Added kernel submission mode for tracers, PTI starts to require Level…

… Zero 1.2 runtime, update MDAPI dependencies

SOFTWARE

@@ -18,7 +18,7 @@ metrics-discovery:
     - sudo make install
     - sudo make package
     - sudo dpkg -i *.deb
-    commit: 7c416548fe13274effb795fcee55d099aee41b3f
+    commit: 8334347b2e61c339e1118680938470172dd38cb4
     url: https://github.com/intel/metrics-discovery
 metrics-library:
   github:
@@ -28,7 +28,7 @@ metrics-library:
     - make -j$(nproc)
     - sudo make package
     - sudo dpkg -i *.deb
-    commit: 0b05eb3138d3f0492c20111599d9a9a5af4a065f
+    commit: 3fd6eb0544fadcec2ac762aedee7c2d5d6479feb
     url: https://github.com/intel/metrics-library
 gtpin:
   archive:

VERSION

@@ -1 +1 @@
-0.41.1
+0.42.0

build_utils/get_md_headers.py

@@ -4,7 +4,7 @@
 import build_utils
 
 url = "https://github.com/intel/metrics-discovery.git"
-commit = "7c416548fe13274effb795fcee55d099aee41b3f"
+commit = "8334347b2e61c339e1118680938470172dd38cb4"
 
 def main():
   if len(sys.argv) < 3:

chapters/device_activity_tracing/LevelZero.md

@@ -85,17 +85,16 @@ Host timestamp value corresponds to `CLOCK_MONOTONIC_RAW` on Linux or `QueryPerf
 
 Note that the number of valid bits for the device timestamp returned by `zeDeviceGetGlobalTimestamps` is `timestampValidBits`, while the global kernel timastamp returned by `zeEventQueryKernelTimestamp` has `kernelTimestampValidBits` (both values are fields of `ze_device_properties_t`). And currently `kernelTimestampValidBits` is less then `timestampValidBits`, so to map kernels into CPU timeline one may need to truncate device timestamp to `kernelTimestampValidBits`:
 ```cpp
-ze_device_properties_t props{ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES, };
+ze_device_properties_t props{ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES_1_2, };
 ze_result_t status = zeDeviceGetProperties(device, &props);
 assert(status == ZE_RESULT_SUCCESS);
 
 uint64_t mask = (1ull << props.kernelTimestampValidBits) - 1ull;
 uint64_t kernel_timestamp = (device_timestamp & mask);
 ```
-To convert GPU cycles into seconds one may use `timerResolution` field from `ze_device_properties_t` structure, that represents cycles per second starting from Level Zero 1.1:
+To convert GPU cycles into seconds one may use `timerResolution` field from `ze_device_properties_t` structure, that represents cycles per second starting from Level Zero 1.2:
 ```cpp
-ze_device_properties_t props{};
-props.stype = ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES;
+ze_device_properties_t props{ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES_1_2, };
 ze_result_t status = zeDeviceGetProperties(device, &props);
 assert(status == ZE_RESULT_SUCCESS);
 

chapters/metrics_collection/MetricsDiscoveryAPI.md

@@ -50,11 +50,11 @@ md::OpenAdapterGroup_fn OpenAdapterGroup =
   (md::OpenAdapterGroup_fn)dlsym(handle, "OpenAdapterGroup");
 assert(OpenAdapterGroup != nullptr);
 
-md::IAdapterGroup_1_9* adapter_group = nullptr;
+md::IAdapterGroupLatest* adapter_group = nullptr;
 status = OpenAdapterGroup(&adapter_group);
 PTI_ASSERT(status == md::CC_OK);
 
-md::IAdapter_1_9* adapter = adapter_group->GetAdapter(0 /* device id*/);
+md::IAdapterLatest* adapter = adapter_group->GetAdapter(0 /* device id*/);
 PTI_ASSERT(adapter != nullptr);
 
 uint32_t sub_devices_count = adapter->GetParams()->SubDevicesCount;
@@ -113,20 +113,20 @@ Every group, set and seprate metric contain a list of properties that helps to d
 In addition to metrics, each set includes so called *information* items, that will be collected along with the metrics and show e.g. time when sample was collected, GPU core frequency, report identifier and others.
 ```cpp
 for (uint32_t gid = 0; gid < device->GetParams()->ConcurrentGroupsCount; ++gid) {
-  md::IConcurrentGroup_1_5* group = device->GetConcurrentGroup(gid);
+  md::IConcurrentGroupLatest* group = device->GetConcurrentGroup(gid);
   assert(group != nullptr);
   std::cout << "Metric Group " << gid << ": " <<
       group->GetParams()->SymbolName << std::endl;
 
