CPU - Init amdsmi with CPUs enabled

CMakeLists.txt

@@ -115,6 +115,11 @@ option(BUILD_EXAMPLES "Build examples" OFF)
 # Enable shared libraries for gtest
 option(BUILD_SHARED_LIBS "Build shared library (.so) or not." ON)
 
+option(BUILD_ESMI "Enable AMDSMI ESMI Library" ON)
+if(BUILD_ESMI)
+    add_definitions("-DENABLE_ESMI_LIB=1")
+endif()
+
 # Enable address sanitizer
 set(ADDRESS_SANITIZER_DEFAULT OFF)
 if(DEFINED ENV{ADDRESS_SANITIZER})

common/rdc_field.data

@@ -238,3 +238,56 @@ FLD_DESC_ENT(RDC_HEALTH_RETIRED_PAGE_LIMIT,       "Retired page limit",
 FLD_DESC_ENT(RDC_HEALTH_EEPROM_CONFIG_VALID,      "Verify checksum of EEPROM",                  "EEPROM_CONFIG_VALID",      true)
 FLD_DESC_ENT(RDC_HEALTH_POWER_THROTTLE_TIME,      "Power throttle status counter",              "POWER_THROTTLE_TIME",      true)
 FLD_DESC_ENT(RDC_HEALTH_THERMAL_THROTTLE_TIME,    "Total time(ms) in thermal throttle status",  "THERMAL_THROTTLE_TIME",    true)
+
+// CPU-related fields description
+FLD_DESC_ENT(RDC_FI_CPU_COUNT,              "The number of CPU cores",                         "CPU_COUNT",              false)
+FLD_DESC_ENT(RDC_FI_CPU_UTIL_TOTAL,         "CPU total percentage of time in use",             "CPU_UTIL_TOTAL",         false)
+FLD_DESC_ENT(RDC_FI_CPU_UTIL_USER,          "Percent of time in use by the user",              "CPU_UTIL_USER",          false)
+FLD_DESC_ENT(RDC_FI_CPU_UTIL_NICE,          "Percent of time in use by low priority programs", "CPU_UTIL_NICE",          false)
+FLD_DESC_ENT(RDC_FI_CPU_UTIL_SYS,           "Percent of time in use by the system",            "CPU_UTIL_SYS",           false)
+FLD_DESC_ENT(RDC_FI_CPU_UTIL_IRQ,           "Percent of time in use by interrupts",            "CPU_UTIL_IRQ",           false)
+FLD_DESC_ENT(RDC_FI_CPU_TEMP_CURRENT,       "Temperature (Celsius)",                           "CPU_TEMP_CURRENT",       false)
+FLD_DESC_ENT(RDC_FI_CPU_CLOCK_CURRENT,      "Clock speed (KHz)",                               "CPU_CLOCK_CURRENT",      false)
+FLD_DESC_ENT(RDC_FI_CPU_POWER_UTIL_CURRENT, "Power usage (watts)",                             "CPU_POWER_UTIL_CURRENT", false)
+FLD_DESC_ENT(RDC_FI_CPU_POWER_LIMIT,        "Power limit (watts)",                             "CPU_POWER_LIMIT",        false)
+FLD_DESC_ENT(RDC_FI_CPU_VENDOR,             "Name of the vendor",                              "CPU_VENDOR",             false)
+FLD_DESC_ENT(RDC_FI_CPU_MODEL,              "Name of the model",                               "CPU_MODEL",              false)
