A geometric computational multi-stage pipeline circuit and its parameterization architecture

By cascading the five basic operations of geometric computation into a multi-stage hard-wired pipeline with a fixed topology, data pass-through and hardware parallel computing are achieved, solving the problems of fragmented operations and high latency in existing chips, and improving computing efficiency and real-time performance.

CN123114553APending Publication Date: 2026-11-17ZHENGZHOU XINGYI MEIYA BIOTECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611318997.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing chips, the five basic operations in geometric computing are scattered, and intermediate results need to be transferred through off-chip or shared caches, resulting in large relocation overhead and unpredictable latency. Field quantity translation is not independently hardware-based, and each translation requires reading and writing from the register file, forming a throughput bottleneck. Boundary determination and time mode switching lack direct hardware connection, resulting in high latency. Path integration relies on software Monte Carlo, with latency reaching the millisecond level. The output end lacks multi-condition topology identity determination.

Method used

The five basic operations of geometric computation are cascaded in a fixed topology to form a multi-level hard-wired pipeline with direct data flow between each level. Field quantity translation is implemented as an independent hardware component. Boundary determination and time mode switching are directly connected to hardware. The output provides three-condition topology identity determination, and configuration parameters support switching between different application scenarios.

Benefits of technology

It achieves deterministic latency from field quantity input to topology identity output, eliminates data transfer overhead, improves the hardware parallel computing efficiency of field quantity translation, reduces the latency of boundary determination, and realizes real-time path integration and multi-condition topology determination.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114553A_ABST
    Figure CN123114553A_ABST
Patent Text Reader

Abstract

This invention discloses a multi-stage pipelined circuit and parameterized architecture for geometric computation, belonging to the field of digital integrated circuit technology. The circuit includes a configurable geometric space storage unit, a curvature calculation unit, a field quantity translation unit, a boundary detection and time mode switching unit, a path summation unit, and a three-condition topology detection unit. The field quantity dimension, mesh type, external field weights, and decision thresholds are all configurable, and each stage is hardwired directly, requiring no software migration. This invention achieves a complete computation chain from field quantity input to topology identity output with a fixed pipeline delay, and can be integrated as a standardized IP core into various digital chips.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of digital integrated circuits and dedicated computing chips, specifically relating to a multi-stage pipeline circuit and parameterized architecture for geometric calculations. Background Technology

[0002] In computational tasks such as robot motion planning, physics simulation, molecular dynamics simulation, and financial geometry modeling, the core computational patterns are highly consistent: calculating the local curvature of field quantities in space, translating field quantities along spatial connections, determining the evolution direction based on energy boundaries, accumulating and summing phases along multiple candidate paths, and determining the topological properties of the final state. These five types of operations belong to different processing units in existing chips, and they exchange intermediate data with each other through off-chip memory or shared cache.

[0003] Currently, chips handle this type of calculation primarily in three ways:

[0004] (I) General-purpose processor / GPU software solution. Curvature calculation, field translation, and path sampling are all mapped to instruction sequences on a general-purpose ALU, with intermediate data exchanged between each level of operation via register files and shared memory. The problems with this solution are: "When multiple levels of operations are chained together, the overhead of moving intermediate data accounts for the majority of the computation time; the point-by-point iterative operation of field translation is executed sequentially, which cannot meet the kilohertz-level update rate requirements of real-time control scenarios; path integral sampling requires repeated readings and writing to external storage, resulting in millisecond-level latency for a single simulation, making it unsuitable for deployment in real-time closed-loop systems."

[0005] (ii) Dedicated accelerator solutions for single scenarios. This involves customizing a computational pipeline for a specific dimension and potential function form, hard-wired and solidifying the computational dimension, field form, and boundary conditions. The problem with this solution is that once fabricated, the computational dimension and field form are completely locked, making it unable to cover different application scenarios (two-dimensional curved surfaces / three-dimensional space / high-dimensional space, force field / electric field / temperature field), resulting in a narrow application scope for a single chip and high fabrication risk.

