A wiring method for FPGA prototype verification system
By combining signals with similar properties and the same power consumption, and optimizing the time division multiplexing ratio with the minimum Steiner tree approximation algorithm and mathematical planning, the problem of signal delay in the FPGA prototype verification system is solved, and the system performance and speed are improved.
Patent Information
- Application Number
- CN202110526352.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-05-14
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2041-05-14
AI Technical Summary
In the FPGA prototype verification system, although TDM technology increases the number of signal transmissions, it leads to signal delay growth and affects system performance.
By combining signals with similar properties and the same power consumption, the minimum Steiner tree approximation algorithm and mathematical planning are used to optimize the time division multiplexing ratio, reduce signal delay, and improve system performance.
It effectively reduces signal delay on critical paths, improves the performance of FPGA systems, and improves the speed of verification systems.
Smart Images

Figure CN113392615B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of electronic design automation, and in particular relates to a wiring method for an FPGA prototype verification system. Background Art
[0002] As the manufacturing process of very large scale integrated circuits (VLSI) becomes more and more advanced, if some loopholes are found in actual manufacturing, the cost of correcting the design and remanufacturing is huge, so verification is very necessary at each stage of the VLSI design process, and the cost and time required for verification account for half of chip design.
[0003] Currently, there are three methods for performing chip logic verification, namely software simulation, processor-based hardware simulation accelerated software simulation, and FPGA prototype verification system. Among them, the verification method based on FPGA prototype has high execution speed, low cost and rich testing experience, so it has high application value. However, the capacity of a single FPGA is limited, so a Multi-FPGA prototype system composed of multiple FPGAs is used to verify the logic of large circuit designs.
[0004] During verification, large circuits are divided into multiple sub-circuits, but the number of I / Os used by signals transmitted between FPGAs exceeds the number of I / Os of the FPGA itself, so Babb et al. introduced the time division multiplexing (TDM) technology of I / O pins, which divides a channel between FPGAs into several segments according to time and uses them for multiple signals in turn. This technology greatly increases the number of signals that can be transmitted between FPGAs, allowing the wiring to be successfully laid out.
[0005] Although TDM technology increases the number of signal transmissions between FPGAs, it also has negative effects, the most important of which is the increase in signal delay. This type of delay accounts for a large proportion of the delay of the entire FPGA verification system, causing an increase in delay on the critical path, which in turn leads to a decrease in the performance of the entire FPGA prototype system. Summary of the invention
[0006] Purpose of the invention: The present invention provides a wiring method for an FPGA prototype verification system. Given all signals, signals with similar properties and the same power consumption are combined into a group for wiring and resource allocation according to the requirements of industrial design. This not only enables the successful wiring of the FPGA prototype system, but also minimizes the group with the greatest negative effects among all groups, thereby greatly reducing system delay.
[0007] Technical solution: The present invention provides a wiring method for an FPGA prototype verification system, which specifically includes the following steps:
[0008] (1) Convert the wiring diagram of the FPGA prototype verification system into an undirected connected graph;
[0009] (2) With the goal of shortening the total length of the network and reducing the number of edge reuses, the network is routed based on the approximate algorithm of the minimum Steiner tree;
[0010] (3) According to the network wiring conditions and constraints, list the mathematical programming and pre-allocate the time division multiplexing ratio;
[0011] (4) Accurately allocate the time division multiplexing ratio and further optimize it.
[0012] Furthermore, the implementation process of step (1) is as follows:
[0013] Convert the wiring diagram of the FPGA prototype system into an undirected connected graph G(V,E), where the i-th FPGA is represented by point v i (i=1,2,3…), the jth edge is e j (j=1,2,3…); V={v1,v2,v3…} is the point set composed of all FPGAs, E={e1,e2,e3…} is the edge set composed of all connections between FGPAs, there can only be one edge between a pair of FGPAs, and the capacity of each edge is 1; multiple points constitute a network, and the set of all networks is recorded as N={n1,n2,n3,…}; if some networks have similar properties or the same power consumption, they form a group, and the set of all groups is recorded as G={g1,g2,g3…}.
