tilelang.carver.roller.hint¶

Hint definition for schedule

Classes¶

`TensorCoreExtraConfig`	This class is used to store extra information for tensorcore
`Stride`	Manages stride information for a given axis of a tensor.
`TileDict`	Manages tiling information and configurations for computational tasks.
`IntrinInfo`	The information of tensorcore intrinsic related information
`Hint`	Central configuration class for managing various parameters of computational tasks.

Module Contents¶

class tilelang.carver.roller.hint.TensorCoreExtraConfig(AS_shape, BS_shape, AF_shape, BF_shape, tc_axis)¶

This class is used to store extra information for tensorcore

Parameters:

AS_shape (Tuple[int])
BS_shape (Tuple[int])
AF_shape (Tuple[int])
BF_shape (Tuple[int])
tc_axis (Tuple[int])

AS_shape: Tuple[int]¶

BS_shape: Tuple[int]¶

AF_shape: Tuple[int]¶

BF_shape: Tuple[int]¶

tc_axis: Tuple[int]¶

class tilelang.carver.roller.hint.Stride(stride=1, ax=-1)¶

Manages stride information for a given axis of a tensor.

Parameters:

stride (int)
ax (int)

property ax: int¶

Return type:: int

property stride: int¶

Return type:: int

compute_strides_from_shape(shape)¶

Parameters:: shape (List[int])
Return type:: List[int]

compute_elements_from_shape(shape)¶

Parameters:: shape (List[int])
Return type:: int

is_valid()¶

Return type:: bool

__repr__()¶

Return type:: str

class tilelang.carver.roller.hint.TileDict(output_tile)¶

Manages tiling information and configurations for computational tasks.

output_tile¶

tile_map¶

rstep_map¶

cached_tensors_map¶

output_strides_map¶

tensor_strides_map¶

traffic = -1¶

smem_cost = -1¶

block_per_SM = -1¶

num_wave = -1¶

grid_size = -1¶

valid = True¶

get_tile(func)¶

Return type:: List[int]

get_rstep(node)¶

Return type:: Dict[str, int]

__hash__()¶

Return type:: int

class tilelang.carver.roller.hint.IntrinInfo(in_dtype, out_dtype, trans_b, input_transform_kind=0, weight_transform_kind=0)¶

The information of tensorcore intrinsic related information

Parameters:

in_dtype (str)
out_dtype (str)
trans_b (bool)
input_transform_kind (int)
weight_transform_kind (int)

in_dtype¶

out_dtype¶

trans_a = False¶

trans_b¶

input_transform_kind = 0¶

weight_transform_kind = 0¶

__repr__()¶

Return type:: str

is_input_8bit()¶

Return type:: bool

property smooth_a: bool¶

Return type:: bool

property smooth_b: bool¶

Return type:: bool

property inter_transform_a: bool¶

Return type:: bool

property inter_transform_b: bool¶

Return type:: bool

class tilelang.carver.roller.hint.Hint¶

Bases: object

Central configuration class for managing various parameters of computational tasks.

arch = None¶

use_tc = None¶

block = []¶

thread = []¶

warp = []¶

rstep = []¶

reduce_thread = []¶

rasterization_plan¶

cached_tensors = []¶

output_strides¶

schedule_stages = None¶

block_reduction_depth: int = None¶

split_k_factor: int = 1¶

vectorize: Dict[str, int]¶

pipeline_stage = 1¶

use_async = False¶

opt_shapes: Dict[str, int]¶

intrin_info¶

shared_scope: str = 'shared'¶

pass_context: Dict¶

to_dict()¶

Return type:: Dict

classmethod from_dict(dic)¶

Parameters:: dic (Dict)
Return type:: Hint

tensorcore_legalization()¶

property raxis_order: List[int]¶

Return type:: List[int]

property step: List[int]¶

Return type:: List[int]

__repr__()¶

Return type:: str

complete_config(node)¶

Parameters:: node (tilelang.carver.roller.PrimFuncNode)