Select correct heap base address for ISA

Change-Id: I400f965faa4615519729756daa78350a46c46ff2
Signed-off-by: Bartosz Dunajski <bartosz.dunajski@intel.com>

level_zero/core/source/cmdqueue/cmdqueue_hw_base.inl

@@ -43,6 +43,9 @@ void CommandQueueHw<gfxCoreFamily>::programGeneralStateBaseAddress(uint64_t gsba
     *pcCmd = cmd;
 
     NEO::Device *neoDevice = device->getNEODevice();
+    auto &hwInfo = neoDevice->getHardwareInfo();
+    auto &hwHelper = NEO::HwHelper::get(hwInfo.platform.eRenderCoreFamily);
+
     NEO::EncodeWA<GfxFamily>::encodeAdditionalPipelineSelect(*neoDevice, commandStream, true);
 
     auto pSbaCmd = static_cast<STATE_BASE_ADDRESS *>(commandStream.getSpace(sizeof(STATE_BASE_ADDRESS)));
@@ -56,6 +59,7 @@ void CommandQueueHw<gfxCoreFamily>::programGeneralStateBaseAddress(uint64_t gsba
                                                                     true,
                                                                     (device->getMOCS(true, false) >> 1),
                                                                     neoDevice->getMemoryManager()->getInternalHeapBaseAddress(device->getRootDeviceIndex(), useLocalMemoryForIndirectHeap),
+                                                                    neoDevice->getMemoryManager()->getInternalHeapBaseAddress(device->getRootDeviceIndex(), !hwHelper.useSystemMemoryPlacementForISA(hwInfo)),
                                                                     true,
                                                                     neoDevice->getGmmHelper(),
                                                                     false);

level_zero/core/test/unit_tests/sources/cmdqueue/test_cmdqueue.cpp

@@ -108,13 +108,38 @@ HWTEST2_F(CommandQueueProgramSBATest, whenCreatingCommandQueueThenItIsInitialize
     uint32_t alignedSize = 4096u;
     NEO::LinearStream child(commandQueue->commandStream->getSpace(alignedSize), alignedSize);
 
-    EXPECT_CALL(*memoryManager, getInternalHeapBaseAddress(rootDeviceIndex, true))
+    auto &hwHelper = HwHelper::get(neoDevice->getHardwareInfo().platform.eRenderCoreFamily);
-        .Times(1);
+    bool isaInLocalMemory = !hwHelper.useSystemMemoryPlacementForISA(neoDevice->getHardwareInfo());
+
+    if (isaInLocalMemory) {
+        EXPECT_CALL(*memoryManager, getInternalHeapBaseAddress(rootDeviceIndex, true))
+            .Times(2);
+
+        EXPECT_CALL(*memoryManager, getInternalHeapBaseAddress(rootDeviceIndex, false))
+            .Times(0);
+    } else {
+        EXPECT_CALL(*memoryManager, getInternalHeapBaseAddress(rootDeviceIndex, true))
+            .Times(1); // IOH
+
+        EXPECT_CALL(*memoryManager, getInternalHeapBaseAddress(rootDeviceIndex, false))
+            .Times(1); // instruction heap
+    }
 
     commandQueue->programGeneralStateBaseAddress(0u, true, child);
 
-    EXPECT_CALL(*memoryManager, getInternalHeapBaseAddress(rootDeviceIndex, false))
+    if (isaInLocalMemory) {
-        .Times(1);
+        EXPECT_CALL(*memoryManager, getInternalHeapBaseAddress(rootDeviceIndex, false))
+            .Times(1); // IOH
+
+        EXPECT_CALL(*memoryManager, getInternalHeapBaseAddress(rootDeviceIndex, true))
+            .Times(1); // instruction heap
+    } else {
+        EXPECT_CALL(*memoryManager, getInternalHeapBaseAddress(rootDeviceIndex, true))
+            .Times(0);
+
+        EXPECT_CALL(*memoryManager, getInternalHeapBaseAddress(rootDeviceIndex, false))
+            .Times(2);
+    }
 
     commandQueue->programGeneralStateBaseAddress(0u, false, child);
 

opencl/test/unit_test/command_queue/enqueue_copy_buffer_rect_tests.cpp

@@ -223,7 +223,11 @@ HWTEST_F(EnqueueCopyBufferRectTest, WhenCopyingBufferRect2DThenL3ProgrammingIsCo
 HWCMDTEST_F(IGFX_GEN8_CORE, EnqueueCopyBufferRectTest, When2DEnqueueIsDoneThenStateBaseAddressIsProperlyProgrammed) {
     enqueueCopyBufferRect2D<FamilyType>();
     auto &ultCsr = this->pDevice->getUltCommandStreamReceiver<FamilyType>();
+
+    auto &hwHelper = HwHelper::get(pDevice->getHardwareInfo().platform.eRenderCoreFamily);
+
     validateStateBaseAddress<FamilyType>(ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, pIOH->getGraphicsAllocation()->isAllocatedInLocalMemoryPool()),
+                                         ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, !hwHelper.useSystemMemoryPlacementForISA(pDevice->getHardwareInfo())),
                                          pDSH, pIOH, pSSH, itorPipelineSelect, itorWalker, cmdList, 0llu);
 }
 
