sysman-metrics updated

argonne-lcf · Sep 25, 2023 · 9c15987 · 9c15987
1 parent 8b5c3b2
commit 9c15987
Show file tree

Hide file tree

Showing 7 changed files with 68 additions and 64 deletions.
diff --git a/utils/xprof_utils.cpp b/utils/xprof_utils.cpp
@@ -119,7 +119,7 @@ bt_message* create_frequency_message(const char* hostname, const process_id_t pr
 }
 
 bt_message* create_computeEU_message(const char* hostname, const process_id_t process_id, const thread_id_t thread_id,
-                                     const uintptr_t hDevice, const uint32_t subDevice, const uint64_t activeTime, const uint64_t ts,
+                                     const uintptr_t hDevice, const uint32_t subDevice, const float activeTime, const uint64_t ts,
                                      bt_event_class *event_class, bt_self_message_iterator *message_iterator, bt_stream *stream, backend_t backend) {
 
      /* Message creation */
@@ -161,13 +161,13 @@ bt_message* create_computeEU_message(const char* hostname, const process_id_t pr
 
      //activeTime
      bt_field *activeTime_field = bt_field_structure_borrow_member_field_by_index(payload_field,2);
-     bt_field_integer_unsigned_set_value(activeTime_field, activeTime);
+     bt_field_real_single_precision_set_value(activeTime_field, activeTime);
 
      return message;
 }
 
 bt_message* create_copyEU_message(const char* hostname, const process_id_t process_id, const thread_id_t thread_id,
-                                     const uintptr_t hDevice, const uint32_t subDevice, const uint64_t activeTime, const uint64_t ts,
+                                     const uintptr_t hDevice, const uint32_t subDevice, const float activeTime, const uint64_t ts,
                                      bt_event_class *event_class, bt_self_message_iterator *message_iterator, bt_stream *stream, backend_t backend) {
 
      /* Message creation */
@@ -209,7 +209,7 @@ bt_message* create_copyEU_message(const char* hostname, const process_id_t proce
 
      //activeTime
      bt_field *activeTime_field = bt_field_structure_borrow_member_field_by_index(payload_field,2);
-     bt_field_integer_unsigned_set_value(activeTime_field, activeTime);
+     bt_field_real_single_precision_set_value(activeTime_field, activeTime);
 
      return message;
 }

diff --git a/utils/xprof_utils.hpp b/utils/xprof_utils.hpp
@@ -123,11 +123,11 @@ bt_message* create_frequency_message(const char* hostname, const process_id_t pr
                                      bt_event_class *event_class, bt_self_message_iterator *message_iterator, bt_stream *stream, backend_t backend = BACKEND_UNKNOWN);
 
 bt_message* create_computeEU_message(const char* hostname, const process_id_t proprocess_id, const thread_id_t thread_id,
-                                     const uintptr_t hDevice, const uint32_t subDevice, const uint64_t activeTime, const uint64_t ts,
+                                     const uintptr_t hDevice, const uint32_t subDevice, const float activeTime, const uint64_t ts,
                                      bt_event_class *event_class, bt_self_message_iterator *message_iterator, bt_stream *stream, backend_t backend = BACKEND_UNKNOWN);
 
 bt_message* create_copyEU_message(const char* hostname, const process_id_t proprocess_id, const thread_id_t thread_id,
-                                  const uintptr_t hDevice, const uint32_t subDevice, const uint64_t activeTime, const uint64_t ts,
+                                  const uintptr_t hDevice, const uint32_t subDevice, const float activeTime, const uint64_t ts,
                                   bt_event_class *event_class, bt_self_message_iterator *message_iterator, bt_stream *stream, backend_t backend = BACKEND_UNKNOWN);
 
 bt_message *create_host_message(const char *hostname, const process_id_t, const thread_id_t,

diff --git a/xprof/btx_interval_model.yaml b/xprof/btx_interval_model.yaml
@@ -143,9 +143,8 @@
                 :cast_type: uint32_t
             - :name: activeTime
               :field_class:
-                :type: integer_unsigned
-                :field_value_range: 64
-                :cast_type: uint64_t
+                :type: single
+                :cast_type: float
       - :name: lttng:copyEU
         :payload_field_class:
           :type: structure
@@ -162,6 +161,5 @@
                 :cast_type: uint32_t
             - :name: activeTime
               :field_class:
-                :type: integer_unsigned
-                :field_value_range: 64
-                :cast_type: uint64_t
+                :type: single
+                :cast_type: float
diff --git a/xprof/btx_timeline.cpp b/xprof/btx_timeline.cpp
@@ -149,7 +149,7 @@ static void add_event_power(timeline_dispatch_t *dispatch, std::string hostname,
 
