ilusha/nanovg_vulkan
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
nanovg_vulkan/README.md

12 KiB
Raw Permalink Blame History

Background

This repo is based on the amazing work done by danilw in the archived repository danilw/nanovg_vulkan

Goal

For this library to be a Vulkan 1.0 compatible implementation of NanoVG, while using the full featureset of more capable hardware via optional extensions:

  • Ubuntu 24.04.1 LTS. Vulkan = 2857 fps (0.35ms), OpenGL 3.2 = 2632 fps (0.38ms). [1]

  • Windows 11 2024H2. Vulkan = 2273 fps (0.44 ms), OpenGL 3.2 = 2500 fps (0.40 ms). [1]

  • MacOS X. Vulkan = 709 fps (1.41 ms), OpenGL 3.2 = 763 fps (1.31 ms). [2]

  • Optimising pipeline creation and selection via use of supported (and/or enabled) extensions. Specifically:

    • vkCmdSetColorBlendEquationEXT
    • vkCmdSetPrimitiveTopologyEXT
    • vkCmdSetStencilTestEnableEXT
    • vkCmdSetStencilOpEXT
    • vkCmdSetColorWriteMaskEXT

  • Using one fragment shader instance, based on the NVG_ANTIALIAS flag (specialization constants)

  • Using Push constants in the Vertex Shader

  • Bind vertex buffer once per flush; vkCmdDraw uses firstVertex

  • Using SSBO for a single write of fragment uniforms, indexed via push constants (no per-draw UBO offset dance)

  • One texture descriptor set per call; extra stencil/AA uniforms are only a push-constant change

  • Per-frame VKNVGFrame (vertex + fragment SSBO), optional frameCount separate from swapchain image count

  • Texture layout transitions use a dedicated upload command buffer, not the in-flight render CB

  • Descriptor pools are retired instead of destroyed while other frames may still reference them

  • Optimise texture memory flags

  • Notes

    • [1] Tested on RX 6800, R7 7700, 32GB RAM.
    • [2] Tested on M1 Macbook Pro

2024 - Im archive this.

Just use IMGUI

if you want - fork it and update it as you with, this is working in 2024 but im not going to update or change anything anymore.

Note about Vulkan - this nanovg_vulkan is Vulkan 1.0 project.
Vulkan 1.3 already released with about 10 years back GPU support in drivers. Vulkan 1.3 have VK_KHR_dynamic_rendering in core.
As minimal example of setup of complex pipeline look this blog - modern Vulkan with descriptor indexing, dynamic rendering and shader objects tutorial - and this repository vknew.
I think rewriting this entire project from scratch in Vulkan 1.3 - will be faster and easier than fixin some bugs here and improving it.

Notice - I wanted to fix one major bug here - DPI scale does not work, read known-bugs-related-info but I have no motivation to do it, too much effort, pointless no one need task.

Contact me in Discord or anywhere else if you want to ask some tech-info stuff, or I can add your "active" fork of this as link to this page.

added vulkan build, and cmake files, removed premake4.lua

Why I made this - I made it as a "test task for Vulkan related job" a year ago, for me this project has no real use. I would recommend for everyone use imgui if you look for production-ready Vulkan UI, also look on egui. I will not support or update this project at all, the last update was just to fix bugs and Validation errors.

Description - nanovg in C, no external dependencies. GLFW used only by one example (look examples description, there is example without using GLFW).

Contact: Join discord server

To launch - copy builded binary to example folder(or launch when this folder is current as on build example commands below). Because required fonts and images/shaders to load.

Integration

Compile src/nanovg.c and src/nanovg_vk.c as C11. Include nanovg.h then nanovg_vk.h after vulkan.h. Do not #define NANOVG_VULKAN_IMPLEMENTATION into the Vulkan header (that define is only used by example/demo.c to skip GL). C++ hosts should keep nanovg_vk.c as a C translation unit; the header is extern "C".

nvgCreateVk takes const VKNVGCreateInfo * and copies the struct. Pointers inside it must stay valid for the context lifetime:

Field Role
gpu, device, renderpass, queue (create arg) Device + the render pass NanoVG will record into
cmdBuffer Array of command buffers, length = in-flight count
currentFrame Pointer to the in-flight index (0 .. frameCount-1)
frameCount In-flight GPU buffers. 0 means use swapchainImageCount
commandPool One-shot texture uploads (nvgCreateImage). Prefer your graphics pool with RESET_COMMAND_BUFFER
graphicsQueueFamilyIndex Used only if commandPool is VK_NULL_HANDLE (0 is a valid family)
ext Optional VK_EXT_extended_dynamic_state / _3 bits; leave false if unused

Each frame: begin your render pass on cmdBuffer[*currentFrame], then nvgBeginFrame / draw / nvgEndFrame. NanoVG records into that command buffer; it does not begin a second pass.

Pipelines are tied to renderpass. After swapchain resize, wait idle, nvgDeleteVk, nvgCreateVk with the new pass, then recreate fonts/images.

nvgDeleteVk is safe if you never flushed (frames still null).

