• RadeonML.h - main C API
• RadeonML.hpp - main and graph manipulation C++ API
• RadeonML_cl.h - OpenCL interoperation C API
• RadeonML_cl.hpp - OpenCL interoperation C++ API
• RadeonML_d3d12.h - Direct3D 12 interoperation C API
• RadeonML_d3d12.hpp - Direct3D 12 interoperation C++ API
• RadeonML_dml.h - DirectML interoperation C API
• RadeonML_dml.hpp - DirectML interoperation C++ API
• RadeonML_graph.h - graph manipulation C API
• RadeonML_miopen.h - MIOpen interoperation C API
• RadeonML_miopen.hpp - MIOpen interoperation C++ API
• RadeonML_mtl.h - Metal interoperation C API
• RadeonML_mtl.hpp - Metal interoperation C++ API

• Loading a model from a file (C)
• Loading a model from a file (C++)
• graph operation example (C++)
• graph operation example (C)

• vgg16.onnx, vgg19.onnx
  • Supported backends: DirectML
  • Input format: NCHW (Nx3x224x224)
• vgg16_opset11.onnx, vgg19_opset11.onnx
  • Supported backends: DirectML
  • Input format: NHWC (Nx224x224x3)
• mobilenet_v1_1.0_224.onnx, mobilenet_v2_1.0_224.onnx
  • Supported backends: DirectML
  • Input format: NCHW (Nx3x224x224)
• mobilenet_v1_opset11.onnx, mobilenet_v2_opset11.onnx
  • Supported backends: DirectML
  • Input format: NHWC (Nx224x224x3)
• resnet18v1.onnx, resnet34v1.onnx, resnet50v1.onnx, resnet101v1.onnx, resnet152v1.onnx
• resnet18v2.onnx, resnet34v2.onnx, resnet50v2.onnx, resnet101v2.onnx, resnet152v2.onnx
  • Supported backends: DirectML
  • Input format: NCHW (Nx3x224x224)
• inception_v1_opset8.onnx, inception_v2_opset8.onnx
  • Supported backends: DirectML
  • Input format: NCHW (Nx3x224x224)
• inception_v1_opset11.onnx, inception_v2_opset11.onnx
  • Supported backends: DirectML
  • Input format: NHWC (Nx224x224x3)
• inception_v3_opset11.onnx, inception_v4_opset11.onnx
  • Supported backends: DirectML
  • Input format: NHWC (Nx299x299x3)
• denoise_c3_ldr.onnx, denoise_c3_hdr.onnx
  • Supported backends: DirectML, MIOpen, MPS
  • Input format: NHWC (NxHxWx3)
• denoise_c9_ldr.onnx, denoise_c9_hdr.onnx
  • Supported backends: DirectML, MIOpen, MPS
  • Input format: NHWC (NxHxWx9)
• upscale2x_c3_rt.onnx, upscale2x_c3_esrgan_small.onnx
  • Supported backends: DirectML, MIOpen, MPS
  • Input format: NHWC

To inference supported models just substitute 'path/model', 'path/input' and 'path/output' with correct paths in load_model sample. Additional information about supported models: https://github.com/onnx/models

• Windows 10 19H1 or later (for DirectML backend)
• Windows 7 uses the MIOpen backend
• Ubuntu 18.04
• CentOS/RHEL 7.6, 7.7
• OSX Mojave and Catalina

• ONNX support (opset 6-11)
• TensforFlow frozen graph pb files
• FP32 and FP16 for ONNX

For more information, see documentation at this link https://radeon-pro.github.io/RadeonProRenderDocs/rml/about.html

• Windows DirectML supports our denoisers, upscalers and common models like resnet, VGG etc..

• Miopen backend for Windows and Linux only supports our denoisers and upscalers. When creating a RML context if DML is not supported we will fallback automatically to MIOpen

• MPS backend only supports our denoisers

• Model supported by the different backend

Others models may work as they will have similar operators, but we haven't checked them

• A device and command queue can be passed when creating a RML context. We support both async compute queue(D3D12_COMMAND_LIST_TYPE_COMPUTE) and the default command queue(D3D12_COMMAND_LIST_TYPE_DIRECT). Compute queue are preferred as it will run asynchronously with any graphics job. If no queue are passed, RML will create a compute queue

You will need CMake 3.10 at least to build the sample.

The input must contain contiguous data of a tensor with specified dimensions. The input (.bin files) in the repo don't necessarily represent real data at the moment, but just show to format the data

We aim at providing the same level of feature for every back end and will provide updates monthly for that