   for (uint32_t sid = 0; sid < group->GetParams()->MetricSetsCount; ++sid) {
-    md::IMetricSet_1_5* set = group->GetMetricSet(sid);
+    md::IMetricSetLatest* set = group->GetMetricSet(sid);
     assert(set != nullptr);
     std::cout << "\tMetric Set " << sid << ": " <<
       set->GetParams()->SymbolName << " (" << set->GetParams()->ShortName <<
       ")" << std::endl;
 
     for (uint32_t mid = 0; mid < set->GetParams()->MetricsCount; ++mid) {
-      md::IMetric_1_0* metric = set->GetMetric(mid);
+      md::IMetricLatest* metric = set->GetMetric(mid);
       assert(metric != nullptr);
       std::cout << "\t\tMetric " << mid << ": " <<
         metric->GetParams()->SymbolName << " (" <<
@@ -160,7 +160,7 @@ md::TCompletionCode status = md::CC_OK;
 Next, one should set a type of collection for target metric set. MD library allows to work with multiple APIs, like OpenGL, DirectX, Metal and others, so it's needed to determine explicitely which type of API one wants to employ. In addition to exact API, it's possible to enable the
 most general collection using `API_TYPE_IOSTREAM` flag.
 ```cpp
-md::IMetricSet_1_5* set; // target metric set, see Enumeration section
+md::IMetricSetLatest* set; // target metric set, see Enumeration section
 status = set->SetApiFiltering(md::API_TYPE_IOSTREAM);
 assert(status == md::CC_OK);
 ```
@@ -171,7 +171,7 @@ As a result, both `sampling_interval` and `buffer_size` may be updated by the va
 Note, that the call may fail due to lack of access rights for the current user (try "root" on Linux), or too small sampling interval value. Refer to
 [Metrics Discovery (MD) API](https://github.com/intel/metrics-discovery) project to get more information.
 ```cpp
-md::IConcurrentGroup_1_5* group; // target metric group, see Enumeration section
+md::IConcurrentGroupLatest* group; // target metric group, see Enumeration section
 status = group->OpenIoStream(set, 0, &sampling_interval, &buffer_size);
 assert (status == md::CC_OK && buffer_size > 0);
 

samples/cl_gpu_metrics/cl_metric_collector.h

@@ -42,9 +42,9 @@ class ClMetricCollector {
       return nullptr;
     }
 
-    md::IConcurrentGroup_1_5* group =
+    md::IConcurrentGroupLatest* group =
       metric_device->FindMetricGroup(set_name);
-    md::IMetricSet_1_5* set = metric_device->FindMetricSet(set_name);
+    md::IMetricSetLatest* set = metric_device->FindMetricSet(set_name);
     if (group == nullptr || set == nullptr) {
       std::cerr << "[WARNING] Metric set is not found: " <<
         set_name << std::endl;
@@ -87,7 +87,7 @@ class ClMetricCollector {
     PTI_ASSERT(set_ != nullptr);
 
     for (uint32_t mid = 0; mid < set_->GetParams()->MetricsCount; ++mid) {
-      md::IMetric_1_0* metric = set_->GetMetric(mid);
+      md::IMetricLatest* metric = set_->GetMetric(mid);
       PTI_ASSERT(metric != nullptr);
       if (strcmp(metric->GetParams()->SymbolName, name) == 0) {
         return mid;
@@ -155,8 +155,8 @@ class ClMetricCollector {
 
  private: // Implementation Details
   ClMetricCollector(
-      MetricDevice* device, md::IConcurrentGroup_1_5* group,
-      md::IMetricSet_1_5* set)
+      MetricDevice* device, md::IConcurrentGroupLatest* group,
