performance(ocl): enable usm pool allocator

Enable on xe hpg and lpg platforms

Related-To: NEO-9700

Signed-off-by: Dominik Dabek <dominik.dabek@intel.com>

opencl/source/context/context.cpp

@@ -494,8 +494,8 @@ void Context::initializeUsmAllocationPools() {
     if (!(svmMemoryManager && this->isSingleDeviceContext())) {
         return;
     }
-
+    auto &productHelper = getDevices()[0]->getProductHelper();
-    bool enabled = false;
+    bool enabled = productHelper.isUsmPoolAllocatorSupported();
     size_t poolSize = 2 * MemoryConstants::megaByte;
     if (debugManager.flags.EnableDeviceUsmAllocationPool.get() != -1) {
         enabled = debugManager.flags.EnableDeviceUsmAllocationPool.get() > 0;
@@ -511,7 +511,7 @@ void Context::initializeUsmAllocationPools() {
         usmDeviceMemAllocPool.initialize(svmMemoryManager, memoryProperties, poolSize);
     }
 
-    enabled = false;
+    enabled = productHelper.isUsmPoolAllocatorSupported();
     poolSize = 2 * MemoryConstants::megaByte;
     if (debugManager.flags.EnableHostUsmAllocationPool.get() != -1) {
         enabled = debugManager.flags.EnableHostUsmAllocationPool.get() > 0;

opencl/test/unit_test/api/cl_set_kernel_arg_svm_pointer_tests.inl

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2018-2023 Intel Corporation
+ * Copyright (C) 2018-2024 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -221,6 +221,7 @@ TEST_F(clSetKernelArgSVMPointerTests, givenSvmAndValidArgValueWhenSettingSameKer
         void *const ptrSvm = clSVMAlloc(pContext, CL_MEM_READ_WRITE, 256, 4);
         EXPECT_NE(nullptr, ptrSvm);
         auto callCounter = 0u;
+        auto svmData = mockSvmManager->getSVMAlloc(ptrSvm);
         //  first set arg - called
         mockSvmManager->allocationsCounter = 0u;
         auto retVal = clSetKernelArgSVMPointer(
@@ -261,7 +262,7 @@ TEST_F(clSetKernelArgSVMPointerTests, givenSvmAndValidArgValueWhenSettingSameKer
         ++mockSvmManager->allocationsCounter;
 
         // different pointer - called
-        void *const nextPtrSvm = static_cast<char *>(ptrSvm) + 1;
+        void *const nextPtrSvm = ptrOffset(ptrSvm, 1);
         retVal = clSetKernelArgSVMPointer(
             pMockMultiDeviceKernel, // cl_kernel kernel
             0,                      // cl_uint arg_index
@@ -272,18 +273,7 @@ TEST_F(clSetKernelArgSVMPointerTests, givenSvmAndValidArgValueWhenSettingSameKer
         ++mockSvmManager->allocationsCounter;
 
         // different allocId - called
-        pMockKernel->kernelArguments[0].allocId = 2;
+        pMockKernel->kernelArguments[0].allocId = svmData->getAllocId() + 1;
-        retVal = clSetKernelArgSVMPointer(
-            pMockMultiDeviceKernel, // cl_kernel kernel
-            0,                      // cl_uint arg_index
-            nextPtrSvm              // const void *arg_value
-        );
-        EXPECT_EQ(CL_SUCCESS, retVal);
-        EXPECT_EQ(++callCounter, pMockKernel->setArgSvmAllocCalls);
-        ++mockSvmManager->allocationsCounter;
-
-        // allocId = 3 - called
-        pMockKernel->kernelArguments[0].allocId = 3;
         retVal = clSetKernelArgSVMPointer(
             pMockMultiDeviceKernel, // cl_kernel kernel
             0,                      // cl_uint arg_index

opencl/test/unit_test/context/context_negative_tests.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2018-2023 Intel Corporation
+ * Copyright (C) 2018-2024 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -32,6 +32,7 @@ TEST_F(ContextFailureInjection, GivenFailedAllocationInjectionWhenCreatingContex
     debugManager.flags.ExperimentalSmallBufferPoolAllocator.set(0);      // failing to allocate pool buffer is non-critical
     debugManager.flags.SetAmountOfReusableAllocationsPerCmdQueue.set(0); // same for preallocations
     debugManager.flags.EnableDeviceUsmAllocationPool.set(0);             // usm device allocation pooling
+    debugManager.flags.EnableHostUsmAllocationPool.set(0);               // usm host allocation pooling
     auto device = std::make_unique<MockClDevice>(MockDevice::createWithNewExecutionEnvironment<MockDevice>(nullptr));
     cl_device_id deviceID = device.get();
 

opencl/test/unit_test/context/context_usm_memory_pool_tests.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2023 Intel Corporation
+ * Copyright (C) 2023-2024 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -7,6 +7,7 @@
 
