Reduce Load times for Partition CLI (#290)

* Reduced Load times for CLI in partition mode * Change rsmi_dev_id_get() to use KFD, if KGD interface does not exist * Make gpu_device_uuid fallback to rsmi_wrapper * Moved Enumeration info calls in list for more speed * Moved made group check excluded from recursion --------- Signed-off-by: Maisam Arif <Maisam.Arif@amd.com> Signed-off-by: Charis Poag <Charis.Poag@amd.com> Co-authored-by: Charis Poag <Charis.Poag@amd.com> Co-authored-by: gabrpham_amdeng <Gabriel.Pham@amd.com> [ROCm/amdsmi commit: 63b13ecb05]
2025-04-22 22:54:43 -05:00
@@ -198,7 +198,9 @@ class AMDSMICommands():
        if args.gpu == None:
            args.gpu = self.device_handles

-        self.helpers.check_required_groups()
+        if not self.group_check_printed:
+            self.helpers.check_required_groups()
+            self.group_check_printed = True

        # Handle multiple GPUs
        handled_multiple_gpus, device_handle = self.helpers.handle_gpus(args, self.logger, self.list)
@@ -219,15 +221,6 @@ class AMDSMICommands():
        except amdsmi_exception.AmdSmiLibraryException as e:
            uuid = e.get_error_info()

-        try:
-            enumeration_info = amdsmi_interface.amdsmi_get_gpu_enumeration_info(args.gpu)
-        except:
-            enumeration_info = {"drm_render": "N/A",
-                                "drm_card": "N/A",
-                                "hip_id": "N/A",
-                                "hip_uuid": "N/A",
-                                "hsa_id": "N/A"}
-
        try:
            kfd_info = amdsmi_interface.amdsmi_get_gpu_kfd_info(args.gpu)
            kfd_id = kfd_info['kfd_id']
@@ -250,6 +243,14 @@ class AMDSMICommands():
        self.logger.store_output(args.gpu, 'partition_id', partition_id)

        if args.e:
+            try:
+                enumeration_info = amdsmi_interface.amdsmi_get_gpu_enumeration_info(args.gpu)
+            except:
+                enumeration_info = {"drm_render": "N/A",
+                                    "drm_card": "N/A",
+                                    "hip_id": "N/A",
+                                    "hip_uuid": "N/A",
+                                    "hsa_id": "N/A"}
            if enumeration_info['drm_render'] == "N/A":
                self.logger.store_output(args.gpu, 'render', enumeration_info['drm_render'])
            else:
@@ -404,7 +405,9 @@ class AMDSMICommands():
                                   args.vram, args.cache, args.board, args.process_isolation,
                                   args.clock, args.partition]

-        self.helpers.check_required_groups()
+        if not self.group_check_printed:
+            self.helpers.check_required_groups()
+            self.group_check_printed = True

        if self.helpers.is_linux() and self.helpers.is_baremetal():
            if limit:
@@ -3409,7 +3412,9 @@ class AMDSMICommands():
        # Clear the table header
        self.logger.table_header = ''.rjust(12)

-        self.helpers.check_required_groups()
+        if not self.group_check_printed:
+            self.helpers.check_required_groups()
+            self.group_check_printed = True

        # Populate the possible gpus
        topo_values = []
@@ -4698,7 +4703,9 @@ class AMDSMICommands():
        if core:
            args.core = core

-        self.helpers.check_required_groups()
+        if not self.group_check_printed:
+            self.helpers.check_required_groups()
+            self.group_check_printed = True

        # Check if a GPU argument has been set
        gpu_args_enabled = False
@@ -4844,7 +4851,9 @@ class AMDSMICommands():
        if args.gpu == None:
            args.gpu = self.device_handles

-        self.helpers.check_required_groups()
+        if not self.group_check_printed:
+            self.helpers.check_required_groups()
+            self.group_check_printed = True

