triton.experimental.gluon.language.amd.cdna3.scaled_upcast

triton.experimental.gluon.language.amd.cdna3.scaled_upcast(src, scale, elem_type, axis=None, _semantic=None)

Upcast an fp4 or fp8 tensor and fold raw E8M0 scale payload into the CDNA3 scaled-upcast op.

CDNA3 lowers this through the software-emulated scaled-upcast path; it does not use native hardware scaled-upcast instructions.

The scale tensor must use raw E8M0 payload in int8 or uint8, and must already have the expanded output shape and scaled-upcast result layout. For fp4 inputs, that is the canonical unpacked layout implied by src and axis. elem_type must be fp16 or bf16.