Quellcodebibliothek Statistik Leitseite products/Sources/formale Sprachen/C/Firefox/third_party/rust/gpu-allocator/src/vulkan/   (Browser von der Mozilla Stiftung Version 136.0.1©)  Datei vom 10.2.2025 mit Größe 35 kB image not shown  

Quelle  mod.rs   Sprache: unbekannt

 
Spracherkennung für: .rs vermutete Sprache: Unknown {[0] [0] [0]} [Methode: Schwerpunktbildung, einfache Gewichte, sechs Dimensionen]

#![deny(clippy::unimplemented, clippy::unwrap_used, clippy::ok_expect)]

#[cfg(feature = "visualizer")]
mod visualizer;
use std::{backtrace::Backtrace, fmt, marker::PhantomData, sync::Arc};

use ash::vk;
use log::{debug, Level};
#[cfg(feature = "visualizer")]
pub use visualizer::AllocatorVisualizer;

use super::allocator;
use crate::{
    allocator::{AllocatorReport, MemoryBlockReport},
    AllocationError, AllocationSizes, AllocatorDebugSettings, MemoryLocation, Result,
};

#[derive(Copy, Clone, Debug, Eq, PartialEq)]
pub enum AllocationScheme {
    /// Perform a dedicated, driver-managed allocation for the given buffer, allowing
    /// it to perform optimizations on this type of allocation.
    DedicatedBuffer(vk::Buffer),
    /// Perform a dedicated, driver-managed allocation for the given image, allowing
    /// it to perform optimizations on this type of allocation.
    DedicatedImage(vk::Image),
    /// The memory for this resource will be allocated and managed by gpu-allocator.
    GpuAllocatorManaged,
}

#[derive(Clone, Debug)]
pub struct AllocationCreateDesc<'a> {
    /// Name of the allocation, for tracking and debugging purposes
    pub name: &'a str,
    /// Vulkan memory requirements for an allocation
    pub requirements: vk::MemoryRequirements,
    /// Location where the memory allocation should be stored
    pub location: MemoryLocation,
    /// If the resource is linear (buffer / linear texture) or a regular (tiled) texture.
    pub linear: bool,
    /// Determines how this allocation should be managed.
    pub allocation_scheme: AllocationScheme,
}

/// Wrapper type to only mark a raw pointer [`Send`] + [`Sync`] without having to
/// mark the entire [`Allocation`] as such, instead relying on the compiler to
/// auto-implement this or fail if fields are added that violate this constraint
#[derive(Clone, Copy, Debug)]
pub(crate) struct SendSyncPtr(std::ptr::NonNull<std::ffi::c_void>);
// Sending is fine because mapped_ptr does not change based on the thread we are in
unsafe impl Send for SendSyncPtr {}
// Sync is also okay because Sending &Allocation is safe: a mutable reference
// to the data in mapped_ptr is never exposed while `self` is immutably borrowed.
// In order to break safety guarantees, the user needs to `unsafe`ly dereference
// `mapped_ptr` themselves.
unsafe impl Sync for SendSyncPtr {}

pub struct AllocatorCreateDesc {
    pub instance: ash::Instance,
    pub device: ash::Device,
    pub physical_device: vk::PhysicalDevice,
    pub debug_settings: AllocatorDebugSettings,
    pub buffer_device_address: bool,
    pub allocation_sizes: AllocationSizes,
}

