Backend Plugin Architecture

On desktop platforms, backends are dynamically loaded shared libraries (luisa-backend-<name>.dll/.so/.dylib) implementing DeviceInterface and discovered by Context. iOS links the selected backend and core runtime as static slices into the signed app; it calls an explicit static create/destroy bridge while preserving the same DeviceInterface ownership contract.

Header: include/luisa/runtime/rhi/device_interface.h

DeviceInterface API

All backends inherit from DeviceInterface:

Resource Lifecycle (Handle-based)

Execution

Queries & Extensions

Event Sync

Dynamic Loading

File: src/runtime/context.cpp

Context scans the runtime directory for libraries matching luisa-backend-*.{so|dll|dylib} (and libluisa-backend-* for MinGW). Names are lower-cased, sorted, and de-duplicated.

const BackendModule &load_backend(const luisa::string &backend_name) {
    // 1. Error out (LUISA_ERROR) unless the name is in installed_backends
    // 2. Reuse an already loaded BackendModule from loaded_backends if present
    // 3. Static-registry hit -> take creator/deleter/device_names from
    //    find_static_backend() and skip module loading entirely
    // 4. Otherwise DynamicModule::load(runtime_directory,
    //       "luisa-backend-<name>") and validate
    //       backend_version() == LUISA_COMPUTE_VERSION
    // 5. Extract: creator("create"), deleter("destroy"), backend_device_names
}

Device creation:

Device Context::create_device(backend_name, settings, enable_validation) {
    auto &m = _impl->load_backend(backend_name);
    auto interface = m.creator(Context{_impl}, settings);
    interface->_backend_name = backend_name;
    auto handle = Device::Handle{interface, [deleter = m.deleter](auto p) { deleter(p); }};
    if (enable_validation) {
        auto &validation_layer = _impl->load_validation_layer();
        handle = Device::Handle{
            validation_layer.creator(Context{_impl}, std::move(handle)),
            /* validation deleter */};
    }
    return Device{std::move(handle)};
}

Validation can also be enabled via the LUISA_ENABLE_VALIDATION=1 environment variable (Context::create_device overload without the explicit flag). The layer itself is a separate module, luisa-validation-layer (src/backends/validation/), not a luisa-backend-* plugin.

Backend Registration

Every backend must export three C functions with LUISA_EXPORT_API (signatures fixed by Device::Creator / Device::Deleter / Context::StaticBackendDeviceNames in include/luisa/runtime/device.h:102-103 and include/luisa/runtime/context.h:33-34):

// src/backends/<name>/<name>_device.cpp

LUISA_EXPORT_API DeviceInterface *create(Context &&ctx, const DeviceConfig *config) noexcept {
    return new_with_allocator<MyBackendDevice>(std::move(ctx), config);
}
LUISA_EXPORT_API void destroy(DeviceInterface *device) noexcept {
    delete_with_allocator(device);
}
// Fills in the concrete *device/adapter* names shown to the user - NOT the
// backend name: cuda pushes cuDeviceGetName(), metal4 pushes
// device->name(), fallback/simd push luisa::cpu_name().
LUISA_EXPORT_API void backend_device_names(luisa::vector<luisa::string> &names) noexcept {
    names.clear();
    ...
}

Plus version export (src/backends/common/export_version.inl.h, #included at the bottom of every <name>_device.cpp):

LUISA_EXPORT_API int backend_version() { return LUISA_COMPUTE_VERSION; }

LUISA_COMPUTE_VERSION is major * 10000 + minor * 100 + patch (include/luisa/version.h:11).

Static iOS registration

iOS cannot discover an in-bundle backend through desktop MODULE loading. The app calls luisa_compute_metal4_register_static_backend() from src/backends/metal4/metal_static_backend.h (the old MSL backend ships the sibling src/backends/metal/metal_static_backend.h / luisa_compute_metal_register_static_backend()) before device creation. Both headers declare the entry point only under LUISA_PLATFORM_IOS. Each bridge registers the normal create/destroy/device-name functions through Context::register_static_backend("metal4", create, destroy, backend_device_names); Context::init_backends() seeds installed_backends from that registry and load_backend() prefers a registry hit over dynamic loading. Names are lower-cased before lookup, so registration is case-insensitive, and re-registering different entry points under the same name asserts. Keep create_device("metal4"), ordinary DeviceInterface ownership, and the validation-layer path intact instead of constructing a backend directly in a UIKit host.

