v1.26.0

• New SME micro-kernels
  • Added an x8 matmul pack micro-kernel family with 4vsx4 blocked layout, without packing bias.
  • Added SME RHS depthwise packing kernel for FP16
• New SME2 micro-kernels
  • Added SME2 depthwise indirect micro-kernel for FP16.
  • kai_matmul_i32_u8p4vsx4_u8p4vsx4_i32_i32_8vsx8vs_sme2_mopa.
  • kai_matmul_clamp_f32_u8p4vsx4_u8p4vsx4_i32_i32_f32_f32_8vsx8vs_sme2_mopa.
• Extended the following micro-kernels to support variable block length
  • kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4x4_1x4_neon_dotprod
  • kai_matmul_clamp_f32_qsi8d32p1x8_qsi4c32p4x8_1x4x32_neon_dotprod
  • kai_matmul_clamp_f32_qsi8d32p4x4_qsi4c32p4x4_16x4_neon_dotprod
  • kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_8x4x32_neon_i8mm
  • kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm
  • kai_rhs_pack_nxk_qsi4c32ps1s0scalef16_qsu4c32s16s0
  • kai_lhs_quant_pack_qsi8d32p4x8sb_f32_neon
• Fixes
  • Added ZA lazy save to kai_matmul_clamp_f32_qsi8d32p1x4_qsi4c32p4vlx4_1x4vl_sme_dot
  • Fix QAI8/QSI8CXP matmul test failures by constraining generated qsi32 bias values to preserve int32 accumulator headroom.
  • Fix a clamping issue in matmul_clamp_qai8_qai8p_qai8p_test.cpp
  • Fix traditional matmul and imatmul packed offset helpers to use packing panel boundaries.
• New Advanced SIMD micro-kernels
  • Matrix Multiplication MxN and 1xN Micro-Kernels of QAI8DXP LHS and QSU2CXP RHS with F32 output, optimized for FEAT_DotProd, along with RHS packing kernel.
• Documentation
  • Contribution policy updates as part of third party contribution enablement
  • Added coding standard and conventions
• New Transposed-B RHS packing micro-kernel versions of kai_rhs_pack_kxn_x32p16x1b_x32_x32_neon and kai_rhs_pack_kxn_x16p32x1b_x16_x16_neon:
  • kai_rhs_pack_nxk_x16p32x1bx16_x16_x16_neon
  • kai_rhs_pack_nxk_x32p16x1bx32_x32_x32_neon
• New SME2 FP32 GEMV micro-kernel with 4vsx1 RHS format
• New SME2 static Int8 GEMM/GEMV kernels and the RHS packing kernel.
  • kai_matmul_clamp_qai8_qai8p4vsx4_qsi8cxp4vsx4bi32sf32_8vsx8vs_sme2_mopa
  • kai_matmul_clamp_qai8_qai8_qsi8cxp4vsx4bi32sf32_1x32vs_sme2_dot
  • kai_matmul_pack_rhs_kxn_qsi8cxp4vsx4bi32sf32_qsi8_i32_f32_sme