wgpu_hal/vulkan/

adapter.rs

use alloc::{borrow::ToOwned as _, boxed::Box, collections::BTreeMap, sync::Arc, vec::Vec};
use core::{ffi::CStr, marker::PhantomData};

use ash::{ext, google, khr, vk};
use wgpu_sync::Mutex;

use crate::{vulkan::semaphore_list::SemaphoreList, AllocationSizes};

use super::semaphore_list::SemaphoreListMode;

fn depth_stencil_required_flags() -> vk::FormatFeatureFlags {
    vk::FormatFeatureFlags::SAMPLED_IMAGE | vk::FormatFeatureFlags::DEPTH_STENCIL_ATTACHMENT
}

const INDEXING_FEATURES: wgt::Features = wgt::Features::TEXTURE_BINDING_ARRAY
    .union(wgt::Features::BUFFER_BINDING_ARRAY)
    .union(wgt::Features::STORAGE_RESOURCE_BINDING_ARRAY)
    .union(wgt::Features::SAMPLED_TEXTURE_AND_STORAGE_BUFFER_ARRAY_NON_UNIFORM_INDEXING)
    .union(wgt::Features::STORAGE_TEXTURE_ARRAY_NON_UNIFORM_INDEXING)
    .union(wgt::Features::UNIFORM_BUFFER_BINDING_ARRAYS)
    .union(wgt::Features::PARTIALLY_BOUND_BINDING_ARRAY);
#[expect(rustdoc::private_intra_doc_links)]
/// Features supported by a [`vk::PhysicalDevice`] and its extensions.
///
/// This is used in two phases:
///
/// - When enumerating adapters, this represents the features offered by the
///   adapter. [`Instance::expose_adapter`] calls `vkGetPhysicalDeviceFeatures2`
///   (or `vkGetPhysicalDeviceFeatures` if that is not available) to collect
///   this information about the `VkPhysicalDevice` represented by the
///   `wgpu_hal::ExposedAdapter`.
///
/// - When opening a device, this represents the features we would like to
///   enable. At `wgpu_hal::Device` construction time,
///   [`PhysicalDeviceFeatures::from_extensions_and_requested_features`]
///   constructs an value of this type indicating which Vulkan features to
///   enable, based on the `wgpu_types::Features` requested.
///
/// [`Instance::expose_adapter`]: super::Instance::expose_adapter
#[derive(Debug, Default)]
pub struct PhysicalDeviceFeatures {
    /// Basic Vulkan 1.0 features.
    core: vk::PhysicalDeviceFeatures,

    /// Features provided by `VK_EXT_descriptor_indexing`, promoted to Vulkan 1.2.
    pub(super) descriptor_indexing:
        Option<vk::PhysicalDeviceDescriptorIndexingFeaturesEXT<'static>>,

    /// Features provided by `VK_KHR_timeline_semaphore`, promoted to Vulkan 1.2
    timeline_semaphore: Option<vk::PhysicalDeviceTimelineSemaphoreFeaturesKHR<'static>>,

    /// Features provided by `VK_EXT_image_robustness`, promoted to Vulkan 1.3
    image_robustness: Option<vk::PhysicalDeviceImageRobustnessFeaturesEXT<'static>>,

    /// Features provided by `VK_EXT_robustness2`.
    robustness2: Option<vk::PhysicalDeviceRobustness2FeaturesEXT<'static>>,

    /// Features provided by `VK_KHR_multiview`, promoted to Vulkan 1.1.
    multiview: Option<vk::PhysicalDeviceMultiviewFeaturesKHR<'static>>,

    /// Features provided by `VK_KHR_sampler_ycbcr_conversion`, promoted to Vulkan 1.1.
    sampler_ycbcr_conversion: Option<vk::PhysicalDeviceSamplerYcbcrConversionFeatures<'static>>,

    /// Features provided by `VK_EXT_texture_compression_astc_hdr`, promoted to Vulkan 1.3.
    astc_hdr: Option<vk::PhysicalDeviceTextureCompressionASTCHDRFeaturesEXT<'static>>,

    /// Features provided by `VK_KHR_shader_float16_int8`, promoted to Vulkan 1.2
    shader_float16_int8: Option<vk::PhysicalDeviceShaderFloat16Int8Features<'static>>,

    /// Features provided by `VK_KHR_16bit_storage`, promoted to Vulkan 1.1
    _16bit_storage: Option<vk::PhysicalDevice16BitStorageFeatures<'static>>,

    /// Features provided by `VK_KHR_acceleration_structure`.
    acceleration_structure: Option<vk::PhysicalDeviceAccelerationStructureFeaturesKHR<'static>>,

    /// Features provided by `VK_KHR_buffer_device_address`, promoted to Vulkan 1.2.
    ///
    /// We only use this feature for
    /// [`Features::EXPERIMENTAL_RAY_QUERY`], which requires
    /// `VK_KHR_acceleration_structure`, which depends on
    /// `VK_KHR_buffer_device_address`, so [`Instance::expose_adapter`] only
    /// bothers to check if `VK_KHR_acceleration_structure` is available,
    /// leaving this `None`.
    ///
    /// However, we do populate this when creating a device if
    /// [`Features::EXPERIMENTAL_RAY_QUERY`] is requested.
    ///
    /// [`Instance::expose_adapter`]: super::Instance::expose_adapter
    /// [`Features::EXPERIMENTAL_RAY_QUERY`]: wgt::Features::EXPERIMENTAL_RAY_QUERY
    buffer_device_address: Option<vk::PhysicalDeviceBufferDeviceAddressFeaturesKHR<'static>>,

    /// Features provided by `VK_KHR_ray_query`,
    ///
    /// Vulkan requires that the feature be present if the `VK_KHR_ray_query`
    /// extension is present, so [`Instance::expose_adapter`] doesn't bother retrieving
    /// this from `vkGetPhysicalDeviceFeatures2`.
    ///
    /// However, we do populate this when creating a device if ray tracing is requested.
    ///
    /// [`Instance::expose_adapter`]: super::Instance::expose_adapter
    ray_query: Option<vk::PhysicalDeviceRayQueryFeaturesKHR<'static>>,

    /// Features provided by `VK_KHR_ray_tracing_pipeline`.
    ray_tracing_pipeline: Option<vk::PhysicalDeviceRayTracingPipelineFeaturesKHR<'static>>,

    /// Features provided by `VK_KHR_zero_initialize_workgroup_memory`, promoted
    /// to Vulkan 1.3.
    zero_initialize_workgroup_memory:
        Option<vk::PhysicalDeviceZeroInitializeWorkgroupMemoryFeatures<'static>>,
    position_fetch: Option<vk::PhysicalDeviceRayTracingPositionFetchFeaturesKHR<'static>>,

    /// Features provided by `VK_KHR_shader_atomic_int64`, promoted to Vulkan 1.2.
    shader_atomic_int64: Option<vk::PhysicalDeviceShaderAtomicInt64Features<'static>>,

    /// Features provided by `VK_EXT_shader_image_atomic_int64`
    shader_image_atomic_int64: Option<vk::PhysicalDeviceShaderImageAtomicInt64FeaturesEXT<'static>>,

    /// Features provided by `VK_EXT_shader_atomic_float`.
    shader_atomic_float: Option<vk::PhysicalDeviceShaderAtomicFloatFeaturesEXT<'static>>,

    /// Features provided by `VK_EXT_subgroup_size_control`, promoted to Vulkan 1.3.
    subgroup_size_control: Option<vk::PhysicalDeviceSubgroupSizeControlFeatures<'static>>,

    /// Features provided by `VK_KHR_maintenance4`, promoted to Vulkan 1.3.
    maintenance4: Option<vk::PhysicalDeviceMaintenance4FeaturesKHR<'static>>,

    /// Features proved by `VK_EXT_mesh_shader`
    mesh_shader: Option<vk::PhysicalDeviceMeshShaderFeaturesEXT<'static>>,

    /// Features provided by `VK_KHR_shader_integer_dot_product`, promoted to Vulkan 1.3.
    shader_integer_dot_product:
        Option<vk::PhysicalDeviceShaderIntegerDotProductFeaturesKHR<'static>>,

    /// Features provided by `VK_KHR_fragment_shader_barycentric`
    shader_barycentrics: Option<vk::PhysicalDeviceFragmentShaderBarycentricFeaturesKHR<'static>>,

    /// Features provided by `VK_KHR_portability_subset`.
    ///
    /// Strictly speaking this tells us what features we *don't* have compared to core.
    portability_subset: Option<vk::PhysicalDevicePortabilitySubsetFeaturesKHR<'static>>,

    /// Features provided by `VK_KHR_cooperative_matrix`
    cooperative_matrix: Option<vk::PhysicalDeviceCooperativeMatrixFeaturesKHR<'static>>,

    /// Features provided by `VK_KHR_vulkan_memory_model`, promoted to Vulkan 1.2
    vulkan_memory_model: Option<vk::PhysicalDeviceVulkanMemoryModelFeaturesKHR<'static>>,

    shader_draw_parameters: Option<vk::PhysicalDeviceShaderDrawParametersFeatures<'static>>,
}

impl PhysicalDeviceFeatures {
    pub fn get_core(&self) -> vk::PhysicalDeviceFeatures {
        self.core
    }

    /// Add the members of `self` into `info.enabled_features` and its `p_next` chain.
    pub fn add_to_device_create<'a>(
        &'a mut self,
        mut info: vk::DeviceCreateInfo<'a>,
    ) -> vk::DeviceCreateInfo<'a> {
        info = info.enabled_features(&self.core);
        if let Some(ref mut feature) = self.descriptor_indexing {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.timeline_semaphore {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.image_robustness {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.robustness2 {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.multiview {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.astc_hdr {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.shader_float16_int8 {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self._16bit_storage {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.zero_initialize_workgroup_memory {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.acceleration_structure {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.buffer_device_address {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.ray_query {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.ray_tracing_pipeline {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.shader_atomic_int64 {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.position_fetch {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.shader_image_atomic_int64 {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.shader_atomic_float {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.subgroup_size_control {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.maintenance4 {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.mesh_shader {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.shader_integer_dot_product {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.shader_barycentrics {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.portability_subset {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.cooperative_matrix {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.vulkan_memory_model {
            info = info.push_next(feature);
        }
        if let Some(ref mut feature) = self.shader_draw_parameters {
            info = info.push_next(feature);
        }
        info
    }

    fn supports_storage_input_output_16(&self) -> bool {
        self._16bit_storage
            .as_ref()
            .map(|features| features.storage_input_output16 != 0)
            .unwrap_or(false)
    }

    /// Create a `PhysicalDeviceFeatures` that can be used to create a logical
    /// device.
    ///
    /// Return a `PhysicalDeviceFeatures` value capturing all the Vulkan
    /// features needed for the given [`Features`], [`DownlevelFlags`], and
    /// [`PrivateCapabilities`]. You can use the returned value's
    /// [`add_to_device_create`] method to configure a
    /// [`vk::DeviceCreateInfo`] to build a logical device providing those
    /// features.
    ///
    /// To ensure that the returned value is able to select all the Vulkan
    /// features needed to express `requested_features`, `downlevel_flags`, and
    /// `private_caps`:
    ///
    /// - The given `enabled_extensions` set must include all the extensions
    ///   selected by [`Adapter::required_device_extensions`] when passed
    ///   `features`.
    ///
    /// - The given `device_api_version` must be the Vulkan API version of the
    ///   physical device we will use to create the logical device.
    ///
    /// [`Features`]: wgt::Features
    /// [`DownlevelFlags`]: wgt::DownlevelFlags
    /// [`PrivateCapabilities`]: super::PrivateCapabilities
    /// [`add_to_device_create`]: PhysicalDeviceFeatures::add_to_device_create
    /// [`Adapter::required_device_extensions`]: super::Adapter::required_device_extensions
    fn from_extensions_and_requested_features(
        phd_capabilities: &PhysicalDeviceProperties,
        phd_features: &PhysicalDeviceFeatures,
        enabled_extensions: &[&'static CStr],
        requested_features: wgt::Features,
        downlevel_flags: wgt::DownlevelFlags,
        private_caps: &super::PrivateCapabilities,
    ) -> Self {
        let device_api_version = phd_capabilities.device_api_version;
        let needs_bindless = requested_features.intersects(
            wgt::Features::TEXTURE_BINDING_ARRAY
                | wgt::Features::BUFFER_BINDING_ARRAY
                | wgt::Features::STORAGE_RESOURCE_BINDING_ARRAY
                | wgt::Features::STORAGE_TEXTURE_ARRAY_NON_UNIFORM_INDEXING
                | wgt::Features::SAMPLED_TEXTURE_AND_STORAGE_BUFFER_ARRAY_NON_UNIFORM_INDEXING,
        );
        let needs_partially_bound =
            requested_features.intersects(wgt::Features::PARTIALLY_BOUND_BINDING_ARRAY);

