AI chief architect 10-aica-lanxiang, propeller frame design and core technology

0、 Introduce

platform

1、 Overall introduction to the propeller platform

The development process of deep learning

In depth learning framework process

Panoramic view of propeller

Facing the challenge

Leading technology to solve the challenge

Unity of movement and stillness

High low fusion

Simplify development process

Rich model base

Rich course

2、 The design of the training frame and the core technology of the propeller

Users and industry landing

The overall structure of the training framework

IR, Reduce complexity

Unify front-end conversion and performance

First multi-layer IR

for the first time IR

Static diagram

mapping

The second time IR

Dynamic and static conversion

Dynamic graphs contain static graphs

Dynamic tree rotation , Tree optimization , Tree generated static graph code

Tree transformation

Analysis and transcription

Dynamic and static conversion

3、 performance optimization

Optimization of each stage

Mixing accuracy , Reduce space and speed up

Facing problems

solve the problem , Zoom in , Back up high-precision data

Overall hybrid accuracy solution

Sparsity acceleration

Structured hardware acceleration and unstructured software acceleration

Hardware acceleration , Compress

effect

Model quantification

Quantitative training
Quantify after training

Quantify the effect , The data distribution

OP Fusion optimization

Multiple integration of one

Vertical integration and horizontal integration , The horizontal direction can be split at the back

Commonly used

Coding optimization

contrast

Operator optimization , Middle layer , Focus logic is different and common


kernel Optimize the system
effect

Hardware auto sensing optimization

effect

4、 Deep learning compiler

XLA

TVM

Compiler comparison

Baidu compiler

Training and reasoning can be automatically tuned

effect

summary