Native Shader Injection (NativeShaderExt)

NativeShaderExt (in include/luisa/backends/ext/native_shader_ext.h, the canonical reference) hands native source (HLSL / GLSL / CUDA C++) to the backend's own compiler, reflects its bindings, creates a backend compute-shader instance, and dispatches it from a Stream. It bypasses the DSL/AST/XIR path entirely.

• Compute only, JIT only (no ShaderSerializer/AOT, no LUISA_DUMP_SOURCE).
• Only buffers and uniform blocks; textures/samplers/accel reflect but are rejected at load().
• The per-argument Usage you declare is the command-reorder synchronization contract.
• Destroy a NativeShader before the Device that created it.

Backend support matrix

NativeShaderLanguage	dx	vk	cuda
HLSL	DXIL via DXC	SPIR-V via DXC	rejected
GLSL	rejected	SPIR-V via glslang	rejected
CUDA_NVRTC	rejected	rejected	PTX via NVRTC

compile() fails closed (ok() false, binary empty, error names the reason).

Workflow

#include <luisa/backends/ext/native_shader_ext.h>
// ... runtime/context.h, runtime/device.h, runtime/stream.h, runtime/command_list.h ...

auto ext = device.extension<NativeShaderExt>();
if (ext == nullptr) { /* backend has no NativeShaderExt */ }

// 1. compile source -> bytecode + reflection
NativeShaderCompileInfo info;
info.language           = NativeShaderLanguage::HLSL;  // or GLSL / CUDA_NVRTC
info.source             = source;                      // string_view: must outlive compile()
info.entry_point        = "CSMain";                    // "main" for GLSL
info.push_constant_size = 2u * sizeof(float);          // uniform bytes; 0 = reflect
// info.block_size for CUDA when the kernel has no __launch_bounds__
// info.source_type = NativeShaderSourceType::FilePath; info.include_dirs = {...}
// info.shader_model, info.optimize, info.enable_fast_math, info.enable_debug_info
auto result = ext->compile(info);
if (!result.ok()) { /* result.error */ }

// 2. load bytecode -> backend shader instance + resolved binding table
auto metadata = ext->load(result);                     // optional usage_override span
if (!metadata.valid()) { /* contract violation, LUISA_WARNING names it */ }

// 3. RAII owner (destroys the instance before the device)
NativeShader shader{*ext, std::move(metadata)};
// or one-shot: auto shader = ext->create_shader(info);

// 4. bind + dispatch
auto launcher = shader.launcher();                     // handle + block size + bindings
launcher.add_buffer_by_index(0u, src.view(), Usage::READ)
        .add_buffer_by_index(1u, dst.view(), Usage::WRITE)
        .add_uniform(k)
        .add_uniform(c);
auto cmd = std::move(launcher).build(uint3{thread_count, 1u, 1u});
stream << std::move(cmd) << synchronize();

build(thread_count) takes the total thread count (grid × block), not a grid size; the block size comes from the launcher (the shader's reflected workgroup size). Use launcher.validate() for a soft error string; build() asserts the same plan.

Shader sources

HLSL on dx — free entry-point name; [numthreads] drives the reflected block size:

StructuredBuffer<float> src : register(t0);            // SRV
RWStructuredBuffer<float> dst : register(u0);          // UAV
cbuffer Uniforms : register(b0) { float k; float c; }; // uniform block (root constants)
[numthreads(64, 1, 1)]
void CSMain(uint3 tid : SV_DispatchThreadID) { dst[tid.x] = src[tid.x] * k + c; }

The b0 cbuffer is the launcher's uniform block, fed by add_uniform values and removed from load()'s binding table (compile() still reports it). Root-constant limit: 64 × 32-bit; push_constant_size must be a multiple of 4.

HLSL on vk — SPIR-V decorations are the authoritative reflection table:

struct Uniforms { float k; float c; };
[[vk::binding(0, 0)]] StructuredBuffer<float> src;
[[vk::binding(1, 0)]] RWStructuredBuffer<float> dst;
[[vk::push_constant]] ConstantBuffer<Uniforms> uniforms;
[numthreads(64, 1, 1)]
void CSMain(uint3 tid : SV_DispatchThreadID) { dst[tid.x] = src[tid.x] * uniforms.k + uniforms.c; }

Plain registers also work (t<n>/u<n> → binding n, space 0); use [[vk::binding(set, binding)]] for a specific set/binding.

GLSL on vk — single entry point named main:

#version 450
layout(local_size_x = 64) in;
layout(set = 0, binding = 0) readonly buffer A { float a[]; } src;   // SRV
layout(set = 0, binding = 1) buffer B { float b[]; } dst;            // UAV
layout(push_constant) uniform Push { float k; float c; } uniforms;
void main() { uint i = gl_GlobalInvocationID.x; dst.b[i] = src.a[i] * uniforms.k + uniforms.c; }

readonly → StructuredBuffer, otherwise RWStructuredBuffer. layout(push_constant) is the uniform block (size read back from the module when push_constant_size == 0). Push-constant size must be a multiple of 4. For #include, add #extension GL_GOOGLE_include_directive : require after #version.

