PAPI_get_dev_attr - Man Page

typedef enum {
    PAPI_DEV_ATTR__CPU_CHAR_NAME,
    PAPI_DEV_ATTR__CPU_UINT_L1I_CACHE_SIZE,
    PAPI_DEV_ATTR__CPU_UINT_L1D_CACHE_SIZE,
    PAPI_DEV_ATTR__CPU_UINT_L2U_CACHE_SIZE,
    PAPI_DEV_ATTR__CPU_UINT_L3U_CACHE_SIZE,
    PAPI_DEV_ATTR__CPU_UINT_L1I_CACHE_LINE_SIZE,
    PAPI_DEV_ATTR__CPU_UINT_L1D_CACHE_LINE_SIZE,
    PAPI_DEV_ATTR__CPU_UINT_L2U_CACHE_LINE_SIZE,
    PAPI_DEV_ATTR__CPU_UINT_L3U_CACHE_LINE_SIZE,
    PAPI_DEV_ATTR__CPU_UINT_L1I_CACHE_LINE_COUNT,
    PAPI_DEV_ATTR__CPU_UINT_L1D_CACHE_LINE_COUNT,
    PAPI_DEV_ATTR__CPU_UINT_L2U_CACHE_LINE_COUNT,
    PAPI_DEV_ATTR__CPU_UINT_L3U_CACHE_LINE_COUNT,
    PAPI_DEV_ATTR__CPU_UINT_L1I_CACHE_ASSOC,
    PAPI_DEV_ATTR__CPU_UINT_L1D_CACHE_ASSOC,
    PAPI_DEV_ATTR__CPU_UINT_L2U_CACHE_ASSOC,
    PAPI_DEV_ATTR__CPU_UINT_L3U_CACHE_ASSOC,
    PAPI_DEV_ATTR__CPU_UINT_SOCKET_COUNT,
    PAPI_DEV_ATTR__CPU_UINT_NUMA_COUNT,
    PAPI_DEV_ATTR__CPU_UINT_CORE_COUNT,
    PAPI_DEV_ATTR__CPU_UINT_THREAD_COUNT,
    PAPI_DEV_ATTR__CPU_UINT_FAMILY,
    PAPI_DEV_ATTR__CPU_UINT_MODEL,
    PAPI_DEV_ATTR__CPU_UINT_STEPPING,
    PAPI_DEV_ATTR__CPU_UINT_NUMA_MEM_SIZE,
    PAPI_DEV_ATTR__CPU_UINT_THR_NUMA_AFFINITY,
    PAPI_DEV_ATTR__CPU_UINT_THR_PER_NUMA,
    PAPI_DEV_ATTR__CUDA_ULONG_UID,
    PAPI_DEV_ATTR__CUDA_CHAR_DEVICE_NAME,
    PAPI_DEV_ATTR__CUDA_UINT_WARP_SIZE,
    PAPI_DEV_ATTR__CUDA_UINT_SHM_PER_BLK,
    PAPI_DEV_ATTR__CUDA_UINT_SHM_PER_SM,
    PAPI_DEV_ATTR__CUDA_UINT_BLK_DIM_X,
    PAPI_DEV_ATTR__CUDA_UINT_BLK_DIM_Y,
    PAPI_DEV_ATTR__CUDA_UINT_BLK_DIM_Z,
    PAPI_DEV_ATTR__CUDA_UINT_GRD_DIM_X,
    PAPI_DEV_ATTR__CUDA_UINT_GRD_DIM_Y,
    PAPI_DEV_ATTR__CUDA_UINT_GRD_DIM_Z,
    PAPI_DEV_ATTR__CUDA_UINT_THR_PER_BLK,
    PAPI_DEV_ATTR__CUDA_UINT_SM_COUNT,
    PAPI_DEV_ATTR__CUDA_UINT_MULTI_KERNEL,
    PAPI_DEV_ATTR__CUDA_UINT_MAP_HOST_MEM,
    PAPI_DEV_ATTR__CUDA_UINT_MEMCPY_OVERLAP,
    PAPI_DEV_ATTR__CUDA_UINT_UNIFIED_ADDR,
    PAPI_DEV_ATTR__CUDA_UINT_MANAGED_MEM,
    PAPI_DEV_ATTR__CUDA_UINT_COMP_CAP_MAJOR,
    PAPI_DEV_ATTR__CUDA_UINT_COMP_CAP_MINOR,
    PAPI_DEV_ATTR__CUDA_UINT_BLK_PER_SM,
    PAPI_DEV_ATTR__ROCM_ULONG_UID,
    PAPI_DEV_ATTR__ROCM_CHAR_DEVICE_NAME,
    PAPI_DEV_ATTR__ROCM_UINT_WAVEFRONT_SIZE,
    PAPI_DEV_ATTR__ROCM_UINT_WORKGROUP_SIZE,
    PAPI_DEV_ATTR__ROCM_UINT_WAVE_PER_CU,
    PAPI_DEV_ATTR__ROCM_UINT_SHM_PER_WG,
    PAPI_DEV_ATTR__ROCM_UINT_WG_DIM_X,
    PAPI_DEV_ATTR__ROCM_UINT_WG_DIM_Y,
    PAPI_DEV_ATTR__ROCM_UINT_WG_DIM_Z,
    PAPI_DEV_ATTR__ROCM_UINT_GRD_DIM_X,
    PAPI_DEV_ATTR__ROCM_UINT_GRD_DIM_Y,
    PAPI_DEV_ATTR__ROCM_UINT_GRD_DIM_Z,
    PAPI_DEV_ATTR__ROCM_UINT_CU_COUNT,
    PAPI_DEV_ATTR__ROCM_UINT_SIMD_PER_CU,
    PAPI_DEV_ATTR__ROCM_UINT_COMP_CAP_MAJOR,
    PAPI_DEV_ATTR__ROCM_UINT_COMP_CAP_MINOR,
} PAPI_dev_attr_e;

void *handle;
int id;
int count;
int enum_modifier = PAPI_DEV_TYPE_ENUM__CPU | PAPI_DEV_TYPE_ENUM__CUDA;
while (PAPI_OK == PAPI_enum_dev_type(enum_modifier, &handle)) {
    PAPI_get_dev_type_attr(handle, PAPI_DEV_TYPE_ATTR__INT_PAPI_ID, &id);
    PAPI_get_dev_type_attr(handle, PAPI_DEV_TYPE_ATTR__INT_COUNT, &count);
    if (PAPI_DEV_TYPE_ID__CUDA == id) {
        for (int i = 0; i < count; ++i) {
            unsigned int warp_size;
            unsigned int cc_major, cc_minor;
            PAPI_get_dev_attr(handle, i, PAPI_DEV_ATTR__CUDA_UINT_WARP_SIZE, &warp_size);
            PAPI_get_dev_attr(handle, i, PAPI_DEV_ATTR__CUDA_UINT_COMP_CAP_MAJOR, &cc_major);
            PAPI_get_dev_attr(handle, i, PAPI_DEV_ATTR__CUDA_UINT_COMP_CAP_MINOR, &cc_minor);
            ...
        }
    }
}

PAPI_get_dev_type_attr() allows the user to query all device type attributes. It takes a device type handle, returned by PAPI_enum_dev_type, the device sequential id and an attribute to be queried for the device and returns the attribute value.