GPU Plugin¶
The GPU plugin uses the Intel® Compute Library for Deep Neural Networks (clDNN) to infer deep neural networks. clDNN is an open source performance library for Deep Learning (DL) applications intended for acceleration of Deep Learning Inference on Intel® Processor Graphics including Intel® HD Graphics, Intel® Iris® Graphics, Intel® Iris® Xe Graphics, and Intel® Iris® Xe MAX graphics. For an in-depth description of clDNN, see Inference Engine source files and Accelerate Deep Learning Inference with Intel® Processor Graphics.
Device Naming Convention¶
Devices are enumerated as “GPU.X” where
X={0, 1, 2,...}
. Only Intel® GPU devices are considered.If the system has an integrated GPU, it always has id=0 (“GPU.0”).
Other GPUs have undefined order that depends on the GPU driver.
“GPU” is an alias for “GPU.0”
If the system doesn’t have an integrated GPU, then devices are enumerated starting from 0.
For demonstration purposes, see the Hello Query Device C++ Sample that can print out the list of available devices with associated indices. Below is an example output (truncated to the device names only):
./hello_query_device
Available devices:
Device: CPU
...
Device: GPU.0
...
Device: GPU.1
...
Device: HDDL
Optimizations¶
The plugin supports algorithms that fuse several operations into one optimized operation. Refer to the sections below for details.
Note
For operation descriptions, see the IR Notation Reference.
Fusing Convolution and Simple Layers¶
Merge of a Convolution layer and any of the simple layers listed below:
Activation: ReLU, ELU, Sigmoid, Clamp, and others
Depthwise: ScaleShift, PReLU
FakeQuantize
Note
You can have any number and order of simple layers.
A combination of a Convolution layer and simple layers results in a single fused layer called Convolution :
Fusing Pooling and FakeQuantize Layers¶
A combination of Pooling and FakeQuantize layers results in a single fused layer called Pooling :
Fusing Activation Layers¶
Given the linear pattern, an Activation layer can be fused into other layers:
Fusing Convolution and Sum Layers¶
A combination of Convolution, Simple, and Eltwise layers with the sum operation results in a single layer called Convolution :
Fusing a Group of Convolutions¶
If a topology contains the following pipeline, a GPU plugin merges Split, Convolution, and Concatenation layers into a single Convolution layer with the group parameter:
Note
Parameters of the Convolution layers must coincide.
Optimizing Layers Out¶
The following layers are optimized out under certain conditions:
Crop
Concatenate
Reshape
Flatten
Split
Copy
Load-Time Execution¶
Some layers are executed during the load time, not during the inference. One of such layers is PriorBox.
CPU Executed Layers¶
The following layers are not accelerated on the GPU and executed on the host CPU instead:
Proposal
NonMaxSuppression
PriorBox
DetectionOutput
Supported Configuration Parameters¶
The plugin supports the configuration parameters listed below. All parameters must be set before calling InferenceEngine::Core::LoadNetwork()
in order to take effect. When specifying key values as raw strings (that is, when using Python API), omit the KEY_
prefix.
Parameter Name |
Parameter Values |
Default |
Description |
---|---|---|---|
|
|
|
Specifies a directory where compiled OCL binaries can be cached. First model loading generates the cache, and all subsequent LoadNetwork calls use precompiled kernels which significantly improves load time. If empty - caching is disabled |
|
|
|
Collect performance counters during inference |
|
|
|
Load custom layer configuration files |
|
|
|
OpenCL queue priority (before usage, make sure your OpenCL driver supports appropriate extension) Higher value means higher priority for OpenCL queue. 0 disables the setting. |
|
|
|
OpenCL queue throttling (before usage, make sure your OpenCL driver supports appropriate extension) Lower value means lower driver thread priority and longer sleep time for it. 0 disables the setting. |
|
|
|
Allows using FP16+INT8 mixed precision mode, so non-quantized parts of a model will be executed in FP16 precision for FP16 IR. Does not affect quantized FP32 IRs |
|
|
|
Controls preprocessing logic for nv12 input. If it’s set to YES, then device graph will expect that user will set biplanar nv12 blob as input wich will be directly passed to device execution graph. Otherwise, preprocessing via GAPI is used to convert NV12->BGR, thus GPU graph have to expect single input |
|
|
1 |
Specifies a number of GPU “execution” streams for the throughput mode (upper bound for a number of inference requests that can be executed simultaneously). This option is can be used to decrease GPU stall time by providing more effective load from several streams. Increasing the number of streams usually is more effective for smaller topologies or smaller input sizes. Note that your application should provide enough parallel slack (e.g. running many inference requests) to leverage full GPU bandwidth. Additional streams consume several times more GPU memory, so make sure the system has enough memory available to suit parallel stream execution. Multiple streams might also put additional load on CPU. If CPU load increases, it can be regulated by setting an appropriate |
|
|
|
Forces async requests (also from different executable networks) to execute serially. |
|
|
|
Specifies the number of CPU threads that can be used for GPU engine, e.g, JIT compilation of GPU kernels or cpu kernel processing within GPU plugin. The default value is set as the number of maximum available threads in host environment to minimize the time for LoadNetwork, where the GPU kernel build time occupies a large portion. Note that if the specified value is larger than the maximum available # of threads or less than zero, it is set as maximum available # of threads. It can be specified with a smaller number than the available HW threads according to the usage scenario, e.g., when the user wants to assign more CPU threads while GPU plugin is running. Note that setting this value with lower number will affect not only the network loading time but also the cpu layers of GPU networks that are optimized with multi-threading. |
|
|
|
Enables recurrent layers such as TensorIterator or Loop with fixed iteration count to be unrolled. It is turned on by default. Turning this key on will achieve better inference performance for loops with not too many iteration counts (less than 16, as a rule of thumb). Turning this key off will achieve better performance for both graph loading time and inference time with many iteration counts (greater than 16). Note that turning this key on will increase the graph loading time in proportion to the iteration counts. Thus, this key should be turned off if graph loading time is considered to be most important target to optimize. |
|
|
|
OpenCL queue priority (before usage, make sure your OpenCL driver supports appropriate extension) Higher value means higher priority for OpenCL queue. 0 disables the setting. Deprecated . Please use KEY_GPU_PLUGIN_PRIORITY |
|
|
|
OpenCL queue throttling (before usage, make sure your OpenCL driver supports appropriate extension) Lower value means lower driver thread priority and longer sleep time for it. 0 disables the setting. Deprecated . Please use KEY_GPU_PLUGIN_THROTTLE |
|
|
|
clDNN graph optimizer stages dump output directory (in GraphViz format) Deprecated . Will be removed in the next release |
|
|
|
Final optimized clDNN OpenCL sources dump output directory. Deprecated . Will be removed in the next release |
|
|
|
Dump the final kernels used for custom layers. Deprecated . Will be removed in the next release |
|
|
|
Disable inference kernel tuning Create tuning file (expect much longer runtime) Use an existing tuning file. Deprecated . Will be removed in the next release |
|
|
|
Tuning file to create / use. Deprecated . Will be removed in the next release |