HLSL Code Generation

Codegen utilities using vstd::StringBuilder and formatting helpers.

StringBuilder

#include <luisa/vstl/string_builder.h>

vstd::StringBuilder builder;
builder.clear();
size_t sz = builder.size();
vstd::string_view view = builder.view();
char* data = builder.data();
char& ch = builder[0];      // operator[] (both const and non-const)
builder.reserve(1024);
builder.resize(64);

Appending

str.append("Hello");
str += "World"sv;
str.append(' ');
str << '\n';
str.append(other_builder);
str << "func" << '(' << ")";

Numbers

vstd::to_string(42, str); // int (faster than format)
vstd::to_string(3.14f, str); // float
vstd::string s = vstd::to_string(42); // single-arg form -> vstd::string
str << luisa::format("{}", 42);
str << luisa::format("{}, {}!", "Hello", "World");
str << luisa::format("{:016X}", hash); // same as luisa::hash_to_string(hash)

String View Literals

str += "void"sv; // no allocation

<luisa/vstl/common.h> already enables the literals: it includes vstl/vstring.h, which includes vstl/string_hash.h - and that header has using namespace std::literals; at (global) namespace scope. So generator .cpp files write "..."sv without their own using namespace std::string_view_literals;.

Code Generation Patterns

Function Declaration

// real source: CodegenUtility members, codegen_utils/function_codegen.cpp
void GetFunctionDecl(Function func, vstd::StringBuilder &str) {
    vstd::StringBuilder data;
    if (func.return_type()) CodegenUtility::GetTypeName(*func.return_type(), data, Usage::READ);
    else data += "void"sv;
    data += " "sv;
    CodegenUtility::GetFunctionName(func, data);
    data += '(';
    for (auto &&arg : func.arguments()) {
        Usage usage = func.variable_usage(arg.uid());
        CodegenUtility::GetTypeName(*arg.type(), data, usage);
        data << ' ';
        vstd::StringBuilder varName;
        CodegenUtility::GetVariableName(func, arg, varName);
        data << varName << ',';
    }
    if (!func.arguments().empty()) data[data.size() - 1] = ')';
    else data += ')';
    str << '\n' << data;
}

Template Parameters

str << "template<"sv;
for (uint64 i = 0; i < tempIdx; ++i) {
    str << "typename T"sv; vstd::to_string(static_cast<int64_t>(i), str); str << ',';
}
if (tempIdx > 0) *(str.end() - 1) = '>';

Call Expression

str << "func_name"sv << '(';
if (!args.empty()) {
    for (size_t i = 0; i < args.size() - 1; ++i) { args[i]->accept(vis); str << ','; }
    args.back()->accept(vis);
}
str << ')';

Type-Aware Generation

// buffer / texture arguments become template parameters T0, T1, ... (GetFunctionDecl)
if (t->is_texture() || t->is_buffer()) { str << 'T'; vstd::to_string(tempIdx++, str); }
// GetTypeName: element type name followed by the dimension
else if (t->is_vector()) { GetTypeName(*t->element(), str, usage); vstd::to_string(t->dimension(), str); }
// a 3-row matrix is stored as float3x4 (columns are padded to 4)
else if (t->is_matrix()) { GetTypeName(*t->element(), str, usage); vstd::to_string(t->dimension(), str); str << 'x'; vstd::to_string(t->dimension() == 3 ? 4 : t->dimension(), str); }
// matrix element of a buffer/texture is wrapped: "_WrappedFloat" + n + 'x' + n,
// with n = vstd::to_string(t->element()->dimension()) (type_system.cpp, BUFFER case)

RAII Wrapping

// StringStateVisitor::Scope - the real RAII wrapper of the generator
struct Scope {
    StringStateVisitor *self;
    explicit Scope(StringStateVisitor *self);            // self->str << "{\n"sv
    ~Scope() { self->str << "}\n"sv; }
};
Scope scope{this};                                       // braces emitted automatically

// non-RAII case (balanced open/close emitted by hand, e.g. aliased struct casts):
AliasedToOrigin(expr->type(), str);
str << '(';
// ... inner content ...
str << ')';

Trailing Comma Fix

str << '(';
for (auto &&item : items) str << item << ',';
if (!items.empty()) str[str.size() - 1] = ')';
else str << ')';

Iteration & Hash

for (char c : str) {}
str.erase(str.begin() + 5);
vstd::hash<vstd::StringBuilder> hasher; size_t h = hasher(builder);
if (builder1 == builder2) {}
if (builder1 == "literal"sv) {}

Best Practices

1. Use "text"sv literals to avoid allocations
2. vstd::to_string() for integers (faster than luisa::format())
3. builder.reserve(1024) when size known
4. operator<< for chaining, operator+= for single items
5. Check size() before modifying last character

