Add support for OffsetToSkipSetFFIDGP igc parameter

Related-To: NEO-3829

Change-Id: I18b237bac5301f57bbb26636bec94683c3d250a7
Signed-off-by: Filip Hazubski <filip.hazubski@intel.com>

runtime/command_queue/gpgpu_walker_bdw_plus.inl

@@ -7,6 +7,7 @@
 
 #pragma once
 #include "runtime/command_queue/gpgpu_walker_base.inl"
+#include "runtime/helpers/engine_node_helper.h"
 
 namespace NEO {
 
@@ -123,7 +124,7 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
     // Program the walker.  Invokes execution so all state should already be programmed
     auto pGpGpuWalkerCmd = static_cast<GPGPU_WALKER *>(commandStream.getSpace(sizeof(GPGPU_WALKER)));
     *pGpGpuWalkerCmd = GfxFamily::cmdInitGpgpuWalker;
-
+    auto isCcsUsed = isCcs(devQueueHw.getDevice().getDefaultEngine().osContext->getEngineType());
     bool inlineDataProgrammingRequired = HardwareCommandsHelper<GfxFamily>::inlineDataProgrammingRequired(scheduler);
     HardwareCommandsHelper<GfxFamily>::sendIndirectState(
         commandStream,
@@ -138,7 +139,8 @@ void GpgpuWalkerHelper<GfxFamily>::dispatchScheduler(
         preemptionMode,
         pGpGpuWalkerCmd,
         nullptr,
-        true);
+        true,
+        isCcsUsed);
 
     // Implement enabling special WA DisableLSQCROPERFforOCL if needed
     GpgpuWalkerHelper<GfxFamily>::applyWADisableLSQCROPERFforOCL(&commandStream, scheduler, true);

runtime/command_queue/hardware_interface_bdw_plus.inl

@@ -7,6 +7,7 @@
 
 #pragma once
 #include "runtime/command_queue/hardware_interface_base.inl"
+#include "runtime/os_interface/os_context.h"
 
 namespace NEO {
 
@@ -112,6 +113,8 @@ inline void HardwareInterface<GfxFamily>::programWalker(
         GpgpuWalkerHelper<GfxFamily>::setupTimestampPacket(&commandStream, walkerCmd, timestampPacketNode, TimestampPacketStorage::WriteOperationType::AfterWalker, commandQueue.getDevice().getHardwareInfo());
     }
 
+    auto isCcsUsed = isCcs(commandQueue.getGpgpuEngine().osContext->getEngineType());
+
     HardwareCommandsHelper<GfxFamily>::sendIndirectState(
         commandStream,
         dsh,
@@ -125,7 +128,8 @@ inline void HardwareInterface<GfxFamily>::programWalker(
         preemptionMode,
         walkerCmd,
         nullptr,
-        true);
+        true,
+        isCcsUsed);
 
     GpgpuWalkerHelper<GfxFamily>::setGpgpuWalkerThreadData(walkerCmd, globalOffsets, startWorkGroups,
                                                            numWorkGroups, localWorkSizes, simd, dim,

runtime/helpers/CMakeLists.txt

@@ -31,6 +31,8 @@ set(RUNTIME_SRCS_HELPERS_BASE
   ${CMAKE_CURRENT_SOURCE_DIR}/dispatch_info_builder.h
   ${CMAKE_CURRENT_SOURCE_DIR}/enable_product.inl
   ${CMAKE_CURRENT_SOURCE_DIR}/engine_control.h
+  ${CMAKE_CURRENT_SOURCE_DIR}${BRANCH_DIR_SUFFIX}/engine_node_helper.cpp
+  ${CMAKE_CURRENT_SOURCE_DIR}/engine_node_helper.h
   ${CMAKE_CURRENT_SOURCE_DIR}/error_mappers.h
   ${CMAKE_CURRENT_SOURCE_DIR}/file_io.cpp
   ${CMAKE_CURRENT_SOURCE_DIR}/file_io.h

