refactor: Add GDI profiling

Resolves: NEO-9236
Related-To: NEO-10036

Signed-off-by: Lukasz Jobczyk <[email protected]>

Author: LukaszJobczyk

CMakeLists.txt

@@ -579,11 +579,6 @@ else()
   set(NEO_TESTS_LISTENER_OPTION "--enable_default_listener")
 endif()
 
-# Put profiling enable flag into define
-if(KMD_PROFILING)
-  add_definitions(-DKMD_PROFILING=${KMD_PROFILING})
-endif()
-
 if(MSVC)
   # Force to treat warnings as errors
   if(NOT CMAKE_CXX_FLAGS MATCHES "/WX")

opencl/source/api/api_enter.h

@@ -12,10 +12,3 @@
 
 #define API_ENTER(retValPointer) \
     LoggerApiEnterWrapper<NEO::FileLogger<globalDebugFunctionalityLevel>::enabled()> ApiWrapperForSingleCall(__FUNCTION__, retValPointer)
-
-#if KMD_PROFILING == 1
-#undef API_ENTER
-
-#define API_ENTER(x) \
-    PerfProfilerApiWrapper globalPerfProfilersWrapperInstanceForSingleApiFunction(__FUNCTION__)
-#endif

opencl/test/unit_test/linux/main_linux_dll.cpp

@@ -234,7 +234,7 @@ TEST_F(DrmSimpleTests, givenPrintIoctlTimesWhenCallIoctlThenStatisticsAreGathere
     auto drm = DrmWrap::createDrm(*(mockExecutionEnvironment.rootDeviceEnvironments[0].get()));
 
     DebugManagerStateRestore restorer;
-    debugManager.flags.PrintIoctlTimes.set(true);
+    debugManager.flags.PrintKmdTimes.set(true);
     VariableBackup<decltype(forceExtraIoctlDuration)> backupForceExtraIoctlDuration(&forceExtraIoctlDuration, true);
 
     EXPECT_TRUE(drm->ioctlStatistics.empty());

shared/source/debug_settings/debug_variables_base.inl

@@ -308,7 +308,7 @@ DECLARE_DEBUG_VARIABLE(bool, PrintBOPrefetchingResult, false, "tracks the result
 DECLARE_DEBUG_VARIABLE(bool, PrintTagAllocationAddress, false, "Print tag allocation address for each engine")
 DECLARE_DEBUG_VARIABLE(bool, ProvideVerboseImplicitFlush, false, "provides verbose messages about implicit flush mechanism")
 DECLARE_DEBUG_VARIABLE(bool, PrintBlitDispatchDetails, false, "Print blit dispatch details")
-DECLARE_DEBUG_VARIABLE(bool, PrintIoctlTimes, false, "Print ioctl times")
+DECLARE_DEBUG_VARIABLE(bool, PrintKmdTimes, false, "Print ioctl times")
 DECLARE_DEBUG_VARIABLE(bool, PrintIoctlEntries, false, "Print ioctl being called")
 DECLARE_DEBUG_VARIABLE(bool, PrintUmdSharedMigration, false, "Print log message when shared allocation is being migrated by UMD")
 DECLARE_DEBUG_VARIABLE(bool, PrintImageBlitBlockCopyCmdDetails, false, "Prints XY_BLOCK_COPY_BLT command details")

shared/source/helpers/options.h

@@ -8,10 +8,6 @@
 #pragma once
 #include <cstdint>
 
-#ifndef KMD_PROFILING
-#define KMD_PROFILING 0
-#endif
-
 namespace NEO {
 enum CommandStreamReceiverType {
     // Use receiver for real HW

shared/source/os_interface/linux/drm_neo.cpp

@@ -88,7 +88,7 @@ int Drm::ioctl(DrmIoctl request, void *arg) {
     int returnedErrno = 0;
     SYSTEM_ENTER();
     do {
-        auto measureTime = debugManager.flags.PrintIoctlTimes.get();
+        auto measureTime = debugManager.flags.PrintKmdTimes.get();
         std::chrono::steady_clock::time_point start;
         std::chrono::steady_clock::time_point end;
 
