← All resources

Introducing MiniDXNN: MLP library for DirectX 12

Auto-discovered

Blog Intermediate

Takahiro Harada, Sho Ikeda

Visit resource ↗

Description

Introducing MiniDXNN: MLP library for DirectX 12 is an AMD GPUOpen learn article by Takahiro Harada and Sho Ikeda announcing a native HLSL and DirectX 12 library for small MLP inference on Radeon matrix cores via cooperative vector APIs. It avoids HIP interop for graphics-centric engines.

MiniDXNN targets neural radiance caching, neural texture compression, and similar in-shader inference workloads with shipped kernels, unit tests, and a 2D image reconstruction sample project. The post outlines how developers bind weights and activate cooperative vector paths on supported hardware.

Intermediate DirectX 12 graphics programmers exploring hardware-accelerated neural features in rendering pipelines are the audience. Basic MLP layer math and HLSL compute familiarity is assumed.

The announcement is free on gpuopen.com with links to the GitHub repository. GPUOpen neural inference samples are standard starting points for AMD-focused DX12 neural rendering experiments.