Which kernel or algorithm runs, for the knobs outside the route pins.
Every field changes which code path executes. syevx_algorithm and syevx_preconditioner override an explicit SyevxParams field, and gesvd_bidiag changes numerics.
Definition at line 119 of file settings.hh.
#include <batchlas/settings.hh>
| EnvValue batchlas::SelectionSettings::expand_route {} |
BATCHLAS_EXPAND_ROUTE = expand | loop: pins the scratch-expansion route of hemm, herk and her2k, so a test can reach the arm the shape would not have picked (symm and trmm pin expand through their BATCHLAS_<OP>_ROUTE).
Not op-keyed, so RoutingSettings does not hold it.
Definition at line 126 of file settings.hh.
| EnvValue batchlas::SelectionSettings::gemv_segt {} |
BATCHLAS_GEMV_SEGT = off | auto | 2 | 4 | 8: segmented-tail width of the native gemv.
Definition at line 132 of file settings.hh.
| EnvValue batchlas::SelectionSettings::gesvd_bidiag {} |
BATCHLAS_GESVD_BIDIAG = bdsdc | normal | bdsqr.
Definition at line 138 of file settings.hh.
| EnvValue batchlas::SelectionSettings::getrf_leaf {} |
BATCHLAS_GETRF_LEAF = slm | reg (default: the register-resident panel leaf wherever the panel fits).
Re-read per call.
Definition at line 142 of file settings.hh.
| EnvValue batchlas::SelectionSettings::geqrf_leaf {} |
BATCHLAS_GEQRF_LEAF = auto (default) | reg (the register panel leaf wherever it fits).
Re-read per call.
Definition at line 147 of file settings.hh.
| EnvValue batchlas::SelectionSettings::getrf_laswp {} |
BATCHLAS_GETRF_LASWP = inloop | defer_walk | defer_gather (default).
Presence is latched at the call site, the value re-read per call.
Definition at line 151 of file settings.hh.
| EnvValue batchlas::SelectionSettings::getrf_right_laswp {} |
BATCHLAS_GETRF_RIGHT_LASWP = walk | gather; unset takes the measured crossover.
Definition at line 154 of file settings.hh.
| EnvValue batchlas::SelectionSettings::getrs_laswp {} |
BATCHLAS_GETRS_LASWP = walk | gather. Deliberately not latched (as gemv_segt).
Definition at line 157 of file settings.hh.
| EnvValue batchlas::SelectionSettings::iluk_device {} |
BATCHLAS_ILUK_DEVICE: only a first character of 0 or 1 is inspected; true, on and yes fall through to the shape default (batch_size >= 32).
Definition at line 161 of file settings.hh.
| EnvValue batchlas::SelectionSettings::latrd_impl {} |
BATCHLAS_LATRD_IMPL = legacy | device | grid. Re-read per call.
Definition at line 164 of file settings.hh.
| EnvValue batchlas::SelectionSettings::ormqr_impl {} |
BATCHLAS_ORMQR_IMPL: only the exact value device has an effect.
Definition at line 167 of file settings.hh.
| EnvValue batchlas::SelectionSettings::ormqr_wy {} |
BATCHLAS_ORMQR_WY = gemm | trmm | measured: overrides the measured gate.
Definition at line 170 of file settings.hh.
| EnvValue batchlas::SelectionSettings::ortho_gram {} |
BATCHLAS_ORTHO_GRAM: only the exact value gemm has an effect (forces the GEMM gram route in place of the syrk one).
Definition at line 174 of file settings.hh.
| EnvValue batchlas::SelectionSettings::sb2st_back_wave {} |
BATCHLAS_SB2ST_BACK_WAVE: the wave-parallel back-transform is on unless disabled with one of 0, false, off, no, n, disable, disabled (case-insensitive).
Definition at line 180 of file settings.hh.
| EnvValue batchlas::SelectionSettings::sb2st_subgroup {} |
BATCHLAS_SB2ST_SUBGROUP = auto | on | off (case-folded).
on throws when kd > 32 or the device lacks sub-group size 32.
Definition at line 184 of file settings.hh.
| EnvValue batchlas::SelectionSettings::syev_two_stage_chase {} |
BATCHLAS_SYEV_TWO_STAGE_CHASE: only the exact value givens has an effect.
Definition at line 189 of file settings.hh.
| EnvValue batchlas::SelectionSettings::syevx_algorithm {} |
BATCHLAS_SYEVX_ALGORITHM: overrides an explicit SyevxParams::method.
An env-supplied algorithm degrades where an explicit request throws, and an unrecognised value still wins (it parses to Auto).
Definition at line 194 of file settings.hh.
| EnvValue batchlas::SelectionSettings::syevx_preconditioner {} |
BATCHLAS_SYEVX_PRECONDITIONER: overrides the SyevxParams default; loses to an explicit request and to a configured ILU(k) factor.
Definition at line 198 of file settings.hh.
| EnvValue batchlas::SelectionSettings::syevx_filter_degree_auto {} |
BATCHLAS_SYEVX_FILTER_DEGREE_AUTO: derive the Chebyshev filter degree automatically.
Off by default because it wins only at batch 1. Parsed with atoi != 0.
Definition at line 203 of file settings.hh.
| EnvValue batchlas::SelectionSettings::syevx_instr_host {} |
BATCHLAS_SYEVX_INSTR_HOST: forces the host instrumentation path.
First-character parser {1,t,T,y,Y}: true works, on does not.
Definition at line 207 of file settings.hh.
| EnvValue batchlas::SelectionSettings::syevx_projected_vendor {} |
BATCHLAS_SYEVX_PROJECTED_VENDOR: same first-character dialect.
Definition at line 212 of file settings.hh.
| EnvValue batchlas::SelectionSettings::syevx_soft_lock {} |
BATCHLAS_SYEVX_SOFT_LOCK.
Definition at line 218 of file settings.hh.
| bool batchlas::SelectionSettings::sytrd_force_local_small = false |
BATCHLAS_SYTRD_FORCE_LOCAL_SMALL (env_truthy): forces the local-memory small-n path past the tuning heuristic.
It cannot force an unschedulable launch; the device gate still applies.
Definition at line 222 of file settings.hh.
| std::optional<bool> batchlas::SelectionSettings::sytrd_fuse_panel_update {} |
BATCHLAS_SYTRD_FUSE_PANEL_UPDATE: forced on (env_truthy), forced off (env_falsy), or std::nullopt to let the tuned default decide.
A value that is neither reads as nullopt.
Definition at line 226 of file settings.hh.
| EnvValue batchlas::SelectionSettings::sytrd_impl {} |
BATCHLAS_SYTRD_IMPL: only the exact value device has an effect.
Definition at line 229 of file settings.hh.
| EnvValue batchlas::SelectionSettings::sytrd_trailing_update {} |
| EnvValue batchlas::SelectionSettings::tuned_dir {} |
BATCHLAS_TUNED_DIR: a directory of select tables (src/select/select.hh); a file there replaces the built-in table of the same name.
Definition at line 237 of file settings.hh.
1.18.0