@@ -343,7 +347,11 @@ HWTEST_F(EnqueueCopyBufferRectTest, WhenCopyingBufferRect3DThenL3ProgrammingIsCo
 HWCMDTEST_F(IGFX_GEN8_CORE, EnqueueCopyBufferRectTest, When3DEnqueueIsDoneThenStateBaseAddressIsProperlyProgrammed) {
     enqueueCopyBufferRect3D<FamilyType>();
     auto &ultCsr = this->pDevice->getUltCommandStreamReceiver<FamilyType>();
+
+    auto &hwHelper = HwHelper::get(pDevice->getHardwareInfo().platform.eRenderCoreFamily);
+
     validateStateBaseAddress<FamilyType>(ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, pIOH->getGraphicsAllocation()->isAllocatedInLocalMemoryPool()),
+                                         ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, !hwHelper.useSystemMemoryPlacementForISA(pDevice->getHardwareInfo())),
                                          pDSH, pIOH, pSSH, itorPipelineSelect, itorWalker, cmdList, 0llu);
 }
 

opencl/test/unit_test/command_queue/enqueue_copy_buffer_tests.cpp

@@ -192,7 +192,11 @@ HWTEST_F(EnqueueCopyBufferTest, WhenCopyingBufferThenL3ProgrammingIsCorrect) {
 HWCMDTEST_F(IGFX_GEN8_CORE, EnqueueCopyBufferTest, WhenEnqueueIsDoneThenStateBaseAddressIsProperlyProgrammed) {
     enqueueCopyBufferAndParse<FamilyType>();
     auto &ultCsr = this->pDevice->getUltCommandStreamReceiver<FamilyType>();
+
+    auto &hwHelper = HwHelper::get(pDevice->getHardwareInfo().platform.eRenderCoreFamily);
+
     validateStateBaseAddress<FamilyType>(ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, pIOH->getGraphicsAllocation()->isAllocatedInLocalMemoryPool()),
+                                         ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, !hwHelper.useSystemMemoryPlacementForISA(pDevice->getHardwareInfo())),
                                          pDSH, pIOH, pSSH, itorPipelineSelect, itorWalker, cmdList, 0llu);
 }
 

opencl/test/unit_test/command_queue/enqueue_copy_buffer_to_image_tests.cpp

@@ -96,7 +96,11 @@ HWTEST_F(EnqueueCopyBufferToImageTest, WhenCopyingBufferToImageThenL3Programming
 HWCMDTEST_F(IGFX_GEN8_CORE, EnqueueCopyBufferToImageTest, WhenEnqueueIsDoneThenStateBaseAddressIsProperlyProgrammed) {
     enqueueCopyBufferToImage<FamilyType>();
     auto &ultCsr = this->pDevice->getUltCommandStreamReceiver<FamilyType>();
+
+    auto &hwHelper = HwHelper::get(pDevice->getHardwareInfo().platform.eRenderCoreFamily);
+
     validateStateBaseAddress<FamilyType>(ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, pIOH->getGraphicsAllocation()->isAllocatedInLocalMemoryPool()),
+                                         ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, !hwHelper.useSystemMemoryPlacementForISA(pDevice->getHardwareInfo())),
                                          pDSH, pIOH, pSSH, itorPipelineSelect, itorWalker, cmdList, 0llu);
 }
 

opencl/test/unit_test/command_queue/enqueue_copy_image_tests.cpp

@@ -98,7 +98,11 @@ HWTEST_F(EnqueueCopyImageTest, WhenCopyingImageThenL3ProgrammingIsCorrect) {
 HWCMDTEST_F(IGFX_GEN8_CORE, EnqueueCopyImageTest, WhenEnqueueIsDoneThenStateBaseAddressIsProperlyProgrammed) {
     enqueueCopyImage<FamilyType>();
     auto &ultCsr = this->pDevice->getUltCommandStreamReceiver<FamilyType>();
+
+    auto &hwHelper = HwHelper::get(pDevice->getHardwareInfo().platform.eRenderCoreFamily);
+
     validateStateBaseAddress<FamilyType>(ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, pIOH->getGraphicsAllocation()->isAllocatedInLocalMemoryPool()),
+                                         ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, !hwHelper.useSystemMemoryPlacementForISA(pDevice->getHardwareInfo())),
                                          pDSH, pIOH, pSSH, itorPipelineSelect, itorWalker, cmdList, 0llu);
 }
 

opencl/test/unit_test/command_queue/enqueue_copy_image_to_buffer_tests.cpp

@@ -97,7 +97,11 @@ HWTEST_F(EnqueueCopyImageToBufferTest, WhenCopyingImageToBufferThenL3Programming
 HWCMDTEST_F(IGFX_GEN8_CORE, EnqueueCopyImageToBufferTest, WhenEnqueueIsDoneThenStateBaseAddressIsProperlyProgrammed) {
     enqueueCopyImageToBuffer<FamilyType>();
     auto &ultCsr = this->pDevice->getUltCommandStreamReceiver<FamilyType>();
+
+    auto &hwHelper = HwHelper::get(pDevice->getHardwareInfo().platform.eRenderCoreFamily);
+
     validateStateBaseAddress<FamilyType>(ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, pIOH->getGraphicsAllocation()->isAllocatedInLocalMemoryPool()),
+                                         ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, !hwHelper.useSystemMemoryPlacementForISA(pDevice->getHardwareInfo())),
                                          pDSH, pIOH, pSSH, itorPipelineSelect, itorWalker, cmdList, 0llu);
 }
 

