Automatic compilation method and framework for generating integrated circuit layout

Through automatic compilation methods and frameworks, the layout of integrated memory computing circuits is automatically generated, which solves the problem that DCIM design relies on manual operations and circuit layout generators in the existing technology that cannot meet the integrated memory computing requirements, and realizes efficient and accurate DCIM circuit design , improving the application scope of artificial intelligence computing.

CN119067043APending Publication Date: 2024-12-03AI CHIP CENTER FOR EMERGING SMART SYSTEMS LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310944954.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2023-05-31
Filing Date
2023-07-28
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing digital design process based on RTL cannot be directly applied to integrated memory computing circuits, resulting in DCIM circuit design relying on manual operations, and the existing circuit layout generator cannot meet the requirements of integrated memory computing, limiting DCIM in artificial intelligence applications. scope of application.

Method used

Provides an automatic compilation method and framework to automatically generate the layout of integrated memory computing circuits by building a design database, receiving user specifications, selecting reference architecture templates, parameterizing, generating physical layouts, iterative design, optimizing and evaluating hardware indicators.

Benefits of technology

It realizes automatic generation of DCIM circuits, provides fast, accurate and acceptable performance generation results, meets user needs, solves the efficiency and accuracy problems of traditional design processes, and achieves energy efficiency improvement.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119067043A_ABST
    Figure CN119067043A_ABST
Patent Text Reader

Abstract

An automatic compilation method for generating a layout of a circuit design for integrated memory computing. The automatic compiling method comprises the following steps: constructing a design database which is provided with at least one basic design library and a set of reference architecture templates; receiving a user specification associated with a circuit design of the integrated memory computation; selecting a reference architecture template from the design database, and parameterizing the selected architecture template according to the received user specification to generate an architecture of the circuit design of the integrated memory calculation; generating a physical layout based on the generated architecture according to one or more layouts and routing algorithms; iteratively designing an exploration loop on the generated layout with one or more optimization schemes based on expert experience to obtain one or more explored physical layouts; and evaluating the explored hardware metrics of the one or more physical layouts under one or more constraints to determine an optimal physical layout.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to circuit design for integrated in-memory computing, and more particularly, to an automatic compilation method and framework for generating circuit design layout for integrated in-memory computing. Background Art

[0002] Computing-in-Memory (CIM) architecture has been recognized as a promising computing paradigm for efficient artificial intelligence (AI) computing. Traditional digital architectures usually have a lot of data movement between their separate computing and memory. CIM solves the memory wall problem by integrating multiply-accumulation (MAC) operations into static random access memory (SRAM) or non-volatile memory.

[0003] Digital CIM (DCIM) has become an emerging type of CIM architecture. DCIM embeds digital MAC calculation logic into SRAM memory in a distributed manner, which is different from analog CIM (ACIM) that usually relies on the sum of current / voltage along the columns of the memory array for MAC calculation. In addition to the benefits of CIM architecture, DCIM has no DAC or ADC overhead, and all-digital calculations do not encounter non-ideal analog problems. Therefore, DCIM is an efficient and high-precision architecture with strong technical expansion capabilities.

[0004] However, since DCIM-like circuits are compact hybrids of memory and digital logic, the existing RTL-based digital design flows cannot be directly applied to the implementation of such integrated memory computing circuits. Current DCIM designs rely heavily on manual operations. People design a DCIM macrocell architecture, from basic memory cells, bit-by-bit multipliers to the entire array. The physical layout must be carefully optimized to meet certain user specifications (architectural parameters such as macrocell height and width, and hardware limitations such as area and delay). Once the specifications change, the time-consuming manual design process will have to be run again, which will limit the application of DCIM to various AI applications.

[0005] At present, on the one hand, some memory compilers have been used to automatically generate arbitrary memory arrays. However, most of them are foundry-provided or open source SRAM memory compilers (e.g., OpenRAM), which are not applicable because these can only generate SRAM memories for reading and writing without computing capabilities. On the other hand, there are also reports that there are some circuit layout generators that can automatically generate analog circuit modules. For example, OpenSAR was developed for the automatic design of analog building blocks based on SAR ADC, such as sample-and-hold circuits and comparators. However, these existing reports on circuit layout generators only solve traditional circuit design problems and cannot meet the requirements of integrated memory computing circuits.

[0006] Therefore, we still look forward to flexible implementation methods of DCIM-type circuits, so as to establish a good ecosystem for DCIM-based artificial intelligence computing. Summary of the invention

[0007] An object of the present invention is to provide an automatic compilation method and framework to solve the above technical problems.

