DFG Conversion for CGRA Mapping With Fewer Data Transmission Columns
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing coarse-grained reconfigurable architectures (CGRA) face inefficiencies in data processing due to suboptimal mapping of data flow graphs (DFGs), leading to reduced arithmetic operation performance and processing capacity, particularly when data transmission paths are limited to one-way directions, hindering expansion and throughput enhancement.
Innovation Solution
A conversion method and device that optimizes DFG mapping by identifying and converting specific patterns within the DFG to reduce the number of data transmission columns, enhancing the utilization efficiency of processing elements (PEs) in CGRA through pattern-based data flow adjustments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If data transmission between PEs is limited to one-way direction to simplify mapping, then mapping complexity is reduced, but the number of columns increases and processing capacity deteriorates
Solution Approach 1:
The patent introduces dynamic data transmission paths that allow data to flow in multiple directions between PEs based on computational needs. The routing configuration can be dynamically adjusted to optimize data flow, enabling both simplification of mapping and maintenance of processing capacity through adaptive path selection rather than fixed one-way constraints
Solution Approach 2:
The patent utilizes two-dimensional PE arrangement with configurable routing in both horizontal and vertical directions. By leveraging the two-dimensional topology and allowing data transmission in multiple directions across the array, the system reduces the number of columns needed while maintaining processing capacity, effectively using spatial dimensionality to resolve the contradiction
2Ease of operation
If DFG mapping is performed with limited data transmission paths, then routing simplicity is improved, but PE utilization efficiency deteriorates
Solution Approach 1:
The patent implements dynamic routing configurations that adapt to the specific DFG being executed. The routing paths are configured based on the data flow requirements of each computation task, providing simplicity for each specific case while optimizing PE utilization through task-specific path selection rather than rigid predefined paths
Solution Approach 2:
The patent changes routing parameters and data flow characteristics based on the DFG structure and computational requirements. By adjusting transmission paths, directionality, and routing configuration to match the specific computation task, the system achieves both routing simplicity for the given task and high PE utilization efficiency
3Quantity of substance
If the number of columns in DFG mapping is increased to accommodate more data flow, then data transmission capacity is improved, but expansion by combination becomes difficult and device complexity increases
Solution Approach 1:
The patent leverages the two-dimensional PE array topology to provide data transmission capacity in both horizontal and vertical directions. By utilizing the two-dimensional space efficiently with configurable routing, the system achieves high data transmission capacity without increasing the number of columns, as data can flow through multiple paths across the二维 array
Solution Approach 2:
The patent creates a universal routing infrastructure that can handle multiple data flow patterns and computation tasks. The same PE array and routing fabric can be reconfigured for different DFGs, providing high data transmission capacity for various workloads without requiring dedicated hardware for each case, thus avoiding increased device complexity
Data Source
AI summary
A non-transitory computer-readable recording medium stores therein a conversion program that causes a computer to execute a process including, acquiring a predetermined original DFG and a mapping result of mapping of the predetermined original DFG with respect to a CGRA that includes a plurality of arithmetic operation units, extracting, from the predetermined original DFG, a portion corresponding to a pattern that is formed in the DFG and that has been determined in advance, determining a conversion candidate DFG based on a location in which the extraction portion has been allocated and based on the number of transmission paths for data used between the arithmetic operation units indicated in the mapping result, and generating a converted DFG based on the conversion candidate DFG by converting the predetermined original DFG.