3.4

[General]

• removed vcpkg, switched to fetch content for dependencies

[Core]

• renamed some error codes for consistency
• deprecated now irrelevant memory flags
• RingBuffer
  • fixed missing symbol export for function by @Jaisiero
• Optional
  • added missing asserts on access
• SmallString
  • now always null terminated
• TimelineSemaphore
  • added host timeline signalling
• Swapchain
  • fixed bug in acquire causing increased frame in flight delay
  • rewrote internal error mangement
  • added amd driver workaround fixing extreme stutter when resizing
  • added color space to default format score function
• Device
  • fix bugs in device initialization
  • added function to return VkQueue from device
  • fixed bugs around buffer host address creation
  • fixed bugs around error recovery in resource creation
  • fixed bugs within memory block construction causing vma to panic
  • added HostImageCopyProperties
  • added amd driver workaround for proper error reporting
  • added submit index queue sync
    • queue submit now returns a submit index
    • added function to query latest submit index of queue
    • added host submit index wait function
    • added submit parameter: list of submit indices to wait on
  • fixed abi incompatibility between c++ and c api by @GodOfMemes
  • added support for creating tlas from memory blocks
  • updated host image copy api to reflect latest image layout handling changes
  • fixed bug in tlas destruction causing invalid buffer reuse
  • fixed internal synchronization around queues
• CommandRecorder
  • fixed synchronization bug causing deadlocks when submitting while recording commands on the same thread
  • removed functionality to keep recording after completing current commands
  • fixed bug in command pool management
  • removed live time locks in cmd recording. Cmd recorders can now record while submitting commands on another thread.
  • added error reporting for multiple command recorder functions
  • added new validation to multiple command recorder functions
  • added a function to reset assumed state of the cmd recorder (currently bound pipelines etc)
    • useful for tool integration such as DLSS/FSR
• Pipelines
  • added support for line rasterization settings to raster pipelines
  • added raytracing pipeline library support
• Slang
  • added spirv post processing to fix invalid slang output for per primitive output
  • added support to get acceleration structures from address
  • added nonunfirom ext around all image access to support amd
• Cleaned up internal validation macro by @laurelkeys

[TaskGraph]

• fully rewritten task graph internals (disabled for now)

Thanks to the external Contributors!

• @GodOfMemes
• @Jaisiero
• @laurelkeys

3.3.1

[Core]

• Naming scheme update to new image barriers
• Bugfix in swapchain for linux

3.3

General

• fixed links in readme by @FabulousCodingFox

[Core]

• Deprecated image layouts! Only remaining layouts are undefined, general and present_src
• Deprecated image pipeline barrier api
• Added new simpler pipeline image barrier api
• As read only layouts are now deprecated, shader sampled and shader storage access is now allowed simultaneously. This is possible now, as both descriptor arrays now use general layout!
• Fixed incorret host image copy validation
• Moved host image copy from required to implicit feature by @laurelkeys
• Fixed swapchain frame in flight handling. Previously the swapchain always allowed one too many frames in flight, potentially causing invalid semaphore reuse
• Optimized internal handling of hot buffer/blas/tlas data. Hot data is now on a separate array, also optimized accessor functions for hot buffer/blas/tlas data. Results in around perf 40% improvement when accessing hot data.
• Fixed compile errors on latest clang by @Jaisiero
• [BREAKING CHANGE] updated event api to use new image barrier functions. I expect these functions are very uncommonly used so i choose not to keep a deprecated old version of this part of the api.

[TaskGraph]

• Now uses general layout for all operations aside from present
• Sheduling improvements for read -> read access
• Fixed taskgraph callback data passing bugs
• Improved task callbacks to be stack allocatable
• Tasks and task builders are now fully stack allocatable, task building no longer does any hidden heap allocations!
• Added new static validation for tasks
• Moved all internal task data allocations to use flat arena memory allocations
• Added fixed size task graph cpu memory pool size
• Fixed compile errors in taskgraph on linux by@niooii

Thank you to all external Contributors!

• new contributor @niooii made their first contribution in #116
• VK_EXT_host_image_copy is now implicit, not required by @laurelkeys in #122
• Fixed compilation error for clang-win64 by @Jaisiero in #115

3.2

General

• Streamlined README
• updated a few samples and tutorial code to match new api/ removed use of deprecated api features in some samples and tutorials
• fixed the version of some deprecations

[Core]

• added new mandatory extension: HOST_IMAGE_COPY:
  • added host to image device copy function
  • added image to host device copy function
  • added host image layout transition function
• fixed queue count constant in c++ api
• exposed lifetime tracking via submit index
  • added daxa_dvc_latest_submit_index / Device::daxa_dvc_latest_submit_index function
  • added daxa_dvc_oldest_pending_submit_index / Device::oldest_pending_submit_index function
  • both together can be used to track and defer resource destruction based on daxas gpu submit progress tracking

