[Feature Request] Add vulkan execution provider

[thewh1teagle]

Describe the feature request

Please consider adding Vulkan as an execution provider. Vulkan offers several advantages over DirectML, Rocm, Migraph, and Cuda:

• Easier Compilation: Vulkan simplifies the compilation process, requiring fewer and lighter dependencies across both Linux and Windows.
• Lightweight Dependencies: Vulkan eliminates the need for heavy dependencies and large packages required by Cuda on Linux.
• Cross-Platform Efficiency: Unlike Rocm, Vulkan operates efficiently on both Linux and Windows without the burden of extensive dependencies.
• Reduced Footprint: On Windows, Vulkan avoids the need for the 300MB dependencies required by Cuda.
• Wide Compatibility: Vulkan works seamlessly out-of-the-box on Windows with a variety of GPUs.
• Unified Provider: A single Vulkan execution provider can support GPUs/CPUs across AMD, Nvidia, and Intel platforms.

By the way, the Vulkan license is open for both commercial and non-commercial use.

[altunenes]

This is one of my biggest dreams for onnx models...Not having to create different installers for each GPU/os and being able to harness the power of Vulkan: perfect...

[fs-eire]

@altunenes the first version of WebGPU EP is almost ready to merge. Please track: #22591

Currently WebGPU EP uses D3D12 on Windows, vulkan on linux/android and metal on iOS/macOS. Is using vulkan in Windows a requirement?

[jlucaso1]

@fs-eire Is it possible to choose a backend on Windows (either D3D12 or Vulkan)? I'm curious about the performance differences between the two.

[fs-eire]

That makes sense. I can share a change after merging that PR to allow build locally with valkan, but considering the binary size, I will stick to use only one backend unless it's proven useful.

[altunenes]

@altunenes the first version of WebGPU EP is almost ready to merge. Please track: #22591

Currently WebGPU EP uses D3D12 on Windows, vulkan on linux/android and metal on iOS/macOS. Is using vulkan in Windows a requirement?

First of all, I am really grateful for this important contribution. I can't wait to test this!
Vulkan is not a " requirement" on Windows, but as said above, I think it will be very nice for performance comparison.

[ramkrishna2910]

@fs-eire can you please share the change that allows webgpu to build locally with Vulkan on Windows?

[fs-eire]

I made a PR: #23096

with this change you can build with --use_webgpu --config Release --build_shared_lib --cmake_extra_defines onnxruntime_ENABLE_DAWN_BACKEND_D3D12=0 onnxruntime_ENABLE_DAWN_BACKEND_VULKAN=1 to use Vulkan backend of Dawn.

I tested it locally on my Windows devbox and it seems working with f32 models. f16 seems disabled for unknown reason (the same machine can run f16 on D3D12). All other things are just good.

[fs-eire]

f16 is disabled on all nvidia GPUs in Vulkan backend now:

https://github.com/google/dawn/blob/b8306de17c9a5775819a8526802fe85a860f5216/src/dawn/native/vulkan/PhysicalDeviceVk.cpp#L262