Vulkan® Memory Allocator

The industry-leading, open source, memory allocation library for the Vulkan API.

The Vulkan® Memory Allocator (VMA) library provides a simple and easy to integrate API to help you allocate memory for Vulkan® buffer and image storage.

Download the latest version - v3.2.1

View VMA on GitHub
Read the VMA documentation
This release adds the following features:
  • Fixed an assert in vmaCreateAllocator function incorrectly failing when Vulkan version 1.4 is used.
  • Fix for importing function vkGetPhysicalDeviceMemoryProperties2 / vkGetPhysicalDeviceMemoryProperties2KHR when VMA_DYNAMIC_VULKAN_FUNCTIONS macro is enabled.
  • Other minor fixes and improvements.

Benefits

The library is battle-ready, and integrated into the majority of Vulkan® game titles on PC, as well as the Google Filament rendering engine, the official Khronos® Group Vulkan® Samples, and many other open source projects for Android™, Linux, MacOS, and Windows®.

This library can help game developers to manage memory allocations and resource creation by offering some higher-level functions:

  1. Functions that help to choose correct and optimal memory type based on intended usage of the memory.
    • Required or preferred traits of the memory are expressed using higher-level description comparing to Vulkan flags.
  2. Functions that allocate memory blocks, reserve and return parts of them (VkDeviceMemory + offset + size) to the user.
    • Library keeps track of allocated memory blocks, used and unused ranges inside them, finds best matching unused ranges for new allocations, respects all the rules of alignment and buffer/image granularity.
  3. Functions that can create an image/buffer, allocate memory for it and bind them together – all in one call.

Additional features

  • Self-contained C++ library in single header file. No external dependencies other than the standard C and C++ library and Vulkan.
  • Public interface in C, in same convention as Vulkan API. Implementation in C++.
  • Platform-independent, but developed and tested on Windows using Visual Studio. Continuous integration setup for Windows and Linux. Used also on Android, MacOS, and other platforms.
  • Support for memory mapping, reference-counted internally.
  • Support for persistently mapped memory: Just allocate with appropriate flag and access the pointer to already mapped memory.
  • Linear allocator: Create a pool with linear algorithm and use it for much faster allocations and deallocations in free-at-once, stack, double stack, or ring buffer fashion.
  • Defragmentation of GPU and CPU memory: Let the library move data around to free some memory blocks and make your allocations better compacted.
  • Statistics: Obtain brief or detailed statistics about the amount of memory used, unused, number of allocated blocks, number of allocations etc. - globally, per memory heap, and per memory type.
  • Debug annotations: Associate custom void* pUserData and debug char* pName with each allocation.
  • JSON dump: Output a JSON string with detailed map of internal state, including allocations, their string names, and gaps between them.
  • Convert JSON dump into a picture to visualize your memory.
  • Support for interoperability with OpenGL.
  • Virtual allocator: Interface for using core allocation algorithm to allocate any custom data, e.g. pieces of one large buffer.
I can let VMA handle all the hard work and do a better job than I would. On top of that it still lets me easily drop down to raw Vulkan when I need the finest control for something specific. You should absolutely use VMA for your general memory allocation needs in Vulkan, even when you do need something special it will probably do that for you or make it easy to write some special case code.
Baldur Karlsson
Baldur KarlssonCreator, RenderDoc
We integrated AMD's Vulkan Memory Allocator (VMA) and never looked back.
Jean-Noe Morissette
Jean-Noe MorissetteSoftware Architect, Ubisoft
My experience with VMA (…) was that it let you have as much or little control as you wanted. Made Vulkan feel more like D3D11 unless you really wanted to think about heaps. I am doing another Vk project soon and plan to use it.
Timothy Lochner
Timothy LochnerSenior Engine Architect. Rightware Inc.
Previous
Next

Version history

  • Fixed an assert in vmaCreateAllocator function incorrectly failing when Vulkan version 1.4 is used.
  • Fix for importing function vkGetPhysicalDeviceMemoryProperties2 / vkGetPhysicalDeviceMemoryProperties2KHR when VMA_DYNAMIC_VULKAN_FUNCTIONS macro is enabled.
  • Other minor fixes and improvements.

