Finalize the rocprofiler fields

Change-Id: I4ed1c4309f21bdcc7281d911663036caf5947182
Signed-off-by: Galantsev, Dmitrii <dmitrii.galantsev@amd.com>

include/rdc/rdc.h

@@ -150,7 +150,7 @@ typedef enum {
   RDC_FI_GPU_COUNT = 1,  //!< GPU count in the system
   RDC_FI_DEV_NAME,       //!< Name of the device
 
-  /*
+  /**
    * @brief Frequency related fields
    */
   RDC_FI_GPU_CLOCK = 100,  //!< The current clock for the GPU
@@ -163,7 +163,7 @@ typedef enum {
   RDC_FI_GPU_TEMP,           //!< Current temperature for the device
   RDC_FI_POWER_USAGE = 300,  //!< Power usage for the device
 
-  /*
+  /**
    * @brief PCIe related fields
    */
   RDC_FI_PCIE_TX = 400,  //!< PCIe Tx utilization information
@@ -172,7 +172,7 @@ typedef enum {
   // The RDC_FI_PCIE_BANDWIDTH should be used
   RDC_FI_PCIE_BANDWIDTH,  //!< PCIe bandwidth in GB/sec
 
-  /*
+  /**
    * @brief GPU usage related fields
    */
   RDC_FI_GPU_UTIL = 500,    //!< GPU Utilization
@@ -250,21 +250,21 @@ typedef enum {
   /**
    * @brief ROC-profiler related fields
    */
-  RDC_FI_PROF_CU_UTILIZATION = 800,
-  RDC_FI_PROF_CU_OCCUPANCY,
-  RDC_FI_PROF_FLOPS_16,
-  RDC_FI_PROF_FLOPS_32,
-  RDC_FI_PROF_FLOPS_64,
+  RDC_FI_PROF_CU_OCCUPANCY = 800,
+  RDC_FI_PROF_MEAN_OCCUPANCY_PER_CU,
+  RDC_FI_PROF_MEAN_OCCUPANCY_PER_ACTIVE_CU,
   RDC_FI_PROF_ACTIVE_CYCLES,
   RDC_FI_PROF_ACTIVE_WAVES,
   RDC_FI_PROF_ELAPSED_CYCLES,
-  RDC_FI_PROF_FETCH_SIZE,
-  RDC_FI_PROF_WRITE_SIZE,
-  RDC_FI_PROF_GRBM_COUNT,
-  RDC_FI_PROF_SQ_WAVES,
-  RDC_FI_PROF_TA_BUSY_AVR,
 
-  /*
+  // metrics below are divided by time passed
+  RDC_FI_PROF_EVAL_MEM_R_BW,
+  RDC_FI_PROF_EVAL_MEM_W_BW,
+  RDC_FI_PROF_EVAL_FLOPS_16,
+  RDC_FI_PROF_EVAL_FLOPS_32,
+  RDC_FI_PROF_EVAL_FLOPS_64,
+
+  /**
    * @brief Raw XGMI counter events
    */
   RDC_EVNT_XGMI_0_NOP_TX = 1000,  //!< NOPs sent to neighbor 0

include/rdc_modules/rdc_rocp/RdcRocpBase.h

@@ -24,8 +24,10 @@ THE SOFTWARE.
 #define RDC_MODULES_RDC_ROCP_RDCROCPBASE_H_
 #include <rocprofiler/rocprofiler.h>
 
+#include <chrono>
 #include <cstdint>
 #include <map>
+#include <unordered_set>
 #include <vector>
 
 #include "rdc/rdc.h"
@@ -69,12 +71,17 @@ class RdcRocpBase {
   std::map<const char*, double> metric_to_value = {};
   // array of features for each device
   std::map<uint32_t, rocprofiler_feature_t> feature;
-  // rocprofiler_feature_t features[dev_count][features_count] = {};
-  void read_feature(rocprofiler_t* context, const unsigned feature_count);
+  void read_feature(rocprofiler_t* context, const unsigned feature_count, uint32_t gpu_index);
   int run_profiler(uint32_t gpu_index, rdc_field_t field);
   std::vector<hsa_queue_t*> queues;
   hsa_agent_arr_t agent_arr = {};
   std::map<rdc_field_t, const char*> field_to_metric = {};
+  // these fields must be divided by time passed
+  std::unordered_set<rdc_field_t> eval_fields = {
+      RDC_FI_PROF_EVAL_MEM_R_BW, RDC_FI_PROF_EVAL_MEM_W_BW, RDC_FI_PROF_EVAL_FLOPS_16,
+      RDC_FI_PROF_EVAL_FLOPS_32, RDC_FI_PROF_EVAL_FLOPS_64,
+  };
+  std::chrono::time_point<std::chrono::system_clock, std::chrono::nanoseconds> start_time;
 
   /**
    * @brief Convert from rocmtools status into RDC status