feature: add debug flag to control graph patch preamble

Related-To: NEO-16434

Signed-off-by: Zbigniew Zdanowicz <zbigniew.zdanowicz@intel.com>

level_zero/core/test/unit_tests/experimental/test_graph.cpp

@@ -1026,7 +1026,8 @@ TEST_F(GraphTestInstantiationTest, WhenInstantiatingGraphThenBakeCommandsIntoCom
     EXPECT_EQ(ZE_RESULT_SUCCESS, zeCommandListEndGraphCaptureExp(immCmdListHandle, &srcGraphHandle, nullptr));
 
     ctx.cmdListsToReturn.push_back(new Mock<CommandList>());
-    ExecutableGraph execGraph;
+    MockExecutableGraph execGraph;
+    EXPECT_TRUE(execGraph.usePatchingPreamble);
 
     auto *graphHwCommands = ctx.cmdListsToReturn[0];
     EXPECT_EQ(0U, graphHwCommands->appendBarrierCalled);
@@ -1082,6 +1083,16 @@ TEST_F(GraphTestInstantiationTest, WhenInstantiatingGraphThenBakeCommandsIntoCom
     EXPECT_EQ(2U, graphHwCommands->appendLaunchKernelWithParametersCalled); // +1 for zeCommandListAppendLaunchKernelWithArguments
 }
 
+TEST_F(GraphTestInstantiationTest, GivenGraphPatchPreambleDebugFlagWhenInstantiatingGraphThenUseDebugSettingForPatchPreamble) {
+    GraphsCleanupGuard graphCleanup;
+
+    DebugManagerStateRestore restorer;
+    debugManager.flags.ForceDisableGraphPatchPreamble.set(1);
+
+    MockExecutableGraph execGraph;
+    EXPECT_FALSE(execGraph.usePatchingPreamble);
+}
+
 TEST_F(GraphTestInstantiationTest, givenInOrderCmdListAndRegularCbEventWhenInstantiateToGraphThenDoNotRecordExternalCbEvent) {
     GraphsCleanupGuard graphCleanup;
 

level_zero/core/test/unit_tests/mocks/mock_graph.h

@@ -24,6 +24,7 @@ struct MockExecutableGraph : ExecutableGraph {
     using ExecutableGraph::multiEngineGraph;
     using ExecutableGraph::myCommandLists;
     using ExecutableGraph::submissionChain;
+    using ExecutableGraph::usePatchingPreamble;
 };
 
 } // namespace ult

level_zero/experimental/source/graph/graph.cpp

@@ -7,6 +7,7 @@
 
 #include "level_zero/experimental/source/graph/graph.h"
 
+#include "shared/source/debug_settings/debug_settings_manager.h"
 #include "shared/source/helpers/gfx_core_helper.h"
 
 #include "level_zero/core/source/cmdlist/cmdlist.h"
@@ -375,6 +376,13 @@ ze_result_t Closure<CaptureApi::zeCommandListAppendLaunchKernelWithArguments>::i
     return result;
 }
 
+ExecutableGraph::ExecutableGraph() {
+    int32_t overrideDisablePatchingPreamble = NEO::debugManager.flags.ForceDisableGraphPatchPreamble.get();
+    if (overrideDisablePatchingPreamble != -1) {
+        this->usePatchingPreamble = (overrideDisablePatchingPreamble == 0);
+    }
+}
+
 ExecutableGraph::~ExecutableGraph() = default;
 
 L0::CommandList *ExecutableGraph::allocateAndAddCommandListSubmissionNode() {
@@ -494,7 +502,7 @@ ze_result_t ExecutableGraph::execute(L0::CommandList *executionTarget, void *pNe
             auto currSignalEvent = (myLastCommandList == *cmdList) ? hSignalEvent : nullptr;
 
             ze_command_list_handle_t hCmdList = *cmdList;
-            executionTarget->setPatchingPreamble(true, this->multiEngineGraph);
+            executionTarget->setPatchingPreamble(this->usePatchingPreamble, this->multiEngineGraph);
             auto res = executionTarget->appendCommandLists(1, &hCmdList, currSignalEvent, numWaitEvents, phWaitEvents);
             executionTarget->setPatchingPreamble(false, false);
             if (ZE_RESULT_SUCCESS != res) {
@@ -506,7 +514,7 @@ ze_result_t ExecutableGraph::execute(L0::CommandList *executionTarget, void *pNe
             if (L0::CommandList **cmdList = std::get_if<L0::CommandList *>(&this->submissionChain[submissioNodeId])) {
                 auto currSignalEvent = (myLastCommandList == *cmdList) ? hSignalEvent : nullptr;
                 ze_command_list_handle_t hCmdList = *cmdList;
-                executionTarget->setPatchingPreamble(true, this->multiEngineGraph);
+                executionTarget->setPatchingPreamble(this->usePatchingPreamble, this->multiEngineGraph);
                 auto res = executionTarget->appendCommandLists(1, &hCmdList, currSignalEvent, 0, nullptr);
                 executionTarget->setPatchingPreamble(false, false);
                 if (ZE_RESULT_SUCCESS != res) {

level_zero/experimental/source/graph/graph.h

@@ -251,8 +251,7 @@ struct GraphInstatiateSettings {
 };
 
 struct ExecutableGraph : _ze_executable_graph_handle_t {
-    ExecutableGraph() {
-    }
+    ExecutableGraph();
 
     ze_result_t instantiateFrom(Graph &graph, const GraphInstatiateSettings &settings);
     ze_result_t instantiateFrom(Graph &graph) {
@@ -297,6 +296,7 @@ struct ExecutableGraph : _ze_executable_graph_handle_t {
     GraphSubmissionChain submissionChain;
 
     bool multiEngineGraph = false;
+    bool usePatchingPreamble = true;
 };
 
 constexpr size_t maxVariantSize = 2 * 64;

shared/source/debug_settings/debug_variables_base.inl

@@ -320,6 +320,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, OverrideCopyOffloadMode, -1, "-1: default, 0: di
 DECLARE_DEBUG_VARIABLE(int32_t, UseSingleListForTemporaryAllocations, -1, "-1: default, 0: disabled, 0: enabled. If enabled, use single list, instead of per CSR for tracking temporary allocations")
 DECLARE_DEBUG_VARIABLE(int32_t, OverrideMaxMemAllocSizeMb, -1, "-1: default, >=0 override reported max mem alloc size in MB")
 DECLARE_DEBUG_VARIABLE(int32_t, DetectIncorrectPointersOnSetArgCalls, -1, "-1: default do not detect, 0: do not detect, 1: detect incorrect pointers and return error")
+DECLARE_DEBUG_VARIABLE(int32_t, ForceDisableGraphPatchPreamble, -1, "-1: default, 0: enable patch preamble, 1: disable graph patch preamble. If disabled, do not patch preamble graph internal command lists")
 
 /*LOGGING FLAGS*/
 DECLARE_DEBUG_VARIABLE(int32_t, PrintDriverDiagnostics, -1, "prints driver diagnostics messages to standard output, value corresponds to hint level")

shared/test/common/test_files/igdrcl.config

@@ -207,6 +207,7 @@ ReturnRawGpuTimestamps = 0
 EnableDeviceBasedTimestamps = 1
 MaxHwThreadsPercent = 0
 MinHwThreadsUnoccupied = 0
+ForceDisableGraphPatchPreamble = -1
 LimitBlitterMaxWidth = -1
 LimitBlitterMaxHeight = -1
 LimitBlitterMaxSetWidth = -1