Vulkan-Hpp/README.md
asuessenbach db58507063 Introduce VULKAN_HPP_NO_STRUCT_CONSTRUCTORS so support designated initializers.
You have to define VULKAN_HPP_NO_STRUCT_CONSTRUCTORS to remove all the struct constructors from vulkan.hpp, so that you can use designated initializers on them.
Of course, makes only sense with C++20.
2020-06-24 11:28:43 +02:00

26 KiB
Raw Blame History

Vulkan-Hpp: C++ Bindings for Vulkan

The goal of the Vulkan-Hpp is to provide header only C++ bindings for the Vulkan C API to improve the developers Vulkan experience without introducing CPU runtime cost. It adds features like type safety for enums and bitfields, STL container support, exceptions and simple enumerations.

Platform Build Status
Linux Build Status

Getting Started

Vulkan-Hpp is part of the LunarG Vulkan SDK since version 1.0.24. Just #include <vulkan/vulkan.hpp> and you're ready to use the C++ bindings. If you're using a Vulkan version not yet supported by the Vulkan SDK you can find the latest version of the header here.

Minimum Requirements

Vulkan-Hpp requires a C++11 capable compiler to compile. The following compilers are known to work:

  • Visual Studio >=2015
  • GCC >= 4.8.2 (earlier version might work, but are untested)
  • Clang >= 3.3

Optional Features

Formatting

If the program clang-format is found by CMake, the define CLANG_FORMAT_EXECUTABLE is set accordingly. In that case, the generated vulkan.hpp is formatted using the .clang-format file located in the root directory of this project. Otherwise it's formatted as hard-coded in the generator.

Usage

namespace vk

To avoid name collisions with the Vulkan C API the C++ bindings reside in the vk namespace. The following rules apply to the new naming

  • All functions, enums, handles, and structs have the Vk prefix removed. In addition to this the first letter of functions is lower case.
    • vkCreateImage can be accessed as vk::createImage
    • VkImageTiling can be accessed as vk::ImageTiling
    • VkImageCreateInfo can be accessed as vk::ImageCreateInfo
  • Enums are mapped to scoped enums to provide compile time type safety. The names have been changed to 'e' + CamelCase with the VK_ prefix and type infix removed. In case the enum type is an extension the extension suffix has been removed from the enum values.

In all other cases the extension suffix has not been removed.

  • VK_IMAGETYPE_2D is now vk::ImageType::e2D.
  • VK_COLOR_SPACE_SRGB_NONLINEAR_KHR is now vk::ColorSpaceKHR::eSrgbNonlinear.
  • VK_STRUCTURE_TYPE_PRESENT_INFO_KHR is now vk::StructureType::ePresentInfoKHR.
  • Flag bits are handled like scoped enums with the addition that the _BIT suffix has also been removed.

In some cases it might be necessary to move Vulkan-Hpp to a custom namespace. This can be achieved by defining VULKAN_HPP_NAMESPACE before including Vulkan-Hpp.

Handles

Vulkan-Hpp declares a class for all handles to ensure full type safety and to add support for member functions on handles. A member function has been added to a handle class for each function which accepts the corresponding handle as first parameter. Instead of vkBindBufferMemory(device, ...) one can write device.bindBufferMemory(...) or vk::bindBufferMemory(device, ...).

C/C++ Interop for Handles

On 64-bit platforms Vulkan-Hpp supports implicit conversions between C++ Vulkan handles and C Vulkan handles. On 32-bit platforms all non-dispatchable handles are defined as uint64_t, thus preventing type-conversion checks at compile time which would catch assignments between incompatible handle types.. Due to that Vulkan-Hpp does not enable implicit conversion for 32-bit platforms by default and it is recommended to use a static_cast for the conversion like this: VkDevice = static_cast<VkDevice>(cppDevice) to prevent converting some arbitrary int to a handle or vice versa by accident. If you're developing your code on a 64-bit platform, but want compile your code for a 32-bit platform without adding the explicit casts you can define VULKAN_HPP_TYPESAFE_CONVERSION to 1 in your build system or before including vulkan.hpp. On 64-bit platforms this define is set to 1 by default and can be set to 0 to disable implicit conversions.

