intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-25 13:33:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

Introduce barrier tracking mechanism

Browse Source 
Related-To: NEO-7696

Signed-off-by: Lukasz Jobczyk <lukasz.jobczyk@intel.com>
This commit is contained in:
Lukasz Jobczyk
2023-02-06 10:12:34 +00:00
committed by Compute-Runtime-Automation
parent 6f3503af38
commit 9f574b6fba
6 changed files with 114 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
shared/source/command_stream/command_stream_receiver.cpp
View File

@@ -743,6 +743,8 @@ bool CommandStreamReceiver::initializeTagAllocation() {
userPauseConfirmation = Thread::create(CommandStreamReceiver::asyncDebugBreakConfirmation, reinterpret_cast<void *>(this));
}
this->barrierCountTagAddress = ptrOffset(this->tagAddress, TagAllocationLayout::barrierCountOffset);
return true;
}
@@ -1009,7 +1011,7 @@ TaskCountType CompletionStamp::getTaskCountFromSubmissionStatusError(SubmissionS
return 0;
}
}
uint64_t CommandStreamReceiver::getBarrierCountGpuAddress() const { return ptrOffset(this->tagAllocation->getGpuAddress(), TagAllocationLayout::barrierCountOffset); }
uint64_t CommandStreamReceiver::getDebugPauseStateGPUAddress() const { return tagAllocation->getGpuAddress() + TagAllocationLayout::debugPauseStateAddressOffset; }
uint64_t CommandStreamReceiver::getCompletionAddress() const {
uint64_t completionFenceAddress = castToUint64(const_cast<TagAddressType *>(tagAddress));

6
shared/source/command_stream/command_stream_receiver.h
View File

@@ -132,7 +132,11 @@ class CommandStreamReceiver {
return tagsMultiAllocation;
}
MultiGraphicsAllocation &createTagsMultiAllocation();
TaskCountType getNextBarrierCount() { return this->barrierCount.fetch_add(1u); }
volatile TagAddressType *getTagAddress() const { return tagAddress; }
volatile TagAddressType *getBarrierCountTagAddress() const { return this->barrierCountTagAddress; }
uint64_t getBarrierCountGpuAddress() const;
uint64_t getDebugPauseStateGPUAddress() const;
virtual bool waitForFlushStamp(FlushStamp &flushStampToWait) { return true; }
@@ -424,6 +428,7 @@ class CommandStreamReceiver {
uint64_t totalMemoryUsed = 0u;
volatile TagAddressType *tagAddress = nullptr;
volatile TagAddressType *barrierCountTagAddress = nullptr;
volatile DebugPauseState *debugPauseStateAddress = nullptr;
SpinLock debugPauseStateLock;
static void *asyncDebugBreakConfirmation(void *arg);
@@ -445,6 +450,7 @@ class CommandStreamReceiver {
OsContext *osContext = nullptr;
TaskCountType *completionFenceValuePointer = nullptr;
std::atomic<TaskCountType> barrierCount{0};
// current taskLevel. Used for determining if a PIPE_CONTROL is needed.
std::atomic<TaskCountType> taskLevel{0};
std::atomic<TaskCountType> latestSentTaskCount{0};

4
shared/source/command_stream/tag_allocation_layout.h
View File

@@ -1,5 +1,5 @@
/*
* Copyright (C) 2022 Intel Corporation
* Copyright (C) 2022-2023 Intel Corporation
*
* SPDX-License-Identifier: MIT
*
@@ -12,6 +12,6 @@ namespace NEO {
namespace TagAllocationLayout {
inline constexpr uint64_t debugPauseStateAddressOffset = MemoryConstants::kiloByte;
inline constexpr uint64_t completionFenceOffset = 2 * MemoryConstants::kiloByte;
inline constexpr uint64_t barrierCountOffset = 3 * MemoryConstants::kiloByte;
} // namespace TagAllocationLayout
} // namespace NEO