OPENSWMM_PERF=1 launch/fence/deep_copy counters for the Kokkos surface solver, riding the in-process Kokkos::Tools callbacks. More...

#include <Kokkos_Core.hpp>
#include <algorithm>
#include <chrono>
#include <cstdint>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <map>
#include <string>
#include <vector>
#include "../../core/PerfTimers.hpp"

Include dependency graph for KokkosPerfCounters.hpp:

This graph shows which files directly or indirectly include this file:

Go to the source code of this file.

Classes
struct	openswmm::twoD::gpu::kperf::KernelEntry

struct	openswmm::twoD::gpu::kperf::Counters

Namespaces
namespace	openswmm

namespace	openswmm::twoD

namespace	openswmm::twoD::gpu

namespace	openswmm::twoD::gpu::kperf

namespace	openswmm::twoD::gpu::kperf::detail

Typedefs
using	openswmm::twoD::gpu::kperf::detail::clock = std::chrono::steady_clock

Functions
void	openswmm::twoD::gpu::kperf::detail::begin_dispatch (const char name, std::uint32_t, std::uint64_t kID)

void	openswmm::twoD::gpu::kperf::detail::end_dispatch (std::uint64_t kID)

void	openswmm::twoD::gpu::kperf::detail::begin_fence (const char , std::uint32_t, std::uint64_t handle)

void	openswmm::twoD::gpu::kperf::detail::end_fence (std::uint64_t)

void	openswmm::twoD::gpu::kperf::detail::begin_deep_copy (Kokkos_Profiling_SpaceHandle dst, const char , const void , Kokkos_Profiling_SpaceHandle src, const char , const void , std::uint64_t size)

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

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

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

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

Variables
Counters	openswmm::twoD::gpu::kperf::g

std::vector< clock::time_point >	openswmm::twoD::gpu::kperf::detail::dispatch_t0

std::vector< clock::time_point >	openswmm::twoD::gpu::kperf::detail::fence_t0

Detailed Description

OPENSWMM_PERF=1 launch/fence/deep_copy counters for the Kokkos surface solver, riding the in-process Kokkos::Tools callbacks.

Phase A2 of plans/CPU_GPU_PERF_REVIEW_PLAN_2026-09-02.md. The counters answer three questions the wall clock cannot: how many kernel launches an advance() costs, how long the host blocks in fences, and how many bytes cross host<->device per advance. They hook Kokkos::Tools::Experimental::set_*_callback rather than the ~25 dispatch sites so every launch — including implicit fences and the deep_copies inside devCopy/devRefresh — is counted without a line of code in the hot path.

Registration is skipped when an external tool is loaded (KOKKOS_TOOLS_LIBS / KOKKOS_PROFILE_LIBRARY): Kokkos holds ONE callback set, and stealing it would blind nsys/kernel-logger, which attribute by the same kernel labels (plan Phase A3). Use one or the other per run.

Per-kernel s= is HOST-SIDE wall between the begin/end dispatch callbacks: on the OpenMP backend that is the kernel's execution time (dispatch blocks); on CUDA/HIP it is launch overhead only — device time lives in the fence/nsys numbers.

Counters are plain (non-atomic) longs: the engine drives the solver from one thread, and Kokkos dispatch callbacks fire on the calling thread. Zero cost when OPENSWMM_PERF is unset (nothing is registered at all).

Author: Caleb Buahin caleb.nosp@m..bua.nosp@m.hin@g.nosp@m.mail.nosp@m..com

License\n Apache-2.0

Classes

Namespaces

Typedefs

Functions

Variables

Detailed Description