[0006] (III) FPGA Reconfigurable Solution. This solution integrates the computational process and scenario parameters into reconfigurable logic to gain flexibility. The problem with this solution is that the computational density and clock frequency of FPGAs are far lower than those of ASICs at the same process node. More importantly, the configuration overhead of the reconfigurable logic causes the "computational paradigm that must be executed at every step" and the "configuration parameters that only change with the scenario" to be mixed in the same configurable layer. This results in the most frequently executed operations also incurring a reconfigurability cost, leading to a double loss in both area efficiency and power efficiency.

[0007] The above three solutions have the following drawbacks:

[0008] Defect 1: The five basic operations (curvature calculation, field translation, boundary determination, path summation, and topology determination) are scattered in existing chips and lack a dedicated pipeline with a fixed topology. The intermediate results of each operation need to be transferred through off-chip or shared cache, resulting in large relocation overhead and unpredictable latency.

[0009] Defect 2: Field quantity translation is the most frequently executed inner-level iterative operation in multi-stage computation. Existing solutions treat it as a segment of a general instruction sequence rather than an independent operator implemented in hardware. Each iteration of this operation requires multiple multiplications and additions. In software solutions, each translation operation requires reading the connection coefficients and field quantity components from the register file, performing multiplication and addition, and writing back the intermediate results—this "read-calculate-write" loop is repeated tens of thousands of times in each iteration of the pipeline, with data movement overhead accounting for more than half of the computation time. This makes it impossible to achieve dedicated hardware acceleration for field quantity translation, becoming the throughput bottleneck of the entire computation process.

[0010] Defect 3: Lack of a direct hardware connection between boundary determination and time mode switching. In existing solutions, after the boundary detector outputs a determination signal, the switching of time parameters is completed by software or microcode, with a switching delay of hundreds to thousands of clock cycles, which fails to capture the physical process of boundary crossing.

[0011] Defect 4: Phase accumulation and summation of path integrals rely on software Monte Carlo sampling. The phase accumulation of each path is performed point-by-point on a general-purpose ALU in the form of instruction sequences. The sampling of millions of paths requires millisecond-level latency, which is incompatible with the microsecond-level requirements of real-time systems.

[0012] Defect 5: The output only provides a single metric classification judgment (below / critical / exceeding) or a single topological load value, lacking a multi-condition joint topological identity judgment for the final state - it cannot simultaneously confirm "whether it is closed", "whether it is frozen" and "whether it is supported by a sufficient number of paths".

[0013] In summary, there is currently no dedicated circuit in chips that integrates multi-level operations such as curvature calculation, field translation, boundary determination, path summation, and topology determination into a single hard-wired pipeline—with direct data flow between each level, no software migration overhead, and the field dimension and field form can be switched as configuration parameters between different applications. Summary of the Invention

[0014] The technical problem to be solved by the present invention

[0015] This invention aims to solve multiple technical bottlenecks in geometric calculations using existing chips. The specific technical problems are as follows:

[0016] (i) The five basic operations (curvature calculation, field translation, boundary determination, path summation, and topology determination) are scattered across different processing units. Intermediate results need to be transferred through off-chip or shared caches, resulting in large relocation costs and unpredictable latency.

[0017] (ii) Field quantity translation, as the most frequently executed iterative operation, has not been independently hardwareized. Each translation requires reading and writing the connection coefficient and field quantity component from the register file. The "read-calculate-write" loop overhead accounts for more than half of the computation time, forming a throughput bottleneck.

[0018] (iii) There is no direct hardware connection between boundary determination and time mode switching. The determination result needs to be driven by software or microcode to drive the time parameter switching, with a delay of hundreds of clock cycles.

[0019] (iv) Path integral sampling relies on Monte Carlo software for execution, and the path latency of millions of paths reaches the millisecond level, which cannot meet the requirements of real-time systems;

[0020] (v) The output only provides a single metric classification or a single topological load value, lacking the ability to determine the multi-condition joint topological identity of the final state.

[0021] Inventive concept

[0022] This invention includes the following technical solution: five basic geometric calculation operations (curvature calculation, field translation, boundary determination, path summation, and topology determination) are cascaded in a fixed topology as a multi-stage hard-wired pipeline, with direct data flow between each stage, eliminating software migration and cache write-back. The data flow direction of the pipeline is fixed, and the calculation result of each stage is directly sent to the input of the next stage in the next clock cycle.