 static void add_event_computeEU(timeline_dispatch_t *dispatch, std::string hostname,
                                 uint64_t process_id, uint64_t thread_id, uintptr_t did,
-                                uint32_t subDevice, uint64_t timestamp, uint64_t activeTime) {
+                                uint32_t subDevice, uint64_t timestamp, float activeTime) {
   perfetto_uuid_t track_uuid = get_computeEU_track_uuuid(dispatch, hostname, process_id, did, subDevice);
   auto *packet = dispatch->trace.add_packet();
   packet->set_trusted_packet_sequence_id(10000);
@@ -158,12 +158,12 @@ static void add_event_computeEU(timeline_dispatch_t *dispatch, std::string hostn
   track_event->set_type(perfetto_pruned::TrackEvent::TYPE_COUNTER);
   track_event->set_track_uuid(track_uuid);
   track_event->set_name("computeEngine Usage");
-  track_event->set_counter_value(activeTime);
+  track_event->set_double_counter_value(activeTime);
 }
 
 static void add_event_copyEU(timeline_dispatch_t *dispatch, std::string hostname,
                              uint64_t process_id, uint64_t thread_id, uintptr_t did,
-                             uint32_t subDevice, uint64_t timestamp, uint64_t activeTime) {
+                             uint32_t subDevice, uint64_t timestamp, float activeTime) {
   perfetto_uuid_t track_uuid = get_copyEU_track_uuuid(dispatch, hostname, process_id, did, subDevice);
   auto *packet = dispatch->trace.add_packet();
   packet->set_trusted_packet_sequence_id(10000);
@@ -172,7 +172,7 @@ static void add_event_copyEU(timeline_dispatch_t *dispatch, std::string hostname
   track_event->set_type(perfetto_pruned::TrackEvent::TYPE_COUNTER);
   track_event->set_track_uuid(track_uuid);
   track_event->set_name("copyEngine Usage");
-  track_event->set_counter_value(activeTime);
+  track_event->set_double_counter_value(activeTime);
 }
 
 static void add_event_begin(timeline_dispatch_t *dispatch, perfetto_uuid_t uuid, timestamp_t begin,
@@ -394,14 +394,14 @@ static void power_usr_callback(void *btx_handle, void *usr_data, const char *hos
 
 static void computeEU_usr_callback(void *btx_handle, void *usr_data, const char *hostname,
                                    int64_t vpid, uint64_t vtid, int64_t ts, int64_t backend,
-                                   uint64_t did, uint32_t subDevice, uint64_t activeTime) {
+                                   uint64_t did, uint32_t subDevice, float activeTime) {
   auto *dispatch = static_cast<timeline_dispatch_t *>(usr_data);
   add_event_computeEU(dispatch, hostname, vpid, vtid, did, subDevice, ts, activeTime);
 }
 
 static void copyEU_usr_callback(void *btx_handle, void *usr_data, const char *hostname,
                                 int64_t vpid, uint64_t vtid, int64_t ts, int64_t backend,
-                                uint64_t did, uint32_t subDevice, uint64_t activeTime) {
+                                uint64_t did, uint32_t subDevice, float activeTime) {
   auto *dispatch = static_cast<timeline_dispatch_t *>(usr_data);
   add_event_copyEU(dispatch, hostname, vpid, vtid, did, subDevice, ts, activeTime);
 }

diff --git a/ze/tracer_ze_helpers.include.c b/ze/tracer_ze_helpers.include.c
@@ -768,7 +768,6 @@ static uint32_t _sampling_subDeviceCount;
 static uint32_t* _sampling_freqDomainCounts;
 static uint32_t* _sampling_powerDomainCounts;
 static uint32_t* _sampling_engineCounts;
-static uint32_t* _sampling_engineCounts;
 
 typedef struct {
     uint64_t timestamp;
@@ -797,7 +796,7 @@ int initializeHandles() {
 #endif
 