+
+// AI generated:
+// Field Definitions
+FLD_DESC_ENT(RDC_FI_CPU_CORE_ENERGY,              "CPU core energy consumption (microjoules)",          "CPU_CORE_ENERGY",              false)
+FLD_DESC_ENT(RDC_FI_CPU_SOCKET_ENERGY,            "CPU socket energy consumption (microjoules)",        "CPU_SOCKET_ENERGY",            false)
+FLD_DESC_ENT(RDC_FI_CPU_THREADS_PER_CORE,         "Number of threads per CPU core",                     "CPU_THREADS_PER_CORE",         false)
+FLD_DESC_ENT(RDC_FI_CPU_HSMP_DRIVER_VERSION,      "HSMP driver version",                                "CPU_HSMP_DRIVER_VERSION",      false)
+FLD_DESC_ENT(RDC_FI_CPU_SMU_FW_VERSION,           "SMU firmware version",                               "CPU_SMU_FW_VERSION",           false)
+FLD_DESC_ENT(RDC_FI_CPU_HSMP_PROTO_VERSION,       "HSMP protocol version",                              "CPU_HSMP_PROTO_VERSION",       false)
+FLD_DESC_ENT(RDC_FI_CPU_PROCHOT_STATUS,           "CPU PROCHOT status",                                 "CPU_PROCHOT_STATUS",           false)
+FLD_DESC_ENT(RDC_FI_CPU_FCLK_FREQUENCY,           "CPU fabric clock frequency (MHz)",                   "CPU_FCLK_FREQUENCY",           false)
+FLD_DESC_ENT(RDC_FI_CPU_MCLK_FREQUENCY,           "CPU memory clock frequency (MHz)",                   "CPU_MCLK_FREQUENCY",           false)
+FLD_DESC_ENT(RDC_FI_CPU_CCLK_LIMIT,               "CPU core clock limit (MHz)",                         "CPU_CCLK_LIMIT",               false)
+FLD_DESC_ENT(RDC_FI_CPU_SOCKET_ACTIVE_FREQ_LIMIT, "CPU socket active frequency limit (MHz)",            "CPU_SOCKET_ACTIVE_FREQ_LIMIT", false)
+FLD_DESC_ENT(RDC_FI_CPU_SOCKET_FREQ_LIMIT_SRC,    "CPU socket frequency limit source type",             "CPU_SOCKET_FREQ_LIMIT_SRC",    false)
+FLD_DESC_ENT(RDC_FI_CPU_SOCKET_FREQ_RANGE_MAX,    "CPU socket maximum frequency range (MHz)",           "CPU_SOCKET_FREQ_RANGE_MAX",    false)
+FLD_DESC_ENT(RDC_FI_CPU_SOCKET_FREQ_RANGE_MIN,    "CPU socket minimum frequency range (MHz)",           "CPU_SOCKET_FREQ_RANGE_MIN",    false)
+FLD_DESC_ENT(RDC_FI_CPU_CORE_FREQ_LIMIT,          "CPU core current frequency limit (MHz)",             "CPU_CORE_FREQ_LIMIT",          false)
+FLD_DESC_ENT(RDC_FI_CPU_CORE_BOOST_LIMIT,         "CPU core boost limit (MHz)",                         "CPU_CORE_BOOST_LIMIT",         false)
+FLD_DESC_ENT(RDC_FI_CPU_SOCKET_C0_RESIDENCY,      "CPU socket C0 residency percentage",                 "CPU_SOCKET_C0_RESIDENCY",      false)