[0008] According to a first aspect of the present invention, an automatic compilation method is provided for generating a layout of a circuit design for integrated memory computing. The automatic compilation method includes: constructing a design database having at least one basic design library and a set of reference architecture templates; receiving a user specification associated with the circuit design for integrated memory computing; selecting a reference architecture template from the design database, and parameterizing the selected architecture template according to the received user specification to generate an architecture of the circuit design for integrated memory computing; generating a physical layout based on the generated architecture according to one or more placement and routing algorithms; iterating a design exploration cycle on the generated layout with one or more optimization schemes based on expert experience to obtain one or more explored physical layouts; and evaluating the hardware index measurement of the one or more explored physical layouts under one or more constraints to determine the optimal physical layout.

[0009] According to a second aspect of the present invention, an automatic compilation framework is provided for generating a layout of a circuit design for integrated memory computing. The automatic compilation framework includes: a design database, which is constructed by a collection of at least one basic design library and reference architecture templates; an automatic layout generator, which includes: an architecture generator, which is configured to: receive user specifications associated with the circuit design of integrated memory computing; select a reference architecture template from the design database, and parameterize the selected architecture template according to the received user specifications to generate the architecture of the circuit design of integrated memory computing; and a layout and routing processor, which is configured to: generate a physical layout based on the generated architecture according to one or more layout and routing algorithms. The framework also includes a design explorer, which is configured to iterate a design exploration cycle on the generated layout using one or more expert experience-based optimization schemes to obtain one or more explored physical layouts; and a hardware performance evaluator, which is configured to measure the hardware indicators of the one or more explored physical layouts under constraints to determine the optimal physical layout.

[0010] The present invention provides a solution for automatically generating integrated memory computing circuits, which achieves many energy efficiency improvements compared to traditional computing architectures. On the one hand, the present invention provides automatic generation of emerging integrated memory computing circuits, which cannot be achieved with existing memory compilers and other traditional circuit generators. On the other hand, the present invention combines automatic tools with expert experience to quickly and accurately provide generation results with acceptable performance to meet user needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Embodiments of the present invention will be described in more detail below with reference to the accompanying drawings, in which:

[0012] Figure 1 A flow chart of an automatic compilation method according to an embodiment of the present invention is shown, which is a layout for generating a circuit design for integrated in-memory computing;

[0013] Figure 2 A block diagram of an automatic compilation framework for implementing an automatic compilation method is shown;

[0014] Figure 3 An embodiment of a memory cell in the present invention is shown;

[0015] Figure 4 An embodiment of a local logic element in the present invention is shown;

[0016] Figure 5 An embodiment of the peripheral circuit in the present invention is shown;

[0017] Figure 6An exemplary architectural template of a circuit for integrated in-memory computing according to one embodiment of the present invention is shown;

[0018] Fig. 7A A column template having m weight units and one bitwise multiplier is shown; Figure 7B shows how the automatic compilation framework with m=4, n=4 generates a DCIM component layout;

[0019] Fig. 8A shows a local adder template, Figure 8B Shown by following Fig. 8A The template of an exemplary local adder generated by the automatic compilation framework;

[0020] Fig. 9 shows an expert-based design exploration with three different types of array partitioning schemes according to one embodiment of the present invention; and

[0021] Fig. 10A and Fig. 10B Two examples of physical layouts are shown, with and without array partitioning, respectively. DETAILED DESCRIPTION

[0022] In the following description, embodiments of the automatic compilation framework and method are described as preferred examples, which are layouts for generating circuit designs for integrated memory computing. Modifications, including additions and / or substitutions, may be made to those skilled in the art without departing from the scope and spirit of the invention. Specific details may be omitted to avoid obscuring the invention; however, this disclosure is written to enable those skilled in the art to implement the teachings of the invention without undue experimentation.

[0023] Figure 1 A flow chart of an automatic compilation method according to an embodiment of the present invention is shown, which is used to generate a layout of a circuit design for integrated memory computing. Figure 2 A block diagram of an automatic compilation framework for implementing an automatic compilation method is shown.