Flags

The scoped enum feature adds type safety to the flags, but also prevents using the flag bits as input for bitwise operations like & and |.

As solution Vulkan-Hpp provides a template class vk::Flags which brings the standard operations like &=, |=, & and | to our scoped enums. Except for the initialization with 0 this class behaves exactly like a normal bitmask with the improvement that it is impossible to set bits not specified by the corresponding enum by accident. Here are a few examples for the bitmask handling:

vk::ImageUsage iu1; // initialize a bitmask with no bit set
vk::ImageUsage iu2 = {}; // initialize a bitmask with no bit set
vk::ImageUsage iu3 = vk::ImageUsage::eColorAttachment; // initialize with a single value
vk::ImageUsage iu4 = vk::ImageUsage::eColorAttachment | vk::ImageUsage::eStorage; // or two bits to get a bitmask
PipelineShaderStageCreateInfo ci( {} /* pass a flag without any bits set */, ...);

CreateInfo structs

When constructing a handle in Vulkan one usually has to create some CreateInfo struct which describes the new handle. This can result in quite lengthy code as can be seen in the following Vulkan C example:

VkImageCreateInfo ci;
ci.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO;
ci.pNext = nullptr;
ci.flags = ...some flags...;
ci.imageType = VK_IMAGE_TYPE_2D;
ci.format = VK_FORMAT_R8G8B8A8_UNORM;
ci.extent = VkExtent3D { width, height, 1 };
ci.mipLevels = 1;
ci.arrayLayers = 1;
ci.samples = VK_SAMPLE_COUNT_1_BIT;
ci.tiling = VK_IMAGE_TILING_OPTIMAL;
ci.usage = VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT;
ci.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
ci.queueFamilyIndexCount = 0;
ci.pQueueFamilyIndices = 0;
ci.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
vkCreateImage(device, &ci, allocator, &image));

There are two typical issues Vulkan developers encounter when filling out a CreateInfo struct field by field

  • One or more fields are left uninitialized.
  • sType is incorrect.

Especially the first one is hard to detect.

Vulkan-Hpp provides constructors for all CreateInfo objects which accept one parameter for each member variable. This way the compiler throws a compiler error if a value has been forgotten. In addition to this sType is automatically filled with the correct value and pNext set to a nullptr by default. Here's how the same code looks with a constructor:

vk::ImageCreateInfo ci({}, vk::ImageType::e2D, vk::format::eR8G8B8A8Unorm,
                       { width, height, 1 },
                       1, 1, vk::SampleCount::e1,
                       vk::ImageTiling::eOptimal, vk::ImageUsage:eColorAttachment,
                       vk::SharingMode::eExclusive, 0, 0, vk::Imagelayout::eUndefined);

With constructors for CreateInfo structures one can also pass temporaries to Vulkan functions like this:

vk::Image image = device.createImage({{}, vk::ImageType::e2D, vk::format::eR8G8B8A8Unorm,
                                     { width, height, 1 },
                                     1, 1, vk::SampleCount::e1,
                                     vk::ImageTiling::eOptimal, vk::ImageUsage:eColorAttachment,
                                     vk::SharingMode::eExclusive, 0, 0, vk::Imagelayout::eUndefined});

Designated Initializers

Beginning with C++20, C++ supports designated initializers. As that feature requires to not have any user-declared or inherited constructors, you have to #define VULKAN_HPP_NO_STRUCT_CONSTRUCTORS, which removes all the structure constructors from vulkan.hpp. Instead you can then use aggregate initialization. The first few vk-lines in your source might then look like

// initialize the vk::ApplicationInfo structure
vk::ApplicationInfo applicationInfo{ .pApplicationName   = AppName,
                                     .applicationVersion = 1,
                                     .pEngineName        = EngineName,
                                     .engineVersion      = 1,
                                     .apiVersion         = VK_API_VERSION_1_1 };
        
// initialize the vk::InstanceCreateInfo
vk::InstanceCreateInfo instanceCreateInfo{ .pApplicationInfo = & applicationInfo };

instead of

// initialize the vk::ApplicationInfo structure
vk::ApplicationInfo applicationInfo( AppName, 1, EngineName, 1, VK_API_VERSION_1_1 );
        
// initialize the vk::InstanceCreateInfo
vk::InstanceCreateInfo instanceCreateInfo( {}, &applicationInfo );

Note, that the designator order needs to match the declaration order. Note as well, that now you can explicitly set the sType member of vk-structures. This is neither neccessary (as they are correctly initialized by default) nor recommended.