+      md::IMetricSetLatest* set)
       : device_(device), group_(group), set_(set) {
     PTI_ASSERT(device_ != nullptr);
     PTI_ASSERT(group_ != nullptr);
@@ -202,8 +202,8 @@ class ClMetricCollector {
     PTI_ASSERT(collector->group_ != nullptr);
     PTI_ASSERT(collector->set_ != nullptr);
 
-    md::IConcurrentGroup_1_5* group = collector->group_;
-    md::IMetricSet_1_5* set = collector->set_;
+    md::IConcurrentGroupLatest* group = collector->group_;
+    md::IMetricSetLatest* set = collector->set_;
 
     uint32_t sampling_interval = 100000; // nanoseconds
     uint32_t buffer_size = 0; // defined by MDAPI
@@ -249,8 +249,8 @@ class ClMetricCollector {
 
  private: // Data
   MetricDevice* device_ = nullptr;
-  md::IConcurrentGroup_1_5* group_ = nullptr;
-  md::IMetricSet_1_5* set_ = nullptr;
+  md::IConcurrentGroupLatest* group_ = nullptr;
+  md::IMetricSetLatest* set_ = nullptr;
 
   std::atomic<CollectorState> collector_state_{COLLECTOR_STATE_IDLE};
   std::thread* collector_thread_ = nullptr;

samples/cl_gpu_query/cl_metric_collector.h

@@ -78,9 +78,9 @@ class ClMetricCollector {
       return nullptr;
     }
 
-    md::IConcurrentGroup_1_5* group =
+    md::IConcurrentGroupLatest* group =
       metric_device->FindMetricGroup(set_name);
-    md::IMetricSet_1_5* set = metric_device->FindMetricSet(set_name);
+    md::IMetricSetLatest* set = metric_device->FindMetricSet(set_name);
     if (group == nullptr || set == nullptr) {
       std::cerr << "[WARNING] Metric set is not found: " <<
         set_name << std::endl;
@@ -204,8 +204,8 @@ class ClMetricCollector {
  private: // Implementation Details
   ClMetricCollector(
       MetricDevice* device,
-      md::IConcurrentGroup_1_5* group,
-      md::IMetricSet_1_5* set,
+      md::IConcurrentGroupLatest* group,
+      md::IMetricSetLatest* set,
       decltype(clCreatePerfCountersCommandQueueINTEL)* ptr)
           : device_(device), group_(group), set_(set),
             clCreatePerfCountersCommandQueue_(ptr) {
@@ -312,7 +312,7 @@ class ClMetricCollector {
     PTI_ASSERT(set_ != nullptr);
 
     for (uint32_t mid = 0; mid < set_->GetParams()->MetricsCount; ++mid) {
-      md::IMetric_1_0* metric = set_->GetMetric(mid);
+      md::IMetricLatest* metric = set_->GetMetric(mid);
       PTI_ASSERT(metric != nullptr);
       if (strcmp(metric->GetParams()->SymbolName, name) == 0) {
         return mid;
@@ -520,8 +520,8 @@ class ClMetricCollector {
   ClKernelInfoMap kernel_info_map_;
 
   MetricDevice* device_ = nullptr;
-  md::IConcurrentGroup_1_5* group_ = nullptr;
-  md::IMetricSet_1_5* set_ = nullptr;
+  md::IConcurrentGroupLatest* group_ = nullptr;
+  md::IMetricSetLatest* set_ = nullptr;
 
   decltype(clCreatePerfCountersCommandQueueINTEL)*
     clCreatePerfCountersCommandQueue_ = nullptr;

samples/ze_gemm/main.cc

@@ -167,7 +167,7 @@ static float RunAndCheck(ze_kernel_handle_t kernel,
   PTI_ASSERT(status == ZE_RESULT_SUCCESS);
 
   ze_device_properties_t props{};
-  props.stype = ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES;
+  props.stype = ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES_1_2;
   status = zeDeviceGetProperties(device, &props);