[0014] Furthermore, the step (2) comprises the following steps:
[0015] (21) Classify the groups: Let the number of edges required by each network be n pi =num(n i )-1(i∈{1,2,3,…,|N|}), where num(n i ) indicates network i The number of points contained, |N| represents the number of elements in N, that is, the number of networks; the required number of edges for each group is:
[0016] g pi =∑ j n pj (n j ∈g i ,)
[0017] According to g pi Sort all groups from largest to smallest and give them new subscripts, where g1 is the group with the largest edge demand;
[0018] (22) Wiring the network in g1: Initialize the weight of each edge in E to cost = 2, increase the weight by 2 each time the edge is used, set the threshold α, and divide the network into num(n i )>α and num(n i )≤α; for num(n i )>α, the network is solved by using the minimum Steiner tree approximation algorithm based on the shortest path in descending order according to the number of points contained; for num(n i )≤α, use the disjoint shortest path approximation algorithm to solve, and finally get all the paths of g1;
[0019] (23) Reorder the remaining groups: After assigning paths to the nets in g1, retain the weights of all edges; do not consider the nets that have been routed, recount the number of edge demands of the remaining groups, sort them from large to small and give them new subscripts. At this time, g2 is the group with the largest edge demand. Repeat step (22) to assign paths to the unrouted nets in g2;
[0020] (24) Repeat steps (22) and (23) until all nets are allocated;
[0021] (25) Split and reroute the largest group: After all the nets are routed successfully, use the estimation function To predict the group with the largest time division multiplexing ratio, where E n is the set of all edges that the network n passes through, N e is the set of all networks passing through edge e, |N e | is the set N e The number of elements in the group; finally, the minimum Steiner tree approximation algorithm is used to redistribute the paths of the networks in the group.
[0022] Furthermore, the step (3) comprises the following steps:
[0023] (31) After adopting the time division multiplexing technology, the connection between each pair of FPGAs needs to be multiplexed by multiple networks in a single clock cycle. In order to use the multiplexing hardware, the time division multiplexing ratio is required to be an even number. The mathematical planning (P) is established according to the wiring situation:
[0024]
[0025]
[0026]
[0027] k∈N (3)
[0028] Where r(e, n) represents the time division multiplexing ratio allocated when network n uses edge e, and N represents a natural number set;
[0029] (32) Pre-allocate the time division multiplexing ratio for the network in g1: Allocate a time division multiplexing ratio to all networks n that use edge e equal to the number of times e is used in g1: (When n uses e, x(e, n) = 1, otherwise x(e, n) = 0), record where r e Indicates the number of times e is used in g1, r n represents the time division multiplexing ratio of network n, represents the time division multiplexing ratio of g1, Indicates g i Allocatable time division multiplexing ratio;
[0030] (33) As a benchmark, pre-allocate the time division multiplexing ratio for the remaining groups: set a delta value for each group to indicate the priority of allocating the time division multiplexing ratio to each group:
[0031]
[0032] Select the group with the smallest delta, denoted as g2; assign the time division multiplexing ratio to the network using edge e in g2:
[0033] (34) Update and Repeat (32)-(33) until all nets are allocated.
[0034] Furthermore, the step (4) comprises the following steps:
[0035] (41) Adjust the time division multiplexing ratio and calculate the sum of the inverses of the time division multiplexing ratios of all networks on each edge:
[0036]
[0037] The newly allocated time division multiplexing ratio is:
[0038]
[0039] If r(e, n) is an odd number, then (r(e, n)+1) is taken to satisfy the constraints (1) and (2) in the mathematical programming (P); recalculate the time division multiplexing ratio of all groups Sort from largest to smallest and give new subscripts;
[0040] (42) Substitute r(e,n) into the mathematical program (P) described in step (3) to calculate the gap between each component in the restriction (1) and the boundary 1:
[0041]
[0042] Then reduce the amount involved The time division multiplexing ratio corresponding to the network, the new time division multiplexing ratio is:
[0043]
[0044] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are: Compared with the original technology, this technology reduces the signal delay on the critical path and improves the performance of the entire FPGA system; the present invention adopts different Steiner tree approximation algorithms according to the size of the network, selects as few edges as possible with a high number of reuse times and makes the total length of the network shorter, which provides the possibility for better allocation of time division multiplexing ratio; the present invention adopts a combinatorial optimization method to reduce the time division multiplexing ratio of the maximum group as much as possible, thereby reducing signal delay and improving the speed of the verification system. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION
[0046] The present invention will be further described in detail below in conjunction with the accompanying drawings.