opencl/test/unit_test/command_queue/enqueue_fill_buffer_tests.cpp

@@ -213,7 +213,11 @@ HWTEST_F(EnqueueFillBufferCmdTests, WhenFillingBufferThenL3ProgrammingIsCorrect)
 HWCMDTEST_F(IGFX_GEN8_CORE, EnqueueFillBufferCmdTests, WhenEnqueueIsDoneThenStateBaseAddressIsProperlyProgrammed) {
     enqueueFillBuffer<FamilyType>();
     auto &ultCsr = this->pDevice->getUltCommandStreamReceiver<FamilyType>();
+
+    auto &hwHelper = HwHelper::get(pDevice->getHardwareInfo().platform.eRenderCoreFamily);
+
     validateStateBaseAddress<FamilyType>(ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, pIOH->getGraphicsAllocation()->isAllocatedInLocalMemoryPool()),
+                                         ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, !hwHelper.useSystemMemoryPlacementForISA(pDevice->getHardwareInfo())),
                                          pDSH, pIOH, pSSH, itorPipelineSelect, itorWalker, cmdList, 0llu);
 }
 

opencl/test/unit_test/command_queue/enqueue_fill_image_tests.cpp

@@ -105,7 +105,11 @@ HWTEST_F(EnqueueFillImageTest, WhenFillingImageThenL3ProgrammingIsCorrect) {
 HWCMDTEST_F(IGFX_GEN8_CORE, EnqueueFillImageTest, WhenEnqueueIsDoneThenStateBaseAddressIsProperlyProgrammed) {
     enqueueFillImage<FamilyType>();
     auto &ultCsr = this->pDevice->getUltCommandStreamReceiver<FamilyType>();
+
+    auto &hwHelper = HwHelper::get(pDevice->getHardwareInfo().platform.eRenderCoreFamily);
+
     validateStateBaseAddress<FamilyType>(ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, pIOH->getGraphicsAllocation()->isAllocatedInLocalMemoryPool()),
+                                         ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, !hwHelper.useSystemMemoryPlacementForISA(pDevice->getHardwareInfo())),
                                          pDSH, pIOH, pSSH, itorPipelineSelect, itorWalker, cmdList, 0llu);
 }
 

opencl/test/unit_test/command_queue/enqueue_kernel_2_tests.cpp

@@ -186,7 +186,11 @@ HWCMDTEST_P(IGFX_GEN8_CORE, EnqueueWorkItemTestsWithLimitedParamSet, LoadRegiste
 HWCMDTEST_P(IGFX_GEN8_CORE, EnqueueWorkItemTestsWithLimitedParamSet, WhenEnqueueIsDoneThenStateBaseAddressIsProperlyProgrammed) {
     enqueueKernel<FamilyType>();
     auto &ultCsr = this->pDevice->getUltCommandStreamReceiver<FamilyType>();
+
+    auto &hwHelper = HwHelper::get(pDevice->getHardwareInfo().platform.eRenderCoreFamily);
+
     validateStateBaseAddress<FamilyType>(ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, pIOH->getGraphicsAllocation()->isAllocatedInLocalMemoryPool()),
+                                         ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, !hwHelper.useSystemMemoryPlacementForISA(pDevice->getHardwareInfo())),
                                          pDSH, pIOH, pSSH, itorPipelineSelect, itorWalker, cmdList,
                                          context->getMemoryManager()->peekForce32BitAllocations() ? context->getMemoryManager()->getExternalHeapBaseAddress(ultCsr.rootDeviceIndex, false) : 0llu);
 }

opencl/test/unit_test/command_queue/enqueue_read_buffer_rect_tests.cpp

@@ -209,7 +209,11 @@ HWTEST_F(EnqueueReadBufferRectTest, WhenReadingBufferThenL3ProgrammingIsCorrect)
 HWCMDTEST_F(IGFX_GEN8_CORE, EnqueueReadBufferRectTest, When2DEnqueueIsDoneThenStateBaseAddressIsProperlyProgrammed) {
     enqueueReadBufferRect2D<FamilyType>();
     auto &ultCsr = this->pDevice->getUltCommandStreamReceiver<FamilyType>();
+
+    auto &hwHelper = HwHelper::get(pDevice->getHardwareInfo().platform.eRenderCoreFamily);
+
     validateStateBaseAddress<FamilyType>(ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, pIOH->getGraphicsAllocation()->isAllocatedInLocalMemoryPool()),
+                                         ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, !hwHelper.useSystemMemoryPlacementForISA(pDevice->getHardwareInfo())),
                                          pDSH, pIOH, pSSH, itorPipelineSelect, itorWalker, cmdList, 0llu);
 }
 

opencl/test/unit_test/command_queue/enqueue_read_buffer_tests.cpp

@@ -173,7 +173,11 @@ HWCMDTEST_F(IGFX_GEN8_CORE, EnqueueReadBufferTypeTest, WhenEnqueueIsDoneThenStat
     srcBuffer->forceDisallowCPUCopy = true;
     enqueueReadBuffer<FamilyType>();
     auto &ultCsr = this->pDevice->getUltCommandStreamReceiver<FamilyType>();
+
+    auto &hwHelper = HwHelper::get(pDevice->getHardwareInfo().platform.eRenderCoreFamily);
+
     validateStateBaseAddress<FamilyType>(ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, pIOH->getGraphicsAllocation()->isAllocatedInLocalMemoryPool()),
+                                         ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, !hwHelper.useSystemMemoryPlacementForISA(pDevice->getHardwareInfo())),
                                          pDSH, pIOH, pSSH, itorPipelineSelect, itorWalker, cmdList, 0llu);
 }
 