[0024] refer to Figure 1 , the automatic compilation method comprises:

[0025] Step S102: constructing a design database having at least one basic design library and a set of reference architecture templates;

[0026] Step S104: receiving a user specification associated with a circuit design for integrated in-memory computing;

[0027] Step S106: Selecting a reference architecture template from the design database, and parameterizing the selected architecture template according to the received user specification to generate an architecture of the circuit design of integrated memory computing;

[0028] Step S108: generating a physical layout based on the generated architecture according to one or more layout and routing algorithms;

[0029] Step S110: iteratively designing and exploring the generated layout using one or more optimization solutions based on expert experience to obtain one or more explored physical layouts; and

[0030] Step S112: Evaluate hardware metrics of the explored one or more physical layouts to determine an optimal physical layout.

[0031] refer to Figure 2 , the automatic compilation framework may be composed of a design database 210 , an automatic layout generator 220 , a design explorer 230 , and a hardware performance evaluator 240 .

[0032] The design database 210 may include a basic design library, a collection of reference architecture templates, and necessary technical files, such as standard design check rules (DRC) and layout and schematic (LVS) rules.

[0033] The basic design library may include designs of various customized functional units, such as memory cells, local logic elements of partially integrated memory computing functional units, and peripheral computing circuit elements.

[0034] Reference architecture templates can be extracted from expert-based or user-defined designs. Basic design libraries and top-level architecture templates are the basic requirements for providing fast and accurate circuit generation solutions.

[0035] The automatic layout generator 220 may include an architecture generator 221 and a placement and routing processor 222 .

[0036] The architecture generator 221 can be configured to: receive user specifications associated with a circuit design for integrated memory computing; select a reference architecture template from a design database, and parameterize the selected architecture template according to the received user specifications to generate an architecture for the circuit design for integrated memory computing.

[0037] The user specification may include computer readable code and one or more constraints associated with circuit design for integrated in-memory computing. The selected architecture may include memory cells, local logic elements, and peripheral circuit elements.

[0038] The place and route processor 222 may be configured to generate a physical layout based on the generated architecture according to one or more place and route algorithms.

[0039] The design explorer 230 may be configured to iterate the design exploration cycle on the generated layout using one or more expert experience-based optimization schemes to obtain one or more explored physical layouts, thereby optimizing the circuit quality and electrical performance of the physical layout, such as critical paths and parasitic effects, layout density, etc. Various integration schemes of basic modules will affect the overall performance, such as different partition sizes, partition methods, etc.

[0040] The hardware performance evaluator can be configured to measure hardware metrics for one or more explored physical layouts under constraints to determine the best physical layout and generate a corresponding performance report. The constraints can include area priority or critical path delay priority.

[0041] refer to Figure 3 Specifically, the memory cells may include, but are not limited to, volatile memory cells, including static memory, dynamic random-access memory (DRAM) and their variants, and non-volatile memory cells, including flash memory, resistive random-access memory (RRAM), magnetoresistive random-access memory (MRAM) and ferroelectric random-access memory (FRAM) and their variants.

[0042] refer to Figure 4 , the local logic elements may include, but are not limited to, multipliers, partial sum adders, and any other suitable functional logic computing circuits.

[0043] refer to Figure 5 , the peripheral circuits may include but are not limited to row drivers, column drivers, computational processing units, control logic, shift accumulators, precision fusion units, and any other suitable functional modules.

[0044] Figure 6 An exemplary architectural template for an integrated in-memory computing circuit, which is an SRAM-based H×W DCIM macrocell, is shown according to one embodiment of the present invention. The template is parameterized to expose a large design space for various user-defined DCIM specifications.

[0045] The DCIM macrocell consists of a DCIM array and digital peripheral circuits (row / column drivers, shift accumulators, precision fusion units, and controllers). The DCIM array consists of multiple columns of DCIM elements and their local adders. When calculating matrix multiplication in the AI ​​model, the DCIM stores the weight matrix in the SRAM cells of the DCIM array and loads the input matrix through the row driver. The DCIM element receives the input and performs a local MAC operation. The output of the array is sent to the shift accumulator and the post-processing circuit to obtain the result.

[0046] The elements of the H×W DCIM macrocell are further described as follows: The DCIM element is the basic building block of the DCIM array, which has two groups of m×n SRAM-based weight cells and 1b×nb multipliers. The two groups are placed face-to-face to achieve a compact physical layout. The DCIM element width (n) is equivalent to the maximum granularity used to construct the supported weight precision. For example, if the macrocell requires 8b and 12b weight precision, n should be 4 to construct 2 and 3 DCIM elements in the horizontal direction to achieve the mapping of these two precisions. The weight multiplier ratio (m) can be defined in the user specification. This allows m rows of weight cells to reuse a 1b×nb multiplier. Therefore, in a DCIM macrocell with H rows and W columns of SRAM cells, W / n DCIM elements in each row are placed horizontally to match the width of the macrocell, and H / 2m DCIM elements in each column are placed vertically to match the height of the macrocell.