HLSL Builtins

Located in src/backends/common/hlsl/builtin/. Access via:

// included as "../builtin/hlsl_builtin.hpp" by codegen_utils/entry_points.cpp,
// under `#ifndef LC_NO_HLSL_BUILTIN`
auto header = lc_hlsl::get_hlsl_builtin("hlsl_header"); // {void const* ptr; size_t size;}
std::string_view code(static_cast<const char*>(header.ptr), header.size);

Generator and backend code never calls get_hlsl_builtin directly - it goes through CodegenUtility::ReadInternalHLSLFile(name) (codegen_utils/entry_points.cpp), which returns the same bytes as a vstd::string_view. The DX/VK shader loaders also use it for pre-compiled blobs: CacheType::Internal (src/backends/dx/Shader/ComputeShader.cpp) and SerdeType::kBuiltin (src/backends/vk/shader_serializer.cpp) try the embedded asset first, then fall back to BinaryIO::read_internal_shader.

With LC_NO_HLSL_BUILTIN defined (CMake LUISA_COMPUTE_VULKAN_ENABLE_DXC_COMPATIBILITY=OFF, xmake: lc_vk_backend_enable_dxc_compatibility not set) get_hlsl_builtin is stubbed out and returns {}, so no builtin is available.

Header Files (.bytes)

Keys are the strings passed to ReadInternalHLSLFile / get_hlsl_builtin. The backing file is builtin/<key> with .bytes appended unless the key already carries the extension (three keys below keep .bytes in the lookup string), and the BC6/BC7 encoder fragments are extensionless files.

Key	Description
hlsl_header	Main HLSL header
hlsl_header_fallback	Fallback header (noRegister path, e.g. DX without DXC)
spv_alias	SPIR-V aliases (SPIR-V route only)
fallback_rtx_header	Software ray-tracing traversal (DeviceConfigExt::use_fallback_rtx())
dx_linalg	Linear algebra utils (DXIL)
vk_linalg	Linear algebra utils (SPIR-V)
raytracing_header	Ray tracing
raytracing_motion_header	Ray tracing motion blur
tex2d_bindless / tex3d_bindless	Bindless textures
compute_quad	Compute quad ops
determinant / inverse	Matrix ops
indirect	Indirect dispatch/draw
resource_size	Resource queries
accel_header	Acceleration structures
copy_sign	Sign ops
bindless_common	Bindless utils
auto_diff	Autodiff
reduce	Parallel reduction
bindless_upload.bytes	Bindless upload source -> compiled to load_bdls.dxil
accel_process.bytes	Acceleration-structure processing source -> set_accel4.dxil
accel_process_vk_motion.bytes	Acceleration-structure processing with motion blur
bc6_header / bc6_encode_block / bc6_trymode_g10cs / bc6_trymode_le10cs	BC6 encoder HLSL fragments
bc7_header / bc7_encode_block / bc7_trymode_02cs / bc7_trymode_137cs / bc7_trymode_456cs	BC7 encoder HLSL fragments

builtin/bindless_upload, builtin/bindless_upload_vk and builtin/load_bdls_dxil are extensionless files that are not embedded (neither the CMake list nor the xmake globs match them) and have no dict key - the registered sources are bindless_upload.bytes and load_bdls.dxil.

DXIL Files (.dxil)

Pre-compiled builtin shader artifacts (ShaderSerializer containers), keyed with their full file name.

Key	Description
accel_process_vk.dxil	Vulkan acceleration structures
load_bdls.dxil / load_bdls_vk.dxil	Bindless loading
set_accel4.dxil	Set acceleration structure
bc6_encodeblock.dxil / bc6_trymodeg10.dxil / bc6_trymodele10.dxil	BC6 compression
bc7_encodeblock.dxil / bc7_trymode02.dxil / bc7_trymode137.dxil / bc7_trymode456.dxil	BC7 compression

*_vk.dxil are legacy HLSL-to-SPIR-V (dxc -spirv) containers: the current Vulkan builtin kernels load .spv artifacts embedded as luisa_compute_vk_builtin_* (src/backends/vk/builtin_kernel.cpp), and scripts/verify_compile_builtin.py treats load_bdls_vk.dxil as a stale v2 container.