Each iOS bundle is a static application closure (BUILD_SHARED_LIBS=OFF): runtime, DSL, XIR, Metal4, llvm-downgrade, and target LLVM archives are inside the arm64 Mach-O. Audit with scripts/audit_ios_bundles.sh; otool -L should show Apple system paths only.

Command Encoder (Visitor Pattern)

Commands dispatched via MutableCommandVisitor:

class MyCommandEncoder : public MutableCommandVisitor {
    MyStream *_stream;
    void visit(BufferUploadCommand *cmd) noexcept override;
    void visit(BufferDownloadCommand *cmd) noexcept override;
    void visit(BufferCopyCommand *cmd) noexcept override;
    void visit(BufferToTextureCopyCommand *cmd) noexcept override;
    void visit(TextureUploadCommand *cmd) noexcept override;
    void visit(TextureDownloadCommand *cmd) noexcept override;
    void visit(TextureCopyCommand *cmd) noexcept override;
    void visit(TextureToBufferCopyCommand *cmd) noexcept override;
    void visit(ShaderDispatchCommand *cmd) noexcept override;
    void visit(AccelBuildCommand *cmd) noexcept override;
    void visit(MeshBuildCommand *cmd) noexcept override;
    void visit(CurveBuildCommand *cmd) noexcept override;
    void visit(ProceduralPrimitiveBuildCommand *cmd) noexcept override;
    void visit(MotionInstanceBuildCommand *cmd) noexcept override;
    void visit(BindlessArrayUpdateCommand *cmd) noexcept override;
    void visit(CustomCommand *cmd) noexcept override;
    void commit(CommandList::CallbackContainer &&user_callbacks) noexcept;
};

Architecture: each backend has a *Stream class owning the native stream/queue. stream->dispatch(CommandList) visits all commands via an encoder. User callbacks are executed after GPU work completes.

Both visitor flavours exist: CommandVisitor (void visit(const CMD *)) for read-only passes such as reordering, MutableCommandVisitor (void visit(CMD *)) for encoders. Extension commands reach backends through the single CustomCommand hook: CustomCommand::custom_cmd_uuid() identifies them against CustomCommandUUID (include/luisa/backends/ext/registry.h). Dispatch-style commands derive from CustomDispatchCommand : public CustomCommand (VKCustomCmd, DXCustomCmd, NativeShaderDispatchCommand, CudaKernelLaunchCommand); the raster commands (DrawRasterSceneCommand, ClearDepthCommand, ClearRenderTargetCommand) and DStorageReadCommand derive from CustomCommand directly.

Metal4 acceleration capability boundary

Creating an MTL4::Compiler, queue, or AIR pipeline does not prove that every MTL4 encoder feature is executable. Address-driven acceleration-structure builds and component motion require Apple9 — reported by the metal4_address_driven_acceleration_structures / metal4_component_motion query properties and gated by supportsFamily(MTL::GPUFamilyApple9) (src/backends/metal4/metal_device.cpp:878-882). Query the concrete device family (metal4_gpu_family): Apple9+ uses MTL4 primitive/instance descriptors and its compute encoder; Apple7/Apple8 synchronize only AS build/refit/compact through an isolated legacy MTL::CommandQueue opened per stream (_acceleration_structure_compatibility_queue, src/backends/metal4/metal_stream.cpp:213-225, and build_acceleration_structure_compatibility / refit_acceleration_structure_compatibility in src/backends/metal4/metal_acceleration_structure_build.h). User shaders, PSOs, argument tables, command buffers, and dispatch remain MTL4 AIR on both paths.