[0023] The circuit also includes a configurable storage system—spatial grid parameters, field dimensions, external field weights, and various decision thresholds are all stored in the configuration register, supporting the switching of different application scenarios (different dimensions, different physical fields, and different decision criteria) on the same hardware.

[0024] This includes the following three aspects:

[0025] (I) Multi-stage fixed topology pipeline. The first stage uses a differential array to calculate the local curvature of space; the second stage uses a vectorized multiplier array to perform iterative translation of field quantities along spatial connections (including convergence determination and maximum iteration protection); the third stage uses a three-input comparator to calculate the energy boundary criterion and trigger time mode switching; the fourth stage uses a parallel phase accumulation chain to accumulate the phase of multiple candidate paths and sum them; the fifth stage uses a three-condition parallel detection output to determine the topology identity. Data is directly connected between stages without software intervention.

[0026] (ii) Independent hardware implementation of field quantity translation. Field quantity translation, as a fundamental operator shared by all spatial calculations, is implemented independently using a vectorized interconnected multiplier array: multiple multiply-accumulators calculate the translation components in each direction in parallel, including a fixed-point iterative convergence loop, and output the translated field quantity cross section. This operator can be reused as an independent module.

[0027] (III) Configurable Parameter System. Spatial grid type, field dimension, external field weight, and decision thresholds at each level are all stored in configurable registers and weight storage. The weight storage adopts a double-buffered structure, supporting runtime switching without interrupting the pipeline. The same hardware can be configured to cover application scenarios with different dimensions, different external field forms, and different decision criteria.

[0028] Overall technical solution

[0029] This invention hardware-encodes five fundamental operations in geometric computation using a fixed multi-stage pipeline, with direct data flow between stages, completing the entire computation chain from field input to topological identity output with a deterministic delay of a fixed clock cycle. Using configurable geometric space storage units, curvature difference arrays, interconnected multiplier arrays, energy boundary comparators, parallel phase accumulator arrays, and a three-condition topology detection unit as core components, the following computation chain is solved in hardware parallelism:

[0030] ① Curvature calculation: The local curvature κ² is calculated by contraction of the second-order difference chain and the connection term;

[0031] ② Field quantity translation: The field quantity is translated along each direction using a vectorized multiplier array, and the process is iterated at a fixed point until convergence;

[0032] ③ Boundary determination: The boundary criteria are calculated using a three-energy comparator to trigger a time mode switch;

[0033] ④ Path summation: Accumulate the path phase using K parallel phase accumulation chains, sum them using an addition tree, and output the global cross section;

[0034] ⑤ Topology determination: Output topology identity flags by detecting three conditions in parallel.

[0035] Basic pipeline timing rules: After the field quantity is input, curvature calculation is completed by the differential array in cycle 1; field quantity translation is completed iteratively by the interconnected multiplier array from cycle 2 to cycle (2+M) (M is the number of iteration convergence steps, maximum 64); boundary determination is completed by the comparator in the next cycle; path summation is completed by the accumulator chain and the addition tree in subsequent cycles; and topology determination is output by the three-condition detection unit in the final cycle. The initial output delay depends on the number of iteration convergence steps (typically no more than 70 cycles). After the pipeline is full, one topology identity determination result is output every cycle. There are no branches and no software intervention throughout the entire process.

[0036] Specific technical solutions

[0037] (i) Configurable Geometric Space Storage Unit. Used to store the mesh structure parameters and field quantity dimension parameters of the computational space. The mesh structure parameters include a mesh type register (2-bit, encoding two-dimensional surfaces, three-dimensional space, or high-dimensional space) and resolution register groups for each dimension; the field quantity dimension parameters are configurable register values ​​(typically 3, 4, or 8). This unit provides mesh geometry information and field quantity data width information for each stage of the pipeline.

[0038] (ii) First-level curvature calculation unit. Connected to the configurable geometric space storage unit, it receives field cross-sectional values ​​and mesh geometric information, and calculates the local curvature of space using a differential calculation unit array. The differential calculation unit array consists of hard-connected second-order differential chains and connection term contraction units in each coordinate direction. It is a pure combinational logic, outputting curvature values ​​in a single cycle. The curvature values ​​also serve as the curvature energy input for the third-level boundary determination.