[0047] Table 1 shows the meanings of the commonly used parameters of the present invention:
[0048] Table 1 Parameter Description
[0049] symbol meaning G The set of all groups g Group N The set of all nets n Net E The collection of all connections between FPGAs e Connections between FPGAs V The collection of FPGAs v Single FPGA r(e,n) The time division multiplexing ratio allocated when network n uses edge e
[0050] The present invention proposes a wiring method for an FPGA prototype verification system, such as Figure 1 As shown, the specific steps include:
[0051] Step 1: Convert the wiring diagram of the FPGA prototype system into an undirected connected graph.
[0052] Convert the wiring diagram of the FPGA prototype system into an undirected connected graph G(V, E), where the i-th FPGA is represented by point v i (i=1,2,3…), the jth edge is e j(j=1,2,3…); V={v1,v2,v3…} is the point set consisting of all FPGAs, E={e1,e2,e3…} is the edge set consisting of all connections between FGPAs, there can only be one edge between a pair of FGPAs, and the capacity of each edge is 1; multiple points form a network, and the set consisting of all networks is recorded as N={n1,n2,n3,…}; if some networks have similar properties or the same power consumption, they form a group, and the set consisting of all groups is recorded as G={g1,g2,g3…}. The same point can belong to multiple networks, and the same network can belong to multiple groups. The goal of the problem is to assign a path to each network in the shortest possible running time, and at the same time assign a time division multiplexing ratio to the edges used by the network (if the same edge is used by different networks, their respective time division multiplexing ratios need to be assigned separately), and minimize the time division multiplexing ratio of the maximum group, where the time division multiplexing ratio of each edge satisfies the following two constraints:
[0053]
[0054] Among them, N e Represents the set of all networks passing through edge e.
[0055] Step 2: With the goal of selecting as few edges as possible with the highest number of reuses and minimizing the total length of the network, the network is routed based on the approximate algorithm of the minimum Steiner tree.
[0056] (1) Classify the groups:
[0057] Let the number of edges required by each network be n pi =num(n i )-1(i∈{1, 2, 3, …, |N|}), where num(n i ) indicates network i The number of points contained, |N| represents the number of elements in N, that is, the number of networks; the required number of edges for each group is:
[0058] g pi =∑ j n pj (n j ∈g i ,)
[0059] According to g pi Sort all groups from largest to smallest and give them new subscripts, where g1 is the group with the largest edge demand.
[0060] (2) Route the net in g1:
[0061] Initialize the weight of each edge in E to cost = 2. Each time the edge is used, its weight increases by 2. Set the threshold α and divide the network into num(n i )>α and num(ni )≤α. For num(n i )>α, the network is solved by using the minimum Steiner tree approximation algorithm based on the shortest path in descending order according to the number of points contained. i )≤α, use the disjoint shortest path approximation algorithm to solve it, and finally get all the paths of g1.
[0062] (3) Route the nets in the remaining groups:
[0063] After assigning paths to the nets in g1, keep the weights of all edges; do not consider the nets that have been routed, re-count the number of edge demands of the remaining groups, sort them from large to small and give them new subscripts. At this time, g2 is the group with the largest edge demand, and use method (2) to assign paths to the unrouted nets in g2. Repeat (2) and (3) until the nets in all groups are assigned.
[0064] (4) Split and redistribute the largest group:
[0065] After all the nets are routed successfully, use the estimation function To predict the group with the largest time division multiplexing ratio, where E n is the set of all edges that the network n passes through, N e is the set of all networks passing through edge e, |N e | is the set N e The number of elements in the group; finally, the minimum Steiner tree approximation algorithm is used to redistribute the paths of the networks in the group.