 #include "shared/test/common/helpers/debug_manager_state_restore.h"
 #include "shared/test/common/mocks/mock_usm_memory_pool.h"
+#include "shared/test/common/test_macros/hw_test.h"
 
 #include "opencl/source/cl_device/cl_device.h"
 #include "opencl/test/unit_test/mocks/mock_context.h"
@@ -40,7 +41,7 @@ struct ContextUsmPoolFlagValuesTest : public ::testing::Test {
 
 using ContextUsmPoolDefaultFlagsTest = ContextUsmPoolFlagValuesTest<-1, -1>;
 
-TEST_F(ContextUsmPoolDefaultFlagsTest, givenDefaultDebugFlagsWhenCreatingContextThenPoolsAreNotInitialized) {
+HWTEST2_F(ContextUsmPoolDefaultFlagsTest, givenDefaultDebugFlagsWhenCreatingContextThenPoolsAreNotInitialized, IsNotXeHpgCore) {
     EXPECT_FALSE(mockDeviceUsmMemAllocPool->isInitialized());
     EXPECT_EQ(0u, mockDeviceUsmMemAllocPool->poolSize);
     EXPECT_EQ(nullptr, mockDeviceUsmMemAllocPool->pool);
@@ -50,6 +51,18 @@ TEST_F(ContextUsmPoolDefaultFlagsTest, givenDefaultDebugFlagsWhenCreatingContext
     EXPECT_EQ(nullptr, mockHostUsmMemAllocPool->pool);
 }
 
+HWTEST2_F(ContextUsmPoolDefaultFlagsTest, givenDefaultDebugFlagsWhenCreatingContextThenPoolsAreInitialized, IsXeHpgCore) {
+    EXPECT_TRUE(mockDeviceUsmMemAllocPool->isInitialized());
+    EXPECT_EQ(2 * MemoryConstants::megaByte, mockDeviceUsmMemAllocPool->poolSize);
+    EXPECT_NE(nullptr, mockDeviceUsmMemAllocPool->pool);
+    EXPECT_EQ(InternalMemoryType::deviceUnifiedMemory, mockDeviceUsmMemAllocPool->poolMemoryType);
+
+    EXPECT_TRUE(mockHostUsmMemAllocPool->isInitialized());
+    EXPECT_EQ(2 * MemoryConstants::megaByte, mockHostUsmMemAllocPool->poolSize);
+    EXPECT_NE(nullptr, mockHostUsmMemAllocPool->pool);
+    EXPECT_EQ(InternalMemoryType::hostUnifiedMemory, mockHostUsmMemAllocPool->poolMemoryType);
+}
+
 using ContextUsmPoolEnabledFlagsTest = ContextUsmPoolFlagValuesTest<1, 3>;
 TEST_F(ContextUsmPoolEnabledFlagsTest, givenEnabledDebugFlagsWhenCreatingContextThenPoolsAreInitialized) {
     EXPECT_TRUE(mockDeviceUsmMemAllocPool->isInitialized());

opencl/test/unit_test/mem_obj/buffer_pool_alloc_tests.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2022-2023 Intel Corporation
+ * Copyright (C) 2022-2024 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -52,6 +52,7 @@ class AggregatedSmallBuffersTestTemplate : public ::testing::Test {
     void setUpImpl() {
         debugManager.flags.ExperimentalSmallBufferPoolAllocator.set(poolBufferFlag);
         debugManager.flags.EnableDeviceUsmAllocationPool.set(0);
+        debugManager.flags.EnableHostUsmAllocationPool.set(0);
         this->deviceFactory = std::make_unique<UltClDeviceFactory>(2, 0);
         this->device = deviceFactory->rootDevices[rootDeviceIndex];
         this->mockMemoryManager = static_cast<MockMemoryManager *>(device->getMemoryManager());

shared/source/os_interface/product_helper.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2018-2023 Intel Corporation
+ * Copyright (C) 2018-2024 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -161,6 +161,7 @@ class ProductHelper {
     virtual bool isResolveDependenciesByPipeControlsSupported(const HardwareInfo &hwInfo, bool isOOQ, TaskCountType queueTaskCount, const CommandStreamReceiver &queueCsr) const = 0;
     virtual bool isMidThreadPreemptionDisallowedForRayTracingKernels() const = 0;
     virtual bool isBufferPoolAllocatorSupported() const = 0;
+    virtual bool isUsmPoolAllocatorSupported() const = 0;
     virtual bool isTlbFlushRequired() const = 0;
     virtual bool isDummyBlitWaRequired() const = 0;
     virtual bool isDetectIndirectAccessInKernelSupported(const KernelDescriptor &kernelDescriptor, const bool isPrecompiled) const = 0;

shared/source/os_interface/product_helper.inl

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2020-2023 Intel Corporation
+ * Copyright (C) 2020-2024 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -584,6 +584,11 @@ bool ProductHelperHw<gfxProduct>::isBufferPoolAllocatorSupported() const {
     return false;
 }
 