[0039] (III) Second-level field quantity translation unit. Connected to the first-level curvature calculation unit, it receives the current field quantity cross section, connection matrix, and external field weights, and outputs the translation cross section along each direction in parallel using a vectorized connection multiplier array. The parallelism of the vectorized connection multiplier array is equal to the field quantity dimension N. Each multiplier-adder calculates the translation components in each direction in parallel, and outputs the translation cross section after merging through a summation tree.

[0040] This unit executes iteratively at fixed points: the current cross-sectional component is multiplied and added with the connection coefficient to obtain the translational cross-section. Convergence occurs when the residual norm of two adjacent iterations is less than a preset threshold, and the final translational cross-section is output. If convergence fails, the translational cross-section is written back to the cross-section register group via the iteration control unit, and the next iteration continues. The iteration control unit includes a maximum iteration counter (typically 64) to prevent infinite loops in divergent scenarios. The connection matrix is ​​provided by a programmable weight storage, and the external field weights are provided by another programmable weight storage, both of which support runtime rewriting.

[0041] (iv) Third-level boundary detection and time mode switching unit. Connected to the second-level field quantity translation unit, it receives three energy inputs—curvature energy (from the first level). The boundary criterion A = curvature energy + ambient energy − locking energy is calculated using a three-input comparator, and the criterion sign is output. The boundary criterion is calculated using a three-input comparator, which calculates the weighted value of the curvature energy, the locking energy (from the structure locking item in the configuration register), and the ambient energy (from the shrinkage value of the external field weights).

[0042] This unit is also connected to a time-mode switching state machine (containing two stable states: a normal evolution state and a relaxed convergence state). When the criterion symbol flips, it switches the time parameter mode within a single clock cycle and outputs a mode flag and a flip count. The response time of the time parameter mode switching is less than a single clock cycle.

[0043] (V) Fourth-level path summation unit. Connected to the third-level boundary detection and time mode switching unit, it generates K candidate paths using an LFSR pseudo-random path generator (K equals the hardware parallelism, configurable to 64 or 256), and accumulates the connection phase of each path point by point using K parallel phase accumulation chains (including modulo 2π winding processing to prevent overflow of long paths). The complex phase of the order complex number and the addition tree are summed over the complex phase of K paths, and the global cross section is output by normalized division unit.

[0044] (vi) Fifth-level three-condition topology detection unit. Connected to the fourth-level path summation unit, three parallel sub-units detect the three topological conditions of the final cross-section respectively: closure determination sub-unit (combining the fractional part of the accumulated cyclic phase value with the threshold). The comparison is as follows: (1) The sub-unit determines closure if the value is less than a threshold (typically 0.1); the sub-unit determines freezing by comparing the number of time pattern flips with a flip threshold (typically 1); and the sub-unit determines support by comparing the proportion of convergence paths with a threshold. For comparison, the typical value is 0.7 (for decision support). The three decision signals are ANDed by an AND gate to output a topology identity flag. A valid flag is output when all three conditions are met simultaneously.

[0045] The circuit front end can be cascaded with a multi-channel sensor synchronous sensing circuit and a hybrid precision adaptive quantization circuit, and the back end can be cascaded with a collapse projection output operator circuit and a critical trigger freezing circuit. The application layer configuration is completed by the host through the configuration bus, which writes the grid structure, field dimension, connection weight and path number parameters, without entering the hardware pipeline.

[0046] Basic pipeline timing rules: After the field quantity is input, curvature calculation is completed in cycle 1, field quantity translation iterates from cycle 2 until convergence (M cycles, maximum 64), boundary determination is completed in cycle M+1, path summation starts pipeline output from cycle M+2, and topology determination is output in the final cycle. The pipeline depth is fixed throughout, with no software intervention.

[0047] Beneficial effects

[0048] (i) This invention uses a fixed topology to cascade five basic operations in geometric calculations into a multi-level hard-wired pipeline, with data flow between each level and intermediate results not written back to the chip or shared cache, eliminating the data migration overhead in traditional schemes, and completing the complete calculation chain from field quantity input to topology identity output with deterministic delay.

