Quantized MatMul Fusion Patterns

Quantized MatMul Fusion Patterns#

Overview#

oneDNN supports both floating-point and quantized MatMul fusion patterns to optimize performance and reduce memory bandwidth requirements. This document describes the supported quantized fusion patterns for MatMul. For floating-point MatMul fusion patterns, refer to MatMul Fusion Patterns for more details.

Pattern Structure#

oneDNN defines quantized MatMul fusion patterns as follows. The blue nodes are required when defining a quantized MatMul fusion pattern while the brown nodes are optional.

quantized MatMul pattern

  1. Q2F Conversion Subgraph : Converts src and weights tensors from quantized to floating-point. It can be one of the following subgraphs, while the last two subgraphs apply only to weights. See Dequantize, TypeCast and Quantize operations in Graph API.

    q2f_conversion_subgraph

  2. F2F Conversion Subgraph : Converts bias tensor from floating-point to another floating-point. It is constructed by a TypeCast operation.

    f2f_conversion_subgraph

  3. MatMul Operation : Performs matrix multiplication between the src and weights tensors. The bias tensor is optional. See the MatMul operation in the Graph API for more details.

  4. Epilogue Subgraph : Optional and can include the following operations:

    Combination rules:

    epilogue subgraph

    • BiasAdd : If present, must be the first op in the epilogue subgraph and can only appear once.

    • 0 to 4 Binary or Unary operations are supported in the epilogue subgraph.

    • Select : If present, must follow binary/unary operations (if present) and can only appear once.

  5. F2F/F2Q Conversion Subgraph : Converts the output tensor from floating-point to floating-point or quantized data type. It can be one of the following subgraphs, the last two subgraphs are implementations for SmoothQuant[1]. See TypeCast, Quantize and Multiply operations in Graph API.

    f2q_conversion_subgraph

Data Types#

oneDNN supports the following combinations of data types for src, weights, bias and dst:

src

weights

bias

dst

u8,s8

s8,f32

f32,bf16,f16

u8,s8,bf16,f16,f32

The definition of the data types and support status on different CPU and GPU platforms follow the general description in the Data Types Guide.

Limitations#

  • F2F Conversion Subgraph used for bias tensor only supports f32 to bf16 data type conversion.

Reference#

[1] SmoothQuant, https://arxiv.org/abs/2211.10438

Contents