openswmm::twoD::gpu::kperf Namespace Reference

Namespaces
namespace	detail

Classes
struct	Counters
struct	KernelEntry
class	Scope

Functions
void	install ()
	Registers the counting callbacks once per process. No-op unless OPENSWMM_PERF is set; defers to an external Kokkos tool if one is loaded.
void	reset ()
	Replace active solver counters with empty defaults.
void	count_advance ()
	Count one managed solver advance when profiling is enabled.
void	dump ()
	One [PERF-2D-KOKKOS] summary line plus one line per kernel label, sorted by host-side dispatch wall descending. Same key=value contract as [PERF-FV] / [PERF-LOAD]; tests/benchmarks scripts can scrape it.

Variables
thread_local Counters *	active = nullptr

Function Documentation

count_advance()

void openswmm::twoD::gpu::kperf::count_advance ( )

inline

Count one managed solver advance when profiling is enabled.

Postcondition

Increments the active n_advance counter only when both a Scope and perf::enabled() exist. No allocation, synchronization or solver advancement.

Here is the call graph for this function:

Here is the caller graph for this function:

dump()

void openswmm::twoD::gpu::kperf::dump ( )

inline

One [PERF-2D-KOKKOS] summary line plus one line per kernel label, sorted by host-side dispatch wall descending. Same key=value contract as [PERF-FV] / [PERF-LOAD]; tests/benchmarks scripts can scrape it.

Note

No-op without active counters, disabled profiling, or zero advances. Allocates a temporary pointer vector to sort labels and may throw; counters remain unchanged. Output goes to stderr and can interleave with other threads; label seconds measure host dispatch wall, not device kernel time. Requires stable exclusive access while reading counters.

Here is the call graph for this function:

Here is the caller graph for this function:

install()

void openswmm::twoD::gpu::kperf::install ( )

inline

Registers the counting callbacks once per process. No-op unless OPENSWMM_PERF is set; defers to an external Kokkos tool if one is loaded.

Note

First-call initialization is process-wide and synchronized by C++ local-static initialization. Skipping because performance is disabled or an external tool environment variable exists is permanent for later calls. Registers callbacks globally but collects only inside an active Scope; does not own counter objects or initialize Kokkos.

Here is the call graph for this function:

Here is the caller graph for this function:

reset()

void openswmm::twoD::gpu::kperf::reset ( )

inline

Replace active solver counters with empty defaults.

Postcondition

No-op without an active Scope; otherwise clears all counts, labels and kernel entries. Does not clear pairing timestamp stacks or remove global callbacks.

Precondition

No dispatch/fence timing pair may be open across reset. Requires exclusive active-counter access; prior entry references are invalidated.

Here is the caller graph for this function:

Variable Documentation

active

thread_local Counters* openswmm::twoD::gpu::kperf::active = nullptr

inline