Merge pull request #6514 from JJenkx/test/gpu-timing-incomplete-batch

test(agc): cover incomplete GPU timing batch completion

core/libs/prx/libSceAgcDriver/CMakeLists.txt

@@ -1191,8 +1191,12 @@ configure_windows_unwind(agc_driver_recorder_tests)
 if(BUILD_TESTING)
     set_target_properties(agc_driver_recorder_tests PROPERTIES EXCLUDE_FROM_ALL OFF)
     add_test(NAME agc_recorder_device_selection COMMAND agc_driver_recorder_tests --device-selection-only)
+    add_test(NAME agc_gpu_timing_complete COMMAND agc_driver_recorder_tests --gpu-timing-complete-only)
+    add_test(NAME agc_gpu_timing_incomplete COMMAND agc_driver_recorder_tests --gpu-timing-incomplete-only)
+    set_tests_properties(agc_gpu_timing_complete agc_gpu_timing_incomplete PROPERTIES
+        ENVIRONMENT "APS5_PROFILE_GPU=1" TIMEOUT 15 SKIP_RETURN_CODE 77)
     if(WIN32)
-        set_tests_properties(agc_recorder_device_selection PROPERTIES ENVIRONMENT_MODIFICATION
+        set_tests_properties(agc_recorder_device_selection agc_gpu_timing_complete agc_gpu_timing_incomplete PROPERTIES ENVIRONMENT_MODIFICATION
             "PATH=path_list_prepend:$<TARGET_FILE_DIR:libc>")
     endif()
     add_test(NAME agc_pipeline_cache COMMAND agc_driver_recorder_tests --pipeline-cache-only)

core/libs/prx/libSceAgcDriver/tests/Recorder.cpp

@@ -392,6 +392,14 @@ public:
     bool NullDescriptors() const { return nullDescriptors; }
     const Context& GetContext() const { return context; }
     bool Graphics() const { return graphics; }
+    bool Timestamps() const {
+        std::uint32_t count = 0;
+        const auto queues = function<PFN_vkGetPhysicalDeviceQueueFamilyProperties>("vkGetPhysicalDeviceQueueFamilyProperties");
+        queues(context.physical, &count, nullptr);
+        std::vector<VkQueueFamilyProperties> families(count);
+        queues(context.physical, &count, families.data());
+        return families.at(context.queueFamily).timestampValidBits != 0;
+    }
     VkPhysicalDeviceProperties Properties() const {
         VkPhysicalDeviceProperties properties{};
         function<PFN_vkGetPhysicalDeviceProperties>("vkGetPhysicalDeviceProperties")(context.physical, &properties);
@@ -1195,6 +1203,44 @@ void afterRecordedWorkTests(const Device& device, Recorder& recorder) {
     Require(seen == 1, "an action ran before the completion label recorded ahead of it");
 }
 
+int gpuTimingCompletionTests(const Device& device, bool incomplete) {
+    const auto& context = device.GetContext();
+    if (!device.Timestamps()) {
+        std::cout << "skipped, queue timestamp queries unavailable\n";
+        return 77;
+    }
+    Require(Recorder::GpuTimingEnabled(), "GPU timing completion test requires APS5_PROFILE_GPU");
+    std::unique_lock gpu(GpuMutex());
+    int completed = 0;
+    Recorder recorder(context);
+    recorder.Activate();
+    for (int round = 0; round < 3; ++round) {
+        const bool missingEnd = incomplete && round == 1;
+        const auto first = recorder.BeginGpuTiming(0x12340);
+        Require(first != Recorder::NoTiming, "first GPU timing range was not allocated");
+        if (!missingEnd) recorder.EndGpuTiming(first);
+        const auto second = recorder.BeginGpuTiming(0x12341);
+        Require(second != Recorder::NoTiming && second != first, "second GPU timing range was not allocated");
+        recorder.EndGpuTiming(second);
+        recorder.OnComplete([&] { ++completed; });
+        const auto serial = recorder.Submissions() + 1;
+        recorder.Submit();
+        recorder.Sync();
+        Require(completed == round + 1 && recorder.Idle(), "GPU timing prevented batch completion");
+        std::size_t missing = 0;
+        const auto batches = recorder.CompletedBatches(serial - 1, serial, missing);
+        Require(missing == 0 && batches.size() == 1 && batches.front().serial == serial, "GPU timing batch completion record is missing");
+        const auto& batch = batches.front();
+        if (missingEnd) {
+            Require(batch.gpuStartNs == 0 && batch.gpuEndNs == 0, "incomplete query batch published GPU timings");
+        } else {
+            Require(batch.gpuStartNs > 0 && batch.gpuEndNs >= batch.gpuStartNs, "complete query batch lost GPU timings");
+        }
+    }
+    std::cout << "GPU timing batch completion checked, incomplete=" << incomplete << '\n';
+    return 0;
+}
+
 void batchStampTests(Recorder& recorder) {
     Require(recorder.Idle(), "batch stamps: the recorder is busy");
     double previousEnd = 0;
@@ -8960,6 +9006,9 @@ int main(int argc, char** argv) {
         Device& device = *created;
         if (argc == 2 && std::string_view(argv[1]) == "--host-import-lifetime-only") return RunHostImportLifetimeTests(device.GetContext());
         if (argc == 2 && std::string_view(argv[1]) == "--host-import-mutation-race-only") return RunHostImportMutationRaceTests(device.GetContext());
+        if (argc == 2 && (std::string_view(argv[1]) == "--gpu-timing-complete-only" || std::string_view(argv[1]) == "--gpu-timing-incomplete-only")) {
+            return gpuTimingCompletionTests(device, std::string_view(argv[1]) == "--gpu-timing-incomplete-only");
+        }
         if (argc == 2 && std::string_view(argv[1]) == "--import-watch-only") {
             importWatchTests(device);
             return 0;