Media/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-09-20 13:11:34 +08:00
Code Issues Packages Projects Releases Wiki Activity

performance: Enable split for non-usm host

Browse Source 
Signed-off-by: Lukasz Jobczyk <lukasz.jobczyk@intel.com>
This commit is contained in:
Lukasz Jobczyk
2023-04-28 13:09:43 +00:00
committed by Compute-Runtime-Automation
parent 26f16f4e98
commit ff10e400c8
2 changed files with 93 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
level_zero/core/source/cmdlist/cmdlist_hw.inl
View File

@ -2669,11 +2669,9 @@ bool CommandListCoreFamily<gfxCoreFamily>::isAppendSplitNeeded(void *dstPtr, con
bool srcAllocFound = this->device->getDriverHandle()->findAllocationDataForRange(const_cast<void *>(srcPtr), size, &srcAllocData);
bool dstAllocFound = this->device->getDriverHandle()->findAllocationDataForRange(dstPtr, size, &dstAllocData);
if (srcAllocFound && dstAllocFound) {
return this->isAppendSplitNeeded(dstAllocData->gpuAllocations.getDefaultGraphicsAllocation()->getMemoryPool(), srcAllocData->gpuAllocations.getDefaultGraphicsAllocation()->getMemoryPool(), size, directionOut);
}
return false;
auto srcMemoryPool = srcAllocFound ? srcAllocData->gpuAllocations.getDefaultGraphicsAllocation()->getMemoryPool() : NEO::MemoryPool::System4KBPages;
auto dstMemoryPool = dstAllocFound ? dstAllocData->gpuAllocations.getDefaultGraphicsAllocation()->getMemoryPool() : NEO::MemoryPool::System4KBPages;
return this->isAppendSplitNeeded(dstMemoryPool, srcMemoryPool, size, directionOut);
}
template <GFXCORE_FAMILY gfxCoreFamily>

90
level_zero/core/test/unit_tests/xe_hpc_core/test_cmdqueue_xe_hpc_core.cpp
View File