MotionInstanceBuildCommand is host-state capture: validate a built child and copy its matrix/SRT keyframes into the backend resource. Native motion TLAS packing occurs later in AccelBuildCommand. Preserve the shader-visible 72-byte static instance ABI while creating separate 48-byte indirect-motion records and a transform buffer for the build descriptor. The ABIs are static_asserted against MTL::IndirectAccelerationStructureInstanceDescriptor and MTL::IndirectAccelerationStructureMotionInstanceDescriptor in src/backends/metal4/metal_accel.h:28-68; the transform buffer is MetalAccel::_motion_transform_buffer. Matrix motion is available where primitive motion blur is reported (metal_motion_blur → supportsPrimitiveMotionBlur()); component/SRT motion must be rejected before resource creation below Apple9.

On iOS, do not call MTL::Device::isDepth24Stencil8PixelFormatSupported() merely because metal-cpp's safe-send helper finds a method signature. Some AGX devices expose that signature without responding to the selector. Check real Objective-C class selector responsiveness first, then map logical D24S8 storage to D32S8A24 when unavailable; preserve the public logical format and execute both stencil paths.

Command reordering and bindless hazards

src/backends/common/command_reorder_visitor.h plans command layers from the resource accesses visible at each command boundary (CommandReorderVisitor is itself a CommandVisitor; a backend supplies a ReorderFuncTable). The pass is switchable through CommandReorderSwitch (src/backends/common/command_reorder_switch.h): seeded from VulkanDeviceConfigExt::enable_command_reorder() / DirectXDeviceConfigExt::EnableCommandReorder(), changed at runtime by CommandReorderExt::set_command_reorder_enabled(), and forced off process-wide by LUISA_DISABLE_COMMAND_REORDER=1 (which the first two cannot re-enable). A backend instantiating this visitor must preserve these contracts:

• Saved shader argument Usage is authoritative. Do not rediscover or approximate read/write access in the reorder or barrier layer.
• A bindless dispatch reads the bindless index/descriptor object itself.
• Snapshot every resource currently reachable from the bindless array when the dispatch is visited. Register each member — buffer or texture — as a whole-resource read, or as a write when the saved bindless argument usage declares WRITE. The DSL marks the whole array variable WRITE when the shader writes through any slot and tex2d()/tex3d() views support write(), so restricting the write mark to buffers leaves bindless-texture RAW/WAW races unordered (add_bindless_dispatch_handles).
• The snapshot belongs to that dispatch. Never infer an earlier dispatch's hazards from the array's later membership; a subsequent update may replace a slot, and two distinct arrays may still reference the same resource.
• Backend resource-barrier preprocessing must mirror the same exact saved usage and resource snapshot. Correct reordering without matching native API access masks is insufficient, and vice versa.
• A Vulkan VKCustomCmd::ResourceUsage that names a bindless array applies its declared native stage/access/layout contract both to the array's index storage and to every member in the encoded descriptor snapshot at that command boundary. Reordering and native barriers must describe the same aggregate access.
• VKCustomCmd states are exact native contracts, not abstract usage hints. Place them in an isolated resource reorder layer: otherwise a read-only custom layout can merge with another read into one layer even though the two commands pass different VkImageLayout values to Vulkan. Isolation covers the bindless descriptor object and every snapshotted member.
• The simultaneous-access optimization may select GENERAL only for backend-owned commands that query the tracker's selected layout. Never replace a custom command's explicit native layout. When aggregating access, an empty prior access is an identity and repeated equal read layouts remain equal; only incompatible abstract read layouts collapse to GENERAL.
• A VulkanDeviceConfigExt before_states or after_states entry naming a bindless array expands over the descriptor index storage and the exact encoded member snapshot. Preserve the entry's raw stage/access contract for every member and its texture layout for every referenced mip; the texture layout must be defined whenever the snapshot contains an image.
• VulkanDeviceConfigExt::before_states() is expanded against the bindless snapshot at command-list entry. after_states() is expanded only after all commands have been encoded, against the final snapshot after in-list updates. Treating both callbacks as if they saw the same membership loses either newly inserted resources or resources removed during the list.

