libclc

libclc/ is libclc — an implementation of the OpenCL C standard library. It compiles OpenCL kernel source into bitcode that is later linked with user-written kernels to provide the OpenCL built-ins (get_global_id, barrier, mad, dot, math functions, atomics, etc.).

Purpose

OpenCL kernels rely on a large surface of host-provided built-ins. libclc supplies that surface as portable bitcode that can be combined with hardware-specific code-gen output. It targets AMDGPU, NVPTX, SPIR-V, and a "generic" software target. From libclc/README.md:

libclc is intended to be used with the Clang compiler's OpenCL frontend.

Directory layout

libclc/
├── clc/            # The CLC (Common Library for OpenCL) implementation — most of the math
├── opencl/         # OpenCL-specific wrappers and headers
├── amdgcn/, amdgcn-amdhsa/, ptx/, ptx-nvidiacl/, spirv/, spirv64/, generic/
│                  # Per-target overlay directories
├── cmake/
├── test/
├── utils/          # gen_convert.py, gen_format.py, prepare-builtins
├── docs/
└── README.md

The library is built by compiling OpenCL C source to bitcode using the just-built clang. The output is libclc-<target>-<vendor>.bc, suitable for linking into compiled OpenCL kernels.

How it works

graph LR
    src["libclc OpenCL C source<br/>(libclc/clc/, opencl/)"] --> clang[clang -emit-llvm-bc<br/>--target=spirv64/...]
    clang --> bc[Per-target bitcode]
    bc --> link[Linked with user kernel]
    link --> kernel[Final kernel module]

The build invokes clang on each source file with the appropriate target triple, then concatenates and finalizes the bitcode. The result is shipped to OpenCL runtimes (driver-side) or to compilers that target SPIR-V.

What ships

A non-exhaustive list:

• Math (sin, cos, pow, sqrt, mad, fma, etc.) — in IEEE-754-correct and "native_" fast variants
• Vector arithmetic
• Synchronization (barrier, mem_fence)
• Atomics
• Conversion (convert_int4, etc.) — extensively templated
• Geometric, common, integer, relational built-ins
• Workgroup queries (get_global_id, get_local_size, ...)

Integration points

• Clang — produces SPIR-V or per-target bitcode that consumes libclc's symbols.
• AMDGPU / NVPTX backends in llvm/lib/Target/ — the codegen targets.
• OpenCL drivers consume the resulting bitcode at runtime or at JIT time.

Entry points for modification

• Adding a built-in: implement under libclc/clc/ (or libclc/opencl/ for OpenCL-specific shapes), update the per-target manifests, and add a test.
• Updating math precision: math sources are in libclc/clc/lib/math/. The precision policy is documented inline in each file.

Reference

• libclc/README.md
• libclc documentation
• OpenCL C 3.0 specification