Launch geometry, block widths, iteration counts and tuning thresholds.
Where the call site's default is a function of n, the scalar type, an argument or a device property, the field is a sentinel (0, std::nullopt or an unset EnvValue) and the default stays at the call site. Fields with a constant default carry it.
Definition at line 248 of file settings.hh.
#include <batchlas/settings.hh>
| int batchlas::GeometrySettings::latrd_grid_groups = 0 |
BATCHLAS_LATRD_GRID_GROUPS (env_positive_int_or).
0 = auto, min(residency cap, ceil((n-1)/32)). A forced value is still clamped by the residency cap unless UnsafeSettings::latrd_grid_force_unsafe lifts it.
Definition at line 252 of file settings.hh.
| int batchlas::GeometrySettings::latrd_grid_min_n = 768 |
BATCHLAS_LATRD_GRID_MIN_N (env_positive_int_or): the n at which the grid latrd path becomes eligible.
Definition at line 257 of file settings.hh.
| int batchlas::GeometrySettings::latrd_grid_wg = 0 |
BATCHLAS_LATRD_GRID_WG (env_positive_int_or).
0 = computed. Only 32, 64, 128 and 256 are honoured; any other value is silently ignored.
Definition at line 261 of file settings.hh.
| int batchlas::GeometrySettings::latrd_lower_panel_wg_hint = 0 |
BATCHLAS_LATRD_LOWER_PANEL_WG_HINT.
0 = unset; only 64, 128 and 256 take effect, and only on the device path. Wins over BATCHLAS_TUNE_LATRD_WG_HINT.
Definition at line 266 of file settings.hh.
| int batchlas::GeometrySettings::sb2st_back_subs = 0 |
BATCHLAS_SB2ST_BACK_SUBS (env_positive_int_or).
0 = the call site's default chain.
Definition at line 271 of file settings.hh.
| int batchlas::GeometrySettings::sb2st_back_tile_w = 0 |
BATCHLAS_SB2ST_BACK_TILE_W (env_positive_int_or): column tile of the wave kernel.
0 = the tuned heuristic.
Definition at line 275 of file settings.hh.
| EnvValue batchlas::GeometrySettings::sb2st_back_tile {} |
BATCHLAS_SB2ST_BACK_TILE: tile of the tiled back-transform kernel (a different kernel from sb2st_back_tile_w).
0 selects the streaming kernel.
Definition at line 280 of file settings.hh.
| int batchlas::GeometrySettings::potrf_nb = 0 |
BATCHLAS_POTRF_NB.
0 = unset; the default is type-dependent (128/96/96/64 for float/double/cfloat/cdouble).
Definition at line 284 of file settings.hh.
| int batchlas::GeometrySettings::potrf_w = 0 |
BATCHLAS_POTRF_W. 0 = unset; the default is type-dependent (128/32/32/16).
Definition at line 286 of file settings.hh.
| int batchlas::GeometrySettings::syev_two_stage_kd = 32 |
BATCHLAS_SYEV_TWO_STAGE_KD (env_positive_int_or): the band width, clamped to [1, n-1] at the call site.
Definition at line 290 of file settings.hh.
| int batchlas::GeometrySettings::syev_two_stage_sb2st_block = 32 |
BATCHLAS_SYEV_TWO_STAGE_SB2ST_BLOCK (env_positive_int_or).
Read by four solve/sizing pairs.
Definition at line 294 of file settings.hh.
| EnvValue batchlas::GeometrySettings::sy2sb_ormqr_nb {} |
BATCHLAS_SY2SB_ORMQR_NB: unset, off or 0 (never hint), or a positive forced value (clamped to kd at the call site).
A value that does not parse, is negative or is above 1024 reads as unset. Wins over BATCHLAS_TUNE_SY2SB_ORMQR_NB.
Definition at line 299 of file settings.hh.
| int batchlas::GeometrySettings::sytrd_block_size = 0 |
BATCHLAS_SYTRD_BLOCK_SIZE.
0 = unset; the default is n-bucketed and type-dependent. Wins over BATCHLAS_TUNE_SYTRD_BLOCK_SIZE.
Definition at line 303 of file settings.hh.
| int batchlas::GeometrySettings::trmm_tile_m = 0 |
BATCHLAS_TRMM_TILE_M. 0 = unset (a function of m); bucketed to 16, 32, 64 or 128.
Definition at line 306 of file settings.hh.
| int batchlas::GeometrySettings::trsm_outer_nb = 0 |
BATCHLAS_TRSM_OUTER_NB. 0 = unset: 128 for Side::Left, the CTA nb for Side::Right.
Definition at line 309 of file settings.hh.
| EnvValue batchlas::GeometrySettings::expand_max_bytes {} |
BATCHLAS_EXPAND_MAX_BYTES (strtoull): only ever lowers the expansion ceiling, whose default is GLOBAL_MEM_SIZE / 4.
Definition at line 313 of file settings.hh.
| EnvValue batchlas::GeometrySettings::gesvd_blocked_gebrd_min {} |
BATCHLAS_GESVD_BLOCKED_GEBRD_MIN (bare atoi, default 1).
Definition at line 318 of file settings.hh.
| int batchlas::GeometrySettings::syevx_check_every = 4 |
BATCHLAS_SYEVX_CHECK_EVERY: iterations between convergence checks; each check is a full pipeline drain.
Definition at line 322 of file settings.hh.
| std::optional<int> batchlas::GeometrySettings::syevx_extra_directions {} |
BATCHLAS_SYEVX_EXTRA_DIRECTIONS: 0 means "no guard block"; nullopt means the default max(2, k/4).
A negative value reads as unset.
Definition at line 326 of file settings.hh.
| int batchlas::GeometrySettings::syevx_filter_degree = 0 |
BATCHLAS_SYEVX_FILTER_DEGREE.
0 = unset (only > 0 is accepted). Top of a four-level precedence chain; setting it also disables the automatic degree.
Definition at line 330 of file settings.hh.
| std::optional<int> batchlas::GeometrySettings::syevx_init_power {} |
BATCHLAS_SYEVX_INIT_POWER: 0 means "no power iterations"; nullopt means 4 unless SyevxParams supplied a value.
Definition at line 334 of file settings.hh.
| double batchlas::GeometrySettings::syevx_lock_factor = 0.1 |
BATCHLAS_SYEVX_LOCK_FACTOR (atof, > 0 only): a column is masked once its residual is this multiple of the tolerance.
An unparseable value yields the default.
Definition at line 339 of file settings.hh.
| struct batchlas::GeometrySettings::TuneSettings batchlas::GeometrySettings::tune |
The eleven BATCHLAS_TUNE_* overrides.
Classes | |
| struct | TuneSettings |
| The runtime overrides of include/batchlas/tuning_params.hh (BATCHLAS_TUNE_*). More... | |
1.18.0
