reprospect.tools.ncu.metrics module

static create(*, mode: Literal['sass'] | None = 'sass', subs: tuple[MetricCounterRollUp, ...] = (MetricCounterRollUp.SUM,)) → tuple[MetricCounter, ...]View on GitHub 

class reprospect.tools.ncu.metrics.L1TEXCacheGlobalLoadRequestsView on GitHub 

Bases: object

Factory of counter metric l1tex__t_requests_pipe_lsu_mem_global_op_ld.

static create(*, subs: tuple[MetricCounterRollUp, ...] = (MetricCounterRollUp.SUM,)) → tuple[MetricCounter, ...]View on GitHub 

class reprospect.tools.ncu.metrics.L1TEXCacheGlobalLoadSectorHitsView on GitHub 

Bases: object

Factory of counter metric l1tex__t_sectors_pipe_lsu_mem_global_op_ld_lookup_hit.

static create(*, subs: tuple[MetricCounterRollUp, ...] = (MetricCounterRollUp.SUM,)) → tuple[MetricCounter, ...]View on GitHub 

class reprospect.tools.ncu.metrics.L1TEXCacheGlobalLoadSectorMissesView on GitHub 

Bases: object

Factory of counter metric l1tex__t_sectors_pipe_lsu_mem_global_op_ld_lookup_miss.

static create(*, subs: tuple[MetricCounterRollUp, ...] = (MetricCounterRollUp.SUM,)) → tuple[MetricCounter, ...]View on GitHub 

class reprospect.tools.ncu.metrics.L1TEXCacheGlobalLoadSectorsView on GitHub 

Bases: object

Factory of counter metric l1tex__t_sectors_pipe_lsu_mem_global_op_ld.

static create(*, subs: tuple[MetricCounterRollUp, ...] = (MetricCounterRollUp.SUM,), suffix: Literal['hit', 'miss'] | None = None) → tuple[MetricCounter, ...]View on GitHub 

class reprospect.tools.ncu.metrics.L1TEXCacheGlobalLoadWavefrontsView on GitHub 

Bases: object

Factory of counter metric l1tex__t_wavefronts_pipe_lsu_mem_global_op_ld.

static create(*, subs: tuple[MetricCounterRollUp, ...] = (MetricCounterRollUp.SUM,)) → tuple[MetricCounter, ...]View on GitHub 

class reprospect.tools.ncu.metrics.L1TEXCacheGlobalStoreView on GitHub 

Bases: object

InstructionsView on GitHub : alias of L1TEXCacheGlobalStoreInstructions

NAME: Final[str] = 'global store'

RequestsView on GitHub : alias of L1TEXCacheGlobalStoreRequests

SectorsView on GitHub : alias of L1TEXCacheGlobalStoreSectors

class reprospect.tools.ncu.metrics.L1TEXCacheGlobalStoreInstructionsView on GitHub 

Bases: object

Factory of counter metric (unit)__(sass?)_inst_executed_op_global_st.

static create(*, mode: Literal['sass'] | None = 'sass', subs: tuple[MetricCounterRollUp, ...] = (MetricCounterRollUp.SUM,)) → tuple[MetricCounter, ...]View on GitHub 

class reprospect.tools.ncu.metrics.L1TEXCacheGlobalStoreRequestsView on GitHub 

Bases: object

Factory of counter metric l1tex__t_requests_pipe_lsu_mem_global_op_st.

static create(*, subs: tuple[MetricCounterRollUp, ...] = (MetricCounterRollUp.SUM,)) → tuple[MetricCounter, ...]View on GitHub 

class reprospect.tools.ncu.metrics.L1TEXCacheGlobalStoreSectorsView on GitHub 

Bases: object

Factory of counter metric l1tex__t_sectors_pipe_lsu_mem_global_op_st.