   PTI_ASSERT(status == ZE_RESULT_SUCCESS);
 
@@ -248,9 +248,6 @@ static void Compute(ze_device_handle_t device,
 }
 
 int main(int argc, char* argv[]) {
-  utils::SetEnv("NEOReadDebugKeys", "1");
-  utils::SetEnv("UseCyclesPerSecondTimer", "1");
-
   ze_result_t status = ZE_RESULT_SUCCESS;
   status = zeInit(ZE_INIT_FLAG_GPU_ONLY);
   PTI_ASSERT(status == ZE_RESULT_SUCCESS);

samples/ze_hot_kernels/tool.cc

@@ -40,8 +40,6 @@ __declspec(dllexport)
 #endif
 void SetToolEnv() {
   utils::SetEnv("ZE_ENABLE_TRACING_LAYER", "1");
-  utils::SetEnv("NEOReadDebugKeys", "1");
-  utils::SetEnv("UseCyclesPerSecondTimer", "1");
 }
 
 // Internal Tool Functionality ////////////////////////////////////////////////

samples/ze_info/main.cc

@@ -173,8 +173,12 @@ void PrintDeviceInfo() {
       "Number of devices " << device_list.size() << std::endl;
 
     for (size_t j = 0; j < device_list.size(); ++j) {
-      ze_device_properties_t device_props{
-          ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES, };
+      ze_structure_type_t stype =
+        version >= ZE_API_VERSION_1_2 ?
+        ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES_1_2 :
+        ZE_STRUCTURE_TYPE_DEVICE_PROPERTIES;
+
+      ze_device_properties_t device_props{stype, };
       status = zeDeviceGetProperties(device_list[j], &device_props);
       PTI_ASSERT(status == ZE_RESULT_SUCCESS);
 
@@ -267,7 +271,7 @@ void PrintDeviceInfo() {
       std::cout << std::setw(TEXT_WIDTH) << std::left <<
         std::string() + TAB + "Timer Resolution " <<
         device_props.timerResolution;
-      if (version >= ZE_API_VERSION_1_0) {
+      if (version < ZE_API_VERSION_1_2) {
         std::cout << "ns";
       } else {
         std::cout << "clks";
@@ -373,9 +377,6 @@ int main(int argc, char *argv[]) {
     }
   }
 
-  utils::SetEnv("NEOReadDebugKeys", "1");
-  utils::SetEnv("UseCyclesPerSecondTimer", "1");
-
   ze_result_t status = ZE_RESULT_SUCCESS;
   status = zeInit(ZE_INIT_FLAG_GPU_ONLY);
   PTI_ASSERT(status == ZE_RESULT_SUCCESS);

samples/ze_metric_streamer/tool.cc

@@ -72,8 +72,6 @@ __declspec(dllexport)
 void SetToolEnv() {
   utils::SetEnv("ZE_ENABLE_TRACING_LAYER", "1");
   utils::SetEnv("ZET_ENABLE_METRICS", "1");
-  utils::SetEnv("NEOReadDebugKeys", "1");
-  utils::SetEnv("UseCyclesPerSecondTimer", "1");
 }
 
 // Internal Tool Functionality ////////////////////////////////////////////////

tests/run.py

@@ -34,7 +34,7 @@
 tools = [["gpuinfo", "-l", "-i", "-m"],
          ["sysmon", "-p", "-l", "-d"],
          ["onetrace",
-          "-c", "-h", "-d", "-v", "-t",
+          "-c", "-h", "-d", "-v", "-t", "-s",
           "--kernels-per-tile",
           "--chrome-call-logging",
           "--chrome-device-timeline",
@@ -43,15 +43,15 @@
           "--conditional-collection",
           "cl", "ze", "omp"],
          ["cl_tracer",
-          "-c", "-h", "-d", "-v", "-t",
+          "-c", "-h", "-d", "-v", "-t", "-s",
           "--chrome-call-logging",
           "--chrome-device-timeline",
           "--chrome-kernel-timeline",
           "--chrome-device-stages",
           "--conditional-collection",
           "gpu", "dpc", "omp"],
          ["ze_tracer",
