NoGraphicsAPI

NoGraphicsAPI is an experimental Vulkan 1.4 implementation of the ideas in Sebastian Aaltonen's No Graphics API. It explores how much of a conventional graphics API disappears when shaders use 64-bit GPU pointers, texture and sampler descriptors live in application-owned GPU memory, and synchronization describes hazards instead of resource state.

The Vulkan backend is implemented and exercised by three example applications. Metal support is not implemented; the Metal 4 design records the proposed mapping and open issues.

• GPU pointers replace buffer objects and bindings. create_gpu_heap() returns a raw allocation with a GPU address and, for mapped memory, a CPU address. Address-based commands consume GpuRange {gpu, size} directly, and shaders follow typed 64-bit pointers for vertex fetch and arbitrary data structures.
• Applications own descriptor heaps. Texture and sampler descriptor heaps are mapped GPU heaps. The application chooses slots, writes descriptors through the CPU address, binds the GPU range, and passes 32-bit indices to shaders.
• Root data is one small payload. A shared C++/Slang structure is copied with vkCmdPushDataEXT for each draw or dispatch. Its pointer fields are GPU addresses. Unlike the blog's GPU-resident, stage-specific roots, graphics stages share one CPU-supplied root.
• Barriers describe execution and memory hazards. The public API exposes global stage/access barriers, not per-resource transition lists. Normal textures remain in one unified layout.
• Pipeline binding state stays small. There are no public buffer objects, descriptor sets, descriptor layouts, pipeline layouts, or sampler objects. PSOs contain shader and fixed-function state, while data arrives through the root and descriptor heaps.
• Submission is explicit and asynchronous. Applications provide timeline points for reuse and deferred destruction. Submitted command buffers are one-shot.

The design comparison separates faithful mappings, Vulkan-driven differences, and features that remain outside the prototype.

The backend intentionally creates no VkDescriptorSetLayout, VkDescriptorPool, VkDescriptorSet, or VkPipelineLayout. Its central extensions are:

• VK_EXT_descriptor_heap for application-owned resource/sampler heaps and vkCmdPushDataEXT;
• VK_KHR_device_address_commands for address-based index, indirect, and copy commands;
• VK_KHR_shader_untyped_pointers as the descriptor-heap SPIR-V prerequisite;
• VK_KHR_unified_image_layouts, when available, to optimize ordinary texture access in VK_IMAGE_LAYOUT_GENERAL;
• VK_EXT_mesh_shader for mesh pipelines and dispatch.

Debug builds enable VK_EXT_debug_utils and the Khronos validation layer when available.

Vulkan 1.4 supplies buffer device addresses, timeline semaphores, dynamic rendering, synchronization2, scalar block layout, and the remaining core features. See Vulkan support for the concise feature and command mapping.

There are no public buffer objects or internal suballocators. Applications own data and descriptor heaps; optimal-tiled textures use separate GPU-only heaps. The optional utility library provides application-side data and texture allocation policies.

GpuRange is the non-owning address/size view used by commands. Texture and sampler descriptors are addressed in Slang through the standard heap syntax:

Texture2D<float4> texture = ResourceDescriptorHeap[texture_index];
SamplerState sampler = SamplerDescriptorHeap[sampler_index];
float4 texel = texture.Sample(sampler, uv);

Shared scalar, vector, and matrix types come from <NoGraphicsAPIUtility/shader_types.h>. Root structures are declared once and included by C++ and Slang:

struct RootArguments
{
    Vertex* vertices;
    float4x4 mvp;
};

On the CPU, pointer fields receive GPU virtual addresses while ordinary values are copied directly:

GpuCpuRange<Vertex> vertex_memory = bump_allocator.allocate<Vertex>(vertex_count);
RootArguments root{
    .vertices = vertex_memory.gpu,
    .mvp = mvp,
};
gpu::draw(commands, root, vertex_count);

Slang declares the same root as push-constant data and reads both pointer and value fields directly:

[[vk::push_constant]] ConstantBuffer<RootArguments> root;

Vertex vertex = root.vertices[vertex_id];
float4x4 mvp = root.mvp;

The draw or dispatch copies the root bytes immediately through vkCmdPushDataEXT; the root does not need to outlive the call. Shared structures use C layout, and matrix-bearing roots use row-major matrix layout. Root values must be trivially copyable, have a size divisible by four, and be no larger than DeviceCaps::max_push_data_size.

Public descriptor structures have useful defaults. Call sites use C++20 designated initializers to name only fields that differ from those defaults. Span and ByteSpan are non-owning pointer/count views used for function inputs.

The full shader-side contract is in the Slang shader contract.