Images

  • nvgCreateImage — NanoVG owns the VkImage. Fine for demos; the samples reload via loadDemoData after resize.
  • nvgCreateImageFromHandleVk — you own image, view, and sampler. NanoVG will not destroy them. Integer ids die with the context.

Keep GPU images across resize, then re-register:

  1. Keep VkImage / view / sampler alive.
  2. nvgDeleteVk / nvgCreateVk with the new render pass.
  3. Call nvgCreateImageFromHandleVk again; store the new ids (do not assume id 1).

nvgDeleteImage on a borrowed id only drops the NanoVG slot.

Working setup: example/example_vulkan.c (GLFW, multiple frames in flight) and example/example_vulkan_min_no_glfw.c.

Known bugs-related info:

Build:

(use cmake to build on Windows, Windows/Linux supported and tested)

git clone --recursive https://github.com/danilw/nanovg_vulkan
cd nanovg_vulkan/example
mkdir build
cd build
cmake ../
make
cd ../
./build/example-vk
./build/example-vk_min_no_glfw

CMake already compiles src/nanovg_vk.c. See Integration above for embedding in another project.

Look Examples description below there link to repository with C only minimal example without dependencies not using any library.

MoltenVK note - after TOPOLOGY_TRIANGLE_LIST update(look below) this Vulkan port does work on MoltenVK(Mac/etc) but I dont have it to test so you should make cmake config to build it and launch by yourself.

Examples description:

Multiple frames in flight - example_vulkan.c is multiple frames in flight example, example_vulkan_min_no_glfw.c is single frame in flight. Clearly visible on FPS - with multiple frames about 3x better FPS.

example_vulkan.c - GLFW sample: fills VKNVGCreateInfo (including frameCount and commandPool) and recreates the context on resize.

example_vulkan_min_no_glfw.c - same NanoVG wiring without GLFW, Linux and Windows.

Old examples, they not compatible with this latest nanovg_vulkan.

Those linked demos still use the old header-only nvgCreateVk(createInfo by value) API. Prefer this repo’s samples plus Integration above.

Two external examples:

nanovg-vulkan-min-integration-demo - repository with minimal RenderPass integration example. Not using any libraries, no GLFW, only NanoVG code and C. Look description and screenshot on link.

nanovg-vulkan-glfw-integration-demo - repository with example from Vulkan-tutorial Depth buffering modified adding NanoVG integration. Using C++ and GLFW.

Remember NanoVG is not GUI, and examples about is just examples of this NanoVG integration, not GUI examples.

About RenderPass integration - copy paste code from integration examples above after your vkCmdEndRenderPass (or before vkCmdBeginRenderPass) and everything should work.

Framebuffer integration, where UI rendered in its own Framebuffer - I did not add example for this case, because it should be obvious - just replace framebuffer RenderPass with RenderPass of NanoVG integration examples above(look linked commit).

2023 update - huge performance improvement thanks to @SubiyaCryolite

SubiyaCryolite pull request Optimizations > Cached Descriptor Sets, Implied Multiple Frames in Flight, Fencing for faster perf

  • Support for multiple command buffers, specifically one per swap-chain image / buffer
  • Caching of descriptor sets, removing the need to create new ones per draw func or to call vkResetDescriptorPool per frame.
  • Under example_vulkan, using vkWaitForFences to control rendering as opposed to vkQueueWaitIdle (seems to be the biggest perf booster). This change also has implied "multiple frames in flight" as dictated by the swap-chain image count.
  • Using persisted mapped buffers via VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT instead of VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT , allows to skip calls to vkMap/UnmapMemory per frame

Link to version before this change 5ba9d31, this change 10d5211.

2022 update - thanks to @nidefawl pull request with lots of changes:

In this PR:

  1. Merge latest nanovg

  2. Fix Stencil Strokes: I added a new enum to handle the 3 different pipelines used for stencil strokes. I wanted to also use enums for the fill-path (nvgFill) to make the code cleaner, but haven't done that yet.

  3. I combined the 2 CMakeLists.txt into example/CMakeLists.txt and added some bits to make it easier to build and debug in MSVC/VSCode.

  4. I added GLAD as default gl-loader for the GL3 example. It can be switched back to GLEW (I kept glew to allow easier merge with main nanovg)

  5. I Increased the number of swapchain images to minimum 3 and updated the swapchain barrier to gain some extra performance.

  6. Spacebar renders the demo multiple times to add some 'load' when comparing performance

2021 update - fixed all errors, this code has no Validation errors and work correctly everywhere.

Thanks to @fzwoch commits by default used TOPOLOGY_TRIANGLE_LIST, because TOPOLOGY_TRIANGLE_FAN is optional in Vulkan.

To enable TOPOLOGY_TRIANGLE_FAN uncomment #define USE_TOPOLOGY_TRIANGLE_FAN in src/nanovg_vk.c.

Depth order bug on AMD fix by @leranger 6ee1009.

Screenshot of Vulkan version from Linux:

nvgvk

Read original readme for more info about NanoVG API https://github.com/memononen/nanovg