Overview

Per-size tuning lookups, environment parsers and internal hooks behind settings() and configure().

Functions

◆ note_queue_constructed()

void batchlas::detail::note_queue_constructed ( )

noexcept

Latch recording that a Queue exists, which closes configure().

Called only by Queue's constructors.

◆ queue_constructed()

bool batchlas::detail::queue_constructed ( )

noexcept

True once note_queue_constructed() has been called.

◆ tuning_env_override()

int32_t batchlas::tuning::detail::tuning_env_override	(	const EnvValue &	captured,
		int32_t	fallback )

inline

The captured override if it is a clean positive 32-bit integer, else fallback.

Unset, empty, unparseable, trailing garbage ("16x"), non-positive or out-of-range values all yield fallback, the compiled constant.

Parameters

captured	the captured BATCHLAS_TUNE_* value from settings()
fallback	the compiled constant for this n

Definition at line 37 of file tuning_params.hh.

◆ ormqr_block_size_default_for_n()

int32_t batchlas::tuning::ormqr_block_size_default_for_n ( int32_t n )

inlineconstexpr

Compiled ORMQR_BLOCK_SIZE_* bucket for n (no override).

Definition at line 139 of file tuning_params.hh.

◆ ormqr_block_size_for_n()

int32_t batchlas::tuning::ormqr_block_size_for_n ( int32_t n )

inline

ormqr WY block width for order n; override BATCHLAS_TUNE_ORMQR_BLOCK_SIZE.

Definition at line 232 of file tuning_params.hh.

◆ gebrd_block_size_for_n()

int32_t batchlas::tuning::gebrd_block_size_for_n ( int32_t n )

inline

gebrd panel width for order n; override BATCHLAS_TUNE_GEBRD_BLOCK_SIZE.

Definition at line 239 of file tuning_params.hh.

◆ sb2st_back_tile_for_n()

int32_t batchlas::tuning::sb2st_back_tile_for_n ( int32_t n )

inline

sb2st wave back-transform tile for order n, or 0 for "keep the call site's heuristic"; override BATCHLAS_TUNE_SB2ST_BACK_TILE.

Note: 0 is reachable only from the compiled constant: the override forces a geometry, never auto. The same holds for sb2st_back_subs_for_n() and sy2sb_ormqr_nb_for_n().

Definition at line 249 of file tuning_params.hh.

◆ sb2st_back_subs_for_n()

int32_t batchlas::tuning::sb2st_back_subs_for_n ( int32_t n )

inline

sb2st wave back-transform sub-group count for order n, or 0 for the heuristic; override BATCHLAS_TUNE_SB2ST_BACK_SUBS.

Definition at line 257 of file tuning_params.hh.

◆ sy2sb_ormqr_nb_for_n()

int32_t batchlas::tuning::sy2sb_ormqr_nb_for_n ( int32_t n )

inline

sy2sb back-transform WY width for order n, or 0 for the shape gate; override BATCHLAS_TUNE_SY2SB_ORMQR_NB.

Definition at line 265 of file tuning_params.hh.

◆ sytrd_block_size_for_n()

int32_t batchlas::tuning::sytrd_block_size_for_n ( int32_t n )

inline

sytrd panel width for order n; override BATCHLAS_TUNE_SYTRD_BLOCK_SIZE.

Definition at line 272 of file tuning_params.hh.

◆ latrd_lower_panel_wg_hint_for_n()

int32_t batchlas::tuning::latrd_lower_panel_wg_hint_for_n ( int32_t n )

inline

latrd lower-panel work-group size hint for order n, 0 for none; override BATCHLAS_TUNE_LATRD_WG_HINT.

Definition at line 280 of file tuning_params.hh.

◆ latrd_lower_panel_wg_hint()

int32_t batchlas::tuning::latrd_lower_panel_wg_hint ( )

inline

latrd_lower_panel_wg_hint_for_n() at n = 256, for the legacy path.

Definition at line 287 of file tuning_params.hh.

◆ sytrd_fuse_panel_update_for_n()

bool batchlas::tuning::sytrd_fuse_panel_update_for_n ( int32_t n )

inlineconstexpr

Whether sytrd fuses the panel update at order n.

No BATCHLAS_TUNE_* override; see SelectionSettings::sytrd_fuse_panel_update.

Definition at line 292 of file tuning_params.hh.

◆ stedc_recursion_threshold_for_n()

int32_t batchlas::tuning::stedc_recursion_threshold_for_n ( int32_t n )

inline

stedc leaf size for order n; override BATCHLAS_TUNE_STEDC_RECURSION_THRESHOLD.

Note: 32 is the CTA invariant (sub-group width), not a tuning result.

Definition at line 303 of file tuning_params.hh.

◆ stedc_merge_variant_for_n()

int32_t batchlas::tuning::stedc_merge_variant_for_n ( int32_t n )

inline

stedc merge variant for order n, as a StedcMergeVariant value (1 Fused, 2 FusedCta); override BATCHLAS_TUNE_STEDC_MERGE_VARIANT.

Definition at line 313 of file tuning_params.hh.

◆ stedc_threads_per_root_for_n()

int32_t batchlas::tuning::stedc_threads_per_root_for_n ( int32_t n )

inline

stedc secular-solve threads per root for order n; override BATCHLAS_TUNE_STEDC_THREADS_PER_ROOT.

Definition at line 321 of file tuning_params.hh.

◆ stedc_wg_multiplier_for_n()

int32_t batchlas::tuning::stedc_wg_multiplier_for_n ( int32_t n )

inline

stedc merge work-group multiplier for order n; override BATCHLAS_TUNE_STEDC_WG_MULTIPLIER.

Definition at line 329 of file tuning_params.hh.

◆ env_truthy()

bool batchlas::env_truthy ( const char * v )

inline

True when v is exactly one of 1, true, TRUE, on, ON.

Parameters

v	the variable's value (std::getenv(...)), not its name; nullptr means unset

Returns: false for an unset variable and for every other spelling

Definition at line 29 of file env.hh.

◆ env_falsy()

bool batchlas::env_falsy ( const char * v )

inline

True when v is exactly one of 0, false, FALSE, off, OFF.

Not !env_truthy(v): an unset variable is neither truthy nor falsy, which lets a caller distinguish "forced off" from "not specified".

Parameters

v	the variable's value, or nullptr when unset

Returns: false for an unset variable and for every other spelling

Definition at line 44 of file env.hh.

◆ env_int_or()

int batchlas::env_int_or	(	const char *	name,
		int	fallback )

inline

The integer value of environment variable name.

Parameters

name	the variable's name
fallback	returned when the variable is unset or std::stoi rejects it

Returns: the parsed value; note that std::stoi accepts a numeric prefix ("16x" is 16)

Definition at line 55 of file env.hh.

◆ env_positive_int_or()

int batchlas::env_positive_int_or	(	const char *	name,
		int	fallback )

inline

As env_int_or(), but a value <= 0 also yields fallback.

The shape of every kernel-geometry knob: a forced work-group count, tile width or sub-group count is meaningless at zero or below.

Parameters

name	the variable's name
fallback	returned when unset, unparseable or non-positive

Definition at line 72 of file env.hh.

◆ env_string_or()

std::string batchlas::env_string_or	(	const char *	name,
		const std::string &	fallback )

inline

The value of environment variable name, or fallback when it is unset or empty.

Definition at line 79 of file env.hh.