@@ -644,7 +644,7 @@ std::vector<DataType> Drm::query(uint32_t queryId, uint32_t queryItemFlags) {
 }
 
 void Drm::printIoctlStatistics() {
-    if (!debugManager.flags.PrintIoctlTimes.get()) {
+    if (!debugManager.flags.PrintKmdTimes.get()) {
         return;
     }
 

shared/source/os_interface/windows/CMakeLists.txt

@@ -68,6 +68,7 @@ set(NEO_CORE_OS_INTERFACE_WDDM
     ${CMAKE_CURRENT_SOURCE_DIR}/gdi_interface.h
     ${CMAKE_CURRENT_SOURCE_DIR}/gdi_interface_logging.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/gdi_interface_logging.h
+    ${CMAKE_CURRENT_SOURCE_DIR}/gdi_profiling.h
     ${CMAKE_CURRENT_SOURCE_DIR}/hw_device_id_win.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/hw_device_id.h
     ${CMAKE_CURRENT_SOURCE_DIR}/product_helper_wddm.cpp

shared/source/os_interface/windows/gdi_interface.cpp

@@ -25,6 +25,7 @@ Gdi::~Gdi() {
     if constexpr (GdiLogging::gdiLoggingSupport) {
         GdiLogging::close();
     }
+    this->profiler.printGdiTimes();
 }
 
 bool Gdi::setupHwQueueProcAddresses() {

shared/source/os_interface/windows/gdi_interface.h

@@ -8,67 +8,73 @@
 #pragma once
 #include "shared/source/os_interface/os_library.h"
 #include "shared/source/os_interface/windows/d3dkmthk_wrapper.h"
+#include "shared/source/os_interface/windows/gdi_profiling.h"
 #include "shared/source/os_interface/windows/thk_wrapper.h"
 
 #include <memory>
 
 namespace NEO {
 
+#define DEFINE_THK_WRAPPER(TYPE, VAR) ThkWrapper<TYPE> VAR = ThkWrapper<TYPE>(this->profiler, #TYPE, this->gdiId++);
+
 class Gdi {
+    uint32_t gdiId = 0;
+    GdiProfiler profiler;
+
   public:
     Gdi();
     MOCKABLE_VIRTUAL ~Gdi();
 