opencl/test/unit_test/command_queue/enqueue_read_image_tests.cpp

@@ -108,7 +108,11 @@ HWTEST_F(EnqueueReadImageTest, WhenReadingImageThenL3ProgrammingIsCorrect) {
 HWCMDTEST_F(IGFX_GEN8_CORE, EnqueueReadImageTest, WhenEnqueueIsDoneThenStateBaseAddressIsProperlyProgrammed) {
     enqueueReadImage<FamilyType>();
     auto &ultCsr = this->pDevice->getUltCommandStreamReceiver<FamilyType>();
+
+    auto &hwHelper = HwHelper::get(pDevice->getHardwareInfo().platform.eRenderCoreFamily);
+
     validateStateBaseAddress<FamilyType>(ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, pIOH->getGraphicsAllocation()->isAllocatedInLocalMemoryPool()),
+                                         ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, !hwHelper.useSystemMemoryPlacementForISA(pDevice->getHardwareInfo())),
                                          pDSH, pIOH, pSSH, itorPipelineSelect, itorWalker, cmdList, 0llu);
 }
 

opencl/test/unit_test/command_queue/enqueue_write_buffer_rect_tests.cpp

@@ -183,7 +183,11 @@ HWTEST_F(EnqueueWriteBufferRectTest, WhenWritingBufferThenL3ProgrammingIsCorrect
 HWCMDTEST_F(IGFX_GEN8_CORE, EnqueueWriteBufferRectTest, When2DEnqueueIsDoneThenStateBaseAddressIsProperlyProgrammed) {
     enqueueWriteBufferRect2D<FamilyType>();
     auto &ultCsr = this->pDevice->getUltCommandStreamReceiver<FamilyType>();
+
+    auto &hwHelper = HwHelper::get(pDevice->getHardwareInfo().platform.eRenderCoreFamily);
+
     validateStateBaseAddress<FamilyType>(ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, pIOH->getGraphicsAllocation()->isAllocatedInLocalMemoryPool()),
+                                         ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, !hwHelper.useSystemMemoryPlacementForISA(pDevice->getHardwareInfo())),
                                          pDSH, pIOH, pSSH, itorPipelineSelect, itorWalker, cmdList, 0llu);
 }
 

opencl/test/unit_test/command_queue/enqueue_write_buffer_tests.cpp

@@ -171,7 +171,11 @@ HWCMDTEST_F(IGFX_GEN8_CORE, EnqueueWriteBufferTypeTest, WhenEnqueueIsDoneThenSta
     srcBuffer->forceDisallowCPUCopy = true;
     enqueueWriteBuffer<FamilyType>();
     auto &ultCsr = this->pDevice->getUltCommandStreamReceiver<FamilyType>();
+
+    auto &hwHelper = HwHelper::get(pDevice->getHardwareInfo().platform.eRenderCoreFamily);
+
     validateStateBaseAddress<FamilyType>(ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, pIOH->getGraphicsAllocation()->isAllocatedInLocalMemoryPool()),
+                                         ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, !hwHelper.useSystemMemoryPlacementForISA(pDevice->getHardwareInfo())),
                                          pDSH, pIOH, pSSH, itorPipelineSelect, itorWalker, cmdList, 0llu);
 }
 

opencl/test/unit_test/command_queue/enqueue_write_image_tests.cpp

@@ -108,7 +108,11 @@ HWTEST_F(EnqueueWriteImageTest, WhenWritingImageThenL3ProgrammingIsCorrect) {
 HWCMDTEST_F(IGFX_GEN8_CORE, EnqueueWriteImageTest, WhenEnqueueIsDoneThenStateBaseAddressIsProperlyProgrammed) {
     enqueueWriteImage<FamilyType>();
     auto &ultCsr = this->pDevice->getUltCommandStreamReceiver<FamilyType>();
+
+    auto &hwHelper = HwHelper::get(pDevice->getHardwareInfo().platform.eRenderCoreFamily);
+
     validateStateBaseAddress<FamilyType>(ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, pIOH->getGraphicsAllocation()->isAllocatedInLocalMemoryPool()),
+                                         ultCsr.getMemoryManager()->getInternalHeapBaseAddress(ultCsr.rootDeviceIndex, !hwHelper.useSystemMemoryPlacementForISA(pDevice->getHardwareInfo())),
                                          pDSH, pIOH, pSSH, itorPipelineSelect, itorWalker, cmdList, 0llu);
 }
 

opencl/test/unit_test/command_stream/command_stream_receiver_flush_task_1_tests.cpp

@@ -517,8 +517,11 @@ HWCMDTEST_F(IGFX_GEN8_CORE, CommandStreamReceiverFlushTaskTests, stateBaseAddres
     ASSERT_NE(nullptr, cmdStateBaseAddress);
     auto &cmd = *reinterpret_cast<STATE_BASE_ADDRESS *>(cmdStateBaseAddress);
 
