Media
/
compute-runtime
mirror of https://github.com/intel/compute-runtime.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

fix: use submit timestamp in kernel start time calculation 

Related-To: NEO-8394

Signed-off-by: Mateusz Jablonski <mateusz.jablonski@intel.com>
This commit is contained in:
Mateusz Jablonski 2023-10-20 11:10:54 +00:00 committed by Compute-Runtime-Automation
parent b8831e87b0
commit ad3aeb6eea
3 changed files with 56 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
opencl/source/event/event.cpp
View File

@ -370,17 +370,17 @@ void Event::calculateProfilingDataInternal(uint64_t contextStartTS, uint64_t con
auto &device = this->cmdQueue->getDevice();
auto &gfxCoreHelper = device.getGfxCoreHelper();
auto frequency = device.getDeviceInfo().profilingTimerResolution;
auto gpuQueueTimeStamp = gfxCoreHelper.getGpuTimeStampInNS(queueTimeStamp.gpuTimeStamp, frequency);
auto gpuSubmitTimeStamp = gfxCoreHelper.getGpuTimeStampInNS(submitTimeStamp.gpuTimeStamp, frequency);
if (DebugManager.flags.EnableDeviceBasedTimestamps.get()) {
startTimeStamp = static_cast<uint64_t>(globalStartTS * frequency);
while (startTimeStamp < gpuQueueTimeStamp) {
while (startTimeStamp < gpuSubmitTimeStamp) {
startTimeStamp += static_cast<uint64_t>((1ULL << gfxCoreHelper.getGlobalTimeStampBits()) * frequency);
}
} else {
int64_t c0 = queueTimeStamp.cpuTimeinNS - gpuQueueTimeStamp;
int64_t c0 = submitTimeStamp.cpuTimeinNS - gpuSubmitTimeStamp;
startTimeStamp = static_cast<uint64_t>(globalStartTS * frequency) + c0;
if (startTimeStamp < queueTimeStamp.cpuTimeinNS) {
if (startTimeStamp < submitTimeStamp.cpuTimeinNS) {
c0 += static_cast<uint64_t>((1ULL << (gfxCoreHelper.getGlobalTimeStampBits())) * frequency);
startTimeStamp = static_cast<uint64_t>(globalStartTS * frequency) + c0;
}

11
opencl/test/unit_test/event/event_tests.cpp
View File

@ -825,6 +825,10 @@ TEST_F(InternalsEventTest, givenDeviceTimestampBaseNotEnabledWhenCalculateStartT
timestamp.globalStartTS = 2;
event.queueTimeStamp.gpuTimeStamp = 1;
event.queueTimeStamp.cpuTimeinNS = 100;
event.queueTimeStamp.gpuTimeStamp = 2;
event.queueTimeStamp.cpuTimeinNS = 200;
TagNode<HwTimeStamps> timestampNode{};
timestampNode.tagForCpuAccess = &timestamp;
event.timeStampNode = &timestampNode;
@ -834,7 +838,7 @@ TEST_F(InternalsEventTest, givenDeviceTimestampBaseNotEnabledWhenCalculateStartT
auto resolution = pClDevice->getDevice().getDeviceInfo().profilingTimerResolution;
auto &gfxCoreHelper = pClDevice->getGfxCoreHelper();
auto c0 = event.queueTimeStamp.cpuTimeinNS - gfxCoreHelper.getGpuTimeStampInNS(event.queueTimeStamp.gpuTimeStamp, resolution);
auto c0 = event.submitTimeStamp.cpuTimeinNS - gfxCoreHelper.getGpuTimeStampInNS(event.submitTimeStamp.gpuTimeStamp, resolution);
EXPECT_EQ(start, static_cast<uint64_t>(timestamp.globalStartTS * resolution) + c0);
event.timeStampNode = nullptr;
@ -883,8 +887,9 @@ TEST_F(InternalsEventTest, givenDeviceTimestampBaseEnabledAndGlobalStartTSSmalle
MockEvent<Event> event(&cmdQ, CL_COMPLETE, 0, 0);
HwTimeStamps timestamp{};
timestamp.globalStartTS = 1;
timestamp.globalStartTS = 3;
event.queueTimeStamp.gpuTimeStamp = 2;
event.submitTimeStamp.gpuTimeStamp = 4;
TagNode<HwTimeStamps> timestampNode{};
timestampNode.tagForCpuAccess = &timestamp;
event.timeStampNode = &timestampNode;
@ -894,7 +899,7 @@ TEST_F(InternalsEventTest, givenDeviceTimestampBaseEnabledAndGlobalStartTSSmalle
auto &gfxCoreHelper = pClDevice->getGfxCoreHelper();
auto resolution = pClDevice->getDevice().getDeviceInfo().profilingTimerResolution;
auto refStartTime = static_cast<uint64_t>(timestamp.globalStartTS * resolution + (1ULL << gfxCoreHelper.getGlobalTimeStampBits()) * resolution);
auto refStartTime = static_cast<uint64_t>(timestamp.globalStartTS * resolution) + static_cast<uint64_t>((1ULL << gfxCoreHelper.getGlobalTimeStampBits()) * resolution);
EXPECT_EQ(start, refStartTime);
event.timeStampNode = nullptr;