   // Query driver
-  uint32_t driverCount;
+  uint32_t driverCount = 0;
   res = ZE_DRIVER_GET_PTR(&driverCount, NULL);
   if (res != ZE_RESULT_SUCCESS) {
     _ZE_ERROR_MSG("1st ZE_DRIVER_GET_PTR", res);
@@ -851,7 +850,6 @@ int initializeHandles() {
       printf("zesDeviceEnumFrequencyDomains failed for device %d: %d\n", i, res);
       return(-1);
     }
-
     // Get power domains for each device
     res = zesDeviceEnumPowerDomains(_sampling_hDevices[i], &_sampling_powerDomainCounts[i], NULL);
     if (res != ZE_RESULT_SUCCESS) {
@@ -871,6 +869,7 @@ int initializeHandles() {
       printf("No engine groups found\n");
       return(-1);
     }
+
     _sampling_engineHandles[i] = (zes_engine_handle_t*)malloc(_sampling_engineCounts[i] * sizeof(zes_engine_handle_t));
     res = zesDeviceEnumEngineGroups(_sampling_hDevices[i], &_sampling_engineCounts[i], _sampling_engineHandles[i]);
     if (res != ZE_RESULT_SUCCESS) {
@@ -889,7 +888,6 @@ void readFrequency(uint32_t deviceIdx, uint32_t domainIdx, uint32_t *frequency)
   *frequency=0;
   zes_freq_state_t freqState;
   if (zesFrequencyGetState(_sampling_hFrequencies[deviceIdx][domainIdx], &freqState) == ZE_RESULT_SUCCESS) {
-    // printf("Device %d - Frequency Domain %d: Current frequency: %lf MHz\n", deviceIdx, domainIdx, freqState.actual);
     *frequency = freqState.actual;
   }
 }
@@ -900,7 +898,6 @@ void readEnergy(uint32_t deviceIdx, uint32_t domainIdx, uint64_t *ts_us, uint64_
   *energy_uj = 0;
   zes_power_energy_counter_t energyCounter;
   if (zesPowerGetEnergyCounter(_sampling_hPowers[deviceIdx][domainIdx], &energyCounter) == ZE_RESULT_SUCCESS) {
-    // printf("Device %d - Power Domain %d: Total energy consumption: %lu Joules\n", deviceIdx, domainIdx, energyCounter.energy);
     *ts_us = energyCounter.timestamp;
     *energy_uj = energyCounter.energy;
   }
@@ -953,16 +950,16 @@ static void thapi_sampling_energy() {
   for (uint32_t i = 0; i < _sampling_deviceCount; i++) {
     for (uint32_t j = 0; j < _sampling_freqDomainCounts[i]; j++) {
       readFrequency(i, j, &frequency);
-      do_tracepoint(lttng_ust_ze_sampling, gpu_frequency, (ze_device_handle_t)_sampling_hDevices[i], j, ts_us, frequency);
+      tracepoint(lttng_ust_ze_sampling, gpu_frequency, (ze_device_handle_t)_sampling_hDevices[i], j, ts_us, frequency);
     }
     for (uint32_t j = 0; j < _sampling_powerDomainCounts[i]; j++) {
       readEnergy(i, j, &ts_us, &energy_uj);
-      do_tracepoint(lttng_ust_ze_sampling, gpu_energy, (ze_device_handle_t)_sampling_hDevices[i], j, (uint64_t)energy_uj, ts_us);
+      tracepoint(lttng_ust_ze_sampling, gpu_energy, (ze_device_handle_t)_sampling_hDevices[i], j, (uint64_t)energy_uj, ts_us);
     }
     readPerformance(i, computeE, copyE);
     for (uint32_t k=0; k<_sampling_subDeviceCount; k++){
-      do_tracepoint(lttng_ust_ze_sampling, computeEngine, (ze_device_handle_t)_sampling_hDevices[i], k, computeE[k].computeActive, computeE[k].timestamp);
-      do_tracepoint(lttng_ust_ze_sampling, copyEngine, (ze_device_handle_t)_sampling_hDevices[i], k, copyE[k].copyActive, copyE[k].timestamp);
+      tracepoint(lttng_ust_ze_sampling, computeEngine, (ze_device_handle_t)_sampling_hDevices[i], k, computeE[k].computeActive, computeE[k].timestamp);
+      tracepoint(lttng_ust_ze_sampling, copyEngine, (ze_device_handle_t)_sampling_hDevices[i], k, copyE[k].copyActive, copyE[k].timestamp);
     }
   }
 }