[MemUtil]

• remaned to RingBuffer
• [BREAKING CHANGE] uses new core api submit index tracking to defer deallocations instead of requireing the signaling of a timeline semaphore
• [BREAKING CHANGE] removed timeline semaphore
• [BREAKING CHANGE] simplified api

[TaskGraph]

• deprecated and removed old TG syntax
  • [BREAKING CHANGE] removed short hand versions of stage accessors
  • [BREAKING CHANGE] removed view overrides
  • [BREAKING CHANGE] removed implicit TaskResource -> TaskResourceView conversions in view assignment
  • deprecated .view() function of task resource views
  • deprecated free attachment_view function
  • deprecated old style inline tasks and InlineTaskInfo
  • deprecated old style task access consts
  • [BREAKING CHANGE] removed runtime image index for layout (layout is the same for all runtime images of an attachment)
  • [CAUTION] changed inline task builder to return references instead of values
  • [BREAKING CHANGE] renamed readwrites_concurrent -> reads_writes_concurrent
• added support for task heads in inline tasks
  • added new Info struct containing all info about a given task head
    • Info field is very convenient as its possible to pass it as template parameters
    • deprecated old head constants and usings in favor of new Info type within the task head namespace
    • added span and at accessors to ATTACHMENTS constant in task head info
  • this might lead to the deprecation of static tasks
  • added new name overload for inline tasks "Task"
  • added new function to add a head to a inline task
  • added new head_uses function to inline tasks to allow assignment to attachment views of head attachments
  • added overloads for all accessors, overriding head attachment accessors
  • added new debug checks in new head attachment inline functions
  • added new overload for executes builder call taking function ptr and parameters instead of lambda only
  • [BREAKING CHANGE] removed experimental HeadTask class
  • [BREAKING CHANGE] removed experimental InlineTaskWithHead class
  • refactored assignment and internal template logic in multiple places
• changed multi queue api (the multi queue api unlikely to change much from the new syntax)
  • [BREAKING CHANGE] removed old secondary parameter in add task
  • added new Task::queue() function allowing tasks to specify what queue they want to run on
• fixed inconsistent stage overrides within taskgraph backend
• refactored multiple sections within the taskgraph headers to ensure some complex parts of the code are c++ spec compliant
• bumped attachment maximum to 48

3.1

Versioning has changed. From now on releases will be more regular and purely time based. A version will be released every few weeks no matter how much changes are made in that time. This will make deprecations and package manager updates easier over time.

Changelog:

[TaskGraph]

General

• fixed warnings building cl-x86_64-windows-msvc-debug by @laurelkeys
• removed pessimizing moves by @laurelkeys
• fixed bugs in resource attachment overlap validation
• added new mips and layers functions to create slices of image views
• added new id and view quick access functions to TaskInterface
• added new Views initializer type to task head made tasks
• added new shortened versions of task access names
• syntax changes to allow for intellisense to properly parse task heads
• added better syntax to initialize attachment views (Task::Views{.attachment_name=...})

Refactoring of TaskAccess

• added struct type TaskAccess
  • contains stage, access type and type restriction
• converted TaskBufferAccess and TaskImageAccess into struct containing consants of type TaskAccess
• converted all TaskBufferAccess and TaskImageAccess to TaskAccess
• added new validation for attachments based on the type restriction field in TaskAccess
• change allows us to programmatically create accesses, this is required for sane code in builder pattern
• change allowed us to remove some old clutter switch tables with more procedural code internally
• Added new shorthand initializer constants in TaskAccessConsts, that first take stage and then take access.

Added builder pattern InlineTasks V1

• InlineTask now has builter pattern functions to construct a task
• uinion containing typed accessors over internal storage make construction syntax very nice and consice:
  task.stage.access(resource), example: task.compute_shader.writes(hdr_image)
• all stages have shorthand and long names. Example: compute_shader and cs
• stages are typed, each stage only allows proper access. For example the stage index_input and present only allow for read

Task Types

• added concept of a task type
• types are Raster,Compute,Raytracing,Transfer and General
• per default a task is General
• typed tasks have a few benefits:
  • allows to omit the stage of the access. More concise syntax
  • CHANGED MEANING OF ACCESS WITHOUT STAGE! TYPED TASKS NOW OVERWRITE THE ANY STAGE WITH THE TASKS STAGE!
  • TO GET OLD GENERIC BEHAVIOR USE THE STAGE ANY NOW!
  • compute tasks for example usually all have the stage compute_shader for all attachments
  • they allow for type checking in async compute.
• added stage typed inline task constructor parameter
• added new stage less access functions (reads, writes...)
• added new generic attachment function (uses) rename?

