tilelang.carver.template.conv module#
- class tilelang.carver.template.conv.ConvTemplate(N: int, C: int, H: int, W: int, F: int, K: int, S: int, D: int, P: int, in_dtype: str = 'float16', out_dtype: str = 'float16', accum_dtype: str = 'float16', with_bias: bool = False)#
Bases:
BaseTemplateA template for convolution (Conv).
This class defines the computation for a matrix-matrix convolution with configurable parameters such as transposition, data types, and bias addition.
- N#
The number of input samples processed simultaneously in a batch.
- Type:
int
- C#
The number of input feature maps.
- Type:
int
- H#
The height of the input feature maps.
- Type:
int
- W#
The width of the input feature maps.
- Type:
int
- F#
The number of filters (kernels) applied, determining output depth.
- Type:
int
- K#
The spatial dimensions of each convolutional filter.
- Type:
int
- S#
The step size by which the kernel slides across the input.
- Type:
int
- D#
The spacing between kernel elements, controlling receptive field expansion.
- Type:
int
- P#
The number of pixels added to input borders to control output spatial dimensions.
- Type:
int
- in_dtype#
Data type of input matrices.
- Type:
str
- out_dtype#
Data type of output matrix.
- Type:
str
- accum_dtype#
Data type used for accumulation.
- Type:
str
- with_bias#
Whether to add a bias term.
- Type:
bool
- C: int#
- D: int#
- F: int#
- H: int#
- K: int#
- N: int#
- P: int#
- S: int#
- W: int#
- accum_dtype: str = 'float16'#
- property class_attributes#
Returns the class attributes in dictionary form.
- Returns:
Dictionary of class attributes.
- Return type:
dict
- get_hardware_aware_configs(arch=None, topk=10) List[Hint]#
Retrieves optimized hardware-aware configurations.
- Parameters:
arch (TileDevice, optional) – The target hardware architecture.
topk (int, optional) – Number of top configurations to consider.
- Returns:
A list of optimization hints for hardware acceleration.
- Return type:
List[Hint]
- in_dtype: str = 'float16'#
- initialize_function() None#
Defines and initializes the convolution computation.
This method sets up placeholders for input matrices, computes the convolution using TVM’s compute API, and optionally applies bias and type casting.
- Raises:
AssertionError – If N, C, H, W, F, K, S, D, P are not positive integers.
- out_dtype: str = 'float16'#
- params_as_dict()#
Returns the template parameters as a dictionary.
- Returns:
Dictionary containing template parameter values.
- Return type:
dict
- with_bias: bool = False#