Intermediate Representation Suitable for INT8 Inference

Introduction

OpenVINO Runtime CPU and GPU devices can infer models in the low precision. For details, refer to Model Optimization Guide.

Intermediate Representation (IR) should be specifically formed to be suitable for low precision inference. Such an IR is called a Low Precision IR and you can generate it in two ways:

For an operation to be executed in INT8, it must have FakeQuantize operations as inputs. See the specification of `FakeQuantize operation <doxid-openvino_docs_ops_quantization__fake_quantize_1>` for details.

To execute the Convolution operation in INT8 on CPU, both data and weight inputs should have FakeQuantize as an input operation:

_images/expanded_int8_Convolution_weights.png

Low precision IR is also suitable for FP32 and FP16 inference if a chosen plugin supports all operations of the IR, because the only difference between a Low Precision IR and FP16 or FP32 IR is the existence of FakeQuantize in the Low Precision IR. Plugins with Low Precision Inference support recognize these sub-graphs and quantize them during the inference time. Plugins without Low Precision support execute all operations, including FakeQuantize, as is in the FP32 or FP16 precision.

Accordingly, the presence of FakeQuantize operations in the IR is a recommendation for a plugin on how to quantize particular operations in the model. If capable, a plugin accepts the recommendation and performs Low Precision Inference, otherwise, the plugin ignores the recommendation and executes a model in the floating-point precision.

Compressed Low Precision Weights

Weighted operations, like Convolution, MatMul, and others, store weights as floating-point Constant in the graph followed by the FakeQuantize operation. Constant followed by the FakeQuantize operation could be optimized memory-wise due to the FakeQuantize operation semantics. The resulting weights sub-graph stores weights in Low Precision Constant, which gets unpacked back to floating point with the Convert operation. Weights compression replaces FakeQuantize with optional Subtract and Multiply operation leaving output arithmetically the same and weights storing takes four times less memory.

See the visualization of Convolution with the compressed weights:

_images/compressed_int8_Convolution_weights.png

Both Model Optimizer and Post-Training Optimization tool generate a compressed IR by default.