Align sync buffer address to max atomic type size

Signed-off-by: Filip Hazubski <filip.hazubski@intel.com>

Author: fhazubski-Intel

opencl/test/unit_test/command_queue/sync_buffer_handler_tests.cpp

@@ -104,10 +104,9 @@ class SyncBufferHandlerTest : public SyncBufferEnqueueHandlerTest {
 
     const cl_uint workDim = 1;
     const size_t gwOffset[3] = {0, 0, 0};
-    const size_t lws[3] = {10, 1, 1};
-    size_t workgroupCount[3] = {10, 1, 1};
-    size_t globalWorkSize[3] = {100, 1, 1};
-    size_t workItemsCount = 10;
+    const size_t workItemsCount = 16;
+    const size_t lws[3] = {workItemsCount, 1, 1};
+    size_t workgroupCount[3] = {workItemsCount, 1, 1};
     std::unique_ptr<MockKernelWithInternals> kernelInternals;
     MockKernel *kernel;
     MockCommandQueue *commandQueue;
@@ -128,6 +127,19 @@ HWTEST_TEMPLATED_F(SyncBufferHandlerTest, GivenAllocateSyncBufferPatchAndConcurr
               static_cast<UltCommandStreamReceiver<FamilyType> *>(pCsr)->latestSentTaskCount);
 }
 
+HWTEST_TEMPLATED_F(SyncBufferHandlerTest, GivenAllocateSyncBufferPatchAndConcurrentKernelWhenEnqueuingKernelThenSyncBufferOffsetIsProperlyAligned) {
+    patchAllocateSyncBuffer();
+
+    workgroupCount[0] = 1;
+    enqueueNDCount();
+
+    auto syncBufferHandler = getSyncBufferHandler();
+    EXPECT_EQ(CommonConstants::maximalSizeOfAtomicType, syncBufferHandler->usedBufferSize);
+
+    enqueueNDCount();
+    EXPECT_EQ(2u * CommonConstants::maximalSizeOfAtomicType, syncBufferHandler->usedBufferSize);
+}
+
 HWTEST_TEMPLATED_F(SyncBufferHandlerTest, GivenConcurrentKernelWithoutAllocateSyncBufferPatchWhenEnqueuingConcurrentKernelThenSyncBufferIsNotCreated) {
     auto retVal = enqueueNDCount();
     EXPECT_EQ(CL_SUCCESS, retVal);
@@ -153,7 +165,6 @@ HWTEST_TEMPLATED_F(SyncBufferHandlerTest, GivenConcurrentKernelWithAllocateSyncB
 HWTEST_TEMPLATED_F(SyncBufferHandlerTest, GivenMaxWorkgroupCountWhenEnqueuingConcurrentKernelThenSuccessIsReturned) {
     auto maxWorkGroupCount = kernel->getMaxWorkGroupCount(workDim, lws, commandQueue);
     workgroupCount[0] = maxWorkGroupCount;
-    globalWorkSize[0] = maxWorkGroupCount * lws[0];
 
     auto retVal = enqueueNDCount();
     EXPECT_EQ(CL_SUCCESS, retVal);
@@ -162,7 +173,6 @@ HWTEST_TEMPLATED_F(SyncBufferHandlerTest, GivenMaxWorkgroupCountWhenEnqueuingCon
 HWTEST_TEMPLATED_F(SyncBufferHandlerTest, GivenTooHighWorkgroupCountWhenEnqueuingConcurrentKernelThenErrorIsReturned) {
     size_t maxWorkGroupCount = kernel->getMaxWorkGroupCount(workDim, lws, commandQueue);
     workgroupCount[0] = maxWorkGroupCount + 1;
-    globalWorkSize[0] = maxWorkGroupCount * lws[0];
 
     auto retVal = enqueueNDCount();
     EXPECT_EQ(CL_INVALID_VALUE, retVal);

shared/source/helpers/constants.h

@@ -88,6 +88,7 @@ namespace CommonConstants {
 constexpr uint32_t unspecifiedDeviceIndex = std::numeric_limits<uint32_t>::max();
 constexpr uint32_t invalidStepping = std::numeric_limits<uint32_t>::max();
 constexpr uint32_t maximalSimdSize = 32;
+constexpr uint32_t maximalSizeOfAtomicType = 8;
 constexpr uint32_t engineGroupCount = static_cast<uint32_t>(NEO::EngineGroupType::MaxEngineGroups);
 constexpr uint32_t partitionAddressOffsetDwords = 2u;
 constexpr uint32_t partitionAddressOffset = sizeof(uint32_t) * partitionAddressOffsetDwords;

shared/source/program/sync_buffer_handler.inl

@@ -9,7 +9,7 @@
 
 template <typename KernelT>
 void NEO::SyncBufferHandler::prepareForEnqueue(size_t workGroupsCount, KernelT &kernel) {
-    auto requiredSize = workGroupsCount;
+    auto requiredSize = alignUp(workGroupsCount, CommonConstants::maximalSizeOfAtomicType);
     std::lock_guard<std::mutex> guard(this->mutex);
 
     bool isCurrentBufferFull = (usedBufferSize + requiredSize > bufferSize);