runtime/helpers/engine_node_helper.cpp

@@ -0,0 +1,16 @@
+/*
+ * Copyright (C) 2019 Intel Corporation
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ */
+
+#include "engine_node.h"
+
+namespace NEO {
+
+bool isCcs(aub_stream::EngineType engineType) {
+    return engineType == aub_stream::ENGINE_CCS;
+}
+
+} // namespace NEO

runtime/helpers/engine_node_helper.h

@@ -0,0 +1,16 @@
+/*
+ * Copyright (C) 2019 Intel Corporation
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ */
+
+#pragma once
+
+#include "engine_node.h"
+
+namespace NEO {
+
+bool isCcs(aub_stream::EngineType engineType);
+
+} // namespace NEO

runtime/helpers/hardware_commands_helper.h

@@ -112,7 +112,8 @@ struct HardwareCommandsHelper : public PerThreadDataHelper {
         PreemptionMode preemptionMode,
         WALKER_TYPE<GfxFamily> *walkerCmd,
         INTERFACE_DESCRIPTOR_DATA *inlineInterfaceDescriptor,
-        bool localIdsGenerationByRuntime);
+        bool localIdsGenerationByRuntime,
+        bool isCcsUsed);
 
     static void programPerThreadData(
         size_t &sizePerThreadData,
@@ -140,7 +141,8 @@ struct HardwareCommandsHelper : public PerThreadDataHelper {
         const KernelInfo &kernelInfo,
         const bool &localIdsGenerationByRuntime,
         const bool &kernelUsesLocalIds,
-        Kernel &kernel);
+        Kernel &kernel,
+        bool isCssUsed);
 
     static size_t getSizeRequiredCS(const Kernel *kernel);
     static size_t getSizeRequiredForCacheFlush(const CommandQueue &commandQueue, const Kernel *kernel, uint64_t postSyncAddress);

runtime/helpers/hardware_commands_helper.inl

@@ -239,7 +239,8 @@ size_t HardwareCommandsHelper<GfxFamily>::sendIndirectState(
     PreemptionMode preemptionMode,
     WALKER_TYPE<GfxFamily> *walkerCmd,
     INTERFACE_DESCRIPTOR_DATA *inlineInterfaceDescriptor,
-    bool localIdsGenerationByRuntime) {
+    bool localIdsGenerationByRuntime,
+    bool isCcsUsed) {
 
     using SAMPLER_STATE = typename GfxFamily::SAMPLER_STATE;
 
@@ -252,7 +253,8 @@ size_t HardwareCommandsHelper<GfxFamily>::sendIndirectState(
     const auto &kernelInfo = kernel.getKernelInfo();
     auto kernelAllocation = kernelInfo.getGraphicsAllocation();
     DEBUG_BREAK_IF(!kernelAllocation);
-    setKernelStartOffset(kernelStartOffset, kernelAllocation, kernelInfo, localIdsGenerationByRuntime, kernelUsesLocalIds, kernel);
+    setKernelStartOffset(kernelStartOffset, kernelAllocation, kernelInfo, localIdsGenerationByRuntime,
+                         kernelUsesLocalIds, kernel, isCcsUsed);
 
     const auto &patchInfo = kernelInfo.patchInfo;
 

runtime/helpers/hardware_commands_helper_base.inl

@@ -95,12 +95,21 @@ void HardwareCommandsHelper<GfxFamily>::setKernelStartOffset(
     const KernelInfo &kernelInfo,
     const bool &localIdsGenerationByRuntime,
     const bool &kernelUsesLocalIds,
-    Kernel &kernel) {
+    Kernel &kernel,
+    bool isCssUsed) {
 
     if (kernelAllocation) {
         kernelStartOffset = kernelInfo.getGraphicsAllocation()->getGpuAddressToPatch();
     }
     kernelStartOffset += kernel.getStartOffset();
+
+#ifdef WIN32
+    if ((kernel.getDevice().getHardwareInfo().platform.eProductFamily == IGFX_TIGERLAKE_LP) &&
+        (kernel.getDevice().getHardwareInfo().platform.usRevId == REVISION_A0) &&
+        isCssUsed) {
+        kernelStartOffset += kernelInfo.patchInfo.threadPayload->OffsetToSkipSetFFIDGP;
+    }
+#endif
 }
 