-    ThkWrapper<IN OUT CONST D3DKMT_OPENADAPTERFROMLUID *> openAdapterFromLuid{};
-    ThkWrapper<IN OUT D3DKMT_CREATEALLOCATION *> createAllocation{};
-    ThkWrapper<IN OUT D3DKMT_CREATEALLOCATION *> createAllocation2{};
+    DEFINE_THK_WRAPPER(IN OUT CONST D3DKMT_OPENADAPTERFROMLUID *, openAdapterFromLuid);
+    DEFINE_THK_WRAPPER(IN OUT D3DKMT_CREATEALLOCATION *, createAllocation);
+    DEFINE_THK_WRAPPER(IN OUT D3DKMT_CREATEALLOCATION *, createAllocation2);
     NTSTATUS(APIENTRY *shareObjects)
     (UINT cObjects, const D3DKMT_HANDLE *hObjects, POBJECT_ATTRIBUTES pObjectAttributes, DWORD dwDesiredAccess, HANDLE *phSharedNtHandle) = {};
-    ThkWrapper<IN CONST D3DKMT_DESTROYALLOCATION *> destroyAllocation{};
-    ThkWrapper<IN CONST D3DKMT_DESTROYALLOCATION2 *> destroyAllocation2{};
-    ThkWrapper<IN CONST D3DKMT_QUERYADAPTERINFO *> queryAdapterInfo{};
-    ThkWrapper<IN CONST D3DKMT_CLOSEADAPTER *> closeAdapter{};
-    ThkWrapper<IN OUT D3DKMT_CREATEDEVICE *> createDevice{};
-    ThkWrapper<IN CONST D3DKMT_DESTROYDEVICE *> destroyDevice{};
-    ThkWrapper<IN CONST D3DKMT_ESCAPE *> escape{};
-    ThkWrapper<IN D3DKMT_CREATECONTEXTVIRTUAL *> createContext{};
-    ThkWrapper<IN CONST D3DKMT_DESTROYCONTEXT *> destroyContext{};
-    ThkWrapper<IN OUT D3DKMT_OPENRESOURCE *> openResource{};
-    ThkWrapper<IN OUT D3DKMT_OPENRESOURCEFROMNTHANDLE *> openResourceFromNtHandle{};
-    ThkWrapper<IN OUT D3DKMT_QUERYRESOURCEINFO *> queryResourceInfo{};
-    ThkWrapper<IN OUT D3DKMT_QUERYRESOURCEINFOFROMNTHANDLE *> queryResourceInfoFromNtHandle{};
-    ThkWrapper<IN OUT D3DKMT_CREATESYNCHRONIZATIONOBJECT *> createSynchronizationObject{};
-    ThkWrapper<IN OUT D3DKMT_CREATESYNCHRONIZATIONOBJECT2 *> createSynchronizationObject2{};
-    ThkWrapper<IN CONST D3DKMT_DESTROYSYNCHRONIZATIONOBJECT *> destroySynchronizationObject{};
-    ThkWrapper<IN CONST D3DKMT_SIGNALSYNCHRONIZATIONOBJECT *> signalSynchronizationObject{};
-    ThkWrapper<IN CONST_FROM_WDK_10_0_18328_0 D3DKMT_WAITFORSYNCHRONIZATIONOBJECT *> waitForSynchronizationObject{};
-    ThkWrapper<IN CONST D3DKMT_WAITFORSYNCHRONIZATIONOBJECTFROMCPU *> waitForSynchronizationObjectFromCpu{};
-    ThkWrapper<IN CONST D3DKMT_SIGNALSYNCHRONIZATIONOBJECTFROMCPU *> signalSynchronizationObjectFromCpu{};
-    ThkWrapper<IN CONST D3DKMT_WAITFORSYNCHRONIZATIONOBJECTFROMGPU *> waitForSynchronizationObjectFromGpu{};
-    ThkWrapper<IN CONST D3DKMT_SIGNALSYNCHRONIZATIONOBJECTFROMGPU *> signalSynchronizationObjectFromGpu{};
-    ThkWrapper<IN OUT D3DKMT_CREATEPAGINGQUEUE *> createPagingQueue{};
-    ThkWrapper<IN OUT D3DDDI_DESTROYPAGINGQUEUE *> destroyPagingQueue{};
-    ThkWrapper<IN OUT D3DKMT_LOCK2 *> lock2{};
-    ThkWrapper<IN CONST D3DKMT_UNLOCK2 *> unlock2{};
-    ThkWrapper<IN OUT D3DDDI_MAPGPUVIRTUALADDRESS *> mapGpuVirtualAddress{};
-    ThkWrapper<IN OUT D3DDDI_RESERVEGPUVIRTUALADDRESS *> reserveGpuVirtualAddress{};
-    ThkWrapper<IN CONST D3DKMT_FREEGPUVIRTUALADDRESS *> freeGpuVirtualAddress{};
-    ThkWrapper<IN CONST D3DKMT_UPDATEGPUVIRTUALADDRESS *> updateGpuVirtualAddress{};
-    ThkWrapper<IN CONST D3DKMT_SUBMITCOMMAND *> submitCommand{};
-    ThkWrapper<IN OUT D3DDDI_MAKERESIDENT *> makeResident{};
-    ThkWrapper<IN D3DKMT_EVICT *> evict{};
-    ThkWrapper<IN D3DKMT_REGISTERTRIMNOTIFICATION *> registerTrimNotification{};
-    ThkWrapper<IN D3DKMT_UNREGISTERTRIMNOTIFICATION *> unregisterTrimNotification{};
-    ThkWrapper<IN CONST D3DKMT_SETALLOCATIONPRIORITY *> setAllocationPriority{};