This release adds the following features:

  • Additions to the library API:
    • Added support for Vulkan 1.4.
    • Added support for VK_KHR_external_memory_win32 extension – VMA_ALLOCATOR_CREATE_KHR_EXTERNAL_MEMORY_WIN32_BIT flag, vmaGetMemoryWin32Handle function, and a whole new documentation chapter about it.
  • Other changes:
    • Fixed thread safety issue.
    • Many other bug fixes and improvements in the library code, documentation, sample app, Cmake script, mostly to improve compatibility with various compilers and GPUs.

This release adds the following features:

  • Added convenience functions vmaCopyMemoryToAllocation , vmaCopyAllocationToMemory .
  • Added functions vmaCreateAliasingBuffer2 , vmaCreateAliasingImage2 that offer creating a buffer/image in an existing allocation with additional allocationLocalOffset .
  • Added function vmaGetAllocationInfo2 , structure VmaAllocationInfo2 that return additional information about an allocation, useful for interop with other APIs.
  • Added callback VmaDefragmentationInfo::pfnBreakCallback that allows breaking long execution of vmaBeginDefragmentation . Also added PFN_vmaCheckDefragmentationBreakFunction , VmaDefragmentationInfo::pBreakCallbackUserData .
  • Added support for VK_KHR_maintenance4 extension – VMA_ALLOCATOR_CREATE_KHR_MAINTENANCE4_BIT flag.
  • Added support for VK_KHR_maintenance5 extension – VMA_ALLOCATOR_CREATE_KHR_MAINTENANCE5_BIT flag.
  • Changes in debug and configuration macros:
    • Split macros into separate VMA_DEBUG_LOG and VMA_DEBUG_LOG_FORMAT .
    • Added macros VMA_ASSERT_LEAK , VMA_LEAK_LOG_FORMAT separate from normal VMA_ASSERT , VMA_DEBUG_LOG_FORMAT .
    • Added macro VMA_EXTENDS_VK_STRUCT .
  • Countless bug fixes and improvements in the code and documentation, mostly to improve compatibility with various compilers and GPUs, including:
    • Fixed missing #include that resulted in compilation error about snprintf not declared on some compilers.
    • Fixed main memory type selection algorithm for GPUs that have no HOST_CACHED memory type, like Raspberry Pi.
  • Major changes in Cmake script.
  • Fixes in GpuMemDumpVis.py script.
  • Fixes in defragmentation algorithm.
  • Fixes in GpuMemDumpVis.py regarding image height calculation.
  • Other bug fixes, optimizations, and improvements in the code and documentation.
  • Added new API for selecting preferred memory type.
  • Added new defragmentation API and algorithm, replacing the old one.
  • Redesigned API for statistics, replacing the old one.
  • Added “Virtual allocator” feature – possibility to use core allocation algorithms for allocation of custom memory, not necessarily Vulkan device memory.
  • VmaAllocation now keeps both void* pUserData and char* pName . Added function vmaSetAllocationName , member VmaAllocationInfo::pName .
  • Clarified and cleaned up various ways of importing Vulkan functions. Added members VmaVulkanFunctions::vkGetInstanceProcAddr , and vkGetDeviceProcAddr , which are now required when using VMA_DYNAMIC_VULKAN_FUNCTIONS .
  • Added support for Vulkan 1.1.
    • Added member VmaAllocatorCreateInfo::vulkanApiVersion .
    • When Vulkan 1.1 is used, there is no need to enable VK_KHR_dedicated_allocation or VK_KHR_bind_memory2 extensions, as they are promoted to Vulkan itself.
  • Added support for query for memory budget and staying within the budget.
    • Added function vmaGetBudget , structure VmaBudget . This can also serve as simple statistics, more efficient than vmaCalculateStats .
    • By default the budget it is estimated based on memory heap sizes. It may be queried from the system using VK_EXT_memory_budget extension if you use VMA_ALLOCATOR_CREATE_EXT_MEMORY_BUDGET_BIT flag and VmaAllocatorCreateInfo::instance member.
    • Added flag VMA_ALLOCATION_CREATE_WITHIN_BUDGET_BIT that fails an allocation if it would exceed the budget.
  • Added new memory usage options:
    • VMA_MEMORY_USAGE_CPU_COPY for memory that is preferably not DEVICE_LOCAL but not guaranteed to be HOST_VISIBLE .
    • VMA_MEMORY_USAGE_GPU_LAZILY_ALLOCATED for memory that is LAZILY_ALLOCATED .
  • Added support for VK_KHR_bind_memory2 extension:
    • Added VMA_ALLOCATION_CREATE_DONT_BIND_BIT flag that lets you create both buffer/image and allocation, but don’t bind them together.
    • Added flag VMA_ALLOCATOR_CREATE_KHR_BIND_MEMORY2_BIT , functions vmaBindBufferMemory2 , vmaBindImageMemory2 that let you specify additional local offset and pNext pointer while binding.
