tvm_tirx.backend.cuda.op.tensormap_encode_tiled

tvm_tirx.backend.cuda.op.tensormap_encode_tiled(*args, descriptor_dtype, rank, interleave=0, swizzle=0, l2_promotion=0, oob_fill=0, force_cu_dtype=-1, loc=UNKNOWN_LOC)

Encode a tiled tensor map using runtime pointers and shape operands.

Arguments are the descriptor and data pointers, global dimensions (rank), byte strides (rank - 1), box dimensions (rank), and element strides (rank). The dtype describes the final descriptor units, including any promotion. CUDA-host codegen encodes directly; other hosts use the runtime packed call.

Parameters:

loc (Location)