Does ash use function pointers or link against the loader directly?

[Jerrody]

For example, in C/C++ headers if we call a PFN we skip the step when loader choose what driver to call.

Do I need to call a PFN or ash does all work?

[MarijnS95]

@Jerrody Are you referring whether a function is loaded generically through vkGetInstanceProcAddr with or without VkInstance, either resulting in an ICD loader dispatch wrapper or a function specific for that instance?

https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/vkGetInstanceProcAddr.html#_description

Same for vkGetDeviceProcAddr: https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/vkGetDeviceProcAddr.html#_description

[Jerrody]

@MarijnS95

I'm not big professional I refer to this article where shown that calling of vkCmdDraw is an overhead, instead we call directly PFN_vkCmdDraw at this case we remove this overhead.

https://gpuopen.com/learn/reducing-vulkan-api-call-overhead/

So, my question is does ash call directly this PFN where we call device. cmd_draw()?

[MarijnS95]

@Jerrody PFN_vkCmdDraw it the type of a function pointer. The loader exports a function with this signature as vkCmdDraw, to be linked against directly but with aforementioned dispatch table overhead. Ash does not use those and instead loads it via vkGetDeviceProcAddr(device, "vkCmdDraw") directly, as written in that article:

ash/ash/src/vk/features.rs

Lines 3051 to 3052 in 29b935b

	let cname = ::std::ffi::CStr::from_bytes_with_nul_unchecked(b"vkCmdDraw\0");
	let val = _f(cname);

_f here is the loader function:

ash/ash/src/device.rs

Lines 22 to 24 in 29b935b

	let load_fn = |name: &std::ffi::CStr| {
	mem::transmute((instance_fn.get_device_proc_addr)(device, name.as_ptr()))
	};

Which is used when creating a device:

ash/ash/src/instance.rs

Lines 341 to 357 in 29b935b

	#[inline]
	pub unsafe fn create_device(
	&self,
	physical_device: vk::PhysicalDevice,
	create_info: &vk::DeviceCreateInfo,
	allocation_callbacks: Option<&vk::AllocationCallbacks>,
	) -> VkResult<Device> {
	let mut device = mem::zeroed();
	(self.instance_fn_1_0.create_device)(
	physical_device,
	create_info,
	allocation_callbacks.as_raw_ptr(),
	&mut device,
	)
	.result()?;
	Ok(Device::load(&self.instance_fn_1_0, device))
	}

tl;dr: Ash uses optimized function pointers (specific to an instance/device) in almost all cases ¹.

Footnotes

1. Some extensions combine device-level and instance-level functions: the latter cannot be loaded for a device, hence the entire extension is loaded via the instance and we take the dispatch table hit until our generator is capable of separating them out (https://github.com/ash-rs/ash/pull/556#discussion_r784237840). ↩

[Jerrody]

@MarijnS95

Thanks for the answer!