+    auto &hwHelper = HwHelper::get(pDevice->getHardwareInfo().platform.eRenderCoreFamily);
+    auto instructionHeapBaseAddress = commandStreamReceiver.getMemoryManager()->getInternalHeapBaseAddress(commandStreamReceiver.rootDeviceIndex, !hwHelper.useSystemMemoryPlacementForISA(pDevice->getHardwareInfo()));
+
     EXPECT_EQ(dsh.getCpuBase(), reinterpret_cast<void *>(cmd.getDynamicStateBaseAddress()));
-    EXPECT_EQ(commandStreamReceiver.getMemoryManager()->getInternalHeapBaseAddress(commandStreamReceiver.rootDeviceIndex, ioh.getGraphicsAllocation()->isAllocatedInLocalMemoryPool()), cmd.getInstructionBaseAddress());
+    EXPECT_EQ(instructionHeapBaseAddress, cmd.getInstructionBaseAddress());
     EXPECT_EQ(ioh.getCpuBase(), reinterpret_cast<void *>(cmd.getIndirectObjectBaseAddress()));
     EXPECT_EQ(ssh.getCpuBase(), reinterpret_cast<void *>(cmd.getSurfaceStateBaseAddress()));
 

opencl/test/unit_test/command_stream/command_stream_receiver_flush_task_2_tests.cpp

@@ -1134,7 +1134,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, CommandStreamReceiverFlushTaskTests, givenCsrInNonDi
     EXPECT_EQ(0u, surfacesForResidency.size());
 }
 
-HWTEST_F(CommandStreamReceiverFlushTaskTests, givenCsrWhenGeneralStateBaseAddressIsProgrammedThenDecanonizedAddressIsWritten) {
+HWCMDTEST_F(IGFX_GEN8_CORE, CommandStreamReceiverFlushTaskTests, givenCsrWhenGeneralStateBaseAddressIsProgrammedThenDecanonizedAddressIsWritten) {
     uint64_t generalStateBaseAddress = 0xffff800400010000ull;
 
     DispatchFlags dispatchFlags = DispatchFlagsHelper::createDefaultDispatchFlags();
@@ -1147,6 +1147,7 @@ HWTEST_F(CommandStreamReceiverFlushTaskTests, givenCsrWhenGeneralStateBaseAddres
                                                                 generalStateBaseAddress,
                                                                 true,
                                                                 0,
+                                                                0,
                                                                 generalStateBaseAddress,
                                                                 true,
                                                                 pDevice->getGmmHelper(),
@@ -1167,6 +1168,7 @@ HWTEST_F(CommandStreamReceiverFlushTaskTests, givenNonZeroGeneralStateBaseAddres
                                                                 generalStateBaseAddress,
                                                                 false,
                                                                 0,
+                                                                0,
                                                                 generalStateBaseAddress,
                                                                 true,
                                                                 pDevice->getGmmHelper(),
@@ -1190,6 +1192,7 @@ HWTEST_F(CommandStreamReceiverFlushTaskTests, givenNonZeroInternalHeapBaseAddres
                                                                 true,
                                                                 0,
                                                                 internalHeapBaseAddress,
+                                                                0,
                                                                 false,
                                                                 pDevice->getGmmHelper(),
                                                                 false);
@@ -1205,6 +1208,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, CommandStreamReceiverFlushTaskTests, givenSbaProgram
     DispatchFlags dispatchFlags = DispatchFlagsHelper::createDefaultDispatchFlags();
 
     uint64_t internalHeapBase = 0x10000;
+    uint64_t instructionHeapBase = 0x10000;
     uint64_t generalStateBase = 0x30000;
     typename FamilyType::STATE_BASE_ADDRESS sbaCmd;
 
@@ -1216,6 +1220,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, CommandStreamReceiverFlushTaskTests, givenSbaProgram
                                                                 true,
                                                                 0,
                                                                 internalHeapBase,
+                                                                instructionHeapBase,
                                                                 true,
                                                                 pDevice->getGmmHelper(),
                                                                 false);
@@ -1234,7 +1239,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, CommandStreamReceiverFlushTaskTests, givenSbaProgram
     EXPECT_EQ(0u, sbaCmd.getSurfaceStateBaseAddress());
 
     EXPECT_TRUE(sbaCmd.getInstructionBaseAddressModifyEnable());
-    EXPECT_EQ(internalHeapBase, sbaCmd.getInstructionBaseAddress());
+    EXPECT_EQ(instructionHeapBase, sbaCmd.getInstructionBaseAddress());
     EXPECT_TRUE(sbaCmd.getInstructionBufferSizeModifyEnable());
     EXPECT_EQ(MemoryConstants::sizeOf4GBinPageEntities, sbaCmd.getInstructionBufferSize());
 

opencl/test/unit_test/gen_common/gen_commands_common_validation.h

@@ -19,7 +19,9 @@
 
 namespace NEO {
 template <typename FamilyType>
-void validateStateBaseAddress(uint64_t internalHeapBase, IndirectHeap *pDSH,
+void validateStateBaseAddress(uint64_t indirectObjectHeapBase,
+                              uint64_t instructionHeapBaseAddress,
+                              IndirectHeap *pDSH,
                               IndirectHeap *pIOH,
                               IndirectHeap *pSSH,
                               GenCmdList::iterator &startCommand,
@@ -45,7 +47,7 @@ void validateStateBaseAddress(uint64_t internalHeapBase, IndirectHeap *pDSH,
     EXPECT_EQ(expectedGeneralStateHeapBaseAddress, cmd->getGeneralStateBaseAddress());
     EXPECT_EQ(pSSH->getGraphicsAllocation()->getGpuAddress(), cmd->getSurfaceStateBaseAddress());
     EXPECT_EQ(pIOH->getGraphicsAllocation()->getGpuBaseAddress(), cmd->getIndirectObjectBaseAddress());
-    EXPECT_EQ(internalHeapBase, cmd->getInstructionBaseAddress());
+    EXPECT_EQ(instructionHeapBaseAddress, cmd->getInstructionBaseAddress());
 
