This is a low-abstraction, high-performance Vulkan API with interfaces for JavaScript and TypeScript.
nvk comes with pre-built N-API binaries for the following platforms:
OS | Status |
---|---|
Windows | ✔ |
Linux | ✔ |
MacOS | ✔ |
Real-Time RTX Ray Tracer |
---|
- Vulkan is a binding friendly API
- Less overhead than WebGL/OpenGL
- Essential features like Compute, Geometry and Tesselation shaders
- Support for Real-Time Ray Tracing, Mesh shaders, ...
- Supports Multithreading
- Low-level memory control using ArrayBuffers
This project is a thin layer on top of native Vulkan, built with simplicity and performance in mind. Native memory for Vulkan gets constructed entirely within JavaScript to reduce trampolining overhead.
Bounding checks and type validations are enabled by default, but can be disabled using the --disable-validation-checks
flag.
- Installation
- Example
- TypeScript
- Syntactic Sugar
- Project Structure
- Binding Code Generator
- Linking
- Build Instructions
- CLI
- RenderDoc
- TODOs
npm install nvk
In most cases the bindings match the C99 style of Vulkan. This allows you to follow existing C/C++ tutorials, but write the implementation itself with nvk. Note that both interfaces end up with a similar amount of code. Optionally you can use some syntactic sugar to write things quicker.
JavaScript/TypeScript:
let instance = new VkInstance();
let appInfo = new VkApplicationInfo();
appInfo.sType = VK_STRUCTURE_TYPE_APPLICATION_INFO;
appInfo.pApplicationName = "App";
appInfo.applicationVersion = VK_MAKE_VERSION(1, 0, 0);
appInfo.pEngineName = "Engine";
appInfo.engineVersion = VK_MAKE_VERSION(1, 0, 0);
appInfo.apiVersion = VK_API_VERSION_1_0;
let validationLayers = [
"VK_LAYER_LUNARG_standard_validation"
];
let instanceInfo = new VkInstanceCreateInfo();
instanceInfo.sType = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO;
instanceInfo.pApplicationInfo = appInfo;
instanceInfo.ppEnabledLayerNames = validationLayers;
instanceInfo.enabledLayerCount = validationLayers.length;
vkCreateInstance(instanceInfo, null, instance);
C++:
VkInstance instance;
VkApplicationInfo appInfo = {};
appInfo.sType = VK_STRUCTURE_TYPE_APPLICATION_INFO;
appInfo.pApplicationName = "App";
appInfo.applicationVersion = VK_MAKE_VERSION(1, 0, 0);
appInfo.pEngineName = "Engine";
appInfo.engineVersion = VK_MAKE_VERSION(1, 0, 0);
appInfo.apiVersion = VK_API_VERSION_1_0;
const std::vector<const char*> validationLayers = {
"VK_LAYER_LUNARG_standard_validation"
};
VkInstanceCreateInfo instanceInfo = {};
instanceInfo.sType = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO;
instanceInfo.pApplicationInfo = &appInfo;
instanceInfo.ppEnabledLayerNames = validationLayers.data();
instanceInfo.enabledLayerCount = static_cast<uint32_t>(validationLayers.size());
vkCreateInstance(&instanceInfo, nullptr, &instance);
To use the TypeScript definition file, simply follow the installation steps above or use this example as a reference. Afterwards in your .ts
file, import and use nvk as follows:
import * as nvk from "nvk";
Object.assign(global, nvk);
let win = new VulkanWindow({
width: 480,
height: 320,
title: "typescript-example"
});
let appInfo = new VkApplicationInfo({
pApplicationName: "Hello!",
applicationVersion: VK_MAKE_VERSION(1, 0, 0),
pEngineName: "No Engine",
engineVersion: VK_MAKE_VERSION(1, 0, 0),
apiVersion: VK_API_VERSION_1_0
});
Also note, that it is recommended to enable the --strict
mode in the TS compiler options and use the latest version of the TS compiler.
The API gives you some sugar to write things quicker, but still gives you the option to write everything explicitly
sType
members get auto-filled, but you can still set them yourself
let appInfo = new VkApplicationInfo();
appInfo.sType = VK_STRUCTURE_TYPE_APPLICATION_INFO;
Becomes:
let appInfo = new VkApplicationInfo(); // sType auto-filled
Instead of:
let offset = new VkOffset2D();
offset.x = 0;
offset.y = 0;
let extent = new VkExtent2D();
extent.width = 640;
extent.height = 480;
let renderArea = new VkRect2D();
renderArea.offset = offset;
renderArea.extent = extent;
You can write:
let renderArea = new VkRect2D({
offset: new VkOffset2D({ x: 0, y: 0 }),
extent: new VkExtent2D({ width: 640, height: 480 })
});
nvk allows to use nested structures to improve memory usage and performance. A nested structure is pre-allocated automatically and shares the native memory of it's top-level structure.
You can use the --enable-shared-memory-hints
flag, to get hints where you could've used a nested structure in your code.
Instead of:
let scissor = new VkRect2D();
scissor.offset = new VkOffset2D();
scissor.extent = new VkExtent2D();
scissor.offset.x = 0;
scissor.offset.y = 0;
scissor.extent.width = 480;
scissor.extent.height = 320;
You can write:
let scissor = new VkRect2D();
scissor.offset.x = 0;
scissor.offset.y = 0;
scissor.extent.width = 480;
scissor.extent.height = 320;
To reduce GC pressure, nvk allows to use cached structures. Instead of having to allocate a structure every time on the heap, nvk pre-allocates a cached version of each available structure.
