Developer Guide

Deep Neural Network Library (DNNL) is an open-source performance library for deep learning applications. The library includes basic building blocks for neural networks optimized for Intel Architecture Processors and Intel Processor Graphics. DNNL is intended for deep learning applications and framework developers interested in improving application performance on Intel CPUs and GPUs.

Building and Linking

• Build from Source
  • Build Options
• Linking to the Library

Programming Model

• Basic Concepts
• Getting started
• Memory format propagation
• Inference and Training Aspects
  • Inference
  • Int8 Inference
  • Bfloat16 Training
• Primitive Attributes
  • Primitive Attributes: Scratchpad
  • Primitive Attributes: Quantization
  • Primitive Attributes: Post-ops
• Data Types
• Reorder between CPU and GPU engines
• API

Primitives

Compute intensive operations:

• (De-)Convolution: Direct 1D/2D/3D, Winograd 2D
• Inner Product
• Matrix Multiplication
• RNN: LSTM, Vanilla RNN, GRU

Memory bandwidth limited operations:

• Batch Normalization
• Binary
• Concat
• Elementwise: ReLU, Tanh, ELU, Abs, and other
• Layer Normalization
• Local Response Normalization
• LogSoftmax
• Pooling
• Resampling
• Shuffle
• Softmax
• Sum

Data manipulation:

• Reorder

Performance Benchmarking and Inspection

• Verbose Mode
• Benchmarking Performance
• Profiling DNNL performance
• Inspecting JIT Code
• Performance Profiling Example
• CPU dispatcher control

Advanced topics

• Version 1.0 Transition Guide
• Intel MKL-DNN to DNNL Transition Guide
• Understanding Memory Formats
• Int8 Computation Aspects
• OpenCL Interoperability
• Primitive Cache

Examples

Topic	Engine	C++ API	C API
Tutorials	CPU/GPU	Getting started
	CPU/GPU	Memory format propagation
	CPU/GPU	Performance Profiling Example
	CPU/GPU	Reorder between CPU and GPU engines	Reorder between CPU and GPU engines
	GPU	Getting started on GPU with OpenCL extensions API
f32 inference	CPU/GPU	CNN f32 inference example	CNN f32 inference example
	CPU	RNN f32 inference example
int8 inference	CPU/GPU	CNN int8 inference example
	CPU	RNN int8 inference example
f32 training	CPU/GPU	CNN f32 training example
	CPU		CNN f32 training example
	CPU/GPU	RNN f32 training example
bf16 training	CPU	CNN bf16 training example