An adaptive protocol reduction calculation method for multi-core vectorized computing platform

By optimizing the data volume combination of data cutting and iterative operations on a multi-core vectorized computing platform, the problem of low efficiency in reduction computation is solved, and efficient reduction algorithm optimization is achieved.

CN118689833BActive Publication Date: 2024-11-01北京麟卓信息科技有限公司
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202411175723.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-26
Publication Date
2024-11-01
Estimated Expiration
2044-08-26

AI Technical Summary

Technical Problem

Existing multi-core vectorized computing platforms are inefficient in reduction computation, especially when dealing with large-scale data, and fail to fully utilize the advantages of multi-core parallelism.

Method used

By segmenting the data of the multi-core vectorized computing platform, setting the data volume combination of each core iterative transmission and iterative vector operation, recording the total execution time under different combinations, and selecting the optimal parameters to optimize the reduction calculation process.

Benefits of technology

It improves the computational efficiency of the multi-core vectorization computing platform and realizes a high-performance reduction algorithm.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118689833B_ABST
    Figure CN118689833B_ABST
Patent Text Reader

Abstract

The present invention discloses an adaptive protocol calculation method for a multi-core vectorized computing platform. The method divides the data to be processed according to the number of cores of the multi-core vectorized computing platform, and then sets different combinations of the amount of data iteratively transmitted by each core and the amount of data for iterative vector operations. The method records the total execution time of a protocol optimization processing process determined for data with different total data amounts under different combinations, and selects the combination with the shortest total execution time as the optimal parameters for the protocol optimization processing determined on the multi-core vectorized computing platform. The method optimizes the parameters of the multi-core vectorized computing platform while determining the protocol optimization processing process, and effectively improves the computing efficiency of the platform.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of computer software development, and in particular relates to an adaptive protocol calculation method for a multi-core vectorized computing platform. Background Art

[0002] Vectorized computing is a method of improving computing efficiency by operating all elements of an array or vector at the same time. It can avoid traditional element-by-element processing to increase computing speed and is widely used in many fields, including but not limited to data processing, numerical computing, machine learning, and graphics processing. Multi-core vectorized computing platform is a commonly used design architecture for heterogeneous acceleration chips. Multi-core parallelization can further improve processing performance and facilitate efficient algorithm optimization. In addition, reduction computing is a process of merging multiple values ​​into one value, which is usually used for data reduction and simplified calculation. In the field of high-performance computing and parallel computing, reduction refers to computing operations that can be accelerated by parallelization, such as summation and maximum value. These operations are widely used in scientific computing, image processing, machine learning and other fields. Since the multi-core vectorized computing platform has high processing performance, it is particularly important to build a reduction computing method on the multi-core vectorized computing platform to further improve computing efficiency, especially when processing large-scale data. Summary of the invention

[0003] In view of this, the present invention provides an adaptive reduction calculation method for a multi-core vectorized computing platform, which realizes a high-performance reduction algorithm on a multi-core platform.

[0004] The present invention provides a multi-core vectorized computing platform adaptive protocol calculation method, which specifically includes the following steps:

[0005] Step 1, the data to be processed by the multi-core vectorized computing platform is the first data, and the total data volume is N; the data volume of the first-layer data segment that each core needs to process is lenPerT, the data volume that needs to be transmitted each time is TILE, the element number range in the first-layer data segment is [snum, snum+lenPerT), and snum is the data segment benchmark;

[0006] Step 2: determine the second-layer data segment that each core needs to process for iterative transmission, where the element number range in the second-layer data segment is [snum+i*TILE, snum+i*TILE+TILE), where i is the number of iterative transmissions; each core performs iterative transmission, and after each iterative transmission is executed, the initial value of the corresponding result vector VS_i is set to zero;

[0007] Step 3, each core performs iterative vector operations on the second-layer data segment, and the total number of iterative vector operations is inloop, the value of inloop is TILE / VLEN, VLEN is the amount of data for each iterative vector operation, and j is the number of iterative vector operations; the data processed each time is the third-layer data segment V_j, and the element number range in the third-layer data segment is [snum+i*TILE+j*VLEN, snum+i* TILE+j*VLEN+VLEN), and iterative vector operation reduction calculation is performed on VS_i and V_j, and VS_i is updated with the operation result until j equals inloop; let i increase by 1, if i is less than loop, execute step 2, otherwise perform reduction calculation on all VS_i to obtain the first calculation result VS of the current core; loop is the total number of iterative transmissions of each core, and the value is lenPerT / TILE;