  • Added functions vmaSetPoolName , vmaGetPoolName that let you assign string names to custom pools. JSON dump file format and VmaDumpVis tool is updated to show these names.
  • Defragmentation is legal only on buffers and images in VK_IMAGE_TILING_LINEAR . This is due to the way it is currently implemented in the library and the restrictions of the Vulkan specification. Clarified documentation in this regard.
  • New, more powerful defragmentation:
    • Added structure VmaDefragmentationInfo2, functions vmaDefragmentationBegin, vmaDefragmentationEnd.
    • Added support for defragmentation of GPU memory.
    • Defragmentation of CPU memory now uses memmove, so it can move data to overlapping regions.
    • Defragmentation of CPU memory is now available for memory types that are HOST_VISIBLE but not HOST_COHERENT.
    • Added structure member VmaVulkanFunctions::vkCmdCopyBuffer.
    • Major internal changes in defragmentation algorithm.
    • VmaReplay: added parameters: --DefragmentAfterLine, --DefragmentationFlags.
    • Old interface (structure VmaDefragmentationInfo, function vmaDefragment) is now deprecated.
  • Added buddy algorithm, available for custom pools – flag VMA_POOL_CREATE_BUDDY_ALGORITHM_BIT.
  • Added convenience functions for multiple allocations and deallocations at once, intended for sparse binding resources – functions vmaAllocateMemoryPages, vmaFreeMemoryPages.
  • Added function that tries to resize existing allocation in place: vmaResizeAllocation.
  • Added flags for allocation strategy: VMA_ALLOCATION_CREATE_STRATEGY_BEST_FIT_BIT, VMA_ALLOCATION_CREATE_STRATEGY_WORST_FIT_BIT, 
    VMA_ALLOCATION_CREATE_STRATEGY_FIRST_FIT_BIT, and their aliases: VMA_ALLOCATION_CREATE_STRATEGY_MIN_MEMORY_BIT, 
    VMA_ALLOCATION_CREATE_STRATEGY_MIN_TIME_BIT, VMA_ALLOCATION_CREATE_STRATEGY_MIN_FRAGMENTATION_BIT.
  • Added linear allocation algorithm, accessible for custom pools, that can be used as free-at-once, stack, double stack, or ring buffer. See “Linear allocation algorithm” documentation chapter.
    • Added VMA_POOL_CREATE_LINEAR_ALGORITHM_BIT, VMA_ALLOCATION_CREATE_UPPER_ADDRESS_BIT.
  • Added feature to record sequence of calls to the library to a file and replay it using dedicated application. See documentation chapter “Record and replay”.
    • Recording: added VmaAllocatorCreateInfo::pRecordSettings.
    • Replaying: added VmaReplay project.
    • Recording file format: added document “docs/Recording file format.md”.
  • Improved support for non-coherent memory.
    • Added functions: vmaFlushAllocation, vmaInvalidateAllocation.
    • nonCoherentAtomSize is now respected automatically.
    • Added VmaVulkanFunctions::vkFlushMappedMemoryRanges, vkInvalidateMappedMemoryRanges.
  • Improved debug features related to detecting incorrect mapped memory usage. See documentation chapter “Debugging incorrect memory usage”.
    • Added debug macro VMA_DEBUG_DETECT_CORRUPTION, functions vmaCheckCorruption, vmaCheckPoolCorruption.
    • Added debug macro VMA_DEBUG_INITIALIZE_ALLOCATIONS to initialize contents of allocations with a bit pattern.
    • Changed behavior of VMA_DEBUG_MARGIN macro – it now adds margin also before first and after last allocation in a block.
  • Changed format of JSON dump returned by vmaBuildStatsString (not backward compatible!).
    • Custom pools and memory blocks now have IDs that don’t change after sorting.
    • Added properties: “CreationFrameIndex”, “LastUseFrameIndex”, “Usage”.
    • Changed VmaDumpVis tool to use these new properties for better coloring.
    • Changed behavior of vmaGetAllocationInfo and vmaTouchAllocation to update allocation.lastUseFrameIndex even if allocation cannot become lost.
  • Introduction of VmaAllocation handle that you must retrieve from allocation functions and pass to deallocation functions next to normal VkBuffer and VkImage.
  • Introduction of VmaAllocationInfo structure that you can retrieve from VmaAllocation handle to access parameters of the allocation (like VkDeviceMemory and offset) instead of retrieving them directly from allocation functions.
  • Support for reference-counted mapping and persistently mapped allocations – see vmaMapMemory, VMA_ALLOCATION_CREATE_MAPPED_BIT.
  • Support for custom memory pools – see VmaPool handle, VmaPoolCreateInfo structure, vmaCreatePool function.
  • Support for defragmentation (compaction) of allocations – see function vmaDefragment and related structures.
  • Support for “lost allocations” – see appropriate chapter on documentation Main Page.
  • Initial release.