-    ThkWrapper<IN CONST D3DKMT_SETCONTEXTSCHEDULINGPRIORITY *> setSchedulingPriority{};
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_DESTROYALLOCATION *, destroyAllocation);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_DESTROYALLOCATION2 *, destroyAllocation2);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_QUERYADAPTERINFO *, queryAdapterInfo);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_CLOSEADAPTER *, closeAdapter);
+    DEFINE_THK_WRAPPER(IN OUT D3DKMT_CREATEDEVICE *, createDevice);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_DESTROYDEVICE *, destroyDevice);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_ESCAPE *, escape);
+    DEFINE_THK_WRAPPER(IN D3DKMT_CREATECONTEXTVIRTUAL *, createContext);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_DESTROYCONTEXT *, destroyContext);
+    DEFINE_THK_WRAPPER(IN OUT D3DKMT_OPENRESOURCE *, openResource);
+    DEFINE_THK_WRAPPER(IN OUT D3DKMT_OPENRESOURCEFROMNTHANDLE *, openResourceFromNtHandle);
+    DEFINE_THK_WRAPPER(IN OUT D3DKMT_QUERYRESOURCEINFO *, queryResourceInfo);
+    DEFINE_THK_WRAPPER(IN OUT D3DKMT_QUERYRESOURCEINFOFROMNTHANDLE *, queryResourceInfoFromNtHandle);
+    DEFINE_THK_WRAPPER(IN OUT D3DKMT_CREATESYNCHRONIZATIONOBJECT *, createSynchronizationObject);
+    DEFINE_THK_WRAPPER(IN OUT D3DKMT_CREATESYNCHRONIZATIONOBJECT2 *, createSynchronizationObject2);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_DESTROYSYNCHRONIZATIONOBJECT *, destroySynchronizationObject);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_SIGNALSYNCHRONIZATIONOBJECT *, signalSynchronizationObject);
+    DEFINE_THK_WRAPPER(IN CONST_FROM_WDK_10_0_18328_0 D3DKMT_WAITFORSYNCHRONIZATIONOBJECT *, waitForSynchronizationObject);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_WAITFORSYNCHRONIZATIONOBJECTFROMCPU *, waitForSynchronizationObjectFromCpu);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_SIGNALSYNCHRONIZATIONOBJECTFROMCPU *, signalSynchronizationObjectFromCpu);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_WAITFORSYNCHRONIZATIONOBJECTFROMGPU *, waitForSynchronizationObjectFromGpu);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_SIGNALSYNCHRONIZATIONOBJECTFROMGPU *, signalSynchronizationObjectFromGpu);
+    DEFINE_THK_WRAPPER(IN OUT D3DKMT_CREATEPAGINGQUEUE *, createPagingQueue);
+    DEFINE_THK_WRAPPER(IN OUT D3DDDI_DESTROYPAGINGQUEUE *, destroyPagingQueue);
+    DEFINE_THK_WRAPPER(IN OUT D3DKMT_LOCK2 *, lock2);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_UNLOCK2 *, unlock2);
+    DEFINE_THK_WRAPPER(IN OUT D3DDDI_MAPGPUVIRTUALADDRESS *, mapGpuVirtualAddress);
+    DEFINE_THK_WRAPPER(IN OUT D3DDDI_RESERVEGPUVIRTUALADDRESS *, reserveGpuVirtualAddress);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_FREEGPUVIRTUALADDRESS *, freeGpuVirtualAddress);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_UPDATEGPUVIRTUALADDRESS *, updateGpuVirtualAddress);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_SUBMITCOMMAND *, submitCommand);
+    DEFINE_THK_WRAPPER(IN OUT D3DDDI_MAKERESIDENT *, makeResident);
+    DEFINE_THK_WRAPPER(IN D3DKMT_EVICT *, evict);
+    DEFINE_THK_WRAPPER(IN D3DKMT_REGISTERTRIMNOTIFICATION *, registerTrimNotification);
+    DEFINE_THK_WRAPPER(IN D3DKMT_UNREGISTERTRIMNOTIFICATION *, unregisterTrimNotification);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_SETALLOCATIONPRIORITY *, setAllocationPriority);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_SETCONTEXTSCHEDULINGPRIORITY *, setSchedulingPriority);
 
