Move hWnd association for GPU-GDI blt model presents to token propaga…

…tion instead of creation. This aligns to when DWM receives the token and decides which frame should be composed into a window, and is where similar logic already existed for composition buffer presents. This was causing PresentMon's latencies for this mode to be off by ~1 VSync for some apps.

Also fixed several regressions in ETL file processing introduced by refactoring:
1. ETL file is not mutually exclusive with filename.
2. ETL files can cause the ETW thread to exit without an external signal. This was triggering asserts. When this happens, the app should exit cleanly (at least, that's how it used to behave).
3. The start time for a trace session was uninitialized and caused CSV files to have impossible "time in second" values for each frame.

README.md

@@ -29,18 +29,18 @@ SOFTWARE.
 ## Command line options
 
 ```html
-Capture target options (use one of the following):
+Capture target options:
     -captureall                Record all processes (default).
     -process_name [exe name]   Record specific process specified by name.
     -process_id [integer]      Record specific process specified by ID.
-    -etl_file [path]           Consume events from an ETL file instead of a running process.
 
 Output options:
     -no_csv                    Do not create any output file.
     -output_file [path]        Write CSV output to specified path. Otherwise, the default is
                                PresentMon-PROCESSNAME-TIME.csv.
 
 Control and filtering options:
+    -etl_file [path]           Consume events from an ETL file instead of a running process.
     -scroll_toggle             Only record events while scroll lock is enabled.
     -scroll_indicator          Set scroll lock while recording events.
     -hotkey [key]              Use specified key to start and stop recording, writing to a

source/CommandLine.cpp

@@ -220,18 +220,18 @@ void PrintHelp()
     fprintf(stderr,
         "PresentMon version 1.1.0\n"
         "\n"
-        "Capture target options (use one of the following):\n"
+        "Capture target options:\n"
         "    -captureall                Record all processes (default).\n"
         "    -process_name [exe name]   Record specific process specified by name.\n"
         "    -process_id [integer]      Record specific process specified by ID.\n"
-        "    -etl_file [path]           Consume events from an ETL file instead of a running process.\n"
         "\n"
         "Output options:\n"
         "    -no_csv                    Do not create any output file.\n"
         "    -output_file [path]        Write CSV output to specified path. Otherwise, the default is\n"
         "                               PresentMon-PROCESSNAME-TIME.csv.\n"
         "\n"
         "Control and filtering options:\n"
+        "    -etl_file [path]           Consume events from an ETL file instead of a running process.\n"
         "    -scroll_toggle             Only record events while scroll lock is enabled.\n"
         "    -scroll_indicator          Set scroll lock while recording events.\n"
         "    -hotkey [key]              Use specified key to start and stop recording, writing to a\n"
@@ -303,9 +303,8 @@ bool ParseCommandLine(int argc, char** argv, CommandLineArgs* args)
 
     // Validate command line arguments
     if (((args->mTargetProcessName == nullptr) ? 0 : 1) +
-        ((args->mTargetPid         <= 0      ) ? 0 : 1) +
-        ((args->mEtlFileName       == nullptr) ? 0 : 1) > 1) {
-        fprintf(stderr, "error: only specify one of -captureall, -process_name, -process_id, or -etl_file.\n");
+        ((args->mTargetPid         <= 0      ) ? 0 : 1) > 1) {
+        fprintf(stderr, "error: only specify one of -captureall, -process_name, or -process_id.\n");
         PrintHelp();
         return false;
     }

source/PresentMon.cpp

@@ -523,7 +523,11 @@ void EtwConsumingThread(const CommandLineArgs& args)
                 }
 
                 if (doneProcessingEvents) {
-                    assert(EtwThreadsShouldQuit());
+                    assert(EtwThreadsShouldQuit() || args.mEtlFileName);
+                    if (!EtwThreadsShouldQuit()) {
+                        PostStopRecording();
+                        PostQuitProcess();
+                    }
                     break;
                 }
 

source/PresentMonTraceConsumer.cpp

@@ -323,12 +323,6 @@ void PMTraceConsumer::OnDXGKrnlEvent(PEVENT_RECORD pEventRecord)
             CompletePresent(eventIter->second);
         }
 
-        if (eventIter->second->PresentMode == PresentMode::Composed_Copy_GPU_GDI) {
-            // Manipulate the map here
-            // When DWM is ready to present, we'll query for the most recent blt targeting this window and take it out of the map
-            mPresentByWindow[hWnd] = eventIter->second;
-        }
-
         // For all other events, just remember the hWnd, we might need it later
         eventIter->second->Hwnd = hWnd;
         eventIter->second->SeenDxgkPresent = true;
@@ -433,6 +427,12 @@ void PMTraceConsumer::OnDXGKrnlEvent(PEVENT_RECORD pEventRecord)
             mPresentsWaitingForDWM.emplace_back(eventIter->second);
         }
 
+        if (eventIter->second->PresentMode == PresentMode::Composed_Copy_GPU_GDI) {
+            // Manipulate the map here
+            // When DWM is ready to present, we'll query for the most recent blt targeting this window and take it out of the map
+            mPresentByWindow[eventIter->second->Hwnd] = eventIter->second;
+        }
+
         mDxgKrnlPresentHistoryTokens.erase(eventIter);
         break;
     }

source/TraceSession.hpp

@@ -50,6 +50,7 @@ struct TraceSession {
     TraceSession()
         : sessionHandle_(0)
         , traceHandle_(INVALID_PROCESSTRACE_HANDLE)
+        , startTime_(0)
         , pmData_(nullptr)
         , pmTraceConsumer_(nullptr)
     {