TVM 各个模块总体架构
Deploy Deep Learning Everywhere
Existing Deep Learning Frameworks
Limitations of Existing Approach
Learning-based Learning System
Problem Setting
Example Instance in a Search Space
Optimization Choices in a Search Space
Problem Formalization
Black-box Optimization
Cost-model Driven Approach
Statistical Cost Model
Unique Problem Characteristics
Vanilla Cost Modeling
Program-aware Modeling: Tree-based Approach
Program-aware Modeling: Neural Approach
Comparisons of Models
Unique Problem Characteristics
Transferable Cost Model
Impact of Transfer Learning
Learning to Optimize Tensor Programs
Device Fleet: Distributed Test Bed for AutoTVM
TVM: End to End Deep Learning Compiler
Tensor Expression and Optimization Search Space
Search Space for CPUs
Hardware-aware Search Space
Search Space for GPUs
Search Space for TPU-like Specialized Accelerators
Tensorization Challenge
Tensorization Challenge
Search Space for TPU-like Specialized Accelerators
Software Support for Latency Hiding
Summary: Hardware-aware Search Space
VTA: Open & Flexible Deep Learning Accelerator
TVM: End to End Deep Learning Compiler
Need for More Dynamism
Relay Virtual Machine
uTVM: TVM on bare-metal Devices
Core Infrastructure
TSIM: Support for Future Hardware
Unified Runtime For Heterogeneous Devices
Unified Runtime Benefit
Effectiveness of ML based Model
Comparisons of Models
Device Fleet in Action
End to End Inference Performance (Nvidia Titan X)
Portable Performance Across Hardware Platforms