[0008] Step 4. If lenPerT is not divisible by TILE, perform iterative vector operations on the remaining data segment drm with an element number range of [snum+lenPerT-drm, snum+lenPerT); if drm is not divisible by VLEN, load the data segment with an element number range of [snum+lenPerT-drrm, snum+lenPerT) into [0:drrm] of the vector VM, set the element at [drrm:VLEN] of VM to 0, and perform reduction calculations on VS and VM;

[0009] Step 5: The core stores VS in an array space VA of size NCORE*VLEN allocated in the off-chip global storage, performs a reduction calculation on VA with NCORE*VLEN as the unit length, and saves the result as VSL; performs a reduction calculation on the elements in the vector VSL to obtain the final reduction result; NCORE is the total number of cores;

[0010] Step 6, determine the combination of VLEN and TILE values, traverse the total data volume N to execute steps 1 to 5 to obtain the execution time corresponding to all the total data volumes, sum these execution times to obtain the total execution time under the current combination of VLEN and TILE values, traverse the value range of VLEN and TILE to obtain the total execution time of all combinations of VLEN and TILE values; select the combination of VLEN and TILE with the shortest total execution time as the optimal parameter of the multi-core vectorized computing platform.

[0011] Furthermore, the first layer data segment is determined in the following manner:

[0012] Let the element number of the element in the first data be dataId, and the value of dataId be 0, 1, 2, ..., (N-1); let the total number of cores be NCORE, the core number be taskId, and the value of taskId be 0, 1, 2, ..., (NCORE-1); let the value of the numbering base rmn be the remainder of N divided by NCORE;

[0013] If the core's taskId value is less than rmn, lenPerT is N / (NCORE+1) and snum is lenPerT*taskId; otherwise, lenPerT is N / NCORE and snum is lenPerT*taskId+rmn.

[0014] Furthermore, the value of TILE is 4*BW / DSIZE, BW is the data transmission bandwidth of the multi-core vectorized computing platform, and DSIZE is the number of bytes corresponding to the data type.

[0015] Furthermore, the value of VLEN is the bit width of the vector register.

[0016] Furthermore, TILE is set to an integer multiple of VLEN by padding with zeros.

[0017] Furthermore, the reduced calculation is a reduced summation.

[0018] Furthermore, when determining the value combination of VLEN and TILE in step 6, the value of VLEN is an integer multiple of 2, and the value range of the integer exponent is [4, 14]; the value of TILE is an integer multiple of VLEN; the value of N is an integer multiple of 2, and the value range of the integer exponent is [10, 26]. Beneficial Effects

[0019] The present invention divides the data to be processed according to the number of cores of a multi-core vectorized computing platform, and then sets different combinations of the amount of data iteratively transmitted by each core and the amount of data iteratively operated by vectors, and records the total execution time of the determined protocol optimization processing process for data with different total data amounts under different combinations, and selects the combination with the shortest total execution time as the optimal parameters for the protocol optimization processing determined on the multi-core vectorized computing platform. While determining the protocol optimization processing process, the multi-core vectorized computing platform parameters are optimized, thereby effectively improving the computing efficiency of the platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A schematic diagram of the calculation flow of an adaptive reduction calculation method for a multi-core vectorized computing platform provided by the present invention. DETAILED DESCRIPTION

[0021] The present invention is described in detail below with reference to the accompanying drawings and with reference to the embodiments.

[0022] The present invention provides an adaptive protocol calculation method for a multi-core vectorized computing platform, the core idea of ​​which is: dividing the data to be processed according to the number of cores of the multi-core vectorized computing platform, and then setting different combinations of the amount of data iteratively transmitted by each core and the amount of data iteratively operated by vectors, executing a determined protocol optimization processing process for data with different total data amounts under different combinations, recording the total execution time, and selecting the combination with the shortest total execution time as the optimal parameters for the protocol optimization processing determined on the multi-core vectorized computing platform.

[0023] The present invention provides a multi-core vectorized computing platform adaptive protocol calculation method, which specifically includes the following steps:

[0024] Step 1: Evenly deploy computing tasks according to the number of cores of the multi-core vectorized computing platform, so that the amount of data to be calculated on each core is basically the same. Let the data related to the reduced calculation, that is, the data to be processed by the multi-core vectorized computing platform, be the first data, the total amount of the first data be N, set the element number dataId for the elements in the first data, and the values ​​of dataId are 0, 1, 2, ..., (N-1) respectively; the total number of cores of the multi-core vectorized computing platform is NCORE, set the core with a continuous core number taskId, and the values ​​of taskId are 0, 1, 2, ..., (NCORE-1) respectively, where N can be the length of the array for the reduced calculation; determine the amount of data TILE required to be transmitted by each core each time.

