v0.3.0

MiniDXNN v0.3.0

Upgraded MiniDXNN from Cooperative Vector (Shader Model 6.9) to LinAlg Matrix (Shader Model 6.10), with input encoding support and a new texture compression example.

Features

• HLSL MLP with LinAlg Matrix (include/minidxnn/hlsl/mlp.hlsl): Forward and backward passes using Shader Model 6.10 LinAlg Matrix operations
• Texture compression example (example/03_texture_compression_with_input_encoding): MLP training with positional and grid input encoding for texture compression
• Matrix conversion: Host-side matrix format conversion via GetLinearAlgebraMatrixConversionDestinationInfo / ConvertLinearAlgebraMatrix
• C++ fallback (include/minidxnn/cpp/hlsl_compat.hpp): Updated CPU execution path compatible with LinAlg Matrix API

Changes from v0.2.0

• Migrated from Cooperative Vector (SM 6.9) to LinAlg Matrix (SM 6.10)
• Added 03_texture_compression_with_input_encoding example with positional and grid encoding
• Added matrix conversion utilities for host-side weight format conversion
• Added LinAlg Matrix MLP guide (docs/linalg_matrix_mlp.md)
• Updated Agility SDK to 1.720-preview, DXC to v1.10.2605.2
• Updated examples and unit tests for SM 6.10 API

Requirements

• Windows 11 with Developer Mode
• GPU supports Shader Model 6.10 and LinAlg Matrix in D3D12 (AMD Radeon™ RX 9000 Series GPUs or equivalent NVIDIA)
• CMake 3.21+, Visual Studio 2022 (C++20)
• Agility SDK 1.720-preview, DXC v1.10.2605.2
• Python 3.8+ with PyTorch (optional, for reference training)

License

MIT — Copyright (c) 2026 Advanced Micro Devices, Inc.

v0.2.0

MiniDXNN v0.2.0

MLP training support for MiniDXNN — GPU-accelerated forward and backward passes using DirectX 12 Cooperative Vector, with a C++ fallback path.

Features

• HLSL MLP training (include/minidxnn/hlsl/mlp.hlsl): Forward and backward passes with mininn::forward() and mininn::backward()
• Texture training example (example/02_texture_training): End-to-end MLP training on GPU with SGD, Adam, and Lion optimizers
• C++ fallback (include/minidxnn/cpp/hlsl_compat.hpp): CPU execution path that compiles mlp.hlsl as C++ for environments without Cooperative Vector support

Changes from v0.1.0

• Added MLP backward pass in HLSL
• Added 02_texture_training example with GPU training pipeline
• Added C++ fallback infrastructure (hlsl_compat.hpp)
• Added training-related unit tests (atomic operations, MLP training)
• Restructured HLSL header path from include/hlsl/ to include/minidxnn/hlsl/
• Updated documentation and README

Requirements

• Windows 11 with Developer Mode
• GPU supports Shader Model 6.9 and Cooperative Vector in D3D12 (AMD Radeon™ RX 9000 Series GPUs or equivalent NVIDIA)
• CMake 3.21+, Visual Studio 2022 (C++20)
• Agility SDK 1.717.1-preview, DXC v1.8.2505.1
• Python 3.8+ with PyTorch (optional, for reference training)

License

MIT — Copyright (c) 2026 Advanced Micro Devices, Inc.

v0.1.0

MiniDXNN v0.1.0

Initial release of MiniDXNN — a header-only HLSL library for GPU-accelerated MLP inference using DirectX 12 Cooperative Vector.

Features

• HLSL MLP inference (include/hlsl/mlp.hlsl): Configurable forward pass with mininn::forward()
• Texture inference example: Train with PyTorch, export weights, run GPU inference

Requirements

• Windows 10/11 with Developer Mode
• GPU supports Shader Model 6.9 and Cooperative Vector in D3D12 (AMD Radeon™ RX 9000 Series GPUs or equivalent NVIDIA)
• CMake 3.21+, Visual Studio 2022 (C++20)
• Agility SDK 1.717.1-preview, DXC v1.8.2505.1

License

MIT — Copyright (c) 2026 Advanced Micro Devices, Inc.