* Update BERT inference instructions * Preserve order of operations in graph when folding batch norm ops in CNN inference * Various README typo and documentation fixes * Tidy PopART pipelining example In PopART pipelining code example - replace popart.SessionOptionsCore() with popart.SessionOptions() * Test fixesClearer accuracy assertion error in public example tests * Fixes to avoid warnings in kernel benchmarks run harness * Add a more advanced/complete Poplar example * Add a simple Tensorflow custom op example. * BERT Inference bug fixes
Advanced Poplar Example では、Poplar APIにて、Graphcore内のTileにマッピングするにはどうしているかも多少分かります。Custom Opを読み込んでいるが、このCuston Opについては、Creating a simple Tensorflow Custom Op にあるね。
This paper describes a high-level polyhedral compiler that takes a high-level algorithm description that can be written manually or extracted from a TensorFlow computation graph and generates input to the low-level C-based compiler.