Conv3DQuantProcessFusionPass

Description

Performs quantization fallback or bias optimizations when the shape is static.

• Bias optimization: When the c0 value of AscendQuant is greater than that of fp16 or Conv3D does not support the fp16 format, bias optimization is required. Otherwise, perform quantization fallback. For bias optimization, the graph structure does not need to be modified.
• During quantization fallback, constant folding operators are inserted to eliminate AscendQuant and AscendDequant operators. For details, see the following figures.

After: (unquantization)

Restrictions

None