[SWDEV-357472] Add evicted_ms metric (#620)

- **Added evicted_time metric for kfd processes**.  
  - Time that queues are evicted on a GPU in milliseconds
  - Added to CLI in `amd-smi monitor -q` and `amd-smi process`
  - Added to C API and Python API:
    - amdsmi_get_gpu_process_list()
    - amdsmi_get_gpu_compute_process_info()
    - amdsmi_get_gpu_compute_process_info_by_pid()

---------

Signed-off-by: Pryor, Adam <Adam.Pryor@amd.com>
This commit is contained in:
Pryor, Adam
2025-10-28 14:49:03 -05:00
committed by GitHub
parent 6f222c11a6
commit 2144cfbba4
11 changed files with 95 additions and 31 deletions
+2 -1
View File
@@ -178,8 +178,9 @@ int32_t AMDSmiGPUDevice::get_compute_process_list_impl(GPUComputeProcessList_t&
amdsmi_proc_info.memory_usage.vram_mem = rsmi_proc_info.vram_usage;
}
// Copy the cu occupancy from rsmi_process_info_t to amdsmi_proc_info_t
// Copy the kfd stats from rsmi_process_info_t to amdsmi_proc_info_t
amdsmi_proc_info.cu_occupancy = rsmi_proc_info.cu_occupancy;
amdsmi_proc_info.evicted_time = rsmi_proc_info.evicted_time;
// Safely handle KFD processes to get total memory_usage of the process
uint64_t kfd_gpu_id = get_kfd_gpu_id();