Related to Vulkan® Memory Allocator

Vulkan Memory Allocator 2.3.0

The latest Vulkan Memory Allocator v2.3.0 adds support for Vulkan 1.1, support for query for memory budget, and support for VK_KHR_bind_memory2 extension.

AMD GPUOpen Direct3D12 Memory Allocator (D3D12MA)

D3D12 Memory Allocator

The D3D12 Memory Allocator (D3D12MA) is a C++ library that provides a simple and easy-to-integrate API to help you allocate memory for DirectX®12 buffers and textures.

Our other SDKs

AMD Dense Geometry Format

Dense Geometry Compression Format (DGF) is our block-based geometry compression technology. It is a hardware-friendly format, supported by future GPU architectures.

AMD Schola

AMD Schola is a library for developing reinforcement learning (RL) agents in Unreal Engine and training with your favorite python-based RL Frameworks.

Anti-Lag 2 SDK

AMD Radeon™ Anti-Lag 2 reduces the system latency by applying frame alignment between the CPU and GPU jobs.

AMD Capsaicin Framework

Capsaicin is a Direct3D12 framework for real-time graphics research which implements the GI-1.0 technique and a reference path-tracer.

AMD Render Pipeline Shaders (RPS) SDK

The Render Pipeline Shaders (RPS) SDK provides a framework for graphics engines to use Render Graphs with explicit APIs.

AMD Device Library eXtra

ADLX is a modern library designed to access features and functionality of AMD systems such as Display, 3D graphics, Performance Monitoring, GPU Tuning, and more.

AMD Brotli-G SDK

Brotli-G is an open-source compression/decompression standard for digital assets (based on Brotli) that is compatible with GPU hardware.

AMD HIP Ray Tracing

HIP RT is a ray tracing library for HIP, making it easy to write ray tracing applications in HIP.

AMD Orochi

Orochi is a library which loads HIP and CUDA® APIs dynamically, allowing the user to switch APIs at runtime.

AMD Radeon ProRender Developer Suite

AMD Radeon™ ProRender is our fast, easy, and incredible physically-based rendering engine built on industry standards that enables accelerated rendering on virtually any GPU, any CPU, and any OS in over a dozen leading digital content creation and CAD applications.

AMD Radeon ML

Radeon™ Machine Learning (Radeon™ ML or RML) is an AMD SDK for high-performance deep learning inference on GPUs.

AMD Radeon Image Filter

Harness the power of machine learning to enhance images with denoising, enabling your application to produce high quality images in a fraction of the time traditional denoising filters take.