tvm_tirx.backend.cuda.op.TensorMapEncodeTiledAttr

tvm_tirx.backend.cuda.op.TensorMapEncodeTiledAttr#

class tvm_tirx.backend.cuda.op.TensorMapEncodeTiledAttr(descriptor_dtype, rank, interleave=0, swizzle=0, l2_promotion=0, oob_fill=0, force_cu_dtype=-1)#

Bases: Attrs

Fixed options for tensormap_encode_tiled().

Parameters:

  • descriptor_dtype (str or tvm_ffi.dtype) – Tensor element dtype encoded in the descriptor.

  • rank (int) – Number of tensor dimensions. Dimensions, strides and pointers are separate operands of the encoding call, not fields of this object.

  • interleave (int) – CUDA tensor-map enum values for the corresponding encoding modes. Each defaults to zero.

  • swizzle (int) – CUDA tensor-map enum values for the corresponding encoding modes. Each defaults to zero.

  • l2_promotion (int) – CUDA tensor-map enum values for the corresponding encoding modes. Each defaults to zero.

  • oob_fill (int) – CUDA tensor-map enum values for the corresponding encoding modes. Each defaults to zero.

  • force_cu_dtype (int) – Explicit CUDA tensor-map dtype enum, or -1 to derive it from descriptor_dtype.

property descriptor_dtype#
property force_cu_dtype#
property interleave#
property l2_promotion#
property oob_fill#
property rank#
property swizzle#