44
opencl/test/unit_test/profiling/profiling_tests.cpp
View File

@ -625,6 +625,50 @@ HWCMDTEST_F(IGFX_GEN8_CORE, EventProfilingTests, givenRawTimestampsDebugModeWhen
event.timeStampNode = nullptr;
}
TEST_F(EventProfilingTests, givenSubmitTimeMuchGreaterThanQueueTimeWhenCalculatingStartTimeThenItIsGreaterThanSubmitTime) {
auto device = std::make_unique<MockClDevice>(MockDevice::createWithNewExecutionEnvironment<MockDevice>(nullptr));
MockContext context(device.get());
MockCommandQueue cmdQ(&context, device.get(), nullptr, false);
cmdQ.setProfilingEnabled();
HwTimeStamps timestamp{};
timestamp.globalStartTS = 10;
timestamp.contextStartTS = 20;
timestamp.globalEndTS = 80;
timestamp.contextEndTS = 56;
MockTagNode<HwTimeStamps> timestampNode{};
timestampNode.tagForCpuAccess = &timestamp;
MockEvent<Event> event(&cmdQ, CL_COMPLETE, 0, 0);
cl_event clEvent = &event;
event.queueTimeStamp.cpuTimeinNS = 1;
event.queueTimeStamp.gpuTimeStamp = 2;
event.submitTimeStamp.cpuTimeinNS = (1ull << 33) + 3;
event.submitTimeStamp.gpuTimeStamp = (1ull << 33) + 4;
event.timeStampNode = &timestampNode;
cl_ulong queued = 0ul;
cl_ulong submited = 0ul;
cl_ulong start = 0ul;
cl_ulong end = 0ul;
clGetEventProfilingInfo(clEvent, CL_PROFILING_COMMAND_QUEUED, sizeof(cl_ulong), &queued, nullptr);
clGetEventProfilingInfo(clEvent, CL_PROFILING_COMMAND_SUBMIT, sizeof(cl_ulong), &submited, nullptr);
clGetEventProfilingInfo(clEvent, CL_PROFILING_COMMAND_START, sizeof(cl_ulong), &start, nullptr);
clGetEventProfilingInfo(clEvent, CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &end, nullptr);
EXPECT_LT(0ull, queued);
EXPECT_LT(queued, submited);
EXPECT_LT(submited, start);
EXPECT_LT(start, end);
event.timeStampNode = nullptr;
}
HWCMDTEST_F(IGFX_GEN8_CORE, EventProfilingTest, givenRawTimestampsDebugModeWhenStartTimeStampLTQueueTimeStampThenIncreaseStartTimeStamp) {
DebugManagerStateRestore stateRestore;
DebugManager.flags.ReturnRawGpuTimestamps.set(1);