intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-30 09:58:55 +08:00
Code Issues Packages Projects Releases Wiki Activity

Make migration for indirect allocations

Browse Source 
Signed-off-by: Maciej Plewka <maciej.plewka@intel.com>
This commit is contained in:
Maciej Plewka
2022-10-21 15:12:24 +00:00
committed by Compute-Runtime-Automation
parent 4f9498476d
commit 8a9ea9afd7
11 changed files with 211 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
level_zero/core/source/cmdlist/cmdlist.cpp
View File

@@ -144,19 +144,6 @@ NEO::PreemptionMode CommandList::obtainKernelPreemptionMode(Kernel *kernel) {
return NEO::PreemptionHelper::taskPreemptionMode(device->getDevicePreemptionMode(), flags);
}
void CommandList::makeResidentAndMigrate(bool performMigration) {
for (auto alloc : commandContainer.getResidencyContainer()) {
csr->makeResident(*alloc);
if (performMigration &&
(alloc->getAllocationType() == NEO::AllocationType::SVM_GPU ||
alloc->getAllocationType() == NEO::AllocationType::SVM_CPU)) {
auto pageFaultManager = device->getDriverHandle()->getMemoryManager()->getPageFaultManager();
pageFaultManager->moveAllocationToGpuDomain(reinterpret_cast<void *>(alloc->getGpuAddress()));
}
}
}
void CommandList::migrateSharedAllocations() {
auto deviceImp = static_cast<DeviceImp *>(device);
DriverHandleImp *driverHandleImp = static_cast<DriverHandleImp *>(deviceImp->getDriverHandle());

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

@@ -278,7 +278,6 @@ struct CommandList : _ze_command_list_handle_t {
return static_cast<uint32_t>(returnPoints.size());
}
void makeResidentAndMigrate(bool);
void migrateSharedAllocations();
bool getSystolicModeSupport() const {

4
level_zero/core/source/cmdlist/cmdlist_hw_immediate.inl
View File

@@ -110,7 +110,7 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::executeCommandListImm
std::unique_lock<std::mutex> lockForIndirect;
if (this->hasIndirectAllocationsAllowed()) {
this->cmdQImmediate->handleIndirectAllocationResidency(this->getUnifiedMemoryControls(), lockForIndirect);
this->cmdQImmediate->handleIndirectAllocationResidency(this->getUnifiedMemoryControls(), lockForIndirect, performMigration);
}
this->csr->setRequiredScratchSizes(this->getCommandListPerThreadScratchSize(), this->getCommandListPerThreadPrivateScratchSize());
@@ -123,7 +123,7 @@ ze_result_t CommandListCoreFamilyImmediate<gfxCoreFamily>::executeCommandListImm
}
}
this->makeResidentAndMigrate(performMigration);
this->cmdQImmediate->makeResidentAndMigrate(performMigration, this->commandContainer.getResidencyContainer());
if (performMigration) {
this->migrateSharedAllocations();

19
level_zero/core/source/cmdqueue/cmdqueue.cpp
View File

@@ -5,6 +5,7 @@
*
*/
#include "shared/source/command_container/cmdcontainer.h"
#include "shared/source/command_stream/command_stream_receiver.h"
#include "shared/source/command_stream/csr_definitions.h"
#include "shared/source/command_stream/linear_stream.h"
@@ -264,7 +265,7 @@ NEO::WaitStatus CommandQueueImp::CommandBufferManager::switchBuffers(NEO::Comman
return waitStatus;
}
void CommandQueueImp::handleIndirectAllocationResidency(UnifiedMemoryControls unifiedMemoryControls, std::unique_lock<std::mutex> &lockForIndirect) {
void CommandQueueImp::handleIndirectAllocationResidency(UnifiedMemoryControls unifiedMemoryControls, std::unique_lock<std::mutex> &lockForIndirect, bool performMigration) {
NEO::Device *neoDevice = this->device->getNEODevice();
auto svmAllocsManager = this->device->getDriverHandle()->getSvmAllocsManager();
auto submitAsPack = this->device->getDriverHandle()->getMemoryManager()->allowIndirectAllocationsAsPack(neoDevice->getRootDeviceIndex());
@@ -276,9 +277,23 @@ void CommandQueueImp::handleIndirectAllocationResidency(UnifiedMemoryControls un
svmAllocsManager->makeIndirectAllocationsResident(*(this->csr), this->csr->peekTaskCount() + 1u);
} else {
lockForIndirect = this->device->getDriverHandle()->getSvmAllocsManager()->obtainOwnership();
NEO::ResidencyContainer residencyAllocations;
svmAllocsManager->addInternalAllocationsToResidencyContainer(neoDevice->getRootDeviceIndex(),
this->csr->getResidencyAllocations(),
residencyAllocations,
unifiedMemoryControls.generateMask());
makeResidentAndMigrate(performMigration, residencyAllocations);
}
}
void CommandQueueImp::makeResidentAndMigrate(bool performMigration, const NEO::ResidencyContainer &residencyContainer) {
for (auto alloc : residencyContainer) {
csr->makeResident(*alloc);
if (performMigration &&
(alloc->getAllocationType() == NEO::AllocationType::SVM_GPU ||
alloc->getAllocationType() == NEO::AllocationType::SVM_CPU)) {
auto pageFaultManager = device->getDriverHandle()->getMemoryManager()->getPageFaultManager();
pageFaultManager->moveAllocationToGpuDomain(reinterpret_cast<void *>(alloc->getGpuAddress()));
}
}
}

