The aim of this project is to provide low-level low-overhead haskell bindings to vulkan api. Features of the bindings:
- Keep as close as possible to vulkan naming conventions unless they violate
haskell syntax. This involves heavy usage of
PatternSynonyms
extension, and occasional violation of camel case. - Provide as much as possible information at type level, but allow avoiding any overheads related to it. Compile-time constants are duplicated at type level, but it is not necessary to use them.
- Do not introduce type marshalling overheads.
All vulkan structures have
ByteArray#
runtime representation, allowing zero-copy conversion to and from pointers. Moreover, it is not necessary to convert them at all, if one prefers to manage corresponding memory manually. - Use no dependencies except
base
.
Generated haskell bindings for vulkan api.
-
The generated library is rather big; consider using
-split-objs
or-split-sections
to reduce the size of a project. Note, enabling one of these options can make the library compiling painfully long time (take some coffee... or watch a movie). -
By default, the library loads vulkan symbols explicitly dynamically at runtime. Therefore, it does not even link to the vulkan loader library at compile time.
-
The library provides
useNativeFFI-x-y
flags that enable haskell FFI functions for vulkan corex.y
symbols. Turning on any of these flags enables compile-time linking to the vulkan loader library. -
All available extension functions can be found at runtime using simple lookup functions in
Graphics.Vulkan.Marshal.Proc
module.
Tested using stack
on:
- Windows 10 x64 with LunarG Vulkan SDK
- Mac OS High Sierra 10.13.4 with
LunarG Vulkan SDK
and
MoltenVK.
See
README-macOS.md
for the Mac OS setup tutorial. - Ubuntu 17.10 x64 with LunarG Vulkan SDK
Vulkan-Docs changed between version 1.1 and 1.2 a lot, which made adapting genvulkan rather hard. At this point, I decided to modify the generated code manually until I come up with a better way to generate haskell code fully automatically (I expect this would require a rather large refactoring).
The current semi-generated code matches v. 1.2.174 of Vulkan-Docs vk.xml. Here are some manual adjustments I've had to make:
VkAccelerationStructureInstanceKHR
has bitfields and not processed by hsc2hs and does not fitVulkanMarshal.StructRep
; the manual class instance workarounds this (rather inconveniently).- A few new cyclic module dependencies must have been fixed with manual .hs-boot
Graphics.Vulkan.Ext.VK_NV_ray_tracing
and some related structs are hidden behindenableBetaExtensions
flag (seems to compile with the flag enabled though)
Generate haskell vulkan sources using vk.xml file.
To update the api bindings, run genvulkan
using stack with this project folder:
cd genvulkan
stack build
stack exec genvulkan
Examples of programs using vulkan-api bindings. Consists of several executables implementing steps of vulkan-tutorial.com. This is the easiest way to familiarize yourself with the library.
Prerequisites
- For validation layers to work, you need to have Vulkan SDK installed, get it on vulkan.lunarg.com.
- Some examples compile shaders using
glslangValidator
via TH, so the tool must be in yourPATH
(it is included in Vulkan SDK). - Windowing is done via GLFW, so you may need to have it on your system, version 3.2 or newer.
A more haskell-style example of a vulkan program.
This is a combined result of programs in vulkan-examples
with a little cleaner code.
- Try to build it on various platforms, check if specifying foreign code calling convention is necessary.
- Remove unsafe FFI call to functions that could break at runtime. Currently we have both safe and unsafe versions for every function.
- Figure out if it is necessary to have
extra-libraries: vulkan
on various platforms (or, maybe,extra-ghci-libraries
is enough?). An alternative would be to make a C stub to get all functions viavk***ProcAddr
, which seems not the best option due to performance considerations of doing dynamic wrapping FFI. - Make
Graphics.Vulkan.Marshal.Create
fillsType
fields automatically, together with optional fields - Make
Graphics.Vulkan.Marshal.Create
provide more meaningful error messages when types of fields mismatch. - Check whether we can disallow writing
returnedonly
fields. - Investigate the need to use the extension loader (
vulkan-docs/src/ext_loader
).Graphics.Vulkan.Marshal.Proc
seems to be good enough for this low-level binding. -
vkGetProc
andvkLookupProc
currently lookup functions in a shared library, even if vulkan is linked statically. This can be dangerous! Need to check it.
-
VkXml.Sections.Commands
: parse command parameters more robustly, maybe uselanguage-c
package for that. Make parsing more compliant with the registry spec. -
VkXml.Sections.Types
parseVkTypeData
needs a cleaner rewrite. Especially, check if type and member names are parsed correctly.
The generated bindings vulkan-api
are not the only Haskell bindings for Vulkan API.
There is another package, called vulkan
that started in 2016.
The main reason for me to write this new package two years later was that vulkan
package was abandoned for a while and required significant efforts to be compiled
at the time this project started in January 2018
(as of April 2018 things seem to have changed and that package is great again :) ).
However, the are a few design decisions that render vulkan
and vulkan-api
quite different.
The main difference is that vulkan
uses regular Haskell data types plus DuplicateRecordFields
to manipulate Vulkan objects,
whereas vulkan-api
uses wrapped pinned byte arrays plus type classes and TypeApplications
;
as a result:
-
Creating and composing data types in
vulkan
is very close to normal haskell way of doing that (modulo the need to manually allocate pointers). Creating and composing data types invulkan-api
is done viaVulkanMarshal
class. There are helpers for managing memory inGraphics.Vulkan.Marshal.Create
module, you can find some examples in the repository. -
Duplicate field names in
vulkan
structure, such assType
useDuplicateRecordFields
and often require you writing a lot of type signatures explicitly, which can be very annoying. Things will become better with record type inference andOverloadedRecordFields
extension; but this is not implemented even in GHC 8.4 yet. -
Writing structure fields in
vulkan-api
is done via type classes (and heavy inlining); thus, overloading with custom data types is extremely easy (e.g. writing vectors or bytearrays directly into vulkan structures). That comes at the cost of a not particularly novice-friendly interface. -
Low overheads:
vulkan-api
structures can be converted to and from C pointers for FFI doing zero copying. There is no need topeek
all fields of a structure to read one of them.
There is a number of smaller things:
-
vulkan-api
has differentvkGetXxxProc
machinery for loading Vulkan symbols dynamically, check outGraphics.Vulkan.Marshal.Proc
for that. -
vulkan-api
keeps all Vulkan extension names inPtr CString
bi-directional patterns, which eliminates the need toalloca
when feeding them to Vulkan functions. -
Most of the constants in
vulkan-api
are duplicated at type level usingNat
andSymbol
, which should allow more type-level programming and fancy high-level wrapppers.