+FLD_DESC_ENT(RDC_FI_CPU_DDR_BW_MAX_BW,            "CPU DDR maximum bandwidth (MB/s)",                   "CPU_DDR_BW_MAX_BW",            false)
+FLD_DESC_ENT(RDC_FI_CPU_DDR_BW_UTILIZED_BW,       "CPU DDR utilized bandwidth (MB/s)",                  "CPU_DDR_BW_UTILIZED_BW",       false)
+FLD_DESC_ENT(RDC_FI_CPU_DDR_BW_UTILIZED_PCT,      "CPU DDR utilized bandwidth percentage",              "CPU_DDR_BW_UTILIZED_PCT",      false)
+FLD_DESC_ENT(RDC_FI_CPU_SOCKET_TEMPERATURE,       "CPU socket temperature (millidegrees Celsius)",      "CPU_SOCKET_TEMPERATURE",       false)
+FLD_DESC_ENT(RDC_FI_CPU_DIMM_TEMP_RANGE,          "CPU DIMM temperature range",                         "CPU_DIMM_TEMP_RANGE",          false)
+FLD_DESC_ENT(RDC_FI_CPU_DIMM_REFRESH_RATE,        "CPU DIMM refresh rate",                              "CPU_DIMM_REFRESH_RATE",        false)
+FLD_DESC_ENT(RDC_FI_CPU_DIMM_POWER_CONSUMPTION,   "CPU DIMM power consumption (milliwatts)",            "CPU_DIMM_POWER_CONSUMPTION",   false)
+FLD_DESC_ENT(RDC_FI_CPU_DIMM_THERMAL_SENSOR,      "CPU DIMM thermal sensor temperature (millidegrees)", "CPU_DIMM_THERMAL_SENSOR",      false)
+FLD_DESC_ENT(RDC_FI_CPU_SOCKET_LCLK_DPM_LEVEL,    "CPU socket LCLK DPM level",                          "CPU_SOCKET_LCLK_DPM_LEVEL",    false)
+FLD_DESC_ENT(RDC_FI_CPU_IO_BANDWIDTH,             "CPU I/O bandwidth (MB/s)",                           "CPU_IO_BANDWIDTH",             false)
+FLD_DESC_ENT(RDC_FI_CPU_XGMI_BANDWIDTH,           "CPU XGMI bandwidth (MB/s)",                          "CPU_XGMI_BANDWIDTH",           false)
+FLD_DESC_ENT(RDC_FI_CPU_HSMP_METRICS_VERSION,     "HSMP metrics table version",                         "CPU_HSMP_METRICS_VERSION",     false)
+FLD_DESC_ENT(RDC_FI_CPU_HSMP_METRICS_TABLE,       "HSMP metrics table data",                            "CPU_HSMP_METRICS_TABLE",       false)
+FLD_DESC_ENT(RDC_FI_CPU_FIRST_ONLINE_CORE,        "First online core on CPU socket",                    "CPU_FIRST_ONLINE_CORE",        false)
+FLD_DESC_ENT(RDC_FI_CPU_FAMILY,                   "CPU family identifier",                              "CPU_FAMILY",                   false)
+FLD_DESC_ENT(RDC_FI_CPU_MODEL_ID,                 "CPU model identifier",                               "CPU_MODEL_ID",                 false)
+FLD_DESC_ENT(RDC_FI_CPU_CORES_PER_SOCKET,         "Number of CPU cores per socket",                     "CPU_CORES_PER_SOCKET",         false)
+FLD_DESC_ENT(RDC_FI_CPU_SOCKET_COUNT,             "Number of CPU sockets",                              "CPU_SOCKET_COUNT",             false)
+

