Add StateComputeModeProperties to StreamProperties

Related-To: NEO-4940, NEO-4574


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

level_zero/core/source/cmdlist/cmdlist.h

@@ -9,6 +9,7 @@
 
 #include "shared/source/command_container/cmdcontainer.h"
 #include "shared/source/command_stream/preemption_mode.h"
+#include "shared/source/command_stream/stream_properties.h"
 
 #include "level_zero/core/source/cmdqueue/cmdqueue.h"
 #include "level_zero/core/source/device/device.h"
@@ -16,8 +17,6 @@
 #include <level_zero/ze_api.h>
 #include <level_zero/zet_api.h>
 
-#include "stream_properties.h"
-
 #include <vector>
 
 struct _ze_command_list_handle_t {};

level_zero/core/source/cmdlist/cmdlist_hw.h

@@ -7,11 +7,12 @@
 
 #pragma once
 
+#include "shared/source/command_stream/stream_properties.h"
+
 #include "level_zero/core/source/builtin/builtin_functions_lib.h"
 #include "level_zero/core/source/cmdlist/cmdlist_imp.h"
 
 #include "igfxfmid.h"
-#include "stream_properties.h"
 
 namespace NEO {
 enum class ImageType;
@@ -219,7 +220,7 @@ struct CommandListCoreFamily : CommandListImp {
                                                               bool isCooperative);
     ze_result_t appendLaunchKernelSplit(ze_kernel_handle_t hKernel, const ze_group_count_t *pThreadGroupDimensions, ze_event_handle_t hEvent);
     ze_result_t prepareIndirectParams(const ze_group_count_t *pThreadGroupDimensions);
-    void updateStreamProperties(Kernel &kernel);
+    void updateStreamProperties(Kernel &kernel, bool isMultiOsContextCapable);
     void clearCommandsToPatch();
 
     void applyMemoryRangesBarrier(uint32_t numRanges, const size_t *pRangeSizes,

level_zero/core/source/cmdlist/cmdlist_hw.inl

@@ -1847,7 +1847,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::prepareIndirectParams(const ze
 }
 
 template <GFXCORE_FAMILY gfxCoreFamily>
-void CommandListCoreFamily<gfxCoreFamily>::updateStreamProperties(Kernel &kernel) {
+void CommandListCoreFamily<gfxCoreFamily>::updateStreamProperties(Kernel &kernel, bool isMultiOsContextCapable) {
     using GfxFamily = typename NEO::GfxFamilyMapper<gfxCoreFamily>::GfxFamily;
     using VFE_STATE_TYPE = typename GfxFamily::VFE_STATE_TYPE;
 
@@ -1855,7 +1855,6 @@ void CommandListCoreFamily<gfxCoreFamily>::updateStreamProperties(Kernel &kernel
         requiredStreamState.setCooperativeKernelProperties(kernel.usesSyncBuffer(), device->getHwInfo());
         finalStreamState = requiredStreamState;
         containsAnyKernel = true;
-        return;
     }
 
     auto &hwInfo = device->getHwInfo();
@@ -1867,6 +1866,17 @@ void CommandListCoreFamily<gfxCoreFamily>::updateStreamProperties(Kernel &kernel
                                                         NEO::AdditionalKernelExecInfo::NotApplicable, finalStreamState);
         commandsToPatch.push_back({pVfeStateAddress, pVfeState, CommandToPatch::FrontEndState});
     }
+
+    auto &kernelAttributes = kernel.getKernelDescriptor().kernelAttributes;
+    auto &neoDevice = *device->getNEODevice();
+    finalStreamState.setStateComputeModeProperties(false, kernelAttributes.numGrfRequired, isMultiOsContextCapable,
+                                                   kernelAttributes.flags.useGlobalAtomics,
+                                                   (neoDevice.getNumAvailableDevices() > 1));
+    if (finalStreamState.stateComputeMode.isDirty()) {
+        NEO::EncodeWA<GfxFamily>::encodeAdditionalPipelineSelect(neoDevice, *commandContainer.getCommandStream(), true);
+        NEO::EncodeComputeMode<GfxFamily>::adjustComputeMode(*commandContainer.getCommandStream(), nullptr, finalStreamState.stateComputeMode);
+        NEO::EncodeWA<GfxFamily>::encodeAdditionalPipelineSelect(neoDevice, *commandContainer.getCommandStream(), false);
+    }
 }
 
 template <GFXCORE_FAMILY gfxCoreFamily>

level_zero/core/source/cmdlist/cmdlist_hw_base.inl

@@ -101,7 +101,7 @@ ze_result_t CommandListCoreFamily<gfxCoreFamily>::appendLaunchKernelWithParams(z
             kernel->getKernelDescriptor().kernelMetadata.kernelName.c_str());
     }
 
-    updateStreamProperties(*kernel);
+    updateStreamProperties(*kernel, false);
 
     NEO::EncodeDispatchKernel<GfxFamily>::encode(commandContainer,
                                                  reinterpret_cast<const void *>(pThreadGroupDimensions),

level_zero/core/source/cmdqueue/cmdqueue_hw.h

@@ -8,6 +8,7 @@
 #pragma once
 #include "shared/source/command_stream/command_stream_receiver.h"
 #include "shared/source/command_stream/scratch_space_controller.h"
+#include "shared/source/command_stream/stream_properties.h"
 #include "shared/source/command_stream/submissions_aggregator.h"
 #include "shared/source/helpers/constants.h"
 #include "shared/source/memory_manager/graphics_allocation.h"
@@ -16,7 +17,6 @@
 #include "level_zero/core/source/cmdqueue/cmdqueue_imp.h"
 
 #include "igfxfmid.h"
-#include "stream_properties.h"
 
 namespace L0 {
 

level_zero/core/source/cmdqueue/cmdqueue_hw.inl

@@ -13,6 +13,7 @@
 #include "shared/source/command_stream/command_stream_receiver_hw.h"
 #include "shared/source/command_stream/linear_stream.h"
 #include "shared/source/command_stream/preemption.h"
+#include "shared/source/command_stream/stream_properties.h"
 #include "shared/source/command_stream/thread_arbitration_policy.h"
 #include "shared/source/device/device.h"
 #include "shared/source/helpers/hw_helper.h"
@@ -34,7 +35,6 @@
 #include "level_zero/tools/source/metrics/metric.h"
 
 #include "pipe_control_args.h"
-#include "stream_properties.h"
 
 #include <limits>
 #include <thread>

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

@@ -132,11 +132,12 @@ HWTEST_F(CommandListAppendLaunchKernel, givenNotEnoughSpaceInCommandStreamWhenAp
     const auto stream = commandContainer.getCommandStream();
     const auto streamCpu = stream->getCpuBase();
 
+    ze_group_count_t groupCount{1, 1, 1};
+    commandList->appendLaunchKernel(kernel->toHandle(), &groupCount, nullptr, 0, nullptr);
     auto available = stream->getAvailableSpace();
     stream->getSpace(available - sizeof(MI_BATCH_BUFFER_END) - 16);
     auto bbEndPosition = stream->getSpace(0);
 
-    ze_group_count_t groupCount{1, 1, 1};
     commandList->appendLaunchKernel(kernel->toHandle(), &groupCount, nullptr, 0, nullptr);
 
     auto usedSpaceAfter = commandContainer.getCommandStream()->getUsed();