/// A piece of allocated memory.
///
/// Could be contained in its own individual underlying memory object or as a sub-region
/// of a larger allocation.
///
/// # Copying data into a CPU-mapped [`Allocation`]
///
/// You'll very likely want to copy data into CPU-mapped [`Allocation`]s in order to send that data to the GPU.
/// Doing this data transfer correctly without invoking undefined behavior can be quite fraught and non-obvious<sup>[\[1\]]</sup>.
///
/// To help you do this correctly, [`Allocation`] implements [`presser::Slab`], which means you can directly
/// pass it in to many of `presser`'s [helper functions] (for example, [`copy_from_slice_to_offset`]).
///
/// In most cases, this will work perfectly. However, note that if you try to use an [`Allocation`] as a
/// [`Slab`] and it is not valid to do so (if it is not CPU-mapped or if its `size > isize::MAX`),
/// you will cause a panic. If you aren't sure about these conditions, you may use [`Allocation::try_as_mapped_slab`].
///
/// ## Example
///
/// Say we've created an [`Allocation`] called `my_allocation`, which is CPU-mapped.
/// ```ignore
/// let mut my_allocation: Allocation = my_allocator.allocate(...)?;
/// ```
///
/// And we want to fill it with some data in the form of a `my_gpu_data: Vec<MyGpuVector>`, defined as such:
///
/// ```ignore
/// // note that this is size(12) but align(16), thus we have 4 padding bytes.
/// // this would mean a `&[MyGpuVector]` is invalid to cast as a `&[u8]`, but
/// // we can still use `presser` to copy it directly in a valid manner.
/// #[repr(C, align(16))]
/// #[derive(Clone, Copy)]
/// struct MyGpuVertex {
///     x: f32,
///     y: f32,
///     z: f32,
/// }
///
/// let my_gpu_data: Vec<MyGpuData> = make_vertex_data();
/// ```
///
/// Depending on how the data we're copying will be used, the Vulkan device may have a minimum
/// alignment requirement for that data:
///
/// ```ignore
/// let min_gpu_align = my_vulkan_device_specifications.min_alignment_thing;
/// ```
///
/// Finally, we can use [`presser::copy_from_slice_to_offset_with_align`] to perform the copy,
/// simply passing `&mut my_allocation` since [`Allocation`] implements [`Slab`].
///
/// ```ignore
/// let copy_record = presser::copy_from_slice_to_offset_with_align(
///     &my_gpu_data[..], // a slice containing all elements of my_gpu_data
///     &mut my_allocation, // our Allocation
///     0, // start as close to the beginning of the allocation as possible
///     min_gpu_align, // the minimum alignment we queried previously
/// )?;
/// ```
///
/// It's important to note that the data may not have actually been copied starting at the requested
/// `start_offset` (0 in the example above) depending on the alignment of the underlying allocation
/// as well as the alignment requirements of `MyGpuVector` and the `min_gpu_align` we passed in. Thus,
/// we can query the `copy_record` for the actual starting offset:
///
/// ```ignore
/// let actual_data_start_offset = copy_record.copy_start_offset;
/// ```
///
/// ## Safety
///
/// It is technically not fully safe to use an [`Allocation`] as a [`presser::Slab`] because we can't validate that the
/// GPU is not using the data in the buffer while `self` is borrowed. However, trying
/// to validate this statically is really hard and the community has basically decided that
/// requiring `unsafe` for functions like this creates too much "unsafe-noise", ultimately making it
/// harder to debug more insidious unsafety that is unrelated to GPU-CPU sync issues.
///
/// So, as would always be the case, you must ensure the GPU
/// is not using the data in `self` for the duration that you hold the returned [`MappedAllocationSlab`].
///
/// [`Slab`]: presser::Slab
/// [`copy_from_slice_to_offset`]: presser::copy_from_slice_to_offset
/// [helper functions]: presser#functions
/// [\[1\]]: presser#motivation
#[derive(Debug)]
pub struct Allocation {
    chunk_id: Option<std::num::NonZeroU64>,
    offset: u64,
    size: u64,
    memory_block_index: usize,
    memory_type_index: usize,
    device_memory: vk::DeviceMemory,
    mapped_ptr: Option<SendSyncPtr>,
    dedicated_allocation: bool,
    memory_properties: vk::MemoryPropertyFlags,
    name: Option<Box<str>>,
}

impl Allocation {
    /// Tries to borrow the CPU-mapped memory that backs this allocation as a [`presser::Slab`], which you can then
    /// use to safely copy data into the raw, potentially-uninitialized buffer.
    /// See [the documentation of Allocation][Allocation#example] for an example of this.
    ///
    /// Returns [`None`] if `self.mapped_ptr()` is `None`, or if `self.size()` is greater than `isize::MAX` because
    /// this could lead to undefined behavior.
    ///
    /// Note that [`Allocation`] implements [`Slab`] natively, so you can actually pass this allocation as a [`Slab`]
    /// directly. However, if `self` is not actually a valid [`Slab`] (this function would return `None` as described above),
    /// then trying to use it as a [`Slab`] will panic.
    ///
    /// # Safety
    ///
    /// See the note about safety in [the documentation of Allocation][Allocation#safety]
    ///
    /// [`Slab`]: presser::Slab
    // best to be explicit where the lifetime is coming from since we're doing unsafe things
    // and relying on an inferred lifetime type in the PhantomData below
    #[allow(clippy::needless_lifetimes)]
    pub fn try_as_mapped_slab<'a>(&'a mut self) -> Option<MappedAllocationSlab<'a>> {
        let mapped_ptr = self.mapped_ptr()?.cast().as_ptr();

        if self.size > isize::MAX as _ {
            return None;
        }

        // this will always succeed since size is <= isize::MAX which is < usize::MAX
        let size = self.size as usize;

