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

Reuse command buffer for immediate cmdlist

Browse Source 
Signed-off-by: Szymon Morek <szymon.morek@intel.com>
This commit is contained in:
Szymon Morek
2022-09-27 11:46:03 +00:00
committed by Compute-Runtime-Automation
parent e11bcb9882
commit 6073702941
7 changed files with 153 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

62
shared/source/command_container/cmdcontainer.cpp
View File

@@ -272,8 +272,9 @@ void CommandContainer::handleCmdBufferAllocations(size_t startIndex) {
if (isHandleFenceCompletionRequired) {
this->device->getMemoryManager()->handleFenceCompletion(cmdBufferAllocations[i]);
}
reusableAllocationList->pushFrontOne(*cmdBufferAllocations[i]);
if (!this->reusableAllocationList->peekContains(*cmdBufferAllocations[i])) {
reusableAllocationList->pushFrontOne(*cmdBufferAllocations[i]);
}
} else {
this->device->getMemoryManager()->freeGraphicsMemory(cmdBufferAllocations[i]);
}
@@ -281,22 +282,14 @@ void CommandContainer::handleCmdBufferAllocations(size_t startIndex) {
}
GraphicsAllocation *CommandContainer::obtainNextCommandBufferAllocation() {
size_t alignedSize = alignUp<size_t>(this->getTotalCmdBufferSize(), MemoryConstants::pageSize64k);
GraphicsAllocation *cmdBufferAllocation = nullptr;
if (this->reusableAllocationList) {
size_t alignedSize = alignUp<size_t>(this->getTotalCmdBufferSize(), MemoryConstants::pageSize64k);
cmdBufferAllocation = this->reusableAllocationList->detachAllocation(alignedSize, nullptr, nullptr, AllocationType::COMMAND_BUFFER).release();
}
if (!cmdBufferAllocation) {
AllocationProperties properties{device->getRootDeviceIndex(),
true /* allocateMemory*/,
alignedSize,
AllocationType::COMMAND_BUFFER,
(device->getNumGenericSubDevices() > 1u) /* multiOsContextCapable */,
false,
device->getDeviceBitfield()};
cmdBufferAllocation = device->getMemoryManager()->allocateGraphicsMemoryWithProperties(properties);
cmdBufferAllocation = this->allocateCommandBuffer();
}
return cmdBufferAllocation;
@@ -308,13 +301,7 @@ void CommandContainer::allocateNextCommandBuffer() {
cmdBufferAllocations.push_back(cmdBufferAllocation);
size_t alignedSize = alignUp<size_t>(this->getTotalCmdBufferSize(), MemoryConstants::pageSize64k);
commandStream->replaceBuffer(cmdBufferAllocation->getUnderlyingBuffer(), alignedSize - cmdBufferReservedSize);
commandStream->replaceGraphicsAllocation(cmdBufferAllocation);
if (!getFlushTaskUsedForImmediate()) {
addToResidencyContainer(cmdBufferAllocation);
}
setCmdBuffer(cmdBufferAllocation);
}
void CommandContainer::closeAndAllocateNextCommandBuffer() {
@@ -361,4 +348,41 @@ void CommandContainer::ensureHeapSizePrepared(size_t sshRequiredSize, size_t dsh
}
}
GraphicsAllocation *CommandContainer::reuseExistingCmdBuffer() {
size_t alignedSize = alignUp<size_t>(this->getTotalCmdBufferSize(), MemoryConstants::pageSize64k);
auto cmdBufferAllocation = this->reusableAllocationList->detachAllocation(alignedSize, nullptr, this->immediateCmdListCsr, AllocationType::COMMAND_BUFFER).release();
return cmdBufferAllocation;
}
void CommandContainer::addCurrentCommandBufferToReusableAllocationList() {
auto taskCount = this->immediateCmdListCsr->peekTaskCount() + 1;
auto osContextId = this->immediateCmdListCsr->getOsContext().getContextId();
commandStream->getGraphicsAllocation()->updateTaskCount(taskCount, osContextId);
commandStream->getGraphicsAllocation()->updateResidencyTaskCount(taskCount, osContextId);
this->reusableAllocationList->pushTailOne(*this->commandStream->getGraphicsAllocation());
}
void CommandContainer::setCmdBuffer(GraphicsAllocation *cmdBuffer) {
size_t alignedSize = alignUp<size_t>(this->getTotalCmdBufferSize(), MemoryConstants::pageSize64k);
commandStream->replaceBuffer(cmdBuffer->getUnderlyingBuffer(), alignedSize - cmdBufferReservedSize);
commandStream->replaceGraphicsAllocation(cmdBuffer);
if (!getFlushTaskUsedForImmediate()) {
addToResidencyContainer(cmdBuffer);
}
}
GraphicsAllocation *CommandContainer::allocateCommandBuffer() {
size_t alignedSize = alignUp<size_t>(this->getTotalCmdBufferSize(), MemoryConstants::pageSize64k);
AllocationProperties properties{device->getRootDeviceIndex(),
true /* allocateMemory*/,
alignedSize,
AllocationType::COMMAND_BUFFER,
(device->getNumGenericSubDevices() > 1u) /* multiOsContextCapable */,
false,
device->getDeviceBitfield()};
return device->getMemoryManager()->allocateGraphicsMemoryWithProperties(properties);
}
} // namespace NEO