[0047] A local adder is a building block placed next to each column of DCIM elements, which is used to vertically add their products. The local adder has a tree structure determined by H and n. Specifically, the input product of the local adder has n bits, and the last stage output has log_2(H / m)+n bits.

[0048] The row driver provides wordline direction input signals for weight programming and parallel memory calculations. It may include a row address decoder, a signal buffer, and a multiplexer to select a row from m×n weight units. The column driver writes the bitline signal to the SRAM cell during the weight programming phase of DCIM.

[0049] The shift accumulator receives the output from the local adder in the memory calculation stage of the DCIM. Since the input is fed to the DCIM in a bit-serial manner with multiple cycles, the shift accumulator collects the partial sum value of one input bit in each cycle and then performs the shift and add function to obtain the MAC value for the full input precision.

[0050] A precision fusion unit is designed at the bottom of the DCIM macrocell to support variable weight precision defined in the user specification. Since the DCIM element only performs calculations on n-bit weights, the unit fuses the outputs from multiple shift accumulators to obtain the final MAC result of full weight precision. Taking 8b weight precision as an example, the precision fusion unit is configured to fuse the outputs of every 2 shift accumulators.

[0051] The controller (CTR.) configures the DCIM macrocell into weight programming or memory computing stages and enables the corresponding components in each stage.

[0052] Fig. 7A and 7B Shows the Figure 6 Example of custom placement and routing of DCIM component circuits in . Fig. 7A A column template with m weight units and one bit-wise multiplier is shown. Figure 7B It shows how the automatic compilation framework with m=4, n=4 generates a DCIM element layout.

[0053] The DCIM element is a basic building block with two sets of m×n weight cells and 1b×nb multipliers. The automatic compilation framework generates the DCIM element by following a template of SRAM-based weight cells and bit-wise multipliers from a custom cell database.

[0054] Fig. 7A A column template with m weight cells and one bitwise multiplier is shown. The weight bits (Q) are stored in 6T SRAM cells. The bitwise multiplication is implemented by 4T NOR gates. To support a variable weight multiplier ratio of m in the same template, an NMOS-based selection transistor is added between the Q_B node of the weight cell and one input port of the bitwise multiplier. The other input port is the inversion of the input signal (IN_B).

[0055] Figure 7BIt shows how the automatic compilation framework with m=4, n=4 generates a DCIM component layout. The automatic compilation framework first generates a set of m×n weight cells and a 1b×nb multiplier, and then places the two groups of cells face to face. Each group is generated by setting m=4 for the column template and repeating n=4 multiple times horizontally. In the DCIM component template, the weight cells in a row can share the same selection signal controlled by the row driver. In the memory calculation stage, a row of weights can be selected and multiplied with the input through a 1b×nb multiplier. To ensure that the automatic DCIM component generation procedure can be highly compact, the layout of the SRAM cells and NOR gates is designed to have the same width between the power lines. Therefore, they can be connected vertically and extended horizontally with almost no waste of area. The automatic compilation framework is highly scalable for new DCIM technologies in this emerging field. The DCIM component template can be easily extended by adding new cell designs to the database.

[0056] Fig. 8A and 8B Shows Figure 6 An example of custom placement and routing of a local adder tree circuit in . Fig. 8A shows a local adder template, while Figure 8B Shown by following Fig. 8A This is an example of a local adder generated by the automatic compilation framework using a template.

[0057] The local adder is the basic building block for adding the products of DCIM elements in a column. Fig. 8A , the automatic compilation framework generates local adders by following a template of a tree structure determined by H, m, n. The local adder template is derived from an optimized adder tree with interleaved 28T and 14T full adders (which is in TSMC's 22nm DCIM design), which improves energy efficiency by 30% over using only 28T adders. The 28T and 14T-bit full adders are created in the custom cell database, so once the parameters H, m, n are given, the automatic compilation framework can generate the architecture of any local adder. In order to ensure high compactness after stacking the 28T and 14T adders in one level of the adder tree, the two custom cells are designed to have the same width in layout.

[0058] Design rules are set for the layout of local adders to shape them into compact rectangles. The total height of the DCIM is the same as the height of the DCIM column in one subarray. The intuitive adder tree layout usually looks like a pyramid, with large areas for the early stages and small areas for the later stages, which will lead to large area waste if the layout is directly integrated between DCIM elements.