     // Verify all sizes are getting programmed
     EXPECT_TRUE(cmd->getDynamicStateBufferSizeModifyEnable());

opencl/test/unit_test/helpers/hw_helper_tests.cpp

@@ -994,10 +994,14 @@ TEST(HwInfoConfigCommonHelperTest, givenBlitterPreferenceWhenEnablingBlitterOper
     EXPECT_EQ(expectedBlitterSupport, hardwareInfo.capabilityTable.blitterOperationsSupported);
 }
 
-HWTEST_F(HwHelperTest, givenHwHelperWhenAskingForIsaSystemMemoryPlacementThenReturnFalse) {
+HWTEST_F(HwHelperTest, givenHwHelperWhenAskingForIsaSystemMemoryPlacementThenReturnFalseIfLocalMemorySupported) {
     HwHelper &hwHelper = HwHelper::get(hardwareInfo.platform.eRenderCoreFamily);
 
+    hardwareInfo.featureTable.ftrLocalMemory = true;
     EXPECT_FALSE(hwHelper.useSystemMemoryPlacementForISA(hardwareInfo));
+
+    hardwareInfo.featureTable.ftrLocalMemory = false;
+    EXPECT_TRUE(hwHelper.useSystemMemoryPlacementForISA(hardwareInfo));
 }
 
 TEST(HwInfoConfigCommonHelperTest, givenDebugFlagSetWhenEnablingBlitterOperationsSupportThenHonorTheFlag) {

opencl/test/unit_test/memory_manager/memory_manager_allocate_in_preferred_pool_tests.inl

@@ -633,7 +633,7 @@ HWTEST_F(GetAllocationDataTestHw, givenKernelIsaTypeWhenGetAllocationDataIsCalle
     MockMemoryManager mockMemoryManager;
     AllocationProperties properties{mockRootDeviceIndex, 1, GraphicsAllocation::AllocationType::KERNEL_ISA, mockDeviceBitfield};
     mockMemoryManager.getAllocationData(allocData, properties, nullptr, mockMemoryManager.createStorageInfoFromProperties(properties));
-    EXPECT_FALSE(allocData.flags.useSystemMemory);
+    EXPECT_NE(defaultHwInfo->featureTable.ftrLocalMemory, allocData.flags.useSystemMemory);
 }
 
 HWTEST_F(GetAllocationDataTestHw, givenLinearStreamWhenGetAllocationDataIsCalledThenSystemMemoryIsNotRequested) {

shared/source/command_container/cmdcontainer.cpp

@@ -87,7 +87,11 @@ ErrorCode CommandContainer::initialize(Device *device) {
         indirectHeaps[i] = std::make_unique<IndirectHeap>(allocationIndirectHeaps[i], requireInternalHeap);
     }
 
-    instructionHeapBaseAddress = device->getMemoryManager()->getInternalHeapBaseAddress(device->getRootDeviceIndex(), allocationIndirectHeaps[IndirectHeap::Type::INDIRECT_OBJECT]->isAllocatedInLocalMemoryPool());
+    auto &hwHelper = HwHelper::get(getDevice()->getHardwareInfo().platform.eRenderCoreFamily);
+
+    indirectObjectHeapBaseAddress = device->getMemoryManager()->getInternalHeapBaseAddress(device->getRootDeviceIndex(), allocationIndirectHeaps[IndirectHeap::Type::INDIRECT_OBJECT]->isAllocatedInLocalMemoryPool());
+
+    instructionHeapBaseAddress = device->getMemoryManager()->getInternalHeapBaseAddress(device->getRootDeviceIndex(), !hwHelper.useSystemMemoryPlacementForISA(getDevice()->getHardwareInfo()));
 
     reserveBindlessOffsets(*indirectHeaps[IndirectHeap::Type::SURFACE_STATE]);
 

shared/source/command_container/cmdcontainer.h

@@ -74,6 +74,8 @@ class CommandContainer : public NonCopyableOrMovableClass {
 
     uint64_t getInstructionHeapBaseAddress() const { return instructionHeapBaseAddress; }
 
+    uint64_t getIndirectObjectHeapBaseAddress() const { return indirectObjectHeapBaseAddress; }
+
     void *getHeapSpaceAllowGrow(HeapType heapType, size_t size);
 