        # Handle multiple GPUs
        handled_multiple_gpus, device_handle = self.helpers.handle_gpus(args, self.logger, self.reset)
@@ -5722,7 +5731,9 @@ class AMDSMICommands():
        # Clear the table header
        self.logger.table_header = ''.rjust(7)

-        self.helpers.check_required_groups()
+        if not self.group_check_printed:
+            self.helpers.check_required_groups()
+            self.group_check_printed = True

        # Populate the possible gpus and their bdfs
        xgmi_values = []
@@ -5968,7 +5979,9 @@ class AMDSMICommands():
        if accelerator:
            args.accelerator = accelerator

-        self.helpers.check_required_groups()
+        if not self.group_check_printed:
+            self.helpers.check_required_groups()
+            self.group_check_printed = True

        ###########################################
        # amd-smi partition (no args)             #
@@ -6296,7 +6309,10 @@ class AMDSMICommands():
        if args.gpu == None:
            args.gpu = self.device_handles

-        self.helpers.check_required_groups()
+        if not self.group_check_printed:
+            self.helpers.check_required_groups()
+            self.group_check_printed = True
+
        handled_multiple_gpus, device_handle = self.helpers.handle_gpus(args, self.logger, self.ras)
        if handled_multiple_gpus:
            return
@@ -799,6 +799,9 @@ rsmi_topo_numa_affinity_get(uint32_t dv_ind, int32_t *numa_node) {
  DEVICE_MUTEX
  std::string str_val;
  ret = get_dev_value_str(amd::smi::kDevNumaNode, dv_ind, &str_val);
+  if (ret != RSMI_STATUS_SUCCESS){
+    return ret;
+  }
  *numa_node = std::stoi(str_val, nullptr);

  return ret;
@@ -913,11 +916,41 @@ rsmi_dev_id_get(uint32_t dv_ind, uint16_t *id) {
  ss << __PRETTY_FUNCTION__ << "| ======= start =======";
  LOG_TRACE(ss);
  CHK_SUPPORT_NAME_ONLY(id)
+  *id = std::numeric_limits<uint16_t>::max();

+  // Get the device ID from KGD
  ret = get_id(dv_ind, amd::smi::kDevDevID, id);
-  ss << __PRETTY_FUNCTION__ << " | ======= end ======="
-     << ", reporting " << amd::smi::getRSMIStatusString(ret);
  LOG_TRACE(ss);
+  ss << __PRETTY_FUNCTION__
+     << (ret == RSMI_STATUS_SUCCESS ?
+          " | No fall back needed retrieved from KGD" : " | fall back needed")
+     << " | Device #: " << std::to_string(dv_ind)
+     << " | Data: device_id = " << std::to_string(*id)
+     << " | ret = " << getRSMIStatusString(ret, false);
+  LOG_DEBUG(ss);
+  // If the device ID is not supported, use KFD's device ID
+  if (ret != RSMI_STATUS_SUCCESS) {
+    GET_DEV_AND_KFDNODE_FROM_INDX
+    uint32_t node_id;
+    uint64_t kfd_device_id;
+    int ret_kfd = kfd_node->get_node_id(&node_id);
+    ret_kfd = amd::smi::read_node_properties(node_id, "device_id", &kfd_device_id);
+    if (ret_kfd == 0) {
+      *id = kfd_device_id;
+      ret = RSMI_STATUS_SUCCESS;
+    } else {
+      *id = std::numeric_limits<uint16_t>::max();
+      ret = RSMI_STATUS_NOT_SUPPORTED;
+    }
+    ss << __PRETTY_FUNCTION__
+       << " | Issue: Could not read device from sysfs, falling back to KFD" << "\n"
+       << " ; Device #: " << std::to_string(dv_ind) << "\n"
+       << " ; ret_kfd: " << std::to_string(ret_kfd) << "\n"
+       << " ; node: " << std::to_string(node_id) << "\n"
+       << " ; Data: device_id (from KFD)= " << std::to_string(*id) << "\n"
+       << " ; ret = " << getRSMIStatusString(ret, false);
+    LOG_DEBUG(ss);
+  }
  return ret;
 }