8
level_zero/core/source/cmdqueue/cmdqueue.h
View File

@@ -13,12 +13,15 @@
#include <atomic>
#include <mutex>
#include <vector>
struct _ze_command_queue_handle_t {};
namespace NEO {
class CommandStreamReceiver;
}
class GraphicsAllocation;
using ResidencyContainer = std::vector<GraphicsAllocation *>;
} // namespace NEO
struct UnifiedMemoryControls;
@@ -52,7 +55,8 @@ struct CommandQueue : _ze_command_queue_handle_t {
return static_cast<CommandQueue *>(handle);
}
virtual void handleIndirectAllocationResidency(UnifiedMemoryControls unifiedMemoryControls, std::unique_lock<std::mutex> &lockForIndirect) = 0;
virtual void handleIndirectAllocationResidency(UnifiedMemoryControls unifiedMemoryControls, std::unique_lock<std::mutex> &lockForIndirect, bool performMigration) = 0;
virtual void makeResidentAndMigrate(bool performMigration, const NEO::ResidencyContainer &residencyContainer) = 0;
ze_command_queue_handle_t toHandle() { return this; }

4
level_zero/core/source/cmdqueue/cmdqueue_hw.inl
View File

@@ -119,7 +119,7 @@ ze_result_t CommandQueueHw<gfxCoreFamily>::executeCommandListsRegular(
std::unique_lock<std::mutex> lockForIndirect;
if (ctx.hasIndirectAccess) {
handleIndirectAllocationResidency(ctx.unifiedMemoryControls, lockForIndirect);
handleIndirectAllocationResidency(ctx.unifiedMemoryControls, lockForIndirect, ctx.isMigrationRequested);
}
size_t linearStreamSizeEstimate = this->estimateLinearStreamSizeInitial(ctx, phCommandLists, numCommandLists);
@@ -563,7 +563,7 @@ void CommandQueueHw<gfxCoreFamily>::setupCmdListsAndContextParams(
}
this->partitionCount = std::max(this->partitionCount, commandList->partitionCount);
commandList->makeResidentAndMigrate(ctx.isMigrationRequested);
makeResidentAndMigrate(ctx.isMigrationRequested, commandList->commandContainer.getResidencyContainer());
}
ctx.isDispatchTaskCountPostSyncRequired = isDispatchTaskCountPostSyncRequired(hFence, ctx.containsAnyRegularCmdList);

3
level_zero/core/source/cmdqueue/cmdqueue_imp.h
View File

@@ -83,7 +83,8 @@ struct CommandQueueImp : public CommandQueue {
MOCKABLE_VIRTUAL NEO::WaitStatus reserveLinearStreamSize(size_t size);
ze_command_queue_mode_t getSynchronousMode() const;
virtual bool getPreemptionCmdProgramming() = 0;
void handleIndirectAllocationResidency(UnifiedMemoryControls unifiedMemoryControls, std::unique_lock<std::mutex> &lockForIndirect) override;
void handleIndirectAllocationResidency(UnifiedMemoryControls unifiedMemoryControls, std::unique_lock<std::mutex> &lockForIndirect, bool performMigration) override;
void makeResidentAndMigrate(bool performMigration, const NEO::ResidencyContainer &residencyContainer) override;
protected:
MOCKABLE_VIRTUAL NEO::SubmissionStatus submitBatchBuffer(size_t offset, NEO::ResidencyContainer &residencyContainer, void *endingCmdPtr,