        Self {
            // vk::PhysicalDeviceFeatures is a struct composed of Bool32's while
            // Features is a bitfield so we need to map everything manually
            core: vk::PhysicalDeviceFeatures::default()
                .robust_buffer_access(private_caps.robust_buffer_access)
                .independent_blend(downlevel_flags.contains(wgt::DownlevelFlags::INDEPENDENT_BLEND))
                .sample_rate_shading(
                    downlevel_flags.contains(wgt::DownlevelFlags::MULTISAMPLED_SHADING),
                )
                .image_cube_array(
                    downlevel_flags.contains(wgt::DownlevelFlags::CUBE_ARRAY_TEXTURES),
                )
                .draw_indirect_first_instance(
                    requested_features.contains(wgt::Features::INDIRECT_FIRST_INSTANCE),
                )
                //.dual_src_blend(requested_features.contains(wgt::Features::DUAL_SRC_BLENDING))
                .multi_draw_indirect(phd_features.core.multi_draw_indirect != 0)
                .fill_mode_non_solid(requested_features.intersects(
                    wgt::Features::POLYGON_MODE_LINE | wgt::Features::POLYGON_MODE_POINT,
                ))
                //.depth_bounds(requested_features.contains(wgt::Features::DEPTH_BOUNDS))
                //.alpha_to_one(requested_features.contains(wgt::Features::ALPHA_TO_ONE))
                //.multi_viewport(requested_features.contains(wgt::Features::MULTI_VIEWPORTS))
                .sampler_anisotropy(
                    downlevel_flags.contains(wgt::DownlevelFlags::ANISOTROPIC_FILTERING),
                )
                .texture_compression_etc2(
                    requested_features.contains(wgt::Features::TEXTURE_COMPRESSION_ETC2),
                )
                .texture_compression_astc_ldr(
                    requested_features.contains(wgt::Features::TEXTURE_COMPRESSION_ASTC),
                )
                .texture_compression_bc(
                    requested_features.contains(wgt::Features::TEXTURE_COMPRESSION_BC),
                    // BC provides formats for Sliced 3D
                )
                //.occlusion_query_precise(requested_features.contains(wgt::Features::PRECISE_OCCLUSION_QUERY))
                .pipeline_statistics_query(
                    requested_features.contains(wgt::Features::PIPELINE_STATISTICS_QUERY),
                )
                .vertex_pipeline_stores_and_atomics(
                    requested_features.contains(wgt::Features::VERTEX_WRITABLE_STORAGE),
                )
                .fragment_stores_and_atomics(
                    downlevel_flags.contains(wgt::DownlevelFlags::FRAGMENT_WRITABLE_STORAGE),
                )
                //.shader_image_gather_extended(
                //.shader_storage_image_extended_formats(
                .shader_uniform_buffer_array_dynamic_indexing(
                    requested_features.contains(wgt::Features::BUFFER_BINDING_ARRAY),
                )
                .shader_storage_buffer_array_dynamic_indexing(requested_features.contains(
                    wgt::Features::BUFFER_BINDING_ARRAY
                        | wgt::Features::STORAGE_RESOURCE_BINDING_ARRAY,
                ))
                .shader_sampled_image_array_dynamic_indexing(
                    requested_features.contains(wgt::Features::TEXTURE_BINDING_ARRAY),
                )
                .shader_storage_buffer_array_dynamic_indexing(requested_features.contains(
                    wgt::Features::TEXTURE_BINDING_ARRAY
                        | wgt::Features::STORAGE_RESOURCE_BINDING_ARRAY,
                ))
                //.shader_storage_image_array_dynamic_indexing(
                .shader_clip_distance(requested_features.contains(wgt::Features::CLIP_DISTANCES))
                //.shader_cull_distance(requested_features.contains(wgt::Features::SHADER_CULL_DISTANCE))
                .shader_float64(requested_features.contains(wgt::Features::SHADER_F64))
                .shader_int64(requested_features.contains(wgt::Features::SHADER_INT64))
                .shader_int16(requested_features.contains(wgt::Features::SHADER_I16))
                //.shader_resource_residency(requested_features.contains(wgt::Features::SHADER_RESOURCE_RESIDENCY))
                .geometry_shader(requested_features.contains(wgt::Features::PRIMITIVE_INDEX))
                .depth_clamp(requested_features.contains(wgt::Features::DEPTH_CLIP_CONTROL))
                .dual_src_blend(requested_features.contains(wgt::Features::DUAL_SOURCE_BLENDING)),
            descriptor_indexing: if requested_features.intersects(INDEXING_FEATURES) {
                Some(
                    vk::PhysicalDeviceDescriptorIndexingFeaturesEXT::default()
                        .shader_sampled_image_array_non_uniform_indexing(needs_bindless)
                        .shader_storage_image_array_non_uniform_indexing(needs_bindless)
                        .shader_storage_buffer_array_non_uniform_indexing(needs_bindless)
                        .descriptor_binding_sampled_image_update_after_bind(needs_bindless)
                        .descriptor_binding_storage_image_update_after_bind(needs_bindless)
                        .descriptor_binding_storage_buffer_update_after_bind(needs_bindless)
                        .descriptor_binding_partially_bound(needs_partially_bound),
                )
            } else {
                None
            },
            timeline_semaphore: if device_api_version >= vk::API_VERSION_1_2
                || enabled_extensions.contains(&khr::timeline_semaphore::NAME)
            {
                Some(
                    vk::PhysicalDeviceTimelineSemaphoreFeaturesKHR::default()
                        .timeline_semaphore(private_caps.timeline_semaphores),
                )
            } else {
                None
            },
            image_robustness: if device_api_version >= vk::API_VERSION_1_3
                || enabled_extensions.contains(&ext::image_robustness::NAME)
            {
                Some(
                    vk::PhysicalDeviceImageRobustnessFeaturesEXT::default()
                        .robust_image_access(private_caps.robust_image_access),
                )
            } else {
                None
            },
            robustness2: if enabled_extensions.contains(&ext::robustness2::NAME) {
                Some(
                    vk::PhysicalDeviceRobustness2FeaturesEXT::default()
                        .robust_buffer_access2(private_caps.robust_buffer_access2)
                        .robust_image_access2(private_caps.robust_image_access2),
                )
            } else {
                None
            },
            multiview: if device_api_version >= vk::API_VERSION_1_1
                || enabled_extensions.contains(&khr::multiview::NAME)
            {
                Some(
                    vk::PhysicalDeviceMultiviewFeatures::default()
                        .multiview(requested_features.contains(wgt::Features::MULTIVIEW)),
                )
            } else {
                None
            },
            sampler_ycbcr_conversion: if device_api_version >= vk::API_VERSION_1_1
                || enabled_extensions.contains(&khr::sampler_ycbcr_conversion::NAME)
            {
                Some(
                    vk::PhysicalDeviceSamplerYcbcrConversionFeatures::default(), // .sampler_ycbcr_conversion(requested_features.contains(wgt::Features::TEXTURE_FORMAT_NV12))
                )
            } else {
                None
            },
            astc_hdr: if enabled_extensions.contains(&ext::texture_compression_astc_hdr::NAME) {
                Some(
                    vk::PhysicalDeviceTextureCompressionASTCHDRFeaturesEXT::default()
                        .texture_compression_astc_hdr(true),
                )
            } else {
                None
            },
            shader_float16_int8: match requested_features.contains(wgt::Features::SHADER_F16) {
                shader_float16 if shader_float16 || private_caps.shader_int8 => Some(
                    vk::PhysicalDeviceShaderFloat16Int8Features::default()
                        .shader_float16(shader_float16)
                        .shader_int8(private_caps.shader_int8),
                ),
                _ => None,
            },
            _16bit_storage: if requested_features
                .intersects(wgt::Features::SHADER_F16 | wgt::Features::SHADER_I16)
            {
                Some(
                    vk::PhysicalDevice16BitStorageFeatures::default()
                        .storage_buffer16_bit_access(true)
                        .storage_input_output16(phd_features.supports_storage_input_output_16())
                        .uniform_and_storage_buffer16_bit_access(true),
                )
            } else {
                None
            },
            acceleration_structure: if enabled_extensions
                .contains(&khr::acceleration_structure::NAME)
            {
                Some(
                    vk::PhysicalDeviceAccelerationStructureFeaturesKHR::default()
                        .acceleration_structure(true)
                        .descriptor_binding_acceleration_structure_update_after_bind(
                            requested_features
                                .contains(wgt::Features::ACCELERATION_STRUCTURE_BINDING_ARRAY),
                        ),
                )
            } else {
                None
            },
            buffer_device_address: if enabled_extensions.contains(&khr::buffer_device_address::NAME)
            {
                Some(
                    vk::PhysicalDeviceBufferDeviceAddressFeaturesKHR::default()
                        .buffer_device_address(true),
                )
            } else {
                None
            },
            ray_query: if enabled_extensions.contains(&khr::ray_query::NAME) {
                Some(vk::PhysicalDeviceRayQueryFeaturesKHR::default().ray_query(true))
            } else {
                None
            },
            ray_tracing_pipeline: if enabled_extensions.contains(&khr::ray_tracing_pipeline::NAME) {
                Some(
                    vk::PhysicalDeviceRayTracingPipelineFeaturesKHR::default()
                        .ray_tracing_pipeline(true),
                )
            } else {
                None
            },
            zero_initialize_workgroup_memory: if device_api_version >= vk::API_VERSION_1_3
                || enabled_extensions.contains(&khr::zero_initialize_workgroup_memory::NAME)
            {
                Some(
                    vk::PhysicalDeviceZeroInitializeWorkgroupMemoryFeatures::default()
                        .shader_zero_initialize_workgroup_memory(
                            private_caps.zero_initialize_workgroup_memory,
                        ),
                )
            } else {
                None
            },
            shader_atomic_int64: if device_api_version >= vk::API_VERSION_1_2
                || enabled_extensions.contains(&khr::shader_atomic_int64::NAME)
            {
                let needed = requested_features.intersects(
                    wgt::Features::SHADER_INT64_ATOMIC_ALL_OPS
                        | wgt::Features::SHADER_INT64_ATOMIC_MIN_MAX,
                );
                Some(
                    vk::PhysicalDeviceShaderAtomicInt64Features::default()
                        .shader_buffer_int64_atomics(needed)
                        .shader_shared_int64_atomics(needed),
                )
            } else {
                None
            },
            shader_image_atomic_int64: if enabled_extensions
                .contains(&ext::shader_image_atomic_int64::NAME)
            {
                let needed = requested_features.intersects(wgt::Features::TEXTURE_INT64_ATOMIC);
                Some(
                    vk::PhysicalDeviceShaderImageAtomicInt64FeaturesEXT::default()
                        .shader_image_int64_atomics(needed),
                )
            } else {
                None
            },
            shader_atomic_float: if enabled_extensions.contains(&ext::shader_atomic_float::NAME) {
                let needed = requested_features.contains(wgt::Features::SHADER_FLOAT32_ATOMIC);
                Some(
                    vk::PhysicalDeviceShaderAtomicFloatFeaturesEXT::default()
                        .shader_buffer_float32_atomics(needed)
                        .shader_buffer_float32_atomic_add(needed),
                )
            } else {
                None
            },
            subgroup_size_control: if device_api_version >= vk::API_VERSION_1_3
                || enabled_extensions.contains(&ext::subgroup_size_control::NAME)
            {
                Some(
                    vk::PhysicalDeviceSubgroupSizeControlFeatures::default()
                        .subgroup_size_control(true),
                )
            } else {
                None
            },
            position_fetch: if enabled_extensions.contains(&khr::ray_tracing_position_fetch::NAME) {
                Some(
                    vk::PhysicalDeviceRayTracingPositionFetchFeaturesKHR::default()
                        .ray_tracing_position_fetch(true),
                )
            } else {
                None
            },
            mesh_shader: if enabled_extensions.contains(&ext::mesh_shader::NAME) {
                let needed = requested_features.contains(wgt::Features::EXPERIMENTAL_MESH_SHADER);
                let multiview_needed =
                    requested_features.contains(wgt::Features::EXPERIMENTAL_MESH_SHADER_MULTIVIEW);
                Some(
                    vk::PhysicalDeviceMeshShaderFeaturesEXT::default()
                        .mesh_shader(needed)
                        .task_shader(needed)
                        .multiview_mesh_shader(multiview_needed),
                )
            } else {
                None
            },
            maintenance4: if device_api_version >= vk::API_VERSION_1_3
                || enabled_extensions.contains(&khr::maintenance4::NAME)
            {
                let needed = requested_features.contains(wgt::Features::EXPERIMENTAL_MESH_SHADER);
                Some(vk::PhysicalDeviceMaintenance4Features::default().maintenance4(needed))
            } else {
                None
            },
            shader_integer_dot_product: if device_api_version >= vk::API_VERSION_1_3
                || enabled_extensions.contains(&khr::shader_integer_dot_product::NAME)
            {
                Some(
                    vk::PhysicalDeviceShaderIntegerDotProductFeaturesKHR::default()
                        .shader_integer_dot_product(private_caps.shader_integer_dot_product),
                )
            } else {
                None
            },
            shader_barycentrics: if enabled_extensions
                .contains(&khr::fragment_shader_barycentric::NAME)
            {
                let needed = requested_features.intersects(
                    wgt::Features::SHADER_BARYCENTRICS | wgt::Features::SHADER_PER_VERTEX,
                );
                Some(
                    vk::PhysicalDeviceFragmentShaderBarycentricFeaturesKHR::default()
                        .fragment_shader_barycentric(needed),
                )
            } else {
                None
            },
            portability_subset: if enabled_extensions.contains(&khr::portability_subset::NAME) {
                let image_view_format_swizzle_needed =
                    requested_features.intersects(wgt::Features::TEXTURE_COMPONENT_SWIZZLE);
                let multisample_array_needed =
                    requested_features.intersects(wgt::Features::MULTISAMPLE_ARRAY);

                Some(
                    vk::PhysicalDevicePortabilitySubsetFeaturesKHR::default()
                        .image_view_format_swizzle(image_view_format_swizzle_needed)
                        .multisample_array_image(multisample_array_needed),
                )
            } else {
                None
            },
            cooperative_matrix: if enabled_extensions.contains(&khr::cooperative_matrix::NAME) {
                let needed =
                    requested_features.contains(wgt::Features::EXPERIMENTAL_COOPERATIVE_MATRIX);
                Some(
                    vk::PhysicalDeviceCooperativeMatrixFeaturesKHR::default()
                        .cooperative_matrix(needed),
                )
            } else {
                None
            },
            vulkan_memory_model: if device_api_version >= vk::API_VERSION_1_2
                || enabled_extensions.contains(&khr::vulkan_memory_model::NAME)
            {
                let needed =
                    requested_features.contains(wgt::Features::EXPERIMENTAL_COOPERATIVE_MATRIX);
                Some(
                    vk::PhysicalDeviceVulkanMemoryModelFeaturesKHR::default()
                        .vulkan_memory_model(needed)
                        // The SPIR-V backend emits storage atomics with `Device`
                        // memory scope, so whenever the Vulkan memory model is
                        // enabled we must also enable device scope, otherwise the
                        // validation layers report
                        // `VUID-RuntimeSpirv-vulkanMemoryModel-06265`.
                        .vulkan_memory_model_device_scope(needed),
                )
            } else {
                None
            },
            shader_draw_parameters: if device_api_version >= vk::API_VERSION_1_1 {
                let needed = requested_features.contains(wgt::Features::SHADER_DRAW_INDEX);
                Some(
                    vk::PhysicalDeviceShaderDrawParametersFeatures::default()
                        .shader_draw_parameters(needed),
                )
            } else {
                None
            },
        }
    }

    /// Compute the wgpu [`Features`] and [`DownlevelFlags`] supported by a physical device.
    ///
    /// Given `self`, together with the instance and physical device it was
    /// built from, and a `caps` also built from those, determine which wgpu
    /// features and downlevel flags the device can support.
    ///
    /// [`Features`]: wgt::Features
    /// [`DownlevelFlags`]: wgt::DownlevelFlags
    fn to_wgpu(
        &self,
        instance: &ash::Instance,
        phd: vk::PhysicalDevice,
        caps: &PhysicalDeviceProperties,
        queue_props: &vk::QueueFamilyProperties,
    ) -> (wgt::Features, wgt::DownlevelFlags) {
        use wgt::{DownlevelFlags as Df, Features as F};
        let mut features = F::empty()
            | F::ADDRESS_MODE_CLAMP_TO_BORDER
            | F::ADDRESS_MODE_CLAMP_TO_ZERO
            | F::CLEAR_TEXTURE
            | F::IMMEDIATES
            | F::MAPPABLE_PRIMARY_BUFFERS
            | F::MEMORY_DECORATION_COHERENT
            | F::MEMORY_DECORATION_VOLATILE
            | F::PASSTHROUGH_SHADERS
            | F::PIPELINE_CACHE
            | F::SHADER_EARLY_DEPTH_TEST
            | F::TEXTURE_ADAPTER_SPECIFIC_FORMAT_FEATURES
            | F::TEXTURE_ATOMIC;

        let mut dl_flags = Df::empty()
            | Df::BASE_VERTEX
            | Df::BUFFER_BINDINGS_NOT_16_BYTE_ALIGNED
            | Df::COMPARISON_SAMPLERS
            | Df::COMPUTE_SHADERS
            | Df::DEPTH_TEXTURE_AND_BUFFER_COPIES
            | Df::FRAGMENT_STORAGE
            | Df::INDIRECT_EXECUTION
            | Df::LINEAR_INTERPOLATION
            | Df::MSL2_1
            | Df::NONBLOCKING_QUERY_RESOLVE
            | Df::NON_POWER_OF_TWO_MIPMAPPED_TEXTURES
            | Df::SHADER_F16_IN_F32
            | Df::UNRESTRICTED_EXTERNAL_TEXTURE_COPIES
            | Df::UNRESTRICTED_INDEX_BUFFER
            | Df::VERTEX_STORAGE
            | Df::VIEW_FORMATS;

        // `VK_IMAGE_LAYOUT_DEPTH_READ_ONLY_STENCIL_ATTACHMENT_OPTIMAL`
        // and `VK_IMAGE_LAYOUT_DEPTH_ATTACHMENT_STENCIL_READ_ONLY_OPTIMAL`
        // is required for separate read-only depth stencil.
        dl_flags.set(
            Df::READ_ONLY_DEPTH_STENCIL,
            caps.device_api_version >= vk::API_VERSION_1_1
                || caps.supports_extension(khr::maintenance2::NAME),
        );

        dl_flags.set(
            Df::SURFACE_VIEW_FORMATS,
            caps.supports_extension(khr::swapchain_mutable_format::NAME)
                || !caps.supports_extension(khr::swapchain::NAME),
        );
        dl_flags.set(Df::CUBE_ARRAY_TEXTURES, self.core.image_cube_array != 0);
        dl_flags.set(Df::ANISOTROPIC_FILTERING, self.core.sampler_anisotropy != 0);
        dl_flags.set(
            Df::FRAGMENT_WRITABLE_STORAGE,
            self.core.fragment_stores_and_atomics != 0,
        );
        dl_flags.set(Df::MULTISAMPLED_SHADING, self.core.sample_rate_shading != 0);
        dl_flags.set(Df::INDEPENDENT_BLEND, self.core.independent_blend != 0);
        dl_flags.set(
            Df::FULL_DRAW_INDEX_UINT32,
            self.core.full_draw_index_uint32 != 0,
        );
        dl_flags.set(Df::DEPTH_BIAS_CLAMP, self.core.depth_bias_clamp != 0);

        features.set(
            F::TIMESTAMP_QUERY
                | F::TIMESTAMP_QUERY_INSIDE_ENCODERS
                | F::TIMESTAMP_QUERY_INSIDE_PASSES,
            // Vulkan strictly defines this as either 36-64, or zero.
            queue_props.timestamp_valid_bits >= 36,
        );
        features.set(
            F::INDIRECT_FIRST_INSTANCE,
            self.core.draw_indirect_first_instance != 0,
        );
        //if self.core.dual_src_blend != 0
        features.set(F::POLYGON_MODE_LINE, self.core.fill_mode_non_solid != 0);
        features.set(F::POLYGON_MODE_POINT, self.core.fill_mode_non_solid != 0);
        //if self.core.depth_bounds != 0 {
        //if self.core.alpha_to_one != 0 {
        //if self.core.multi_viewport != 0 {
        features.set(
            F::TEXTURE_COMPRESSION_ETC2,
            self.core.texture_compression_etc2 != 0,
        );
        features.set(
            F::TEXTURE_COMPRESSION_ASTC,
            self.core.texture_compression_astc_ldr != 0,
        );
        features.set(
            F::TEXTURE_COMPRESSION_BC,
            self.core.texture_compression_bc != 0,
        );
        features.set(
            F::TEXTURE_COMPRESSION_BC_SLICED_3D,
            self.core.texture_compression_bc != 0, // BC guarantees Sliced 3D
        );
        features.set(
            F::PIPELINE_STATISTICS_QUERY,
            self.core.pipeline_statistics_query != 0,
        );
        features.set(
            F::VERTEX_WRITABLE_STORAGE,
            self.core.vertex_pipeline_stores_and_atomics != 0,
        );

        features.set(F::SHADER_F64, self.core.shader_float64 != 0);
        features.set(F::SHADER_INT64, self.core.shader_int64 != 0);
        if let Some(ref bit16) = self._16bit_storage {
            features.set(
                F::SHADER_I16,
                self.core.shader_int16 != 0
                    && bit16.storage_buffer16_bit_access != 0
                    && bit16.uniform_and_storage_buffer16_bit_access != 0,
            );
        }

        features.set(F::PRIMITIVE_INDEX, self.core.geometry_shader != 0);

        if let Some(ref shader_atomic_int64) = self.shader_atomic_int64 {
            features.set(
                F::SHADER_INT64_ATOMIC_ALL_OPS | F::SHADER_INT64_ATOMIC_MIN_MAX,
                shader_atomic_int64.shader_buffer_int64_atomics != 0
                    && shader_atomic_int64.shader_shared_int64_atomics != 0,
            );
        }

        if let Some(ref shader_image_atomic_int64) = self.shader_image_atomic_int64 {
            features.set(
                F::TEXTURE_INT64_ATOMIC,
                shader_image_atomic_int64
                    .shader_image_int64_atomics(true)
                    .shader_image_int64_atomics
                    != 0,
            );
        }

        if let Some(ref shader_atomic_float) = self.shader_atomic_float {
            features.set(
                F::SHADER_FLOAT32_ATOMIC,
                shader_atomic_float.shader_buffer_float32_atomics != 0
                    && shader_atomic_float.shader_buffer_float32_atomic_add != 0,
            );
        }

        if let Some(ref shader_barycentrics) = self.shader_barycentrics {
            features.set(
                F::SHADER_BARYCENTRICS | F::SHADER_PER_VERTEX,
                shader_barycentrics.fragment_shader_barycentric != 0,
            );
        }