     ErrorCode initialize(Device *device);
@@ -108,6 +110,7 @@ class CommandContainer : public NonCopyableOrMovableClass {
     CmdBufferContainer cmdBufferAllocations;
     GraphicsAllocation *allocationIndirectHeaps[HeapType::NUM_TYPES] = {};
     uint64_t instructionHeapBaseAddress = 0u;
+    uint64_t indirectObjectHeapBaseAddress = 0u;
     uint32_t dirtyHeaps = std::numeric_limits<uint32_t>::max();
     uint32_t numIddsPerBlock = 64;
 

shared/source/command_container/command_encoder_base.inl

@@ -351,6 +351,7 @@ void EncodeStateBaseAddress<Family>::encode(CommandContainer &container, STATE_B
         0,
         false,
         (gmmHelper->getMOCS(GMM_RESOURCE_USAGE_OCL_BUFFER) >> 1),
+        container.getIndirectObjectHeapBaseAddress(),
         container.getInstructionHeapBaseAddress(),
         false,
         gmmHelper,

shared/source/command_stream/command_stream_receiver_hw_base.inl

@@ -342,9 +342,9 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
     auto isStateBaseAddressDirty = dshDirty || iohDirty || sshDirty || stateBaseAddressDirty;
 
     auto mocsIndex = latestSentStatelessMocsConfig;
+    auto &hwHelper = HwHelper::get(peekHwInfo().platform.eRenderCoreFamily);
 
     if (dispatchFlags.l3CacheSettings != L3CachingSettings::NotApplicable) {
-        auto &hwHelper = HwHelper::get(peekHwInfo().platform.eRenderCoreFamily);
         auto l3On = dispatchFlags.l3CacheSettings != L3CachingSettings::l3CacheOff;
         auto l1On = dispatchFlags.l3CacheSettings == L3CachingSettings::l3AndL1On;
         mocsIndex = hwHelper.getMocsIndex(*device.getGmmHelper(), l3On, l1On);
@@ -375,6 +375,7 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
         auto stateBaseAddressCmdOffset = commandStreamCSR.getUsed();
         auto pCmd = static_cast<STATE_BASE_ADDRESS *>(commandStreamCSR.getSpace(sizeof(STATE_BASE_ADDRESS)));
         STATE_BASE_ADDRESS cmd;
+        auto instructionHeapBaseAddress = getMemoryManager()->getInternalHeapBaseAddress(rootDeviceIndex, !hwHelper.useSystemMemoryPlacementForISA(peekHwInfo()));
         StateBaseAddressHelper<GfxFamily>::programStateBaseAddress(
             &cmd,
             &dsh,
@@ -384,6 +385,7 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTask(
             true,
             mocsIndex,
             getMemoryManager()->getInternalHeapBaseAddress(rootDeviceIndex, ioh.getGraphicsAllocation()->isAllocatedInLocalMemoryPool()),
+            instructionHeapBaseAddress,
             true,
             device.getGmmHelper(),
             isMultiOsContextCapable());

shared/source/helpers/hw_helper_base.inl

@@ -494,6 +494,6 @@ bool HwHelperHw<GfxFamily>::useOnlyGlobalTimestamps() const {
 
 template <typename GfxFamily>
 bool HwHelperHw<GfxFamily>::useSystemMemoryPlacementForISA(const HardwareInfo &hwInfo) const {
-    return false;
+    return !hwInfo.featureTable.ftrLocalMemory;
 }
 } // namespace NEO

shared/source/helpers/state_base_address.h

@@ -28,7 +28,8 @@ struct StateBaseAddressHelper {
         uint64_t generalStateBase,
         bool setGeneralStateBaseAddress,
         uint32_t statelessMocsIndex,
-        uint64_t internalHeapBase,
+        uint64_t indirectObjectHeapBaseAddress,
+        uint64_t instructionHeapBaseAddress,
         bool setInstructionStateBaseAddress,
         GmmHelper *gmmHelper,
         bool isMultiOsContextCapable);
@@ -37,7 +38,7 @@ struct StateBaseAddressHelper {
         STATE_BASE_ADDRESS *stateBaseAddress,
         const IndirectHeap *ssh,
         bool setGeneralStateBaseAddress,
-        uint64_t internalHeapBase,
+        uint64_t indirectObjectHeapBaseAddress,
         GmmHelper *gmmHelper,
         bool isMultiOsContextCapable);
 

shared/source/helpers/state_base_address_base.inl

@@ -23,7 +23,8 @@ void StateBaseAddressHelper<GfxFamily>::programStateBaseAddress(
     uint64_t generalStateBase,
     bool setGeneralStateBaseAddress,
     uint32_t statelessMocsIndex,
-    uint64_t internalHeapBase,
+    uint64_t indirectObjectHeapBaseAddress,
+    uint64_t instructionHeapBaseAddress,
     bool setInstructionStateBaseAddress,
     GmmHelper *gmmHelper,
     bool isMultiOsContextCapable) {
@@ -51,7 +52,7 @@ void StateBaseAddressHelper<GfxFamily>::programStateBaseAddress(
 
     if (setInstructionStateBaseAddress) {
         stateBaseAddress->setInstructionBaseAddressModifyEnable(true);
-        stateBaseAddress->setInstructionBaseAddress(internalHeapBase);
+        stateBaseAddress->setInstructionBaseAddress(instructionHeapBaseAddress);
         stateBaseAddress->setInstructionBufferSizeModifyEnable(true);
         stateBaseAddress->setInstructionBufferSize(MemoryConstants::sizeOf4GBinPageEntities);
         stateBaseAddress->setInstructionMemoryObjectControlState(gmmHelper->getMOCS(GMM_RESOURCE_USAGE_OCL_STATE_HEAP_BUFFER));
@@ -73,7 +74,7 @@ void StateBaseAddressHelper<GfxFamily>::programStateBaseAddress(
 
     stateBaseAddress->setStatelessDataPortAccessMemoryObjectControlState(statelessMocsIndex);
 
