intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2026-01-08 05:56:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

feature: enable BCS split for copy offload queues

Browse Source 
Related-To: NEO-14557

Signed-off-by: Bartosz Dunajski <bartosz.dunajski@intel.com>
This commit is contained in:
Bartosz Dunajski
2025-08-27 11:01:59 +00:00
committed by Compute-Runtime-Automation
parent e01c47152e
commit 2e58669fe9
9 changed files with 48 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
level_zero/core/source/cmdlist/cmdlist_hw.inl
View File

@@ -353,6 +353,8 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::initialize(Device *device, NEO
enableCopyOperationOffload();
}
enableImmediateBcsSplit();
return returnType;
}

12
level_zero/core/source/cmdlist/cmdlist_imp.cpp
View File

@@ -257,10 +257,6 @@ CommandList *CommandList::createImmediate(uint32_t productFamily, Device *device
commandList->isTbxMode = csr->isTbxMode();
commandList->commandListPreemptionMode = device->getDevicePreemptionMode();
if (!internalUsage) {
commandList->isBcsSplitNeeded = deviceImp->bcsSplit->setupDevice(csr);
}
commandList->copyThroughLockedPtrEnabled = gfxCoreHelper.copyThroughLockedPtrEnabled(hwInfo, productHelper);
const bool cmdListSupportsCopyOffload = commandList->isInOrderExecutionEnabled() && !productHelper.isDcFlushAllowed();
@@ -269,9 +265,17 @@ CommandList *CommandList::createImmediate(uint32_t productFamily, Device *device
commandList->enableCopyOperationOffload();
}
commandList->enableImmediateBcsSplit();
return commandList;
}
void CommandListImp::enableImmediateBcsSplit() {
if (device->getNEODevice()->isBcsSplitSupported() && isImmediateType() && !internalUsage && !isBcsSplitNeeded) {
isBcsSplitNeeded = static_cast<DeviceImp *>(getDevice())->bcsSplit->setupDevice(getCsr(false), isCopyOffloadEnabled());
}
}
void CommandListImp::enableCopyOperationOffload() {
if (isCopyOnly(false) || !static_cast<DeviceImp *>(device)->tryGetCopyEngineOrdinal().has_value()) {
return;

1
level_zero/core/source/cmdlist/cmdlist_imp.h
View File

@@ -56,6 +56,7 @@ struct CommandListImp : public CommandList {
uint64_t getInOrderExecDeviceGpuAddress() const;
size_t getInOrderExecHostRequiredSize() const;
uint64_t getInOrderExecHostGpuAddress() const;
void enableImmediateBcsSplit();
protected:
std::shared_ptr<NEO::InOrderExecInfo> inOrderExecInfo;

12
level_zero/core/source/device/bcs_split.cpp
View File

@@ -19,7 +19,7 @@
namespace L0 {
bool BcsSplit::setupDevice(NEO::CommandStreamReceiver *csr) {
bool BcsSplit::setupDevice(NEO::CommandStreamReceiver *csr, bool copyOffloadEnabled) {
auto &productHelper = this->device.getProductHelper();
auto bcsSplitSettings = productHelper.getBcsSplitSettings(this->device.getHwInfo());
@@ -29,12 +29,12 @@ bool BcsSplit::setupDevice(NEO::CommandStreamReceiver *csr) {
// If expectedTileCount==1, route root device to Tile0, otherwise use all Tiles
bool tileCountMatch = (bcsSplitSettings.requiredTileCount == 1) || (this->device.getNEODevice()->getNumSubDevices() == bcsSplitSettings.requiredTileCount);
bool engineMatch = (csr->getOsContext().getEngineType() == productHelper.getDefaultCopyEngine());
if (copyOffloadEnabled && NEO::debugManager.flags.SplitBcsForCopyOffload.get() != 0) {
engineMatch = NEO::EngineHelpers::isComputeEngine(csr->getOsContext().getEngineType());
}
auto initializeBcsSplit = this->device.getNEODevice()->isBcsSplitSupported() &&
(csr->getOsContext().getEngineType() == productHelper.getDefaultCopyEngine()) &&
tileCountMatch;
if (!initializeBcsSplit) {
if (!(engineMatch && tileCountMatch)) {
return false;
}

