KleidiAI follows the Semantic Versioning specification for releases.
- Added demonstration of integration using CMake in F16 Arm® Neon™ matrix multiplication example.
- Update FP16 example to use NHWC input
- Fixes:
- Fix build error on MSVC for some kai_matmul_clamp_f32_qai8dxp_qsi4c32p micro-kernels
- Fix compilation warnings detected by
-Wcast-qual -Wmissing-prototypes -Wstrict-prototypes -Woverlength-strings
compiler options.- Support compiling the project with the above compilation options enabled.
- Remove
-Werror
from default build flags as to not cause integration problems - Expose the rhs_packed_stride in the header file
- Fix validation error when n > nr in kai_matmul_clamp_f32_qai8dxp1vlx8_qsi4cxp4vlx8_1vlx4vl_sme2_mopa
- New SME micro-kernels:
- Matrix multiplication (MxN) for BF16 inputs with F32 output.
- Add MSVC support for test framework
- Fixes:
- Fix several CPU feature check issues affecting test framework
- Fix the LHS/RHS packed offset calculation in matmul get_offset methods
- New Advanced SIMD micro-kernels:
- New 16x4 and 1x4 block size variants of matrix multiplication of QAI8DXP LHS and QSI4C32P RHS with F32 output.
- Optimizations for FEAT_DotProd.
- New 16x4 and 1x4 block size variants of matrix multiplication of QAI8DXP LHS and QSI4C32P RHS with F32 output.
- New SME micro-kernels:
- Matrix multiplication (MxN and 1xN) of QAI8DXP LHS and QSI4CXP RHS to produce F32 output.
- Packing micro-kernels for QSI4CXP RHS to work with the SME matrix multiplication (MxN and 1xN) micro-kernels.
- Fixes:
- Fix out-of-bounds read in
kai_lhs_quant_pack_qai8dxp_f32
packing micro-kernel. - Unit test improvements.
- Fix out-of-bounds read in
- Breaking changes:
- Change the F16 matrix multiplication function signature to use single-precision floating-point for the clamp values.
- Optimizations:
- Optimize QAI8DXP LHS quant and pack micro-kernel using Arm® Neon™
- Optimize the NxK scalar RHS packing function for QSU4C32 with BF16 quantization scales
- Add initial Microsoft® Visual C++™ build support
- API for querying library version
- Fixes:
- Update QSI8CX tests
- Asserts will call
abort()
instead ofexit(...)
- Changed invalid assertion in F16 kernel
- Build system improvements
- Unit test improvements
- New Advanced SIMD micro-kernels:
- Matrix multiplication (MxN and 1xN) of QSI8D32 LHS (dynamic 8-bit integer per-block quantized) and QSI4C32 RHS (4-bit integer per-block quantized) to produce F32 output.
- Optimizations for FEAT_DotProd.
- Matrix multiplication (MxN and 1xN) of QAI8DX LHS (dynamic 8-bit integer per-row quantized) and QSI4CX RHS (4-bit integer per-channel quantized) to produce F32 output.
- Optimizations for FEAT_DotProd and FEAT_I8MM.
- Packing micro-kernels for LHS and non-transposed and transposed RHS.
- Matrix multiplication (MxN) of BF16 LHS and BF16 RHS to produce F16 output.
- Packing micro-kernels for LHS and non-transposed RHS.
- Matrix multiplication (MxN and 1xN) of QSI8D32 LHS (dynamic 8-bit integer per-block quantized) and QSI4C32 RHS (4-bit integer per-block quantized) to produce F32 output.
- New SME micro-kernels:
- Matrix multiplication (MxN and 1xN) of F16 LHS and F16 RHS to produce F16 output.
- Packing micro-kernels for LHS and non-transposed and transposed RHS.
- Matrix multiplication (MxN) of QAI8 LHS and QSI8 RHS to produce QAI8 output.
- Packing micro-kernels for LHS and non-transposed RHS.
- Matrix multiplication (MxN and 1xN) of QSI8D32 LHS and QSI4C32 RHS to produce F32 output
- Matrix multiplication (MxN and 1xN) of F16 LHS and F16 RHS to produce F16 output.
- Packing micro-kernels for QSI8D32 LHS and non-transposed QSI4C32 RHS, to work with the SME matrix multiplication (MxN and 1xN) micro-kernels.