        Some(MappedAllocationSlab {
            _borrowed_alloc: PhantomData,
            mapped_ptr,
            size,
        })
    }

    pub fn chunk_id(&self) -> Option<std::num::NonZeroU64> {
        self.chunk_id
    }

    ///Returns the [`vk::MemoryPropertyFlags`] of this allocation.
    pub fn memory_properties(&self) -> vk::MemoryPropertyFlags {
        self.memory_properties
    }

    /// Returns the [`vk::DeviceMemory`] object that is backing this allocation.
    /// This memory object can be shared with multiple other allocations and shouldn't be freed (or allocated from)
    /// without this library, because that will lead to undefined behavior.
    ///
    /// # Safety
    /// The result of this function can safely be used to pass into [`ash::Device::bind_buffer_memory()`],
    /// [`ash::Device::bind_image_memory()`] etc. It is exposed for this reason. Keep in mind to also
    /// pass [`Self::offset()`] along to those.
    pub unsafe fn memory(&self) -> vk::DeviceMemory {
        self.device_memory
    }

    /// Returns [`true`] if this allocation is using a dedicated underlying allocation.
    pub fn is_dedicated(&self) -> bool {
        self.dedicated_allocation
    }

    /// Returns the offset of the allocation on the [`vk::DeviceMemory`].
    /// When binding the memory to a buffer or image, this offset needs to be supplied as well.
    pub fn offset(&self) -> u64 {
        self.offset
    }

    /// Returns the size of the allocation
    pub fn size(&self) -> u64 {
        self.size
    }

    /// Returns a valid mapped pointer if the memory is host visible, otherwise it will return None.
    /// The pointer already points to the exact memory region of the suballocation, so no offset needs to be applied.
    pub fn mapped_ptr(&self) -> Option<std::ptr::NonNull<std::ffi::c_void>> {
        self.mapped_ptr.map(|SendSyncPtr(p)| p)
    }

    /// Returns a valid mapped slice if the memory is host visible, otherwise it will return None.
    /// The slice already references the exact memory region of the allocation, so no offset needs to be applied.
    pub fn mapped_slice(&self) -> Option<&[u8]> {
        self.mapped_ptr().map(|ptr| unsafe {
            std::slice::from_raw_parts(ptr.cast().as_ptr(), self.size as usize)
        })
    }

    /// Returns a valid mapped mutable slice if the memory is host visible, otherwise it will return None.
    /// The slice already references the exact memory region of the allocation, so no offset needs to be applied.
    pub fn mapped_slice_mut(&mut self) -> Option<&mut [u8]> {
        self.mapped_ptr().map(|ptr| unsafe {
            std::slice::from_raw_parts_mut(ptr.cast().as_ptr(), self.size as usize)
        })
    }

    pub fn is_null(&self) -> bool {
        self.chunk_id.is_none()
    }
}

impl Default for Allocation {
    fn default() -> Self {
        Self {
            chunk_id: None,
            offset: 0,
            size: 0,
            memory_block_index: !0,
            memory_type_index: !0,
            device_memory: vk::DeviceMemory::null(),
            mapped_ptr: None,
            memory_properties: vk::MemoryPropertyFlags::empty(),
            name: None,
            dedicated_allocation: false,
        }
    }
}

/// A wrapper struct over a borrowed [`Allocation`] that infallibly implements [`presser::Slab`].
///
/// This type should be acquired by calling [`Allocation::try_as_mapped_slab`].
pub struct MappedAllocationSlab<'a> {
    _borrowed_alloc: PhantomData<&'a mut Allocation>,
    mapped_ptr: *mut u8,
    size: usize,
}