        //if caps.supports_extension(khr::sampler_mirror_clamp_to_edge::NAME) {
        //if caps.supports_extension(ext::sampler_filter_minmax::NAME) {
        features.set(
            F::MULTI_DRAW_INDIRECT_COUNT,
            caps.supports_extension(khr::draw_indirect_count::NAME),
        );
        features.set(
            F::CONSERVATIVE_RASTERIZATION,
            caps.supports_extension(ext::conservative_rasterization::NAME),
        );
        features.set(
            F::EXPERIMENTAL_RAY_HIT_VERTEX_RETURN,
            caps.supports_extension(khr::ray_tracing_position_fetch::NAME),
        );

        if let Some(ref descriptor_indexing) = self.descriptor_indexing {
            // We use update-after-bind descriptors for all bind groups containing binding arrays.
            //
            // In those bind groups, we allow all binding types except uniform buffers to be present.
            //
            // As we can only switch between update-after-bind and not on a per bind group basis,
            // all supported binding types need to be able to be marked update after bind.
            //
            // As such, we enable all features as a whole, rather individually.
            let supports_descriptor_indexing =
                // Sampled Images
                descriptor_indexing.shader_sampled_image_array_non_uniform_indexing != 0
                    && descriptor_indexing.descriptor_binding_sampled_image_update_after_bind != 0
                    // Storage Images
                    && descriptor_indexing.shader_storage_image_array_non_uniform_indexing != 0
                    && descriptor_indexing.descriptor_binding_storage_image_update_after_bind != 0
                    // Storage Buffers
                    && descriptor_indexing.shader_storage_buffer_array_non_uniform_indexing != 0
                    && descriptor_indexing.descriptor_binding_storage_buffer_update_after_bind != 0;

            let descriptor_indexing_features = F::BUFFER_BINDING_ARRAY
                | F::TEXTURE_BINDING_ARRAY
                | F::STORAGE_RESOURCE_BINDING_ARRAY
                | F::SAMPLED_TEXTURE_AND_STORAGE_BUFFER_ARRAY_NON_UNIFORM_INDEXING
                | F::STORAGE_TEXTURE_ARRAY_NON_UNIFORM_INDEXING;

            features.set(descriptor_indexing_features, supports_descriptor_indexing);

            let supports_partially_bound =
                descriptor_indexing.descriptor_binding_partially_bound != 0;

            features.set(F::PARTIALLY_BOUND_BINDING_ARRAY, supports_partially_bound);
        }

        features.set(F::DEPTH_CLIP_CONTROL, self.core.depth_clamp != 0);
        features.set(F::DUAL_SOURCE_BLENDING, self.core.dual_src_blend != 0);
        features.set(F::CLIP_DISTANCES, self.core.shader_clip_distance != 0);

        if let Some(ref multiview) = self.multiview {
            features.set(F::MULTIVIEW, multiview.multiview != 0);
            features.set(F::SELECTIVE_MULTIVIEW, multiview.multiview != 0);
        }

        features.set(
            F::TEXTURE_FORMAT_16BIT_NORM,
            is_format_16bit_norm_supported(instance, phd),
        );

        if let Some(ref astc_hdr) = self.astc_hdr {
            features.set(
                F::TEXTURE_COMPRESSION_ASTC_HDR,
                astc_hdr.texture_compression_astc_hdr != 0,
            );
        }

        if self.core.texture_compression_astc_ldr != 0 {
            features.set(
                F::TEXTURE_COMPRESSION_ASTC_SLICED_3D,
                supports_astc_3d(instance, phd),
            );
        }

        if let (Some(ref f16_i8), Some(ref bit16)) = (self.shader_float16_int8, self._16bit_storage)
        {
            // Note `storage_input_output16` is not required, we polyfill `f16` I/O using `f32`
            // types when this capability is not available
            features.set(
                F::SHADER_F16,
                f16_i8.shader_float16 != 0
                    && bit16.storage_buffer16_bit_access != 0
                    && bit16.uniform_and_storage_buffer16_bit_access != 0,
            );
        }

        if let Some(ref subgroup) = caps.subgroup {
            if (caps.device_api_version >= vk::API_VERSION_1_3
                || caps.supports_extension(ext::subgroup_size_control::NAME))
                && subgroup.supported_operations.contains(
                    vk::SubgroupFeatureFlags::BASIC
                        | vk::SubgroupFeatureFlags::VOTE
                        | vk::SubgroupFeatureFlags::ARITHMETIC
                        | vk::SubgroupFeatureFlags::BALLOT
                        | vk::SubgroupFeatureFlags::SHUFFLE
                        | vk::SubgroupFeatureFlags::SHUFFLE_RELATIVE
                        | vk::SubgroupFeatureFlags::QUAD,
                )
            {
                features.set(
                    F::SUBGROUP,
                    subgroup
                        .supported_stages
                        .contains(vk::ShaderStageFlags::COMPUTE | vk::ShaderStageFlags::FRAGMENT),
                );
                features.set(
                    F::SUBGROUP_VERTEX,
                    subgroup
                        .supported_stages
                        .contains(vk::ShaderStageFlags::VERTEX),
                );
                features.insert(F::SUBGROUP_BARRIER);
            }
        }

        let supports_depth_format = |format| {
            supports_format(
                instance,
                phd,
                format,
                vk::ImageTiling::OPTIMAL,
                depth_stencil_required_flags(),
            )
        };

        let texture_s8 = supports_depth_format(vk::Format::S8_UINT);
        let texture_d32 = supports_depth_format(vk::Format::D32_SFLOAT);
        let texture_d24_s8 = supports_depth_format(vk::Format::D24_UNORM_S8_UINT);
        let texture_d32_s8 = supports_depth_format(vk::Format::D32_SFLOAT_S8_UINT);

        let stencil8 = texture_s8 || texture_d24_s8;
        let depth24_plus_stencil8 = texture_d24_s8 || texture_d32_s8;

        dl_flags.set(
            Df::WEBGPU_TEXTURE_FORMAT_SUPPORT,
            stencil8 && depth24_plus_stencil8 && texture_d32,
        );

        features.set(F::DEPTH32FLOAT_STENCIL8, texture_d32_s8);

        let supports_acceleration_structures = caps
            .supports_extension(khr::deferred_host_operations::NAME)
            && caps.supports_extension(khr::acceleration_structure::NAME)
            && caps.supports_extension(khr::buffer_device_address::NAME);

        let supports_ray_query =
            supports_acceleration_structures && caps.supports_extension(khr::ray_query::NAME);
        let supports_acceleration_structure_binding_array = supports_ray_query
            && self
                .acceleration_structure
                .as_ref()
                .is_some_and(|features| {
                    features.descriptor_binding_acceleration_structure_update_after_bind != 0
                });

        features.set(
            F::EXPERIMENTAL_RAY_QUERY
            // Although this doesn't really require ray queries, it does not make sense to be enabled if acceleration structures
            // aren't enabled.
                | F::EXTENDED_ACCELERATION_STRUCTURE_VERTEX_FORMATS,
            supports_ray_query,
        );

        // Binding arrays of TLAS are supported on Vulkan when ray queries are supported.
        //
        // Note: this flag is used for shader-side `binding_array<acceleration_structure>` as well as
        // allowing `BindGroupLayoutEntry::count = Some(...)` for `BindingType::AccelerationStructure`.
        features.set(
            F::ACCELERATION_STRUCTURE_BINDING_ARRAY,
            supports_acceleration_structure_binding_array,
        );

        if supports_acceleration_structures
            && caps.supports_extension(khr::ray_tracing_pipeline::NAME)
        {
            features.insert(
                F::EXPERIMENTAL_RAY_TRACING_PIPELINES
                        // Same reason as for ray queries.
                        | F::EXTENDED_ACCELERATION_STRUCTURE_VERTEX_FORMATS,
            );
        }

        let rg11b10ufloat_renderable = supports_format(
            instance,
            phd,
            vk::Format::B10G11R11_UFLOAT_PACK32,
            vk::ImageTiling::OPTIMAL,
            vk::FormatFeatureFlags::COLOR_ATTACHMENT
                | vk::FormatFeatureFlags::COLOR_ATTACHMENT_BLEND,
        );
        features.set(F::RG11B10UFLOAT_RENDERABLE, rg11b10ufloat_renderable);

        features.set(
            F::BGRA8UNORM_STORAGE,
            supports_bgra8unorm_storage(instance, phd, caps.device_api_version),
        );

        features.set(
            F::FLOAT32_FILTERABLE,
            is_float32_filterable_supported(instance, phd),
        );

        features.set(
            F::FLOAT32_BLENDABLE,
            is_float32_blendable_supported(instance, phd),
        );

        if let Some(ref _sampler_ycbcr_conversion) = self.sampler_ycbcr_conversion {
            features.set(
                F::TEXTURE_FORMAT_NV12,
                supports_format(
                    instance,
                    phd,
                    vk::Format::G8_B8R8_2PLANE_420_UNORM,
                    vk::ImageTiling::OPTIMAL,
                    vk::FormatFeatureFlags::SAMPLED_IMAGE
                        | vk::FormatFeatureFlags::TRANSFER_SRC
                        | vk::FormatFeatureFlags::TRANSFER_DST,
                ) && !caps.is_driver(vk::DriverId::MOLTENVK),
            );
        }

        if let Some(ref _sampler_ycbcr_conversion) = self.sampler_ycbcr_conversion {
            features.set(
                F::TEXTURE_FORMAT_P010,
                supports_format(
                    instance,
                    phd,
                    vk::Format::G10X6_B10X6R10X6_2PLANE_420_UNORM_3PACK16,
                    vk::ImageTiling::OPTIMAL,
                    vk::FormatFeatureFlags::SAMPLED_IMAGE
                        | vk::FormatFeatureFlags::TRANSFER_SRC
                        | vk::FormatFeatureFlags::TRANSFER_DST,
                ) && !caps.is_driver(vk::DriverId::MOLTENVK),
            );
        }

        features.set(
            F::VULKAN_GOOGLE_DISPLAY_TIMING,
            caps.supports_extension(google::display_timing::NAME),
        );

        features.set(
            F::VULKAN_EXTERNAL_MEMORY_WIN32,
            caps.supports_extension(khr::external_memory_win32::NAME),
        );
        features.set(
            F::VULKAN_EXTERNAL_MEMORY_FD,
            caps.supports_extension(khr::external_memory_fd::NAME),
        );
        features.set(
            F::VULKAN_EXTERNAL_MEMORY_DMA_BUF,
            caps.supports_extension(khr::external_memory_fd::NAME)
                && caps.supports_extension(ext::external_memory_dma_buf::NAME)
                && caps.supports_extension(ext::image_drm_format_modifier::NAME),
        );
        features.set(
            F::EXPERIMENTAL_MESH_SHADER,
            caps.supports_extension(ext::mesh_shader::NAME),
        );
        features.set(
            F::EXPERIMENTAL_MESH_SHADER_POINTS,
            caps.supports_extension(ext::mesh_shader::NAME),
        );
        if let Some(ref mesh_shader) = self.mesh_shader {
            features.set(
                F::EXPERIMENTAL_MESH_SHADER_MULTIVIEW,
                mesh_shader.multiview_mesh_shader != 0,
            );
        }

        // Not supported by default by `VK_KHR_portability_subset`, which we use on apple platforms.
        features.set(
            F::TEXTURE_COMPONENT_SWIZZLE,
            self.portability_subset
                .map(|p| p.image_view_format_swizzle == vk::TRUE)
                .unwrap_or(true),
        );
        features.set(
            F::MULTISAMPLE_ARRAY,
            self.portability_subset
                .map(|p| p.multisample_array_image == vk::TRUE)
                .unwrap_or(true),
        );

        // Enable cooperative matrix if any configuration is supported. The SPIR-V
        // we emit for it uses `Device`-scope atomics under the Vulkan memory model,
        // so the device must also support `vulkanMemoryModelDeviceScope`, otherwise
        // those shaders trip `VUID-RuntimeSpirv-vulkanMemoryModel-06265`.
        features.set(
            F::EXPERIMENTAL_COOPERATIVE_MATRIX,
            !caps.cooperative_matrix_properties.is_empty()
                && self.vulkan_memory_model.is_some_and(|m| {
                    m.vulkan_memory_model == vk::TRUE
                        && m.vulkan_memory_model_device_scope == vk::TRUE
                }),
        );

        features.set(
            F::SHADER_DRAW_INDEX,
            self.shader_draw_parameters
                .is_some_and(|a| a.shader_draw_parameters != 0)
                || caps.supports_extension(c"VK_KHR_shader_draw_parameters"),
        );

        features.set(
            F::DEBUG_PRINTF,
            caps.device_api_version >= vk::API_VERSION_1_3
                || caps.supports_extension(khr::shader_non_semantic_info::NAME),
        );

        (features, dl_flags)
    }
}

/// Vulkan "properties" structures gathered about a physical device.
///
/// This structure holds the properties of a [`vk::PhysicalDevice`]:
/// - the standard Vulkan device properties
/// - the `VkExtensionProperties` structs for all available extensions, and
/// - the per-extension properties structures for the available extensions that
///   `wgpu` cares about.
///
/// Generally, if you get it from any of these functions, it's stored
/// here:
/// - `vkEnumerateDeviceExtensionProperties`
/// - `vkGetPhysicalDeviceProperties`
/// - `vkGetPhysicalDeviceProperties2`
///
/// This also includes a copy of the device API version, since we can
/// use that as a shortcut for searching for an extension, if the
/// extension has been promoted to core in the current version.
///
/// This does not include device features; for those, see
/// [`PhysicalDeviceFeatures`].
#[derive(Default, Debug)]
pub struct PhysicalDeviceProperties {
    /// Extensions supported by the `vk::PhysicalDevice`,
    /// as returned by `vkEnumerateDeviceExtensionProperties`.
    supported_extensions: Vec<vk::ExtensionProperties>,

    /// Properties of the `vk::PhysicalDevice`, as returned by
    /// `vkGetPhysicalDeviceProperties`.
    properties: vk::PhysicalDeviceProperties,

    /// Additional `vk::PhysicalDevice` properties from the
    /// `VK_KHR_maintenance3` extension, promoted to Vulkan 1.1.
    maintenance_3: Option<vk::PhysicalDeviceMaintenance3Properties<'static>>,

    /// Additional `vk::PhysicalDevice` properties from the
    /// `VK_KHR_maintenance4` extension, promoted to Vulkan 1.3.
    maintenance_4: Option<vk::PhysicalDeviceMaintenance4Properties<'static>>,

    /// Additional `vk::PhysicalDevice` properties from the
    /// `VK_KHR_maintenance5` extension, promoted to Vulkan 1.4.
    maintenance_5: Option<vk::PhysicalDeviceMaintenance5PropertiesKHR<'static>>,

    /// Additional `vk::PhysicalDevice` properties from the
    /// `VK_EXT_descriptor_indexing` extension, promoted to Vulkan 1.2.
    descriptor_indexing: Option<vk::PhysicalDeviceDescriptorIndexingPropertiesEXT<'static>>,

    /// Additional `vk::PhysicalDevice` properties from the
    /// `VK_KHR_acceleration_structure` extension.
    acceleration_structure: Option<vk::PhysicalDeviceAccelerationStructurePropertiesKHR<'static>>,

    /// Additional `vk::PhysicalDevice` properties from the
    /// `VK_KHR_ray_tracing_pipeline` extension.
    ray_tracing_pipeline: Option<vk::PhysicalDeviceRayTracingPipelinePropertiesKHR<'static>>,

    /// Additional `vk::PhysicalDevice` properties from the
    /// `VK_KHR_driver_properties` extension, promoted to Vulkan 1.2.
    driver: Option<vk::PhysicalDeviceDriverPropertiesKHR<'static>>,

    /// Additional `vk::PhysicalDevice` properties from Vulkan 1.1.
    subgroup: Option<vk::PhysicalDeviceSubgroupProperties<'static>>,

    /// Additional `vk::PhysicalDevice` properties from the
    /// `VK_EXT_subgroup_size_control` extension, promoted to Vulkan 1.3.
    subgroup_size_control: Option<vk::PhysicalDeviceSubgroupSizeControlProperties<'static>>,

    /// Additional `vk::PhysicalDevice` properties from the
    /// `VK_EXT_robustness2` extension.
    robustness2: Option<vk::PhysicalDeviceRobustness2PropertiesEXT<'static>>,

    /// Additional `vk::PhysicalDevice` properties from the
    /// `VK_EXT_mesh_shader` extension.
    mesh_shader: Option<vk::PhysicalDeviceMeshShaderPropertiesEXT<'static>>,

    /// Additional `vk::PhysicalDevice` properties from the
    /// `VK_KHR_multiview` extension.
    multiview: Option<vk::PhysicalDeviceMultiviewPropertiesKHR<'static>>,

    /// `VK_EXT_pci_bus_info` extension.
    pci_bus_info: Option<vk::PhysicalDevicePCIBusInfoPropertiesEXT<'static>>,

    /// The device API version.
    ///
    /// Which is the version of Vulkan supported for device-level functionality.
    ///
    /// It is associated with a `VkPhysicalDevice` and its children.
    device_api_version: u32,

    /// Supported cooperative matrix configurations.
    ///
    /// This is determined by querying `vkGetPhysicalDeviceCooperativeMatrixPropertiesKHR`.
    cooperative_matrix_properties: Vec<wgt::CooperativeMatrixProperties>,
}

impl PhysicalDeviceProperties {
    pub fn properties(&self) -> vk::PhysicalDeviceProperties {
        self.properties
    }

    pub fn supports_extension(&self, extension: &CStr) -> bool {
        self.supported_extensions
            .iter()
            .any(|ep| ep.extension_name_as_c_str() == Ok(extension))
    }

    pub fn is_driver(&self, id: vk::DriverId) -> bool {
        self.driver.is_some_and(|driver| driver.driver_id == id)
    }

    /// Map `requested_features` to the list of Vulkan extension strings required to create the logical device.
    fn get_required_extensions(&self, requested_features: wgt::Features) -> Vec<&'static CStr> {
        let mut extensions = Vec::new();

        // Note that quite a few extensions depend on the `VK_KHR_get_physical_device_properties2` instance extension.
        // We enable `VK_KHR_get_physical_device_properties2` unconditionally (if available).