- Fixes:
- Fixes relating to illegal instruction errors on systems with SME but without SVE support:
- Contain SME assembly inside the SMSTART and SMSTOP boundary.
- Disable compiler generated SVE instructions by adding the -fno-tree-vectorize compiler option to the build.
- Fix build warnings in the core library introduced by the -Wpedantic compiler option.
- Fix typos in the micro-kernel interface files.
- Fixes relating to illegal instruction errors on systems with SME but without SVE support:
- New Advanced SIMD micro-kernels:
- Matrix multiplication (MxN) of QAI8DX (dynamically quantized 8-bit integer) LHS and QSI4CX (quantized 4-bit integer) RHS with F32 output.
- Matrix multiplication (MxN and 1xN) of BF16 LHS and RHS with F32 output.
- New SME micro-kernels:
- SME2 F32 matrix multiplication (1xN) micro-kernels:
- Compatible with 2VL RHS packing, for sharing one packed RHS with SME2 F32 GEMM micro-kernel.
- Compatible with 16VL RHS packing.
- SME F32 packing function for transposed RHS matrix.
- SME2 F32 matrix multiplication (1xN) micro-kernels:
- Enhancements to existing micro-kernels:
- Port several quantized micro-kernels to optimized Advanced SIMD assembly.
- Register SME F32 matrix multiplication micro-kernel in the benchmark suite.
- Enable air gapped CMake builds through local third-party dependencies.
- Advanced SIMD FP32 GEMM micro-kernel.
- Micro-kernels to compute the matrix multiplication of dynamically quantized asymmetric signed 8-bit integer with per-row quantization (QAI8DX) LHS and quantized symmetric 4-bit signed integer with per-block quantization (QSI4C32) RHS. The destination matrix data type is single-precision floating-point (F32). The micro-kernels have been optimized using the Arm® CPU feature FEAT_I8MM for the matrix-by-matrix cases and the FEAT_DotProd for the vector-by-matrix cases.
- RHS matrix packing micro-kernels to pack the RHS matrix holding the QSI4C32 values.
- Unit test and example for integer micro-kernels.
- Extend support for signed 4-bit integer inputs in quantized symmetric 4-bit signed integer with per-channel quantization (QSI4CXP) RHS packing micro-kernel.
- kai_rhs_pack_nxk_qsi4cxp_qsu4cxs1s0 renamed to kai_rhs_pack_nxk_qsi4cxp_qs4cxs1s0.
- kai_rhs_pack_kxn_qsi4cxp_qsu4cxs1s0 renamed to kai_rhs_pack_kxn_qsi4cxp_qs4cxs1s0.
- Remove FP16 GEMV micro-kernel optimized for Advanced SIMD.
- Where a dedicated GEMV micro-kernel is not provided, it is recommended to use existing GEMM micro-kernels which have dedicated paths for M=1 (a "GEMV" operation).
- Micro-kernels to compute the matrix multiplication of dynamically quantized symmetric signed 8-bit integer with per-block quantization (QSI8D32) activations and quantized symmetric 4-bit signed integer with per-block quantization (QSI4C32) weights and the accumulation of the result into a single-precision (F32) output, optimized for Arm® Neon™ technology.
- Tensor packing micro-kernels to prepare the activations and weights for input to the above matrix multiplication micro-kernel.
- Unit test and example for integer micro-kernels.
The first release of KleidiAI includes:
- Micro-kernels to compute the matrix multiplication of:
- Dynamically quantized 8-bit integer (QAI8DX) activations and quantized 4-bit integer (QSI4CX) weights and the accumulation of the result into a single-precision (F32) output, optimized for Arm® Neon™ technology.
- Half precision floating-point (F16) activations and weights and the accumulation of the result into an F16 output, optimized for Neon technology.
- F32 activations and weights and the accumulation of the result into an F32 output, optimized for SME2 technology.
- Tensor packing micro-kernels to prepare the activations and weights for input to the above matrix multiplication micro-kernels.
- Examples and documentation demonstrating the usage of the 4-bit integer and 16-bit floating point matrix multiplication micro-kernels.
- Testing suite.
- CMake and Bazel build system for micro kernels.