[0059] Figure 8BAn example of a local adder generated by the automatic compilation framework is shown. The adder tree structure has four 4b full adders in the first level, two 5b full adders in the second level, and one 6b full adder in the third level. The design rules of the automatic compilation framework restrict the placement of late adders, and the rules place 5b and 6b adders in the same column, matching the total height of the first-level 4b adders. In this way, the area in the DCIM macrocell can be fully utilized, and the layout of the adder tree can be compressed into a rectangle.

[0060] Fig. 9 An expert-based design exploration with three different types of array partitioning schemes according to one embodiment of the present invention is shown.

[0061] By partitioning the DCIM array into different sub-array granularities, a trade-off between area, latency, and energy efficiency can be achieved. Multiple sub-arrays can be used to implement an H×W DCIM macrocell. After the array is partitioned, the row / column drivers and shift accumulators are distributed near the sub-arrays, and the DCIM functionality of each sub-array remains the same as the basic architecture.

[0062] The architectural changes lead to new DCIM layouts with different area, latency, and energy efficiency. In smaller sub-array granularity, the critical path latency of memory computation is usually shorter. However, this generates more wiring and peripheral costs between sub-arrays, which may increase power consumption and reduce energy efficiency. Therefore, an automatic compilation framework is developed to compare different sub-array granularities through layout exploration cycles. For each iteration, the automatic compilation framework runs a macro cell generation process. The hardware metric measurements of each generated DCIM macro cell are evaluated. After the exploration is completed, the automatic compilation framework finds the optimal DCIM macro cell that meets the hardware constraints.

[0063] refer to Fig. 9 , for easier exploration and comparison, three types of partitioning schemes are shown here (W-oriented, H-oriented, H&W-oriented). W-oriented partitioning only splits in the horizontal direction, so the wordline critical path is reduced. H-oriented partitioning only splits in the vertical direction, which reduces the input product count and stage count of the local adder. H&W-oriented partitioning splits in both vertical and horizontal directions with more complex trade-offs.

[0064] Fig. 10A and Fig. 10B Two physical layout examples are shown, with and without array partitioning, respectively. Fig. 10A A 1024×256 DCIM macrocell without array partitioning is shown, and its macrocell area is 1.777×1.667=2.962 mm 2The critical path delay is 55.95ns, and the energy efficiency is 43.4TOPS / W at 1.1V. Fig. 10B A 1024×256 H-oriented partitioning scheme is shown that achieves the highest energy efficiency (74.7 TOPS / W) at the same voltage.

[0065] The efficiency improvement mainly comes from the shorter critical path (25.98ns) after H-directed partitioning, which simplifies the hierarchy of local adders. Overall, the various results measured by the above hardware indicators prove the necessity of the layout exploration loop of the automatic compilation framework. For some user specifications, the design space should be carefully explored, which is difficult for traditional manual DCIM design.

[0066] The functional units and modules according to the embodiments disclosed herein can be implemented using computing devices, computer processors or electronic circuits, including but not limited to application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), microcontrollers, and other programmable logic devices configured or programmed according to the teachings of the present disclosure. Based on the teachings of the present disclosure, computer instructions or software codes running in computing devices, computer processors or programmable logic devices can be easily prepared by technicians in the software or electronics fields.

[0067] All or part of the method according to the embodiment may be performed in one or more computing devices, including server computers, personal computers, laptop computers, mobile computing devices such as smart phones and tablet computers.

[0068] Embodiments may include computer storage media, transient and non-transitory memory devices having computer instructions or software codes stored therein, which may be used to program or configure a computing device, a computer processor, or an electronic circuit to perform any of the processes of the present invention. Storage media, transient and non-transitory storage devices may include, but are not limited to, floppy disks, optical disks, Blu-ray disks, DVDs, CD-ROMs and magneto-optical disks, ROMs, RAMs, flash memory devices, or any type of media or device suitable for storing instructions, codes, and / or data.

[0069] Each functional unit and module according to various embodiments may also be implemented in a distributed computing environment and / or a cloud computing environment, where all or part of the machine instructions may be executed in a distributed manner by one or more processing devices interconnected by a communication network, such as an intranet, a wide area network (WAN), a local area network (LAN), the Internet, and other forms of data transmission media.

