fix: Align IOH entry

Related-To: NEO-10036

Signed-off-by: Lukasz Jobczyk <lukasz.jobczyk@intel.com>

level_zero/core/test/unit_tests/sources/cmdlist/test_cmdlist_append_launch_kernel_2.cpp

@@ -1000,7 +1000,7 @@ struct CmdlistAppendLaunchKernelWithImplicitArgsTests : CmdlistAppendLaunchKerne
         implicitArgsProgrammingSize = ImplicitArgsHelper::getSizeForImplicitArgsPatching(&expectedImplicitArgs, *kernelDescriptor, !kernelRequiresGenerationOfLocalIdsByRuntime, gfxCoreHelper);
         auto sizeCrossThreadData = kernel->getCrossThreadDataSize();
         auto sizePerThreadDataForWholeGroup = kernel->getPerThreadDataSizeForWholeThreadGroup();
-        EXPECT_EQ(indirectHeap->getUsed(), sizeCrossThreadData + sizePerThreadDataForWholeGroup + implicitArgsProgrammingSize);
+        EXPECT_EQ(indirectHeap->getUsed(), alignUp(sizeCrossThreadData + sizePerThreadDataForWholeGroup + implicitArgsProgrammingSize, gfxCoreHelper.getIOHAlignment()));
 
         if (FamilyType::supportsCmdSet(IGFX_XE_HP_CORE)) {
             expectedImplicitArgs.localIdTablePtr = indirectHeapAllocation->getGpuAddress();

opencl/source/helpers/hardware_commands_helper_base.inl

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2019-2023 Intel Corporation
+ * Copyright (C) 2019-2024 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -327,6 +327,8 @@ size_t HardwareCommandsHelper<GfxFamily>::sendIndirectState(
         auto indirectDataLength = alignUp(static_cast<uint32_t>(sizeCrossThreadData + sizePerThreadDataTotal),
                                           WalkerType::INDIRECTDATASTARTADDRESS_ALIGN_SIZE);
         walkerCmd->setIndirectDataLength(indirectDataLength);
+
+        ioh.align(kernel.getGfxCoreHelper().getIOHAlignment());
     }
 
     return offsetCrossThreadData;

shared/source/command_container/command_encoder_xehp_and_later.inl

@@ -343,6 +343,8 @@ void EncodeDispatchKernel<Family>::encode(CommandContainer &container, EncodeDis
     } else {
         walkerCmd.setIndirectDataStartAddress(static_cast<uint32_t>(offsetThreadData));
         walkerCmd.setIndirectDataLength(sizeThreadData);
+
+        container.getIndirectHeap(HeapType::indirectObject)->align(rootDeviceEnvironment.getHelper<GfxCoreHelper>().getIOHAlignment());
     }
 
     EncodeDispatchKernel<Family>::encodeThreadData(walkerCmd,

shared/source/debug_settings/debug_variables_base.inl

@@ -574,6 +574,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, ForceSemaphoreDelayBetweenWaits, -1, "Specifies
 DECLARE_DEBUG_VARIABLE(int32_t, ForceLocalMemoryAccessMode, -1, "-1: don't override, 0: default rules apply, 1: CPU can access local memory, 3: CPU never accesses local memory")
 DECLARE_DEBUG_VARIABLE(int32_t, ForceUserptrAlignment, -1, "-1: no force (4kb), >0: n kb alignment")
 DECLARE_DEBUG_VARIABLE(int32_t, ForceCommandBufferAlignment, -1, "-1: no force (64kb), >0: n kb alignment")
+DECLARE_DEBUG_VARIABLE(int32_t, ForceIOHAlignment, -1, "-1: no force, >0: n byte alignment")
 DECLARE_DEBUG_VARIABLE(int32_t, ForceDefaultHeapSize, -1, "-1: no force (64kb), >0: n kb size")
 DECLARE_DEBUG_VARIABLE(int32_t, PreferCopyEngineForCopyBufferToBuffer, -1, "-1: default, 0: prefer EUs, 1: prefer blitter")
 DECLARE_DEBUG_VARIABLE(int64_t, ForceSystemMemoryPlacement, 0, "0: default,  >0: (bitmask) for given Graphics Allocation Type, force system memory placement")

