gfx-rs/portability

Discussing the C bindings for Vulkan Portability with gfx-rs

Shawn Thompson

@Thomspoon

Is that specific layer message coming from gfx, or is it intrinsic to the Vulkan Loader?

Shawn Thompson

@Thomspoon

Ah, they're specified in descriptorsets.txt

Dzmitry Malyshau

@kvark

Check gfxGetPhysicalDeviceFeatures

Shawn Thompson

@Thomspoon

@kvark I've created #188 which depends on gfx-rs/gfx#2814, the only issue I had was !0 didn't seem to work for max_memory_allocation_count so I just set it to what my physical device queried, and a couple other limits I set arbitrarily based on what my machine returned.

It turned 52 validation errors into 6! Now what I'm dealing with is:

[ERROR] "Invalid DescriptorSet Object 0x23. The Vulkan spec states: pDescriptorSets must be a valid pointer to an array of descriptorSetCount valid VkDescriptorSet handles (https://www.khronos.org/registry/vulkan/specs/1.1-extensions/html/vkspec.html#VUID-vkCmdBindDescriptorSets-pDescriptorSets-parameter)"
[ERROR] "Invalid DescriptorSet Object 0x23. The Vulkan spec states: pDescriptorSets must be a valid pointer to an array of descriptorSetCount valid VkDescriptorSet handles (https://www.khronos.org/registry/vulkan/specs/1.1-extensions/html/vkspec.html#VUID-vkCmdBindDescriptorSets-pDescriptorSets-parameter)"
[ERROR] "Attempt to bind descriptor set 0x23 that doesn\'t exist!"
[ERROR] "Attempt to bind descriptor set 0x23 that doesn\'t exist!"
thread '<unnamed>' panicked at 'assertion failed: !self.0.is_null()', libportability-gfx/src/handle.rs:91:9
note: Run with `RUST_BACKTRACE=1` environment variable to display a backtrace.
[ERROR] "Invalid Framebuffer Object 0x11. The Vulkan spec states: If framebuffer is not VK_NULL_HANDLE, framebuffer must be a valid VkFramebuffer handle (https://www.khronos.org/registry/vulkan/specs/1.1-extensions/html/vkspec.html#VUID-vkDestroyFramebuffer-framebuffer-parameter)"
[ERROR] "Invalid Framebuffer Object 0x11. The Vulkan spec states: If framebuffer is not VK_NULL_HANDLE, framebuffer must be a valid VkFramebuffer handle (https://www.khronos.org/registry/vulkan/specs/1.1-extensions/html/vkspec.html#VUID-vkDestroyFramebuffer-framebuffer-parameter)"
[1]    18306 segmentation fault  VK_INSTANCE_LAYERS=VK_LAYER_LUNARG_standard_validation cargo run

Which is roughly what I had with MoltenVK

I'm not sure what to do about the first error, I know it's operator error, but I'm not sure where.

Dzmitry Malyshau

@kvark

great, thank you!

these errors look legitimately on your side :)

Shawn Thompson

@Thomspoon

Can vkUpdateDescriptorSets fail? It doesn't return vk::Result, and validation layers aren't spitting anything out.

Dzmitry Malyshau

@kvark

if you are not following the "valid use" of this function, validation layers may be incomplete to tell you that

Shawn Thompson

@Thomspoon

Ah, user error. My descriptor pool was getting deleted which invalidates the descriptor sets associated with it.

Dzmitry Malyshau

@kvark

:+1:

Shawn Thompson

@Thomspoon

What can cause VK_ERROR_SURFACE_LOST_KHR on queue_present?

Dzmitry Malyshau

@kvark

change of surface dimensions

window system integration that we have for Metal is pretty picky...

Shawn Thompson

@Thomspoon

Interesting, so we should just recreate the surface and swapchain?

Dzmitry Malyshau

@kvark

IIRC just the swapchain

Dzmitry Malyshau

@kvark

My sdl patch got finally merged upstream, yet they are still keeping libmoltenvk search path for compatibility reasons:/

Josh Groves

@grovesNL

A step in the right direction at least :)

Dzmitry Malyshau

@kvark

Just updated to macOS 10.15 and XCode 11, was finally able to capture vkQuake running with argument buffers!

