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
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
|
/*
* Copyright (C) 2020-2024 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
*/
#include "shared/source/os_interface/os_time.h"
#include "shared/source/debug_settings/debug_settings_manager.h"
#include "shared/source/helpers/debug_helpers.h"
#include "shared/source/helpers/hw_info.h"
#include <mutex>
namespace NEO {
double OSTime::getDeviceTimerResolution(HardwareInfo const &hwInfo) {
return hwInfo.capabilityTable.defaultProfilingTimerResolution;
};
bool DeviceTime::getGpuCpuTimeImpl(TimeStampData *pGpuCpuTime, OSTime *osTime) {
pGpuCpuTime->cpuTimeinNS = 0;
pGpuCpuTime->gpuTimeStamp = 0;
return true;
}
double DeviceTime::getDynamicDeviceTimerResolution(HardwareInfo const &hwInfo) const {
return OSTime::getDeviceTimerResolution(hwInfo);
}
uint64_t DeviceTime::getDynamicDeviceTimerClock(HardwareInfo const &hwInfo) const {
return static_cast<uint64_t>(1000000000.0 / OSTime::getDeviceTimerResolution(hwInfo));
}
void DeviceTime::setDeviceTimerResolution(HardwareInfo const &hwInfo) {
deviceTimerResolution = getDynamicDeviceTimerResolution(hwInfo);
if (debugManager.flags.OverrideProfilingTimerResolution.get() != -1) {
deviceTimerResolution = static_cast<double>(debugManager.flags.OverrideProfilingTimerResolution.get());
}
}
bool DeviceTime::isTimestampsRefreshEnabled() const {
bool timestampsRefreshEnabled = true;
if (debugManager.flags.EnableReusingGpuTimestamps.get() != -1) {
timestampsRefreshEnabled = debugManager.flags.EnableReusingGpuTimestamps.get();
}
return timestampsRefreshEnabled;
}
/**
* @brief If this method is called within interval, GPU timestamp
* will be calculated based on CPU timestamp and previous GPU ticks
* to reduce amount of internal KMD calls. Interval is selected
* adaptively, based on misalignment between calculated ticks and actual ticks.
*
* @return returns false if internal call to KMD failed. True otherwise.
*/
bool DeviceTime::getGpuCpuTimestamps(TimeStampData *timeStamp, OSTime *osTime, bool forceKmdCall) {
uint64_t cpuTimeinNS;
osTime->getCpuTime(&cpuTimeinNS);
auto cpuTimeDiffInNS = cpuTimeinNS - fetchedTimestamps.cpuTimeinNS;
if (forceKmdCall || cpuTimeDiffInNS >= timestampRefreshTimeoutNS) {
refreshTimestamps = true;
}
bool reusingTimestampsEnabled = isTimestampsRefreshEnabled();
if (!reusingTimestampsEnabled || refreshTimestamps) {
if (!getGpuCpuTimeImpl(timeStamp, osTime)) {
return false;
}
if (!reusingTimestampsEnabled) {
return true;
}
if (initialGpuTimeStamp) {
UNRECOVERABLE_IF(deviceTimerResolution == 0);
auto calculatedTimestamp = fetchedTimestamps.gpuTimeStamp + static_cast<uint64_t>(cpuTimeDiffInNS / deviceTimerResolution);
auto diff = abs(static_cast<int64_t>(timeStamp->gpuTimeStamp - calculatedTimestamp));
auto elapsedTicks = timeStamp->gpuTimeStamp - fetchedTimestamps.gpuTimeStamp;
int64_t adaptValue = static_cast<int64_t>(diff * deviceTimerResolution);
adaptValue = std::min(adaptValue, static_cast<int64_t>(timestampRefreshMinTimeoutNS));
if (diff * 1.0f / elapsedTicks > 0.05) {
adaptValue = adaptValue * (-1);
}
timestampRefreshTimeoutNS += adaptValue;
timestampRefreshTimeoutNS = std::max(timestampRefreshMinTimeoutNS, std::min(timestampRefreshMaxTimeoutNS, timestampRefreshTimeoutNS));
}
fetchedTimestamps = *timeStamp;
refreshTimestamps = false;
} else {
timeStamp->cpuTimeinNS = cpuTimeinNS;
UNRECOVERABLE_IF(deviceTimerResolution == 0);
timeStamp->gpuTimeStamp = fetchedTimestamps.gpuTimeStamp + static_cast<uint64_t>(cpuTimeDiffInNS / deviceTimerResolution);
}
return true;
}
bool DeviceTime::getGpuCpuTime(TimeStampData *pGpuCpuTime, OSTime *osTime, bool forceKmdCall) {
if (!getGpuCpuTimestamps(pGpuCpuTime, osTime, forceKmdCall)) {
return false;
}
auto maxGpuTimeStampValue = osTime->getMaxGpuTimeStamp();
static std::mutex gpuTimeStampOverflowCounterMutex;
std::lock_guard<std::mutex> lock(gpuTimeStampOverflowCounterMutex);
pGpuCpuTime->gpuTimeStamp &= (maxGpuTimeStampValue - 1);
if (!initialGpuTimeStamp) {
initialGpuTimeStamp = pGpuCpuTime->gpuTimeStamp;
waitingForGpuTimeStampOverflow = true;
} else {
if (waitingForGpuTimeStampOverflow && pGpuCpuTime->gpuTimeStamp < *initialGpuTimeStamp) {
gpuTimeStampOverflowCounter++;
waitingForGpuTimeStampOverflow = false;
}
if (!waitingForGpuTimeStampOverflow && pGpuCpuTime->gpuTimeStamp > *initialGpuTimeStamp) {
waitingForGpuTimeStampOverflow = true;
}
pGpuCpuTime->gpuTimeStamp += gpuTimeStampOverflowCounter * maxGpuTimeStampValue;
}
return true;
}
bool OSTime::getCpuTime(uint64_t *timeStamp) {
*timeStamp = 0;
return true;
}
double OSTime::getHostTimerResolution() const {
return 0;
}
uint64_t OSTime::getCpuRawTimestamp() {
return 0;
}
OSTime::OSTime(std::unique_ptr<DeviceTime> deviceTime) {
this->deviceTime = std::move(deviceTime);
}
} // namespace NEO
|