Passing Arrays to Functions using ArrayProxy

The Vulkan API has several places where which require (count,pointer) as two function arguments and C++ has a few containers which map perfectly to this pair. To simplify development the Vulkan-Hpp bindings have replaced those argument pairs with the ArrayProxy template class which accepts empty arrays and a single value as well as STL containers std::initializer_list, std::array and std::vector as argument for construction. This way a single generated Vulkan version can accept a variety of inputs without having the combinatoric explosion which would occur when creating a function for each container type.

Here are some code samples on how to use the ArrayProxy:

vk::CommandBuffer c;

// pass an empty array
c.setScissor(0, nullptr);

// pass a single value. Value is passed as reference
vk::Rect2D scissorRect = { {0, 0}, {640, 480} };
c.setScissor(0, scissorRect);

// pass a temporary value.
c.setScissor(0, { { 0, 0 },{ 640, 480 } });

// generate a std::initializer_list using two rectangles from the stack. This might generate a copy of the rectangles.
vk::Rect2D scissorRect1 = { { 0, 0 },{ 320, 240 } };
vk::Rect2D scissorRect2 = { { 320, 240 },{ 320, 240 } };
c.setScissor(0, { scissorRect, scissorRect2 });

// construct a std::initializer_list using two temporary rectangles.
c.setScissor(0, { { {   0,   0 },{ 320, 240 } },
                { { 320, 240 },{ 320, 240 } }
}
);

// pass a std::array
std::array<vk::Rect2D, 2> arr{ scissorRect1, scissorRect2 };
c.setScissor(0, arr);

// pass a std::vector of dynamic size
std::vector<vk::Rect2D> vec;
vec.push_back(scissorRect1);
vec.push_back(scissorRect2);
c.setScissor(0, vec);

Passing Structs to Functions

Vulkan-Hpp generates references for pointers to structs. This conversion allows passing temporary structs to functions which can result in shorter code. In case the input is optional and thus accepting a null pointer the parameter type will be a vk::Optional<T> const& type. This type accepts either a reference to T or nullptr as input and thus allows optional temporary structs.

// C
ImageSubResource subResource;
subResource.aspectMask = 0;
subResource.mipLevel = 0;
subResource.arrayLayer = 0;
vkSubresourceLayout layout = vkGetImageSubResourceLayout(image, subresource);

// C++
auto layout = device.getImageSubResourceLayout(image, { {} /* flags*/, 0 /* miplevel */, 0 /* layout */ });

Structure Pointer Chains

Vulkan allows chaining of structures through the pNext pointer. Vulkan-Hpp has a variadic template class which allows constructing of such structure chains with minimal efforts. In addition to this it checks at compile time if the spec allows the construction of such a pNext chain.

// This will compile successfully.
vk::StructureChain<vk::MemoryAllocateInfo, vk::ImportMemoryFdInfoKHR> c;
vk::MemoryAllocateInfo &allocInfo = c.get<vk::MemoryAllocateInfo>();
vk::ImportMemoryFdInfoKHR &fdInfo = c.get<vk::ImportMemoryFdInfoKHR>();

// This will fail compilation since it's not valid according to the spec.
vk::StructureChain<vk::MemoryAllocateInfo, vk::MemoryDedicatedRequirementsKHR> c;
vk::MemoryAllocateInfo &allocInfo = c.get<vk::MemoryAllocateInfo>();
vk::ImportMemoryFdInfoKHR &fdInfo = c.get<vk::ImportMemoryFdInfoKHR>();