 template <typename GfxFamily>

unit_tests/helpers/hardware_commands_helper_tests.cpp

@@ -14,6 +14,7 @@
 #include "runtime/api/api.h"
 #include "runtime/built_ins/builtins_dispatch_builder.h"
 #include "runtime/command_queue/command_queue_hw.h"
+#include "runtime/helpers/engine_node_helper.h"
 #include "runtime/helpers/hardware_commands_helper.h"
 #include "unit_tests/fixtures/execution_model_kernel_fixture.h"
 #include "unit_tests/fixtures/hello_world_fixture.h"
@@ -326,6 +327,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, sendIndirectStateResourceUsage
         IDToffset,
         sizeof(INTERFACE_DESCRIPTOR_DATA));
     uint32_t interfaceDescriptorIndex = 0;
+    auto isCcsUsed = isCcs(cmdQ.getGpgpuEngine().osContext->getEngineType());
     HardwareCommandsHelper<FamilyType>::sendIndirectState(
         commandStream,
         dsh,
@@ -339,7 +341,8 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, sendIndirectStateResourceUsage
         pDevice->getPreemptionMode(),
         pWalkerCmd,
         nullptr,
-        true);
+        true,
+        isCcsUsed);
 
     // It's okay these are EXPECT_GE as they're only going to be used for
     // estimation purposes to avoid OOM.
@@ -376,6 +379,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, givenKernelWithFourBindingTabl
     const size_t localWorkSize = 256;
     const size_t localWorkSizes[3]{localWorkSize, 1, 1};
     uint32_t interfaceDescriptorIndex = 0;
+    auto isCcsUsed = isCcs(cmdQ.getGpgpuEngine().osContext->getEngineType());
     HardwareCommandsHelper<FamilyType>::sendIndirectState(
         commandStream,
         dsh,
@@ -389,7 +393,8 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, givenKernelWithFourBindingTabl
         pDevice->getPreemptionMode(),
         pWalkerCmd,
         nullptr,
-        true);
+        true,
+        isCcsUsed);
 
     auto interfaceDescriptor = reinterpret_cast<INTERFACE_DESCRIPTOR_DATA *>(dsh.getCpuBase());
     if (HardwareCommandsHelper<FamilyType>::doBindingTablePrefetch()) {
@@ -419,6 +424,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, givenKernelThatIsSchedulerWhen
     const size_t localWorkSize = 256;
     const size_t localWorkSizes[3]{localWorkSize, 1, 1};
     uint32_t interfaceDescriptorIndex = 0;
+    auto isCcsUsed = isCcs(cmdQ.getGpgpuEngine().osContext->getEngineType());
     HardwareCommandsHelper<FamilyType>::sendIndirectState(
         commandStream,
         dsh,
@@ -432,7 +438,8 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, givenKernelThatIsSchedulerWhen
         pDevice->getPreemptionMode(),
         pWalkerCmd,
         nullptr,
-        true);
+        true,
+        isCcsUsed);
 
     auto interfaceDescriptor = reinterpret_cast<INTERFACE_DESCRIPTOR_DATA *>(dsh.getCpuBase());
     EXPECT_EQ(0u, interfaceDescriptor->getBindingTableEntryCount());
@@ -456,6 +463,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, givenKernelWith100BindingTable
     const size_t localWorkSize = 256;
     const size_t localWorkSizes[3]{localWorkSize, 1, 1};
     uint32_t interfaceDescriptorIndex = 0;
+    auto isCcsUsed = isCcs(cmdQ.getGpgpuEngine().osContext->getEngineType());
     HardwareCommandsHelper<FamilyType>::sendIndirectState(
         commandStream,
         dsh,
@@ -469,7 +477,8 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, givenKernelWith100BindingTable
         pDevice->getPreemptionMode(),
         pWalkerCmd,
         nullptr,
-        true);
+        true,
+        isCcsUsed);
 