When a backend permits multiple opaque handles to wrap the same native resource, the native resource identity is an additional alias boundary. Either canonicalize hazards by that identity or enforce a documented external-sync contract; pointer equality between backend wrappers is not enough.

Acceleration-structure (BLAS/TLAS) hazards

A TLAS build dereferences every instanced BLAS through the raw device addresses baked into the instance buffer (instance AABBs derive from child BLAS contents). Two generic consequences:

• Barrier coverage must be per-referenced-BLAS, not per-modified-instance. Recording the child-BLAS read barrier only for instances appearing in the current modification list misses the in-place BLAS update case: it leaves both modifications and the pending-refresh slots empty, so the TLAS update build races the BLAS build and picks up stale geometry (or device-lost). Iterate the full instance table and record a read on every live child BLAS buffer at every TLAS build (src/backends/vk/tlas.cpp:583-595).
• BLAS destruction must unlink the BLAS from every referencing TLAS's pending refresh bookkeeping (the refresh queued by a BLAS recreate to rewrite its device address, which now lives in the TLAS's dense array slot — Instance::refresh_blas in src/backends/vk/tlas.h:39-47, BottomAccel *refresh in src/backends/dx/Resource/TopAccel.h:35-43 — replacing the former per-TLAS hash map). Store the stable BLAS, never the pooled handle, and drop the entry from the destructor. Only clearing the live-instance slot leaves a dangling pooled handle the next TLAS build dereferences — a UAF that typically manifests as a hang (spin lock inside freed memory), not a crash. Applies to any backend with update-handle bookkeeping (vk, dx).

For Vulkan-owned buffers and images, queue-family sharing is a separate creation-time contract. If graphics, compute, and copy select more than one unique queue family, create resources in concurrent sharing mode over exactly those unique families; a single-family device may retain exclusive sharing. Timeline events still establish execution and memory ordering between streams: concurrent sharing removes queue-family ownership transfers, not synchronization. Sparse resources follow the same graphics/compute/copy sharing plan. A backend-owned logical device additionally requests one canonical queue whose family advertises VK_QUEUE_SPARSE_BINDING_BIT, preferring a family without graphics or compute capability and falling back to a shared stream family. Route every vkQueueBindSparse through that queue and its canonical per-handle mutex. Sparse binding changes page tables but does not access the resource in a pipeline stage, so a dedicated sparse-only family is not added to the resource-access sharing list. Imported logical devices keep sparse support disabled until the import ABI can attest enabled sparse features and supply an actual sparse queue handle/family; physical-device support alone is not enough. Creation also requires the corresponding enabled logical-device features (sparseBinding plus buffer/2D/3D residency), and the format-specific sparse image query must succeed. Vulkan forbids sparse-residency 1D images. Do not encode a guessed sparse page size or image tile shape. Query each created resource's VkMemoryRequirements, query every sparse image's VkSparseImageMemoryRequirements, and retain the selected color-aspect requirement with the resource. The non-tail image tile shape is formatProperties.imageGranularity; each tile consumes one full VkMemoryRequirements::alignment block even when an edge bind clips its texel extent against the selected mip. Reject mip-tail levels in the public tile-bind path, and reject creation when any requested mip begins at or beyond imageMipTailFirstLod: those levels require opaque binds that SparseTextureMapOperation cannot express, so returning a partly usable image would make the public mip range dishonest. Likewise reject metadata or ambiguous aspect requirements until the public API can describe their opaque bindings.

Sparse-buffer binds use VkMemoryRequirements::alignment for resource offset, memory offset, and bind size. Preserve the API-visible VkBufferCreateInfo size and use VkMemoryRequirements::size as the aligned sparse virtual-address range; the requirements already include any padding needed by the final page. Validate map and unmap ranges with checked arithmetic against this sparse tile grid; do not clip a final buffer page as if it were an image edge.