Vulkan-Hpp provides a constructor for these chains similar to the CreateInfo objects which accepts a list of all structures part of the chain. The pNext field is automatically set to the correct value:

vk::StructureChain<vk::MemoryAllocateInfo, vk::MemoryDedicatedAllocateInfo> c = {
  vk::MemoryAllocateInfo(size, type),
  vk::MemoryDedicatedAllocateInfo(image)
};

If one of the structures of a StructureChain is to be removed, maybe due to some optional settings, you can use the function vk::StructureChain::unlink<ClassType>(). It modifies the StructureChain such that the specified structure isn't part of the pNext-chain any more. Note, that the actual memory layout of the StructureChain is not modified by that function. In case that very same structure has to be re-added to the StructureChain again, use vk::StructureChain::relink<ClassType>().

Sometimes the user has to pass a preallocated structure chain to query information. For those cases there are two corresponding getter functions. One with a variadic template generating a structure chain of at least two elements to construct the return value:

// Query vk::MemoryRequirements2HR and vk::MemoryDedicatedRequirementsKHR when calling Device::getBufferMemoryRequirements2KHR:
auto result = device.getBufferMemoryRequirements2KHR<vk::MemoryRequirements2KHR, vk::MemoryDedicatedRequirementsKHR>({});
vk::MemoryRequirements2KHR &memReqs = result.get<vk::MemoryRequirements2KHR>();
vk::MemoryDedicatedRequirementsKHR &dedMemReqs = result.get<vk::MemoryDedicatedRequirementsKHR>();

To get just the base structure, without chaining, the other getter function provided does not need a template argument for the structure to get:

// Query just vk::MemoryRequirements2KHR
vk::MemoryRequirements2KHR memoryRequirements = device.getBufferMemoryRequirements2KHR({});

Return values, Error Codes & Exceptions

By default Vulkan-Hpp has exceptions enabled. This means that Vulkan-Hpp checks the return code of each function call which returns a Vk::Result. If Vk::Result is a failure a std::runtime_error will be thrown. Since there is no need to return the error code anymore the C++ bindings can now return the actual desired return value, i.e. a vulkan handle. In those cases ResultValue ::type is defined as the returned type.

To create a device you can now just write:

vk::Device device = physicalDevice.createDevice(createInfo);

Some functions allow more than just vk::Result::eSuccess to be considered as a success code. For those functions, we always return a ResultValue<SomeType>. An example is acquireNextImage2KHR, that can be used like this:

vk::ResultValue<uint32_t> result = device->acquireNextImage2KHR(acquireNextImageInfo);
switch (result.result)
{
	case vk::Result::eSuccess:
		currentBuffer = result.value;
		break;
	case vk::Result::eTimeout:
	case vk::Result::eNotReady:
	case vk::Result::eSuboptimalKHR:
		// do something meaningfull
		break;
	default:
		// should not happen, as other return codes are considered to be an error and throw an exception
		break;
}

Note, that there are implicit cast operators in vk::ResultValue, which allow direct assignments of the returned value. That way, any success code is imlicitly ignored. If you want to make sure, you don't miss any such cases and correctly handle or explicitly ignore the success codes, define VULKAN_HPP_DISABLE_IMPLICIT_RESULT_VALUE_CAST before including vulkan.hpp. As time passes, some vulkan functions might change, such that they start to support more result codes than vk::Result::eSuccess as a success code. That logical change would not be visible in the C-API, but in the C++-API, as such a function would now return a vk::ResultValue<SomeType> instead of just SomeType. If you have VULKAN_HPP_DISABLE_IMPLICIT_RESULT_VALUE_CAST defined, you suddenly get compiler errors, that force you to handle the changed logic.

If exception handling is disabled by defining VULKAN_HPP_NO_EXCEPTIONS the type of ResultValue<SomeType>::type is a struct holding a vk::Result and a SomeType. This struct supports unpacking the return values by using std::tie.

