MIGraphX v0.6

Notes

MIGraphX 0.6 is primarily a bugfix release. The release includes:

• Added a C API in addition to existing C++
• Support multiple outputs
• Additional operators including ArgMax, ArgMin, Ceil, ConvTranspose, InstanceNormalization, ReduceL1, ReduceL2, ReduceLogSum, ReduceProd, ReduceSum, ReduceSumSquare, Shape

MIGraphX v0.5

Notes

MIGraphX 0.5 is primarily a performance and bugfix release. The release includes:

• Additional operators: split, ceil, floor
• Support for additional models including NASNet-a_Large for Tensorflow
• Simplified python interface
• Performance improvements
• A driver for exercising migraphx

See https://github.com/ROCmSoftwarePlatform/AMDMIGraphX/wiki/Getting-started:-using-the-new-features-of-MIGraphX-0.5 for more getting started information

MIGraphX v0.4

Notes

Updated release

• Additional operators support for new models and features
• Quantization support for fp16 and int8
• Support for NLP models, particularly BERT with both Tensorflow and ONNX examples

See https://github.com/ROCmSoftwarePlatform/AMDMIGraphX/wiki/Getting-started:-using-the-new-features-of-MIGraphX-0.4 for more getting started information

MIGraphX v0.3

Notes

Updated release

• Tensorflow Support
• Quantization Part 1 (fp16)
  See https://github.com/ROCmSoftwarePlatform/AMDMIGraphX/wiki/Getting-started:-using-the-new-features-of-MIGraphX-0.3 for more getting started information

MIGraphX v0.2

Notes

Updated release

• New Python API
• Support for additional ONNX operators and fixes that now enable a large set of Imagenet models
• Support for RNN, LSTM and GRU operators
• Experimental support for Tensorflow frozen protobuf
  See https://github.com/ROCmSoftwarePlatform/AMDMIGraphX/wiki/Getting-started:-using-the-new-features-of-MIGraphX-0.2 for more getting started information.

MIGraphX v0.1

Notes

Initial release of MIGraphX library that includes basic functionality:

• Reading models saved in ONNX file format
• Performing basic optimizations including
  • graph transformations, e.g. rewriting or combining nodes
  • fusion of operators
  • memory coloring to reduce memory usage
• Executing resultant models on AMD GPUs including
  • support for fp16, float32, ...
  • calls to optimized MIOpen library routines
  • generated code