The public sparse heap is allocated before a consuming Vulkan resource exposes its memory type mask and alignment. Defer its physical VkDeviceMemory allocation until the first map, allocate for that resource's real requirements, exclude lazily allocated memory types, and require all later consumers to be compatible with the chosen memory type and alignment. Until that first map no VkDeviceMemory exists, and the immutable ResourceCreationInfo::native_handle remains null: native sparse-heap interop is intentionally unsupported on Vulkan. A map region consumes tile_count * alignment bytes even when its final image tile is texel-clipped; heap-capacity checks must use that full-block size rather than copied pixel size. Within one VkBindSparseInfo, reject overlapping ranges of the same resource, including a map overlapping an unmap. The public sparse API supplies no heap offset or alias opt-in, and the Vulkan device does not enable aliased sparse residency. Maintain a mutex-protected, Device-wide residency registry across all streams: a heap may back only one live mapping (possibly split into fragments by partial unmaps), mapped resource ranges may not overlap, and a heap released by an unmap may not be recycled by another bind in the same unordered batch. Require maps to target an inactive registered heap and an entirely unmapped resource range. Require unmaps to be fully covered by live mappings, subtract partial buffer intervals or image texel boxes exactly, and keep the heap active until its last fragment is removed. Hold the registry transaction through native submission and host completion, commit only after success, and reject destruction of a heap or sparse resource with live mappings.

Sparse-binding submissions also have no implicit execution dependency with adjacent command-buffer submissions on the same queue. A stream can also contain external-event wait/signal submissions that do not advance its internal timeline. Immediately before sparse binding, submit a bridge signal of the internal timeline on the ordinary stream queue; make the sparse queue wait for that exact bridge value and signal a second value. Query the current semaphore counter and reject a handoff whose farther signal would exceed the physical device's maxTimelineSemaphoreValueDifference; host-only logical fence gaps are not a license to violate the timeline value limit. Establish an explicit completion boundary before later stream work may be submitted. The current Vulkan backend host-waits that second value because sparse updates are rare and the public API exposes no asynchronous residency lifetime token; do not replace either edge with inferred submission order.

The same timeline-value window applies to ordinary vkQueueSubmit event waits and signals, not only vkQueueBindSparse. Validate public and internal event values through the common timeline planner. Keep a GPU-only completion watermark so the hot dispatch path can validate conservatively without calling vkGetSemaphoreCounterValue every time; query the driver only on a limit-boundary slow path. Logical stream fences may include callback-only or skipped-present work, so an embedded Vulkan wait (for example the standalone present handoff) must use the last value actually submitted to the semaphore, never the last logical fence. The image simultaneous_access flag controls the layout/access policy and is not a substitute for queue-family sharing. Imported native resources retain their external creation-time sharing contract, while aliases of a Luisa-owned native image must share its canonical per-mip layout state.

An imported Vulkan logical device must provide the exact family index for each graphics, compute, and copy queue. A VkQueue handle cannot be queried for its family after creation, so rediscovering a preferred family from the physical device is not equivalent. The graphics family must support both graphics and compute because a Luisa graphics stream records both command classes. It must also provide all three actual queue handles: calling vkGetDeviceQueue would otherwise assume queue index zero was present in the imported device's queue creation infos. Roles may share one handle; equal handles share one backend mutex because Vulkan host synchronization is per VkQueue identity. Every path that may submit through a VulkanDeviceConfigExt callback must hold that same canonical queue mutex while invoking the callback; custom submission is not outside Vulkan's host-synchronization rules. A callback that handles an event signal must still advance the backend's timeline-fence bookkeeping. Returning true from a queue semaphore callback means the requested operation was submitted on the supplied queue before return and has independent forward progress; returning true from the host-sync callback means the requested value is already complete. A callback must not report mere intent or defer submission to work that depends on the blocked backend call. borrow_command_buffer() is a one-shot ownership boundary: every non-null result must be a fresh primary command buffer in the initial state for the selected stream family, and its owner must keep both it and its pool alive until stream completion. The backend records it but never resets, frees, or recycles it; only backend-owned command buffers enter the reset/reuse pool.

