Add debug flag to bind at creation time

Signed-off-by: Lukasz Jobczyk <[email protected]>

Author: LukaszJobczyk

opencl/test/unit_test/test_files/igdrcl.config

@@ -369,6 +369,7 @@ Force2dImageAsArray = -1
 ForceExtendedBufferSize = -1
 ForceExtendedUSMBufferSize = -1
 MakeIndirectAllocationsResidentAsPack = -1
+MakeEachAllocationResident = -1
 EnableChipsetUniqueUUID = -1
 ForceSimdMessageSizeInWalker = -1
 UseNewQueryTopoIoctl = 1

shared/source/command_stream/command_stream_receiver.cpp

@@ -115,7 +115,16 @@ void CommandStreamReceiver::makeResident(MultiGraphicsAllocation &gfxAllocation)
 void CommandStreamReceiver::makeResident(GraphicsAllocation &gfxAllocation) {
     auto submissionTaskCount = this->taskCount + 1;
     if (gfxAllocation.isResidencyTaskCountBelow(submissionTaskCount, osContext->getContextId())) {
-        this->getResidencyAllocations().push_back(&gfxAllocation);
+        auto pushAllocations = true;
+
+        if (DebugManager.flags.MakeEachAllocationResident.get() != -1) {
+            pushAllocations = !DebugManager.flags.MakeEachAllocationResident.get();
+        }
+
+        if (pushAllocations) {
+            this->getResidencyAllocations().push_back(&gfxAllocation);
+        }
+
         checkForNewResources(submissionTaskCount, gfxAllocation.getTaskCount(osContext->getContextId()), gfxAllocation);
         gfxAllocation.updateTaskCount(submissionTaskCount, osContext->getContextId());
         if (!gfxAllocation.isResident(osContext->getContextId())) {

shared/source/debug_settings/debug_variables_base.inl

@@ -242,6 +242,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, OverrideKernelSizeLimitForSmallDispatch, -1, "-1
 DECLARE_DEBUG_VARIABLE(int32_t, OverrideUseKmdWaitFunction, -1, "-1: default (L0: disabled), 0: disabled, 1: enabled. It uses only busy loop to wait or busy loop with KMD wait function, when KMD fallback is enabled")
 DECLARE_DEBUG_VARIABLE(int32_t, ResolveDependenciesViaPipeControls, -1, "-1: default , 0: disabled, 1: enabled. If enabled, instead of programming semaphores, dependencies are resolved using task levels")
 DECLARE_DEBUG_VARIABLE(int32_t, MakeIndirectAllocationsResidentAsPack, -1, "-1: default, 0:disabled, 1: enabled. If enabled, driver handles all indirect allocations as one pack instead of making them resident individually.")
+DECLARE_DEBUG_VARIABLE(int32_t, MakeEachAllocationResident, -1, "-1: default, 0: disabled, 1: bind every allocation at creation time, 2: bind all created allocations in flush")
 
 /*DIRECT SUBMISSION FLAGS*/
 DECLARE_DEBUG_VARIABLE(bool, DirectSubmissionPrintBuffers, false, "Print address of submitted command buffers")

shared/source/os_interface/linux/drm_memory_manager.cpp

@@ -1046,15 +1046,28 @@ std::vector<GraphicsAllocation *> &DrmMemoryManager::getLocalMemAllocs(uint32_t
     return this->localMemAllocs[rootDeviceIndex];
 }
 
+void DrmMemoryManager::makeAllocationResident(GraphicsAllocation *allocation) {
+    if (DebugManager.flags.MakeEachAllocationResident.get() == 1) {
+        auto drmAllocation = static_cast<DrmAllocation *>(allocation);
+        for (uint32_t i = 0; getDrm(allocation->getRootDeviceIndex()).getVirtualMemoryAddressSpace(i) > 0u; i++) {
+            drmAllocation->makeBOsResident(registeredEngines[defaultEngineIndex[allocation->getRootDeviceIndex()]].osContext, i, nullptr, true);
+            getDrm(allocation->getRootDeviceIndex()).waitForBind(i);
+        }
+    }
+}
+
 void DrmMemoryManager::registerSysMemAlloc(GraphicsAllocation *allocation) {
+    makeAllocationResident(allocation);
     std::lock_guard<std::mutex> lock(this->allocMutex);
     this->sysMemAllocs.push_back(allocation);
 }
 
 void DrmMemoryManager::registerLocalMemAlloc(GraphicsAllocation *allocation, uint32_t rootDeviceIndex) {
+    makeAllocationResident(allocation);
     std::lock_guard<std::mutex> lock(this->allocMutex);
     this->localMemAllocs[rootDeviceIndex].push_back(allocation);
 }
+
 void DrmMemoryManager::unregisterAllocation(GraphicsAllocation *allocation) {
     std::lock_guard<std::mutex> lock(this->allocMutex);
     sysMemAllocs.erase(std::remove(sysMemAllocs.begin(), sysMemAllocs.end(), allocation),

shared/source/os_interface/linux/drm_memory_manager.h

@@ -123,6 +123,7 @@ class DrmMemoryManager : public MemoryManager {
     void registerAllocationInOs(GraphicsAllocation *allocation) override;
     void waitOnCompletionFence(GraphicsAllocation *allocation);
     bool allocationTypeForCompletionFence(AllocationType allocationType);
+    void makeAllocationResident(GraphicsAllocation *allocation);
 
     Drm &getDrm(uint32_t rootDeviceIndex) const;
     uint32_t getRootDeviceIndex(const Drm *drm);

shared/source/os_interface/linux/drm_memory_operations_handler_bind.cpp

@@ -104,7 +104,15 @@ MemoryOperationsStatus DrmMemoryOperationsHandlerBind::isResident(Device *device
 }
 
 MemoryOperationsStatus DrmMemoryOperationsHandlerBind::mergeWithResidencyContainer(OsContext *osContext, ResidencyContainer &residencyContainer) {
-    MemoryOperationsStatus retVal = this->makeResidentWithinOsContext(osContext, ArrayRef<GraphicsAllocation *>(residencyContainer), true);
+    if (DebugManager.flags.MakeEachAllocationResident.get() == 2) {
+        auto memoryManager = static_cast<DrmMemoryManager *>(this->rootDeviceEnvironment.executionEnvironment.memoryManager.get());
+
+        auto allocLock = memoryManager->acquireAllocLock();
+        this->makeResidentWithinOsContext(osContext, ArrayRef<GraphicsAllocation *>(memoryManager->getSysMemAllocs()), true);
+        this->makeResidentWithinOsContext(osContext, ArrayRef<GraphicsAllocation *>(memoryManager->getLocalMemAllocs(this->rootDeviceIndex)), true);
+    }
+
+    auto retVal = this->makeResidentWithinOsContext(osContext, ArrayRef<GraphicsAllocation *>(residencyContainer), true);
     if (retVal != MemoryOperationsStatus::SUCCESS) {
         return retVal;
     }