static create(*, subs: tuple[MetricCounterRollUp, ...] = (MetricCounterRollUp.SUM,)) → tuple[MetricCounter, ...]View on GitHub 

class reprospect.tools.ncu.metrics.L1TEXCacheLocalStoreView on GitHub 

Bases: object

InstructionsView on GitHub : alias of L1TEXCacheLocalStoreInstructions

NAME: Final[str] = 'local store'

class reprospect.tools.ncu.metrics.L1TEXCacheLocalStoreInstructionsView on GitHub 

Bases: object

Factory of counter metric (unit)__(sass?)_inst_executed_op_local_st.

static create(*, mode: Literal['sass'] | None = 'sass', subs: tuple[MetricCounterRollUp, ...] = (MetricCounterRollUp.SUM,)) → tuple[MetricCounter, ...]View on GitHub 

class reprospect.tools.ncu.metrics.LaunchBlockView on GitHub 

Bases: XYZBase

Factory of metrics launch__block_dim_x, launch__block_dim_y and launch__block_dim_z.

PREFIX: ClassVar[str] = 'launch__block_dim_'

PRETTY_PREFIX: ClassVar[str | None] = 'Launch block size'

class reprospect.tools.ncu.metrics.LaunchGridView on GitHub 

Bases: XYZBase

Factory of metrics launch__grid_dim_x, launch__grid_dim_y and launch__grid_dim_z.

PREFIX: ClassVar[str] = 'launch__grid_dim_'

PRETTY_PREFIX: ClassVar[str | None] = 'Launch grid size'

class reprospect.tools.ncu.metrics.Metric(name: str, pretty_name: str | None = None, subs: tuple[str | tuple[str, ...], ...] | None = None)View on GitHub 

Bases: object

Used to represent a ncu metric.

If subs is not given, it is assumed that name is a valid metric that can be directly evaluated by ncu.

References:

https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#metrics-structure

Note

It is not decorated with dataclasses.dataclass() because of https://github.com/mypyc/mypyc/issues/1061.

SILENT_SUBS: ClassVar[frozenset[str]] = frozenset({}): Sub-metric components implied by the metric kind, omitted from labels.

__init__(name: str, pretty_name: str | None = None, subs: tuple[str | tuple[str, ...], ...] | None = None) → NoneView on GitHub 

Parameters:

name – The base name of the metric.
pretty_name – Human readable name; defaults to name.
subs – Sub-metric names. The constructor normalizes each to a path of components.

gather() → tuple[str, ...]View on GitHub : Get the list of sub-metric names or the metric name itself if no sub-metrics are defined.

labels() → tuple[str, ...]View on GitHub : Get the list of sub-metric labels. Parallel to gather(), but uses the pretty name.

name: Final[str]

pretty_name: Final[str | None]

subs: Final[tuple[tuple[str, ...], ...] | None]

class reprospect.tools.ncu.metrics.MetricCorrelation(name: str)View on GitHub 

Bases: object

A metric with correlations, like sass__inst_executed_per_opcode.

References:

https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#metrics-structure

__init__(name: str) → None

gather() → tuple[str]View on GitHub 

labels() → tuple[str]View on GitHub 

name: str

class reprospect.tools.ncu.metrics.MetricCorrelationData(correlated: dict[str | int, int | float], value: int | float | None = None)View on GitHub 

Bases: object

Data for MetricCorrelation.

References:

https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#metrics-structure

__init__(correlated: dict[str | int, int | float], value: int | float | None = None) → None

correlated: dict[str | int, int | float]

value: int | float | None

class reprospect.tools.ncu.metrics.MetricCounter(name: str, pretty_name: str | None = None, subs: tuple[str | tuple[str, ...], ...] | None = None)View on GitHub 

Bases: Metric

A counter metric.

The sub-metric names are expected to be from MetricCounterRollUp.

