Set Uncached MOCS for List only given stateless uncached args

nrspruit · Compute-Runtime-Automation · commit 6d8502847ebe · 2021-11-18T03:35:48.000+01:00
Signed-off-by: Spruit, Neil R &lt;neil.r.spruit@intel.com&gt;
diff --git a/level_zero/core/source/cmdlist/cmdlist.h b/level_zero/core/source/cmdlist/cmdlist.h
@@ -250,7 +250,7 @@ struct CommandList : _ze_command_list_handle_t {
     bool isFlushTaskSubmissionEnabled = false;
     bool isSyncModeQueue = false;
     bool commandListSLMEnabled = false;
-    bool requiresUncachedMOCS = false;
+    bool requiresQueueUncachedMocs = false;
 
   protected:
     NEO::GraphicsAllocation *getAllocationFromHostPtrMap(const void *buffer, uint64_t bufferSize);
diff --git a/level_zero/core/source/cmdlist/cmdlist_hw_base.inl b/level_zero/core/source/cmdlist/cmdlist_hw_base.inl
@@ -109,7 +109,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(z
 
     KernelImp *kernelImp = static_cast<KernelImp *>(kernel);
     this->containsStatelessUncachedResource |= kernelImp->getKernelRequiresUncachedMocs();
-    this->requiresUncachedMOCS = this->containsStatelessUncachedResource;
+    this->requiresQueueUncachedMocs |= kernelImp->getKernelRequiresQueueUncachedMocs();
     uint32_t partitionCount = 0;
 
     NEO::Device *neoDevice = device->getNEODevice();
diff --git a/level_zero/core/source/cmdlist/cmdlist_hw_xehp_and_later.inl b/level_zero/core/source/cmdlist/cmdlist_hw_xehp_and_later.inl
@@ -225,6 +225,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(z
 
     KernelImp *kernelImp = static_cast<KernelImp *>(kernel);
     this->containsStatelessUncachedResource |= kernelImp->getKernelRequiresUncachedMocs();
+    this->requiresQueueUncachedMocs |= kernelImp->getKernelRequiresQueueUncachedMocs();
     this->threadArbitrationPolicy = kernelImp->getSchedulingHintExp();
 
     uint32_t partitionCount = 0;
diff --git a/level_zero/core/source/cmdqueue/cmdqueue_hw.h b/level_zero/core/source/cmdqueue/cmdqueue_hw.h
@@ -35,7 +35,7 @@ struct CommandQueueHw : public CommandQueueImp {
 
     void dispatchTaskCountWrite(NEO::LinearStream &commandStream, bool flushDataCache) override;
 
-    void programStateBaseAddress(uint64_t gsba, bool useLocalMemoryForIndirectHeap, NEO::LinearStream &commandStream);
+    void programStateBaseAddress(uint64_t gsba, bool useLocalMemoryForIndirectHeap, NEO::LinearStream &commandStream, bool cachedMOCSAllowed);
     size_t estimateStateBaseAddressCmdSize();
     MOCKABLE_VIRTUAL void programFrontEnd(uint64_t scratchAddress, uint32_t perThreadScratchSpaceSize, NEO::LinearStream &commandStream);
 
diff --git a/level_zero/core/source/cmdqueue/cmdqueue_hw.inl b/level_zero/core/source/cmdqueue/cmdqueue_hw.inl
@@ -82,7 +82,7 @@ ze_result_t CommandQueueHw<gfxCoreFamily>::executeCommandLists(
     auto anyCommandListWithCooperativeKernels = false;
     auto anyCommandListWithoutCooperativeKernels = false;
 
-    cachedMOCSAllowed = true;
+    bool cachedMOCSAllowed = true;
 
     for (auto i = 0u; i < numCommandLists; i++) {
         auto commandList = CommandList::fromHandle(phCommandLists[i]);
@@ -100,7 +100,7 @@ ze_result_t CommandQueueHw<gfxCoreFamily>::executeCommandLists(
             anyCommandListWithoutCooperativeKernels = true;
         }
         // If the Command List has commands that require uncached MOCS, then any changes to the commands in the queue requires the uncached MOCS
-        if (commandList->requiresUncachedMOCS && cachedMOCSAllowed == true) {
+        if (commandList->requiresQueueUncachedMocs && cachedMOCSAllowed == true) {
             cachedMOCSAllowed = false;
         }
     }
@@ -314,7 +314,7 @@ ze_result_t CommandQueueHw<gfxCoreFamily>::executeCommandLists(
 