-          "-c", "-h", "-d", "-v", "-t",
+          "-c", "-h", "-d", "-v", "-t", "-s",
           "--kernels-per-tile",
           "--chrome-call-logging",
           "--chrome-device-timeline",

tests/tools/cl_tracer.py

@@ -51,7 +51,7 @@ def run(path, option):
     if option == "gpu":
       command = [file_name_prefix + "cl_tracer" + file_extention,\
         "-h", "-d", "-t", app_file, "gpu", "1024", "1"]
-    elif option == "--conditional-collection":
+    elif option == "-v" or option == "--conditional-collection":
       command = [file_name_prefix + "cl_tracer" + file_extention,\
         "-d", option, app_file, "cpu", "1024", "1"]
     else:
@@ -98,6 +98,10 @@ def main(option):
     option = "-d"
   if len(sys.argv) > 1 and sys.argv[1] == "-t":
     option = "-t"
+  if len(sys.argv) > 1 and sys.argv[1] == "-s":
+    option = "-s"
+  if len(sys.argv) > 1 and sys.argv[1] == "-v":
+    option = "-v"
   if len(sys.argv) > 1 and sys.argv[1] == "--chrome-call-logging":
     option = "--chrome-call-logging"
   if len(sys.argv) > 1 and sys.argv[1] == "--chrome-device-timeline":

tests/tools/onetrace.py

@@ -35,7 +35,7 @@ def run(path, option):
     app_folder = utils.get_sample_executable_path("omp_gemm")
     app_file = os.path.join(app_folder, "omp_gemm")
     command = ["./onetrace", "-h", "-d", "-t", app_file, "gpu", "1024", "1"]
-  elif option == "--kernels-per-tile" or option == "--conditional-collection":
+  elif option == "-v" or option == "--kernels-per-tile" or option == "--conditional-collection":
     app_folder = utils.get_sample_executable_path("dpc_gemm")
     app_file = os.path.join(app_folder, "dpc_gemm")
     command = ["./onetrace", "-d", option, app_file, "gpu", "1024", "1"]
@@ -84,6 +84,10 @@ def main(option):
     option = "-d"
   if len(sys.argv) > 1 and sys.argv[1] == "-t":
     option = "-t"
+  if len(sys.argv) > 1 and sys.argv[1] == "-s":
+    option = "-s"
+  if len(sys.argv) > 1 and sys.argv[1] == "-v":
+    option = "-v"
   if len(sys.argv) > 1 and sys.argv[1] == "--kernels-per-tile":
     option = "--kernels-per-tile"
   if len(sys.argv) > 1 and sys.argv[1] == "--chrome-call-logging":

tests/tools/ze_tracer.py

@@ -30,7 +30,7 @@ def run(path, option):
     app_folder = utils.get_sample_executable_path("omp_gemm")
     app_file = os.path.join(app_folder, "omp_gemm")
     command = ["./ze_tracer", "-h", "-d", "-t", app_file, "gpu", "1024", "1"]
-  elif option == "--kernels-per-tile" or option == "--conditional-collection":
+  elif option == "-v" or option == "--kernels-per-tile" or option == "--conditional-collection":
     app_folder = utils.get_sample_executable_path("ze_gemm")
     app_file = os.path.join(app_folder, "ze_gemm")
     command = ["./ze_tracer", "-d", option, app_file, "1024", "1"]
@@ -77,6 +77,10 @@ def main(option):
     option = "-d"
   if len(sys.argv) > 1 and sys.argv[1] == "-t":
     option = "-t"
+  if len(sys.argv) > 1 and sys.argv[1] == "-s":
+    option = "-s"
+  if len(sys.argv) > 1 and sys.argv[1] == "-v":
+    option = "-v"
   if len(sys.argv) > 1 and sys.argv[1] == "--kernels-per-tile":
     option = "--kernels-per-tile"
   if len(sys.argv) > 1 and sys.argv[1] == "--chrome-call-logging":