Targeting the IPU from TensorFlow 1

• 1. Introduction
  • 1.1. Document overview
• 2. Tutorial
  • 2.1. Preliminary graphs
  • 2.2. A basic graph
    • 2.2.1. Selecting hardware to run on
    • 2.2.2. Running on the IPU Model simulator
  • 2.3. Compiling the graph for the IPU
  • 2.4. Sharding a graph
  • 2.5. Adding variables
    • 2.5.1. Troubleshooting
    • 2.5.2. Note on the global_step counter
• 3. Targeting the Poplar XLA device
  • 3.1. Supported types
  • 3.2. Device selection
  • 3.3. Configuring system options
    • 3.3.1. TF_POPLAR_FLAGS environment variable
  • 3.4. Supported operations
  • 3.5. Unsupported operations
  • 3.6. Error Handling
    • 3.6.1. Construction and compilation errors
    • 3.6.2. Runtime errors
• 4. Compiling and pre-compiling executables
  • 4.1. Caching of compiled executables
  • 4.2. Pre-compiling executables
    • 4.2.1. Unsupported Operations
• 5. Training a model
  • 5.1. Training loops, data sets and feed queues
  • 5.2. Accessing outfeed queue results during execution
  • 5.3. Replicated graphs
    • 5.3.1. Selecting the number of replicas
    • 5.3.2. Performing parameter updates
  • 5.4. Pipelined training
    • 5.4.1. Sequential scheduling
    • 5.4.2. Interleaved scheduling
    • 5.4.3. Grouped scheduling
    • 5.4.4. Pipeline stage inputs and outputs
    • 5.4.5. Applying an optimiser to the graph
    • 5.4.6. Device mapping
    • 5.4.7. Concurrent pipeline stages
  • 5.5. Gradient accumulation
    • 5.5.1. Optimizers
    • 5.5.2. Pipelining
    • 5.5.3. Accumulation data type
  • 5.6. Optimizer state offloading
  • 5.7. Dataset benchmarking
    • 5.7.1. Accessing the JSON data
• 6. Efficient IPU I/O
  • 6.1. Prefetch elements
  • 6.2. I/O Tiles
• 7. Example using IPUEstimator
• 8. Example using IPUPipelineEstimator
• 9. Distributed training
  • 9.1. Example using IPUMultiWorkerStrategy
    • 9.1.1. The input function
    • 9.1.2. The model function
    • 9.1.3. Cluster definition
    • 9.1.4. Complete example
  • 9.2. Distributed training with Horovod
  • 9.3. Launching Horovod training
  • 9.4. Complete Horovod example
• 10. Half-precision floating point and stochastic rounding
  • 10.1. Controlling the half-precision floating-point unit
  • 10.2. Resetting the global random number seed
  • 10.3. Debugging numerical issues
• 11. IPU-optimised operations
  • 11.1. LSTM and GRU
  • 11.2. Dropout
  • 11.3. Embedding lookup
  • 11.4. Group normalisation
  • 11.5. Instance normalisation
  • 11.6. Layer normalisation
  • 11.7. GeLU activation
  • 11.8. Sequence slice
  • 11.9. Histogram
• 12. IPU Outlined Functions
  • 12.1. Usage
  • 12.2. Examples
    • 12.2.1. Models with common structures
    • 12.2.2. Serializing large operations
• 13. Writing custom operations
  • 13.1. Custom operation on the IPU
    • 13.1.1. Building the Poplar graph
    • 13.1.2. Gradient builders
    • 13.1.3. Metadata
    • 13.1.4. Compiling the IPU code
      • API level
      • PopLibs library code
      • Compiling the library file
    • 13.1.5. Using the custom op in TensorFlow
    • 13.1.6. Tensor allocation
    • 13.1.7. Examples
      • In-place operations
      • Operation attributes
      • Custom codelet
  • 13.2. Custom host CPU operations
    • 13.2.1. Gradient callback
• 14. IPU host embeddings
  • 14.1. Usage
  • 14.2. Example
  • 14.3. Experimental functionality: IPU embeddings in remote buffers
    • 14.3.1. Partitioning strategies
      • Token strategy
      • Encoding strategy
      • Choosing a strategy for your application
• 15. Retrieving information about compilation and execution
  • 15.1. TensorFlow options for reporting
  • 15.2. Dumping auxiliary Poplar information
    • 15.2.1. Poplar vertex graph
    • 15.2.2. Poplar interval report
  • 15.3. XLA graph file naming