shared/source/helpers/gfx_core_helper.h

@@ -159,6 +159,7 @@ class GfxCoreHelper {
     virtual size_t getMax3dImageWidthOrHeight() const = 0;
     virtual uint64_t getMaxMemAllocSize() const = 0;
     virtual uint64_t getPatIndex(CacheRegion cacheRegion, CachePolicy cachePolicy) const = 0;
+    virtual size_t getIOHAlignment() const = 0;
     virtual bool isStatelessToStatefulWithOffsetSupported() const = 0;
     virtual void encodeBufferSurfaceState(EncodeSurfaceStateArgs &args) const = 0;
     virtual bool platformSupportsImplicitScaling(const NEO::RootDeviceEnvironment &rootDeviceEnvironment) const = 0;
@@ -383,6 +384,7 @@ class GfxCoreHelperHw : public GfxCoreHelper {
     size_t getMax3dImageWidthOrHeight() const override;
     uint64_t getMaxMemAllocSize() const override;
     uint64_t getPatIndex(CacheRegion cacheRegion, CachePolicy cachePolicy) const override;
+    size_t getIOHAlignment() const override;
     bool isStatelessToStatefulWithOffsetSupported() const override;
     void encodeBufferSurfaceState(EncodeSurfaceStateArgs &args) const override;
     bool platformSupportsImplicitScaling(const NEO::RootDeviceEnvironment &rootDeviceEnvironment) const override;

shared/source/helpers/gfx_core_helper_base.inl

@@ -668,6 +668,11 @@ uint64_t GfxCoreHelperHw<GfxFamily>::getPatIndex(CacheRegion cacheRegion, CacheP
     return -1;
 }
 
+template <typename GfxFamily>
+size_t GfxCoreHelperHw<GfxFamily>::getIOHAlignment() const {
+    return 1;
+}
+
 template <typename GfxFamily>
 bool GfxCoreHelperHw<GfxFamily>::copyThroughLockedPtrEnabled(const HardwareInfo &hwInfo, const ProductHelper &productHelper) const {
     if (debugManager.flags.ExperimentalCopyThroughLock.get() != -1) {

shared/test/common/test_files/igdrcl.config

@@ -279,6 +279,7 @@ OverrideSlmSize = -1
 UseCyclesPerSecondTimer = 0
 PrintOsContextInitializations = 0
 WaitLoopCount = -1
+ForceIOHAlignment = -1
 DebuggerLogBitmask = 0
 GTPinAllocateBufferInSharedMemory = -1
 DeferOsContextInitialization = -1

shared/test/unit_test/encoders/test_encode_dispatch_kernel_xehp_and_later.cpp

@@ -592,9 +592,10 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, CommandEncodeStatesTest, givenInlineDataRequiredWhe
     EXPECT_EQ(1u, cmd->getEmitInlineParameter());
 
     const uint32_t inlineDataSize = sizeof(InlineData);
-    size_t expectedSizeIOH = dispatchInterface->getCrossThreadDataSize() +
+    size_t expectedSizeIOH = alignUp(dispatchInterface->getCrossThreadDataSize() +
-                             dispatchInterface->getPerThreadDataSizeForWholeThreadGroup() -
+                                         dispatchInterface->getPerThreadDataSizeForWholeThreadGroup() -
-                             inlineDataSize;
+                                         inlineDataSize,
+                                     this->getHelper<GfxCoreHelper>().getIOHAlignment());
     auto heap = cmdContainer->getIndirectHeap(HeapType::indirectObject);
     EXPECT_EQ(expectedSizeIOH, heap->getUsed());
 }
@@ -619,8 +620,9 @@ HWCMDTEST_F(IGFX_XE_HP_CORE, CommandEncodeStatesTest, givenInlineDataRequiredIsF
     auto cmd = genCmdCast<DefaultWalkerType *>(*itor);
     EXPECT_EQ(0u, cmd->getEmitInlineParameter());
 
-    size_t expectedSizeIOH = dispatchInterface->getCrossThreadDataSize() +
+    size_t expectedSizeIOH = alignUp(dispatchInterface->getCrossThreadDataSize() +
-                             dispatchInterface->getPerThreadDataSizeForWholeThreadGroup();
+                                         dispatchInterface->getPerThreadDataSizeForWholeThreadGroup(),
+                                     this->getHelper<GfxCoreHelper>().getIOHAlignment());
     auto heap = cmdContainer->getIndirectHeap(HeapType::indirectObject);
     EXPECT_EQ(expectedSizeIOH, heap->getUsed());
 }