配置设备

推理引擎 API 提供通过配置密钥来 配置设备 及 获取设备特定指标 的功能。按字符串名称请求从 InferenceEngine::Core::GetConfig 中提取的值，而返回类型为 InferenceEngine::Parameter ，导致用户无法确定此参数中存储的到底是哪种类型。

API 2.0 通过引入 属性 来统一指标和配置密钥概念， 从而解决了这些问题。主要优势在于它们具有 C++ 类型：

static constexpr Property<std::string> full_name{"FULL_DEVICE_NAME"};

其中，可以从推理设备请求属性，如下所示：

// 'auto' is automatically deduced as std::string
// since the type is stored in the property
auto full_device_name = core.get_property("CPU", ov::device::full_name);

以下部分中的代码片段演示了用于从推理引擎迁移到 API 2.0 的设备配置。

设置配置值

推理引擎 API

C++

Devices

core.SetConfig({ { CONFIG_KEY(PERF_COUNT), CONFIG_VALUE(YES) } }, "CPU");

Model Loading

auto exec_network = core.LoadNetwork(model, "MULTI", {
    { MULTI_CONFIG_KEY(DEVICE_PRIORITIES), "CPU, GPU" },
    { CONFIG_KEY(PERFORMANCE_HINT), CONFIG_VALUE(THROUGHPUT) },
    { CONFIG_KEY(ENFORCE_BF16), CONFIG_VALUE(NO) } });

Execution

// turn CPU off for multi-device execution
exec_network.SetConfig({ { MULTI_CONFIG_KEY(DEVICE_PRIORITIES), "GPU" } });

Python

Devices

core.set_config({"PERF_COUNT": "YES"}, "CPU")

Model Loading

exec_network = core.load_network(net, "MULTI", {"DEVICE_PRIORITIES": "CPU, GPU",
                                                "PERFORMANCE_HINT": "THROUGHPUT",
                                                "ENFORCE_BF16": "NO"})

Execution

# turn CPU off for multi-device execution
exec_network.set_config({"DEVICE_PRIORITIES": "GPU"})

C

Devices

ie_config_t config = {"PERF_COUNT", "YES", NULL};

Model Loading

ie_config_t config_1 = {"DEVICE_PRIORITIES", "CPU, GPU", NULL};
ie_config_t config_2 = {"PERFORMANCE_HINT", "THROUGHPUT", &config_1};
ie_config_t config_3 = {"ENFORCE_BF16", "NO", &config_2};
ie_executable_network_t *exe_network = NULL;

Execution

// turn CPU off for multi-device executio
ie_config_t config_param = {"DEVICE_PRIORITIES", "GPU", NULL};
ie_exec_network_set_config(exe_network, &config_param);

API 2.0

Python

Devices

core.set_property(device_name="CPU", properties={"PERF_COUNT": "YES"})

Model Loading

compiled_model = core.compile_model(model=model, device_name="MULTI", config=
    {
        "MULTI_DEVICE_PRIORITIES": "GPU,CPU",
        "PERFORMANCE_HINT": "THROUGHPUT",
        "INFERENCE_PRECISION_HINT": "f32"
    })

Execution

# turn CPU off for multi-device execution
compiled_model.set_property(properties={"MULTI_DEVICE_PRIORITIES": "GPU"})

C++

Devices

core.set_property("CPU", ov::enable_profiling(true));

Model Loading

auto compiled_model = core.compile_model(model, "MULTI",
    ov::device::priorities("GPU", "CPU"),
    ov::hint::performance_mode(ov::hint::PerformanceMode::THROUGHPUT),
    ov::hint::inference_precision(ov::element::f32));

Execution

// turn CPU off for multi-device execution
compiled_model.set_property(ov::device::priorities("GPU"));

C

Devices

ov_core_set_property(core, "CPU", ov_property_key_enable_profiling, "TRUE");

Model Loading

ov_compiled_model_t* compiled_model = NULL;

Execution

// turn CPU off for multi-device execution

获取信息

推理引擎 API

Python

Device Configuration

num_streams = core.get_config("CPU", "CPU_THROUGHPUT_STREAMS")

Device metrics

full_device_name = core.get_metric("CPU", "FULL_DEVICE_NAME")

Execution config

# turn CPU off for multi-device execution
exec_network.set_config({"DEVICE_PRIORITIES": "GPU"})

Execution metrics

nireq = exec_network.get_metric("OPTIMAL_NUMBER_OF_INFER_REQUESTS")

C++

Device Configuration

// a user has to parse std::string after
auto num_streams = core.GetConfig("CPU", CONFIG_KEY(CPU_THROUGHPUT_STREAMS)).as<std::string>();

Device metrics

auto full_device_name = core.GetMetric("CPU", METRIC_KEY(FULL_DEVICE_NAME)).as<std::string>();

Execution config

// turn CPU off for multi-device execution
exec_network.SetConfig({ { MULTI_CONFIG_KEY(DEVICE_PRIORITIES), "GPU" } });

Execution metrics

auto nireq = exec_network.GetMetric(EXEC_NETWORK_METRIC_KEY(OPTIMAL_NUMBER_OF_INFER_REQUESTS)).as<uint32_t>();

C

Device Configuration

ie_param_t num_streams;
num_streams.params = NULL;

Device metrics

ie_param_t full_device_name;
full_device_name.params = NULL;

Execution config

// turn CPU off for multi-device executio
ie_config_t config_param = {"DEVICE_PRIORITIES", "GPU", NULL};
ie_exec_network_set_config(exe_network, &config_param);

Execution metrics

ie_param_t nireq;
nireq.params = NULL;
ie_exec_network_get_metric(exe_network, "OPTIMAL_NUMBER_OF_INFER_REQUESTS", &nireq);
ie_param_free(&nireq);

API 2.0

Python

Device Configuration

num_streams = core.get_property("CPU", "NUM_STREAMS")

Device metrics

full_device_name = core.get_property("CPU", "FULL_DEVICE_NAME")

Execution config

perf_mode = compiled_model.get_property("PERFORMANCE_HINT")

Execution metrics

nireq = compiled_model.get_property("OPTIMAL_NUMBER_OF_INFER_REQUESTS")

C++

Device Configuration

// 'auto' is automatically deduced as ov::streams::Num
// since the type is stored in the property
auto num_streams = core.get_property("CPU", ov::streams::num);

Device metrics

// 'auto' is automatically deduced as std::string
// since the type is stored in the property
auto full_device_name = core.get_property("CPU", ov::device::full_name);

Execution config

ov::hint::PerformanceMode perf_mode = compiled_model.get_property(ov::hint::performance_mode);

Execution metrics

// 'auto' is deduced to 'uint32_t'
auto nireq = compiled_model.get_property(ov::optimal_number_of_infer_requests);

C

Device Configuration

char* num_streams = NULL;
ov_core_get_property(core, "CPU", ov_property_key_num_streams, &num_streams);
ov_free(num_streams);

Device metrics

char* full_device_name = NULL;
ov_core_get_property(core, "CPU", ov_property_key_device_full_name, &full_device_name);
ov_free(full_device_name);

Execution config

char* perf_mode = NULL;
ov_compiled_model_get_property(compiled_model, ov_property_key_hint_performance_mode, &perf_mode);
ov_free(perf_mode);

Execution metrics

char* nireq = NULL;
ov_compiled_model_get_property(compiled_model, ov_property_key_hint_num_requests, &nireq);
ov_free(nireq);