 ± 1.60891 (W 0.32377) 296.758 ± 1.55635 (W 0.422526) 🔵 tlinalg-pool 8.56× faster than lapack-openblas ! between-start range over 5%
svd_thin/n4/b1024 0.645722 ± 0.0104946 (W 0.00255229) 7.92581 ± 0.0137568 (W 0.0147519) 🔵 tlinalg-pool 12.27× faster than lapack-openblas
svd_thin/n8/b1024 2.16798 ± 0.0391072 (W 0.0374402) 24.96 ± 0.178769 (W 0.0294621) 🔵 tlinalg-pool 11.51× faster than lapack-openblas
svd_values/n16/b1024 5.86249 ± 0.0470878 (W 0.0267828) 33.4194 ± 0.0681238 (W 0.0414114) 🔵 tlinalg-pool 5.70× faster than lapack-openblas
svd_values/n32/b1024 22.3545 ± 0.0631473 (W 0.131033) 110.463 ± 0.195405 (W 0.166232) 🔵 tlinalg-pool 4.94× faster than lapack-openblas
svd_values/n4/b1024 0.496807 ± 0.00212269 (W 0.00167365) 4.54826 ± 0.00964077 (W 0.0056216) 🔵 tlinalg-pool 9.15× faster than lapack-openblas
svd_values/n8/b1024 1.64685 ± 0.0135313 (W 0.00483269) 11.2404 ± 0.0389778 (W 0.0175071) 🔵 tlinalg-pool 6.83× faster than lapack-openblas
triangular_solve/n16/b1024 0.144795 ± 0.0614267 (W 0.00410531) 1.36192 ± 0.0753983 (W 0.00508851) 🔵 tlinalg-pool 9.41× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
triangular_solve/n32/b1024 0.534079 ± 0.0526721 (W 0.0112238) 3.09951 ± 0.122402 (W 0.0125576) 🔵 tlinalg-pool 5.80× faster than lapack-openblas ! between-start range over 5%; p position effect over 5%
triangular_solve/n4/b1024 0.028343 ± 0.000481197 (W 0.000853476) 0.754563 ± 0.0727299 (W 0.00358742) 🔵 tlinalg-pool 26.62× faster than lapack-openblas ! between-start range over 5%
triangular_solve/n8/b1024 0.065547 ± 0.00041511 (W 0.00109842) 1.01691 ± 0.0554418 (W 0.00464818) 🔵 tlinalg-pool 15.51× faster than lapack-openblas ! between-start range over 5%