• 16. API changes
  • 16.1. Release 2.2
    • 16.1.1. Breaking changes
      • C++ Poplar TensorFlow libraries are private by default
      • Reports removed from ipu events
    • 16.1.2. Non-breaking changes
      • IPULoggingTensorHook replication_factor deprecated
      • IPUInfeedQueue/IPUOutfeedQueue/IPULoggingTensorHook feed_name deprecated
      • Change of output location for profiling information
      • IPU Keras Layers deprecation in TensorFlow 1.15
  • 16.2. Release 2.1
    • 16.2.1. Breaking changes
      • IPUPipelineEstimator change
      • Autosharding removed
      • Old IPU option configuration API changes
      • IPU Keras changes [TensorFlow 2]
    • 16.2.2. Non-breaking changes
      • Recompute suggestions deprecated
      • IPUInfeedQueue/IPUOutfeedQueue replication_factor deprecated
      • IPUInfeedQueue data_to_prefetch deprecated
      • IPUOutfeedQueue data_to_prefetch deprecated
      • CTC loss ops deprecated
      • New configuration API
      • Support for grouped collectives
      • Environment variable changes
  • 16.3. Release 2.0
    • 16.3.1. Breaking changes
    • 16.3.2. Non-breaking changes
      • IPUPipelineEstimator change
      • Autosharding deprecated
      • IPU config change
      • IPU Keras changes [TensorFlow 2]
• 17. Deprecated profiling functionality
  • 17.1. Adding an operation to get compilation and execution events
    • 17.1.1. ipu_event_trace()
    • 17.1.2. ipu_compile_summary(name, [op list])
  • 17.2. Enabling tracing in the hardware configuration options
  • 17.3. Extract the reports from the returned events
  • 17.4. Producing reports for use with the PopVision Graph Analyser
    • 17.4.1. COMPILE_BEGIN
    • 17.4.2. COMPILE_END
      • Tensor map
    • 17.4.3. EXECUTE
  • 17.5. Using the IPU Model device for debugging
  • 17.6. Reading the Poplar textual summary report
    • 17.6.1. Target
    • 17.6.2. Graph
    • 17.6.3. Memory usage
  • 17.7. Producing an ELF image of the compilation
• 18. Python API
  • 18.1. Operations and utilities related to the Graphcore IPU
  • 18.2. Compiler interface
  • 18.3. Scoping contexts
  • 18.4. Infeed queue
  • 18.5. Outfeed queue
  • 18.6. General utilities
  • 18.7. Configuration utilities
  • 18.8. Looping utilities
  • 18.9. Distributed training
  • 18.10. Horovod
  • 18.11. Datasets
    • 18.11.1. Dataset benchmarking
    • 18.11.2. Dataset wrappers
  • 18.12. Estimators
    • 18.12.1. IPUEstimator
    • 18.12.2. IPUPipelineEstimator
    • 18.12.3. Run configs
    • 18.12.4. Session run hooks
  • 18.13. Keras layers
    • 18.13.1. Keras layer specializations for the Graphcore IPU
  • 18.14. Operators
    • 18.14.1. Custom operations
    • 18.14.2. Functional operators
    • 18.14.3. Image operations
    • 18.14.4. Graphcore utility operations
    • 18.14.5. IPU specific maths operations
    • 18.14.6. Pipelining operators
    • 18.14.7. Popnn primitive neural network operators
    • 18.14.8. Popnn normalization operators
    • 18.14.9. Popnn recurrent neural network operators
    • 18.14.10. Popops all to all and all gather operators
    • 18.14.11. Popops cross replica operators
    • 18.14.12. Popops embedding operators
    • 18.14.13. Popops reduce scatter operator
    • 18.14.14. Poprand operators
    • 18.14.15. Utility operations to be used in replicated mode
    • 18.14.16. Slicing operators
    • 18.14.17. Statistics operators
    • 18.14.18. Summary operations for IPUs
  • 18.15. Optimisers
    • 18.15.1. Optimizer classes for the Graphcore IPU
  • 18.16. Sharding
    • 18.16.1. Utility functions for sharding graphs
• 19. TensorFlow operators supported by the IPU
• 20. Resources
  • 20.1. Graphcore
  • 20.2. TensorFlow
  • 20.3. Other
• 21. Index
• 22. Trademarks & copyright