AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Don't use hwmon# as indicator of gpu

Browse Source 
Previously, during the rsmi_init discovery process, the existence
of an hwmon# directory was used to distinguish between gpus nodes
and non-gpu nodes. This isn't reliable in some scenarios. Instead,
the existence of the vbios_version file is used as an
indicator that the node is indeed a gpu.

Change-Id: Icfbe5c42ed0970077b05f25c3d209308a31bec85
This commit is contained in:
Chris Freehill
2021-01-28 18:25:53 -06:00
parent 20e2d260fb
commit ff9546aa62
7 changed files with 111 additions and 143 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/rocm_smi.cc
+17 -17
View File
@@ -391,7 +391,7 @@ static rsmi_status_t get_power_mon_value(amd::smi::PowerMonTypes type,
uint32_t dv_ind, uint64_t *val) {
amd::smi::RocmSMI& smi = amd::smi::RocmSMI::getInstance();
if (dv_ind >= smi.monitor_devices().size() || val == nullptr) {
if (dv_ind >= smi.devices().size() || val == nullptr) {
return RSMI_STATUS_INVALID_ARGS;
}
@@ -400,7 +400,7 @@ static rsmi_status_t get_power_mon_value(amd::smi::PowerMonTypes type,
return amd::smi::ErrnoToRsmiStatus(ret);
}
std::shared_ptr<amd::smi::Device> dev = smi.monitor_devices()[dv_ind];
std::shared_ptr<amd::smi::Device> dev = smi.devices()[dv_ind];
assert(dev != nullptr);
assert(dev->monitor() != nullptr);
@@ -462,9 +462,9 @@ rsmi_shut_down(void) {
#if DEBUG
int ret = 0;
#endif
for (uint32_t i = 0; i < smi.monitor_devices().size(); ++i) {
for (uint32_t i = 0; i < smi.devices().size(); ++i) {
#if DEBUG
ret = pthread_mutex_unlock(smi.monitor_devices()[i]->mutex());
ret = pthread_mutex_unlock(smi.devices()[i]->mutex());
if (ret != EPERM) { // We expect to get EPERM if the lock has already
// been released
if (ret == 0) {
@@ -476,7 +476,7 @@ rsmi_shut_down(void) {
}
}
#else
(void)pthread_mutex_unlock(smi.monitor_devices()[i]->mutex());
(void)pthread_mutex_unlock(smi.devices()[i]->mutex());
#endif
}
@@ -499,7 +499,7 @@ rsmi_num_monitor_devices(uint32_t *num_devices) {
amd::smi::RocmSMI& smi = amd::smi::RocmSMI::getInstance();
*num_devices = static_cast<uint32_t>(smi.monitor_devices().size());
*num_devices = static_cast<uint32_t>(smi.devices().size());
return RSMI_STATUS_SUCCESS;
CATCH
}
@@ -1458,13 +1458,13 @@ rsmi_dev_gpu_clk_freq_set(uint32_t dv_ind,
// Above call to rsmi_dev_get_gpu_clk_freq should have emitted an error if
// assert below is not true
assert(dv_ind < smi.monitor_devices().size());
assert(dv_ind < smi.devices().size());
std::string freq_enable_str =
bitfield_to_freq_string(freq_bitmask, freqs.num_supported);
std::shared_ptr<amd::smi::Device> dev = smi.monitor_devices()[dv_ind];
std::shared_ptr<amd::smi::Device> dev = smi.devices()[dv_ind];
assert(dev != nullptr);
ret = rsmi_dev_perf_level_set_v1(dv_ind, RSMI_DEV_PERF_LEVEL_MANUAL);
@@ -1912,12 +1912,12 @@ rsmi_dev_pci_bandwidth_set(uint32_t dv_ind, uint64_t bw_bitmask) {
// Above call to rsmi_dev_pci_bandwidth_get() should have emitted an error
// if assert below is not true
assert(dv_ind < smi.monitor_devices().size());
assert(dv_ind < smi.devices().size());
std::string freq_enable_str =
bitfield_to_freq_string(bw_bitmask, bws.transfer_rate.num_supported);
std::shared_ptr<amd::smi::Device> dev = smi.monitor_devices()[dv_ind];
std::shared_ptr<amd::smi::Device> dev = smi.devices()[dv_ind];
assert(dev != nullptr);
ret = rsmi_dev_perf_level_set_v1(dv_ind, RSMI_DEV_PERF_LEVEL_MANUAL);
@@ -2994,7 +2994,7 @@ rsmi_compute_process_gpus_get(uint32_t pid, uint32_t *dv_indices,
}
*num_devices = static_cast<uint32_t>(gpu_set.size());
if (gpu_set.size() > smi.monitor_devices().size()) {
if (gpu_set.size() > smi.devices().size()) {
return RSMI_STATUS_UNEXPECTED_SIZE;
}
@@ -3672,11 +3672,11 @@ rsmi_event_notification_get(int timeout_ms,
amd::smi::RocmSMI& smi = amd::smi::RocmSMI::getInstance();
std::vector<uint32_t> fd_indx_to_dev_id;
for (uint32_t i = 0; i < smi.monitor_devices().size(); ++i) {
if (smi.monitor_devices()[i]->evt_notif_anon_fd() == -1) {
for (uint32_t i = 0; i < smi.devices().size(); ++i) {
if (smi.devices()[i]->evt_notif_anon_fd() == -1) {
continue;
}
fds.push_back({smi.monitor_devices()[i]->evt_notif_anon_fd(),
fds.push_back({smi.devices()[i]->evt_notif_anon_fd(),
POLLIN | POLLRDNORM, 0});
fd_indx_to_dev_id.push_back(i);
}
@@ -3694,7 +3694,7 @@ rsmi_event_notification_get(int timeout_ms,
}
FILE *anon_fp =
smi.monitor_devices()[fd_indx_to_dev_id[i]]->evt_notif_anon_file_ptr();
smi.devices()[fd_indx_to_dev_id[i]]->evt_notif_anon_file_ptr();
data_item =
reinterpret_cast<rsmi_evt_notification_data_t *>(&data[*num_elem]);
@@ -3754,7 +3754,7 @@ rsmi_status_t rsmi_event_notification_stop(uint32_t dv_ind) {
return RSMI_STATUS_INVALID_ARGS;
}
// close(dev->evt_notif_anon_fd());
FILE *anon_fp = smi.monitor_devices()[dv_ind]->evt_notif_anon_file_ptr();
FILE *anon_fp = smi.devices()[dv_ind]->evt_notif_anon_file_ptr();
fclose(anon_fp);
assert(errno == 0 || errno == EAGAIN);
dev->set_evt_notif_anon_file_ptr(nullptr);
@@ -3800,7 +3800,7 @@ rsmi_test_refcount(uint64_t refcnt_type) {
amd::smi::RocmSMI& smi = amd::smi::RocmSMI::getInstance();
std::lock_guard<std::mutex> guard(*smi.bootstrap_mutex());
if (smi.ref_count() == 0 && smi.monitor_devices().size() != 0) {
if (smi.ref_count() == 0 && smi.devices().size() != 0) {
return -1;
}