'transforms' Submodule

• Introduction
•  apply_bufferization
•  apply_dataflow_optimization_transforms
•  apply_greedy_kernel_fusion
•  apply_hls_codegen
•  apply_intensity_aware_linalg_tiling
•  apply_linalg_optimization_transforms
•  apply_linear_programming_resource_allocation
•  apply_runtime_codegen
•  apply_tensor_to_dataflow_conversion
•  construct_kernel_fusion_transform_sequence
•  construct_linalg_tiling_transform_sequence

Introduction

from streamtensor import transforms

Transformations for StreamTensor compiler.

This module provides a set of transformations for the StreamTensor compiler. These transformations are used to optimize the input MLIR module targeting spatial accelerators. The module is transformed in place.

apply_bufferization(module, task_partition_merging=True, task_chain_merging=False, data_driven_task_merging=False, max_stream_depth=None)

Apply bufferization.

Parameters:

Name	Type	Description	Default
module	Module	The module to be transformed.	required
task_partition_merging	bool	Whether to merge task partitions.	True
task_chain_merging	bool	Whether to merge task chains.	False
data_driven_task_merging	bool	Whether to merge data-driven tasks.	False
max_stream_depth	Optional[int]	The maximum stream depth.	None

apply_dataflow_optimization_transforms(module, max_widen_bitwidth=512, constant_folding=False)

Apply dataflow optimization transforms.

Parameters:

Name	Type	Description	Default
module	Module	The module to be transformed.	required
max_widen_bitwidth	int	The maximum bitwidth to widen.	512
constant_folding	bool	Whether to perform constant folding.	False

apply_greedy_kernel_fusion(module, entry_name, max_fusion_cost, dot_file=None, fused_dot_file=None, delete_sequence=True)

Apply greedy kernel fusion.

Parameters:

Name	Type	Description	Default
module	Module	The module to be transformed.	required
entry_name	str	The entry function name.	required
max_fusion_cost	int	The maximum fusion cost.	required
dot_file	Optional[str]	The DOT file to print the kernel graph.	None
fused_dot_file	Optional[str]	The DOT file to print the post-fusion kernel graph.	None
delete_sequence	bool	Whether to delete the transform sequence after transform.	True

apply_hls_codegen(module, num_hbm_port=32, create_kernel_wrapper=True, assign_slr=False)

Apply HLS C++ code generation transforms.

Parameters:

Name	Type	Description	Default
module	Module	The module to be transformed.	required
num_hbm_port	int	The number of HBM ports.	32
create_kernel_wrapper	bool	Whether to create kernel wrapper function.	True
assign_slr	bool	Whether to assign SLR for each task.	False

apply_intensity_aware_linalg_tiling(module, entry_name, default_tile_size, overall_unroll_size, max_vec_bitwidth, convert_linalg_to_dataflow, op_ii_map=None, dot_file=None, delete_sequence=True)

Apply intensity-aware linalg tiling.

Parameters:

Name	Type	Description	Default
module	Module	The module to be transformed.	required
entry_name	str	The entry function name.	required
default_tile_size	int	The default tile size.	required
overall_unroll_size	int	The overall unroll size.	required
max_vec_bitwidth	int	The maximum vectorize bitwidth.	required
convert_linalg_to_dataflow	bool	Whether to convert tiled linalg op to dataflow kernel.	required
op_ii_map	Optional[Dict[str, int]]	The map from op name to its initiation interval on hardware.	None
dot_file	Optional[str]	The DOT file to print the linalg graph.	None
delete_sequence	bool	Whether to delete the transform sequence after transform.	True

apply_linalg_optimization_transforms(module, fake_quantize=False, quantize_bitwidth=8, constant_quantize_bitwidth=8)

Apply linalg optimization transforms.

Parameters:

Name	Type	Description	Default
module	Module	The module to be transformed.	required
fake_quantize	bool	Whether to apply fake quantization.	False
quantize_bitwidth	int	The bitwidth for fake quantization.	8
constant_quantize_bitwidth	int	The bitwidth of constants for fake quantization.	8

apply_linear_programming_resource_allocation(module, entry_name, pipeline_rewinding=True, stream_size_map=None, dot_file=None, partition_dot_file=None, partition=3, num_dsp=[2664, 2784, 2928], num_uram=[320, 320, 320], num_bram=[1200, 1152, 1200], num_lut=[386880, 364320, 395040], max_widen_bitwidth=512, imbalance_threshold=0.3)

Apply linear programming resource allocation.

Parameters:

Name	Type	Description	Default
module	Module	The module to be transformed.	required
entry_name	str	The entry function name.	required
pipeline_rewinding	bool	Whether to enable pipeline rewinding.	True
stream_size_map	Optional[Dict[Tuple[str, str], int]]	The overriding map from edge to stream size.	None
dot_file	Optional[str]	The DOT file to print the stream-sized graph.	None
partition_dot_file	Optional[str]	The DOT file to print the partitioned graph.	None
partition	int	The number of partitions on the target FPGA.	3
num_dsp	List[int]	The number of DSPs on the target FPGA for each partition.	[2664, 2784, 2928]
num_uram	List[int]	The number of URAMs on the target FPGA for each partition.	[320, 320, 320]
num_bram	List[int]	The number of BRAMs on the target FPGA for each partition.	[1200, 1152, 1200]
num_lut	List[int]	The number of LUTs on the target FPGA for each partition.	[386880, 364320, 395040]
max_widen_bitwidth	int	The maximum bitwidth to widen.	512
imbalance_threshold	float	The imbalance threshold (>= 0) of graph partitioning, the larger the value, the more imbalance allowed.	0.3

apply_runtime_codegen(module, host_func_name, timeout_in_ms=1000)

Apply runtime C++ code generation transforms.

Parameters:

Name	Type	Description	Default
module	Module	The module to be transformed.	required
host_func_name	str	The name of the host function.	required
timeout_in_ms	int	The timeout in milliseconds.	1000

apply_tensor_to_dataflow_conversion(module)

Apply tensor to dataflow conversion.

Parameters:

Name	Type	Description	Default
module	Module	The module to be transformed.	required

construct_kernel_fusion_transform_sequence(target, design_space)

Construct a transform sequence for applying kernel fusion.

Parameters:

Name	Type	Description	Default
target	BlockArgument	The handle of the target op.	required
design_space	KernelFusionDesignSpace	The design space of kernel fusion.	required

Returns:

Type	Description
List[Value]	A list of transformed values, which is empty in this case.

Raises:

Type	Description
ValueError	If the kernel is not a KernelOp.

construct_linalg_tiling_transform_sequence(target, design_space, max_vec_bitwidth, convert_linalg_to_dataflow=True)

Construct a transform sequence for applying linalg tiling.

The following parameters are required in the input design space

• parallel_tile_sizes: The parallel tile sizes.
• reduction_tile_sizes: The reduction tile sizes.
• unroll_sizes: The unroll sizes.
• inputs_vec_sizes: The input vector sizes.
• outputs_vec_sizes: The output vector sizes.
• permutation: The permutation of the loop order.

Parameters:

Name	Type	Description	Default
target	BlockArgument	The handle of the target op.	required
design_space	LinalgTilingDesignSpace	The design space of linalg tiling.	required
max_vec_bitwidth	int	The maximum vectorize bitwidth.	required
convert_linalg_to_dataflow	bool	Whether to convert tiled linalg op to dataflow kernel.	True

Returns:

Type	Description
List[Value]	A list of transformed values, which is empty in this case.

Raises:

Type	Description
ValueError	If the required attributes are not found.