Huawei Open-Sources SimpleGPULayer: Full-Scenario GPU Acceleration for HarmonyOS

Huawei's HarmonyOS development team has officially open-sourced SimpleGPULayer (SGL) on GitCode, a high-performance GPU acceleration framework built on Vulkan that targets HarmonyOS native applications. SGL encapsulates complex GPU pipeline details — device initialization, memory management, pipeline creation, command submission — into concise APIs, allowing developers to integrate GPU acceleration with only a few lines of code.

Four Core Capabilities

1. Image Processing: 52 Filters + Large-Image Solution + Layer Blending

SGL ships with 52 built-in GPU image filters covering blur, artistic styles, color adjustment, and geometric deformation. It supports large-radius real-time preview. For ultra-large image editing, a tiled processing solution controls video memory usage while avoiding boundary artifacts. An asynchronous prefetch mechanism enables parallel CPU reading and GPU computation, significantly improving large-image editing responsiveness. Additionally, SGL provides 4 layer blending modes (multiply, overlay, dissolve, lighten) for multi-layer editing, material overlay, and complex visual compositing.

2. AI Inference Computing: 28 Operators Covering Mainstream Model Pre/Post-Processing

The ComputeEngine module provides 28 compute operators : 22 GPU operators and 6 CPU reduction operators . These cover matrix operations, activation functions, normalization, positional encoding, element-wise operations, and reduction statistics. They efficiently support on-device AI model pre/post-processing and general GPU compute scenarios. The repository includes an LLM inference demo demonstrating how to quickly build a Transformer inference pipeline using SGL.

3. 2D/3D Rendering: Off-Screen Rendering + Model Loading

The Renderer module supports 3D mesh rendering and 3D Gaussian splatting rendering . It can load mainstream 3D formats such as glTF and includes a complete PBR texture pipeline . Off-screen rendering is supported, allowing render results to be output directly as images — suitable for server-side rendering and automated processing scenarios. This dramatically lowers the development and integration barrier for 2D/3D rendering on HarmonyOS.

4. Vector Graphics Generation: GPU Bezier Thick Line Generation

SGL provides GPUBezierThickLineGenerator, which moves Bezier curve generation to GPU compute shaders for parallel execution. Developers only need to pass control points, start/end widths, and endpoint sampling parameters; the GPU compute shader quickly generates triangle-strip vertex sets for tapered thick lines. This capability has been validated in office software for node-to-node tapered connection lines, showing significant performance gains over traditional CPU per-segment calculation in dense-node, frequent-drag, real-time canvas refresh scenarios.

Minimal Integration: Three Lines for a Gaussian Blur

SGL deeply encapsulates the complex GPU pipeline. Developers only prepare input image, output image, and filter parameters. Example for Gaussian blur:

const auto filter = std::make_shared<FastGaussianBlurFilter>();
filter->SetRadius(120);
sgl::image::ImageEngine::Process(inputImage, outputImage, filter);

For HarmonyOS app integration, SGL provides C API and NAPI interfaces , allowing developers to embed it as a native acceleration library and expose business interfaces to the ArkTS/JS layer with minimal integration cost.

Open-Source Collaboration & Repository

The project is led by the HarmonyOS Strike Team, with image processing filters co-built by the Wukong Image partner, and Bezier tapered-line capability jointly validated with an office software vendor. SGL is already running stably in production apps including Wukong Image. The open-source repository is available at:

https://gitcode.com/OpenHarmonyPerfSquad/SGL

Developers can also search "SimpleGPULayer" or "SGL" on GitCode to obtain the source code. The team welcomes community contributions to continuously expand the framework's capabilities.