Skip to content

v1.33.0

Latest

Choose a tag to compare

@SuhailMunshi SuhailMunshi released this 30 Oct 09:35
· 36 commits to main since this release
  • New Advanced SIMD micro-kernels:
    • Statically quantized INT8 GEMV using Advanced SIMD dot-product instructions, with the struct-based API.
  • Optimizations:
    • Improve FP32-to-QAI8DXP LHS quantisation and packing performance in kai_lhs_quant_pack_qai8dxp_f32 by shortening Advanced SIMD reduction dependencies and reducing redundant clamping and padding work.
  • Fixes:
    • Correct the documented K requirement for kai_lhs_quant_pack_qai8dxp_f32: input K has no divisibility requirement; only the packed K dimension is rounded up to a multiple of 32.
  • New SME micro-kernels:
    • Native MxN SME MOPA matrix multiplication for FP16 x signed QSI4C32P with F32 output and the s4s0sf16 RHS packing format.
    • F32 to QAI8DXP LHS quant and pack kernel kai_lhs_quant_pack_qai8dxp1vlx4_f32_sme optimized with SME
  • New SME2 micro-kernels:
    • F32 to QAI8DXP LHS quant and pack kernel kai_lhs_quant_pack_qai8dxp1vlx4_f32_sme2 optimized with SME2
  • New SME2.1 micro-kernels:
    • New SME2.1 Fexpa micro-kernel for FP32 datatype.
  • New SVE micro-kernels:
    • New SVE Fexpa Softmax FP32 micro-kernel
    • FP32 matrix multiplication using MOP4A instructions for small-M and small-N cases.
  • New SVE micro-kernel: per-channel INT4 matrix multiplication with FP32 output, kai_matmul_clamp_f32_qai8dxp2vsx8sf32_qsi4cxp8x8s16s0sf32bf32_8vsx8_sve_i8mm, requiring SVE I8MM and exactly 256-bit vectors.