@ -431,6 +431,96 @@ HWTEST2_F(CommandQueueCommandsXeHpc, givenSplitBcsCopyAndImmediateCommandListWhe
context->freeMem(dstPtr);
}
HWTEST2_F(CommandQueueCommandsXeHpc, givenSplitBcsCopyAndImmediateCommandListWhenAppendingMemoryCopyFromNonUsmHostToHostThenDoSplit, IsXeHpcCore) {
DebugManagerStateRestore restorer;
DebugManager.flags.SplitBcsCopy.set(1);
DebugManager.flags.EnableFlushTaskSubmission.set(0);
ze_result_t returnValue;
auto hwInfo = *NEO::defaultHwInfo;
hwInfo.featureTable.ftrBcsInfo = 0b111111111;
hwInfo.capabilityTable.blitterOperationsSupported = true;
auto testNeoDevice = NEO::MockDevice::createWithNewExecutionEnvironment<NEO::MockDevice>(&hwInfo);
auto testL0Device = std::unique_ptr<L0::Device>(L0::Device::create(driverHandle.get(), testNeoDevice, false, &returnValue));
ze_command_queue_desc_t desc = {};
desc.ordinal = static_cast<uint32_t>(testNeoDevice->getEngineGroupIndexFromEngineGroupType(NEO::EngineGroupType::Copy));
std::unique_ptr<L0::CommandList> commandList0(CommandList::createImmediate(productFamily,
testL0Device.get(),
&desc,
false,
NEO::EngineGroupType::Copy,
returnValue));
ASSERT_NE(nullptr, commandList0);
EXPECT_EQ(static_cast<DeviceImp *>(testL0Device.get())->bcsSplit.cmdQs.size(), 4u);
EXPECT_EQ(static_cast<CommandQueueImp *>(static_cast<DeviceImp *>(testL0Device.get())->bcsSplit.cmdQs[0])->getTaskCount(), 0u);
EXPECT_EQ(static_cast<CommandQueueImp *>(static_cast<DeviceImp *>(testL0Device.get())->bcsSplit.cmdQs[1])->getTaskCount(), 0u);
EXPECT_EQ(static_cast<CommandQueueImp *>(static_cast<DeviceImp *>(testL0Device.get())->bcsSplit.cmdQs[2])->getTaskCount(), 0u);
EXPECT_EQ(static_cast<CommandQueueImp *>(static_cast<DeviceImp *>(testL0Device.get())->bcsSplit.cmdQs[3])->getTaskCount(), 0u);
constexpr size_t alignment = 4096u;
constexpr size_t size = 8 * MemoryConstants::megaByte;
void *srcPtr = reinterpret_cast<void *>(0x1234);
void *dstPtr;
ze_host_mem_alloc_desc_t hostDesc = {};
context->allocHostMem(&hostDesc, size, alignment, &dstPtr);
auto result = commandList0->appendMemoryCopy(dstPtr, srcPtr, size, nullptr, 0, nullptr, false);
ASSERT_EQ(ZE_RESULT_SUCCESS, result);
EXPECT_EQ(static_cast<CommandQueueImp *>(static_cast<DeviceImp *>(testL0Device.get())->bcsSplit.cmdQs[0])->getTaskCount(), 1u);
EXPECT_EQ(static_cast<CommandQueueImp *>(static_cast<DeviceImp *>(testL0Device.get())->bcsSplit.cmdQs[1])->getTaskCount(), 1u);
EXPECT_EQ(static_cast<CommandQueueImp *>(static_cast<DeviceImp *>(testL0Device.get())->bcsSplit.cmdQs[2])->getTaskCount(), 1u);
EXPECT_EQ(static_cast<CommandQueueImp *>(static_cast<DeviceImp *>(testL0Device.get())->bcsSplit.cmdQs[3])->getTaskCount(), 1u);
context->freeMem(dstPtr);
}
HWTEST2_F(CommandQueueCommandsXeHpc, givenSplitBcsCopyAndImmediateCommandListWhenAppendingMemoryCopyFromHostToNonUsmHostThenDoSplit, IsXeHpcCore) {
DebugManagerStateRestore restorer;
DebugManager.flags.SplitBcsCopy.set(1);
DebugManager.flags.EnableFlushTaskSubmission.set(0);
ze_result_t returnValue;
auto hwInfo = *NEO::defaultHwInfo;
hwInfo.featureTable.ftrBcsInfo = 0b111111111;
hwInfo.capabilityTable.blitterOperationsSupported = true;
auto testNeoDevice = NEO::MockDevice::createWithNewExecutionEnvironment<NEO::MockDevice>(&hwInfo);
auto testL0Device = std::unique_ptr<L0::Device>(L0::Device::create(driverHandle.get(), testNeoDevice, false, &returnValue));
ze_command_queue_desc_t desc = {};
desc.ordinal = static_cast<uint32_t>(testNeoDevice->getEngineGroupIndexFromEngineGroupType(NEO::EngineGroupType::Copy));
std::unique_ptr<L0::CommandList> commandList0(CommandList::createImmediate(productFamily,
testL0Device.get(),
&desc,
false,
NEO::EngineGroupType::Copy,
returnValue));
ASSERT_NE(nullptr, commandList0);
EXPECT_EQ(static_cast<DeviceImp *>(testL0Device.get())->bcsSplit.cmdQs.size(), 4u);
EXPECT_EQ(static_cast<CommandQueueImp *>(static_cast<DeviceImp *>(testL0Device.get())->bcsSplit.cmdQs[0])->getTaskCount(), 0u);
EXPECT_EQ(static_cast<CommandQueueImp *>(static_cast<DeviceImp *>(testL0Device.get())->bcsSplit.cmdQs[1])->getTaskCount(), 0u);
EXPECT_EQ(static_cast<CommandQueueImp *>(static_cast<DeviceImp *>(testL0Device.get())->bcsSplit.cmdQs[2])->getTaskCount(), 0u);
EXPECT_EQ(static_cast<CommandQueueImp *>(static_cast<DeviceImp *>(testL0Device.get())->bcsSplit.cmdQs[3])->getTaskCount(), 0u);
constexpr size_t alignment = 4096u;
constexpr size_t size = 8 * MemoryConstants::megaByte;
void *srcPtr;
void *dstPtr = reinterpret_cast<void *>(0x1234);
ze_host_mem_alloc_desc_t hostDesc = {};
context->allocHostMem(&hostDesc, size, alignment, &srcPtr);
auto result = commandList0->appendMemoryCopy(dstPtr, srcPtr, size, nullptr, 0, nullptr, false);
ASSERT_EQ(ZE_RESULT_SUCCESS, result);
EXPECT_EQ(static_cast<CommandQueueImp *>(static_cast<DeviceImp *>(testL0Device.get())->bcsSplit.cmdQs[0])->getTaskCount(), 1u);
EXPECT_EQ(static_cast<CommandQueueImp *>(static_cast<DeviceImp *>(testL0Device.get())->bcsSplit.cmdQs[1])->getTaskCount(), 1u);
EXPECT_EQ(static_cast<CommandQueueImp *>(static_cast<DeviceImp *>(testL0Device.get())->bcsSplit.cmdQs[2])->getTaskCount(), 1u);
EXPECT_EQ(static_cast<CommandQueueImp *>(static_cast<DeviceImp *>(testL0Device.get())->bcsSplit.cmdQs[3])->getTaskCount(), 1u);
context->freeMem(srcPtr);
}
HWTEST2_F(CommandQueueCommandsXeHpc, givenSplitBcsCopyAndImmediateCommandListWhenAppendingMemoryCopyFromDeviceToDeviceThenDoNotSplit, IsXeHpcCore) {
using MI_FLUSH_DW = typename FamilyType::MI_FLUSH_DW;