In case you dont want to use the vk::ArrayProxy and return value transformation you can still call the plain C-style function. Below are three examples showing the 3 ways to use the API:

The first snippet shows how to use the API without exceptions and the return value transformation:

// No exceptions, no return value transformation
ShaderModuleCreateInfo createInfo(...);
ShaderModule shader1;
Result result = device.createShaderModule(&createInfo, allocator, &shader1);
if (result.result != VK_SUCCESS)
{
    handle error code;
    cleanup?
    return?
}

ShaderModule shader2;
Result result = device.createShaderModule(&createInfo, allocator, &shader2);
if (result != VK_SUCCESS)
{
    handle error code;
    cleanup?
    return?
}

The second snippet shows how to use the API using return value transformation, but without exceptions. Its already a little bit shorter than the original code:

ResultValue<ShaderModule> shaderResult1 = device.createShaderModule({...} /* createInfo temporary */);
if (shaderResult1.result != VK_SUCCESS)
{
  handle error code;
  cleanup?
  return?
}

// std::tie support.
vk::Result result;
vk::ShaderModule shaderModule2;
std::tie(result, shaderModule2)  = device.createShaderModule({...} /* createInfo temporary */);
if (shaderResult2.result != VK_SUCCESS)
{
  handle error code;
  cleanup?
  return?
}

A nicer way to unpack the result is provided by the structured bindings of C++17. They will allow us to get the result with a single line of code:

auto [result, shaderModule2] = device.createShaderModule({...} /* createInfo temporary */);

Finally, the last code example is using exceptions and return value transformation. This is the default mode of the API.

 ShaderModule shader1;
 ShaderModule shader2;
 try {
   shader1 = device.createShaderModule({...});
   shader2 = device.createShaderModule({...});
 } catch(std::exception const &e) {
   // handle error and free resources
 }

Keep in mind that Vulkan-Hpp does not support RAII style handles and that you have to cleanup your resources in the error handler!

Enumerations

For the return value transformation, there's one special class of return values which require special handling: Enumerations. For enumerations you usually have to write code like this:

std::vector<LayerProperties,Allocator> properties;
uint32_t propertyCount;
Result result;
do
{
  // determine number of elements to query
  result = static_cast<Result>( vk::enumerateDeviceLayerProperties( m_physicalDevice, &propertyCount, nullptr ) );
  if ( ( result == Result::eSuccess ) && propertyCount )
  {
    // allocate memory & query again
    properties.resize( propertyCount );
    result = static_cast<Result>( vk::enumerateDeviceLayerProperties( m_physicalDevice, &propertyCount, reinterpret_cast
     <VkLayerProperties*>( properties.data() ) ) );
  }
} while ( result == Result::eIncomplete );
// it's possible that the count has changed, start again if properties was not big enough
properties.resize(propertyCount);

Since writing this loop over and over again is tedious and error prone the C++ binding takes care of the enumeration so that you can just write:

std::vector<LayerProperties> properties = physicalDevice.enumerateDeviceLayerProperties();

UniqueHandle for automatic resource management

Vulkan-Hpp provides a vk::UniqueHandle<Type, Deleter> interface. For each Vulkan handle type vk::Type there is a unique handle vk::UniqueType which will delete the underlying Vulkan resource upon destruction, e.g. vk::UniqueBuffer is the unique handle for vk::Buffer.

For each function which constructs a Vulkan handle of type vk::Type Vulkan-Hpp provides a second version which returns a vk::UniqueType. E.g. for vk::Device::createBuffer there is vk::Device::createBufferUnique and for vk::allocateCommandBuffers there is vk::allocateCommandBuffersUnique.

Note that using vk::UniqueHandle comes at a cost since most deleters have to store the vk::AllocationCallbacks and parent handle used for construction because they are required for automatic destruction.