References:

https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#metrics-structure

SILENT_SUBS: ClassVar[frozenset[str]] = frozenset({MetricCounterRollUp.SUM}): Sub-metric components implied by the metric kind, omitted from labels.

name: Final[str]

pretty_name: Final[str | None]

subs: Final[tuple[tuple[str, ...], ...] | None]

class reprospect.tools.ncu.metrics.MetricCounterRollUp(*values)View on GitHub 

Bases: StrEnum

Available roll-ups for MetricCounter.

AVG = 'avg'

MAX = 'max'

MIN = 'min'

SUM = 'sum'

__str__(): Return str(self).

property label: strView on GitHub 

class reprospect.tools.ncu.metrics.MetricCounterRollUpQuantity(*values)View on GitHub 

Bases: StrEnum

Available quantities for MetricCounterRollUp.

PCT_OF_PEAK_SUSTAINED_ACTIVE = 'pct_of_peak_sustained_active'

PCT_OF_PEAK_SUSTAINED_ELAPSED = 'pct_of_peak_sustained_elapsed'

__str__(): Return str(self).

property label: strView on GitHub 

reprospect.tools.ncu.metrics.MetricData: TypeAlias = int | float | dict[str, int | float]: A single metric value type or a dictionary of submetric values of such type.

class reprospect.tools.ncu.metrics.MetricDeviceAttribute(name: str)View on GitHub 

Bases: object

ncu device attribute metric, such as:

device__attribute_architecture

Note

Available device attribute metrics can be queryied with:

ncu --query-metrics-collection=device

__init__(name: str) → None

property full_name: strView on GitHub 

gather() → tuple[str]View on GitHub 

labels() → tuple[str]View on GitHub 

name: str

class reprospect.tools.ncu.metrics.MetricRatio(name: str, pretty_name: str | None = None, subs: tuple[str | tuple[str, ...], ...] | None = None)View on GitHub 

Bases: Metric

A ratio metric.

The sub-metric names are expected to be from MetricRatioRollUp.

References:

https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#metrics-structure

SILENT_SUBS: ClassVar[frozenset[str]] = frozenset({MetricRatioRollUp.RATIO}): Sub-metric components implied by the metric kind, omitted from labels.

name: Final[str]

pretty_name: Final[str | None]

subs: Final[tuple[tuple[str, ...], ...] | None]

class reprospect.tools.ncu.metrics.MetricRatioRollUp(*values)View on GitHub 

Bases: StrEnum

Available roll-ups for MetricRatio.

MAX_RATE = 'max_rate'

PCT = 'pct'

RATIO = 'ratio'

__str__(): Return str(self).

property label: strView on GitHub 

class reprospect.tools.ncu.metrics.PipeStage(*values)View on GitHub 

Bases: StrEnum

Available pipe stages.

References:

https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#metrics-decoder

TAG = 't'

TAG_OUTPUT = 't_output'

__str__(): Return str(self).

class reprospect.tools.ncu.metrics.Quantity(*values)View on GitHub 

Bases: StrEnum

Available quantities.

References:

https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#metrics-decoder

INSTRUCTION = 'inst'

REQUEST = 'requests'

SECTOR = 'sectors'

WAVEFRONT = 'wavefronts'

__str__(): Return str(self).

class reprospect.tools.ncu.metrics.Unit(*values)View on GitHub 

Bases: StrEnum

Available units.

References:

https://docs.nvidia.com/nsight-compute/ProfilingGuide/#metrics-decoder

L1TEX = 'l1tex'

SM = 'sm'

SMSP = 'smsp'

__str__(): Return str(self).

reprospect.tools.ncu.metrics.ValueType: TypeAlias = int | float: A single metric value type.

class reprospect.tools.ncu.metrics.WarpStallView on GitHub 

Bases: object

A selection of metrics related to warp stall reasons.

References:

https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#warp-stall-reasons

LGThrottleView on GitHub : alias of WarpStallLGThrottle

LongScoreboardView on GitHub : alias of WarpStallLongScoreboard

MIOThrottleView on GitHub : alias of WarpStallMIOThrottle

ShortScoreboardView on GitHub : alias of WarpStallShortScoreboard

class reprospect.tools.ncu.metrics.WarpStallBaseView on GitHub 

Bases: object

Base class for factories of warp-stall ratio metrics.