[0025] Specifically include:

[0026] Step 1.1, let the value of the numbering base rmn be the remainder of N divided by NCORE, that is, rmn = N % NCORE;

[0027] Step 1.2, for cores whose taskId value is less than rmn, the value of lenPerT of the first-layer data segment that each core needs to process is N / (NCORE+1); for cores whose taskId value is not less than rmn, the value of lenPerT of the core data volume is N / NCORE, where N / NCORE means NCORE divides N, and lenPerT means the amount of data allocated to each core, that is, the number of elements in the array corresponding to the total data volume allocated to the core for processing;

[0028] Step 1.3, calculate the data segment in the first data that each core needs to process, and record the data segment as the first-level data segment. The data segment is represented by an element number. The value range of the element number in the first-level data segment is [snum, snum+lenPerT), where snum is the data segment reference; for the core whose taskId value is less than rmn, the value of snum is lenPerT*taskId, and for the core whose taskId value is not less than rmn, the value of snum is lenPerT * taskId+rmn.

[0029] Step 2: record the data segments that each core needs to process in each iterative transmission as the second-layer data segments. The size of the second-layer data segments is the same as TILE. The value range of the element number in the second-layer data segments is [snum+i*TILE, snum+i*TILE+TILE), where i is the number of iterative transmissions, and the value of i is 0, 1, 2, ..., (loop-1), that is, the i-th iterative transmission, loop is the total number of iterative transmissions of each core, and the value of loop is lenPerT / TILE.

[0030] The method for determining the amount of data TILE in the first layer data segment required for iterative transmission of each core is as follows: according to the data transmission bandwidth BW of the multi-core vectorized computing platform and the number of bytes DSIZE corresponding to the data type, the value of TILE is determined to be 4*BW / DSIZE.

[0031] Step 3: Each core performs an iterative transmission operation according to the determined data volume TILE and the total number of iterative transmissions loop. After each iterative transmission is executed, the initial value of the result vector VS_i with a length VLEN corresponding to the iterative transmission is set to zero and step 4 is executed.

[0032] During the calculation process of the multi-core vectorized computing platform, some of the chips selected need to transfer the data to be processed from the global storage outside the chip to the local storage on the chip before the vectorized computing core can access the data for the calculation, such as the Kunlun core AI accelerator card R200. Among them, the memory access bit width between the global storage outside the chip and the local storage on the chip, that is, the data transmission bandwidth, is represented by BW, in bytes; the number of bytes corresponding to the data type of the data to be processed is represented by DSIZE. Therefore, according to the data transmission bandwidth of the multi-core vectorized computing platform, the reasonable setting of the amount of data transmitted in a single iteration can effectively improve the data access efficiency.

[0033] Step 4: According to the amount of data TILE that each core needs to iterate and transmit and the amount of data VLEN that each core needs to calculate for each iterative vector operation, determine the total number of times inloop that each core performs iterative vector operations on the data amount TILE. The value of inloop is TILE / VLEN, and TILE is set to an integer multiple of VLEN. The data segment that needs to be processed for each iterative vector operation is recorded as the third-layer data segment V_j. The value range of the element number in the third-layer data segment is [snum+i*TILE+j*VLEN, snum+i*T ILE+j*VLEN+VLEN), where j is the number of iterative vector operations, and the value of j is 0, 1, 2, ..., (inloop-1); according to the data volume VLEN and the determined total number of iterations inloop, iteratively determine the third-layer data segment V_j, and perform iterative vector operation reduction calculation on VS_i and V_j, and update VS_i with the operation result; let i increase by 1, if i is less than loop, execute step 3, otherwise perform reduction calculation on all VS_i to obtain the first calculation result VS of the current core and then execute step 5.

[0034] The amount of data VLEN required to calculate a single iterative vector operation is related to the vector register and can be determined by obtaining the bit width and other properties of the vector register, which is generally 16 elements. When TILE cannot be divided by VLEN, TILE can be set to an integer multiple of VLEN by padding with zeros.

[0035] For the reduced sum, the iterative vector operation reduction calculation for VS_i and V_j is VS_i[0:VLEN] =VS_i[0:VLEN] + V_j [0:VLEN].