[0049] (ii) Field translation, as the most frequently executed iterative operation, is implemented independently in hardware: the vectorized interconnected multiplier array calculates the translation components in each direction in parallel, and each multiplier-adder outputs in a single cycle; it includes a fixed-point iterative convergence loop, and automatically outputs the final cross section after convergence. This operator can be reused in all spatial computing scenarios, fundamentally eliminating the throughput bottleneck of the "read, calculate, write" loop.

[0050] (iii) A hard-wired direct connection is used between boundary determination and time mode switching. After the criterion symbol is flipped, the state machine completes the time parameter mode switching within a single clock cycle. The response time is less than a single clock cycle (typical value of 1ns at 1GHz), which solves the problem of software / microcode switching delay of hundreds of clock cycles.

[0051] (iv) The K phase accumulation chains of path summation work in parallel. When the pipeline is fully loaded, the phase accumulation and summation of one path is completed in each cycle. The sampling of millions of paths is completed in microseconds (the software solution requires milliseconds), which is about three orders of magnitude faster.

[0052] (v) The output provides a three-condition joint topology identity determination (closure degree + freeze degree + support degree). When all three conditions are met, a valid flag is output. This is different from the existing single-measure hierarchical determination and provides complete topology status information for downstream execution units.

[0053] (vi) Spatial grid type, field dimension, external field weight, and decision threshold are all configurable parameters; the weight storage adopts a double-buffered structure to support runtime switching without interrupting the pipeline. The same chip can be configured to cover different application scenarios such as two-dimensional curved surfaces, three-dimensional space, high-dimensional space, and force fields, electric fields, and temperature fields.

[0054] Terminology Definition

[0055] Geometric space: refers to the underlying spatial structure in which calculations occur. In robot motion planning, it is the configuration space; in physics simulation, it is the parameter space. Different application scenarios correspond to different dimensions and metric structures.

[0056] Field cross section: refers to the value of a field quantity defined in space at the current location. It consists of multiple components, and the number of components is equal to the field quantity dimension N.

[0057] curvature : A scalar measure of the degree of local curvature in space, calculated from the metric tensor and its first and second partial derivatives, characterizing the geometric curvature intensity at the current position.

[0058] Field quantity translation: refers to the operation of moving a field quantity defined in space to an adjacent position along the connection direction. It is the most basic and frequently performed iterative operation in spatial computing.

[0059] Connection matrix: A coefficient matrix that describes how field quantities are related between adjacent locations in space. It is determined by the metric structure of space and corresponds to different physical coupling relationships in different physical fields.

[0060] Energy boundary criterion A: refers to the boundary scalar obtained by weighted combination of curvature energy, locking energy and ambient energy. The sign of A determines whether the evolving system should be in normal evolution mode or relaxed convergence mode.

[0061] Time mode switching: refers to the switching between the normal evolution mode and the relaxation convergence mode of the evolving system. The normal evolution mode is used for the propagation of the field along the current direction, while the relaxation convergence mode is used for the rapid convergence of the field towards the energy minimum.

[0062] Path integral: refers to the operation of summing the phase contributions of all possible paths according to their weights, and is a core mathematical tool in Monte Carlo sampling.

[0063] Topological identity: This refers to whether the final cross section simultaneously satisfies three topological conditions: closure (forming an integer number of loops), freezing (locked in a critical state), and support (convergence of sufficient paths). A valid topological identity flag is output when all three conditions are met. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the overall architecture of the multi-stage pipeline circuit for geometric computation of the present invention. The left side of the diagram shows the field quantity input and configuration parameter input; the solid arrow indicates the positive data flow direction; the dotted-dash arrow indicates the configuration control signal; the six modules marked with circles ① to ⑥ are labeled sequentially (① configurable geometric space storage unit, ② first-stage curvature calculation unit, ③ second-stage field quantity translation unit, ④ third-stage boundary detection and time mode switching unit, ⑤ fourth-stage path summation unit, ⑥ fifth-stage three-condition topology detection unit); the pipeline boundaries are separated by vertical dashed lines and labeled T1 / T2 / T3 / T4 / T5 (where T2 includes multiple cycles of iterative convergence); the data flow direction is fixed, and the hardwired connections between each stage are straight through.