PRETTY_REASON: ClassVar[str]: The stall reason as it appears in the label.

REASON: ClassVar[str]: The stall reason as it appears in the ncu metric name.

TEMPLATE_LABEL: Final[str] = 'Warp stall {reason}'

TEMPLATE_NAME: Final[str] = 'smsp__average_warps_issue_stalled_{reason}_per_issue_active'

classmethod create(*, subs: tuple[MetricRatioRollUp, ...] = (MetricRatioRollUp.RATIO,)) → tuple[MetricRatio, ...]View on GitHub 

class reprospect.tools.ncu.metrics.WarpStallLGThrottleView on GitHub 

Bases: WarpStallBase

Factory of ratio metric smsp__average_warps_issue_stalled_lg_throttle_per_issue_active.

PRETTY_REASON: ClassVar[str] = 'LG throttle': The stall reason as it appears in the label.

REASON: ClassVar[str] = 'lg_throttle': The stall reason as it appears in the ncu metric name.

class reprospect.tools.ncu.metrics.WarpStallLongScoreboardView on GitHub 

Bases: WarpStallBase

Factory of ratio metric smsp__average_warps_issue_stalled_long_scoreboard_per_issue_active.

PRETTY_REASON: ClassVar[str] = 'long scoreboard': The stall reason as it appears in the label.

REASON: ClassVar[str] = 'long_scoreboard': The stall reason as it appears in the ncu metric name.

class reprospect.tools.ncu.metrics.WarpStallMIOThrottleView on GitHub 

Bases: WarpStallBase

Factory of ratio metric smsp__average_warps_issue_stalled_mio_throttle_per_issue_active.

PRETTY_REASON: ClassVar[str] = 'MIO throttle': The stall reason as it appears in the label.

REASON: ClassVar[str] = 'mio_throttle': The stall reason as it appears in the ncu metric name.

class reprospect.tools.ncu.metrics.WarpStallShortScoreboardView on GitHub 

Bases: WarpStallBase

Factory of ratio metric smsp__average_warps_issue_stalled_short_scoreboard_per_issue_active.

PRETTY_REASON: ClassVar[str] = 'short scoreboard': The stall reason as it appears in the label.

REASON: ClassVar[str] = 'short_scoreboard': The stall reason as it appears in the ncu metric name.

class reprospect.tools.ncu.metrics.XYZBaseView on GitHub 

Bases: object

Base class for factories used to represent a triplet of metrics in x, y and z dimensions.

References:

https://docs.nvidia.com/nsight-compute/ProfilingGuide/#metrics-reference

PREFIX: ClassVar[str]

PRETTY_PREFIX: ClassVar[str | None] = None

classmethod create(dims: Iterable[str] | None = None) → tuple[Metric, ...]View on GitHub 

reprospect.tools.ncu.metrics.counter_name_from(*, unit: Unit, pipestage: PipeStage | None = None, quantity: Quantity | str, qualifier: str | None = None) → strView on GitHub 

Based on ncu metrics naming convention:

unit__(subunit?)_(pipestage?)_quantity_(qualifiers?)

References:

https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#metrics-structure

reprospect.tools.ncu.metrics.gather(metrics: Iterable[Metric | MetricCorrelation | MetricDeviceAttribute]) → tuple[str, ...]View on GitHub 

Retrieve all sub-metric names, e.g. to pass them to ncu.

Order follows metrics, then each metric’s subs; parallel to labels().

reprospect.tools.ncu.metrics.labels(metrics: Iterable[Metric | MetricCorrelation | MetricDeviceAttribute]) → tuple[str, ...]View on GitHub 

Retrieve all sub-metric labels, e.g. to lookup collected data by label.

Order follows metrics, then each metric’s subs; parallel to gather().