Custom allocators

Sometimes it is required to use std::vector with custom allocators. Vulkan-Hpp supports vectors with custom allocators as input for vk::ArrayProxy and for functions which do return a vector. For the latter case, add your favorite custom allocator as template argument to the function call like this:

std::vector<LayerProperties, MyCustomAllocator> properties = physicalDevice.enumerateDeviceLayerProperties<MyCustomAllocator>();

You can as well use a stateful custom allocator by providing it as an argument to those functions. Unfortunately, to make the compilers happy, you also need to explicitly set the Dispatch argument. To get the default there, a simple ´´´{}´´´ would suffice:

MyStatefulCustomAllocator allocator;
std::vector<LayerProperties, MyStatefulCustomAllocator> properties = physicalDevice.enumerateDeviceLayerProperties( allocator, {} );

Custom assertions

All over vulkan.hpp, there are a couple of calls to an assert function. By defining VULKAN_HPP_ASSERT, you can specifiy your own custom assert function to be called instead.

Extensions / Per Device function pointers

The Vulkan loader exposes only the Vulkan core functions and a limited number of extensions. To use Vulkan-Hpp with extensions it's required to have either a library which provides stubs to all used Vulkan functions or to tell Vulkan-Hpp to dispatch those functions pointers. Vulkan-Hpp provides a per-function dispatch mechanism by accepting a dispatch class as last parameter in each function call. The dispatch class must provide a callable type for each used Vulkan function. Vulkan-Hpp provides one implementation, DispatchLoaderDynamic, which fetches all function pointers known to the library.

// This dispatch class will fetch all function pointers through the passed instance
vk::DispatchLoaderDynamic dldi(instance);

// This dispatch class will fetch function pointers for the passed device if possible, else for the passed instance
vk::DispatchLoaderDynamic dldid(instance, device);

// Pass dispatch class to function call as last parameter
device.getQueue(graphics_queue_family_index, 0, &graphics_queue, dldid);

To use the DispatchLoaderDynamic as the default dispatcher (means: you don't need to explicitly add it to every function call), you need to #define VULKAN_HPP_DISPATCH_LOADER_DYNAMIC 1, and have the macro VULKAN_HPP_DEFAULT_DISPATCH_LOADER_DYNAMIC_STORAGE excactly once in your source code to provide storage for that default dispatcher. Then you can use it by the macro VULKAN_HPP_DEFAULT_DISPATCHER, as is shown in the code snippets below. To ease creating such a DispatchLoaderDynamic, there is a little helper class DynamicLoader. Creating a full featured DispatchLoaderDynamic is a two- to three-step process:

  1. initialize it with a function pointer of type PFN_vkGetInstanceProcAddr, to get the instance independent function pointers:
    vk::DynamicLoader dl;
    PFN_vkGetInstanceProcAddr vkGetInstanceProcAddr = dl.getProcAddress<PFN_vkGetInstanceProcAddr>("vkGetInstanceProcAddr");
    VULKAN_HPP_DEFAULT_DISPATCHER.init(vkGetInstanceProcAddr);
  1. initialize it with a vk::Instance to get all the other function pointers:
    vk::Instance instance = vk::createInstance({}, nullptr);
    VULKAN_HPP_DEFAULT_DISPATCHER.init(instance);
  1. optionally initialize it with a vk::Device to get device-specific function pointers
    std::vector<vk::PhysicalDevice> physicalDevices = instance.enumeratePhysicalDevices();
    assert(!physicalDevices.empty());
    vk::Device device = physicalDevices[0].createDevice({}, nullptr);
    VULKAN_HPP_DEFAULT_DISPATCHER.init(device);

After the second step above, the dispatcher is fully functional. Adding the third step can potentially result in more efficient code.

Samples and Tests

When you configure your project using CMake, you can enable SAMPLES_BUILD to add some sample projects to your solution. Most of them are ports from the LunarG samples, but there are some more, like CreateDebugUtilsMessenger, InstanceVersion, PhysicalDeviceDisplayProperties, PhysicalDeviceExtensions, PhysicalDeviceFeatures, PhysicalDeviceGroups, PhysicalDeviceMemoryProperties, PhysicalDeviceProperties, PhysicalDeviceQueueFamilyProperties, and RayTracing. All those samples should just compile and run. When you configure your project using CMake, you can enable TESTS_BUILD to add some test projects to your solution. Those tests are just compilation tests and are not required to run.

See Also

Feel free to submit a PR to add to this list.

License

Copyright 2015-2017 The Khronos Group Inc.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.