GitHub - Devsh-Graphics-Programming/Nabla: Vulkan, OptiX and CUDA Interoperation Modular Rendering Library and Framework for PC/Linux/Android
Table of Contents
Physical Device Selection and Filtering
SPIR-V and Vulkan as First-Class Citizens
Nabla Event Handler: Seamless GPU-CPU Synchronization
HLSL2021 Standard Template Library
Full Embrace of Buffer Device Address and Descriptor Indexing
Cancellable Future-based Async I/O
Unit-Tested BxDFs for Physically Based Rendering
Property Pools (GPU Entity Component System)
SPIR-V Introspection and Layout Creation
Showcase
Nabla Shader Compiler & Godbolt docker integration
ImGUI render backend & extensions
Features
🧩 The Nabla Core Profile
Nabla exposes
a curated set of Vulkan extensions and features
compatible across the GPUs we aim to support on Windows, Linux, (coming soon MacOS, iOS as well as Android)
Vulkan evolves fast—just when you think you've figured out
, you realize there's
. Keeping up with new extensions, best practices, and hardware quirks is exhausting. Instead of digging through
or
, Nabla gives you a well-thought-out set of extensions—so you can focus on what you want to achieve, not get stuck in an eternal loop of:
mastering a feature
finding out about a new feature
assesing whether obsoletes or just adds the one you've just mastered
working if the feature is ubiquitous on the devices you target
rewriting what you've just polished
🧩 Physical Device Selection and Filteration
Nabla allows you to select the best GPU for your compute or graphics workload.
voidfilterDevices(core::set<video::IPhysicalDevice*>& physicalDevices) { nbl::video::SPhysicalDeviceFilter deviceFilter = {}; deviceFilter.minApiVersion = { 1,3,0 }; deviceFilter.minConformanceVersion = {1,3,0,0}; deviceFilter.requiredFeatures.rayQuery = true; deviceFilter(physicalDevices); }🧩 SPIR-V and Vulkan as First-Class Citizens
Nabla treats SPIR-V and Vulkan as the preferred, reference standard—everything else is built around them, with all other backends adapting to them.
🧩 Integration of Renderdoc
Built-in support for capturing frames and debugging with
. This is how one debugs headless or async GPU workloads that are not directly involved in producing a swapchain frame to be captured by Renderdoc.
const IQueue::SSubmitInfo submitInfo = { .waitSemaphores = {}, .commandBuffers = {&cmdbufInfo,1}, .signalSemaphores = {&signalInfo,1} }; m_api->startCapture(); // Start Renderdoc Capture queue->submit({&submitInfo,1}); m_api->endCapture(); // End Renderdoc Capture🧩 Nabla Event Handler: Seamless GPU-CPU Synchronization
Nabla Event Handler's extensive usage of
enables CPU Callbacks on GPU conditions.
You can enqueue callbacks that trigger upon submission completion (workload finish), enabling amongst others, async readback of submission side effects, or deallocating an allocation after a workload is finished.
// This doesn't actually free the memory from the pool, the memory is queued up to be freed only after the `scratchSemaphore` reaches a value a future submit will signal memory_pool->deallocate(&offset,&size,nextSubmit.getFutureScratchSemaphore());🧩 GPU Object Lifecycle Tracking
Nabla uses
to track the lifecycle of GPU objects. Descriptor sets and command buffers are responsible for maintaining reference counts on the resources (e.g., buffers, textures) they use. The queue itself also tracks command buffers, ensuring that objects remain alive as long as they are pending execution. This system guarantees the correct order of deletion and makes it difficult for GPU objects to go out of scope and be destroyed before the GPU has finished using them.
🧩 HLSL2021 Standard Template Library
🔄 Reusable: Unified single-source C++/HLSL libraries eliminate code duplication with reimplementation of STL's type_traits, limits, functional, tgmath, etc.
🐞 Shader Logic, CPU-Tested: A subset of HLSL compiles as both C++ and SPIR-V, enabling CPU-side debugging of GPU logic, ensuring correctness in complex tasks like FFT, Prefix Sum, etc. (See our examples:
,
)
🔮 Future-Proof: C++20
in HLSL enable safe and documented polymorphism.
🧠 Insane: Boost Preprocessor and Template Metaprogramming in HLSL!
🛠️ Real-World Problem Solvers: The library offers GPU-optimized solutions for tasks like Prefix Sum, Binary Search, FFT, Global Sort, and even emulated shaderFloat64 when native GPU support is unavailable!