Requirements:

• CMake 3.24+, a C++20 compiler, and Vulkan SDK headers and development libraries version 1.4.357 or newer;
• a little-endian x86-64 target; the optional utility math target additionally requires AVX2 and FMA;
• a Vulkan 1.4 loader and device exposing the required descriptor-heap, device-address-command, untyped-pointer, and mesh extensions above, plus their required BDA, synchronization, and 16-bit/scalar-layout features;
• coherent CPU-visible GPU memory through a PCIe BAR on a discrete GPU or UMA on an integrated GPU, plus device-local memory compatible with the supported buffers and textures;
• Slang 2026.14.1+ and SPIRV-Tools 2026.3+ when building the examples.

MSVC and clang-cl are supported on Windows. GNU and Clang can build the headless library on other platforms. MinGW, 32-bit x86, and ARM targets are not supported.

The utility implementation is shipped in-tree, so configuration has no Git or network dependency. The default configuration builds NoGraphicsAPI and its companion utility library. Examples and tests are opt-in so embedding NoGraphicsAPI with add_subdirectory() does not add development targets:

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build
cmake --install build --prefix path/to/install

The same install provides independent NoGraphicsAPI and NoGraphicsAPIUtility packages:

find_package(NoGraphicsAPI CONFIG REQUIRED)
find_package(NoGraphicsAPIUtility CONFIG REQUIRED)
target_link_libraries(my_application PRIVATE
    NoGraphicsAPI::NoGraphicsAPI
    NoGraphicsAPIUtility::math
    NoGraphicsAPIUtility::textures)

NoGraphicsAPIUtility provides shared C++/Slang types, math, data and texture suballocation, and a timeline-driven DeleteQueue. These are optional application-side policies; NoGraphicsAPI does not depend on them. The queue delays allocator reuse and resource destruction until the application timeline completes.

For repository development on Windows, enable the examples and tests explicitly. Building examples requires the Slang and SPIR-V Tools versions listed above.

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DNOGRAPHICSAPI_BUILD_EXAMPLES=ON -DNOGRAPHICSAPI_BUILD_TESTS=ON
cmake --build build
ctest --test-dir build --output-on-failure

The Windows examples use the normal swapchain path:

• triangle: the minimum vertex/fragment draw and presentation path;
• cube: typed GPU-pointer vertex fetch plus application-owned texture and sampler heaps;
• deferred_renderer: GPU-compute simulation and mesh-shader rendering in a multi-pass workload with pointer-based scene data, barriers between passes, and timeline-managed CPU/GPU overlap.

They can be run from their corresponding directories under build/examples when NOGRAPHICSAPI_BUILD_EXAMPLES is enabled.

Driver support was checked on 5 September 2026 against the latest available packages; linked reports are representative snapshots. create_device() remains authoritative; Vulkan 1.4 alone does not provide the required new extensions.

The latest available Windows drivers are AMD Adrenalin 26.9.1 and NVIDIA 616.64 WHQL.

🔴 marks missing required extensions or a capacity-limited fixed BAR. All checked ReBAR GPUs expose their main VRAM heap as CPU-visible; availability depends on platform firmware. A fixed BAR can still satisfy the memory requirement, but limits the default CPU-visible heaps; GPU-only allocations can use separate VRAM.

Checked UMA heap sizes are 5.1 GiB on a Windows Radeon 680M, 5.8 GiB on Steam Deck, and 21.2 GiB on a Windows Radeon 780M; they vary with system configuration.

Current Windows RDNA 2 reports lack VK_EXT_descriptor_heap; the checked RDNA 3 and RDNA 4 AMD GPUs and all listed NVIDIA GPUs advertise every required extension. With Mesa RADV 26.2 or newer, Steam Deck also has all required extensions. Linux/SteamOS window presentation is not implemented yet and is planned for a near-term update.

🔴 Pascal / GeForce GTX 10 is below the NVIDIA floor: it lacks all four new extensions and exposes a 214 MiB fixed BAR.

Intel Windows support is not verified. The latest public Arc report lacks the descriptor-heap, device-address-command, and shader-untyped-pointer extensions. Mesa ANV 26.2 or newer exposes the required extensions on Linux, but this repository currently lacks Linux swap chain support (to be implemented).

Implemented today: graphics, mesh, and compute PSOs; direct and indirect work; GPU-address copies; application-owned descriptor heaps; common texture types and views; dynamic rendering; global barriers; timeline submission; deferred destruction; and Win32 presentation.

This is a deliberately single-threaded, single-queue prototype and a low-level thin wrapper, not a validation layer. API preconditions are generally enforced by debug assertions; use the Vulkan validation layer during development. Ray tracing, task shaders, sparse memory, device-generated command graphs beyond the existing indirect operations, pipeline caching, MSAA, non-Win32 presentation, and a Metal backend are outside the current implementation. The public header remains the source of truth for the exact API surface.

Further reading:

• Comparison with No Graphics API
• Vulkan extension and command mapping
• Slang shader contract and root ABI
• Proposed Metal 4 port

NoGraphicsAPI and NoGraphicsAPIUtility are distributed under the MIT License. The cube example's texture is derived from Vulkan-Tools and is redistributed under Apache-2.0. See third-party notices for complete attribution and license details.