-    appendStateBaseAddressParameters(stateBaseAddress, ssh, setGeneralStateBaseAddress, internalHeapBase, gmmHelper, isMultiOsContextCapable);
+    appendStateBaseAddressParameters(stateBaseAddress, ssh, setGeneralStateBaseAddress, indirectObjectHeapBaseAddress, gmmHelper, isMultiOsContextCapable);
 }
 
 } // namespace NEO

shared/source/helpers/state_base_address_bdw.inl

@@ -14,7 +14,7 @@ void StateBaseAddressHelper<GfxFamily>::appendStateBaseAddressParameters(
     STATE_BASE_ADDRESS *stateBaseAddress,
     const IndirectHeap *ssh,
     bool setGeneralStateBaseAddress,
-    uint64_t internalHeapBase,
+    uint64_t indirectObjectHeapBaseAddress,
     GmmHelper *gmmHelper,
     bool isMultiOsContextCapable) {
 }

shared/source/helpers/state_base_address_skl_plus.inl

@@ -14,7 +14,7 @@ void StateBaseAddressHelper<GfxFamily>::appendStateBaseAddressParameters(
     STATE_BASE_ADDRESS *stateBaseAddress,
     const IndirectHeap *ssh,
     bool setGeneralStateBaseAddress,
-    uint64_t internalHeapBase,
+    uint64_t indirectObjectHeapBaseAddress,
     GmmHelper *gmmHelper,
     bool isMultiOsContextCapable) {
 

shared/test/unit_test/command_container/command_container_tests.cpp

@@ -127,8 +127,11 @@ TEST_F(CommandContainerTest, givenCommandContainerWhenInitializeThenEverythingIs
         auto heapAllocation = cmdContainer.getIndirectHeapAllocation(static_cast<HeapType>(i));
         EXPECT_EQ(indirectHeap->getGraphicsAllocation(), heapAllocation);
     }
+
+    auto &hwHelper = HwHelper::get(pDevice->getHardwareInfo().platform.eRenderCoreFamily);
+
     EXPECT_EQ(cmdContainer.getInstructionHeapBaseAddress(),
-              pDevice->getMemoryManager()->getInternalHeapBaseAddress(0, cmdContainer.getIndirectHeap(HeapType::INDIRECT_OBJECT)->getGraphicsAllocation()->isAllocatedInLocalMemoryPool()));
+              pDevice->getMemoryManager()->getInternalHeapBaseAddress(0, !hwHelper.useSystemMemoryPlacementForISA(pDevice->getHardwareInfo())));
 }
 
 TEST_F(CommandContainerTest, givenCommandContainerWhenInitializeWithoutDeviceThenReturnedFalse) {
@@ -497,15 +500,18 @@ TEST_F(CommandContainerHeaps, givenCommandContainerForDifferentRootDevicesThenIn
     auto device0 = std::unique_ptr<MockDevice>(Device::create<MockDevice>(executionEnvironment, 0u));
     auto device1 = std::unique_ptr<MockDevice>(Device::create<MockDevice>(executionEnvironment, 1u));
 
+    auto &hwHelper0 = HwHelper::get(device0->getHardwareInfo().platform.eRenderCoreFamily);
+    auto &hwHelper1 = HwHelper::get(device1->getHardwareInfo().platform.eRenderCoreFamily);
+
     CommandContainer cmdContainer0;
     cmdContainer0.initialize(device0.get());
-    bool useLocalMemory0 = cmdContainer0.getIndirectHeap(HeapType::INDIRECT_OBJECT)->getGraphicsAllocation()->isAllocatedInLocalMemoryPool();
+    bool useLocalMemory0 = !hwHelper0.useSystemMemoryPlacementForISA(device0->getHardwareInfo());
     uint64_t baseAddressHeapDevice0 = device0.get()->getMemoryManager()->getInternalHeapBaseAddress(device0->getRootDeviceIndex(), useLocalMemory0);
     EXPECT_EQ(cmdContainer0.getInstructionHeapBaseAddress(), baseAddressHeapDevice0);
 
     CommandContainer cmdContainer1;
     cmdContainer1.initialize(device1.get());
-    bool useLocalMemory1 = cmdContainer0.getIndirectHeap(HeapType::INDIRECT_OBJECT)->getGraphicsAllocation()->isAllocatedInLocalMemoryPool();
+    bool useLocalMemory1 = !hwHelper1.useSystemMemoryPlacementForISA(device0->getHardwareInfo());
     uint64_t baseAddressHeapDevice1 = device1.get()->getMemoryManager()->getInternalHeapBaseAddress(device1->getRootDeviceIndex(), useLocalMemory1);
     EXPECT_EQ(cmdContainer1.getInstructionHeapBaseAddress(), baseAddressHeapDevice1);
 }

shared/test/unit_test/helpers/state_base_address_tests.cpp

@@ -55,6 +55,7 @@ HWTEST2_F(SBATest, WhenProgramStateBaseAddressParametersIsCalledThenSBACmdHasBin
         false,
         0,
         0,
+        0,
         false,
         pDevice->getGmmHelper(),
         true);