         if (gsbaStateDirty) {
             auto indirectHeap = CommandList::fromHandle(phCommandLists[0])->commandContainer.getIndirectHeap(NEO::HeapType::INDIRECT_OBJECT);
-            programStateBaseAddress(scratchSpaceController->calculateNewGSH(), indirectHeap->getGraphicsAllocation()->isAllocatedInLocalMemoryPool(), child);
+            programStateBaseAddress(scratchSpaceController->calculateNewGSH(), indirectHeap->getGraphicsAllocation()->isAllocatedInLocalMemoryPool(), child, cachedMOCSAllowed);
         }
 
         if (initialPreemptionMode) {
diff --git a/level_zero/core/source/cmdqueue/cmdqueue_hw_base.inl b/level_zero/core/source/cmdqueue/cmdqueue_hw_base.inl
@@ -30,7 +30,7 @@
 namespace L0 {
 
 template <GFXCORE_FAMILY gfxCoreFamily>
-void CommandQueueHw<gfxCoreFamily>::programStateBaseAddress(uint64_t gsba, bool useLocalMemoryForIndirectHeap, NEO::LinearStream &commandStream) {
+void CommandQueueHw<gfxCoreFamily>::programStateBaseAddress(uint64_t gsba, bool useLocalMemoryForIndirectHeap, NEO::LinearStream &commandStream, bool cachedMOCSAllowed) {
     using GfxFamily = typename NEO::GfxFamilyMapper<gfxCoreFamily>::GfxFamily;
     using STATE_BASE_ADDRESS = typename GfxFamily::STATE_BASE_ADDRESS;
 
diff --git a/level_zero/core/source/cmdqueue/cmdqueue_imp.h b/level_zero/core/source/cmdqueue/cmdqueue_imp.h
@@ -81,7 +81,6 @@ struct CommandQueueImp : public CommandQueue {
     ze_command_queue_mode_t getSynchronousMode() const;
     virtual void dispatchTaskCountWrite(NEO::LinearStream &commandStream, bool flushDataCache) = 0;
     virtual bool getPreemptionCmdProgramming() = 0;
-    bool cachedMOCSAllowed = true;
 
   protected:
     MOCKABLE_VIRTUAL int submitBatchBuffer(size_t offset, NEO::ResidencyContainer &residencyContainer, void *endingCmdPtr,
diff --git a/level_zero/core/source/cmdqueue/cmdqueue_xe_hp_core_and_later.inl b/level_zero/core/source/cmdqueue/cmdqueue_xe_hp_core_and_later.inl
@@ -21,7 +21,7 @@
 namespace L0 {
 
 template <GFXCORE_FAMILY gfxCoreFamily>
-void CommandQueueHw<gfxCoreFamily>::programStateBaseAddress(uint64_t gsba, bool useLocalMemoryForIndirectHeap, NEO::LinearStream &commandStream) {
+void CommandQueueHw<gfxCoreFamily>::programStateBaseAddress(uint64_t gsba, bool useLocalMemoryForIndirectHeap, NEO::LinearStream &commandStream, bool cachedMOCSAllowed) {
     using GfxFamily = typename NEO::GfxFamilyMapper<gfxCoreFamily>::GfxFamily;
     using STATE_BASE_ADDRESS = typename GfxFamily::STATE_BASE_ADDRESS;
     if (NEO::ApiSpecificConfig::getBindlessConfiguration()) {
diff --git a/level_zero/core/source/kernel/kernel_hw.h b/level_zero/core/source/kernel/kernel_hw.h
@@ -74,7 +74,7 @@ struct KernelHw : public KernelImp {
         }
 
         if (l3Enabled == false) {
-            this->kernelRequiresUncachedMocsCount++;
+            this->kernelRequiresQueueUncachedMocsCount++;
         }
 