CUDA C++ on cuda — the kernel signature is the binding declaration:

extern "C" __global__ void scale(const float *src, float *dst, float k, float c) {
    auto i = blockIdx.x * blockDim.x + threadIdx.x;
    dst[i] = src[i] * k + c;
}

• extern "C" keeps the name unmangled; __global__ must not be templated/qualified.
• Pointer params are buffer bindings in order, each bound as the buffer's 64-bit device address: const T * → StructuredBuffer/READ; T * → RWStructuredBuffer/READ_WRITE.
• Non-pointer params are scalar kernel params fed by add_uniform in order; there is no push-constant block, "uniform bytes" is just their total size (push_constant_size cross-checks, 0 = reflect it). A by-value struct is valid (add_uniform uses sizeof).
• entry_point selects the __global__ (default main = "the only kernel").
• block_size must be supplied unless the kernel declares __launch_bounds__(N).

Reflection & binding

result.bindings / shader.bindings() are reported in canonical order (space_index, register_index, kind); the launcher's positional overload and reorder order follow it too.

struct NativeShaderResourceBinding {
    NativeShaderResourceKind kind;   // StructuredBuffer, RWStructuredBuffer, ConstantBuffer, ...
    uint register_index;             // HLSL register / GLSL binding
    uint space_index;                // HLSL space / GLSL set
    uint array_size;                 // 1 == non-array
    Usage usage;                     // effective usage
    uint32_t stride;                 // element stride when known
    uint32_t size_bytes;             // constant-buffer / block size when known
};

Supported kinds at load(): dx ConstantBuffer, StructuredBuffer/RWStructuredBuffer, ByteAddressBuffer/RWByteAddressBuffer; vk the same plus TypedBuffer/RWTypedBuffer; cuda buffers only. Default usages: CBV/UBO and SRV classes → READ; UAV classes → READ_WRITE.

The three bind overloads (prefer index when HLSL namespaces collide — register(t0) and register(b0) are both bind point 0):

• add_buffer_by_index(index, view, usage) — index into the reflection table (recommended).
• add_buffer(register, space, view, usage) — usage disambiguates a shared (register, space).
• add_buffer(view, usage) — positional, canonical order.

Every reflected binding must be supplied exactly once; missing/duplicate/extra args and null handles are rejected by validate()/build().

Launcher & dispatch

template<typename T> NativeShaderLauncher &add_uniform(const T &value);
NativeShaderLauncher &add_uniform(const void *data, size_t size, size_t alignment);
auto cmd = std::move(launcher).build(uint3{thread_count, 1u, 1u});
// or: std::move(launcher).build(shader_handle, thread_count, block_size);

Uniforms are packed into a blob (offsets shifted past the argument header; alignment clamped to ≤ 16). NativeShaderDispatchCommand is a CustomDispatchCommand (CustomCommandUUID::NATIVE_SHADER_DISPATCH, 0x0600) tagged StreamTag::COMPUTE; enqueue it like any other command.

Usage contract

The declared per-argument usage is the synchronization contract: READ (mergeable with other readers), WRITE (exclusive), READ_WRITE (both); NONE is never valid. build() (and the backend at load()) cross-check it against the reflected class:

• An SRV/CBV bound as WRITE is always rejected.
• A UAV bound as READ is rejected by default; allow it with launcher.set_allow_usage_override(true).

Declaring READ for a resource the shader writes silently breaks synchronization. load(result, usage_override) can override usages up-front (one entry per reflected binding; a const T * binding cannot be declared writable).

Include dirs & file sources

NativeShaderCompileInfo carries include_dirs; a FilePath source also resolves relative to its own directory:

info.source_type  = NativeShaderSourceType::FilePath;
info.source       = luisa::string_view{source_path};   // must outlive compile()
info.include_dirs = { scratch_dir };

Includes go through DXC / glslang's includer / NVRTC -I respectively. A bogus include directory fails closed on every route: the standalone luisa_nvrtc child prints the compiler log to its stderr, exits non-zero (it never aborts), and the parent surfaces the log plus the exit code in result.error.

Lifetime

{
    NativeShader shader{*ext, std::move(metadata)};   // RAII; move-only
    ...
}   // destroy_shader() here
// or: ext->destroy_shader(handle); / shader.reset();

Synchronize dispatches before destruction. A leftover instance is released at extension teardown with a warning, never silently.

Pitfalls

• info.source / info.entry_point are string_view — the memory must outlive compile().
• build(thread_count) is a thread count, not a grid size (1024 threads @ block 64 → build(uint3{1024,1,1})).
• On dx, bind by the load()/metadata.bindings table (compile() reports the raw DXC table including the b0 cbuffer).
• Block size must match the reflected workgroup size, or build() asserts. dx/vk derive it from [numthreads]/local_size_*; cuda requires it (or __launch_bounds__).
• UAV-as-READ needs set_allow_usage_override(true); SRV-as-WRITE is never allowed.
• Destroy the shader before the device and before its dispatches are torn down.