tvm_tirx.backend.cuda.op.ptx_cp_async_legacy

tvm_tirx.backend.cuda.op.ptx_cp_async_legacy#

tvm_tirx.backend.cuda.op.ptx_cp_async_legacy(dst_ptr, dst_offset, src_ptr, src_offset, cp_size, *, elem_dtype='int8', ty=None, loc=UNKNOWN_LOC)#

Fold element offsets into the raw cp.async pointer operands.

elem_dtype scales element offsets independently of the call’s result ty.

Parameters:

loc (Location)