blob: 0d11220dabab24ef502badf8908b4561f8c342d9 (
plain) (
blame)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
|
/* Copyright 2019-2020 Michael Rowan, Axel Huebl, Kevin Gott
*
* This file is part of WarpX.
*
* License: BSD-3-Clause-LBNL
*/
#ifndef ABLASTR_KERNELTIMER_H_
#define ABLASTR_KERNELTIMER_H_
#include "ablastr/utils/TextMsg.H"
#include <AMReX.H>
#include <AMReX_BLassert.H>
#include <AMReX_GpuAtomic.H>
#include <AMReX_GpuQualifiers.H>
#include <AMReX_REAL.H>
#include <climits>
namespace ablastr::parallelization
{
/**
* \brief Defines a timer object to be used on GPU; measures summed thread cycles.
*/
class KernelTimer
{
public:
/** Constructor.
* \param[in] do_timing Controls whether timer is active.
* \param[in,out] cost Pointer to cost which holds summed thread cycles
* (for performance, it is recommended to allocate pinned host memory).
*/
AMREX_GPU_DEVICE
KernelTimer
(const bool do_timing, amrex::Real* cost)
#if (defined AMREX_USE_GPU)
: m_do_timing(do_timing)
#endif
{
#if (defined AMREX_USE_GPU)
if (m_do_timing && cost) {
m_cost = cost;
# if defined(AMREX_USE_CUDA) || defined(AMREX_USE_HIP)
// Start the timer
m_wt = clock64();
# elif defined(AMREX_USE_DPCPP)
// To be updated
ABLASTR_ALWAYS_ASSERT_WITH_MESSAGE(m_do_timing == false,
"KernelTimer not yet supported for this hardware.");
# endif
}
#else // AMREX_USE_GPU
amrex::ignore_unused(do_timing, cost);
#endif // AMREX_USE_GPU
}
//! Destructor.
AMREX_GPU_DEVICE
~KernelTimer () {
#if (defined AMREX_USE_GPU)
if (m_do_timing && m_cost) {
# if defined(AMREX_USE_CUDA) || defined(AMREX_USE_HIP)
m_wt = clock64() - m_wt;
amrex::Gpu::Atomic::Add( m_cost, amrex::Real(m_wt));
# elif defined(AMREX_USE_DPCPP)
// To be updated
# endif
}
#endif //AMREX_USE_GPU
}
#if (defined AMREX_USE_GPU)
private:
//! Stores whether kernel timer is active.
bool m_do_timing;
//! Location in which to accumulate costs from all threads.
amrex::Real* m_cost = nullptr;
//! Store the time difference (cost) from a single thread.
long long int m_wt;
#endif //AMREX_USE_GPU
};
} // namespace ablastr::parallelization
#endif // ABLASTR_KERNELTIMER_H_
|