        // Require `VK_KHR_swapchain`
        extensions.push(khr::swapchain::NAME);

        if self.device_api_version < vk::API_VERSION_1_1 {
            // Require `VK_KHR_maintenance1`
            extensions.push(khr::maintenance1::NAME);

            // Optional `VK_KHR_maintenance2`
            if self.supports_extension(khr::maintenance2::NAME) {
                extensions.push(khr::maintenance2::NAME);
            }

            // Optional `VK_KHR_maintenance3`
            if self.supports_extension(khr::maintenance3::NAME) {
                extensions.push(khr::maintenance3::NAME);
            }

            // Require `VK_KHR_storage_buffer_storage_class`
            extensions.push(khr::storage_buffer_storage_class::NAME);

            // Require `VK_KHR_multiview` if the associated feature was requested
            if requested_features.contains(wgt::Features::MULTIVIEW) {
                extensions.push(khr::multiview::NAME);
            }

            // Require `VK_KHR_sampler_ycbcr_conversion` if the associated feature was requested
            if requested_features.contains(wgt::Features::TEXTURE_FORMAT_NV12) {
                extensions.push(khr::sampler_ycbcr_conversion::NAME);
            }

            // Require `VK_KHR_16bit_storage` if `SHADER_F16` or `SHADER_I16` was requested
            if requested_features.intersects(wgt::Features::SHADER_F16 | wgt::Features::SHADER_I16)
            {
                // - Feature `SHADER_F16` also requires `VK_KHR_shader_float16_int8`, but we always
                //   require that anyway (if it is available) below.
                // - `VK_KHR_16bit_storage` requires `VK_KHR_storage_buffer_storage_class`, however
                //   we require that one already.
                extensions.push(khr::_16bit_storage::NAME);
            }

            if requested_features.contains(wgt::Features::SHADER_DRAW_INDEX) {
                extensions.push(khr::shader_draw_parameters::NAME);
            }
        }

        if self.device_api_version < vk::API_VERSION_1_2 {
            // Optional `VK_KHR_image_format_list`
            if self.supports_extension(khr::image_format_list::NAME) {
                extensions.push(khr::image_format_list::NAME);
            }

            // Optional `VK_KHR_driver_properties`
            if self.supports_extension(khr::driver_properties::NAME) {
                extensions.push(khr::driver_properties::NAME);
            }

            // Optional `VK_KHR_timeline_semaphore`
            if self.supports_extension(khr::timeline_semaphore::NAME) {
                extensions.push(khr::timeline_semaphore::NAME);
            }

            // Require `VK_EXT_descriptor_indexing` if one of the associated features was requested
            if requested_features.intersects(INDEXING_FEATURES) {
                extensions.push(ext::descriptor_indexing::NAME);
            }

            // Always require `VK_KHR_shader_float16_int8` if available as it enables
            // Int8 optimizations. Also require it even if it's not available but
            // requested so that we get a corresponding error message.
            if requested_features.contains(wgt::Features::SHADER_F16)
                || self.supports_extension(khr::shader_float16_int8::NAME)
            {
                extensions.push(khr::shader_float16_int8::NAME);
            }

            // `SPV_EXT_mesh_shader` and `SPV_KHR_ray_tracing` both require SPIR-V 1.4.
            if requested_features.intersects(
                wgt::Features::EXPERIMENTAL_MESH_SHADER
                    | wgt::Features::EXPERIMENTAL_RAY_TRACING_PIPELINES,
            ) {
                extensions.push(khr::spirv_1_4::NAME);
            }

            //extensions.push(khr::sampler_mirror_clamp_to_edge::NAME);
            //extensions.push(ext::sampler_filter_minmax::NAME);
        }

        if self.device_api_version < vk::API_VERSION_1_3 {
            // Optional `VK_KHR_maintenance4`
            if self.supports_extension(khr::maintenance4::NAME) {
                extensions.push(khr::maintenance4::NAME);
            }

            // Optional `VK_EXT_image_robustness`
            if self.supports_extension(ext::image_robustness::NAME) {
                extensions.push(ext::image_robustness::NAME);
            }

            // Require `VK_EXT_subgroup_size_control` if the associated feature was requested
            if requested_features.contains(wgt::Features::SUBGROUP) {
                extensions.push(ext::subgroup_size_control::NAME);
            }

            // Optional `VK_KHR_shader_integer_dot_product`
            if self.supports_extension(khr::shader_integer_dot_product::NAME) {
                extensions.push(khr::shader_integer_dot_product::NAME);
            }

            // Optional `VK_KHR_dynamic_rendering`.
            // Depends on:
            // - `VK_KHR_get_physical_device_properties2` or Vulkan 1.1, and `VK_KHR_depth_stencil_resolve`
            // - or Vulkan 1.2
            //
            // We only check Vulkan 1.2 for now, as `VK_KHR_depth_stencil_resolve`
            // also depends a bunch of extensions.
            if self.device_api_version >= vk::API_VERSION_1_2
                && self.supports_extension(khr::dynamic_rendering::NAME)
            {
                extensions.push(khr::dynamic_rendering::NAME);
            }

            // Optional `VK_KHR_load_store_op_none`
            if self.supports_extension(khr::load_store_op_none::NAME) {
                extensions.push(khr::load_store_op_none::NAME);
            }

            // Optional `VK_QCOM_render_pass_store_ops`
            if self.supports_extension(ash::qcom::render_pass_store_ops::NAME) {
                extensions.push(ash::qcom::render_pass_store_ops::NAME);
            }

            // Optional `VK_EXT_load_store_op_none`
            if self.supports_extension(ext::load_store_op_none::NAME) {
                extensions.push(ext::load_store_op_none::NAME);
            }

            // Require `VK_KHR_shader_non_semantic_info` if the associated feature was requested
            if requested_features.contains(wgt::Features::DEBUG_PRINTF) {
                extensions.push(khr::shader_non_semantic_info::NAME);
            }
        }

        // Optional `VK_KHR_swapchain_mutable_format`
        if self.supports_extension(khr::swapchain_mutable_format::NAME) {
            extensions.push(khr::swapchain_mutable_format::NAME);
        }

        // Optional `VK_EXT_robustness2`
        if self.supports_extension(ext::robustness2::NAME) {
            extensions.push(ext::robustness2::NAME);
        }

        // Optional `VK_KHR_external_memory_win32`
        if self.supports_extension(khr::external_memory_win32::NAME) {
            extensions.push(khr::external_memory_win32::NAME);
        }

        // Optional `VK_KHR_external_memory_fd`
        if self.supports_extension(khr::external_memory_fd::NAME) {
            extensions.push(khr::external_memory_fd::NAME);
        }

        // Optional `VK_EXT_external_memory_dma`
        if self.supports_extension(ext::external_memory_dma_buf::NAME) {
            extensions.push(ext::external_memory_dma_buf::NAME);
        }

        // Optional `VK_EXT_image_drm_format_modifier`
        if self.supports_extension(ext::image_drm_format_modifier::NAME) {
            extensions.push(ext::image_drm_format_modifier::NAME);
        }

        // Optional `VK_EXT_memory_budget`
        if self.supports_extension(ext::memory_budget::NAME) {
            extensions.push(ext::memory_budget::NAME);
        } else {
            log::debug!("VK_EXT_memory_budget is not available.")
        }

        // Require `VK_KHR_draw_indirect_count` if the associated feature was requested
        // Even though Vulkan 1.2 has promoted the extension to core, we must require the extension to avoid
        // large amounts of spaghetti involved with using PhysicalDeviceVulkan12Features.
        if requested_features.contains(wgt::Features::MULTI_DRAW_INDIRECT_COUNT) {
            extensions.push(khr::draw_indirect_count::NAME);
        }

        // Require `VK_KHR_deferred_host_operations`, `VK_KHR_acceleration_structure` `VK_KHR_buffer_device_address` (for acceleration structures) if either `EXPERIMENTAL_RAY_QUERY` or `EXPERIMENTAL_RAY_TRACING_PIPELINES` were requested.
        if requested_features.intersects(
            wgt::Features::EXPERIMENTAL_RAY_QUERY
                | wgt::Features::EXPERIMENTAL_RAY_TRACING_PIPELINES,
        ) {
            extensions.push(khr::deferred_host_operations::NAME);
            extensions.push(khr::acceleration_structure::NAME);
            extensions.push(khr::buffer_device_address::NAME);
        }

        // Require `VK_KHR_ray_query` if `EXPERIMENTAL_RAY_QUERY` was requested
        if requested_features.contains(wgt::Features::EXPERIMENTAL_RAY_QUERY) {
            extensions.push(khr::ray_query::NAME);
        }

        // Require `VK_KHR_ray_tracing_pipeline` if `EXPERIMENTAL_RAY_TRACING_PIPELINES` was requested
        if requested_features.contains(wgt::Features::EXPERIMENTAL_RAY_TRACING_PIPELINES) {
            extensions.push(khr::ray_tracing_pipeline::NAME);
        }

        if requested_features.contains(wgt::Features::EXPERIMENTAL_RAY_HIT_VERTEX_RETURN) {
            extensions.push(khr::ray_tracing_position_fetch::NAME)
        }

        // Require `VK_EXT_conservative_rasterization` if the associated feature was requested
        if requested_features.contains(wgt::Features::CONSERVATIVE_RASTERIZATION) {
            extensions.push(ext::conservative_rasterization::NAME);
        }

        // Require `VK_KHR_portability_subset` on macOS/iOS
        #[cfg(target_vendor = "apple")]
        extensions.push(khr::portability_subset::NAME);

        // Require `VK_EXT_texture_compression_astc_hdr` if the associated feature was requested
        if requested_features.contains(wgt::Features::TEXTURE_COMPRESSION_ASTC_HDR) {
            extensions.push(ext::texture_compression_astc_hdr::NAME);
        }

        // Require `VK_KHR_shader_atomic_int64` if the associated feature was requested
        if requested_features.intersects(
            wgt::Features::SHADER_INT64_ATOMIC_ALL_OPS | wgt::Features::SHADER_INT64_ATOMIC_MIN_MAX,
        ) {
            extensions.push(khr::shader_atomic_int64::NAME);
        }

        // Require `VK_EXT_shader_image_atomic_int64` if the associated feature was requested
        if requested_features.intersects(wgt::Features::TEXTURE_INT64_ATOMIC) {
            extensions.push(ext::shader_image_atomic_int64::NAME);
        }

        // Require `VK_EXT_shader_atomic_float` if the associated feature was requested
        if requested_features.contains(wgt::Features::SHADER_FLOAT32_ATOMIC) {
            extensions.push(ext::shader_atomic_float::NAME);
        }

        // Require VK_GOOGLE_display_timing if the associated feature was requested
        if requested_features.contains(wgt::Features::VULKAN_GOOGLE_DISPLAY_TIMING) {
            extensions.push(google::display_timing::NAME);
        }

        if requested_features.contains(wgt::Features::EXPERIMENTAL_MESH_SHADER) {
            extensions.push(ext::mesh_shader::NAME);
        }

        // Require `VK_KHR_fragment_shader_barycentric` if an associated feature was requested
        // Vulkan bundles both barycentrics and per-vertex attributes under the same feature.
        if requested_features
            .intersects(wgt::Features::SHADER_BARYCENTRICS | wgt::Features::SHADER_PER_VERTEX)
        {
            extensions.push(khr::fragment_shader_barycentric::NAME);
        }

        // Require `VK_KHR_cooperative_matrix` if the associated feature was requested
        if requested_features.contains(wgt::Features::EXPERIMENTAL_COOPERATIVE_MATRIX) {
            extensions.push(khr::cooperative_matrix::NAME);
        }

        extensions
    }

    fn to_wgpu_limits(&self) -> wgt::Limits {
        let limits = &self.properties.limits;

        // Default is only implemented for tuples up to a certain size.
        let (
            mut max_task_workgroup_total_count,
            mut max_task_workgroups_per_dimension,
            mut max_mesh_workgroup_total_count,
            mut max_mesh_workgroups_per_dimension,
        ) = Default::default();
        let (
            mut max_task_invocations_per_workgroup,
            mut max_task_invocations_per_dimension,
            mut max_mesh_invocations_per_workgroup,
            mut max_mesh_invocations_per_dimension,
            mut max_task_payload_size,
            mut max_mesh_output_vertices,
            mut max_mesh_output_primitives,
            mut max_mesh_output_layers,
            mut max_mesh_multiview_view_count,
        ) = Default::default();
        if let Some(m) = self.mesh_shader {
            max_task_workgroup_total_count = m.max_task_work_group_total_count;
            max_task_workgroups_per_dimension =
                m.max_task_work_group_count.into_iter().min().unwrap();
            max_mesh_workgroup_total_count = m.max_mesh_work_group_total_count;
            max_mesh_workgroups_per_dimension =
                m.max_mesh_work_group_count.into_iter().min().unwrap();
            max_task_invocations_per_workgroup = m.max_task_work_group_invocations;
            max_task_invocations_per_dimension =
                m.max_task_work_group_size.into_iter().min().unwrap();
            max_mesh_invocations_per_workgroup = m.max_mesh_work_group_invocations;
            max_mesh_invocations_per_dimension =
                m.max_mesh_work_group_size.into_iter().min().unwrap();
            max_task_payload_size = m.max_task_payload_size;
            max_mesh_output_vertices = m.max_mesh_output_vertices;
            max_mesh_output_primitives = m.max_mesh_output_primitives;
            max_mesh_output_layers = m.max_mesh_output_layers;
            max_mesh_multiview_view_count = m.max_mesh_multiview_view_count;
        }

        let max_memory_allocation_size = self
            .maintenance_3
            .map(|maintenance_3| maintenance_3.max_memory_allocation_size)
            .unwrap_or(u64::MAX);
        let max_buffer_size = self
            .maintenance_4
            .map(|maintenance_4| maintenance_4.max_buffer_size)
            .unwrap_or(u64::MAX);
        let max_buffer_size = max_buffer_size.min(max_memory_allocation_size);

        // Prevent very large buffers on mesa and most android devices, and in all cases
        // don't risk confusing JS by exceeding the range of a double.
        let is_nvidia = self.properties.vendor_id == crate::auxil::db::nvidia::VENDOR;
        let max_buffer_size_cap =
            if (cfg!(target_os = "linux") || cfg!(target_os = "android")) && !is_nvidia {
                i32::MAX as u64
            } else {
                1u64 << 52
            };

        let max_buffer_size = max_buffer_size.min(max_buffer_size_cap);

        let mut max_binding_array_elements = 0;
        let mut max_sampler_binding_array_elements = 0;
        if let Some(ref descriptor_indexing) = self.descriptor_indexing {
            max_binding_array_elements = descriptor_indexing
                .max_descriptor_set_update_after_bind_sampled_images
                .min(descriptor_indexing.max_descriptor_set_update_after_bind_storage_images)
                .min(descriptor_indexing.max_descriptor_set_update_after_bind_storage_buffers)
                .min(descriptor_indexing.max_per_stage_descriptor_update_after_bind_sampled_images)
                .min(descriptor_indexing.max_per_stage_descriptor_update_after_bind_storage_images)
                .min(
                    descriptor_indexing.max_per_stage_descriptor_update_after_bind_storage_buffers,
                );

            max_sampler_binding_array_elements = descriptor_indexing
                .max_descriptor_set_update_after_bind_samplers
                .min(descriptor_indexing.max_per_stage_descriptor_update_after_bind_samplers);
        }

        const MAX_SHADER_STAGES_PER_PIPELINE: u32 = 2;

        // When summed, the 3 limits below must be under Vulkan's maxFragmentCombinedOutputResources.
        // https://gpuweb.github.io/gpuweb/correspondence/#vulkan-maxFragmentCombinedOutputResources
        //
        // - maxStorageTexturesPerShaderStage, WebGPU default: 4
        // - maxStorageBuffersPerShaderStage, WebGPU default: 8
        // - maxColorAttachments, WebGPU default: 8
        //
        // However, maxFragmentCombinedOutputResources should be ignored on
        // intel/nvidia/amd/imgtec since it's not reported correctly.
        //
        // https://github.com/gpuweb/gpuweb/issues/3631#issuecomment-1498747606
        // https://github.com/gpuweb/gpuweb/issues/4018
        let mut max_storage_textures_per_shader_stage = limits
            .max_per_stage_descriptor_storage_images
            .min(limits.max_descriptor_set_storage_images / MAX_SHADER_STAGES_PER_PIPELINE);
        let mut max_storage_buffers_per_shader_stage = limits
            .max_per_stage_descriptor_storage_buffers
            .min(limits.max_descriptor_set_storage_buffers / MAX_SHADER_STAGES_PER_PIPELINE);
        let mut max_color_attachments = limits
            .max_color_attachments
            .min(limits.max_fragment_output_attachments);

        let ignore_max_fragment_combined_output_resources_by_device = [
            crate::auxil::db::intel::VENDOR,
            crate::auxil::db::nvidia::VENDOR,
            crate::auxil::db::amd::VENDOR,
            crate::auxil::db::imgtec::VENDOR,
        ]
        .contains(&self.properties.vendor_id);
        let ignore_max_fragment_combined_output_resources_by_driver =
            self.is_driver(vk::DriverId::MESA_AGXV);
        let ignore_max_fragment_combined_output_resources =
            ignore_max_fragment_combined_output_resources_by_device
                || ignore_max_fragment_combined_output_resources_by_driver;

        if !ignore_max_fragment_combined_output_resources {
            crate::auxil::cap_limits_to_be_under_the_sum_limit(
                [
                    &mut max_storage_textures_per_shader_stage,
                    &mut max_storage_buffers_per_shader_stage,
                    &mut max_color_attachments,
                ],
                limits.max_fragment_combined_output_resources,
            );
        }

        // When summed, the 5 limits below must be under Vulkan's maxPerStageResources.
        //
        // - maxUniformBuffersPerShaderStage, WebGPU default: 12
        // - maxSampledTexturesPerShaderStage, WebGPU default: 16
        // - maxStorageTexturesPerShaderStage, WebGPU default: 4
        // - maxStorageBuffersPerShaderStage, WebGPU default: 8
        // - maxColorAttachments, WebGPU default: 8
        //
        // Note: Vulkan's texel buffers and input attachments also count towards
        // maxPerStageResources but we don't make use of them.
        let mut max_sampled_textures_per_shader_stage = limits
            .max_per_stage_descriptor_sampled_images
            .min(limits.max_descriptor_set_sampled_images / MAX_SHADER_STAGES_PER_PIPELINE);
        let mut max_uniform_buffers_per_shader_stage = limits
            .max_per_stage_descriptor_uniform_buffers
            .min(limits.max_descriptor_set_uniform_buffers / MAX_SHADER_STAGES_PER_PIPELINE);

        crate::auxil::cap_limits_to_be_under_the_sum_limit(
            [
                &mut max_sampled_textures_per_shader_stage,
                &mut max_uniform_buffers_per_shader_stage,
                &mut max_storage_textures_per_shader_stage,
                &mut max_storage_buffers_per_shader_stage,
                &mut max_color_attachments,
            ],
            limits.max_per_stage_resources,
        );

        // Acceleration structure limits
        let mut max_blas_geometry_count = 0;
        let mut max_blas_primitive_count = 0;
        let mut max_tlas_instance_count = 0;
        let mut max_acceleration_structures_per_shader_stage = 0;
        if let Some(properties) = self.acceleration_structure {
            max_blas_geometry_count = properties.max_geometry_count as u32;
            max_blas_primitive_count = properties.max_primitive_count as u32;
            max_tlas_instance_count = properties.max_instance_count as u32;
            max_acceleration_structures_per_shader_stage = properties
                .max_per_stage_descriptor_acceleration_structures
                .min(
                    properties.max_descriptor_set_acceleration_structures
                        / MAX_SHADER_STAGES_PER_PIPELINE,
                );
        }

        // When summed, the 6 limits below must be under Vulkan's
        // maxPerSetDescriptors / MAX_SHADER_STAGES_PER_PIPELINE.
        //
        // - maxUniformBuffersPerShaderStage, WebGPU default: 12
        // - maxSampledTexturesPerShaderStage, WebGPU default: 16
        // - maxStorageTexturesPerShaderStage, WebGPU default: 4
        // - maxStorageBuffersPerShaderStage, WebGPU default: 8
        // - maxSamplersPerShaderStage, WebGPU default: 16
        // - maxAccelerationStructuresPerShaderStage, Native only
        //
        // Note: All Vulkan's descriptor types count towards maxPerSetDescriptors but
        // we don't use all of them.
        // See https://registry.khronos.org/vulkan/specs/latest/html/vkspec.html#interfaces-resources-limits
        let max_per_set_descriptors = self
            .maintenance_3
            .map(|maintenance_3| maintenance_3.max_per_set_descriptors)
            // The lowest value seen in reports is 312, use 256 as a safe default.
            // https://vulkan.gpuinfo.org/displayextensionproperty.php?extensionname=VK_KHR_maintenance3&extensionproperty=maxPerSetDescriptors&platform=all
            // https://vulkan.gpuinfo.org/displaycoreproperty.php?core=1.1&name=maxPerSetDescriptors&platform=all
            .unwrap_or(256);

        let mut max_samplers_per_shader_stage = limits
            .max_per_stage_descriptor_samplers
            .min(limits.max_descriptor_set_samplers / MAX_SHADER_STAGES_PER_PIPELINE);

        crate::auxil::cap_limits_to_be_under_the_sum_limit(
            [
                &mut max_sampled_textures_per_shader_stage,
                &mut max_uniform_buffers_per_shader_stage,
                &mut max_storage_textures_per_shader_stage,
                &mut max_storage_buffers_per_shader_stage,
                &mut max_samplers_per_shader_stage,
                &mut max_acceleration_structures_per_shader_stage,
            ],
            max_per_set_descriptors / MAX_SHADER_STAGES_PER_PIPELINE,
        );