🎤 Talks from us:
Vulkanised 2024: Beyond SPIR-V: Single Source C++ and Shader Programming
Vulkanised 2023: HLSL202x like its C++, building an std:: like Library
🧩 Full Embrace of
and
By utilizing Buffer Device Addresses (BDAs), Nabla enables more direct access to memory through 64-bit GPU virtual addresses. Synergized with Descriptor Indexing, this approach enhances flexibility by enabling more dynamic, scalable resource binding without relying on traditional descriptor sets.
🧩 Minimally Invasive Design
No Singletons, No Main Thread—Nabla allows multiple instances of every object (including Vulkan devices) without assuming a main thread or thread-local contexts. Thread-agnostic by design, it avoids global state and explicitly passes contexts for easy multithreading.
Nabla's minimally invasive and flexible design with api handle acquisitions and multi-window support make it ideal for custom rendering setups and low-level GPU programming without unnecessary constraints such as assuming a main thread or a single window.
Even Win32 windowing is wrapped for use across multiple threads, breaking free traditional single-thread limitations.
This allows simpler porting of legacy OpenGL and DirectX applications.
🧩 Designed for Interoperation
Nabla is built with interoperation in mind, supporting memory export and import between different compute and graphics APIs.
🧩 Cancellable Future based Async I/O
File I/O is fully asynchronous, using
, a cancellable MPSC circular buffer-based future implementation.
Requests start in a PENDING state and can be invalidated before execution if needed. This enables efficient async file reads and GPU memory writes, ensuring non-blocking execution:
ISystem::future_t<size_t> bytesActuallyWritten; file->read(bytesActuallyWritten, gpuMemory->getMappedPointer(), offsetInFile, 2*1024*1024*1024); while (!bytesActuallyWritten.ready()) { /* Do other work */ }🧩 Data Transfer Utilities
Nabla's
streamlines the process of pushing/pulling arbitrary-sized buffers and images with fixed staging memory to/from the GPU, ensuring seamless data transfers. The system automatically handles submission when buffer memory overflows, while
during upload to handle color format conversions. By leveraging device-specific properties, the system respects alignment limits and ensures deterministic behavior. The user only provides initial submission info through
, and the utility manages subsequent submissions automatically.
Learn more: 🎤 Our Talk at Vulkanised:
Vulkanised 2023: Keeping your staging buffer fixed size!
📚 Our Blog post:
Uploading Textures to GPU - The Good Way
🧩 Virtual File System
Nabla provides a [unified Virtual File System] (
) that supports mounting archives and folders under different virtual paths. This enables access to both external and embedded assets while preserving original relative paths.
For embedding, we provide an alternative to C++23's #embed, which allows embedding files directly into compiled binaries. Instead of relying on compiler support, we use Python + CMake to generate what we call built-in resource archives—packing files (e.g., images, shaders, .obj, .mtl, .dds) into DLLs as memory-mapped
objects ensuring that dependent assets (e.g., models and their textures) retain their correct relative paths even when embedded.
The embedding process:
At build time, Python reads an input path table (generated by CMake).
It serializes files into constexpr arrays with metadata (key + timestamps).
The output C++ source + header define a built-in resource library, linked into Nabla or examples.
This approach keeps assets self-contained, making file access efficient while maintaining asset dependencies.
🧩 Asset System
The asset system in Nabla maintains a 1:1 mapping between CPU and GPU representations, where every CPU asset has a direct GPU counterpart. The system also allows for coordination between loaders—for instance, the OBJ loader can trigger the MTL loader, and the MTL loader in turn invokes image loaders, ensuring smooth asset dependency management.
🧩 Asset Converter (CPU to GPU)
The Asset Converter transforms CPU objects (asset::IAsset) into GPU objects (video::IBackendObject) while eliminating duplicates with Merkle Trees. Instead of relying on pointer comparisons, it hashes asset contents to detect and reuse identical GPU objects.
🧩 Unit-Tested BxDFs for Physically Based Rendering
A statically polymorphic library for defining Bidirectional Scattering Distribution Functions (BxDFs) in HLSL and C++. Each BxDF is rigorously unit-tested in C++ as well as HLSL. This is part of Nabla’s HLSL-C++ compatible library.
