WMMA builtins

Wave-Matrix Multiply-Accumulate (WMMA) builtins let you issue hardware matrix multiply-accumulate operations directly from HIP device code. This reference covers both dense and sparse WMMA variants on AMD RDNA architectures (AMD Radeon GPUs). Coverage is organized first by operand density – dense or sparse – and then by RDNA generation.

• Dense WMMA builtins

  • RDNA3 dense WMMA builtins

  • RDNA4 dense WMMA builtins

• Sparse WMMA builtins

  • RDNA4 sparse WMMA builtins

• RDNA4 WMMA transpose load builtins