triton.experimental.gluon.language.amd.cdna3.mfma

triton.experimental.gluon.language.amd.cdna3.mfma(a, b, acc, cd_regclass=None, _semantic: GluonSemantic = None)

Computes matrix multiplication a * b + acc using AMD native matrix core units.

Parameters:

• a (tensor) – The first operand of mfma.

• b (tensor) – The second operand of mfma.

• acc (tensor) – The accumulator tensor.

• cd_regclass (str, optional) – Experimental. Register class for the accumulator input (C) and result (D) of the MFMA instructions: "a" for AGPRs or "v" for VGPRs. Requires an AMDMFMALayout accumulator. None (default) leaves the choice to the compiler.