PopART User Guide

• 1. Introduction
• 2. Importing graphs
  • 2.1. Creating a session
  • 2.2. Session control options
  • 2.3. Executing an imported graph
• 3. Building graphs in PopART
  • 3.1. Adding operations to the graph
  • 3.2. Adding parameters to the graph
  • 3.3. Setting outputs
  • 3.4. Setting the IPU number for operations
• 4. Executing graphs
  • 4.1. Setting input and output data buffers for an execution
    • 4.1.1. Retrieving results
  • 4.2. Session options
    • 4.2.1. Stochastic rounding
  • 4.3. Selecting a device for execution
  • 4.4. Executing a session
  • 4.5. Saving and loading a model
  • 4.6. Retrieving profiling reports
  • 4.7. Turning on execution tracing
    • 4.7.1. Programming interface
    • 4.7.2. Output format
  • 4.8. Errors
    • 4.8.1. Application errors
    • 4.8.2. System errors
• 5. Performance optimisation
  • 5.1. Pipelined execution
  • 5.2. Graph replication
    • 5.2.1. Local replication
    • 5.2.2. Global replication
  • 5.3. Sync configuration
    • 5.3.1. Sync patterns
• 6. Grouping graph replicas
  • 6.1. Concept
  • 6.2. VariableSettings
  • 6.3. Instantiating variables with VariableSettings
  • 6.4. Weight input/output
  • 6.5. ONNX checkpoints
• 7. Supported operators
  • 7.1. Domain: ai.graphcore
  • 7.2. Domain: ai.onnx
  • 7.3. Limitations
    • 7.3.1. Limitations of ai.onnx operators
    • 7.3.2. Limitations of ai.graphcore operators
  • 7.4. Converting ONNX models with opset versions not supported by PopART
• 8. Custom operators
  • 8.1. Overview
    • 8.1.1. Custom op classes
  • 8.2. Implementing a custom op
    • 8.2.1. The Op class
      • LeakyReluOp example
    • 8.2.2. The grad op class
    • 8.2.3. The Opx class
    • 8.2.4. The grad opx class
  • 8.3. Making the op available to PopART
    • 8.3.1. Define the op identifier
    • 8.3.2. Define the Op creator
    • 8.3.3. Define the Opx creator
  • 8.4. ONNX schema and shape inference
  • 8.5. Compiling and using the custom op
    • 8.5.1. Compiling the code
    • 8.5.2. Using the compiled custom op in an application
• 9. Environment variables
  • 9.1. Logging
    • 9.1.1. POPART_LOG_LEVEL
    • 9.1.2. POPART_LOG_DEST
    • 9.1.3. POPART_LOG_CONFIG
  • 9.2. Generating DOT files
    • 9.2.1. POPART_DOT_CHECKS
  • 9.3. Caching of compiled executables
  • 9.4. Inspecting the IR
    • 9.4.1. POPART_IR_DUMP
  • 9.5. Manipulating outliner behaviour for PopLiner
    • 9.5.1. POPART_POPLINER_OUTLINER_REGEX
• 10. Overlapping IO with compute
  • 10.1. Configuring IO tiles
  • 10.2. Specify overlap strategy for inputs
  • 10.3. Specify overlap strategy for anchors
  • 10.4. Exchange strategies
• 11. Tensor locations
  • 11.1. Streaming Memory
  • 11.2. Replicated tensor sharding
  • 11.3. RTS sharding domains and distributed instances
• 12. PopART Python API
  • 12.1. Sessions
    • 12.1.1. Training session
    • 12.1.2. Inference session
    • 12.1.3. Session Options
  • 12.2. Data input and output
  • 12.3. Tensors
  • 12.4. Optimizers
    • 12.4.1. SGD
    • 12.4.2. ConstSGD
    • 12.4.3. Adam
    • 12.4.4. AdaDelta, RMSProp & AdaGrad
  • 12.5. Builder
    • 12.5.1. AiGraphcoreOpset1
  • 12.6. Data flow
  • 12.7. Device manager
  • 12.8. Ops
    • 12.8.1. Op definition for PopART IR
  • 12.9. Patterns
  • 12.10. Utility classes
    • 12.10.1. Writer
    • 12.10.2. Error handling
    • 12.10.3. Debug context
    • 12.10.4. Input shape information
    • 12.10.5. Type definitions
    • 12.10.6. Enums
• 13. PopART C++ API
  • 13.1. Sessions
    • 13.1.1. Training session
    • 13.1.2. Inference session
    • 13.1.3. Session options
  • 13.2. Data input and output (IStepIO)
  • 13.3. Tensors
  • 13.4. Optimizers
    • 13.4.1. Stochastic Gradient Descent (SGD)
    • 13.4.2. Adam, AdaMax & Lamb
    • 13.4.3. AdaDelta, RMSProp & AdaGrad
  • 13.5. Builder
  • 13.6. Data flow
  • 13.7. Device manager
  • 13.8. Ops
    • 13.8.1. Op definition for PopART IR
    • 13.8.2. Op definition for Poplar implementation
    • 13.8.3. Available Ops (Op class)
    • 13.8.4. Available Ops (Opx class)
  • 13.9. Patterns
    • 13.9.1. Available patterns
  • 13.10. Transforms
    • 13.10.1. Available transforms
  • 13.11. Utility classes
    • 13.11.1. Graph
    • 13.11.2. Region
    • 13.11.3. Error handling
    • 13.11.4. Debug context
    • 13.11.5. Attributes
    • 13.11.6. Void data
    • 13.11.7. Input shape information
    • 13.11.8. Profiling
    • 13.11.9. Task information
    • 13.11.10. Type definitions
    • 13.11.11. Enums
    • 13.11.12. Structs
    • 13.11.13. Other classes
• 14. Glossary
• 15. References
  • 15.1. Documentation
  • 15.2. Running applications in Docker
  • 15.3. Tutorials and simple examples
  • 15.4. Application and code examples
  • 15.5. Other Support
• 16. Index
• 17. Trademarks & copyright
• Doxygen