An imported logical device must attest that timelineSemaphore and synchronization2 were enabled in its creation feature chain. Physical-device queries prove only support, not logical-device enablement. All optional logical-device features remain disabled on imports until the public import contract carries a corresponding enabled-feature attestation; never infer enablement from physical support. A borrowed instance must also report the effective API version used at vkCreateInstance and must be Vulkan 1.3 or newer. It is stored on that backend device and remains caller-owned; it must never be installed in the process-owned internal-instance slot. All imported Vulkan handles are borrowed, including the instance, physical device, logical device, and queues. Their complete ancestry and originating loader must outlive the Luisa Vulkan Device; backend teardown must finish before the importer invalidates any of them. Because enabled instance extensions are likewise not queryable, borrowed instances are compute-only until the import contract carries an explicit surface-extension attestation.

Volk's default dispatch mode stores instance and device entry points in process-global tables. The Vulkan backend therefore permits one live backend Device at a time and device enumeration/instance initialization must not run while it is live. Keep the Vulkan loader module resident for process lifetime, and reload the appropriate instance table whenever switching between a borrowed instance and the process-owned instance. Pin loader identity independently of DynamicModule ownership: Volk's default volkInitialize() path owns its module internally, so an empty client-side module handle does not mean that Volk is uninitialized. The first Vulkan backend operation fixes either the default loader or a custom identity made of the normalized absolute search directory plus exact library name. Reuse only that identity and fail closed on default/custom, path, or name mismatches; never replace vkGetInstanceProcAddr underneath existing instance state. Device enumeration requests the default loader, so enumerating before a custom-loader Device intentionally makes the later custom request invalid. Client code linked to a separate Volk copy must initialize that copy through VulkanDeviceConfigExt::init_volk(handler), then call volkLoadInstanceOnly(instance) from readback_vulkan_device() before using volkLoadDeviceTable(). volkInitialize() or volkInitializeCustom() alone loads only loader-level entry points and leaves vkGetDeviceProcAddr unset.

Native interop allocations must derive API creation fields and runtime wrapper metadata from one plan. In particular, image type/dimension, extent, mip count, format capabilities, and simultaneous-access policy must not diverge between the native image and the wrapper returned to LuisaCompute.

Vulkan cannot query an imported image's current layout. A new native-image identity therefore starts with every mip tracked as VK_IMAGE_LAYOUT_UNDEFINED; the importer must publish the real external layout/access state through VulkanDeviceConfigExt::before_states() before preserving or consuming existing contents. Multiple wrappers created by the same backend device for the same VkImage share one per-mip layout state, so aliases cannot publish contradictory native metadata.

Resource Handle Pattern

Resources are created as backend-specific classes and returned as opaque uint64_t handles (allocation/deallocation wrapped in the device's with_handle(...) so the native context is current). Buffer creation returns extra stride/size info, computed by the device rather than read back from the buffer:

// CUDADevice::create_buffer (src/backends/cuda/cuda_device.cpp:548)
BufferCreationInfo info{};
info.element_stride = CUDACompiler::type_size(element);      // 1 for Type::of<void>()
info.total_size_bytes = info.element_stride * elem_count;
auto buffer = with_handle([size = info.total_size_bytes, em = external_memory] {
    return em ? new_with_allocator<CUDABuffer>(reinterpret_cast<CUdeviceptr>(em), size)
              : new_with_allocator<CUDABuffer>(size);
});
info.handle = reinterpret_cast<uint64_t>(buffer);
info.native_handle = reinterpret_cast<void *>(buffer->device_address());
return info;

For most other resources, return a ResourceCreationInfo (or derived type such as ShaderCreationInfo/SwapchainCreationInfo/SparseTextureCreationInfo):

ResourceCreationInfo create_texture(...) noexcept {
    auto tex = with_handle([=] { return new_with_allocator<CUDATexture>(...); });
    return {.handle = reinterpret_cast<uint64_t>(tex),
            .native_handle = reinterpret_cast<void *>(tex->handle())};
}