// SAFETY: See the safety comment of Allocation::as_mapped_slab above.
unsafe impl<'a> presser::Slab for MappedAllocationSlab<'a> {
    fn base_ptr(&self) -> *const u8 {
        self.mapped_ptr
    }

    fn base_ptr_mut(&mut self) -> *mut u8 {
        self.mapped_ptr
    }

    fn size(&self) -> usize {
        self.size
    }
}

// SAFETY: See the safety comment of Allocation::as_mapped_slab above.
unsafe impl presser::Slab for Allocation {
    fn base_ptr(&self) -> *const u8 {
        self.mapped_ptr
            .expect("tried to use a non-mapped Allocation as a Slab")
            .0
            .as_ptr()
            .cast()
    }

    fn base_ptr_mut(&mut self) -> *mut u8 {
        self.mapped_ptr
            .expect("tried to use a non-mapped Allocation as a Slab")
            .0
            .as_ptr()
            .cast()
    }

    fn size(&self) -> usize {
        if self.size > isize::MAX as _ {
            panic!("tried to use an Allocation with size > isize::MAX as a Slab")
        }
        // this will always work if the above passed
        self.size as usize
    }
}

#[derive(Debug)]
pub(crate) struct MemoryBlock {
    pub(crate) device_memory: vk::DeviceMemory,
    pub(crate) size: u64,
    pub(crate) mapped_ptr: Option<SendSyncPtr>,
    pub(crate) sub_allocator: Box<dyn allocator::SubAllocator>,
    #[cfg(feature = "visualizer")]
    pub(crate) dedicated_allocation: bool,
}

impl MemoryBlock {
    fn new(
        device: &ash::Device,
        size: u64,
        mem_type_index: usize,
        mapped: bool,
        buffer_device_address: bool,
        allocation_scheme: AllocationScheme,
        requires_personal_block: bool,
    ) -> Result<Self> {
        let device_memory = {
            let alloc_info = vk::MemoryAllocateInfo::default()
                .allocation_size(size)
                .memory_type_index(mem_type_index as u32);

            let allocation_flags = vk::MemoryAllocateFlags::DEVICE_ADDRESS;
            let mut flags_info = vk::MemoryAllocateFlagsInfo::default().flags(allocation_flags);
            // TODO(manon): Test this based on if the device has this feature enabled or not
            let alloc_info = if buffer_device_address {
                alloc_info.push_next(&mut flags_info)
            } else {
                alloc_info
            };

            // Flag the memory as dedicated if required.
            let mut dedicated_memory_info = vk::MemoryDedicatedAllocateInfo::default();
            let alloc_info = match allocation_scheme {
                AllocationScheme::DedicatedBuffer(buffer) => {
                    dedicated_memory_info = dedicated_memory_info.buffer(buffer);
                    alloc_info.push_next(&mut dedicated_memory_info)
                }
                AllocationScheme::DedicatedImage(image) => {
                    dedicated_memory_info = dedicated_memory_info.image(image);
                    alloc_info.push_next(&mut dedicated_memory_info)
                }
                AllocationScheme::GpuAllocatorManaged => alloc_info,
            };

            unsafe { device.allocate_memory(&alloc_info, None) }.map_err(|e| match e {
                vk::Result::ERROR_OUT_OF_DEVICE_MEMORY => AllocationError::OutOfMemory,
                e => AllocationError::Internal(format!(
                    "Unexpected error in vkAllocateMemory: {:?}",
                    e
                )),
            })?
        };

        let mapped_ptr = mapped
            .then(|| {
                unsafe {
                    device.map_memory(
                        device_memory,
                        0,
                        vk::WHOLE_SIZE,
                        vk::MemoryMapFlags::empty(),
                    )
                }
                .map_err(|e| {
                    unsafe { device.free_memory(device_memory, None) };
                    AllocationError::FailedToMap(e.to_string())
                })
                .and_then(|p| {
                    std::ptr::NonNull::new(p).map(SendSyncPtr).ok_or_else(|| {
                        AllocationError::FailedToMap("Returned mapped pointer is null".to_owned())
                    })
                })
            })
            .transpose()?;

        let sub_allocator: Box<dyn allocator::SubAllocator> = if allocation_scheme
            != AllocationScheme::GpuAllocatorManaged
            || requires_personal_block
        {
            Box::new(allocator::DedicatedBlockAllocator::new(size))
        } else {
            Box::new(allocator::FreeListAllocator::new(size))
        };