common/rdc_fields_supported.cc

@@ -38,7 +38,7 @@ static const fld_id2name_map_t field_id_to_descript = {
 
 #define FLD_DESC_ENT(ID, DESC, LABEL, DISPLAY) {#ID, (ID)},
 static fld_name2id_map_t field_name_to_id = {
-#include "common/rdc_field.data"  // NOLINT
+#include "common/rdc_field.data"
 };
 #undef FLD_DESC_ENT
 

include/rdc/rdc.h

@@ -424,6 +424,61 @@ typedef enum {
   RDC_HEALTH_EEPROM_CONFIG_VALID,    //!< Reads the EEPROM and verifies the checksums
   RDC_HEALTH_POWER_THROTTLE_TIME,    //!< Power throttle status counter
   RDC_HEALTH_THERMAL_THROTTLE_TIME,  //!< Total time in thermal throttle status (microseconds)
+  /**
+   * @brief RDC CPU related fields
+   */
+  RDC_FI_CPU_COUNT = 10000,       //!< CPU count
+  RDC_FI_CPU_UTIL_TOTAL,          //!< CPU total percentage of time in use
+  RDC_FI_CPU_UTIL_USER,           //!< Percent of time in use by the user
+  RDC_FI_CPU_UTIL_NICE,           //!< Percent of time in use by low priority programs
+  RDC_FI_CPU_UTIL_SYS,            //!< Percent of time in use by the system
+  RDC_FI_CPU_UTIL_IRQ,            //!< Percent of time in use by interrupts
+  RDC_FI_CPU_TEMP_CURRENT,        //!< Temperature (Celsius)
+  RDC_FI_CPU_CLOCK_CURRENT,       //!< Clock speed (KHz)
+  RDC_FI_CPU_POWER_UTIL_CURRENT,  //!< Power usage (watts)
+  RDC_FI_CPU_POWER_LIMIT,         //!< Power limit (watts)
+  RDC_FI_CPU_VENDOR,              //!< Name of the vendor
+  RDC_FI_CPU_MODEL,               //!< Name of the model
+
+  /** TODO: Clean up duplicates
+   * @brief Below are AI generated from functions
+   */
+  RDC_FI_CPU_CORE_ENERGY,               //!< CPU core energy consumption (microjoules)
+  RDC_FI_CPU_SOCKET_ENERGY,             //!< CPU socket energy consumption (microjoules)
+  RDC_FI_CPU_THREADS_PER_CORE,          //!< Number of threads per CPU core
+  RDC_FI_CPU_HSMP_DRIVER_VERSION,       //!< HSMP driver version
+  RDC_FI_CPU_SMU_FW_VERSION,            //!< SMU firmware version
+  RDC_FI_CPU_HSMP_PROTO_VERSION,        //!< HSMP protocol version
+  RDC_FI_CPU_PROCHOT_STATUS,            //!< CPU PROCHOT status
+  RDC_FI_CPU_FCLK_FREQUENCY,            //!< CPU fabric clock frequency (MHz)
+  RDC_FI_CPU_MCLK_FREQUENCY,            //!< CPU memory clock frequency (MHz)
+  RDC_FI_CPU_CCLK_LIMIT,                //!< CPU core clock limit (MHz)
+  RDC_FI_CPU_SOCKET_ACTIVE_FREQ_LIMIT,  //!< CPU socket active frequency limit (MHz)
+  RDC_FI_CPU_SOCKET_FREQ_LIMIT_SRC,     //!< CPU socket frequency limit source type
+  RDC_FI_CPU_SOCKET_FREQ_RANGE_MAX,     //!< CPU socket maximum frequency range (MHz)
+  RDC_FI_CPU_SOCKET_FREQ_RANGE_MIN,     //!< CPU socket minimum frequency range (MHz)
+  RDC_FI_CPU_CORE_FREQ_LIMIT,           //!< CPU core current frequency limit (MHz)
+  RDC_FI_CPU_CORE_BOOST_LIMIT,          //!< CPU core boost limit (MHz)
+  RDC_FI_CPU_SOCKET_C0_RESIDENCY,       //!< CPU socket C0 residency percentage
+  RDC_FI_CPU_DDR_BW_MAX_BW,             //!< CPU DDR maximum bandwidth (MB/s)
+  RDC_FI_CPU_DDR_BW_UTILIZED_BW,        //!< CPU DDR utilized bandwidth (MB/s)
+  RDC_FI_CPU_DDR_BW_UTILIZED_PCT,       //!< CPU DDR utilized bandwidth percentage
+  RDC_FI_CPU_SOCKET_TEMPERATURE,        //!< CPU socket temperature (millidegrees Celsius)
+  RDC_FI_CPU_DIMM_TEMP_RANGE,           //!< CPU DIMM temperature range
+  RDC_FI_CPU_DIMM_REFRESH_RATE,         //!< CPU DIMM refresh rate
+  RDC_FI_CPU_DIMM_POWER_CONSUMPTION,    //!< CPU DIMM power consumption (milliwatts)
+  RDC_FI_CPU_DIMM_THERMAL_SENSOR,       //!< CPU DIMM thermal sensor temperature (millidegrees)
+  RDC_FI_CPU_SOCKET_LCLK_DPM_LEVEL,     //!< CPU socket LCLK DPM level
+  RDC_FI_CPU_IO_BANDWIDTH,              //!< CPU I/O bandwidth (MB/s)
+  RDC_FI_CPU_XGMI_BANDWIDTH,            //!< CPU XGMI bandwidth (MB/s)
+  RDC_FI_CPU_HSMP_METRICS_VERSION,      //!< HSMP metrics table version
+  RDC_FI_CPU_HSMP_METRICS_TABLE,        //!< HSMP metrics table data
+  RDC_FI_CPU_FIRST_ONLINE_CORE,         //!< First online core on CPU socket
+  RDC_FI_CPU_FAMILY,                    //!< CPU family identifier
+  RDC_FI_CPU_MODEL_ID,                  //!< CPU model identifier
+  RDC_FI_CPU_CORES_PER_SOCKET,          //!< Number of CPU cores per socket
+  RDC_FI_CPU_SOCKET_COUNT,              //!< Number of CPU sockets
+
 } rdc_field_t;
 