         NEO::Device *neoDevice = module->getDevice()->getNEODevice();
diff --git a/level_zero/core/source/kernel/kernel_imp.h b/level_zero/core/source/kernel/kernel_imp.h
@@ -119,6 +119,7 @@ struct KernelImp : Kernel {
     uint32_t getRequiredWorkgroupOrder() const override { return requiredWorkgroupOrder; }
     bool requiresGenerationOfLocalIdsByRuntime() const override { return kernelRequiresGenerationOfLocalIdsByRuntime; }
     bool getKernelRequiresUncachedMocs() { return (kernelRequiresUncachedMocsCount > 0); }
+    bool getKernelRequiresQueueUncachedMocs() { return (kernelRequiresQueueUncachedMocsCount > 0); }
     void setKernelArgUncached(uint32_t index, bool val) { isArgUncached[index] = val; }
 
     uint32_t *getGlobalOffsets() override {
@@ -202,6 +203,7 @@ struct KernelImp : Kernel {
 
     bool kernelRequiresGenerationOfLocalIdsByRuntime = true;
     uint32_t kernelRequiresUncachedMocsCount = false;
+    uint32_t kernelRequiresQueueUncachedMocsCount = false;
     std::vector<bool> isArgUncached;
 
     uint32_t globalOffsets[3] = {};
diff --git a/level_zero/core/test/unit_tests/sources/cmdqueue/test_cmdqueue.cpp b/level_zero/core/test/unit_tests/sources/cmdqueue/test_cmdqueue.cpp
@@ -444,7 +444,7 @@ HWTEST2_F(CommandQueueProgramSBATest, whenCreatingCommandQueueThenItIsInitialize
             .Times(1); // instruction heap
     }
 
-    commandQueue->programStateBaseAddress(0u, true, child);
+    commandQueue->programStateBaseAddress(0u, true, child, true);
 
     if (isaInLocalMemory) {
         EXPECT_CALL(*memoryManager, getInternalHeapBaseAddress(rootDeviceIndex, false))
@@ -460,7 +460,7 @@ HWTEST2_F(CommandQueueProgramSBATest, whenCreatingCommandQueueThenItIsInitialize
             .Times(2);
     }
 
-    commandQueue->programStateBaseAddress(0u, false, child);
+    commandQueue->programStateBaseAddress(0u, false, child, true);
 
     commandQueue->destroy();
 }
@@ -477,7 +477,7 @@ HWTEST2_F(CommandQueueProgramSBATest,
     uint32_t alignedSize = 4096u;
     NEO::LinearStream child(commandQueue->commandStream->getSpace(alignedSize), alignedSize);
 
-    commandQueue->programStateBaseAddress(0u, true, child);
+    commandQueue->programStateBaseAddress(0u, true, child, true);
     auto pSbaCmd = static_cast<STATE_BASE_ADDRESS *>(commandQueue->commandStream->getSpace(sizeof(STATE_BASE_ADDRESS)));
     uint32_t statelessMocsIndex = pSbaCmd->getStatelessDataPortAccessMemoryObjectControlState();
 
@@ -510,7 +510,7 @@ HWTEST2_F(CommandQueueProgramSBATest,
     uint32_t alignedSize = 4096u;
     NEO::LinearStream child(commandQueue->commandStream->getSpace(alignedSize), alignedSize);
 
-    commandQueue->programStateBaseAddress(0u, true, child);
+    commandQueue->programStateBaseAddress(0u, true, child, true);
 
     auto usedSpaceAfter = commandQueue->commandStream->getUsed();
 
@@ -549,7 +549,7 @@ HWTEST2_F(CommandQueueProgramSBATest,
     uint32_t alignedSize = 4096u;
     NEO::LinearStream child(commandQueue->commandStream->getSpace(alignedSize), alignedSize);
 
-    commandQueue->programStateBaseAddress(0u, true, child);
+    commandQueue->programStateBaseAddress(0u, true, child, true);
 
     auto usedSpaceAfter = commandQueue->commandStream->getUsed();
 
diff --git a/level_zero/core/test/unit_tests/sources/cmdqueue/test_cmdqueue_enqueue_cmdlist.cpp b/level_zero/core/test/unit_tests/sources/cmdqueue/test_cmdqueue_enqueue_cmdlist.cpp
@@ -91,7 +91,7 @@ struct MultiDeviceCommandQueueExecuteCommandLists : public Test<MultiDeviceFixtu
     ze_command_list_handle_t commandLists[numCommandLists];
 };
 