Destruction recovers the typed pointer — as the shared base class when several resource classes share one handle space (CUDA buffers may be CUDABuffer or CUDAIndirectDispatchBuffer, both deleted through CUDABufferBase):

void destroy_buffer(uint64_t handle) noexcept {
    with_handle([buffer = reinterpret_cast<CUDABufferBase *>(handle)] {
        delete_with_allocator(buffer);
    });
}

Files Per Backend

Minimal

Common Helpers

<backend>_command_encoder.h/cpp (command visitor), <backend>_accel.h/cpp (ray tracing), <backend>_mesh.h/cpp, <backend>_curve.h/cpp, <backend>_procedural_primitive.h/cpp (fallback: fallback_proc_prim.*), <backend>_motion_instance.h/cpp, <backend>_bindless_array.h/cpp, <backend>_swapchain.h/cpp, sparse resources (only cuda_sparse_heap.* and vk sparse_heap.*/sparse_buffer.h exist; everything else inherits the invalid DeviceInterface default).

Naming is a convention, not a rule — match the backend you are editing:

Only common, cuda, dx, fallback, metal, tools, validation and vk ship a src/backends/<name>/xmake.lua; hip, metal4, simd and remote are CMake-only. src/backends/validation/ builds the luisa-validation-layer MODULE (a wrapping layer, not a luisa-backend-* plugin) and src/backends/tools/ is the lc_compile_builtin host tool — neither is a device backend.

Shared (src/backends/common/)

• default_binary_io.h/cpp — DefaultBinaryIO shader-cache I/O
• export_version.inl.h — Version export (#included by every backend device TU)
• command_reorder_visitor.h — CommandReorderVisitor + ReorderFuncTable
• command_reorder_switch.h — CommandReorderSwitch / LUISA_DISABLE_COMMAND_REORDER
• tile_xir_kernel.h — shared TileIR → XIR → AST fallback for create_tile_kernel
• rtx/fallback_rtx_* — software ray tracing (lc::fallback_rtx::FallbackRtxDevice, target luisa-fallback-rtx, opted into by VulkanDeviceConfigExt::use_fallback_rtx() / DirectXDeviceConfigExt::use_fallback_rtx())
• hlsl/ — shared HLSL/DXIL codegen + hlsl/builtin/ headers/bytecode
• spirv/, spirv_llvm/ — shared native and LLVM-based SPIR-V codegen
• vulkan_swapchain.cpp (public header: include/luisa/backends/common/vulkan_swapchain.h) — shared Vulkan swapchain

CMake Patterns

function(luisa_compute_add_backend name)
    cmake_parse_arguments(BACKEND "" "SUPPORT_DIR;BUILTIN_DIR" "SOURCES" ${ARGN})
    if (CMAKE_SYSTEM_NAME STREQUAL "iOS")
        set(_LUISA_BACKEND_LIBRARY_TYPE STATIC)
    else ()
        set(_LUISA_BACKEND_LIBRARY_TYPE MODULE)
    endif ()
    add_library(luisa-compute-backend-${name}
        ${_LUISA_BACKEND_LIBRARY_TYPE} ${BACKEND_SOURCES})
    target_link_libraries(luisa-compute-backend-${name} PRIVATE
        luisa-compute-ast
        luisa-compute-runtime
        luisa-compute-gui)
    if (LUISA_COMPUTE_ENABLE_DSL)
        target_link_libraries(luisa-compute-backend-${name} PRIVATE luisa-compute-dsl)
    endif ()
    add_dependencies(luisa-compute-backends luisa-compute-backend-${name})
    set_target_properties(luisa-compute-backend-${name} PROPERTIES
        UNITY_BUILD ${LUISA_COMPUTE_ENABLE_UNITY_BUILD}
        DEBUG_POSTFIX ""
        OUTPUT_NAME luisa-backend-${name})
    install(TARGETS luisa-compute-backend-${name}
        LIBRARY DESTINATION ${CMAKE_INSTALL_BINDIR}
        ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
        RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR})
    if (BACKEND_SUPPORT_DIR)
        add_custom_target(luisa-compute-backend-${name}-copy-support ALL ...)
    endif ()
endfunction()

