Creating a memory allocator

Before you can create buffers in memory, you have to request (allocate) some memory first. It turns out allocating memory efficiently and dynamically is challenging. Luckily, in vulkano, we have several kinds of memory allocators that we can pick from depending on our use case. Since we don't have any special needs, we can use the StandardMemoryAllocator with default settings, that kind of allocator is general-purpose and will be your go-to option in most cases.

#![allow(unused)]
fn main() {
use vulkano::memory::allocator::StandardMemoryAllocator;

let memory_allocator = Arc::new(StandardMemoryAllocator::new_default(device.clone()));
}

Since device is actually an Arc<Device>, the call to .clone() only clones the Arc which isn't expensive. You should get used to passing the device as a parameter, as you will need to do so for most of the Vulkan objects that you create. We encapsulate the memory allocator with an atomic reference counter since Buffer::from_data requires an Arc.

Creating a buffer

When using Vulkan, you will very often need the GPU to read or write data in memory. In fact there isn't much point in using the GPU otherwise, as there is nothing you can do with the results of its work except write them to memory.

In order for the GPU to be able to access some data (either for reading, writing or both), we first need to create a buffer object and put the data in it.

Memory type filter

A Vulkan implementation might (and most often does) have multiple memory types, each being best suited to certain tasks. There are many possible arrangements of memory types a Vulkan implementation might have, and picking the right one is important to ensure most optimal performance.

When allocating memory for a buffer in vulkano, you have to provide a memory type filter, which tells the memory allocator which memory types it should prefer, and which ones it should avoid, when picking the right one. For example, if you want to continuously upload data to the GPU, you should use MemoryTypeFilter::PREFER_DEVICE | MemoryTypeFilter::HOST_SEQUENTIAL_WRITE; on the other hand, if you have some data that will largely remain visible only to the GPU, using MemoryTypeFilter::PREFER_DEVICE brings increased performance at the cost of more complicated data access from the CPU. For staging buffers, you should use MemoryTypeFilter::PREFER_HOST | MemoryTypeFilter::HOST_SEQUENTIAL_WRITE.

The simplest way to create a buffer is to create it in CPU-accessible memory, by using MemoryTypeFilter::HOST_SEQUENTIAL_WRITE or MemoryTypeFilter::HOST_RANDOM_ACCESS, together with one of the other filters depending of whether host or device-local memory is preferred.

#![allow(unused)]
fn main() {
use vulkano::buffer::{Buffer, BufferCreateInfo, BufferUsage};
use vulkano::memory::allocator::{AllocationCreateInfo, MemoryTypeFilter};

let data: i32 = 12;
let buffer = Buffer::from_data(
    memory_allocator.clone(),
    BufferCreateInfo {
        usage: BufferUsage::UNIFORM_BUFFER,
        ..Default::default()
    },
    AllocationCreateInfo {
        memory_type_filter: MemoryTypeFilter::PREFER_DEVICE
            | MemoryTypeFilter::HOST_SEQUENTIAL_WRITE,
        ..Default::default()
    },
    data,
)
.expect("failed to create buffer");
}

We have to indicate several things when creating the buffer. The first parameter is an Arc of the memory allocator to use.

The second parameter is the create info for the buffer. The only field that you have to override is the usage for which we are creating the buffer for, which can help the implementation perform some optimizations. Trying to use a buffer in a way that wasn't indicated when creating it will result in an error. For the sake of the example, we just create a buffer that supports being used as a uniform buffer.

The third parameter is the create info for the allocation. The field of interest is the memory type filter. When creating a CPU-accessible buffer, you will most commonly use MemoryTypeFilter::PREFER_HOST | MemoryTypeFilter::HOST_SEQUENTIAL_WRITE, but in cases where the application is writing data through this buffer continuously, using MemoryTypeFilter::PREFER_HOST | MemoryTypeFilter::HOST_RANDOM_ACCESS is preferred as it may yield some performance gain. Using MemoryTypeFilter::PREFER_DEVICE will get you a buffer that is inaccessible from the CPU when such a memory type exists. Therefore, you can't use this memory usage together with Buffer::from_data directly, and instead have to create a staging buffer whose content is then copied to the device-local buffer.