@@ -570,31 +570,39 @@ amdsmi_get_gpu_device_uuid(amdsmi_processor_handle processor_handle,
                           char *uuid) {
    AMDSMI_CHECK_INIT();

-    if (uuid_length == nullptr || uuid == nullptr || uuid_length == nullptr || *uuid_length < AMDSMI_GPU_UUID_SIZE) {
+    if (uuid_length == nullptr || uuid == nullptr || *uuid_length < AMDSMI_GPU_UUID_SIZE) {
        return AMDSMI_STATUS_INVAL;
    }

-    amd::smi::AMDSmiGPUDevice* gpu_device = nullptr;
-    amdsmi_status_t r = get_gpu_device_from_handle(processor_handle, &gpu_device);
-    if (r != AMDSMI_STATUS_SUCCESS)
-        return r;
+    uint64_t device_uuid = 0;
+    uint16_t device_id = std::numeric_limits<uint16_t>::max();
+    amdsmi_status_t status;

-    amdsmi_status_t status = AMDSMI_STATUS_SUCCESS;
-    SMIGPUDEVICE_MUTEX(gpu_device->get_mutex())
-
-    amdsmi_asic_info_t asic_info = {};
-    const uint8_t fcn = 0xff;
-
-    status = amdsmi_get_gpu_asic_info(processor_handle, &asic_info);
+    status = rsmi_wrapper(rsmi_dev_id_get, processor_handle, 0, &device_id);
    if (status != AMDSMI_STATUS_SUCCESS) {
-        printf("Getting asic info failed. Return code: %d", status);
+        std::ostringstream ss;
+        ss << __PRETTY_FUNCTION__
+        << " | rsmi_dev_id_get(): "
+        << smi_amdgpu_get_status_string(status, true);
+        LOG_INFO(ss);
+        device_id = std::numeric_limits<uint16_t>::max();
+    }
+
+    status = rsmi_wrapper(rsmi_dev_unique_id_get, processor_handle, 0,
+                            &device_uuid);
+    if (status != AMDSMI_STATUS_SUCCESS) {
+        std::ostringstream ss;
+        ss << __PRETTY_FUNCTION__
+        << " | rsmi_dev_unique_id_get(): "
+        << smi_amdgpu_get_status_string(status, true);
+        LOG_INFO(ss);
        return status;
    }

+    const uint8_t fcn = 0xff;
+
    /* generate random UUID */
-    status = amdsmi_uuid_gen(uuid,
-                strtoull(asic_info.asic_serial, nullptr, 16),
-                (uint16_t)asic_info.device_id, fcn);
+    status = amdsmi_uuid_gen(uuid, device_uuid, device_id, fcn);
    return status;
 }

@@ -648,10 +656,10 @@ amdsmi_get_gpu_enumeration_info(amdsmi_processor_handle processor_handle,

    // Retrieve HIP UUID
    std::string hip_uuid_str = "GPU-";
-    amdsmi_asic_info_t asic_info = {};
-    status = amdsmi_get_gpu_asic_info(processor_handle, &asic_info);
+    char asic_serial[AMDSMI_GPU_UUID_SIZE];
+    status = amdsmi_get_gpu_device_uuid(processor_handle, 0, asic_serial);
    if (status == AMDSMI_STATUS_SUCCESS) {
-        hip_uuid_str += std::string(asic_info.asic_serial).substr(0, sizeof(info->hip_uuid) - hip_uuid_str.size() - 1);
+        hip_uuid_str += std::string(asic_serial).substr(0, sizeof(info->hip_uuid) - hip_uuid_str.size() - 1);
        std::strncpy(info->hip_uuid, hip_uuid_str.c_str(), sizeof(info->hip_uuid) - 1);
        info->hip_uuid[sizeof(info->hip_uuid) - 1] = '\0'; // Ensure null termination
    }