-HWTEST_F(CommandQueueExecuteCommandLists, whenACommandListExecutedRequiresUncachedMOCSThenCachedMOCSAllowedIsFalse) {
+HWTEST_F(CommandQueueExecuteCommandLists, whenACommandListExecutedRequiresUncachedMOCSThenSuccessisReturned) {
     using MI_BATCH_BUFFER_START = typename FamilyType::MI_BATCH_BUFFER_START;
     using MI_BATCH_BUFFER_END = typename FamilyType::MI_BATCH_BUFFER_END;
     using PARSE = typename FamilyType::PARSE;
@@ -109,12 +109,10 @@ HWTEST_F(CommandQueueExecuteCommandLists, whenACommandListExecutedRequiresUncach
 
     auto commandList1 = whitebox_cast(CommandList::fromHandle(commandLists[0]));
     auto commandList2 = whitebox_cast(CommandList::fromHandle(commandLists[1]));
-    commandList1->requiresUncachedMOCS = true;
-    commandList2->requiresUncachedMOCS = true;
+    commandList1->requiresQueueUncachedMocs = true;
+    commandList2->requiresQueueUncachedMocs = true;
     auto result = commandQueue->executeCommandLists(numCommandLists, commandLists, nullptr, true);
     ASSERT_EQ(ZE_RESULT_SUCCESS, result);
-    ASSERT_EQ(commandQueue->cachedMOCSAllowed, false);
-
     commandQueue->destroy();
 }
 
diff --git a/level_zero/core/test/unit_tests/sources/kernel/test_kernel.cpp b/level_zero/core/test/unit_tests/sources/kernel/test_kernel.cpp
@@ -1529,7 +1529,7 @@ HWTEST2_F(KernelImpL3CachingTests, GivenKernelImpWhenSetSurfaceStateWithUnaligne
 
     memset(expectedSsInHeap.ssPtr, 0, size);
     mockKernel.setBufferSurfaceState(0, buffer, &mockAllocation);
-    EXPECT_EQ(mockKernel.getKernelRequiresUncachedMocs(), true);
+    EXPECT_EQ(mockKernel.getKernelRequiresQueueUncachedMocs(), true);
 }
 
 struct MyMockKernel : public Mock<Kernel> {

Original file line number	Diff line number	Diff line change
`@@ -82,7 +82,7 @@ ze_result_t CommandQueueHw<gfxCoreFamily>::executeCommandLists(`
`82`	`82`	`auto anyCommandListWithCooperativeKernels = false;`
`83`	`83`	`auto anyCommandListWithoutCooperativeKernels = false;`
`84`	`84`
`85`		`- cachedMOCSAllowed = true;`
	`85`	`+ bool cachedMOCSAllowed = true;`
`86`	`86`
`87`	`87`	`for (auto i = 0u; i < numCommandLists; i++) {`
`88`	`88`	`auto commandList = CommandList::fromHandle(phCommandLists[i]);`
`@@ -100,7 +100,7 @@ ze_result_t CommandQueueHw<gfxCoreFamily>::executeCommandLists(`
`100`	`100`	`anyCommandListWithoutCooperativeKernels = true;`
`101`	`101`	`}`
`102`	`102`	`// If the Command List has commands that require uncached MOCS, then any changes to the commands in the queue requires the uncached MOCS`
`103`		`- if (commandList->requiresUncachedMOCS && cachedMOCSAllowed == true) {`
	`103`	`+ if (commandList->requiresQueueUncachedMocs && cachedMOCSAllowed == true) {`
`104`	`104`	`cachedMOCSAllowed = false;`
`105`	`105`	`}`
`106`	`106`	`}`
`@@ -314,7 +314,7 @@ ze_result_t CommandQueueHw<gfxCoreFamily>::executeCommandLists(`
`314`	`314`
`315`	`315`	`if (gsbaStateDirty) {`
`316`	`316`	`auto indirectHeap = CommandList::fromHandle(phCommandLists[0])->commandContainer.getIndirectHeap(NEO::HeapType::INDIRECT_OBJECT);`
`317`		`- programStateBaseAddress(scratchSpaceController->calculateNewGSH(), indirectHeap->getGraphicsAllocation()->isAllocatedInLocalMemoryPool(), child);`
	`317`	`+ programStateBaseAddress(scratchSpaceController->calculateNewGSH(), indirectHeap->getGraphicsAllocation()->isAllocatedInLocalMemoryPool(), child, cachedMOCSAllowed);`
`318`	`318`	`}`
`319`	`319`
`320`	`320`	`if (initialPreemptionMode) {`
Original file line number	Diff line number	Diff line change
`@@ -74,7 +74,7 @@ struct KernelHw : public KernelImp {`
`74`	`74`	`}`
`75`	`75`
`76`	`76`	`if (l3Enabled == false) {`
`77`		`- this->kernelRequiresUncachedMocsCount++;`
	`77`	`+ this->kernelRequiresQueueUncachedMocsCount++;`
`78`	`78`	`}`
`79`	`79`
`80`	`80`	`NEO::Device *neoDevice = module->getDevice()->getNEODevice();`