Dzmitry Malyshau

@kvark

it even lists indirect resources in a dedicated section... holy greatness

Dzmitry Malyshau

@kvark

going to try out https://github.com/coelckers/gzdoom, which apparently has a Vulkan backend

Dzmitry Malyshau

@kvark

Ok, gzdoom kinda works, we are just missing the timestamp queries

Dzmitry Malyshau

@kvark

we need to force MVK to implement KhronosGroup/MoltenVK#653

otherwise it's hard to update the apps. I don't want our gfx-portability to hackily check if NSView is actually just a layer, yikes

Dzmitry Malyshau

@kvark

my gzdoom PR is merged :tada: coelckers/gzdoom#884

Josh Groves

@grovesNL

sweet! I also asked about portability support as godot adds vulkan support (https://twitter.com/grovesNL/status/1146031845106212864)

Dzmitry Malyshau

@kvark

with MoltenVK we get everything we need so far

what an ignorant response

@grovesNL would you mind replying specifically with the extensions VK_EXTX_portability_subset and VK_EXT_metal_surface?

Dzmitry Malyshau

@kvark

holy crap, what is this for ^^^

Josh Groves

@grovesNL

yeah I can mention the extensions, although maybe I should just join their irc and ask about it. twitter probably isn't the best channel

Dzmitry Malyshau

@kvark

my hopes for them are pretty low... best we can do is to grab an early build and make some benchmarks, maybe submit a PR changing the extensions they use

Dzmitry Malyshau

@kvark

argh, I bought Forsaken on gog and Turok2 on steam, neither seem to support Vulkan on macOS

Dzmitry Malyshau

@kvark

I updated #125 with the missing parts for Forsaken... there is quite a few extensions, surprisingly

one of the things blocking this is our Vulkan bindings - they need to be updated. It would be great to auto-generate them with some of the infrastructure @MaikKlein set up for Ash

Dzmitry Malyshau

@kvark

someone is running gzDoom on Molten - https://github.com/KhronosGroup/MoltenVK/issues/651#issue-462400700

Dzmitry Malyshau

@kvark

heh, VkPI group got shocked to hear that we support vkEvents :P

Josh Groves

@grovesNL

nice :)

Dzmitry Malyshau

@kvark

head's up! I want to do a portability release this weekend (just before SIGGRAPH)

Dota2 doesn't work, since it now asks for MoltenVK-specific entry points :/
but I don't think it matters much. It will run at some point, and there isn't a line of people who want to mess with it today anyway

Josh Groves

@grovesNL

alright. That's unfortunate, which entry points?

Dzmitry Malyshau

@kvark

Missing entrypoints for instance extension 'VK_KHR_get_physical_device_properties2', disabling.
dyld: lazy symbol binding failed: Symbol not found: _vkGetPhysicalDeviceMetalFeaturesMVK
Referenced from: /Users/dmalyshau/Code/dota2/bin/osx64/librendersystemvulkan.dylib
Expected in: /Users/dmalyshau/Code/portability/target/debug/libMoltenVK.dylib

looks like some poorly written code, since the entry point is not a part of the extension they mention

Josh Groves

@grovesNL

ah looks like https://github.com/KhronosGroup/MoltenVK/issues/556#issuecomment-478843505

Still a bit frustrating though, imagine if all drivers did this instead of bothering with extensions

Dzmitry Malyshau

@kvark

this will not go unnoticed :)

I'm providing the list of biggest things we did since GDC, and I mentioned the current Dota situation

Dzmitry Malyshau

@kvark

https://www.phoronix.com/scan.php?page=news_item&px=GFX-RS-0.7-Released

Dzmitry Malyshau

@kvark

oh boy, LunarG examples are being modified on macOS to serve MVK limitations - KhronosGroup/MoltenVK#687

People

Repo info

Activity