Snippet of our
:
TestJacobian<bxdf::reflection::SLambertianBxDF<sample_t, iso_interaction, aniso_interaction, spectral_t>>::run(initparams, cb); TestJacobian<bxdf::reflection::SOrenNayarBxDF<sample_t, iso_interaction, aniso_interaction, spectral_t>>::run(initparams, cb); TestJacobian<bxdf::reflection::SBeckmannBxDF<sample_t, iso_cache, aniso_cache, spectral_t>, false>::run(initparams, cb); TestJacobian<bxdf::reflection::SBeckmannBxDF<sample_t, iso_cache, aniso_cache, spectral_t>, true>::run(initparams, cb); TestJacobian<bxdf::reflection::SGGXBxDF<sample_t, iso_cache, aniso_cache, spectral_t>, false>::run(initparams, cb); TestJacobian<bxdf::reflection::SGGXBxDF<sample_t, iso_cache, aniso_cache, spectral_t>,true>::run(initparams, cb); TestJacobian<bxdf::transmission::SLambertianBxDF<sample_t, iso_interaction, aniso_interaction, spectral_t>>::run(initparams, cb); TestJacobian<bxdf::transmission::SSmoothDielectricBxDF<sample_t, iso_cache, aniso_cache, spectral_t>>::run(initparams, cb); TestJacobian<bxdf::transmission::SSmoothDielectricBxDF<sample_t, iso_cache, aniso_cache, spectral_t, true>>::run(initparams, cb); TestJacobian<bxdf::transmission::SBeckmannDielectricBxDF<sample_t, iso_cache, aniso_cache, spectral_t>, false>::run(initparams, cb); TestJacobian<bxdf::transmission::SBeckmannDielectricBxDF<sample_t, iso_cache, aniso_cache, spectral_t>, true>::run(initparams, cb); TestJacobian<bxdf::transmission::SGGXDielectricBxDF<sample_t, iso_cache, aniso_cache, spectral_t>, false>::run(initparams, cb); TestJacobian<bxdf::transmission::SGGXDielectricBxDF<sample_t, iso_cache, aniso_cache, spectral_t>,true>::run(initparams, cb);🔧 In Progress: Property Pools (GPU Entity Component System)
Property Pools group related properties together in a Structure Of Arrays (SoA) manner, allowing efficient, cache-friendly access to data on the GPU. The system enables transferring properties (Components) between the CPU and GPU, with the PropertyPoolHandler managing scattered updates with a special compute shader. Handles are assigned for each object and remain constant as data is added or removed.
🧩 SPIR-V Introspection and Layout Creation
SPIR-V introspection in Nabla eliminates most of the boilerplate code required to set up descriptor and pipeline layouts, simplifying resource binding to shaders.
🧩 Nabla Extensions
– MultiDrawIndirect based and draws in as little as a single drawcall.
Fast Fourier Transform Extension
– for image processing and all kind of frequncy-domain fun.
– Efficient parallel prefix sum computation.
– Optimized GPU-based image blurring.
– High-performance, GPU-accelerated sorting algorithm.
[WIP] Autoexposure – Adaptive brightness adjustment for HDR rendering.
[WIP] Tonemapping
[WIP] GPU MPMC Queue – Multi-producer, multi-consumer GPU queue.
[WIP] OptiX interoperability for ray tracing.
[WIP] Global Scan – High-speed parallel scanning across large datasets.
🚀 Coming Soon
Full CUDA interoperability support.
Scene Loaders
GPU-Driven Scene Graph
Material Compiler 2.0 for efficient scheduling of BxDF graph evaluation
🤝 Need Our Expertise?
We specialize in:
High-performance computing and performance optimization
Path Tracing and Physically Based Rendering
CAD Rendering
Audio Programming and Digital Signal Processing
Porting and Optimizing legacy Renderers
Graphics and Compute APIs: Vulkan, D3D12, CUDA, OpenCL, WebGPU, D3D11, OpenGL
Whether you're optimizing your renderer or compute workloads, looking to port your legacy renderer, or integrating complex visual effects into your product, our team can help you. As a specialized team, we're constantly learning, evolving, and discussing matters with each other.
brings unique insights to the table, ensuring we approach every project from multiple angles to achieve the best possible solution.
Our primary language is C++20, but we also work with C#, Java, Python, and other related technologies.
If you're already here reading this, We want to hear from you and learn more about what you're building.
Contact us at
.
The members of Devsh Graphics Programming Sp. z O.O. (Company Registration (KRS) #: 0000764661) are available (individually or collectively) for contracts on projects of various scopes and timescales.
———
👥 Join Our Team
💼 Jobs
If you are:
a programmer with a passion for High Performance Computing, Mathematics and Computer Graphics
able to be in charge of your own time management and work 4-day work weeks fully remotely
Then:
Make something impressive using Nabla.
Open a Pull Request
Contact us either on discord or through the following e-mail address
You may check out our
for contribution ideas
🎓 Internships
In case of internships contact us with:
CV in English
A neat description of any paperwork you'll need (schools/universities). Most importantly contact us at least 5 weeks in advance of your intenship organisation deadline!
[TODO]: also link to achievements, personal blogs, websites, linkedin and presentations of each member