     // HW queue
-    ThkWrapper<IN OUT D3DKMT_CREATEHWQUEUE *> createHwQueue{};
-    ThkWrapper<IN CONST D3DKMT_DESTROYHWQUEUE *> destroyHwQueue{};
-    ThkWrapper<IN CONST D3DKMT_SUBMITCOMMANDTOHWQUEUE *> submitCommandToHwQueue{};
+    DEFINE_THK_WRAPPER(IN OUT D3DKMT_CREATEHWQUEUE *, createHwQueue);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_DESTROYHWQUEUE *, destroyHwQueue);
+    DEFINE_THK_WRAPPER(IN CONST D3DKMT_SUBMITCOMMANDTOHWQUEUE *, submitCommandToHwQueue);
 
     // For debug purposes
-    ThkWrapper<IN OUT D3DKMT_GETDEVICESTATE *> getDeviceState{};
+    DEFINE_THK_WRAPPER(IN OUT D3DKMT_GETDEVICESTATE *, getDeviceState);
 
     bool isInitialized() {
         return initialized;
@@ -79,7 +85,7 @@ class Gdi {
   protected:
     OsLibrary *createGdiDLL();
     MOCKABLE_VIRTUAL bool getAllProcAddresses();
-    std::unique_ptr<NEO::OsLibrary> gdiDll;
+    std::unique_ptr<NEO::OsLibrary> gdiDll = nullptr;
     bool initialized = false;
 };
 } // namespace NEO

shared/source/os_interface/windows/gdi_profiling.h

@@ -0,0 +1,75 @@
+/*
+ * Copyright (C) 2018-2024 Intel Corporation
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ */
+
+#pragma once
+
+#include <algorithm>
+#include <limits>
+#include <vector>
+
+namespace NEO {
+
+class GdiProfiler {
+
+    struct StatisticsEntry {
+        long long totalTime = 0;
+        uint64_t count = 0;
+        long long minTime = std::numeric_limits<long long>::max();
+        long long maxTime = 0;
+        const char *gdiCall = nullptr;
+        size_t getLength() const {
+            return this->gdiCall ? strlen(this->gdiCall) : 0u;
+        }
+    };
+
+  public:
+    void printGdiTimes() {
+        if (this->gdiStatistics.empty()) {
+            return;
+        }
+
+        auto maxCallLengthIt = std::max_element(this->gdiStatistics.begin(), this->gdiStatistics.end(), [](const auto &gdiData1, const auto &gdiData2) {
+            return gdiData1.getLength() < gdiData2.getLength();
+        });
+        auto maxCallLength = static_cast<int>(strlen(maxCallLengthIt->gdiCall));
+
+        printf("\n--- Gdi statistics ---\n");
+        printf("%*s %15s %10s %25s %15s %15s", maxCallLength, "Request", "Total time(ns)", "Count", "Avg time per gdi call", "Min", "Max\n");
+        for (const auto &gdiData : this->gdiStatistics) {
+            if (gdiData.count == 0) {
+                continue;
+            }
+            printf("%*s %15llu %10lu %25f %15lld %15lld\n",
+                   maxCallLength,
+                   gdiData.gdiCall,
+                   gdiData.totalTime,
+                   static_cast<unsigned long>(gdiData.count),
+                   gdiData.totalTime / static_cast<double>(gdiData.count),
+                   gdiData.minTime,
+                   gdiData.maxTime);
+        }
+        printf("\n");
+    }
+
+    void recordElapsedTime(long long elapsedTime, const char *name, uint32_t id) {
+        if (this->gdiStatistics.size() <= id) {
+            this->gdiStatistics.resize(id + 1u);
+        }
+
+        auto &gdiData = this->gdiStatistics[id];
+
+        gdiData.gdiCall = name;
+        gdiData.totalTime += elapsedTime;
+        gdiData.count++;
+        gdiData.minTime = std::min(gdiData.minTime, elapsedTime);
+        gdiData.maxTime = std::max(gdiData.maxTime, elapsedTime);
+    }
+
+  protected:
+    std::vector<StatisticsEntry> gdiStatistics{};
+};
+} // namespace NEO