Added BETA Multi-Queue support

• add_task takes queue
  • dead simple to use, simply send tasks to other queues
• currently a submit flushes recorded tasks in all queues and all queues wait on every other queue every submit for now
• validation has patches
• generates slightly illegal sync at times

Internals Refactor

• added a mk2 file containing all new taskgraph code
  • current taskgraph used as testbed for new code in mk2
• moved all string inputs to be string_views
• allocated and copy all input strings within each call
• added linear allocator for task and string data
• move task, name, view cache memory into linear allocator
  • saves around a thousand allocations in timberdoodle for tg construction
• converted many std::vectors to be linearly allocated DynamicVectors
• removed virtual calls, tasks no longer need virtual tables
• reduced nr of transient allocations by over 60% (measured in timberdoodle) in taskgraph construction

[Core]

• removed transfer and compute command recorder
  • was mostly useless over typing
  • more consistent with c api
  • makes async compute api easier
• added queue family to command recorder constructor
• added formatted bindless textures for uint images in daxa.slang
• improved view assignment validation
• added implicit feature for shader clock
• added implicit feature for shaderImageGatherExtended
• updated mesh shader sample
• added manual swapchain cpu frames in flight sync function
• added memory report device feature
• fixed illegal const casts within internals
• fixed handling of SUBOPTIMAL_KHR on resize event by @Derpius

3.0.4 Fixes and Cleanup

• added initial new flat types for pipeline manager
• renamed language define
• [TaskGraph] Fix access api in task interface
• [Shader] added .get() function to sampler ids in slang
• fix mandelbrot
• fix typed index handle name in slang header
• Fix some warnings
• [General] fix linking issues for types.hpp to_string functions
• added graph global additional transient image usage flags
• {TaskGraph] added task name to task interface
• [TaskGraph] added pre and post task callbacks for debugging
• fix various to_string functions
• removed assert, they return UNKNOWN CASE on default now
• added cases to some to_string enums
• [TaskGraph] Redo api for copy and clear on buffer and image. Add check in core api for buffer fill.
• [TaskGraph] Added InlineTaskWithHead
• fix image id array typed handles
• [CommandRecorder] refactor all error returns to use new return if error macro
• fix for inline names
• [MemUtil] Improve default alignment to be alignof(T) for allocations
• [TaskGraph] added task index to task interface for reflection purposes
• [TaskGraph] fixed sheduling behavior for concurrent access. Concurrent accesses are now allowed to be reordered completely freely across the whole range of batches where the concurrent access is used.
• Added new more generic access types for images in taskgraph
• enable debug info in slang
• added tlas .get() function in slang hlsl
• fix internal syntax for slang AS.get()
  *[TaskGraph] fiximage view cache
  • previously generated views for attachments with unspecified view type
  • lead to invalid views beeing created
  • disabled view creation for attachments with no view type specified
• Add missing include
• Fix missing header for std::sort (#96)
• Add missing include to fix Linux build (#97)
• fixed invalid enable of explicit feature
• Fix bug in device creation validation
• Fix minor bug in default SBT
• Added new raytracing pipeline manager functions
• added two missing nodiscards
• added new pipeline manager raster api, cleaned deprecation
• fix index type in slang
• cleaned deprecations
• fix bug in default image view creation
• [TaskGraph] optimize task graph backend ~20% saving in execute for tido
• Added push constant size default value for pipeline creation
• Added debug check for image view creation

3.0.3

Fix compilation issue from ImGui breaking change

3.0.1

minor release for vcpkg. No new additions, just bug fixes

3.0.0

• added raytracing api
• c api
• improved task head interface
• reworked task graph interface code
• reworked command recording
• performance improvements of device code
• added a lot of debug asserts and safety checks
• replaced hlsl with Slang in pipeline manager as a code compiler
• lots of bug fixes

2.0.0

• improved TaskGraph (formerly called TaskList):
  • completely replaced front end api
  • greatly simplified
  • replaced resource model (now there are transient and persistent resources)
  • optimized synchronization and ordering of tasks
  • added new runtime sync engine that can seemlessly jit sync task graphs connecting at runtime
  • added many convenience declarations and makros
  • added new statically declared tasks for better syntax and abstraction
  • added new functionalities to TaskInterface
  • added shader integration to task uses list via shader shared files
  • made transient images and buffers alias memory within task graphs
  • gave persistent images and buffers better interfaces for resource backing
  • added lots of validation and error detailed messages for errornous use
  • redid TaskUse syntax and semantics.
  • added TaskResourceViews to control subregion access in a more consistent and intuitive way.
  • lots of bugfixes
• improved core api:
  • removed dependency on vkBootstrap
  • rewrote bindless resource management code for better efficiency
  • renamings for consistency with vulkan
  • added mesh shader support
  • added optional manual memory management
  • replaced image shader integration completely (now consistent between hlsl and glsl)
  • renamed shader integration makros for consistency
  • completely abstracted image aspect from all operations
  • added new flag parameters to some api objects for more control in edge cases
  • added "bindy" uniform buffers
  • lots of bugfixes