        // Use max(default, maxPerSetDescriptors) since the spec requires this
        // limit to be at least 1000. This is ok because we already lowered
        // all the other relevant per stage limits so their sum is lower
        // than maxPerSetDescriptors.
        let max_bindings_per_bind_group = 1000.max(max_per_set_descriptors);

        // TODO: programmatically determine this, if possible. It's unclear whether we can
        // as of https://github.com/gpuweb/gpuweb/issues/2965#issuecomment-1361315447.
        //
        // In theory some tilers may not support this much. We can't tell however, and
        // the driver will throw a DEVICE_REMOVED if it goes too high in usage. This is fine.
        let max_color_attachment_bytes_per_sample =
            max_color_attachments * wgt::TextureFormat::MAX_TARGET_PIXEL_BYTE_COST;

        let mut max_ray_dispatch_count = 0;
        let mut max_ray_recursion_depth = 0;

        if let Some(properties) = self.ray_tracing_pipeline {
            max_ray_dispatch_count = properties.max_ray_dispatch_invocation_count;
            max_ray_recursion_depth = properties.max_ray_recursion_depth;
        }

        let max_multiview_view_count = self
            .multiview
            .map(|a| a.max_multiview_view_count.min(32))
            .unwrap_or(0);

        crate::auxil::adjust_raw_limits(wgt::Limits {
            //
            // WebGPU LIMITS:
            // Based on https://gpuweb.github.io/gpuweb/correspondence/#limits
            //
            max_texture_dimension_1d: limits.max_image_dimension1_d,
            max_texture_dimension_2d: limits
                .max_image_dimension2_d
                .min(limits.max_image_dimension_cube)
                .min(limits.max_framebuffer_width)
                .min(limits.max_framebuffer_height),
            max_texture_dimension_3d: limits.max_image_dimension3_d,
            max_texture_array_layers: limits.max_image_array_layers,
            max_bind_groups: limits.max_bound_descriptor_sets,
            // No limit.
            max_bind_groups_plus_vertex_buffers: u32::MAX,
            max_bindings_per_bind_group,
            max_dynamic_uniform_buffers_per_pipeline_layout: limits
                .max_descriptor_set_uniform_buffers_dynamic,
            max_dynamic_storage_buffers_per_pipeline_layout: limits
                .max_descriptor_set_storage_buffers_dynamic,
            max_samplers_per_shader_stage,
            max_sampled_textures_per_shader_stage,
            max_storage_buffers_per_shader_stage,
            max_storage_buffers_in_vertex_stage: 0,
            max_storage_buffers_in_fragment_stage: 0,
            max_storage_textures_per_shader_stage,
            max_storage_textures_in_vertex_stage: 0,
            max_storage_textures_in_fragment_stage: 0,
            max_uniform_buffers_per_shader_stage,
            max_vertex_buffers: limits.max_vertex_input_bindings,
            max_buffer_size,
            max_uniform_buffer_binding_size: limits
                .max_uniform_buffer_range
                .min(crate::auxil::MAX_I32_BINDING_SIZE)
                .into(),
            max_storage_buffer_binding_size: limits
                .max_storage_buffer_range
                .min(crate::auxil::MAX_I32_BINDING_SIZE)
                .into(),
            min_uniform_buffer_offset_alignment: limits.min_uniform_buffer_offset_alignment as u32,
            min_storage_buffer_offset_alignment: limits.min_storage_buffer_offset_alignment as u32,
            max_vertex_attributes: limits.max_vertex_input_attributes,
            max_vertex_buffer_array_stride: limits.max_vertex_input_binding_stride,
            max_inter_stage_shader_variables: limits
                .max_vertex_output_components
                .min(limits.max_fragment_input_components)
                / 4
                - 1, // -1 for position
            max_color_attachments,
            max_color_attachment_bytes_per_sample,
            max_compute_workgroup_storage_size: limits.max_compute_shared_memory_size,
            max_compute_invocations_per_workgroup: limits.max_compute_work_group_invocations,
            max_compute_workgroup_size_x: limits.max_compute_work_group_size[0],
            max_compute_workgroup_size_y: limits.max_compute_work_group_size[1],
            max_compute_workgroup_size_z: limits.max_compute_work_group_size[2],
            max_compute_workgroups_per_dimension: limits.max_compute_work_group_count[0]
                .min(limits.max_compute_work_group_count[1])
                .min(limits.max_compute_work_group_count[2]),
            max_immediate_size: limits.max_push_constants_size,
            //
            // NATIVE (Non-WebGPU) LIMITS:
            //
            max_non_sampler_bindings: u32::MAX,

            max_binding_array_elements_per_shader_stage: max_binding_array_elements,
            max_binding_array_sampler_elements_per_shader_stage: max_sampler_binding_array_elements,
            max_binding_array_acceleration_structure_elements_per_shader_stage: if self
                .descriptor_indexing
                .is_some()
            {
                max_acceleration_structures_per_shader_stage
            } else {
                0
            },

            max_task_workgroup_total_count,
            max_task_workgroups_per_dimension,
            max_mesh_workgroup_total_count,
            max_mesh_workgroups_per_dimension,

            max_task_invocations_per_workgroup,
            max_task_invocations_per_dimension,

            max_mesh_invocations_per_workgroup,
            max_mesh_invocations_per_dimension,

            max_task_payload_size,
            max_mesh_output_vertices,
            max_mesh_output_primitives,
            max_mesh_output_layers,
            max_mesh_multiview_view_count,

            max_blas_primitive_count,
            max_blas_geometry_count,
            max_tlas_instance_count,
            max_acceleration_structures_per_shader_stage,
            max_buffers_and_acceleration_structures_per_shader_stage: u32::MAX,

            max_multiview_view_count,

            max_ray_dispatch_count,
            max_ray_recursion_depth,
        })
    }

    /// Return a `wgpu_hal::Alignments` structure describing this adapter.
    ///
    /// The `using_robustness2` argument says how this adapter will implement
    /// `wgpu_hal`'s guarantee that shaders can only read the [accessible
    /// region][ar] of bindgroup's buffer bindings:
    ///
    /// - If this adapter will depend on `VK_EXT_robustness2`'s
    ///   `robustBufferAccess2` feature to apply bounds checks to shader buffer
    ///   access, `using_robustness2` must be `true`.
    ///
    /// - Otherwise, this adapter must use Naga to inject bounds checks on
    ///   buffer accesses, and `using_robustness2` must be `false`.
    ///
    /// [ar]: ../../struct.BufferBinding.html#accessible-region
    fn to_hal_alignments(&self, using_robustness2: bool) -> crate::Alignments {
        let limits = &self.properties.limits;
        crate::Alignments {
            buffer_copy_offset: wgt::BufferSize::new(limits.optimal_buffer_copy_offset_alignment)
                .unwrap(),
            buffer_copy_pitch: wgt::BufferSize::new(limits.optimal_buffer_copy_row_pitch_alignment)
                .unwrap(),
            uniform_bounds_check_alignment: {
                let alignment = if using_robustness2 {
                    self.robustness2
                        .unwrap() // if we're using it, we should have its properties
                        .robust_uniform_buffer_access_size_alignment
                } else {
                    // If the `robustness2` properties are unavailable, then `robustness2` is not available either Naga-injected bounds checks are precise.
                    1
                };
                wgt::BufferSize::new(alignment).unwrap()
            },
            raw_tlas_instance_size: 64,
            ray_tracing_scratch_buffer_alignment: self.acceleration_structure.map_or(
                0,
                |acceleration_structure| {
                    acceleration_structure.min_acceleration_structure_scratch_offset_alignment
                },
            ),
            ray_tracing_pipeline_group_data_size: self
                .ray_tracing_pipeline
                .map_or(0, |ray_tracing_pipeline| {
                    ray_tracing_pipeline.shader_group_handle_size
                }),
            ray_tracing_pipeline_group_data_alignment: self
                .ray_tracing_pipeline
                .map_or(0, |ray_tracing_pipeline| {
                    ray_tracing_pipeline.shader_group_handle_alignment
                }),
            ray_tracing_pipeline_data_offset_alignment: self
                .ray_tracing_pipeline
                .map_or(0, |ray_tracing_pipeline| {
                    ray_tracing_pipeline.shader_group_base_alignment
                }),
        }
    }
}

impl super::InstanceShared {
    fn inspect(
        &self,
        phd: vk::PhysicalDevice,
    ) -> (PhysicalDeviceProperties, PhysicalDeviceFeatures) {
        let capabilities = {
            let mut capabilities = PhysicalDeviceProperties::default();
            capabilities.supported_extensions =
                unsafe { self.raw.enumerate_device_extension_properties(phd).unwrap() };
            capabilities.properties = unsafe { self.raw.get_physical_device_properties(phd) };
            capabilities.device_api_version = capabilities.properties.api_version;

            let supports_multiview = capabilities.device_api_version >= vk::API_VERSION_1_1
                || capabilities.supports_extension(khr::multiview::NAME);

            if let Some(ref get_device_properties) = self.get_physical_device_properties {
                // Get these now to avoid borrowing conflicts later
                let supports_maintenance3 = capabilities.device_api_version >= vk::API_VERSION_1_1
                    || capabilities.supports_extension(khr::maintenance3::NAME);
                let supports_maintenance4 = capabilities.device_api_version >= vk::API_VERSION_1_3
                    || capabilities.supports_extension(khr::maintenance4::NAME);
                let supports_maintenance5 = capabilities.device_api_version
                    >= vk::make_api_version(0, 1, 4, 0) // TODO: Use `vk::API_VERSION_1_4` after `ash` is updated.
                    || capabilities.supports_extension(khr::maintenance5::NAME);
                let supports_descriptor_indexing = capabilities.device_api_version
                    >= vk::API_VERSION_1_2
                    || capabilities.supports_extension(ext::descriptor_indexing::NAME);
                let supports_driver_properties = capabilities.device_api_version
                    >= vk::API_VERSION_1_2
                    || capabilities.supports_extension(khr::driver_properties::NAME);
                let supports_subgroup_size_control = capabilities.device_api_version
                    >= vk::API_VERSION_1_3
                    || capabilities.supports_extension(ext::subgroup_size_control::NAME);
                let supports_robustness2 = capabilities.supports_extension(ext::robustness2::NAME);
                let supports_pci_bus_info =
                    capabilities.supports_extension(ext::pci_bus_info::NAME);

                let supports_acceleration_structure =
                    capabilities.supports_extension(khr::acceleration_structure::NAME);

                let supports_ray_tracing_pipeline =
                    capabilities.supports_extension(khr::ray_tracing_pipeline::NAME);

                let supports_mesh_shader = capabilities.supports_extension(ext::mesh_shader::NAME);

                let mut properties2 = vk::PhysicalDeviceProperties2KHR::default();
                if supports_maintenance3 {
                    let next = capabilities
                        .maintenance_3
                        .insert(vk::PhysicalDeviceMaintenance3Properties::default());
                    properties2 = properties2.push_next(next);
                }

                if supports_maintenance4 {
                    let next = capabilities
                        .maintenance_4
                        .insert(vk::PhysicalDeviceMaintenance4Properties::default());
                    properties2 = properties2.push_next(next);
                }

                if supports_maintenance5 {
                    let next = capabilities
                        .maintenance_5
                        .insert(vk::PhysicalDeviceMaintenance5PropertiesKHR::default());
                    properties2 = properties2.push_next(next);
                }

                if supports_descriptor_indexing {
                    let next = capabilities
                        .descriptor_indexing
                        .insert(vk::PhysicalDeviceDescriptorIndexingPropertiesEXT::default());
                    properties2 = properties2.push_next(next);
                }

                if supports_acceleration_structure {
                    let next = capabilities
                        .acceleration_structure
                        .insert(vk::PhysicalDeviceAccelerationStructurePropertiesKHR::default());
                    properties2 = properties2.push_next(next);
                }

                if supports_ray_tracing_pipeline {
                    let next = capabilities
                        .ray_tracing_pipeline
                        .insert(vk::PhysicalDeviceRayTracingPipelinePropertiesKHR::default());
                    properties2 = properties2.push_next(next);
                }

                if supports_driver_properties {
                    let next = capabilities
                        .driver
                        .insert(vk::PhysicalDeviceDriverPropertiesKHR::default());
                    properties2 = properties2.push_next(next);
                }

                if capabilities.device_api_version >= vk::API_VERSION_1_1 {
                    let next = capabilities
                        .subgroup
                        .insert(vk::PhysicalDeviceSubgroupProperties::default());
                    properties2 = properties2.push_next(next);
                }

                if supports_subgroup_size_control {
                    let next = capabilities
                        .subgroup_size_control
                        .insert(vk::PhysicalDeviceSubgroupSizeControlProperties::default());
                    properties2 = properties2.push_next(next);
                }

                if supports_robustness2 {
                    let next = capabilities
                        .robustness2
                        .insert(vk::PhysicalDeviceRobustness2PropertiesEXT::default());
                    properties2 = properties2.push_next(next);
                }

                if supports_pci_bus_info {
                    let next = capabilities
                        .pci_bus_info
                        .insert(vk::PhysicalDevicePCIBusInfoPropertiesEXT::default());
                    properties2 = properties2.push_next(next);
                }

                if supports_mesh_shader {
                    let next = capabilities
                        .mesh_shader
                        .insert(vk::PhysicalDeviceMeshShaderPropertiesEXT::default());
                    properties2 = properties2.push_next(next);
                }

                if supports_multiview {
                    let next = capabilities
                        .multiview
                        .insert(vk::PhysicalDeviceMultiviewProperties::default());
                    properties2 = properties2.push_next(next);
                }

                unsafe {
                    get_device_properties.get_physical_device_properties2(phd, &mut properties2)
                };

                // Query cooperative matrix properties
                if capabilities.supports_extension(khr::cooperative_matrix::NAME) {
                    let coop_matrix =
                        khr::cooperative_matrix::Instance::new(&self.entry, &self.raw);
                    capabilities.cooperative_matrix_properties =
                        query_cooperative_matrix_properties(&coop_matrix, phd);
                }

                // Suppress some capabilities to avoid known problems
                if is_intel_igpu_outdated_for_robustness2(&capabilities) {
                    capabilities
                        .supported_extensions
                        .retain(|&x| x.extension_name_as_c_str() != Ok(ext::robustness2::NAME));
                    capabilities.robustness2 = None;
                }

                // Due to https://gitlab.freedesktop.org/mesa/mesa/-/work_items/15725
                // TODO(https://github.com/gfx-rs/wgpu/issues/9742): enable on
                // fixed driver versions, when available
                if capabilities.is_driver(vk::DriverId::MESA_RADV) {
                    capabilities
                        .supported_extensions
                        .retain(|&x| x.extension_name_as_c_str() != Ok(ext::memory_budget::NAME));
                }
            };
            capabilities
        };

        let mut features = PhysicalDeviceFeatures::default();
        features.core = if let Some(ref get_device_properties) = self.get_physical_device_properties
        {
            let core = vk::PhysicalDeviceFeatures::default();
            let mut features2 = vk::PhysicalDeviceFeatures2KHR::default().features(core);

            // `VK_KHR_multiview` is promoted to 1.1
            if capabilities.device_api_version >= vk::API_VERSION_1_1
                || capabilities.supports_extension(khr::multiview::NAME)
            {
                let next = features
                    .multiview
                    .insert(vk::PhysicalDeviceMultiviewFeatures::default());
                features2 = features2.push_next(next);
            }

            // `VK_KHR_sampler_ycbcr_conversion` is promoted to 1.1
            if capabilities.device_api_version >= vk::API_VERSION_1_1
                || capabilities.supports_extension(khr::sampler_ycbcr_conversion::NAME)
            {
                let next = features
                    .sampler_ycbcr_conversion
                    .insert(vk::PhysicalDeviceSamplerYcbcrConversionFeatures::default());
                features2 = features2.push_next(next);
            }

            if capabilities.supports_extension(ext::descriptor_indexing::NAME) {
                let next = features
                    .descriptor_indexing
                    .insert(vk::PhysicalDeviceDescriptorIndexingFeaturesEXT::default());
                features2 = features2.push_next(next);
            }

            // `VK_KHR_timeline_semaphore` is promoted to 1.2, but has no
            // changes, so we can keep using the extension unconditionally.
            if capabilities.supports_extension(khr::timeline_semaphore::NAME) {
                let next = features
                    .timeline_semaphore
                    .insert(vk::PhysicalDeviceTimelineSemaphoreFeaturesKHR::default());
                features2 = features2.push_next(next);
            }

            // `VK_KHR_shader_atomic_int64` is promoted to 1.2, but has no
            // changes, so we can keep using the extension unconditionally.
            if capabilities.device_api_version >= vk::API_VERSION_1_2
                || capabilities.supports_extension(khr::shader_atomic_int64::NAME)
            {
                let next = features
                    .shader_atomic_int64
                    .insert(vk::PhysicalDeviceShaderAtomicInt64Features::default());
                features2 = features2.push_next(next);
            }

            if capabilities.supports_extension(ext::shader_image_atomic_int64::NAME) {
                let next = features
                    .shader_image_atomic_int64
                    .insert(vk::PhysicalDeviceShaderImageAtomicInt64FeaturesEXT::default());
                features2 = features2.push_next(next);
            }
            if capabilities.supports_extension(ext::shader_atomic_float::NAME) {
                let next = features
                    .shader_atomic_float
                    .insert(vk::PhysicalDeviceShaderAtomicFloatFeaturesEXT::default());
                features2 = features2.push_next(next);
            }
            if capabilities.supports_extension(ext::image_robustness::NAME) {
                let next = features
                    .image_robustness
                    .insert(vk::PhysicalDeviceImageRobustnessFeaturesEXT::default());
                features2 = features2.push_next(next);
            }
            if capabilities.supports_extension(ext::robustness2::NAME) {
                let next = features
                    .robustness2
                    .insert(vk::PhysicalDeviceRobustness2FeaturesEXT::default());
                features2 = features2.push_next(next);
            }
            if capabilities.supports_extension(ext::texture_compression_astc_hdr::NAME) {
                let next = features
                    .astc_hdr
                    .insert(vk::PhysicalDeviceTextureCompressionASTCHDRFeaturesEXT::default());
                features2 = features2.push_next(next);
            }