Fallback (LLVM + Embree):

find_package(LLVM CONFIG)
find_package(embree CONFIG)
if (LLVM_FOUND AND embree_FOUND)
    luisa_compute_add_backend(fallback SOURCES fallback_device.cpp ...)
    target_link_libraries(luisa-compute-backend-fallback PRIVATE
        luisa-compute-xir
        luisa-compute-vulkan-swapchain
        embree)
    luisa_compute_link_llvm_into_backend(fallback REQUIRED
        COMPONENTS core executionengine support orcjit nativecodegen irreader passes analysis coroutines)
endif()

CUDA advanced: luisa_embed_device_lib, device runtime embedding, optional nvCOMP/NVTT/OIDN, standalone NVRTC compiler target.

XMake Note

The project also supports XMake builds. Backend xmake.lua files create shared targets named lc-backend-<name> with set_basename("luisa-backend-<name>") (dx, vk, cuda, metal, fallback), link lc-runtime plus lc-vstl/lc-tile/lc-hlsl-codegen (and lc-spirv/lc-spirv-llvm for the Vulkan codegen routes), and optionally depend on lc-vulkan-swapchain and backend-specific helpers. There is no lc-ir target: XIR is not a standalone xmake target. The validation layer is the separate lc-validation-layer target (set_basename("luisa-validation-layer")).

Codegen Pipeline

DSL Function (an ast::Function) / xir::Module / tile::Function
│
▼
┌─────────────────────────────────┐
│ Backend Codegen │  → native shader source (CUDA C++, HLSL, MSL, SPIR-V)
└───────────────┬─────────────────┘
▼
┌─────────────────────────────────┐
│ Native Compiler │  NVRTC, DXC, Metal, SPIR-V tools
└───────────────┬─────────────────┘
▼
┌─────────────────────────────────┐
│ Shader Object │  PTX, DXIL, metallib, SPIR-V binary
│ (BinaryIO cache) │  Cached via BinaryIO
└─────────────────────────────────┘

create_shader only receives a DSL Function; every other IR above is a backend-internal route. XIR consumers are not Metal-only: CUDA (src/backends/cuda/llvm_codegen/), HIP (src/backends/hip/llvm_codegen/), Metal4 (src/backends/metal4/llvm_codegen/) and the shared src/backends/common/spirv/ codegen all take XIR. The Vulkan route is a mutually exclusive build switch (src/backends/CMakeLists.txt:3-8 FATAL_ERRORs if both are ON): LUISA_COMPUTE_ENABLE_VK_XIR_SPIRV (CMake default ON) → native XIR→SPIR-V in src/backends/common/spirv/, or LUISA_COMPUTE_ENABLE_VK_AST_LLVM_SPIRV (CMake default OFF) → AST→LLVM→SPIR-V in src/backends/common/spirv_llvm/. xmake counterparts: lc_vk_backend_use_xir_spirv, lc_vk_backend_use_ast_llvm_spirv.

Key Design Decisions

1. Backend Boundary — Desktop backends are shared plugins; iOS statically links the selected backend but keeps the same DeviceInterface boundary
2. Handle-based Resources — Opaque uint64_t handles for ABI stability
3. Visitor Pattern — Double dispatch via MutableCommandVisitor for type-safe command handling
4. Version Checking — Strict match between runtime and backend prevents ABI mismatches
5. Validation Layer — Optional wrap-around validation at device creation (or via LUISA_ENABLE_VALIDATION), loaded as the separate luisa-validation-layer module and wrapping the backend Device::Handle
6. BinaryIO — Pluggable shader caching through DeviceConfig::binary_io; backends fall back to DefaultBinaryIO (src/backends/common/default_binary_io.h)
7. Tile kernels are opt-in — create_tile_kernel defaults to an invalid ShaderCreationInfo; backends either implement a native Tile compiler or reuse src/backends/common/tile_xir_kernel.h