        Ok(Self {
            device_memory,
            size,
            mapped_ptr,
            sub_allocator,
            #[cfg(feature = "visualizer")]
            dedicated_allocation: allocation_scheme != AllocationScheme::GpuAllocatorManaged,
        })
    }

    fn destroy(self, device: &ash::Device) {
        if self.mapped_ptr.is_some() {
            unsafe { device.unmap_memory(self.device_memory) };
        }

        unsafe { device.free_memory(self.device_memory, None) };
    }
}

#[derive(Debug)]
pub(crate) struct MemoryType {
    pub(crate) memory_blocks: Vec<Option<MemoryBlock>>,
    pub(crate) memory_properties: vk::MemoryPropertyFlags,
    pub(crate) memory_type_index: usize,
    pub(crate) heap_index: usize,
    pub(crate) mappable: bool,
    pub(crate) active_general_blocks: usize,
    pub(crate) buffer_device_address: bool,
}

impl MemoryType {
    fn allocate(
        &mut self,
        device: &ash::Device,
        desc: &AllocationCreateDesc<'_>,
        granularity: u64,
        backtrace: Arc<Backtrace>,
        allocation_sizes: &AllocationSizes,
    ) -> Result<Allocation> {
        let allocation_type = if desc.linear {
            allocator::AllocationType::Linear
        } else {
            allocator::AllocationType::NonLinear
        };

        let memblock_size = if self
            .memory_properties
            .contains(vk::MemoryPropertyFlags::HOST_VISIBLE)
        {
            allocation_sizes.host_memblock_size
        } else {
            allocation_sizes.device_memblock_size
        };

        let size = desc.requirements.size;
        let alignment = desc.requirements.alignment;

        let dedicated_allocation = desc.allocation_scheme != AllocationScheme::GpuAllocatorManaged;
        let requires_personal_block = size > memblock_size;

        // Create a dedicated block for large memory allocations or allocations that require dedicated memory allocations.
        if dedicated_allocation || requires_personal_block {
            let mem_block = MemoryBlock::new(
                device,
                size,
                self.memory_type_index,
                self.mappable,
                self.buffer_device_address,
                desc.allocation_scheme,
                requires_personal_block,
            )?;

            let mut block_index = None;
            for (i, block) in self.memory_blocks.iter().enumerate() {
                if block.is_none() {
                    block_index = Some(i);
                    break;
                }
            }

            let block_index = match block_index {
                Some(i) => {
                    self.memory_blocks[i].replace(mem_block);
                    i
                }
                None => {
                    self.memory_blocks.push(Some(mem_block));
                    self.memory_blocks.len() - 1
                }
            };

            let mem_block = self.memory_blocks[block_index]
                .as_mut()
                .ok_or_else(|| AllocationError::Internal("Memory block must be Some".into()))?;

            let (offset, chunk_id) = mem_block.sub_allocator.allocate(
                size,
                alignment,
                allocation_type,
                granularity,
                desc.name,
                backtrace,
            )?;

            return Ok(Allocation {
                chunk_id: Some(chunk_id),
                offset,
                size,
                memory_block_index: block_index,
                memory_type_index: self.memory_type_index,
                device_memory: mem_block.device_memory,
                mapped_ptr: mem_block.mapped_ptr,
                memory_properties: self.memory_properties,
                name: Some(desc.name.into()),
                dedicated_allocation,
            });
        }

        let mut empty_block_index = None;
        for (mem_block_i, mem_block) in self.memory_blocks.iter_mut().enumerate().rev() {
            if let Some(mem_block) = mem_block {
                let allocation = mem_block.sub_allocator.allocate(
                    size,
                    alignment,
                    allocation_type,
                    granularity,
                    desc.name,
                    backtrace.clone(),
                );

                match allocation {
                    Ok((offset, chunk_id)) => {
                        let mapped_ptr = if let Some(SendSyncPtr(mapped_ptr)) = mem_block.mapped_ptr
                        {
                            let offset_ptr = unsafe { mapped_ptr.as_ptr().add(offset as usize) };
                            std::ptr::NonNull::new(offset_ptr).map(SendSyncPtr)
                        } else {
                            None
                        };
                        return Ok(Allocation {
                            chunk_id: Some(chunk_id),
                            offset,
                            size,
                            memory_block_index: mem_block_i,
                            memory_type_index: self.memory_type_index,
                            device_memory: mem_block.device_memory,
                            memory_properties: self.memory_properties,
                            mapped_ptr,
                            dedicated_allocation: false,
                            name: Some(desc.name.into()),
                        });
                    }
                    Err(err) => match err {
                        AllocationError::OutOfMemory => {} // Block is full, continue search.
                        _ => return Err(err),              // Unhandled error, return.
                    },
                }
            } else if empty_block_index.is_none() {
                empty_block_index = Some(mem_block_i);
            }
        }