     auto interfaceDescriptor = reinterpret_cast<INTERFACE_DESCRIPTOR_DATA *>(dsh.getCpuBase());
     if (HardwareCommandsHelper<FamilyType>::doBindingTablePrefetch()) {
@@ -528,6 +537,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, whenSendingIndirectStateThenKe
     modifiedKernelInfo.workgroupDimensionsOrder[2] = 0;
     MockKernel mockKernel{kernel->getProgram(), modifiedKernelInfo, kernel->getDevice(), false};
     uint32_t interfaceDescriptorIndex = 0;
+    auto isCcsUsed = isCcs(cmdQ.getGpgpuEngine().osContext->getEngineType());
     HardwareCommandsHelper<FamilyType>::sendIndirectState(
         commandStream,
         dsh,
@@ -541,7 +551,8 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, whenSendingIndirectStateThenKe
         pDevice->getPreemptionMode(),
         pWalkerCmd,
         nullptr,
-        true);
+        true,
+        isCcsUsed);
 
     size_t numThreads = localWorkSizeX * localWorkSizeY * localWorkSizeZ;
     numThreads = Math::divideAndRoundUp(numThreads, modifiedKernelInfo.getMaxSimdSize());
@@ -607,6 +618,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, usedBindingTableStatePointer)
     // force statefull path for buffers
     const_cast<KernelInfo &>(kernelInfo).requiresSshForBuffers = true;
     uint32_t interfaceDescriptorIndex = 0;
+    auto isCcsUsed = isCcs(cmdQ.getGpgpuEngine().osContext->getEngineType());
     HardwareCommandsHelper<FamilyType>::sendIndirectState(
         commandStream,
         dsh,
@@ -620,7 +632,8 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, usedBindingTableStatePointer)
         pDevice->getPreemptionMode(),
         pWalkerCmd,
         nullptr,
-        true);
+        true,
+        isCcsUsed);
 
     EXPECT_EQ(0x00000000u, *(&bindingTableStatesPointers[0]));
     EXPECT_EQ(0x00000040u, *(&bindingTableStatesPointers[1]));
@@ -767,6 +780,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, usedBindingTableStatePointersF
 
         // push surfaces states and binding table to given ssh heap
         uint32_t interfaceDescriptorIndex = 0;
+        auto isCcsUsed = isCcs(cmdQ.getGpgpuEngine().osContext->getEngineType());
         HardwareCommandsHelper<FamilyType>::sendIndirectState(
             commandStream,
             dsh,
@@ -780,7 +794,8 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, usedBindingTableStatePointersF
             pDevice->getPreemptionMode(),
             pWalkerCmd,
             nullptr,
-            true);
+            true,
+            isCcsUsed);
 
         bti = reinterpret_cast<typename FamilyType::BINDING_TABLE_STATE *>(reinterpret_cast<unsigned char *>(ssh.getCpuBase()) + localSshOffset + btiOffset);
         for (uint32_t i = 0; i < numSurfaces; ++i) {
@@ -1005,6 +1020,7 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, GivenKernelWithSamplersWhenInd
     mockKernelWithInternal->mockKernel->setCrossThreadData(mockKernelWithInternal->crossThreadData, sizeof(mockKernelWithInternal->crossThreadData));
     mockKernelWithInternal->mockKernel->setSshLocal(mockKernelWithInternal->sshLocal, sizeof(mockKernelWithInternal->sshLocal));
     uint32_t interfaceDescriptorIndex = 0;
+    auto isCcsUsed = isCcs(cmdQ.getGpgpuEngine().osContext->getEngineType());
     HardwareCommandsHelper<FamilyType>::sendIndirectState(
         commandStream,
         dsh,
@@ -1018,7 +1034,8 @@ HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, GivenKernelWithSamplersWhenInd
         pDevice->getPreemptionMode(),
         pWalkerCmd,
         nullptr,
-        true);
+        true,
+        isCcsUsed);
 
     bool isMemorySame = memcmp(borderColorPointer, mockDsh, borderColorSize) == 0;
     EXPECT_TRUE(isMemorySame);