inference/research
drafthigh · 0.85updated 2026-10-05

KleidiAI README (Arm Kleidi Libraries, the primary repository on GitLab)

repo
This record is a draft. Not yet verified. Provisional. Treat every number on this page as provisional.

Fields

id
sources/w7k-kleidiai-readme
Publisher
Arm
Kind
repo
Accessed
2026-10-05
notes
PRIMARY SOURCE for the Arm CPU kernel path, and the definitional text for what a 'micro-kernel' (ukernel) is here: 'a near-minimum amount of software to accelerate a given ML operator with high performance', designed 'to give the flexibility to process also a portion of the output tensor' so the caller controls dispatch and output slicing. The stated key features are the deployment-relevant part: no dependencies on external libraries, no dynamic memory allocation, no memory management, no scheduling, stateless stable API. Supported instruction surface: Advanced SIMD, SVE, SME, SME2, with specialised micro-kernels for FEAT_DotProd and FEAT_I8MM. Variants are keyed by computational parameter, Arm technology (e.g. Arm Neon) and architecture feature, and each variant is a standalone .c/.h pair depending only on kai_common.h. Two hazards stated by Arm itself: the SME/SME2 micro-kernels 'require compiler support to generate SME ABI-compliant code', this 'is not automatically detected, and you need...' to work, so a toolchain gap surfaces as an ABI problem; and the experimental directory 'holds experimental features. Do not expect production quality. Note that API might change at any point in time'. Licensing is unusual and worth recording: 'KleidiAI is distributed under the software licenses in LICENSES directory' - there is no single root SPDX id, and the direct LICENSES/MIT.txt and LICENSE.txt paths 404 from outside. Cadence is stated: two releases per month. GitHub mirror metadata (GitHub API, 2026-10-05): created 2025-01-23, pushed 2026-10-05, 187 stars, latest release v1.31.0, no license field.

Notes

PRIMARY SOURCE for the Arm CPU kernel path, and the definitional text for what a 'micro-kernel' (ukernel) is here: 'a near-minimum amount of software to accelerate a given ML operator with high performance', designed 'to give the flexibility to process also a portion of the output tensor' so the caller controls dispatch and output slicing. The stated key features are the deployment-relevant part: no dependencies on external libraries, no dynamic memory allocation, no memory management, no scheduling, stateless stable API. Supported instruction surface: Advanced SIMD, SVE, SME, SME2, with specialised micro-kernels for FEAT_DotProd and FEAT_I8MM. Variants are keyed by computational parameter, Arm technology (e.g. Arm Neon) and architecture feature, and each variant is a standalone .c/.h pair depending only on kai_common.h. Two hazards stated by Arm itself: the SME/SME2 micro-kernels 'require compiler support to generate SME ABI-compliant code', this 'is not automatically detected, and you need...' to work, so a toolchain gap surfaces as an ABI problem; and the experimental directory 'holds experimental features. Do not expect production quality. Note that API might change at any point in time'. Licensing is unusual and worth recording: 'KleidiAI is distributed under the software licenses in LICENSES directory' - there is no single root SPDX id, and the direct LICENSES/MIT.txt and LICENSE.txt paths 404 from outside. Cadence is stated: two releases per month. GitHub mirror metadata (GitHub API, 2026-10-05): created 2025-01-23, pushed 2026-10-05, 187 stars, latest release v1.31.0, no license field.

Cross-references

This record references no other records.