[0066] Step 3: According to the wiring conditions and constraints of the network, list the mathematical programming with the goal of minimizing the time division multiplexing ratio of the maximum group and pre-allocate the time division multiplexing ratio.
[0067] (1) Establish mathematical programming:
[0068] After using time division multiplexing technology, the connection between each pair of FPGAs needs to be multiplexed by multiple nets in a single clock cycle. Secondly, in order to use multiplexing hardware, the time division multiplexing ratio is required to be an even number, so a mathematical plan (P) is established based on the wiring situation:
[0069]
[0070]
[0071]
[0072] k∈N (3)
[0073] Wherein, r(e, n) represents the time division multiplexing ratio allocated when network n uses edge e, and N represents a natural number set.
[0074] (2) Pre-allocated time division multiplexing ratio:
[0075] Allocate a time-division multiplexing ratio to all networks n that use edge e equal to the number of times e is used in g1: (When n uses e, x(e, n) = 1, otherwise x(e, n) = 0), record where r e Indicates the number of times e is used in g1, r n represents the time division multiplexing ratio of network n, represents the time division multiplexing ratio of g1, Indicates g i Allocatable time division multiplexing ratio.
[0076] by As a benchmark, pre-allocate the time division multiplexing ratio for the remaining groups: set a delta value for each group to indicate the priority of allocating the time division multiplexing ratio to each group:
[0077]
[0078] Select the group with the smallest delta, denoted as g2; assign the time division multiplexing ratio to the network using edge e in g2: renew and Repeat the above steps until all nets are allocated.
[0079] Step 4: Accurately allocate the time division multiplexing ratio and further optimize it.
[0080] Adjust the time division reuse ratio and calculate the sum of the inverses of the time division reuse ratios of all networks on each edge:
[0081]
[0082] The newly allocated time division multiplexing ratio is:
[0083]
[0084] If r(e, n) is an odd number, then (r(e, n)+1) is taken to satisfy the constraints (1) and (2) in the mathematical programming (P); recalculate the time division multiplexing ratio of all groups Sort from largest to smallest and give new subscripts;
[0085] Substituting r(e,n) into the mathematical program (P), calculate the gap between each component in the constraint (1) and the boundary 1:
[0086]
[0087] Then reduce the amount involved The time division multiplexing ratio corresponding to the network, the new time division multiplexing ratio is:
[0088]
[0089] In order to verify the effect of the present invention, as shown in Table 2:
[0090] Table 2 Experimental test results
[0091]
[0092] The first column in Table 2 is the test case, the second column is the first place result of the 2019 ICCAD competition, the third column is the experimental result of the present invention, and the fourth column is the comparison of the two results. Through the comparison of the experimental results, the present invention has an average improvement of 0.93%. Except for the test cases synopsys01 and synopsys05, the present invention has better performance.