            // `VK_KHR_shader_float16_int8` is promoted to 1.2
            if capabilities.device_api_version >= vk::API_VERSION_1_2
                || capabilities.supports_extension(khr::shader_float16_int8::NAME)
            {
                let next = features
                    .shader_float16_int8
                    .insert(vk::PhysicalDeviceShaderFloat16Int8FeaturesKHR::default());
                features2 = features2.push_next(next);
            }

            if capabilities.supports_extension(khr::_16bit_storage::NAME) {
                let next = features
                    ._16bit_storage
                    .insert(vk::PhysicalDevice16BitStorageFeaturesKHR::default());
                features2 = features2.push_next(next);
            }
            if capabilities.supports_extension(khr::acceleration_structure::NAME) {
                let next = features
                    .acceleration_structure
                    .insert(vk::PhysicalDeviceAccelerationStructureFeaturesKHR::default());
                features2 = features2.push_next(next);
            }

            if capabilities.supports_extension(khr::ray_tracing_position_fetch::NAME) {
                let next = features
                    .position_fetch
                    .insert(vk::PhysicalDeviceRayTracingPositionFetchFeaturesKHR::default());
                features2 = features2.push_next(next);
            }

            // `VK_KHR_maintenance4` is promoted to 1.3
            if capabilities.device_api_version >= vk::API_VERSION_1_3
                || capabilities.supports_extension(khr::maintenance4::NAME)
            {
                let next = features
                    .maintenance4
                    .insert(vk::PhysicalDeviceMaintenance4Features::default());
                features2 = features2.push_next(next);
            }

            // `VK_KHR_zero_initialize_workgroup_memory` is promoted to 1.3
            if capabilities.device_api_version >= vk::API_VERSION_1_3
                || capabilities.supports_extension(khr::zero_initialize_workgroup_memory::NAME)
            {
                let next = features
                    .zero_initialize_workgroup_memory
                    .insert(vk::PhysicalDeviceZeroInitializeWorkgroupMemoryFeatures::default());
                features2 = features2.push_next(next);
            }

            // `VK_EXT_subgroup_size_control` is promoted to 1.3
            if capabilities.device_api_version >= vk::API_VERSION_1_3
                || capabilities.supports_extension(ext::subgroup_size_control::NAME)
            {
                let next = features
                    .subgroup_size_control
                    .insert(vk::PhysicalDeviceSubgroupSizeControlFeatures::default());
                features2 = features2.push_next(next);
            }

            if capabilities.supports_extension(ext::mesh_shader::NAME) {
                let next = features
                    .mesh_shader
                    .insert(vk::PhysicalDeviceMeshShaderFeaturesEXT::default());
                features2 = features2.push_next(next);
            }

            // `VK_KHR_shader_integer_dot_product` is promoted to 1.3
            if capabilities.device_api_version >= vk::API_VERSION_1_3
                || capabilities.supports_extension(khr::shader_integer_dot_product::NAME)
            {
                let next = features
                    .shader_integer_dot_product
                    .insert(vk::PhysicalDeviceShaderIntegerDotProductFeatures::default());
                features2 = features2.push_next(next);
            }

            if capabilities.supports_extension(khr::fragment_shader_barycentric::NAME) {
                let next = features
                    .shader_barycentrics
                    .insert(vk::PhysicalDeviceFragmentShaderBarycentricFeaturesKHR::default());
                features2 = features2.push_next(next);
            }

            if capabilities.supports_extension(khr::portability_subset::NAME) {
                let next = features
                    .portability_subset
                    .insert(vk::PhysicalDevicePortabilitySubsetFeaturesKHR::default());
                features2 = features2.push_next(next);
            }

            if capabilities.supports_extension(khr::cooperative_matrix::NAME) {
                let next = features
                    .cooperative_matrix
                    .insert(vk::PhysicalDeviceCooperativeMatrixFeaturesKHR::default());
                features2 = features2.push_next(next);
            }

            if capabilities.device_api_version >= vk::API_VERSION_1_2
                || capabilities.supports_extension(khr::vulkan_memory_model::NAME)
            {
                let next = features
                    .vulkan_memory_model
                    .insert(vk::PhysicalDeviceVulkanMemoryModelFeaturesKHR::default());
                features2 = features2.push_next(next);
            }

            if capabilities.device_api_version >= vk::API_VERSION_1_1 {
                let next = features
                    .shader_draw_parameters
                    .insert(vk::PhysicalDeviceShaderDrawParametersFeatures::default());
                features2 = features2.push_next(next);
            }

            unsafe { get_device_properties.get_physical_device_features2(phd, &mut features2) };
            features2.features
        } else {
            unsafe { self.raw.get_physical_device_features(phd) }
        };

        (capabilities, features)
    }
}

impl super::Instance {
    pub fn expose_adapter(
        &self,
        phd: vk::PhysicalDevice,
    ) -> Option<crate::ExposedAdapter<super::Api>> {
        use crate::auxil::db;

        let (phd_capabilities, phd_features) = self.shared.inspect(phd);

        let mem_properties = {
            profiling::scope!("vkGetPhysicalDeviceMemoryProperties");
            unsafe { self.shared.raw.get_physical_device_memory_properties(phd) }
        };
        let memory_types = &mem_properties.memory_types_as_slice();
        let supports_lazily_allocated = memory_types.iter().any(|mem| {
            mem.property_flags
                .contains(vk::MemoryPropertyFlags::LAZILY_ALLOCATED)
        });

        let device_type = match phd_capabilities.properties.device_type {
            vk::PhysicalDeviceType::OTHER => wgt::DeviceType::Other,
            vk::PhysicalDeviceType::INTEGRATED_GPU => wgt::DeviceType::IntegratedGpu,
            vk::PhysicalDeviceType::DISCRETE_GPU => wgt::DeviceType::DiscreteGpu,
            vk::PhysicalDeviceType::VIRTUAL_GPU => wgt::DeviceType::VirtualGpu,
            vk::PhysicalDeviceType::CPU => wgt::DeviceType::Cpu,
            _ => wgt::DeviceType::Other,
        };
        let info = wgt::AdapterInfo {
            name: {
                phd_capabilities
                    .properties
                    .device_name_as_c_str()
                    .ok()
                    .and_then(|name| name.to_str().ok())
                    .unwrap_or("?")
                    .to_owned()
            },
            vendor: phd_capabilities.properties.vendor_id,
            device: phd_capabilities.properties.device_id,
            device_pci_bus_id: phd_capabilities
                .pci_bus_info
                .filter(|info| info.pci_bus != 0 || info.pci_device != 0)
                .map(|info| {
                    format!(
                        "{:04x}:{:02x}:{:02x}.{}",
                        info.pci_domain, info.pci_bus, info.pci_device, info.pci_function
                    )
                })
                .unwrap_or_default(),
            driver: {
                phd_capabilities
                    .driver
                    .as_ref()
                    .and_then(|driver| driver.driver_name_as_c_str().ok())
                    .and_then(|name| name.to_str().ok())
                    .unwrap_or("?")
                    .to_owned()
            },
            driver_info: {
                phd_capabilities
                    .driver
                    .as_ref()
                    .and_then(|driver| driver.driver_info_as_c_str().ok())
                    .and_then(|name| name.to_str().ok())
                    .unwrap_or("?")
                    .to_owned()
            },
            subgroup_min_size: phd_capabilities
                .subgroup_size_control
                .map(|subgroup_size| subgroup_size.min_subgroup_size)
                .unwrap_or(wgt::MINIMUM_SUBGROUP_MIN_SIZE),
            subgroup_max_size: phd_capabilities
                .subgroup_size_control
                .map(|subgroup_size| subgroup_size.max_subgroup_size)
                .unwrap_or(wgt::MAXIMUM_SUBGROUP_MAX_SIZE),
            transient_saves_memory: Some(supports_lazily_allocated),
            ..wgt::AdapterInfo::new(device_type, wgt::Backend::Vulkan)
        };
        let mut workarounds = super::Workarounds::empty();
        {
            // TODO: only enable for particular devices
            workarounds |= super::Workarounds::SEPARATE_ENTRY_POINTS;
            workarounds.set(
                super::Workarounds::EMPTY_RESOLVE_ATTACHMENT_LISTS,
                phd_capabilities.properties.vendor_id == db::qualcomm::VENDOR,
            );
            workarounds.set(
                super::Workarounds::FORCE_FILL_BUFFER_WITH_SIZE_GREATER_4096_ALIGNED_OFFSET_16,
                phd_capabilities.properties.vendor_id == db::nvidia::VENDOR,
            );
            workarounds.set(
                super::Workarounds::IGNORED_NEGATIVE_VIEWPORT_HEIGHT,
                phd_capabilities
                    .driver
                    .as_ref()
                    .is_some_and(|driver| driver.driver_id == vk::DriverId::ARM_PROPRIETARY),
            );
        };

        if let Some(driver) = phd_capabilities.driver {
            if driver.conformance_version.major == 0 {
                if driver.driver_id == vk::DriverId::MOLTENVK {
                    log::debug!("Adapter is not Vulkan compliant, but is MoltenVK, continuing");
                } else if self
                    .shared
                    .flags
                    .contains(wgt::InstanceFlags::ALLOW_UNDERLYING_NONCOMPLIANT_ADAPTER)
                {
                    log::debug!("Adapter is not Vulkan compliant: {}", info.name);
                } else {
                    log::debug!(
                        "Adapter is not Vulkan compliant, hiding adapter: {}",
                        info.name
                    );
                    return None;
                }
            }
        }
        if phd_capabilities.device_api_version == vk::API_VERSION_1_0
            && !phd_capabilities.supports_extension(khr::storage_buffer_storage_class::NAME)
        {
            log::debug!(
                "SPIR-V storage buffer class is not supported, hiding adapter: {}",
                info.name
            );
            return None;
        }
        if !phd_capabilities.supports_extension(khr::maintenance1::NAME)
            && phd_capabilities.device_api_version < vk::API_VERSION_1_1
        {
            log::debug!(
                "VK_KHR_maintenance1 is not supported, hiding adapter: {}",
                info.name
            );
            return None;
        }

        let queue_families = unsafe {
            self.shared
                .raw
                .get_physical_device_queue_family_properties(phd)
        };
        let queue_family_properties = queue_families.first()?;
        let queue_flags = queue_family_properties.queue_flags;
        if !queue_flags.contains(vk::QueueFlags::GRAPHICS) {
            log::debug!("The first queue only exposes {queue_flags:?}");
            return None;
        }

        let (available_features, mut downlevel_flags) = phd_features.to_wgpu(
            &self.shared.raw,
            phd,
            &phd_capabilities,
            queue_family_properties,
        );

        if phd_capabilities.is_driver(vk::DriverId::MESA_LLVMPIPE) {
            // The `F16_IN_F32` instructions do not normally require native `F16` support, but on
            // llvmpipe, they do.
            downlevel_flags.set(
                wgt::DownlevelFlags::SHADER_F16_IN_F32,
                available_features.contains(wgt::Features::SHADER_F16),
            );
        }

        downlevel_flags.set(
            wgt::DownlevelFlags::TEXTURE_COMPRESSION,
            available_features.contains(wgt::Features::TEXTURE_COMPRESSION_BC)
                || available_features.contains(
                    wgt::Features::TEXTURE_COMPRESSION_ETC2
                        | wgt::Features::TEXTURE_COMPRESSION_ASTC,
                ),
        );

        let has_robust_buffer_access2 = phd_features
            .robustness2
            .as_ref()
            .map(|r| r.robust_buffer_access2 == 1)
            .unwrap_or_default();

        let alignments = phd_capabilities.to_hal_alignments(has_robust_buffer_access2);

        let private_caps = super::PrivateCapabilities {
            image_view_usage: phd_capabilities.device_api_version >= vk::API_VERSION_1_1
                || phd_capabilities.supports_extension(khr::maintenance2::NAME),
            // VeriSilicon's proprietary driver, as on NXP's i.MX 8 SoCs, advertises timeline
            // semaphores but crashes in `vkQueueSubmit` when wgpu uses them. It reports
            // `VK_DRIVER_ID_MESA_RADV` as its driver ID, so only the vendor ID identifies it.
            timeline_semaphores: phd_capabilities.properties.vendor_id != db::verisilicon::VENDOR
                && match phd_features.timeline_semaphore {
                    Some(features) => features.timeline_semaphore == vk::TRUE,
                    None => phd_features
                        .timeline_semaphore
                        .is_some_and(|ext| ext.timeline_semaphore != 0),
                },
            texture_d24: supports_format(
                &self.shared.raw,
                phd,
                vk::Format::X8_D24_UNORM_PACK32,
                vk::ImageTiling::OPTIMAL,
                depth_stencil_required_flags(),
            ),
            texture_d24_s8: supports_format(
                &self.shared.raw,
                phd,
                vk::Format::D24_UNORM_S8_UINT,
                vk::ImageTiling::OPTIMAL,
                depth_stencil_required_flags(),
            ),
            texture_s8: supports_format(
                &self.shared.raw,
                phd,
                vk::Format::S8_UINT,
                vk::ImageTiling::OPTIMAL,
                depth_stencil_required_flags(),
            ),
            multi_draw_indirect: phd_features.core.multi_draw_indirect != 0,
            max_draw_indirect_count: phd_capabilities.properties.limits.max_draw_indirect_count,
            non_coherent_map_mask: phd_capabilities.properties.limits.non_coherent_atom_size - 1,
            can_present: true,
            //TODO: make configurable
            robust_buffer_access: phd_features.core.robust_buffer_access != 0,
            robust_image_access: phd_features
                .robustness2
                .is_some_and(|f| f.robust_image_access2 != 0)
                || phd_features
                    .image_robustness
                    .is_some_and(|ext| ext.robust_image_access != 0),
            robust_buffer_access2: has_robust_buffer_access2,
            robust_image_access2: phd_features
                .robustness2
                .as_ref()
                .map(|r| r.robust_image_access2 == 1)
                .unwrap_or_default(),
            zero_initialize_workgroup_memory: phd_features
                .zero_initialize_workgroup_memory
                .is_some_and(|ext| ext.shader_zero_initialize_workgroup_memory == vk::TRUE),
            image_format_list: phd_capabilities.device_api_version >= vk::API_VERSION_1_2
                || phd_capabilities.supports_extension(khr::image_format_list::NAME),
            maximum_samplers: phd_capabilities
                .properties
                .limits
                .max_sampler_allocation_count,
            shader_integer_dot_product: phd_features
                .shader_integer_dot_product
                .is_some_and(|ext| ext.shader_integer_dot_product != 0),
            shader_int8: phd_features
                .shader_float16_int8
                .is_some_and(|features| features.shader_int8 != 0),
            multiview_instance_index_limit: phd_capabilities
                .multiview
                .map(|a| a.max_multiview_instance_index)
                .unwrap_or(0),
            scratch_buffer_alignment: alignments.ray_tracing_scratch_buffer_alignment,
            depth_stencil_swizzle_one_support: phd_capabilities
                .maintenance_5
                .map(|maintenance_5| maintenance_5.depth_stencil_swizzle_one_support == vk::TRUE)
                .unwrap_or(false),
            ray_tracing_pipeline_group_data_size: alignments.ray_tracing_pipeline_group_data_size,
            store_op_none: phd_capabilities.device_api_version >= vk::API_VERSION_1_3
                || (phd_capabilities.device_api_version >= vk::API_VERSION_1_2
                    && phd_capabilities.supports_extension(khr::dynamic_rendering::NAME))
                || phd_capabilities.supports_extension(khr::load_store_op_none::NAME)
                || phd_capabilities.supports_extension(ash::qcom::render_pass_store_ops::NAME)
                || phd_capabilities.supports_extension(ext::load_store_op_none::NAME),
        };
        let capabilities = crate::Capabilities {
            limits: phd_capabilities.to_wgpu_limits(),
            alignments,
            downlevel: wgt::DownlevelCapabilities {
                flags: downlevel_flags,
                limits: wgt::DownlevelLimits {},
                shader_model: wgt::ShaderModel::Sm5, //TODO?
            },
            cooperative_matrix_properties: phd_capabilities.cooperative_matrix_properties.clone(),
        };

        let adapter = super::Adapter {
            raw: phd,
            instance: Arc::clone(&self.shared),
            //queue_families,
            known_memory_flags: vk::MemoryPropertyFlags::DEVICE_LOCAL
                | vk::MemoryPropertyFlags::HOST_VISIBLE
                | vk::MemoryPropertyFlags::HOST_COHERENT
                | vk::MemoryPropertyFlags::HOST_CACHED
                | vk::MemoryPropertyFlags::LAZILY_ALLOCATED,
            phd_capabilities,
            phd_features,
            downlevel_flags,
            private_caps,
            workarounds,
        };

        Some(crate::ExposedAdapter {
            adapter,
            info,
            features: available_features,
            capabilities,
        })
    }
}

impl super::Adapter {
    pub fn raw_physical_device(&self) -> vk::PhysicalDevice {
        self.raw
    }

    pub fn get_physical_device_features(&self) -> &PhysicalDeviceFeatures {
        &self.phd_features
    }

    pub fn physical_device_capabilities(&self) -> &PhysicalDeviceProperties {
        &self.phd_capabilities
    }

    pub fn shared_instance(&self) -> &super::InstanceShared {
        &self.instance
    }

    pub fn required_device_extensions(&self, features: wgt::Features) -> Vec<&'static CStr> {
        let (supported_extensions, unsupported_extensions) = self
            .phd_capabilities
            .get_required_extensions(features)
            .iter()
            .partition::<Vec<&CStr>, _>(|&&extension| {
                self.phd_capabilities.supports_extension(extension)
            });

        if !unsupported_extensions.is_empty() {
            log::debug!("Missing extensions: {unsupported_extensions:?}");
        }

        log::debug!("Supported extensions: {supported_extensions:?}");
        supported_extensions
    }

    /// Create a `PhysicalDeviceFeatures` for opening a logical device with
    /// `features` from this adapter.
    ///
    /// The given `enabled_extensions` set must include all the extensions
    /// selected by [`required_device_extensions`] when passed `features`.
    /// Otherwise, the `PhysicalDeviceFeatures` value may not be able to select
    /// all the Vulkan features needed to represent `features` and this
    /// adapter's characteristics.
    ///
    /// Typically, you'd simply call `required_device_extensions`, and then pass
    /// its return value and the feature set you gave it directly to this
    /// function. But it's fine to add more extensions to the list.
    ///
    /// [`required_device_extensions`]: Self::required_device_extensions
    pub fn physical_device_features(
        &self,
        enabled_extensions: &[&'static CStr],
        features: wgt::Features,
    ) -> PhysicalDeviceFeatures {
        PhysicalDeviceFeatures::from_extensions_and_requested_features(
            &self.phd_capabilities,
            &self.phd_features,
            enabled_extensions,
            features,
            self.downlevel_flags,
            &self.private_caps,
        )
    }