 // even and odd numbers are used for correctable and uncorrectable errors

include/rdc_lib/impl/RdcMetricFetcherImpl.h

@@ -71,6 +71,11 @@ class RdcMetricFetcherImpl final : public RdcMetricFetcher {
  public:
   rdc_status_t fetch_smi_field(uint32_t gpu_index, rdc_field_t field_id,
                                rdc_field_value* value) override;
+  rdc_status_t fetch_gpu_field_(uint32_t gpu_index, rdc_field_t field_id, rdc_field_value* value,
+                                amdsmi_processor_handle& processor_handle);
+  rdc_status_t fetch_gpu_partition_field_(uint32_t gpu_index, rdc_field_t field_id,
+                                          rdc_field_value* value);
+  rdc_status_t fetch_cpu_field_(uint32_t gpu_index, rdc_field_t field_id, rdc_field_value* value);
   rdc_status_t bulk_fetch_smi_fields(
       rdc_gpu_field_t* fields, uint32_t fields_count,
       std::vector<rdc_gpu_field_value_t>& results) override;  // NOLINT
@@ -91,6 +96,11 @@ class RdcMetricFetcherImpl final : public RdcMetricFetcher {
   bool async_get_pcie_throughput(uint32_t gpu_index, rdc_field_t field_id, rdc_field_value* value);
   void get_pcie_throughput(const RdcFieldKey& key);
 
+  //!< is ESMI/CPU mode enabled?
+  bool is_cpu_enabled = false;
+
+  bool async_fetching = false;
+
   //!< Async metric retreive
   std::map<RdcFieldKey, MetricValue> async_metrics_;
   std::map<RdcFieldKey, std::shared_ptr<FieldSMIData>> smi_data_;

include/rdc_modules/rdc_rvs/RvsBase.h

@@ -71,8 +71,8 @@ namespace rdc {
 
 inline amdsmi_status_t get_processor_handle_from_id(uint32_t gpu_id,
                                                     amdsmi_processor_handle* processor_handle) {
-  uint32_t socket_count;
-  uint32_t processor_count;
+  uint32_t socket_count = 0;
+  uint32_t processor_count = 0;
   auto ret = amdsmi_get_socket_handles(&socket_count, nullptr);
   if (ret != AMDSMI_STATUS_SUCCESS) {
     return ret;

protos/rdc.proto

@@ -53,6 +53,7 @@ service RdcAPI {
   // Discovery API
   // rdc_status_t rdc_get_all_devices(uint32_t gpu_index_list[RDC_MAX_NUM_DEVICES], uint32_t* count)
   rpc GetAllDevices(Empty) returns (GetAllDevicesResponse) {}
+
   // rdc_status_t rdc_get_device_attributes(uint32_t gpu_index, rdc_device_attributes_t* p_rdc_attr)
   rpc GetDeviceAttributes(GetDeviceAttributesRequest) returns (GetDeviceAttributesResponse) {}
 

python_binding/rdc_bootstrap.py

@@ -250,6 +250,19 @@ class rdc_field_t(c_int):
      RDC_HEALTH_POWER_THROTTLE_TIME = 3006
      RDC_HEALTH_THERMAL_THROTTLE_TIME = 3007
 