        let new_memory_block = MemoryBlock::new(
            device,
            memblock_size,
            self.memory_type_index,
            self.mappable,
            self.buffer_device_address,
            desc.allocation_scheme,
            false,
        )?;

        let new_block_index = if let Some(block_index) = empty_block_index {
            self.memory_blocks[block_index] = Some(new_memory_block);
            block_index
        } else {
            self.memory_blocks.push(Some(new_memory_block));
            self.memory_blocks.len() - 1
        };

        self.active_general_blocks += 1;

        let mem_block = self.memory_blocks[new_block_index]
            .as_mut()
            .ok_or_else(|| AllocationError::Internal("Memory block must be Some".into()))?;
        let allocation = mem_block.sub_allocator.allocate(
            size,
            alignment,
            allocation_type,
            granularity,
            desc.name,
            backtrace,
        );
        let (offset, chunk_id) = match allocation {
            Ok(value) => value,
            Err(err) => match err {
                AllocationError::OutOfMemory => {
                    return Err(AllocationError::Internal(
                        "Allocation that must succeed failed. This is a bug in the allocator."
                            .into(),
                    ))
                }
                _ => return Err(err),
            },
        };

        let mapped_ptr = if let Some(SendSyncPtr(mapped_ptr)) = mem_block.mapped_ptr {
            let offset_ptr = unsafe { mapped_ptr.as_ptr().add(offset as usize) };
            std::ptr::NonNull::new(offset_ptr).map(SendSyncPtr)
        } else {
            None
        };

        Ok(Allocation {
            chunk_id: Some(chunk_id),
            offset,
            size,
            memory_block_index: new_block_index,
            memory_type_index: self.memory_type_index,
            device_memory: mem_block.device_memory,
            mapped_ptr,
            memory_properties: self.memory_properties,
            name: Some(desc.name.into()),
            dedicated_allocation: false,
        })
    }

    #[allow(clippy::needless_pass_by_value)]
    fn free(&mut self, allocation: Allocation, device: &ash::Device) -> Result<()> {
        let block_idx = allocation.memory_block_index;

        let mem_block = self.memory_blocks[block_idx]
            .as_mut()
            .ok_or_else(|| AllocationError::Internal("Memory block must be Some.".into()))?;

        mem_block.sub_allocator.free(allocation.chunk_id)?;

        if mem_block.sub_allocator.is_empty() {
            if mem_block.sub_allocator.supports_general_allocations() {
                if self.active_general_blocks > 1 {
                    let block = self.memory_blocks[block_idx].take();
                    let block = block.ok_or_else(|| {
                        AllocationError::Internal("Memory block must be Some.".into())
                    })?;
                    block.destroy(device);

                    self.active_general_blocks -= 1;
                }
            } else {
                let block = self.memory_blocks[block_idx].take();
                let block = block.ok_or_else(|| {
                    AllocationError::Internal("Memory block must be Some.".into())
                })?;
                block.destroy(device);
            }
        }

        Ok(())
    }
}

pub struct Allocator {
    pub(crate) memory_types: Vec<MemoryType>,
    pub(crate) memory_heaps: Vec<vk::MemoryHeap>,
    device: ash::Device,
    pub(crate) buffer_image_granularity: u64,
    pub(crate) debug_settings: AllocatorDebugSettings,
    allocation_sizes: AllocationSizes,
}

impl fmt::Debug for Allocator {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        self.generate_report().fmt(f)
    }
}

impl Allocator {
    pub fn new(desc: &AllocatorCreateDesc) -> Result<Self> {
        if desc.physical_device == vk::PhysicalDevice::null() {
            return Err(AllocationError::InvalidAllocatorCreateDesc(
                "AllocatorCreateDesc field `physical_device` is null.".into(),
            ));
        }

        let mem_props = unsafe {
            desc.instance
                .get_physical_device_memory_properties(desc.physical_device)
        };

        let memory_types = &mem_props.memory_types_as_slice();
        let memory_heaps = mem_props.memory_heaps_as_slice().to_vec();

        if desc.debug_settings.log_memory_information {
            debug!("memory type count: {}", mem_props.memory_type_count);
            debug!("memory heap count: {}", mem_props.memory_heap_count);