2
level_zero/core/source/device/bcs_split.h
View File

@@ -169,7 +169,7 @@ struct BcsSplit {
return result;
}
bool setupDevice(NEO::CommandStreamReceiver *csr);
bool setupDevice(NEO::CommandStreamReceiver *csr, bool copyOffloadEnabled);
void releaseResources();
std::vector<CommandList *> &getCmdListsForSplit(NEO::TransferDirection direction);
void setupEnginesMask(NEO::BcsSplitSettings &settings);

29
level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_blit.cpp
View File

@@ -1110,13 +1110,38 @@ HWTEST2_F(AggregatedBcsSplitTests, givenLimitedEnginesCountWhenCreatingBcsSplitT
BcsSplit bcsSplit(static_cast<L0::DeviceImp &>(*device));
bcsSplit.setupDevice(cmdList->getCsr(false));
bcsSplit.setupDevice(cmdList->getCsr(false), false);
EXPECT_EQ(expectedEnginesCount, bcsSplit.cmdLists.size());
bcsSplit.releaseResources();
}
HWTEST2_F(AggregatedBcsSplitTests, givenCopyOffloadEnabledWhenCreatingCmdListThenEnableBcsSplit, IsAtLeastXeHpcCore) {
debugManager.flags.ForceCopyOperationOffloadForComputeCmdList.set(1);
ze_result_t returnValue;
ze_command_queue_desc_t desc = {
.flags = ZE_COMMAND_QUEUE_FLAG_IN_ORDER,
};
std::unique_ptr<L0::CommandList> commandList1(CommandList::createImmediate(productFamily, device, &desc, false, NEO::EngineGroupType::compute, returnValue));
auto mockCmdList1 = static_cast<WhiteBox<L0::CommandListCoreFamilyImmediate<FamilyType::gfxCoreFamily>> *>(commandList1.get());
ASSERT_EQ(ZE_RESULT_SUCCESS, returnValue);
EXPECT_NE(device->getProductHelper().isDcFlushAllowed(), commandList1->isCopyOffloadEnabled());
EXPECT_EQ(commandList1->isCopyOffloadEnabled(), mockCmdList1->isBcsSplitNeeded);
debugManager.flags.SplitBcsForCopyOffload.set(0);
std::unique_ptr<L0::CommandList> commandList2(CommandList::createImmediate(productFamily, device, &desc, false, NEO::EngineGroupType::compute, returnValue));
auto mockCmdList2 = static_cast<WhiteBox<L0::CommandListCoreFamilyImmediate<FamilyType::gfxCoreFamily>> *>(commandList2.get());
ASSERT_EQ(ZE_RESULT_SUCCESS, returnValue);
EXPECT_FALSE(mockCmdList2->isBcsSplitNeeded);
}
HWTEST_F(AggregatedBcsSplitTests, givenTransferDirectionWhenAskingIfSplitIsNeededThenReturnCorrectValue) {
debugManager.flags.SplitBcsTransferDirectionMask.set(-1);
@@ -1133,7 +1158,7 @@ HWTEST2_F(AggregatedBcsSplitTests, givenPlatformSupporingAggregatedSplitModeWhen
BcsSplit bcsSplit(static_cast<L0::DeviceImp &>(*device));
bcsSplit.setupDevice(cmdList->getCsr(false));
bcsSplit.setupDevice(cmdList->getCsr(false), false);
EXPECT_EQ(device->getL0GfxCoreHelper().bcsSplitAggregatedModeEnabled(), bcsSplit.events.aggregatedEventsMode);

2
level_zero/core/test/unit_tests/sources/cmdlist/test_in_order_cmdlist_2.cpp
View File

@@ -4042,7 +4042,7 @@ struct BcsSplitInOrderCmdListTests : public InOrderCmdListFixture {
auto bcsSplit = static_cast<DeviceImp *>(device)->bcsSplit.get();
cmdList->isBcsSplitNeeded = bcsSplit->setupDevice(cmdList->getCsr(false));
cmdList->isBcsSplitNeeded = bcsSplit->setupDevice(cmdList->getCsr(false), false);
return cmdList;
}