+     RDC_FI_CPU_COUNT = 10000
+     RDC_FI_CPU_UTIL_TOTAL = 10001
+     RDC_FI_CPU_UTIL_USER = 10002
+     RDC_FI_CPU_UTIL_NICE = 10003
+     RDC_FI_CPU_UTIL_SYS = 10004
+     RDC_FI_CPU_UTIL_IRQ = 10005
+     RDC_FI_CPU_TEMP_CURRENT = 10006
+     RDC_FI_CPU_CLOCK_CURRENT = 10007
+     RDC_FI_CPU_POWER_UTIL_CURRENT = 10008
+     RDC_FI_CPU_POWER_LIMIT = 10009
+     RDC_FI_CPU_VENDOR = 10010
+     RDC_FI_CPU_MODEL = 10011
+
      _rdc_metric_type_lookup = {
         RDC_FI_INVALID: rdc_metric_type_t.INVALID,
         RDC_FI_GPU_COUNT: rdc_metric_type_t.LABEL,
@@ -277,7 +290,7 @@ class rdc_field_t(c_int):
      def get_rdc_metric_type(cls, rdc_metric_t):
         if isinstance(rdc_metric_t, str):
             rdc_metric_t = getattr(cls, rdc_metric_t, None)
-        
+
         # If the metric was found, do the lookup, otherwise default GAUGE
         if rdc_metric_t is not None:
             return cls._rdc_metric_type_lookup.get(rdc_metric_t, rdc_metric_type_t.GAUGE)
@@ -288,7 +301,7 @@ def get_field_name(cls, value):
         for attr_name, attr_value in cls.__dict__.items():
             if isinstance(attr_value, int) and attr_value == value:
                 return attr_name
-        return "Unknown field value"    
+        return "Unknown field value"
 
 rdc_handle_t = c_void_p
 rdc_gpu_group_t = c_uint32

rdc_libs/bootstrap/src/RdcBootStrap.cc

@@ -164,6 +164,7 @@ rdc_status_t rdc_group_gpu_add(rdc_handle_t p_rdc_handle, rdc_gpu_group_t groupI
   return static_cast<amd::rdc::RdcHandler*>(p_rdc_handle)->rdc_group_gpu_add(groupId, gpuIndex);
 }
 
+// TODO: rewrite get_all to allow different types
 rdc_status_t rdc_device_get_all(rdc_handle_t p_rdc_handle,
                                 uint32_t gpu_index_list[RDC_MAX_NUM_DEVICES], uint32_t* count) {
   if (!p_rdc_handle || !count) {

rdc_libs/rdc/src/RdcEmbeddedHandler.cc

@@ -49,8 +49,21 @@ class smi_initializer {
   smi_initializer() {
     // Make sure smi will not be initialized multiple times
     amdsmi_shut_down();
-    amdsmi_status_t ret = amdsmi_init(AMDSMI_INIT_AMD_GPUS);
+    amdsmi_status_t ret = AMDSMI_STATUS_UNKNOWN_ERROR;
+    uint64_t init_flag = AMDSMI_INIT_AMD_GPUS;
+#ifdef ENABLE_ESMI_LIB
+    init_flag |= AMDSMI_INIT_AMD_CPUS;
+#endif
+    ret = amdsmi_init(init_flag);
+#ifdef ENABLE_ESMI_LIB
     if (ret != AMDSMI_STATUS_SUCCESS) {
+      RDC_LOG(RDC_ERROR, "Failed to initalize amdsmi with CPUs enabled.. Disabling CPUs.");
+      init_flag &= ~AMDSMI_INIT_AMD_CPUS;
+      ret = amdsmi_init(init_flag);
+    }
+#endif
+    if (ret != AMDSMI_STATUS_SUCCESS) {
+      RDC_LOG(RDC_ERROR, "SMI FAILED with" << ret);
       throw amd::rdc::RdcException(RDC_ST_FAIL_LOAD_MODULE, "SMI initialize fail");
     }
   }