Xeon Phi toolchain

An experimental LLVM/Clang toolchain for Knights Corner (KNC / K1OM), with native C, C++, and Rust support, runtime math helpers, regression tests, and llama.cpp kernels for KNC and Knights Landing (KNL).

The LLVM patch adds KNC support to LLVM's X86 backend. Rust uses a pinned, unmodified compiler built against that LLVM. KNL uses LLVM's existing native x86-64 backend with KNL-specific build and runtime settings.

This project does not emulate SSE/AVX instructions or expose them as KNC features. Its capability matcher describes native K1OM operations and their requirements; applications supply their own kernels.

Quick start

Use an x86-64 Linux build host with Python 3.12+, Git, CMake, Ninja, Clang, Make, and an independently installed MPSS SDK for Linux/KNC applications. LLVM-only code generation and host contract tests do not require the SDK. Allow substantial disk space and RAM for LLVM/Rust builds; PHI_JOBS controls parallelism and defaults to at most eight jobs.

cd xeon-phi-toolchain
python3 scripts/phi.py doctor
python3 scripts/phi.py test-host
python3 scripts/phi.py fetch llvm
python3 scripts/phi.py build-llvm
python3 scripts/phi.py test-codegen

export PHI_SDK_ROOT=/path/to/your/mpss-sdk
python3 scripts/phi.py build-runtime
bin/clang-knc -O2 examples/hello.c -o build/hello-knc

The SDK root must contain sysroots/k1om-mpss-linux and sysroots/x86_64-mpsssdk-linux. No SDK binaries, headers, firmware, device drivers, model files, credentials, or private connection settings are included. The C compiler produces K1OM executables to run on a KNC card, not the build host.

For modern C++:

python3 scripts/phi.py build-cxx
bin/clang++-knc-modern -std=c++17 -O2 examples/hello.cpp -o build/hello-cxx

For native Rust:

rustup toolchain install nightly-2026-09-01 --profile minimal --component rust-src
python3 scripts/phi.py fetch rust
python3 scripts/phi.py build-rust
bin/cargo-knc build --manifest-path examples/rust-std-demo/Cargo.toml --release

Standard Rust portable SIMD is supported through core::simd / std::simd on the pinned nightly, without custom vector wrappers. The tested matrix spans 13 numeric element types and ten lane counts, including odd-sized vectors. Build the included example:

bin/cargo-knc build --manifest-path examples/rust-portable-simd/Cargo.toml \
  --bin knc-portable-simd --release

The K1OM executable is build/cargo/knc-native/release/knc-portable-simd; copy it to a KNC card and run it there. See portable SIMD for standard-API source examples, the complete test command and the hardware-validated scope.

See building for prerequisites, SDK relocation, CMake, external build paths, Rust's experimental vectorizer, and troubleshooting.

llama.cpp

python3 scripts/phi.py fetch llama
# KNC: requires the SDK and the target C++ library built above.
python3 scripts/phi.py build-llama --arch knc --profile default
# KNL: native Linux x86-64; does not require the MPSS SDK.
python3 scripts/phi.py build-llama --arch knl --profile repack

The executables land in build/llama/<architecture>/<profile>/bin/. Supply your own GGUF model. See llama.cpp usage and measured performance.

Repository layout

Path	Contents
patches/	Complete patches against pinned LLVM and llama.cpp commits
sources.lock.json	Upstream revisions, patch checksums, GCC source checksum
bin/, scripts/, cmake/, targets/	Portable compiler wrappers and build tools
runtime/	Native math/ABI helpers and attributed musl-derived sources
include/k1om/, capabilities/	Metadata-only capability matcher and snapshot
tests/	Compiler, ABI, numerical, operator, and release-tool regressions
examples/	C/C++ examples and Rust crates
docs/, benchmarks/	Contracts, release guidance, and historical measurements

Downloaded sources, generated records, builds, and local SDK copies are ignored under sources/ and build/. Build scripts do not read private lab evidence.

Status and release scope

This is a source release, with bounded hardware validation, not an Intel product or a drop-in replacement for Intel's complete compiler/offload stack. Unsupported instructions remain unsupported. Floating-point contracts and known limits are described in architecture, validation, and the runtime documentation.

The historical KNC 0.8B/4B output divergence from a host reference has not been established as a general model-quality equivalence result. Short inference benchmarks do not establish long-context behavior or a hardware speed ceiling.

Release steps are in docs/releasing.md. The source archive excludes Git history and all private SDK/build artifacts. No release is uploaded by the build or packaging commands.

License

Toolchain work uses Apache-2.0 WITH LLVM-exception, as in LICENSE. Other upstream licenses and notices remain applicable to their components; see NOTICE and THIRD_PARTY.md.