Add resolves for stateless compression with shared buffer

Related-To: NEO-5107

Signed-off-by: Milczarek, Slawomir <[email protected]>

Author: smilczar

opencl/source/kernel/kernel.cpp

@@ -2530,6 +2530,19 @@ void Kernel::fillWithKernelObjsForAuxTranslation(KernelObjsForAuxTranslation &ke
     }
 }
 
+bool Kernel::hasDirectStatelessAccessToSharedBuffer() const {
+    for (uint32_t i = 0; i < getKernelArgsNumber(); i++) {
+        const auto &arg = kernelInfo.kernelDescriptor.payloadMappings.explicitArgs[i];
+        if (BUFFER_OBJ == kernelArguments.at(i).type && !arg.as<ArgDescPointer>().isPureStateful()) {
+            auto buffer = castToObject<Buffer>(getKernelArg(i));
+            if (buffer && buffer->getMultiGraphicsAllocation().getAllocationType() == GraphicsAllocation::AllocationType::SHARED_BUFFER) {
+                return true;
+            }
+        }
+    }
+    return false;
+}
+
 bool Kernel::hasDirectStatelessAccessToHostMemory() const {
     for (uint32_t i = 0; i < getKernelArgsNumber(); i++) {
         const auto &arg = kernelInfo.kernelDescriptor.payloadMappings.explicitArgs[i];
@@ -2848,7 +2861,9 @@ bool Kernel::requiresLimitedWorkgroupSize() const {
 void Kernel::updateAuxTranslationRequired() {
     const auto &hwInfoConfig = *HwInfoConfig::get(getDevice().getHardwareInfo().platform.eProductFamily);
     if (hwInfoConfig.allowStatelessCompression(getDevice().getHardwareInfo())) {
-        if (hasDirectStatelessAccessToHostMemory() || hasIndirectStatelessAccessToHostMemory()) {
+        if (hasDirectStatelessAccessToHostMemory() ||
+            hasIndirectStatelessAccessToHostMemory() ||
+            hasDirectStatelessAccessToSharedBuffer()) {
             setAuxTranslationRequired(true);
         }
     }

opencl/source/kernel/kernel.h

@@ -514,6 +514,7 @@ class Kernel : public ReferenceTrackedObject<Kernel> {
     void resolveArgs();
 
     void reconfigureKernel();
+    bool hasDirectStatelessAccessToSharedBuffer() const;
     bool hasDirectStatelessAccessToHostMemory() const;
     bool hasIndirectStatelessAccessToHostMemory() const;
 

opencl/test/unit_test/kernel/kernel_arg_buffer_tests.cpp

@@ -333,6 +333,23 @@ TEST_F(KernelArgBufferTest, givenBufferWhenHasDirectStatelessAccessToHostMemoryI
     }
 }
 
+TEST_F(KernelArgBufferTest, givenSharedBufferWhenHasDirectStatelessAccessToSharedBufferIsCalledThenReturnCorrectValue) {
+    MockBuffer buffer;
+    buffer.getGraphicsAllocation(mockRootDeviceIndex)->setAllocationType(GraphicsAllocation::AllocationType::SHARED_BUFFER);
+
+    auto val = (cl_mem)&buffer;
+    auto pVal = &val;
+
+    for (auto pureStatefulBufferAccess : {false, true}) {
+        pKernelInfo->setBufferStateful(0, pureStatefulBufferAccess);
+
+        auto retVal = pKernel->setArg(0, sizeof(cl_mem *), pVal);
+        EXPECT_EQ(CL_SUCCESS, retVal);
+
+        EXPECT_EQ(!pureStatefulBufferAccess, pKernel->hasDirectStatelessAccessToSharedBuffer());
+    }
+}
+
 TEST_F(KernelArgBufferTest, givenBufferInHostMemoryWhenHasDirectStatelessAccessToHostMemoryIsCalledThenReturnCorrectValue) {
     MockBuffer buffer;
     buffer.getGraphicsAllocation(mockRootDeviceIndex)->setAllocationType(GraphicsAllocation::AllocationType::BUFFER_HOST_MEMORY);
@@ -462,6 +479,28 @@ TEST_F(KernelArgBufferTest, whenSettingAuxTranslationRequiredThenIsAuxTranslatio
     }
 }
 
+TEST_F(KernelArgBufferTest, givenSetArgBufferOnKernelWithDirectStatelessAccessToSharedBufferWhenUpdateAuxTranslationRequiredIsCalledThenIsAuxTranslationRequiredShouldReturnTrue) {
+    DebugManagerStateRestore debugRestorer;
+    DebugManager.flags.EnableStatelessCompression.set(1);
+
+    MockBuffer buffer;
+    buffer.getGraphicsAllocation(mockRootDeviceIndex)->setAllocationType(GraphicsAllocation::AllocationType::SHARED_BUFFER);
+
+    auto val = (cl_mem)&buffer;
+    auto pVal = &val;
+
+    auto retVal = pKernel->setArg(0, sizeof(cl_mem *), pVal);
+    EXPECT_EQ(CL_SUCCESS, retVal);
+
+    EXPECT_TRUE(pKernel->hasDirectStatelessAccessToSharedBuffer());
+
+    EXPECT_FALSE(pKernel->isAuxTranslationRequired());
+
+    pKernel->updateAuxTranslationRequired();
+
+    EXPECT_TRUE(pKernel->isAuxTranslationRequired());
+}
+
 TEST_F(KernelArgBufferTest, givenSetArgBufferOnKernelWithDirectStatelessAccessToHostMemoryWhenUpdateAuxTranslationRequiredIsCalledThenIsAuxTranslationRequiredShouldReturnTrue) {
     DebugManagerStateRestore debugRestorer;
     DebugManager.flags.EnableStatelessCompression.set(1);

opencl/test/unit_test/mocks/mock_kernel.h

@@ -107,6 +107,7 @@ class MockKernel : public Kernel {
     using Kernel::getDevice;
     using Kernel::getHardwareInfo;
     using Kernel::hasDirectStatelessAccessToHostMemory;
+    using Kernel::hasDirectStatelessAccessToSharedBuffer;
     using Kernel::hasIndirectStatelessAccessToHostMemory;
     using Kernel::isSchedulerKernel;
     using Kernel::kernelArgHandlers;