            for (i, mem_type) in memory_types.iter().enumerate() {
                let flags = mem_type.property_flags;
                debug!(
                    "memory type[{}]: prop flags: 0x{:x}, heap[{}]",
                    i,
                    flags.as_raw(),
                    mem_type.heap_index,
                );
            }
            for (i, heap) in memory_heaps.iter().enumerate() {
                debug!(
                    "heap[{}] flags: 0x{:x}, size: {} MiB",
                    i,
                    heap.flags.as_raw(),
                    heap.size / (1024 * 1024)
                );
            }
        }

        let memory_types = memory_types
            .iter()
            .enumerate()
            .map(|(i, mem_type)| MemoryType {
                memory_blocks: Vec::default(),
                memory_properties: mem_type.property_flags,
                memory_type_index: i,
                heap_index: mem_type.heap_index as usize,
                mappable: mem_type
                    .property_flags
                    .contains(vk::MemoryPropertyFlags::HOST_VISIBLE),
                active_general_blocks: 0,
                buffer_device_address: desc.buffer_device_address,
            })
            .collect::<Vec<_>>();

        let physical_device_properties = unsafe {
            desc.instance
                .get_physical_device_properties(desc.physical_device)
        };

        let granularity = physical_device_properties.limits.buffer_image_granularity;

        Ok(Self {
            memory_types,
            memory_heaps,
            device: desc.device.clone(),
            buffer_image_granularity: granularity,
            debug_settings: desc.debug_settings,
            allocation_sizes: AllocationSizes::default(),
        })
    }

    pub fn allocate(&mut self, desc: &AllocationCreateDesc<'_>) -> Result<Allocation> {
        let size = desc.requirements.size;
        let alignment = desc.requirements.alignment;

        let backtrace = Arc::new(if self.debug_settings.store_stack_traces {
            Backtrace::force_capture()
        } else {
            Backtrace::disabled()
        });

        if self.debug_settings.log_allocations {
            debug!(
                "Allocating `{}` of {} bytes with an alignment of {}.",
                &desc.name, size, alignment
            );
            if self.debug_settings.log_stack_traces {
                let backtrace = Backtrace::force_capture();
                debug!("Allocation stack trace: {}", backtrace);
            }
        }

        if size == 0 || !alignment.is_power_of_two() {
            return Err(AllocationError::InvalidAllocationCreateDesc);
        }

        let mem_loc_preferred_bits = match desc.location {
            MemoryLocation::GpuOnly => vk::MemoryPropertyFlags::DEVICE_LOCAL,
            MemoryLocation::CpuToGpu => {
                vk::MemoryPropertyFlags::HOST_VISIBLE
                    | vk::MemoryPropertyFlags::HOST_COHERENT
                    | vk::MemoryPropertyFlags::DEVICE_LOCAL
            }
            MemoryLocation::GpuToCpu => {
                vk::MemoryPropertyFlags::HOST_VISIBLE
                    | vk::MemoryPropertyFlags::HOST_COHERENT
                    | vk::MemoryPropertyFlags::HOST_CACHED
            }
            MemoryLocation::Unknown => vk::MemoryPropertyFlags::empty(),
        };
        let mut memory_type_index_opt =
            self.find_memorytype_index(&desc.requirements, mem_loc_preferred_bits);

        if memory_type_index_opt.is_none() {
            let mem_loc_required_bits = match desc.location {
                MemoryLocation::GpuOnly => vk::MemoryPropertyFlags::DEVICE_LOCAL,
                MemoryLocation::CpuToGpu | MemoryLocation::GpuToCpu => {
                    vk::MemoryPropertyFlags::HOST_VISIBLE | vk::MemoryPropertyFlags::HOST_COHERENT
                }
                MemoryLocation::Unknown => vk::MemoryPropertyFlags::empty(),
            };

            memory_type_index_opt =
                self.find_memorytype_index(&desc.requirements, mem_loc_required_bits);
        }

        let memory_type_index = match memory_type_index_opt {
            Some(x) => x as usize,
            None => return Err(AllocationError::NoCompatibleMemoryTypeFound),
        };

