Matrix Multiply Accelerate built-in functions

IBM® Open XL C/C++ for Linux® on Power® 17.1.1 adds Matrix Multiple Accelerate (MMA) built-in functions.

MMA is embedded into the IBM Power10 processor and is designed to achieve faster AI inference for FP32, BFloat16, and INT8 calculations.

ACC is an MMA 512-bit accumulator. MMA intrinsic procedures can be used to directly exploit ACC on the Power10 processor and accelerate matrix multiplication computations.

Intrinsic types

IBM Open XL C/C++ for Linux on Power 17.1.1 supports the following intrinsic types:

__vector_pair

32-byte opaque vector type

__vector_quad

64-byte opaque vector type