Adding Builtins

1. Add .bytes or .dxil to src/backends/common/hlsl/builtin/ - e.g. my_helper.bytes. A .bytes asset is hand-written HLSL text that is edited in place (nothing generates it from a .hlsl).
2. Declare the symbol luisa_embed_device_lib / bin2obj produces for that file (the file stem with its extension turned into _, because of --preserve-ext), inside extern "C" { ... } in builtin/hlsl_builtin.hpp: LC_HLSL_DECL_VARNAME(my_helper_bytes)
3. Register the lookup key: LC_HLSL_INSERT_VARNAME(my_helper_bytes, "my_helper")
4. CMake: append my_helper.bytes to LUISA_COMPUTE_HLSL_DEVICE_LIB_SOURCES in src/backends/common/hlsl/CMakeLists.txt - the list is explicit, new files are not globbed.
5. xmake: nothing to do for .bytes/.dxil (src/backends/common/hlsl/xmake.lua globs builtin/*.bytes and builtin/*.dxil); extensionless assets need their own add_files(...) entry with {rules = "utils.bin2obj"}.

Embedding: xmake uses the utils.bin2obj rule plus add_defines('LUISA_BIN_2_OBJ'), so the _binary_<name>_start/_end symbols are produced per object file. CMake runs the host tool luisa_embed_device_lib (utils/embed_device_lib.cpp) with --unsigned ... --preserve-ext in the luisa-compute-backend-hlsl-embed target to regenerate builtin/hlsl_builtin_embed.cpp (gitignored). That target runs the tool twice - once to write, once with --check - and luisa-compute-backend-hlsl-embed-check is the standalone check-only variant; luisa-compute-hlsl-codegen depends on the embed target. --preserve-ext is why hlsl_header.bytes becomes the symbol hlsl_header_bytes.

Regenerating a compiled .dxil builtin from its .bytes source (same compiler as the runtime, src/backends/tools/main.cpp):

xmake build lc_compile_builtin
xmake run lc_compile_builtin dx src/backends/common/hlsl/builtin/bindless_upload.bytes \
  src/backends/common/hlsl/builtin/load_bdls.dxil --name load_bdls.dxil --verify
xmake run lc_compile_builtin vk <input-shader> <output> # dxc -spirv, VK v10 container
xmake run lc_compile_builtin dx|vk inspect <artifact> # decode contract fields

--shader-model <n> is a packed model (62 == cs_6_2, default 62); DX runtime selection is kShaderModel = 65, kHighShaderModel = 66 (allowed warp size), kTensorShaderModel = 69 (cooperative ops / 8-bit, src/backends/dx/DXApi/LCDevice.cpp). Dumped HLSL can also be compiled by hand: python scripts/compile_dxil.py [source] [--shader-model 6_5|6_6|6_9] [-E main] [--dxc <dxc.exe>] [-o out.dxil] (dxc is auto-searched: D:\DirectXShaderCompiler\build\bin\dxc.exe, then PATH, then bin/debug/dxc.exe). scripts/compile_builtin.py dx|vk <output> [--name <kernel>] [--no-build] is a thin wrapper that first builds lc_compile_builtin in release.

Codegen Debug

Set env LUISA_DUMP_SOURCE=1 to dump generated HLSL to per-shader files.
Output: hlsl_output_<shader_name>.hlsl in the working directory.

Naming priority depends on the backend and shader path:

DX compute / raster / save paths:

1. ShaderOption::name / fileName — user-provided name
2. Function::name() — kernel/callable debug name
3. Function::hash() formatted as hex — fallback (e.g. hlsl_output_a1b2c3d4.hlsl)

VK compute on the HLSL-to-SPIR-V compatibility route (ComputeShader::compile, src/backends/vk/compute_shader.cpp):

1. file_name — ShaderOption::name, else the cache name <md5>.spv when enable_cache
2. MD5 of the generated HLSL — fallback

VK ray tracing (RayTracingShader::compile, src/backends/vk/rt_shader.cpp): file_name → code_md5 → literal rt_unknown.

VK raster (VkRasterExt): ShaderOption::name is always required (LUISA_ASSERT(!option.name.empty(), "Raster shader name must not be empty.")), so the dump file uses that name.

VK user compute normally goes through native XIR-to-SPIR-V; it falls back to this HLSL route only for features the native codegen cannot express - native_include, device printing, async_copy/pipeline commit+wait, motion blur, or software (fallback-RTX) ray tracing (lc::vk::detail::plan_user_compute_codegen_route in src/backends/vk/user_compute_codegen_route.h). LUISA_VULKAN_REQUIRE_NATIVE_XIR_SPIRV=1 forbids the HLSL fallback, and LC_NO_HLSL_BUILTIN (no DXC compatibility) removes the HLSL route entirely - then those features must be compiled by LUISA_XIR_TO_SPIRV or LUISA_AST_LLVM_TO_SPIRV.

Per-shader dumps are written with "wb" (overwrite). The one exception is the VK option.compile_only path (src/backends/vk/device.cpp), which still appends everything to a single hlsl_output.hlsl with "ab".