Claims
1. A wiring method for FPGA prototype verification system, It is characterized in that The following steps are involved: (1) Convert the wiring diagram of the FPGA prototype verification system into an undirected connected graph; (2) With the goal of shortening the total length of the network and reducing the number of edge reuses, the network is routed based on the approximate algorithm of the minimum Steiner tree; (3) According to the network wiring conditions and constraints, list the mathematical programming and pre-allocate the time division multiplexing ratio; (4) Accurately allocate the time division multiplexing ratio and further optimize it; The step (2) comprises the following steps: (21) Classify the groups: Let the number of edges required by each network be n pi =num(n i )-1; i∈{1, 2, 3, …, |N|}, where num(n i ) indicates network i The number of points contained, |N| represents the number of elements in N, that is, the number of networks; the required number of edges for each group is: g pi =∑ j n pj (n j ∈g i ) According to g pi Sort all groups from largest to smallest and give them new subscripts, where g 1 The group with the largest demand for the opposite side; (22) Wiring the network in g1: Initialize the weight of each edge in E to cost = 2, increase the weight by 2 each time the edge is used, set the threshold α, and divide the network into num(n i )>α and num(n i )≤α; for num(n i )>α, the network is solved by using the minimum Steiner tree approximation algorithm based on the shortest path in descending order according to the number of points contained; for num(n i )≤α, use the disjoint shortest path approximation algorithm to solve, and finally get all the paths of g1; (23) Reorder the remaining groups: After assigning paths to the nets in g1, retain the weights of all edges; do not consider the nets that have been routed, recount the number of edge requirements of the remaining groups, sort them from large to small and give them new subscripts. At this time, g 2 For the group with the largest demand on the opposite side, repeat step (22) to get g 2 Assign paths to unrouted nets; (24) Repeat steps (22) and (23) until all nets are allocated; (25) Split and reroute the largest group: After all the nets are routed successfully, use the estimation function To predict the group with the largest time division multiplexing ratio, where E n is the set of all edges that the network n passes through, N e is the set of all networks passing through edge e, |N e | is the set N e The number of elements in the group; finally, the minimum Steiner tree approximation algorithm is used to redistribute the paths of the networks in the group; The step (3) comprises the following steps: (31) After adopting the time division multiplexing technology, the connection between each pair of FPGAs needs to be multiplexed by multiple networks in a single clock cycle. In order to use multiplexing hardware, the time division multiplexing ratio is required to be an even number. The mathematical planning P is established according to the wiring situation: k∈N (3) Where r(e, n) represents the time division multiplexing ratio allocated when network n uses edge e, and N represents a natural number set; (32) Pre-allocate the time division multiplexing ratio for the network in g1: Allocate a time division multiplexing ratio to all networks n that use edge e equal to the number of times e is used in g1: When n uses e, x(e, n) = 1, otherwise x(e, n) = 0; where r e Indicates g 1 The number of times e is used, r n represents the time division multiplexing ratio of network n, Indicates g 1 The time division multiplexing ratio, Indicates g i Allocatable time division multiplexing ratio; (33) As a benchmark, pre-allocate the time division multiplexing ratio for the remaining groups: set a delta value for each group to indicate the priority of allocating the time division multiplexing ratio to each group: Select the group with the smallest delta, denoted as g2; assign the time division multiplexing ratio to the network using edge e in g2: (34) Update and Repeat (32)-(33) until all nets are allocated.
2. The wiring method of the FPGA prototype verification system according to claim 1, It is characterized in that The implementation process of step (1) is as follows: Convert the wiring diagram of the FPGA prototype system into an undirected connected graph G(V, E), where the i-th FPGA is represented by point v i , i = 1, 2, 3…; the jth edge is e j , j = 1, 2, 3 ...; V = {v 1 , v 2 , v 3 …} is the point set consisting of all FPGAs, E={e 1 ,e2,e 3 …} is the edge set composed of all connections between FGPAs. There can be only one edge between a pair of FGPAs, and the capacity of each edge is 1. Multiple points form a network, and the set of all networks is recorded as N = {n 1 , n 2 , n 3 , …}; if some networks have similar attributes or the same power consumption, they form a group, and the set of all groups is recorded as G = {g 1 , g 2 , g 3 …}.
3. The wiring method of the FPGA prototype verification system according to claim 1, It is characterized in that The step (4) comprises the following steps: (41) Adjust the time division multiplexing ratio and calculate the sum of the inverses of the time division multiplexing ratios of all networks on each edge: The newly allocated time division multiplexing ratio is: If r(e, n) is an odd number, then (r(e, n)+1) is taken to satisfy the restriction formula (1) and formula (2) in the mathematical programming P; recalculate the time division multiplexing ratio of all groups Sort from largest to smallest and give new subscripts; (42) Substitute r(e, n) into the mathematical program P described in step (3) to calculate the gap between each component in the constraint formula (1) and the boundary 1: Then reduce the amount involved The time division multiplexing ratio corresponding to the network, the new time division multiplexing ratio is:
Citation Information
Patent Citations
Practical construction method of logic verification architecture level FPGA wiring device
CN112364590A
