IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[GPU] Adjust default num streams (#14308)

Browse Source
This commit is contained in:
Vladimir Paramuzov 2022-12-05 09:59:03 +04:00 committed by GitHub
parent b76168b32b
commit 48c0897b0f
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
5 changed files with 32 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
src/plugins/intel_gpu/include/intel_gpu/plugin/device_config.hpp
View File

@ -58,7 +58,7 @@ struct Config {
uint32_t GetDefaultNStreamsForThroughputMode() const {
return 2;
}
void UpdateFromMap(const std::map<std::string, std::string>& configMap);
void UpdateFromMap(const std::map<std::string, std::string>& configMap, const cldnn::device_info& info);
void adjustKeyMapValues();
static bool isNewApiProperty(std::string property);
static std::string ConvertPropertyToLegacy(const std::string& key, const std::string& value);

1
src/plugins/intel_gpu/include/intel_gpu/runtime/device_info.hpp
View File

@ -81,6 +81,7 @@ struct device_info {
uint32_t num_sub_slices_per_slice; ///< Number of subslices in a slice
uint32_t num_eus_per_sub_slice; ///< Number of execution units per subslice
uint32_t num_threads_per_eu; ///< Number of hardware threads per execution unit
uint32_t num_ccs; ///< Number of compute command streamers
device_uuid uuid; ///< UUID of the gpu device
};

4
src/plugins/intel_gpu/src/plugin/device_config.cpp
View File

@ -61,7 +61,7 @@ static int getNumberOfCores(const IStreamsExecutor::Config::PreferredCoreType co
}
IE_SUPPRESS_DEPRECATED_START
void Config::UpdateFromMap(const std::map<std::string, std::string>& configMap) {
void Config::UpdateFromMap(const std::map<std::string, std::string>& configMap, const cldnn::device_info& info) {
OV_ITT_SCOPED_TASK(itt::domains::intel_gpu_plugin, "Config::UpdateFromMap");
for (auto& kvp : configMap) {
std::string key = kvp.first;
@ -250,7 +250,7 @@ void Config::UpdateFromMap(const std::map<std::string, std::string>& configMap)
} else if (key.compare(PluginConfigParams::KEY_GPU_THROUGHPUT_STREAMS) == 0 || key == ov::num_streams) {
if (val.compare(PluginConfigParams::GPU_THROUGHPUT_AUTO) == 0 ||
val.compare(ov::util::to_string(ov::streams::AUTO)) == 0) {
throughput_streams = GetDefaultNStreamsForThroughputMode();
throughput_streams = std::max(GetDefaultNStreamsForThroughputMode(), info.num_ccs);
} else {
int val_i;
try {

20
src/plugins/intel_gpu/src/plugin/plugin.cpp
View File

@ -196,7 +196,7 @@ void Plugin::UpdateConfig(Config& conf, const InferenceEngine::CNNNetwork &netwo
OV_ITT_SCOPED_TASK(itt::domains::intel_gpu_plugin, "Plugin::UpdateConfig");
auto device_info = GetDeviceInfo(params);
conf.enableInt8 = device_info.supports_imad || device_info.supports_immad;
conf.UpdateFromMap(params);
conf.UpdateFromMap(params, device_info);
if (conf.enableDynamicBatch) {
conf.max_dynamic_batch = static_cast<int>(network.getBatchSize());
}
@ -357,17 +357,27 @@ void Plugin::SetConfig(const std::map<std::string, std::string> &config) {
config.find(CLDNNConfigParams::KEY_CLDNN_PLUGIN_THROTTLE) != config.end() ||
config.find(ov::intel_gpu::hint::queue_throttle.name()) != config.end();
std::string device_id;
cldnn::device_info device_info = device_map.begin()->second->get_info();
if (config.find(PluginConfigInternalParams::KEY_CONFIG_DEVICE_ID) != config.end()) {
device_id = config.at(PluginConfigInternalParams::KEY_CONFIG_DEVICE_ID);
_impl->m_configs.GetConfig(device_id).UpdateFromMap(config);
if (!device_id.empty() && device_map.find(device_id) != device_map.end()) {
device_info = device_map.at(device_id)->get_info();
}
_impl->m_configs.GetConfig(device_id).UpdateFromMap(config, device_info);
} else {
device_id = GetDeviceIDFromConfig(config);
if (!device_id.empty()) {
if (device_map.find(device_id) != device_map.end()) {
device_info = device_map.at(device_id)->get_info();
}
_impl->m_configs.SetDefaultDeviceID(device_id);
_impl->m_configs.GetConfig(device_id).UpdateFromMap(config);
_impl->m_configs.GetConfig(device_id).UpdateFromMap(config, device_info);
} else {
for (auto& conf : _impl->m_configs) {
conf.second.UpdateFromMap(config);
if (device_map.find(conf.first) != device_map.end()) {
device_info = device_map.at(conf.first)->get_info();
}
conf.second.UpdateFromMap(config, device_info);
}
}
}
@ -835,7 +845,7 @@ Parameter Plugin::GetMetric(const std::string& name, const std::map<std::string,
n_streams_str != util::to_string(ov::streams::AUTO)) {
IE_THROW() << "[GPU_MAX_BATCH_SIZE] bad casting: GPU_THROUGHPUT_STREAMS should be either of uint32_t type or \"GPU_THROUGHPUT_AUTO\"";
}
n_streams = config.GetDefaultNStreamsForThroughputMode();
n_streams = std::max(config.GetDefaultNStreamsForThroughputMode(), device_info.num_ccs);
} else {
IE_THROW() << "[GPU_MAX_BATCH_SIZE] bad casting: GPU_THROUGHPUT_STREAMS should be either of uint32_t type or \"GPU_THROUGHPUT_AUTO\"";
}

13
src/plugins/intel_gpu/src/runtime/ocl/ocl_device.cpp
View File

@ -230,6 +230,19 @@ device_info init_device_info(const cl::Device& device) {
info.num_threads_per_eu = 0;
}
info.num_ccs = 1;
if (info.supports_queue_families) {
cl_uint num_queues = 0;
std::vector<cl_queue_family_properties_intel> qfprops = device.getInfo<CL_DEVICE_QUEUE_FAMILY_PROPERTIES_INTEL>();
for (cl_uint q = 0; q < qfprops.size(); q++) {
if (qfprops[q].capabilities == CL_QUEUE_DEFAULT_CAPABILITIES_INTEL && qfprops[q].count > num_queues) {
num_queues = qfprops[q].count;
}
}
info.num_ccs = std::max<uint32_t>(num_queues, info.num_ccs);
}
return info;
}