[0065] Figure 2 This is a schematic diagram of the internal structure of the second-stage field quantity translation unit of the present invention. The figure shows the cross-section register group, the connection matrix storage volume, the vectorized connection multiplier array (N-way parallel multiply-adder), the summation tree, the convergence criterion and the iteration control unit; the solid arrows are the main data streams, the dashed arrows are the feedback iteration paths, and the dotted-dash arrows are the configuration control signals. Detailed Implementation

[0066] Example 1: Complete multi-level implementation (three-dimensional space, mechanical field)

[0067] like Figure 1The overall architecture shown in this embodiment uses the calculation of mechanical field quantities in three-dimensional space as an example to verify the functional correctness and timing performance of the geometric calculation multi-stage pipeline circuit.

[0068] Configuration parameters: The mesh type is three-dimensional space, the resolution is 64×64×64, the field dimension N=3, the number of paths K=64, and the data path adopts 16-bit fixed-point Q4.12 format.

[0069] Pipeline timing:

[0070] Period T1: Configurable geometric space storage unit outputs mesh geometric information and field dimension parameters; the first-level curvature calculation unit calculates the local curvature κ² using a differential calculation unit array, and outputs the curvature value to the second and third levels in a single period.

[0071] Period T2 to T(2+M): The second-level field quantity translation unit receives the current field quantity cross section and the connection matrix, and the vectorized connection multiplier array (3-way parallel) calculates the translation cross section in parallel along each direction; the convergence criterion calculates the residual norm of the cross section of two adjacent iterations, with a typical convergence threshold of 1e-4; when the residual norm is less than the threshold, the iteration is terminated and the final translation cross section is output (typical convergence steps M is about 4 to 6 steps), and if it does not converge, the iteration continues (maximum 64 steps protection).

[0072] Period T(M+1): The third-level boundary detection and time mode switching unit receives curvature energy (from...) The boundary criterion A is calculated as follows: A = Curvature Energy + Environmental Energy - Locking Energy; the criterion symbol detector outputs a symbol bit; when the criterion symbol flips, the time mode switching state machine switches modes within a single clock cycle and outputs a mode flag and a flip count. During mode switching, the mode flag and flip count output by the third-level boundary detection and time mode switching unit are simultaneously output to the fourth-level path summing unit. When the flip pulse is valid, a clear signal is synchronously sent to the K parallel phase accumulation chains of the fourth-level path summing unit, and the accumulation length counter of each phase accumulation chain is synchronously reset to zero, and the accumulated phase register is cleared. The clear operation is completed within the same clock cycle of mode switching, and is executed in parallel with the state register flip, without adding additional delay.

[0073] Starting from period T(M+2): the fourth-level path summation unit generates K candidate paths using the LFSR pseudo-random path generator; the K parallel phase accumulation chains accumulate the phase of each path point by point; Pipeline summation of complex numbers and addition trees (6 levels); output of global cross-section for normalized division units.

[0074] Final cycle: The fifth-level three-condition topology detection unit performs parallel detection of closure (phase fractional part < 0.1), freeze degree (number of flips ≥ 1), and support degree (convergence path ratio ≥ 0.7); when all three conditions are met, the topology identity flag is output.

[0075] Actual results: The field quantity translation iteration core is fully hardwired and executed in parallel, reducing data transfer overhead to zero. The end-to-end computation latency is approximately 70 cycles (including iteration convergence). Once the pipeline is full, one topology identity determination result is output per cycle.

[0076] Hardware Resource and Timing Comprehensive Verification: Based on a 28nm standard process and a 1GHz clock frequency, the equivalent gate count of each module in this circuit is estimated as follows: approximately 1k gates for configurable geometric space storage, approximately 2k gates for the first-stage curvature calculation unit, approximately 5k gates for the second-stage field translation unit (N=3, including iterative control), approximately 1k gates for the third-stage boundary detection and time mode switching unit, approximately 8k gates for the fourth-stage path summation unit (K=64), and approximately 1k gates for the fifth-stage three-condition topology detection unit, totaling approximately 18k gates. The pipeline critical path consists of four levels of combinational logic: differential array, interconnected multiplier array, summation tree, and comparator. Static timing analysis shows sufficient establishment and hold margins, with no timing violations. This circuit, with approximately 18k gates, can be integrated as a standardized IP core into robot controllers, scientific computing accelerator boards, or tensor computing chips.

