Skip to content

feat(rm): Add support for MXINT8

Ian Tayler Lessa requested to merge feat/rm/mxint8-support into main
  • Add TOSA_REF_TYPE_MXINT8
  • Implement utilities like IsFloat, DtypeLimits, getFpMinimum, etc.
  • Rewrite roundToNearestFloat to work around a bug in mxint8 cfloat_cast
  • Add support to the ModelRunner
  • Add MXINT8 cases to MATMUL_T_BLOCK_SCALED, CAST_FROM_BLOCK_SCALED and CAST_TO_BLOCK_SCALED
  • Add support in the Subgraph Traverser for allocating mxint8 tensors from constants
  • Add support to Tensor for readFromNpyFile, readfromVector, writeToVector
  • Add support for AccPrecision because it's used in some of the unit tests
  • Add coverage for mxint8 in the existing MXFP-related unit tests

Signed-off-by: Ian Tayler Lessa ian.taylerlessa@arm.com

Merge request reports