To use the cached version of a structure, you have to remove the new
keyword from a structure call:
let applicationInfo = VkApplicationInfo();
When a structure gets created without the new
keyword, a cached version of the structure gets returned. Note that the returned structure is not a new structure, but always the same one. Each time you create a structure without the new
keyword, the cached structure gets reset to it's original state (similar to zero/null a structure in C).
Example:
let applicationInfoA = new VkApplicationInfo();
applicationInfoA.pApplicationName = "My Application";
let applicationInfoB = new VkApplicationInfo();
applicationInfoB.pApplicationName = applicationInfoA.pApplicationName; // "My Application"
But this is not possible with a cached structure:
let applicationInfoA = VkApplicationInfo();
applicationInfoA.pApplicationName = "My Application A";
let applicationInfoB = VkApplicationInfo(); // reset
applicationInfoB.pApplicationName = applicationInfoA.pApplicationName; // ""
applicationInfoA.pApplicationName
is empty, because when the variable applicationInfoB
gets initialised, all members of the cached structure get reset, so applicationInfoA.pApplicationName
doesn't contain anything anymore.
Only use this feature if you are aware of the side effects that might be introduced by this. Also, using this feature only makes sense in frequently called code sections, where the GC pressure is high.
docs
: generated vulkan documentation filesgenerator
: code for binding generationgenerated
: the generated binding codeexamples
: ready-to-run exampleslib
: required third party libssrc
: classes for e.g. window creation
This tool uses a new JavaScript type called BigInt
to represent memory addresses returned by Vulkan. The BigInt
type was recently added, so make sure you use a recent node.js version.
The Generator generates code based on a vk.xml
specification file. It first converts the XML file into an AST, which is then used by the code generator. Currently more than ~300.000
lines of code get generated, where ~60.000
lines are JavaScript, ~50.000
lines are TypeScript, ~40.000
lines are C++ code and the rest code for the documentation and AST.
Starting from version 0.5.0
, nvk now uses a concept called Hybrid bindings, which reduces the overhead of JavaScript<->C++ context switching. Structures tend to have many members, where each member has to be a getter/setter function. Before this change, these getters/setters were written in C++, so there were many tiny context switches. Now the native memory of Structures and Handles just get filled entirely within JavaScript (see the file here), resulting in much less overhead and much simpler binding and generator code.
This section is of interest, if you have an existing C++ project and want to link against this one.
This project mostly doesn't requires to be linked against. All structures and handles have properties to access the underlying memory directly. For example, see VkApplicationInfo (#Default Properties).
Structures and handles come with these 3 properties:
- .memoryBuffer: Reference to the underlying native memory, wrapped inside an ArrayBuffer
- .memoryAddress: Native address (BigInt) of memoryBuffer. To convert BigInt into a native type, see e.g. this document
- .byteLength: Total native bytelength of the structure/handle
Warning: You may want to skip this section, as nvk uses N-API and ships pre-compiled binaries. This section is only of interest if you want to generate and build the bindings yourself, which is likely not your intention!
This project requires two-pass compilation which means, after initially compiling the bindings, a second compilation is required. This is necessary, because this project constructs Vulkan memory entirely from within JavaScript.
- At the first compilation, memory layouts of vulkan structures get stored inside a JSON file
- At the second pass, these memory layout then get used to inline memory offsets inside the JavaScript binding code
- node.js >= v10.9.0 recommended
If you already have Visual Studio >= 15 installed, then just make sure to have Python 2.7.x
installed.
If you don't have Visual Studio, then install the following package:
npm install --global --production windows-build-tools
Now install the corresponding Vulkan SDK version from here.
Next, clone this repository.
To generate and compile the bindings, run:
npm run generate --vkversion=x
npm run build --vkversion=x
Download and setup the corresponding Vulkan SDK version from here.
Follow the guide on how to correctly setup the SDK.
Make sure that the environment variables are correctly set, e.g. echo $VULKAN_SDK
.
Next, clone this repository.
To generate and compile the bindings, run:
npm run generate --vkversion=x
npm run build --vkversion=x
Download and setup the corresponding Vulkan SDK version from here.
Follow the guide on how to correctly setup the SDK.
Make sure that the environment variables are correctly set, e.g. echo $VULKAN_SDK
.
Next, clone this repository.
To generate and compile the bindings, run:
npm run generate --vkversion=x
npm run build --vkversion=x
npm run [script] [flag] [value]
[--disable-validation-checks]: Disables type and bounding checks for better performance
[--enable-shared-memory-hints]: Enables console hints, reporting to use nested structures when possible - useful for performance optimization
You can generate bindings with:
npm run generate --vkversion=1.1.114
The generated bindings can then be found in generated/{vkversion}/${platform}
- Make sure the specified version to generate bindings for can be found here
- The binding specification file gets auto-downloaded and is stored in
generate/specifications/{vkversion}.xml
--incremental
flag should only be used if you're a developer of nvk
[--vkversion]: The Vulkan version to generate bindings for
[--fake-platform]: Allows to specify a fake platform to generate bindings for. Only use this when the native bindings don't have to be recompiled! A useful but dangerous flag
[--disable-minification]: Disables code minification of the JavaScript interfaces
[--incremental]: Enables incremental builds when building the bindings
[--docs]: Generates HTML-based documentation, also used for TypeScript type annotations
You can build the generated bindings with:
npm run build --vkversion=1.1.114
The compiled bindings can then be found in generated/{vkversion}/build
[--vkversion]: The Vulkan version to build bindings for
[--msvsversion]: The Visual Studio version to build the bindings with
Using RenderDoc is simple. Open RenderDoc and in the Launch Application tab, enter e.g.:
- Executable Path:
C:\Program Files\nodejs\node.exe
- Command-line Arguments:
--experimental-modules C:\GitHub\nvk-examples\triangle\index.mjs
- Function generation (~95%)
- Documentation generator (95%)