[0077] Example 2: Dimension / Field Parameterization and Runtime Switching

[0078] This embodiment verifies the ability of the same circuit to switch between different application scenarios by configuration.

[0079] Scenario A (3D mechanical field): The mesh type register is configured as 3D, the field dimension N=3, the external field weights are loaded with the mechanical field weight set, and the threshold is determined. , .

[0080] Scenario B (2D surface): The mesh type register is configured as 2D, the field dimension N=2, the external field weights are loaded into the surface field weight set, and the judgment threshold is adjusted accordingly.

[0081] Switching process: The host computer writes the new scene's mesh parameters, field dimensions, and external field weights (including multiple pre-loaded weight sets in a double-buffered storage) to the configuration register and weight storage via the configuration bus. The weight storage uses a double-buffered structure, ensuring the main pipeline remains unaffected while the new configuration is being written to the back buffer. After writing is complete, the load selection signal is toggled, and the system automatically switches to the new configuration in the next clock cycle. The switching process does not require pausing the pipeline. The same hardware can cover application scenarios with different dimensions and different external field types.

[0082] Example 3: Path integral summation and geodesic utilization calculation

[0083] This embodiment verifies the function of the fourth-level path summation unit in path integral calculation.

[0084] Configuration parameters: number of paths K=256, weight type register configured to Boltzmann weight mode, kT configured to a value corresponding to the system temperature.

[0085] Working principle: The LFSR pseudo-random path generator generates 256 candidate paths; each phase accumulation chain accumulates the connection phase of the path point by point, including modulo 2π winding processing; each chain outputs a weighted complex phase (real part) imaginary part ), where w is the Boltzmann weight. ; The complex number and addition tree (8 levels) performs pipelined summation on 256 complex phases; the normalized division unit outputs the global cross-section.

[0086] When the target path is the optimal path and the other paths are deviation paths, the normalized output directly corresponds to the global response of the path integral, realizing real-time hardware calculation of physical quantities, replacing traditional software table lookup and post-processing.

Claims

1. A geometric computational multi-stage pipeline circuit, characterized in that, include: A configurable geometric space storage unit is used to store the grid structure parameters and field dimension parameters of the computational space. The grid structure parameters include a grid type field and a resolution field, and the field dimension parameters are configurable register values. The first-level curvature calculation unit is connected to the configurable geometric space storage unit, receives field cross-sectional values ​​and mesh geometric information, calculates local curvature in space using a differential calculation unit array and outputs it. The differential calculation unit array is composed of second-order differential chains in each coordinate direction and hard-wired connection terms contraction units. The second-level field quantity translation unit is connected to the first-level curvature calculation unit. It receives the current field quantity cross section, the connection matrix and the external field weights, and outputs the translation cross section along each direction in parallel using a vectorized connection multiplier array. It iterates in a fixed-point iterative manner until the convergence criterion is met, and outputs the final translation cross section. The connection matrix and the external field weights are provided by a programmable weight storage. The third-level boundary detection and time mode switching unit is connected to the second-level field quantity translation unit. It receives three energy inputs: curvature energy, locking energy, and ambient energy. It calculates the boundary criterion symbol using a comparator and a symbol detection unit, and outputs a time mode switching trigger signal. The unit is also connected to a time mode switching state machine. When the trigger signal is valid, it switches the time parameter mode within a single clock cycle and outputs a mode flag and a toggle count. The fourth-level path summation unit is connected to the third-level boundary detection and time mode switching unit. It accumulates and sums the phases of multiple candidate paths using a parallel phase accumulator array and outputs the global cross-section. The number of parallel paths of the path summator is equal to the hardware parallelism. The fifth-level three-condition topology detection unit is connected to the fourth-level path summation unit. It uses three parallel sub-units to detect the three topological conditions of closure, freezing degree and support degree of the final cross section. When all three conditions are met, the topological identity flag is output. The data flow between the first to the fifth level is a hard-wired fixed topology, with direct data flow between levels without intermediate buffers or write-back. The spatial grid structure, field dimension, and external field weight are configurable parameters. The calculation paradigm is fixed in the hardware data path, and the specific dimension and specific field are determined by the configuration parameters.