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

case tlinalg-pool (ms) lapack-openblas (ms) vs lapack-openblas notes
cholesky/n128/b1 0.0655025 ± 0.00214799 (W 0.000704945) 0.0777355 ± 0.00060756 (W 0.00102094) 🔵 tlinalg-pool 1.19× faster than lapack-openblas ! between-start range over 5%
lu/n128/b1 0.184764 ± 0.0018791 (W 0.00405565) 0.146139 ± 0.000275259 (W 0.00210002) 🔴 tlinalg-pool 1.26× slower than lapack-openblas
lu_factor/n128/b1 0.129663 ± 0.00382243 (W 0.00498146) 0.130499 ± 0.000365501 (W 0.00145194) · ! between-start range over 5%
lu_factor_solve/n128/b1 0.134586 ± 0.0021705 (W 0.00485527) 0.146118 ± 0.000597041 (W 0.00125043) 🔵 tlinalg-pool 1.09× faster than lapack-openblas
lu_solve_prepared/n128/b1 0.009903 ± 4.13148e-05 (W 0.000127985) 0.0121225 ± 0.000116076 (W 0.000125554) 🔵 tlinalg-pool 1.22× faster than lapack-openblas dispatch-dominated (row under 20 us)
solve/n128/b1 0.148147 ± 0.00320999 (W 0.00293233) 0.154624 ± 0.000575504 (W 0.00109961) ·
svd_values/n128/b1 2.37897 ± 0.00467959 (W 0.00335758) 1.37847 ± 0.00437886 (W 0.00275707) 🔴 tlinalg-pool 1.73× slower than lapack-openblas
triangular_solve/n128/b1 0.0067625 ± 6.69347e-05 (W 5.53367e-05) 0.0067775 ± 8.67962e-05 (W 0.000186949) · 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 0.187345 ± 0.000973791 (W 0.00208158) 0.164152 ± 0.00130766 (W 0.00155686) 🔴 tlinalg-pool 1.14× slower than lapack-openblas
lu/n128/b1 0.356035 ± 0.00337431 (W 0.00384996) 0.237799 ± 0.00181954 (W 0.0027091) 🔴 tlinalg-pool 1.50× slower than lapack-openblas
lu_factor/n128/b1 0.271502 ± 0.00181032 (W 0.00310908) 0.190099 ± 0.00234732 (W 0.00203639) 🔴 tlinalg-pool 1.43× slower than lapack-openblas
lu_factor_solve/n128/b1 0.299439 ± 0.00176552 (W 0.00260969) 0.221934 ± 0.000770396 (W 0.00173213) 🔴 tlinalg-pool 1.35× slower than lapack-openblas
lu_solve_prepared/n128/b1 0.0312935 ± 0.000121848 (W 0.000127835) 0.0291795 ± 0.000136359 (W 0.00018085) 🔴 tlinalg-pool 1.07× slower than lapack-openblas
solve/n128/b1 0.328063 ± 0.00426999 (W 0.0047647) 0.242903 ± 0.000488008 (W 0.00255534) 🔴 tlinalg-pool 1.35× slower than lapack-openblas
svd_values/n128/b1 3.14675 ± 0.00859905 (W 0.00875446) 2.60483 ± 0.00324328 (W 0.00439927) 🔴 tlinalg-pool 1.21× slower than lapack-openblas
triangular_solve/n128/b1 0.018725 ± 6.45497e-05 (W 0.000122175) 0.0153135 ± 0.000178364 (W 0.00014582) 🔴 tlinalg-pool 1.22× slower than lapack-openblas 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 23.5542 ± 0.105105 (W 0.105037) 23.4583 ± 0.0778266 (W 0.0600774) ·
eigvalsh/n512/b1 18.7042 ± 0.0575184 (W 0.0211457) 11.2348 ± 0.00706165 (W 0.00768317) 🔴 tlinalg-pool 1.66× slower than lapack-openblas
householder_apply/n512/b1 2.86728 ± 0.0458374 (W 0.0577571) 9.89995 ± 0.104697 (W 0.0236682) 🔵 tlinalg-pool 3.45× faster than lapack-openblas
householder_factor/n512/b1 2.73316 ± 0.00813568 (W 0.0391458) 6.78853 ± 0.0215687 (W 0.00965911) 🔵 tlinalg-pool 2.48× faster than lapack-openblas
qr/n512/b1 7.68994 ± 0.055808 (W 0.108951) 13.4764 ± 0.127907 (W 0.0139148) 🔵 tlinalg-pool 1.75× faster than lapack-openblas
rank_revealing_qr/n512/b1 9.55947 ± 0.0761818 (W 0.104882) 23.1913 ± 0.171338 (W 0.0419118) 🔵 tlinalg-pool 2.43× faster than lapack-openblas
svd_thin/n512/b1 60.8953 ± 0.0528764 (W 0.107577) 85.1764 ± 0.818956 (W 0.0819677) 🔵 tlinalg-pool 1.40× 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 38.8685 ± 0.171927 (W 0.423891) 204.264 ± 0.44243 (W 0.346546) 🔵 tlinalg-pool 5.26× faster than lapack-openblas
eigvalsh/n512/b1 28.9399 ± 0.0372377 (W 0.0558823) 29.8373 ± 0.0779838 (W 0.0202827) ·
householder_apply/n512/b1 7.08357 ± 0.0689304 (W 0.0602982) 14.8265 ± 0.0566596 (W 0.0368967) 🔵 tlinalg-pool 2.09× faster than lapack-openblas
householder_factor/n512/b1 6.9061 ± 0.0583879 (W 0.126565) 10.8785 ± 0.651895 (W 0.0142747) 🔵 tlinalg-pool 1.58× faster than lapack-openblas ! between-start range over 5%
qr/n512/b1 16.9656 ± 0.170664 (W 0.238427) 21.8073 ± 0.0423767 (W 0.0220397) 🔵 tlinalg-pool 1.29× faster than lapack-openblas
rank_revealing_qr/n512/b1 19.5559 ± 0.0783467 (W 0.147364) 24.6119 ± 0.230759 (W 0.0236384) 🔵 tlinalg-pool 1.26× faster than lapack-openblas
svd_thin/n512/b1 99.8139 ± 0.307806 (W 0.786339) 98.7976 ± 0.69727 (W 0.0543399) ·

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

case tlinalg-pool (ms) lapack-openblas (ms) vs lapack-openblas notes
householder_factor/4096x256/b1 7.89053 ± 0.0161559 (W 0.0365686) 17.922 ± 0.0872771 (W 0.0212743) 🔵 tlinalg-pool 2.27× faster than lapack-openblas
qr/4096x256/b1 22.8411 ± 0.0325104 (W 0.579827) 35.7013 ± 0.339325 (W 0.0255107) 🔵 tlinalg-pool 1.56× faster than lapack-openblas
rank_revealing_qr/4096x256/b1 24.5597 ± 0.454449 (W 0.305451) 38.973 ± 0.410198 (W 0.0313668) 🔵 tlinalg-pool 1.59× faster than lapack-openblas
svd_thin/4096x256/b1 39.4189 ± 0.279698 (W 0.259719) 60.7983 ± 0.018257 (W 0.0707943) 🔵 tlinalg-pool 1.54× 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 19.9411 ± 0.0476979 (W 0.0553556) 20.1088 ± 0.0483605 (W 0.0243042) ·
qr/4096x256/b1 53.456 ± 0.169055 (W 0.12754) 40.365 ± 0.13962 (W 0.033922) 🔴 tlinalg-pool 1.32× slower than lapack-openblas
rank_revealing_qr/4096x256/b1 60.3822 ± 0.177435 (W 0.265862) 51.4232 ± 0.313834 (W 1.102) 🔴 tlinalg-pool 1.17× slower than lapack-openblas
svd_thin/4096x256/b1 111.219 ± 0.321955 (W 1.02001) 86.0045 ± 0.162196 (W 0.35479) 🔴 tlinalg-pool 1.29× 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.