+template <PRODUCT_FAMILY gfxProduct>
+bool ProductHelperHw<gfxProduct>::isUsmPoolAllocatorSupported() const {
+    return false;
+}
+
 template <PRODUCT_FAMILY gfxProduct>
 void ProductHelperHw<gfxProduct>::fillScmPropertiesSupportStructureBase(StateComputeModePropertiesSupport &propertiesSupport) const {
     propertiesSupport.coherencyRequired = getScmPropertyCoherencyRequiredSupport();

shared/source/os_interface/product_helper_hw.h

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2023 Intel Corporation
+ * Copyright (C) 2023-2024 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -112,6 +112,7 @@ class ProductHelperHw : public ProductHelper {
     bool isResolveDependenciesByPipeControlsSupported(const HardwareInfo &hwInfo, bool isOOQ, TaskCountType queueTaskCount, const CommandStreamReceiver &queueCsr) const override;
     bool isMidThreadPreemptionDisallowedForRayTracingKernels() const override;
     bool isBufferPoolAllocatorSupported() const override;
+    bool isUsmPoolAllocatorSupported() const override;
     bool isTlbFlushRequired() const override;
     bool isDummyBlitWaRequired() const override;
     bool isDetectIndirectAccessInKernelSupported(const KernelDescriptor &kernelDescriptor, const bool isPrecompiled) const override;

shared/source/xe_hpg_core/dg2/os_agnostic_product_helper_dg2.inl

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2021-2023 Intel Corporation
+ * Copyright (C) 2021-2024 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -213,6 +213,11 @@ bool ProductHelperHw<gfxProduct>::isBufferPoolAllocatorSupported() const {
     return true;
 }
 
+template <>
+bool ProductHelperHw<gfxProduct>::isUsmPoolAllocatorSupported() const {
+    return true;
+}
+
 template <>
 std::optional<aub_stream::ProductFamily> ProductHelperHw<gfxProduct>::getAubStreamProductFamily() const {
     return aub_stream::ProductFamily::Dg2;

shared/source/xe_hpg_core/xe_lpg/os_agnostic_product_helper_xe_lpg.inl

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2022-2023 Intel Corporation
+ * Copyright (C) 2022-2024 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -73,6 +73,11 @@ bool ProductHelperHw<gfxProduct>::isBufferPoolAllocatorSupported() const {
     return true;
 }
 
+template <>
+bool ProductHelperHw<gfxProduct>::isUsmPoolAllocatorSupported() const {
+    return true;
+}
+
 template <>
 uint64_t ProductHelperHw<gfxProduct>::overridePatIndex(bool isUncachedType, uint64_t patIndex) const {
     if (isUncachedType) {

shared/test/unit_test/os_interface/product_helper_tests.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2022-2023 Intel Corporation
+ * Copyright (C) 2022-2024 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -772,6 +772,14 @@ HWTEST2_F(ProductHelperTest, givenProductHelperWhenCheckingIsBufferPoolAllocator
     EXPECT_TRUE(productHelper->isBufferPoolAllocatorSupported());
 }
 
+HWTEST2_F(ProductHelperTest, givenProductHelperWhenCheckingIsUsmPoolAllocatorSupportedThenCorrectValueIsReturned, IsNotXeHpgCore) {
+    EXPECT_FALSE(productHelper->isUsmPoolAllocatorSupported());
+}
+
+HWTEST2_F(ProductHelperTest, givenProductHelperWhenCheckingIsUsmPoolAllocatorSupportedThenCorrectValueIsReturned, IsXeHpgCore) {
+    EXPECT_TRUE(productHelper->isUsmPoolAllocatorSupported());
+}
+
 HWTEST_F(ProductHelperTest, givenProductHelperWhenCheckingIsUnlockingLockedPtrNecessaryThenReturnFalse) {
     EXPECT_FALSE(productHelper->isUnlockingLockedPtrNecessary(pInHwInfo));
 }