intel/compute-runtime
1
0
Fork 0
mirror of https://github.com/intel/compute-runtime.git synced 2025-12-24 21:18:24 +08:00
Code Issues Packages Projects Releases Wiki Activity

Prefetch kmd-migrated shared allocation to multiple subdevices

Browse Source 
Related-To: NEO-6740

Signed-off-by: Slawomir Milczarek <slawomir.milczarek@intel.com>
This commit is contained in:
Slawomir Milczarek
2022-11-16 23:43:35 +00:00
committed by Compute-Runtime-Automation
parent 145f249f0c
commit ba003dd7dc
10 changed files with 100 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
shared/source/os_interface/linux/drm_allocation.cpp
View File

@@ -179,6 +179,21 @@ int DrmAllocation::bindBOs(OsContext *osContext, uint32_t vmHandleId, std::vecto
return 0;
}
bool DrmAllocation::prefetchBO(BufferObject *bo, uint32_t subDeviceId) {
auto drm = bo->peekDrm();
auto ioctlHelper = drm->getIoctlHelper();
auto memoryClassDevice = ioctlHelper->getDrmParamValue(DrmParam::MemoryClassDevice);
auto region = static_cast<uint32_t>((memoryClassDevice << 16u) | subDeviceId);
auto vmId = drm->getVirtualMemoryAddressSpace(subDeviceId);
auto result = ioctlHelper->setVmPrefetch(bo->peekAddress(), bo->peekSize(), region, vmId);
PRINT_DEBUG_STRING(DebugManager.flags.PrintBOPrefetchingResult.get(), stdout,
"prefetch BO=%d to VM %u, range: %llx - %llx, size: %lld, region: %x, result: %d\n",
bo->peekHandle(), vmId, bo->peekAddress(), ptrOffset(bo->peekAddress(), bo->peekSize()), bo->peekSize(), region, result);
return result;
}
void DrmAllocation::registerBOBindExtHandle(Drm *drm) {
if (!drm->resourceRegistrationEnabled()) {
return;
@@ -318,15 +333,18 @@ bool DrmAllocation::setMemAdvise(Drm *drm, MemAdviseFlags flags) {
bool DrmAllocation::setMemPrefetch(Drm *drm, uint32_t subDeviceId) {
bool success = true;
auto ioctlHelper = drm->getIoctlHelper();
auto memoryClassDevice = ioctlHelper->getDrmParamValue(DrmParam::MemoryClassDevice);
auto vmId = drm->getVirtualMemoryAddressSpace(subDeviceId);
auto numHandles = GraphicsAllocation::getNumHandlesForKmdSharedAllocation(storageInfo.getNumBanks());
for (auto bo : bufferObjects) {
if (bo != nullptr) {
auto region = static_cast<uint32_t>((memoryClassDevice << 16u) | subDeviceId);
success &= ioctlHelper->setVmPrefetch(bo->peekAddress(), bo->peekSize(), region, vmId);
if (numHandles > 1) {
for (uint8_t subDeviceId = 0u; subDeviceId < EngineLimits::maxHandleCount; subDeviceId++) {
if (storageInfo.memoryBanks.test(subDeviceId)) {
auto bo = this->getBOs()[subDeviceId];
success &= prefetchBO(bo, subDeviceId);
}
}
} else {
auto bo = this->getBO();
success = prefetchBO(bo, subDeviceId);
}
return success;

1
shared/source/os_interface/linux/drm_allocation.h
View File

@@ -109,6 +109,7 @@ class DrmAllocation : public GraphicsAllocation {
MOCKABLE_VIRTUAL int makeBOsResident(OsContext *osContext, uint32_t vmHandleId, std::vector<BufferObject *> *bufferObjects, bool bind);
MOCKABLE_VIRTUAL int bindBO(BufferObject *bo, OsContext *osContext, uint32_t vmHandleId, std::vector<BufferObject *> *bufferObjects, bool bind);
MOCKABLE_VIRTUAL int bindBOs(OsContext *osContext, uint32_t vmHandleId, std::vector<BufferObject *> *bufferObjects, bool bind);
MOCKABLE_VIRTUAL bool prefetchBO(BufferObject *bo, uint32_t subDeviceId);
MOCKABLE_VIRTUAL void registerBOBindExtHandle(Drm *drm);
void freeRegisteredBOBindExtHandles(Drm *drm);
void linkWithRegisteredHandle(uint32_t handle);

11
shared/source/os_interface/linux/drm_memory_manager.cpp
View File

@@ -1859,15 +1859,12 @@ GraphicsAllocation *DrmMemoryManager::createSharedUnifiedMemoryAllocation(const
BufferObjects bos{};
auto currentAddress = cpuPointer;
auto remainingSize = size;
auto getNumHandles = [](uint32_t numBanks) -> uint32_t {
return (numBanks > 1) && (DebugManager.flags.CreateKmdMigratedSharedAllocationWithMultipleBOs.get() != 0) ? numBanks : 1u;
};
auto numHandles = GraphicsAllocation::getNumHandlesForKmdSharedAllocation(allocationData.storageInfo.getNumBanks());
auto handles = getNumHandles(allocationData.storageInfo.getNumBanks());
for (auto handleId = 0u; handleId < handles; handleId++) {
for (auto handleId = 0u; handleId < numHandles; handleId++) {
uint32_t handle = 0;
auto currentSize = alignUp(remainingSize / (handles - handleId), MemoryConstants::pageSize64k);
auto currentSize = alignUp(remainingSize / (numHandles - handleId), MemoryConstants::pageSize64k);
if (currentSize == 0) {
break;
}
@@ -1917,7 +1914,7 @@ GraphicsAllocation *DrmMemoryManager::createSharedUnifiedMemoryAllocation(const
}
return nullptr;
}
if (handles > 1) {
if (numHandles > 1) {
allocation->storageInfo = allocationData.storageInfo;
}