        //Do not try to create a block if the heap is smaller than the required size (avoids validation warnings).
        let memory_type = &mut self.memory_types[memory_type_index];
        let allocation = if size > self.memory_heaps[memory_type.heap_index].size {
            Err(AllocationError::OutOfMemory)
        } else {
            memory_type.allocate(
                &self.device,
                desc,
                self.buffer_image_granularity,
                backtrace.clone(),
                &self.allocation_sizes,
            )
        };

        if desc.location == MemoryLocation::CpuToGpu {
            if allocation.is_err() {
                let mem_loc_preferred_bits =
                    vk::MemoryPropertyFlags::HOST_VISIBLE | vk::MemoryPropertyFlags::HOST_COHERENT;

                let memory_type_index_opt =
                    self.find_memorytype_index(&desc.requirements, mem_loc_preferred_bits);

                let memory_type_index = match memory_type_index_opt {
                    Some(x) => x as usize,
                    None => return Err(AllocationError::NoCompatibleMemoryTypeFound),
                };

                self.memory_types[memory_type_index].allocate(
                    &self.device,
                    desc,
                    self.buffer_image_granularity,
                    backtrace,
                    &self.allocation_sizes,
                )
            } else {
                allocation
            }
        } else {
            allocation
        }
    }

    pub fn free(&mut self, allocation: Allocation) -> Result<()> {
        if self.debug_settings.log_frees {
            let name = allocation.name.as_deref().unwrap_or("<null>");
            debug!("Freeing `{}`.", name);
            if self.debug_settings.log_stack_traces {
                let backtrace = Backtrace::force_capture();
                debug!("Free stack trace: {}", backtrace);
            }
        }

        if allocation.is_null() {
            return Ok(());
        }

        self.memory_types[allocation.memory_type_index].free(allocation, &self.device)?;

        Ok(())
    }

    pub fn rename_allocation(&mut self, allocation: &mut Allocation, name: &str) -> Result<()> {
        allocation.name = Some(name.into());

        if allocation.is_null() {
            return Ok(());
        }

        let mem_type = &mut self.memory_types[allocation.memory_type_index];
        let mem_block = mem_type.memory_blocks[allocation.memory_block_index]
            .as_mut()
            .ok_or_else(|| AllocationError::Internal("Memory block must be Some.".into()))?;

        mem_block
            .sub_allocator
            .rename_allocation(allocation.chunk_id, name)?;

        Ok(())
    }

    pub fn report_memory_leaks(&self, log_level: Level) {
        for (mem_type_i, mem_type) in self.memory_types.iter().enumerate() {
            for (block_i, mem_block) in mem_type.memory_blocks.iter().enumerate() {
                if let Some(mem_block) = mem_block {
                    mem_block
                        .sub_allocator
                        .report_memory_leaks(log_level, mem_type_i, block_i);
                }
            }
        }
    }

    fn find_memorytype_index(
        &self,
        memory_req: &vk::MemoryRequirements,
        flags: vk::MemoryPropertyFlags,
    ) -> Option<u32> {
        self.memory_types
            .iter()
            .find(|memory_type| {
                (1 << memory_type.memory_type_index) & memory_req.memory_type_bits != 0
                    && memory_type.memory_properties.contains(flags)
            })
            .map(|memory_type| memory_type.memory_type_index as _)
    }

    pub fn generate_report(&self) -> AllocatorReport {
        let mut allocations = vec![];
        let mut blocks = vec![];
        let mut total_reserved_bytes = 0;

        for memory_type in &self.memory_types {
            for block in memory_type.memory_blocks.iter().flatten() {
                total_reserved_bytes += block.size;
                let first_allocation = allocations.len();
                allocations.extend(block.sub_allocator.report_allocations());
                blocks.push(MemoryBlockReport {
                    size: block.size,
                    allocations: first_allocation..allocations.len(),
                });
            }
        }

        let total_allocated_bytes = allocations.iter().map(|report| report.size).sum();

        AllocatorReport {
            allocations,
            blocks,
            total_allocated_bytes,
            total_reserved_bytes,
        }
    }
}

impl Drop for Allocator {
    fn drop(&mut self) {
        if self.debug_settings.log_leaks_on_shutdown {
            self.report_memory_leaks(Level::Warn);
        }

        // Free all remaining memory blocks
        for mem_type in self.memory_types.iter_mut() {
            for mem_block in mem_type.memory_blocks.iter_mut() {
                let block = mem_block.take();
                if let Some(block) = block {
                    block.destroy(&self.device);
                }
            }
        }
    }
}

[ Dauer der Verarbeitung: 0.46 Sekunden  ]