Vulkan: Performance Improvements #113918

New Issue

Jeroen Bakker · 2023-10-19T08:24:00+02:00

Jeroen Bakker commented

2023-10-19 08:24:00 +02:00

Vulkan performance isn't what we are targeting. This is by approach where we first focused on quality, support and feature completeness. This tasks collects some ideas that we have to improve the performance.

The Vulkan API is designed on knowing up front what you want to do, and redo this over and over. These recipes are called pipelines in vulkan. This is fine in gaming where every frame is drawn fairly the same. For Blender (and I assume) other professional applications this isn't the case. Blender still uses immediate mode emulation in pretty much everywhere. Where state is set and draw commands are executed on the active state.

Vulkan has an extension that provides this workflow (VK_EXT_shader_object), but isn't widely supported (https://vulkan.gpuinfo.org/listdevicescoverage.php?extension=VK_EXT_shader_object). At the time of writing only supported by NVidia and llvmpipe.

Here is a list of ideas that we can do to improve the performance in the short term. More evasive ideas might require a big overhaul of the GPU module API.

Use multiple VkCommandBuffers.
- One dedicated to data transfers and blit operations
- One command buffer dedicated to compute operations
- One command buffer dedicated to graphics operations
- These command buffers will have some ordering when they will be sent to the device queue. When sent additional synchronization rules can be added for better performance.
Add support for multi draw commands (draw list)
Render passes, subpasses, framebuffers are tightly together would benefit recipes.
- When using framebuffer transitions, first provide the transitions up front when binding the framebuffer. Use GPU_framebuffer_subpass_next` or something to perform the subpass transition.
Add support for VkPipelineCache

Vulkan performance isn't what we are targeting. This is by approach where we first focused on quality, support and feature completeness. This tasks collects some ideas that we have to improve the performance. The Vulkan API is designed on knowing up front what you want to do, and redo this over and over. These recipes are called pipelines in vulkan. This is fine in gaming where every frame is drawn fairly the same. For Blender (and I assume) other professional applications this isn't the case. Blender still uses immediate mode emulation in pretty much everywhere. Where state is set and draw commands are executed on the active state. Vulkan has an extension that provides this workflow (VK_EXT_shader_object), but isn't widely supported (https://vulkan.gpuinfo.org/listdevicescoverage.php?extension=VK_EXT_shader_object). At the time of writing only supported by NVidia and llvmpipe. Here is a list of ideas that we can do to improve the performance in the short term. More evasive ideas might require a big overhaul of the GPU module API. * [ ] Use multiple VkCommandBuffers. * One dedicated to data transfers and blit operations * One command buffer dedicated to compute operations * One command buffer dedicated to graphics operations * These command buffers will have some ordering when they will be sent to the device queue. When sent additional synchronization rules can be added for better performance. * [ ] Add support for multi draw commands (draw list) * [ ] Render passes, subpasses, framebuffers are tightly together would benefit recipes. * When using framebuffer transitions, first provide the transitions up front when binding the framebuffer. Use GPU_framebuffer_subpass_next` or something to perform the subpass transition. * [ ] Add support for VkPipelineCache

Jeroen Bakker added this to the 4.2 LTS milestone 2023-10-19 08:24:00 +02:00

Jeroen Bakker added the

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Vulkan: Performance Improvements #113918