10
shared/source/command_container/cmdcontainer.h
View File

@@ -105,10 +105,17 @@ class CommandContainer : public NonCopyableOrMovableClass {
void setImmediateCmdListCsr(CommandStreamReceiver *newValue) {
this->immediateCmdListCsr = newValue;
}
void enableHeapSharing() { heapSharingEnabled = true; }
bool immediateCmdListSharedHeap(HeapType heapType) {
return (this->immediateCmdListCsr != nullptr && (heapType == HeapType::DYNAMIC_STATE || heapType == HeapType::SURFACE_STATE));
return (heapSharingEnabled && (heapType == HeapType::DYNAMIC_STATE || heapType == HeapType::SURFACE_STATE));
}
void ensureHeapSizePrepared(size_t sshRequiredSize, size_t dshRequiredSize);
GraphicsAllocation *reuseExistingCmdBuffer();
GraphicsAllocation *allocateCommandBuffer();
void setCmdBuffer(GraphicsAllocation *cmdBuffer);
void addCurrentCommandBufferToReusableAllocationList();
HeapContainer sshAllocations;
uint64_t currentLinearStreamStartOffset = 0u;
uint32_t slmSize = std::numeric_limits<uint32_t>::max();
@@ -146,6 +153,7 @@ class CommandContainer : public NonCopyableOrMovableClass {
bool isFlushTaskUsedForImmediate = false;
bool isHandleFenceCompletionRequired = false;
bool heapSharingEnabled = false;
};
} // namespace NEO

67
shared/test/unit_test/command_container/command_container_tests.cpp
View File

@@ -827,6 +827,72 @@ TEST_F(CommandContainerTest, givenCmdContainerWhenCloseAndAllocateNextCommandBuf
EXPECT_EQ(cmdContainer.getCmdBufferAllocations().size(), 2u);
}
TEST_F(CommandContainerTest, givenCmdContainerWhenSetCmdBufferThenCmdBufferSetCorrectly) {
CommandContainer cmdContainer;
cmdContainer.initialize(pDevice, nullptr, true);
AllocationProperties properties{pDevice->getRootDeviceIndex(),
true /* allocateMemory*/,
2048,
AllocationType::COMMAND_BUFFER,
(pDevice->getNumGenericSubDevices() > 1u) /* multiOsContextCapable */,
false,
pDevice->getDeviceBitfield()};
auto alloc = pDevice->getMemoryManager()->allocateGraphicsMemoryWithProperties(properties);
cmdContainer.setCmdBuffer(alloc);
EXPECT_EQ(cmdContainer.getCommandStream()->getGraphicsAllocation(), alloc);
pDevice->getMemoryManager()->freeGraphicsMemory(alloc);
}
TEST_F(CommandContainerTest, givenCmdContainerWhenReuseExistingCmdBufferWithoutAnyAllocationInListThenReturnNullptr) {
auto cmdContainer = std::make_unique<CommandContainer>();
AllocationsList allocList;
cmdContainer->initialize(pDevice, &allocList, false);
auto csr = pDevice->getDefaultEngine().commandStreamReceiver;
cmdContainer->setImmediateCmdListCsr(csr);
EXPECT_EQ(cmdContainer->reuseExistingCmdBuffer(), nullptr);
cmdContainer.reset();
allocList.freeAllGraphicsAllocations(pDevice);
}
HWTEST_F(CommandContainerTest, givenCmdContainerWhenReuseExistingCmdBufferWithAllocationInListAndCsrTaskCountLowerThanAllocationThenReturnNullptr) {
auto cmdContainer = std::make_unique<CommandContainer>();
auto &csr = pDevice->getUltCommandStreamReceiver<FamilyType>();
*csr.tagAddress = 0u;
AllocationsList allocList;
cmdContainer->initialize(pDevice, &allocList, false);
cmdContainer->setImmediateCmdListCsr(&csr);
cmdContainer->getCmdBufferAllocations()[0]->updateTaskCount(10, 0);
cmdContainer->addCurrentCommandBufferToReusableAllocationList();
EXPECT_EQ(cmdContainer->reuseExistingCmdBuffer(), nullptr);
cmdContainer.reset();
allocList.freeAllGraphicsAllocations(pDevice);
}
HWTEST_F(CommandContainerTest, givenCmdContainerWhenReuseExistingCmdBufferWithAllocationInListAndCsrTaskCountSameAsAllocationThenReturnAlloc) {
auto cmdContainer = std::make_unique<CommandContainer>();
auto &csr = pDevice->getUltCommandStreamReceiver<FamilyType>();
*csr.tagAddress = 10u;
AllocationsList allocList;
cmdContainer->initialize(pDevice, &allocList, false);
cmdContainer->setImmediateCmdListCsr(&csr);
cmdContainer->getCmdBufferAllocations()[0]->updateTaskCount(10, 0);
cmdContainer->addCurrentCommandBufferToReusableAllocationList();
EXPECT_NE(cmdContainer->reuseExistingCmdBuffer(), nullptr);
cmdContainer.reset();
allocList.freeAllGraphicsAllocations(pDevice);
}
TEST_F(CommandContainerTest, GivenCmdContainerWhenContainerIsInitializedThenSurfaceStateIndirectHeapSizeIsCorrect) {
MyMockCommandContainer cmdContainer;
cmdContainer.initialize(pDevice, nullptr, true);
@@ -837,6 +903,7 @@ TEST_F(CommandContainerTest, GivenCmdContainerWhenContainerIsInitializedThenSurf
HWTEST_F(CommandContainerTest, givenCmdContainerHasImmediateCsrWhenGettingHeapWithoutEnsuringSpaceThenExpectNullptrReturnedOrUnrecoverable) {
CommandContainer cmdContainer;
cmdContainer.enableHeapSharing();
cmdContainer.setImmediateCmdListCsr(pDevice->getDefaultEngine().commandStreamReceiver);
cmdContainer.setNumIddPerBlock(1);
auto code = cmdContainer.initialize(pDevice, nullptr, true);