    /// # Safety
    ///
    /// - `raw_device` must be created from this adapter.
    /// - `raw_device` must be created using `family_index`, `enabled_extensions` and `physical_device_features()`
    /// - `enabled_extensions` must be a superset of `required_device_extensions()`.
    /// - If `drop_callback` is [`None`], wgpu-hal will take ownership of `raw_device`. If
    ///   `drop_callback` is [`Some`], `raw_device` must be valid until the callback is called.
    #[allow(clippy::too_many_arguments)]
    pub unsafe fn device_from_raw(
        &self,
        raw_device: ash::Device,
        drop_callback: Option<crate::DropCallback>,
        enabled_extensions: &[&'static CStr],
        features: wgt::Features,
        limits: &wgt::Limits,
        memory_hints: &wgt::MemoryHints,
        family_index: u32,
        queue_index: u32,
    ) -> Result<crate::OpenDevice<super::Api>, crate::DeviceError> {
        let mem_properties = {
            profiling::scope!("vkGetPhysicalDeviceMemoryProperties");
            unsafe {
                self.instance
                    .raw
                    .get_physical_device_memory_properties(self.raw)
            }
        };
        let queue_flags = unsafe {
            self.instance
                .raw
                .get_physical_device_queue_family_properties(self.raw)
                .get(family_index as usize)
                .map(|queue_family_properties| queue_family_properties.queue_flags)
                .ok_or(crate::DeviceError::Unexpected)?
        };
        let memory_types = &mem_properties.memory_types_as_slice();
        let valid_ash_memory_types = memory_types.iter().enumerate().fold(0, |u, (i, mem)| {
            if self.known_memory_flags.contains(mem.property_flags) {
                u | (1 << i)
            } else {
                u
            }
        });

        // Note that VK_EXT_debug_utils is an instance extension (enabled at the instance
        // level) but contains a few functions that can be loaded directly on the Device for a
        // dispatch-table-less pointer.
        let debug_utils_fn = if self.instance.extensions.contains(&ext::debug_utils::NAME) {
            Some(ext::debug_utils::Device::new(
                &self.instance.raw,
                &raw_device,
            ))
        } else {
            None
        };
        let indirect_count_fn = if enabled_extensions.contains(&khr::draw_indirect_count::NAME) {
            Some(khr::draw_indirect_count::Device::new(
                &self.instance.raw,
                &raw_device,
            ))
        } else {
            None
        };
        let timeline_semaphore_fn = if enabled_extensions.contains(&khr::timeline_semaphore::NAME) {
            Some(super::ExtensionFn::Extension(
                khr::timeline_semaphore::Device::new(&self.instance.raw, &raw_device),
            ))
        } else if self.phd_capabilities.device_api_version >= vk::API_VERSION_1_2 {
            Some(super::ExtensionFn::Promoted)
        } else {
            None
        };
        let ray_tracing_fns = if enabled_extensions.contains(&khr::acceleration_structure::NAME)
            && enabled_extensions.contains(&khr::buffer_device_address::NAME)
        {
            Some(super::RayTracingDeviceExtensionFunctions {
                acceleration_structure: khr::acceleration_structure::Device::new(
                    &self.instance.raw,
                    &raw_device,
                ),
                buffer_device_address: khr::buffer_device_address::Device::new(
                    &self.instance.raw,
                    &raw_device,
                ),
            })
        } else {
            None
        };
        let ray_tracing_pipeline_fns =
            if enabled_extensions.contains(&khr::ray_tracing_pipeline::NAME) {
                Some(khr::ray_tracing_pipeline::Device::new(
                    &self.instance.raw,
                    &raw_device,
                ))
            } else {
                None
            };
        let mesh_shading_fns = if enabled_extensions.contains(&ext::mesh_shader::NAME) {
            Some(ext::mesh_shader::Device::new(
                &self.instance.raw,
                &raw_device,
            ))
        } else {
            None
        };
        let external_memory_fd_fn = if enabled_extensions.contains(&khr::external_memory_fd::NAME) {
            Some(khr::external_memory_fd::Device::new(
                &self.instance.raw,
                &raw_device,
            ))
        } else {
            None
        };

        let naga_options = {
            use naga::back::spv;

            // The following capabilities are always available
            // see https://registry.khronos.org/vulkan/specs/1.3-extensions/html/chap52.html#spirvenv-capabilities
            let mut capabilities = vec![
                spv::Capability::Shader,
                spv::Capability::Matrix,
                spv::Capability::Sampled1D,
                spv::Capability::Image1D,
                spv::Capability::ImageQuery,
                spv::Capability::DerivativeControl,
                spv::Capability::StorageImageExtendedFormats,
            ];

            if self
                .downlevel_flags
                .contains(wgt::DownlevelFlags::CUBE_ARRAY_TEXTURES)
            {
                capabilities.push(spv::Capability::SampledCubeArray);
            }

            if self
                .downlevel_flags
                .contains(wgt::DownlevelFlags::MULTISAMPLED_SHADING)
            {
                capabilities.push(spv::Capability::SampleRateShading);
            }

            if features.contains(wgt::Features::MULTIVIEW) {
                capabilities.push(spv::Capability::MultiView);
            }

            if features.contains(wgt::Features::PRIMITIVE_INDEX) {
                capabilities.push(spv::Capability::Geometry);
            }

            if features.intersects(wgt::Features::SUBGROUP | wgt::Features::SUBGROUP_VERTEX) {
                capabilities.push(spv::Capability::GroupNonUniform);
                capabilities.push(spv::Capability::GroupNonUniformVote);
                capabilities.push(spv::Capability::GroupNonUniformArithmetic);
                capabilities.push(spv::Capability::GroupNonUniformBallot);
                capabilities.push(spv::Capability::GroupNonUniformShuffle);
                capabilities.push(spv::Capability::GroupNonUniformShuffleRelative);
                capabilities.push(spv::Capability::GroupNonUniformQuad);
            }

            if features.intersects(
                wgt::Features::SAMPLED_TEXTURE_AND_STORAGE_BUFFER_ARRAY_NON_UNIFORM_INDEXING
                    | wgt::Features::STORAGE_TEXTURE_ARRAY_NON_UNIFORM_INDEXING
                    | wgt::Features::UNIFORM_BUFFER_BINDING_ARRAYS,
            ) {
                capabilities.push(spv::Capability::ShaderNonUniform);
            }
            if features.contains(wgt::Features::BGRA8UNORM_STORAGE) {
                capabilities.push(spv::Capability::StorageImageWriteWithoutFormat);
            }

            if features.contains(wgt::Features::EXPERIMENTAL_RAY_QUERY) {
                capabilities.push(spv::Capability::RayQueryKHR);
            }

            if features.contains(wgt::Features::SHADER_INT64) {
                capabilities.push(spv::Capability::Int64);
            }

            if features.contains(wgt::Features::SHADER_F16) {
                capabilities.push(spv::Capability::Float16);
            }

            if features.contains(wgt::Features::SHADER_I16) {
                capabilities.push(spv::Capability::Int16);
            }

            if features.intersects(
                wgt::Features::SHADER_INT64_ATOMIC_ALL_OPS
                    | wgt::Features::SHADER_INT64_ATOMIC_MIN_MAX
                    | wgt::Features::TEXTURE_INT64_ATOMIC,
            ) {
                capabilities.push(spv::Capability::Int64Atomics);
            }

            if features.intersects(wgt::Features::TEXTURE_INT64_ATOMIC) {
                capabilities.push(spv::Capability::Int64ImageEXT);
            }

            if features.contains(wgt::Features::SHADER_FLOAT32_ATOMIC) {
                capabilities.push(spv::Capability::AtomicFloat32AddEXT);
            }

            if features.contains(wgt::Features::CLIP_DISTANCES) {
                capabilities.push(spv::Capability::ClipDistance);
            }

            // Vulkan bundles both barycentrics and per-vertex attributes under the same feature.
            if features
                .intersects(wgt::Features::SHADER_BARYCENTRICS | wgt::Features::SHADER_PER_VERTEX)
            {
                capabilities.push(spv::Capability::FragmentBarycentricKHR);
            }

            if features.contains(wgt::Features::SHADER_DRAW_INDEX) {
                capabilities.push(spv::Capability::DrawParameters);
            }

            let mut flags = spv::WriterFlags::empty();
            flags.set(
                spv::WriterFlags::DEBUG,
                self.instance.flags.contains(wgt::InstanceFlags::DEBUG),
            );
            flags.set(
                spv::WriterFlags::LABEL_VARYINGS,
                self.phd_capabilities.properties.vendor_id != crate::auxil::db::qualcomm::VENDOR,
            );
            flags.set(
                spv::WriterFlags::ADJUST_COORDINATE_SPACE,
                self.workarounds
                    .contains(super::Workarounds::IGNORED_NEGATIVE_VIEWPORT_HEIGHT),
            );
            flags.set(
                spv::WriterFlags::FORCE_POINT_SIZE,
                //Note: we could technically disable this when we are compiling separate entry points,
                // and we know exactly that the primitive topology is not `PointList`.
                // But this requires cloning the `spv::Options` struct, which has heap allocations.
                true, // could check `super::Workarounds::SEPARATE_ENTRY_POINTS`
            );
            flags.set(
                spv::WriterFlags::PRINT_ON_RAY_QUERY_INITIALIZATION_FAIL
                    | spv::WriterFlags::PRINT_ON_TRACE_RAYS_FAIL,
                self.instance.flags.contains(wgt::InstanceFlags::DEBUG)
                    && (self.instance.instance_api_version >= vk::API_VERSION_1_3
                        || enabled_extensions.contains(&khr::shader_non_semantic_info::NAME)),
            );
            if features.contains(wgt::Features::EXPERIMENTAL_RAY_QUERY) {
                capabilities.push(spv::Capability::RayQueryKHR);
            }
            if features.contains(wgt::Features::EXPERIMENTAL_RAY_TRACING_PIPELINES) {
                capabilities.push(spv::Capability::RayTracingKHR);
            }
            if features.contains(wgt::Features::EXPERIMENTAL_RAY_HIT_VERTEX_RETURN) {
                capabilities.push(spv::Capability::RayQueryPositionFetchKHR)
            }
            if features.contains(wgt::Features::EXPERIMENTAL_MESH_SHADER) {
                capabilities.push(spv::Capability::MeshShadingEXT);
            }
            if features.contains(wgt::Features::EXPERIMENTAL_COOPERATIVE_MATRIX) {
                capabilities.push(spv::Capability::CooperativeMatrixKHR);
                // TODO: expose this more generally
                capabilities.push(spv::Capability::VulkanMemoryModel);
            }
            if self.private_caps.shader_integer_dot_product {
                // See <https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/VK_KHR_shader_integer_dot_product.html#_new_spir_v_capabilities>.
                capabilities.extend(&[
                    spv::Capability::DotProductInputAllKHR,
                    spv::Capability::DotProductInput4x8BitKHR,
                    spv::Capability::DotProductInput4x8BitPackedKHR,
                    spv::Capability::DotProductKHR,
                ]);
            }
            if self.private_caps.shader_int8 {
                // See <https://registry.khronos.org/vulkan/specs/latest/man/html/VkPhysicalDeviceShaderFloat16Int8Features.html#extension-features-shaderInt8>.
                capabilities.extend(&[spv::Capability::Int8]);
            }
            spv::Options {
                lang_version: {
                    // Use maximum supported SPIR-V version according to
                    // <https://github.com/KhronosGroup/Vulkan-Docs/blob/19b7651/appendices/spirvenv.adoc?plain=1#L21-L40>.
                    let version = match self.phd_capabilities.device_api_version {
                        vk::API_VERSION_1_0..vk::API_VERSION_1_1 => (1, 0),
                        vk::API_VERSION_1_1..vk::API_VERSION_1_2 => (1, 3),
                        vk::API_VERSION_1_2..vk::API_VERSION_1_3 => (1, 5),
                        vk::API_VERSION_1_3.. => (1, 6),
                        _ => unreachable!(),
                    };
                    // `VK_KHR_spirv_1_4` raises that ceiling on pre-1.2 devices, which
                    // both mesh shaders and ray tracing pipelines need it to do.
                    if enabled_extensions.contains(&khr::spirv_1_4::NAME) {
                        version.max((1, 4))
                    } else {
                        version
                    }
                },
                flags,
                capabilities: Some(capabilities.iter().cloned().collect()),
                bounds_check_policies: naga::proc::BoundsCheckPolicies {
                    index: naga::proc::BoundsCheckPolicy::Restrict,
                    buffer: if self.private_caps.robust_buffer_access2 {
                        naga::proc::BoundsCheckPolicy::Unchecked
                    } else {
                        naga::proc::BoundsCheckPolicy::Restrict
                    },
                    image_load: if self.private_caps.robust_image_access {
                        naga::proc::BoundsCheckPolicy::Unchecked
                    } else {
                        naga::proc::BoundsCheckPolicy::Restrict
                    },
                    // TODO: support bounds checks on binding arrays
                    binding_array: naga::proc::BoundsCheckPolicy::Unchecked,
                },
                zero_initialize_workgroup_memory: if self
                    .private_caps
                    .zero_initialize_workgroup_memory
                {
                    spv::ZeroInitializeWorkgroupMemoryMode::Native
                } else {
                    spv::ZeroInitializeWorkgroupMemoryMode::Polyfill
                },
                force_loop_bounding: true,
                ray_query_initialization_tracking: true,
                use_storage_input_output_16: features.contains(wgt::Features::SHADER_F16)
                    && self.phd_features.supports_storage_input_output_16(),
                fake_missing_bindings: false,
                // We need to build this separately for each invocation, so just default it out here
                binding_map: BTreeMap::default(),
                debug_info: None,
                task_dispatch_limits: Some(naga::back::TaskDispatchLimits {
                    max_mesh_workgroups_per_dim: limits.max_mesh_workgroups_per_dimension,
                    max_mesh_workgroups_total: limits.max_mesh_workgroup_total_count,
                }),
                mesh_shader_primitive_indices_clamp: true,
                trace_ray_argument_validation: true,
                emit_int_div_checks: true,
            }
        };

        let raw_queue = {
            profiling::scope!("vkGetDeviceQueue");
            unsafe { raw_device.get_device_queue(family_index, queue_index) }
        };

        let driver_version = self
            .phd_capabilities
            .properties
            .driver_version
            .to_be_bytes();
        #[rustfmt::skip]
        let pipeline_cache_validation_key = [
            driver_version[0], driver_version[1], driver_version[2], driver_version[3],
            0, 0, 0, 0,
            0, 0, 0, 0,
            0, 0, 0, 0,
        ];

        let drop_guard = crate::DropGuard::from_option(drop_callback);

        let empty_descriptor_set_layout = unsafe {
            raw_device
                .create_descriptor_set_layout(&vk::DescriptorSetLayoutCreateInfo::default(), None)
                .map_err(super::map_host_device_oom_err)?
        };

        let shared = Arc::new(super::DeviceShared {
            raw: raw_device,
            family_index,
            queue_flags,
            queue_index,
            raw_queue,
            drop_guard,
            instance: Arc::clone(&self.instance),
            physical_device: self.raw,
            enabled_extensions: enabled_extensions.into(),
            extension_fns: super::DeviceExtensionFunctions {
                debug_utils: debug_utils_fn,
                draw_indirect_count: indirect_count_fn,
                timeline_semaphore: timeline_semaphore_fn,
                ray_tracing: ray_tracing_fns,
                ray_tracing_pipelines: ray_tracing_pipeline_fns,
                mesh_shading: mesh_shading_fns,
                external_memory_fd: external_memory_fd_fn,
            },
            pipeline_cache_validation_key,
            vendor_id: self.phd_capabilities.properties.vendor_id,
            timestamp_period: self.phd_capabilities.properties.limits.timestamp_period,
            private_caps: self.private_caps.clone(),
            features,
            workarounds: self.workarounds,
            render_passes: Mutex::new(Default::default()),
            sampler_cache: Mutex::new(super::sampler::SamplerCache::new(
                self.private_caps.maximum_samplers,
            )),
            memory_allocations_counter: Default::default(),

            texture_identity_factory: super::ResourceIdentityFactory::new(),
            texture_view_identity_factory: super::ResourceIdentityFactory::new(),
            empty_descriptor_set_layout,
        });

        let relay_semaphores = super::RelaySemaphores::new(&shared)?;

        let queue = super::Queue {
            raw: raw_queue,
            device: Arc::clone(&shared),
            family_index,
            relay_semaphores: Mutex::new(relay_semaphores),
            signal_semaphores: Mutex::new(SemaphoreList::new(SemaphoreListMode::Signal)),
            wait_semaphores: Mutex::new(SemaphoreList::new(SemaphoreListMode::Wait)),
            next_submit_chain: Mutex::new(None),
        };

        let allocation_sizes = AllocationSizes::from_memory_hints(memory_hints);

        let buffer_device_address = enabled_extensions.contains(&khr::buffer_device_address::NAME);

        let mem_allocator = super::MemoryAllocators::new(
            self.instance.raw.clone(),
            shared.raw.clone(),
            self.raw,
            buffer_device_address,
            allocation_sizes,
        )?;

        let desc_allocator = super::descriptor::DescriptorAllocator::new(
            if let Some(di) = self.phd_capabilities.descriptor_indexing {
                di.max_update_after_bind_descriptors_in_all_pools
            } else {
                0
            },
        );

        let device = super::Device {
            shared,
            mem_allocator,
            desc_allocator: Mutex::new(desc_allocator),
            valid_ash_memory_types,
            naga_options,
            #[cfg(feature = "renderdoc")]
            render_doc: Default::default(),
            counters: Default::default(),
        };

        Ok(crate::OpenDevice { device, queue })
    }

    pub fn texture_format_as_raw(&self, texture_format: wgt::TextureFormat) -> vk::Format {
        self.private_caps.map_texture_format(texture_format)
    }

