`neural_compressor.tensorflow.quantization.utils.graph_rewriter.qdq.share_qdq_y_pattern`

Share QDQ for ITEX Y pattern Graph Rewriter.

Module Contents

Insert Q/DQ op before one input of Add to enable Conv/MatMul + BiasAdd + Add + Relu fusion for ITEX.

class neural_compressor.tensorflow.quantization.utils.graph_rewriter.qdq.share_qdq_y_pattern.ShareQDQForItexYPatternOptimizer(model)[source]

Insert Q/DQ op before one input of Add to enable Conv/MatMul + BiasAdd + Add + Relu fusion for ITEX.

Only 1 Q/DQ before Add op need to be inserted. Insert 2 Q/DQ breaks the ITEX fusion pattern.