Finally, the fourth parameter is the content of the buffer. Here we create a buffer that contains a single integer with the value 12.

Note: In a real application you shouldn't create buffers with only 4 bytes of data. Although buffers aren't expensive, you should try to group as much related data as you can in the same buffer.

From_data and from_iter

In the example above we create a buffer that contains the value 12, which is of type i32. but you can put any type you want in a buffer, there is no restriction. In order to give our arbitrary types a representation that can be used in a generic way, we use the crate bytemuck and its "plain old data" trait, AnyBitPattern. Thus, any crate which exposes types with bytemuck support can be used in a buffer. You can also derive AnyBitPattern for you own types, or use the vulkano-provided BufferContents derive macro:

#![allow(unused)]
fn main() {
use vulkano::buffer::BufferContents;

#[derive(BufferContents)]
#[repr(C)]
struct MyStruct {
    a: u32,
    b: u32,
}

let data = MyStruct { a: 5, b: 69 };

let buffer = Buffer::from_data(
    memory_allocator.clone(),
    BufferCreateInfo {
        usage: BufferUsage::UNIFORM_BUFFER,
        ..Default::default()
    },
    AllocationCreateInfo {
        memory_type_filter: MemoryTypeFilter::PREFER_DEVICE
            | MemoryTypeFilter::HOST_SEQUENTIAL_WRITE,
        ..Default::default()
    },
    data,
)
.unwrap();
}

While it is sometimes useful to use a buffer that contains a single struct, in practice it is very common to put an array of values inside a buffer. You can, for example, put an array of fifty i32s in a buffer with the Buffer::from_data function.

However, in practice it is also very common to not know the size of the array at compile-time. In order to handle this, Buffer provides a from_iter constructor that takes an iterator to the data as the last parameter, instead of the data itself.

In the example below, we create a buffer that contains the value 5 of type u8, 128 times. The type of the content of the buffer is [u8], which, in Rust, represents an array of u8s whose size is only known at runtime.

#![allow(unused)]
fn main() {
let iter = (0..128).map(|_| 5u8);
let buffer = Buffer::from_iter(
    memory_allocator.clone(),
    BufferCreateInfo {
        usage: BufferUsage::UNIFORM_BUFFER,
        ..Default::default()
    },
    AllocationCreateInfo {
        memory_type_filter: MemoryTypeFilter::PREFER_DEVICE
            | MemoryTypeFilter::HOST_SEQUENTIAL_WRITE,
        ..Default::default()
    },
    iter,
)
.unwrap();
}

Reading and writing the contents of a buffer

Once a CPU-accessible buffer is created, you can access its content with the read() or write() methods. Using read() will grant you shared access to the content of the buffer, and using write() will grant you exclusive access. This is similar to using a RwLock.

For example if buffer contains a MyStruct (see above):

#![allow(unused)]
fn main() {
let mut content = buffer.write().unwrap();
// `content` implements `DerefMut` whose target is of type `MyStruct` (the content of the buffer)
content.a *= 2;
content.b = 9;
}

Alternatively, suppose that the content of buffer is of type [u8] (like with the example that uses from_iter):

#![allow(unused)]
fn main() {
let mut content = buffer.write().unwrap();
// this time `content` derefs to `[u8]`
content[12] = 83;
content[7] = 3;
}

Just like the constructors, keep in mind that being able to read/write the content of the buffer like this is specific to buffer allocated in CPU-accessible memory. Device-local buffers cannot be accessed in this way.

Next: Example operation

Vulkano

Creating a memory allocator

Creating a buffer

Memory type filter

From_data and from_iter

Reading and writing the contents of a buffer