    /// # Safety:
    /// - Same as `open` plus
    /// - The callback may not change anything that the device does not support.
    /// - The callback may not remove features.
    pub unsafe fn open_with_callback<'a>(
        &self,
        features: wgt::Features,
        limits: &wgt::Limits,
        memory_hints: &wgt::MemoryHints,
        callback: Option<Box<super::CreateDeviceCallback<'a>>>,
    ) -> Result<crate::OpenDevice<super::Api>, crate::DeviceError> {
        let mut enabled_extensions = self.required_device_extensions(features);
        let mut enabled_phd_features = self.physical_device_features(&enabled_extensions, features);

        let default_family_index = 0;
        let mut family_infos = vec![vk::DeviceQueueCreateInfo::default()
            .queue_family_index(default_family_index)
            .queue_priorities(&[1.0])];

        let mut pre_info = vk::DeviceCreateInfo::default();

        if let Some(callback) = callback {
            callback(super::CreateDeviceCallbackArgs {
                extensions: &mut enabled_extensions,
                device_features: &mut enabled_phd_features,
                queue_create_infos: &mut family_infos,
                create_info: &mut pre_info,
                _phantom: PhantomData,
            })
        }

        let str_pointers = enabled_extensions
            .iter()
            .map(|&s| {
                // Safe because `enabled_extensions` entries have static lifetime.
                s.as_ptr()
            })
            .collect::<Vec<_>>();

        let pre_info = pre_info
            .queue_create_infos(&family_infos)
            .enabled_extension_names(&str_pointers);
        let info = enabled_phd_features.add_to_device_create(pre_info);
        let raw_device = {
            profiling::scope!("vkCreateDevice");
            unsafe {
                self.instance
                    .raw
                    .create_device(self.raw, &info, None)
                    .map_err(map_err)?
            }
        };
        fn map_err(err: vk::Result) -> crate::DeviceError {
            match err {
                vk::Result::ERROR_TOO_MANY_OBJECTS => crate::DeviceError::OutOfMemory,
                vk::Result::ERROR_INITIALIZATION_FAILED => crate::DeviceError::Lost,
                vk::Result::ERROR_EXTENSION_NOT_PRESENT | vk::Result::ERROR_FEATURE_NOT_PRESENT => {
                    crate::hal_usage_error(err)
                }
                other => super::map_host_device_oom_and_lost_err(other),
            }
        }

        unsafe {
            self.device_from_raw(
                raw_device,
                None,
                &enabled_extensions,
                features,
                limits,
                memory_hints,
                family_infos[0].queue_family_index,
                0,
            )
        }
    }
}

impl crate::Adapter for super::Adapter {
    type A = super::Api;

    unsafe fn open(
        &self,
        features: wgt::Features,
        limits: &wgt::Limits,
        memory_hints: &wgt::MemoryHints,
    ) -> Result<crate::OpenDevice<super::Api>, crate::DeviceError> {
        unsafe { self.open_with_callback(features, limits, memory_hints, None) }
    }

    unsafe fn texture_format_capabilities(
        &self,
        format: wgt::TextureFormat,
    ) -> crate::TextureFormatCapabilities {
        use crate::TextureFormatCapabilities as Tfc;

        let vk_format = self.private_caps.map_texture_format(format);
        let properties = unsafe {
            self.instance
                .raw
                .get_physical_device_format_properties(self.raw, vk_format)
        };
        let features = properties.optimal_tiling_features;

        let mut flags = Tfc::empty();
        flags.set(
            Tfc::SAMPLED,
            features.contains(vk::FormatFeatureFlags::SAMPLED_IMAGE),
        );
        flags.set(
            Tfc::SAMPLED_LINEAR,
            features.contains(vk::FormatFeatureFlags::SAMPLED_IMAGE_FILTER_LINEAR),
        );
        // flags.set(
        //     Tfc::SAMPLED_MINMAX,
        //     features.contains(vk::FormatFeatureFlags::SAMPLED_IMAGE_FILTER_MINMAX),
        // );
        flags.set(
            Tfc::STORAGE_READ_WRITE
                | Tfc::STORAGE_WRITE_ONLY
                | Tfc::STORAGE_READ_ONLY
                | Tfc::STORAGE_ATOMIC,
            features.contains(vk::FormatFeatureFlags::STORAGE_IMAGE),
        );
        flags.set(
            Tfc::STORAGE_ATOMIC,
            features.contains(vk::FormatFeatureFlags::STORAGE_IMAGE_ATOMIC),
        );
        flags.set(
            Tfc::COLOR_ATTACHMENT,
            features.contains(vk::FormatFeatureFlags::COLOR_ATTACHMENT),
        );
        flags.set(
            Tfc::COLOR_ATTACHMENT_BLEND,
            features.contains(vk::FormatFeatureFlags::COLOR_ATTACHMENT_BLEND),
        );
        flags.set(
            Tfc::DEPTH_STENCIL_ATTACHMENT,
            features.contains(vk::FormatFeatureFlags::DEPTH_STENCIL_ATTACHMENT),
        );
        flags.set(
            Tfc::COPY_SRC,
            features.intersects(vk::FormatFeatureFlags::TRANSFER_SRC),
        );
        flags.set(
            Tfc::COPY_DST,
            features.intersects(vk::FormatFeatureFlags::TRANSFER_DST),
        );
        flags.set(
            Tfc::STORAGE_ATOMIC,
            features.intersects(vk::FormatFeatureFlags::STORAGE_IMAGE_ATOMIC),
        );
        // Vulkan is very permissive about MSAA
        flags.set(Tfc::MULTISAMPLE_RESOLVE, !format.is_compressed());

        // get the supported sample counts
        let format_aspect = crate::FormatAspects::from(format);
        let limits = self.phd_capabilities.properties.limits;

        let sample_flags = if format_aspect.contains(crate::FormatAspects::DEPTH) {
            limits
                .framebuffer_depth_sample_counts
                .min(limits.sampled_image_depth_sample_counts)
        } else if format_aspect.contains(crate::FormatAspects::STENCIL) {
            limits
                .framebuffer_stencil_sample_counts
                .min(limits.sampled_image_stencil_sample_counts)
        } else {
            let first_aspect = format_aspect
                .iter()
                .next()
                .expect("All texture should at least one aspect")
                .map();

            // We should never get depth or stencil out of this, due to the above.
            assert_ne!(first_aspect, wgt::TextureAspect::DepthOnly);
            assert_ne!(first_aspect, wgt::TextureAspect::StencilOnly);

            match format.sample_type(Some(first_aspect), None).unwrap() {
                wgt::TextureSampleType::Float { .. } => limits
                    .framebuffer_color_sample_counts
                    .min(limits.sampled_image_color_sample_counts),
                wgt::TextureSampleType::Sint | wgt::TextureSampleType::Uint => {
                    limits.sampled_image_integer_sample_counts
                }
                _ => unreachable!(),
            }
        };

        flags.set(
            Tfc::MULTISAMPLE_X2,
            sample_flags.contains(vk::SampleCountFlags::TYPE_2),
        );
        flags.set(
            Tfc::MULTISAMPLE_X4,
            sample_flags.contains(vk::SampleCountFlags::TYPE_4),
        );
        flags.set(
            Tfc::MULTISAMPLE_X8,
            sample_flags.contains(vk::SampleCountFlags::TYPE_8),
        );
        flags.set(
            Tfc::MULTISAMPLE_X16,
            sample_flags.contains(vk::SampleCountFlags::TYPE_16),
        );

        flags
    }

    unsafe fn surface_capabilities(
        &self,
        surface: &super::Surface,
    ) -> Option<crate::SurfaceCapabilities> {
        surface.inner.surface_capabilities(self)
    }

    unsafe fn surface_display_hdr_info(
        &self,
        surface: &super::Surface,
    ) -> Option<wgt::DisplayHdrInfo> {
        // Vulkan has no portable luminance query; the Win32 surface reads it
        // through DXGI (see `dxgi::hdr`). Every other surface reports `None`.
        surface.inner.display_hdr_info()
    }

    unsafe fn get_presentation_timestamp(&self) -> wgt::PresentationTimestamp {
        // VK_GOOGLE_display_timing is the only way to get presentation
        // timestamps on vulkan right now and it is only ever available
        // on android and linux. This includes mac, but there's no alternative
        // on mac, so this is fine.
        #[cfg(unix)]
        {
            let mut timespec = libc::timespec::default();
            unsafe {
                libc::clock_gettime(libc::CLOCK_MONOTONIC, &mut timespec);
            }

            wgt::PresentationTimestamp(
                timespec.tv_sec as u128 * 1_000_000_000 + timespec.tv_nsec as u128,
            )
        }
        #[cfg(not(unix))]
        {
            wgt::PresentationTimestamp::INVALID_TIMESTAMP
        }
    }

    fn get_ordered_buffer_usages(&self) -> wgt::BufferUses {
        wgt::BufferUses::INCLUSIVE | wgt::BufferUses::MAP_WRITE
    }

    // Vulkan makes very few execution ordering guarantees
    // see https://registry.khronos.org/vulkan/specs/latest/html/vkspec.html#synchronization-implicit
    // We just don't want to insert barriers between inclusive uses
    // See https://github.com/gfx-rs/wgpu/issues/8853
    fn get_ordered_texture_usages(&self) -> wgt::TextureUses {
        wgt::TextureUses::INCLUSIVE
    }
}

fn is_format_16bit_norm_supported(instance: &ash::Instance, phd: vk::PhysicalDevice) -> bool {
    [
        vk::Format::R16_UNORM,
        vk::Format::R16_SNORM,
        vk::Format::R16G16_UNORM,
        vk::Format::R16G16_SNORM,
        vk::Format::R16G16B16A16_UNORM,
        vk::Format::R16G16B16A16_SNORM,
    ]
    .into_iter()
    .all(|format| {
        supports_format(
            instance,
            phd,
            format,
            vk::ImageTiling::OPTIMAL,
            vk::FormatFeatureFlags::SAMPLED_IMAGE
                | vk::FormatFeatureFlags::STORAGE_IMAGE
                | vk::FormatFeatureFlags::TRANSFER_SRC
                | vk::FormatFeatureFlags::TRANSFER_DST,
        )
    })
}

fn is_float32_filterable_supported(instance: &ash::Instance, phd: vk::PhysicalDevice) -> bool {
    [
        vk::Format::R32_SFLOAT,
        vk::Format::R32G32_SFLOAT,
        vk::Format::R32G32B32A32_SFLOAT,
    ]
    .into_iter()
    .all(|format| {
        supports_format(
            instance,
            phd,
            format,
            vk::ImageTiling::OPTIMAL,
            vk::FormatFeatureFlags::SAMPLED_IMAGE_FILTER_LINEAR,
        )
    })
}

fn is_float32_blendable_supported(instance: &ash::Instance, phd: vk::PhysicalDevice) -> bool {
    [
        vk::Format::R32_SFLOAT,
        vk::Format::R32G32_SFLOAT,
        vk::Format::R32G32B32A32_SFLOAT,
    ]
    .into_iter()
    .all(|format| {
        supports_format(
            instance,
            phd,
            format,
            vk::ImageTiling::OPTIMAL,
            vk::FormatFeatureFlags::COLOR_ATTACHMENT_BLEND,
        )
    })
}

pub(super) fn supports_format(
    instance: &ash::Instance,
    phd: vk::PhysicalDevice,
    format: vk::Format,
    tiling: vk::ImageTiling,
    features: vk::FormatFeatureFlags,
) -> bool {
    let properties = unsafe { instance.get_physical_device_format_properties(phd, format) };
    match tiling {
        vk::ImageTiling::LINEAR => properties.linear_tiling_features.contains(features),
        vk::ImageTiling::OPTIMAL => properties.optimal_tiling_features.contains(features),
        _ => false,
    }
}

fn supports_astc_3d(instance: &ash::Instance, phd: vk::PhysicalDevice) -> bool {
    [
        vk::Format::ASTC_4X4_UNORM_BLOCK,
        vk::Format::ASTC_4X4_SRGB_BLOCK,
        vk::Format::ASTC_5X4_UNORM_BLOCK,
        vk::Format::ASTC_5X4_SRGB_BLOCK,
        vk::Format::ASTC_5X5_UNORM_BLOCK,
        vk::Format::ASTC_5X5_SRGB_BLOCK,
        vk::Format::ASTC_6X5_UNORM_BLOCK,
        vk::Format::ASTC_6X5_SRGB_BLOCK,
        vk::Format::ASTC_6X6_UNORM_BLOCK,
        vk::Format::ASTC_6X6_SRGB_BLOCK,
        vk::Format::ASTC_8X5_UNORM_BLOCK,
        vk::Format::ASTC_8X5_SRGB_BLOCK,
        vk::Format::ASTC_8X6_UNORM_BLOCK,
        vk::Format::ASTC_8X6_SRGB_BLOCK,
        vk::Format::ASTC_8X8_UNORM_BLOCK,
        vk::Format::ASTC_8X8_SRGB_BLOCK,
        vk::Format::ASTC_10X5_UNORM_BLOCK,
        vk::Format::ASTC_10X5_SRGB_BLOCK,
        vk::Format::ASTC_10X6_UNORM_BLOCK,
        vk::Format::ASTC_10X6_SRGB_BLOCK,
        vk::Format::ASTC_10X8_UNORM_BLOCK,
        vk::Format::ASTC_10X8_SRGB_BLOCK,
        vk::Format::ASTC_10X10_UNORM_BLOCK,
        vk::Format::ASTC_10X10_SRGB_BLOCK,
        vk::Format::ASTC_12X10_UNORM_BLOCK,
        vk::Format::ASTC_12X10_SRGB_BLOCK,
        vk::Format::ASTC_12X12_UNORM_BLOCK,
        vk::Format::ASTC_12X12_SRGB_BLOCK,
    ]
    .into_iter()
    .all(|format| {
        unsafe {
            instance.get_physical_device_image_format_properties(
                phd,
                format,
                vk::ImageType::TYPE_3D,
                vk::ImageTiling::OPTIMAL,
                vk::ImageUsageFlags::SAMPLED,
                vk::ImageCreateFlags::empty(),
            )
        }
        .is_ok()
    })
}

fn supports_bgra8unorm_storage(
    instance: &ash::Instance,
    phd: vk::PhysicalDevice,
    device_api_version: u32,
) -> bool {
    // See https://github.com/KhronosGroup/Vulkan-Docs/issues/2027#issuecomment-1380608011

    // This check gates the function call and structures used below.
    // TODO: check for (`VK_KHR_get_physical_device_properties2` or VK1.1) and (`VK_KHR_format_feature_flags2` or VK1.3).
    // Right now we only check for VK1.3.
    if device_api_version < vk::API_VERSION_1_3 {
        return false;
    }

    unsafe {
        let mut properties3 = vk::FormatProperties3::default();
        let mut properties2 = vk::FormatProperties2::default().push_next(&mut properties3);

        instance.get_physical_device_format_properties2(
            phd,
            vk::Format::B8G8R8A8_UNORM,
            &mut properties2,
        );

        let features2 = properties2.format_properties.optimal_tiling_features;
        let features3 = properties3.optimal_tiling_features;

        features2.contains(vk::FormatFeatureFlags::STORAGE_IMAGE)
            && features3.contains(vk::FormatFeatureFlags2::STORAGE_WRITE_WITHOUT_FORMAT)
    }
}

// For https://github.com/gfx-rs/wgpu/issues/4599
// Intel iGPUs with outdated drivers can break rendering if `VK_EXT_robustness2` is used.
// Driver version 31.0.101.2115 works, but there's probably an earlier functional version.
fn is_intel_igpu_outdated_for_robustness2(capabilities: &PhysicalDeviceProperties) -> bool {
    const DRIVER_VERSION_WORKING: u32 = (101 << 14) | 2115; // X.X.101.2115

    let props = &capabilities.properties;

    let is_outdated = props.vendor_id == crate::auxil::db::intel::VENDOR
        && props.device_type == vk::PhysicalDeviceType::INTEGRATED_GPU
        && props.driver_version < DRIVER_VERSION_WORKING
        && capabilities.is_driver(vk::DriverId::INTEL_PROPRIETARY_WINDOWS);

    if is_outdated {
        log::debug!(
            "Disabling robustBufferAccess2 and robustImageAccess2: IntegratedGpu Intel Driver is outdated. Found with version 0x{:X}, less than the known good version 0x{:X} (31.0.101.2115)",
            props.driver_version,
            DRIVER_VERSION_WORKING
        );
    }
    is_outdated
}

/// Convert Vulkan component type to wgt::CooperativeScalarType.
fn map_vk_component_type(ty: vk::ComponentTypeKHR) -> Option<wgt::CooperativeScalarType> {
    match ty {
        vk::ComponentTypeKHR::FLOAT16 => Some(wgt::CooperativeScalarType::F16),
        vk::ComponentTypeKHR::FLOAT32 => Some(wgt::CooperativeScalarType::F32),
        vk::ComponentTypeKHR::SINT32 => Some(wgt::CooperativeScalarType::I32),
        vk::ComponentTypeKHR::UINT32 => Some(wgt::CooperativeScalarType::U32),
        _ => None,
    }
}

/// Convert Vulkan matrix size.
fn map_vk_cooperative_size(size: u32) -> Option<u32> {
    match size {
        8 | 16 => Some(size),
        _ => None,
    }
}

/// Query all supported cooperative matrix configurations from Vulkan.
fn query_cooperative_matrix_properties(
    coop_matrix: &khr::cooperative_matrix::Instance,
    phd: vk::PhysicalDevice,
) -> Vec<wgt::CooperativeMatrixProperties> {
    let vk_properties =
        match unsafe { coop_matrix.get_physical_device_cooperative_matrix_properties(phd) } {
            Ok(props) => props,
            Err(e) => {
                log::warn!("Failed to query cooperative matrix properties: {e:?}");
                return Vec::new();
            }
        };

    log::debug!(
        "Vulkan reports {} cooperative matrix configurations",
        vk_properties.len()
    );

    let mut result = Vec::new();
    for prop in &vk_properties {
        log::debug!(
            "  Vulkan coop matrix: M={} N={} K={} A={:?} B={:?} C={:?} Result={:?} scope={:?} saturating={}",
            prop.m_size,
            prop.n_size,
            prop.k_size,
            prop.a_type,
            prop.b_type,
            prop.c_type,
            prop.result_type,
            prop.scope,
            prop.saturating_accumulation
        );

        // Only include subgroup-scoped operations (the only scope we support)
        if prop.scope != vk::ScopeKHR::SUBGROUP {
            log::debug!("    Skipped: scope is not SUBGROUP");
            continue;
        }

        // Map sizes - skip configurations with sizes we don't support
        let m_size = match map_vk_cooperative_size(prop.m_size) {
            Some(s) => s,
            None => {
                log::debug!("    Skipped: M size {} not supported", prop.m_size);
                continue;
            }
        };
        let n_size = match map_vk_cooperative_size(prop.n_size) {
            Some(s) => s,
            None => {
                log::debug!("    Skipped: N size {} not supported", prop.n_size);
                continue;
            }
        };
        let k_size = match map_vk_cooperative_size(prop.k_size) {
            Some(s) => s,
            None => {
                log::debug!("    Skipped: K size {} not supported", prop.k_size);
                continue;
            }
        };

        // Map the component types - A and B must match, C and Result must match
        let ab_type = match map_vk_component_type(prop.a_type) {
            Some(t) if Some(t) == map_vk_component_type(prop.b_type) => t,
            _ => {
                log::debug!(
                    "    Skipped: A/B types {:?}/{:?} not supported or don't match",
                    prop.a_type,
                    prop.b_type
                );
                continue;
            }
        };
        let cr_type = match map_vk_component_type(prop.c_type) {
            Some(t) if Some(t) == map_vk_component_type(prop.result_type) => t,
            _ => {
                log::debug!(
                    "    Skipped: C/Result types {:?}/{:?} not supported or don't match",
                    prop.c_type,
                    prop.result_type
                );
                continue;
            }
        };

        log::debug!("    Accepted!");
        result.push(wgt::CooperativeMatrixProperties {
            m_size,
            n_size,
            k_size,
            ab_type,
            cr_type,
            saturating_accumulation: prop.saturating_accumulation != 0,
        });
    }

    log::debug!(
        "Found {} cooperative matrix configurations supported by wgpu",
        result.len()
    );
    result
}