2. The geometric computational multi-stage pipeline circuit according to claim 1, characterized in that, The external field communication weight storage adopts a double-buffered structure, including a first buffer and a second buffer. The current iteration uses the weights in the first buffer, and the new configuration is written to the second buffer. After the writing is completed, the load selection signal is flipped to enable the execution unit to switch automatically in the next clock cycle. The switching process does not require pausing the pipeline. The typical encoding of the grid type field includes two-dimensional, three-dimensional and high-dimensional spaces. The typical value of the field quantity dimension parameter is 3, 4 or 8.

3. The geometric computational multi-stage pipeline circuit according to claim 1, characterized in that, The curvature calculation of the differential calculation unit array is a single-step combinational logic: each coordinate direction is composed of two adjacent differential steps connected to form a second-order differential chain, and the first derivative of the metric is contracted by Christofel sign through the connection term contraction unit to output the scalar curvature value. All data paths adopt the 16-bit fixed-point Q4.12 format.

4. The geometric computational multi-stage pipeline circuit according to claim 1, characterized in that, The second-stage field quantity translation unit also includes a convergence criterion and an iteration control unit. The convergence criterion compares the residual norm of two adjacent iteration sections with a preset threshold. The iteration control unit contains a maximum iteration counter. When the residual norm is less than the threshold or the number of iterations reaches the maximum value, the iteration is terminated and the final translation section is output. The typical value of the threshold is 1e-4, and the typical value of the maximum number of iterations is 64. When convergence is not determined, the iteration control unit writes the translation section back to the section register group to continue the next round of iteration. The entire iteration process is completed on-chip without external storage read / write.

5. The geometric computational multi-stage pipeline circuit according to claim 1, characterized in that, The state machine of the third-level boundary detection and time mode switching unit includes two stable states: normal evolution state and relaxed convergence state. When the trigger signal is valid, the time parameter mode is switched within a single clock cycle, and the mode flag and flip count are output. The boundary criterion A = curvature energy + ambient energy − locking energy. Switching is triggered when the sign of A flips. The switching response time is less than a single clock cycle; the mode flag and the toggle count are simultaneously output to the fourth-level path summation unit, which is used to clear each phase accumulation chain and reset the accumulation length counter when switching modes.

6. The geometric computational multi-stage pipeline circuit according to claim 1, characterized in that, The fourth-level path summation unit includes: a pseudo-random path generator for generating candidate paths according to the number of configurable paths K; and K parallel phase accumulation chains, each chain accumulating the connection phase of a single path point by point, including modulo 2π winding processing. The system consists of a complex number and an addition tree, which sums the complex phases of K paths; a normalized division unit that outputs the global cross-section; the typical value of the number of paths K is 64 or 256, and is equal to the hardware parallelism; the weight type register of the fourth-level path summation unit is configured to uniform weight or Boltzmann weight mode, and in Boltzmann mode, the normalized division unit outputs the global cross-section of the path integral.

7. The geometric computational multi-stage pipeline circuit according to claim 1, characterized in that, The fifth-level three-condition topology detection unit includes: a closure determination subunit, which compares the fractional part of the cumulative value of the cyclic phase with the closure threshold, and determines closure if it is less than the threshold; and a freeze determination subunit, which compares the number of time pattern flips with the flip threshold to determine freeze. The support determination subunit compares the proportion of convergence paths with the support threshold to determine support; the three-condition AND gate outputs the topology identity flag. The typical value for the closure threshold is 0.1, the typical value for the support threshold is 0.7, and the typical value for the flip threshold is 1.

8. The geometric computational multi-stage pipeline circuit according to claim 1, characterized in that, The parallelism of the vectorized interconnected multiplier array of the second-stage field quantity translation unit is equal to the field quantity dimension N. Each multiplier-adder calculates the translation components in each direction in parallel. The outputs of each path are merged by a summation tree to output the translation cross section. The entire combinational logic is completed in a single cycle.

9. The geometric computational multi-stage pipeline circuit according to claim 1, characterized in that, The circuit front end is cascaded with a multi-channel sensor synchronous sensing circuit and a hybrid precision adaptive quantization circuit, and the back end is cascaded with a collapse projection output operator circuit and a critical trigger freezing circuit. The application layer configuration is completed by the host through the configuration bus, which writes the grid structure, field dimension, connection weight and path number parameters, without entering the hardware pipeline.