[0070] The foregoing description of the present invention has been provided for the purpose of illustration and description. It is not intended to be exhaustive or to limit the invention to the precise form disclosed. Many modifications and variations will be apparent to those skilled in the art.

[0071] The embodiment was chosen and described in order to best explain the principles of the invention and its practical application, thereby enabling others skilled in the art to understand the invention for various embodiments and with various modifications as are suited to the particular use contemplated.

Claims

1. An automatic compilation method, characterized in that: Used to generate layouts for circuit designs for integrated in-memory computing, including: constructing a design database having at least one basic design library and a collection of reference architecture templates; receiving a user specification associated with a circuit design for the integrated in-memory computing; Selecting a reference architecture template from the design database and parameterizing the selected architecture template according to the received user specification to generate an architecture for the circuit design of the integrated in-memory computing; generating a physical layout based on the generated architecture according to one or more placement and routing algorithms; Iterating the design exploration cycle on the generated layout using one or more expert experience-based optimization solutions to obtain one or more explored physical layouts; and The hardware metrics of the one or more physical layouts explored are evaluated under one or more constraints to determine an optimal physical layout.

2. The automatic compilation method according to claim 1, characterized in that: The basic design library includes designs of memory cells, local logic elements and peripheral circuits.

3. The automatic compilation method according to claim 2, characterized in that: The memory unit is a volatile memory unit or a non-volatile memory unit.

4. The automatic compilation method according to claim 2, characterized in that: The local logic elements include multipliers and / or partial sum adders.

5. The automatic compilation method according to claim 2, characterized in that: The peripheral circuit includes a row / column driver, a shift accumulator, a precision fusion unit and / or a controller.

6. The automatic compilation method according to claim 1, characterized in that: The reference architecture template includes an architecture template extracted from an architecture design based on expert experience or user definition.

7. The automatic compilation method according to claim 1, characterized in that: The user specification includes a computer readable code and one or more constraints associated with a circuit design of the integrated in-memory computing.

8. The automatic compilation method according to claim 1, characterized in that: The expert experience-based optimization scheme includes a partitioning scheme of one or more memory cell arrays.

9. The automatic compilation method according to claim 8, characterized in that: The one or more memory cell array partitioning schemes have different partitioning sub-array granularities or different partitioning orientations.

10. The automatic compilation method according to claim 1, characterized in that: The one or more constraint conditions include an area priority condition or a critical path delay priority condition.

11. An automatic compilation framework, characterized in that: Used to generate layouts of circuit designs for integrated in-memory computing and includes: a design database constructed from at least a collection of basic design libraries and reference architecture templates; Automatic layout generator, including: an architecture generator configured to: receive a user specification associated with a circuit design for integrated in-memory computing; select a reference architecture template from the design database and parameterize the selected architecture template according to the received user specification to generate an architecture for the circuit design for integrated in-memory computing; and a placement and routing processor configured to: generate a physical layout based on the generated architecture according to one or more placement and routing algorithms; a design explorer configured to iterate a design exploration cycle on the generated layout using one or more expert experience-based optimization schemes to obtain one or more explored physical layouts; and The hardware performance evaluator is configured to measure hardware indicators of the one or more explored physical layouts under constraints to determine an optimal physical layout.

12. The automatic compilation framework according to claim 11, characterized in that: The basic design library includes designs of memory cells, local logic elements and peripheral circuits.

13. The automatic compilation framework according to claim 12, characterized in that: The memory unit is a volatile memory unit or a non-volatile memory unit.

14. The automatic compilation framework according to claim 12, characterized in that: The local logic elements include multipliers and / or partial sum adders.

15. The automatic compilation framework according to claim 12, characterized in that: The peripheral circuit includes a row / column driver, a shift accumulator, a precision fusion unit and / or a controller.

16. The automatic compilation framework according to claim 11, characterized in that: The reference architecture template includes an architecture template extracted from an architecture design based on expert experience or user definition.

17. The automatic compilation framework according to claim 11, characterized in that: The user specification includes a computer readable code and one or more constraints associated with a circuit design of the integrated in-memory computing.

18. The automatic compilation framework according to claim 11, characterized in that: The expert experience-based optimization scheme includes a partitioning scheme of one or more memory cell arrays.

19. The automatic compilation framework according to claim 18, characterized in that: The one or more memory cell array partitioning schemes have different partitioning sub-array granularities or different partitioning orientations.

20. The automatic compilation framework according to claim 11, characterized in that: The one or more constraint conditions include an area priority condition or a critical path delay priority condition.