Developer Guide

Intel(R) Math Kernel Library for Deep Neural Networks (Intel(R) MKL-DNN) 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. Intel MKL-DNN is intended for deep learning applications and framework developers intersted 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
• Getting started on GPU
• 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
• C and C++ APIs

Primitives

Compute intensive operations:

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

Memory bandwidth limited operations:

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

Data manipulation:

• Reorder

Performance Benchmarking and Inspection

• Verbose Mode
• Benchmarking Performance
• Profiling with Intel(R) VTune Amplifier
• Inspecting JIT Code
• Performance Profiling Example

Advanced topics

• Version 1.0 Transition Guide
• Understanding Memory Formats
• Int8 Computation Aspects
• OpenCL Interoperability

Examples

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