[0036] Step 5, when lenPerT cannot be divided by TILE, it means that there are unprocessed data in the first layer data segment, and the remaining number needs to be processed. The value of the remaining data segment drm is lenPerT%VLEN, and the value range of the element number in the corresponding data segment is [snum+lenPerT-drm, snum+lenPerT), and then the iterative vector operation is performed on the data segment;

[0037] When drm cannot be divided by VLEN, it means that there is unprocessed data in the second-level data segment, and the remaining number needs to be processed. The value of the remainder drrm is drm % VLEN, and the value range of the element number in the corresponding data segment is [snum+lenPerT-drrm, snum+lenPerT). The data in the data segment is loaded into the vector VM, and the position [0:drrm] in VM saves the data segment. Then the element at [drrm:VLEN] in VM is set to 0, and VS and VM perform reduction calculations. For example, the reduced sum calculation VS[0:VLEN] = VS[0:VLEN]+VM[0:VLEN].

[0038] Step 6. All cores store data in the global storage outside the chip, that is, all cores store VS in the global storage. Specifically, an array VA of size NCORE*VLEN is allocated in the global storage, and each core copies the reduction result VS to VA. For the core numbered taskId, its reduction result VS is copied to VA[taskId*VLEN:(taskId+1)*VLEN].

[0039] A core synchronization operation is further performed to ensure that the calculations of all cores are saved into the array VA.

[0040] Step 7: Use the method in step 4 to perform reduction calculation on the data with a length of NCORE*VLEN in the array VA, and save the result as VSL; perform reduction calculation on the elements in the vector VSL to obtain the final reduction result sum.

[0041] For example, for the reduced sum calculation, the process of performing the reduced calculation on the elements in the vector VSL is to accumulate the vector VSL to obtain the final reduced result sum, specifically sum=VSL[0]+VSL[1]+…+VSL[VLEN-1].

[0042] Step 8. Set the value ranges of the total data volume N, VLEN and TILE. Under any combination of VLEN and TILE, traverse the total data volume N to execute step 1 to obtain the execution time corresponding to all total data volumes. Sum these execution times to obtain the total execution time under the current combination of VLEN and TILE. Traverse the value range of VLEN and TILE to obtain the total execution time of all combinations of VLEN and TILE. Select the combination of VLEN and TILE with the shortest total execution time as the optimal parameters of TILE and VLEN under this platform. Example

[0043] In this implementation, an adaptive reduction calculation method for a multi-core vectorized computing platform provided by the present invention is used to optimize the reduction summation process of the multi-core vectorized computing platform. The calculation process is as follows Figure 1 as shown, specifically including:

[0044] S1. Use the task balance deployment module to evenly deploy the computing tasks to each core.

[0045] According to the number of cores in the multi-core vectorized computing platform, evenly deploy the computing tasks so that the amount of data to be calculated on each core is basically the same.

[0046] S1.1. Let the total amount of data to be reduced and calculated be N, and the total number of cores in the multi-core vectorized computing platform be NCORE;

[0047] S1.2. Continuously number the cores. The numbers of the NCORE cores are 0, 1, 2,..., (NCORE - 1) respectively, and use taskId to identify the core number;

[0048] S1.3. rmn = N % NCORE, where rmn is the remainder of dividing data N by NCORE;

[0049] S1.4. Use lenPerT to represent the amount of tasks (number of array elements) of each core. For cores with taskId < rmn, the amount of tasks lenPerT = N / NCORE + 1; for cores with taskId >= rmn, the amount of tasks lenPerT = N / NCORE; where N / NCORE means NCORE divides N evenly;

[0050] S1.5. Continuously number the data to be calculated. The numbers of the N data are 0, 1, 2,..., (N - 1) respectively, and use dataId to identify the data label;

[0051] S1.6. Calculate the data segments that each core needs to calculate. For cores with taskId < rmn, the data segment is [snum, snum + lenPerT), where snum = lenPerT * taskId; for cores with taskId >= rmn, the data segment is also [snum, snum + lenPerT), where snum = lenPerT * taskId + rmn;

[0052] S2. Use the multi-core parallel vectorized reduction module to implement each core to iterate on the required data segments for calculation. The amount of data for a single iteration TILE = 4 * BW / DSIZE, the total number of iteration transmissions for each core loop = lenPerT / TILE, and the data segment for the i-th iteration calculation is [snum + i * TILE, snum + i * TILE + TILE).

[0053] S3. Each core performs vectorized iterative calculations on a single iteration data segment.

[0054] The total number of iterations that need vectorized calculation for TILE is represented by inloop, inloop=TILE / VLEN, and TILE needs to be set to an integer multiple of VLEN. The data segment of the j-th vectorized calculation is [snum+i*TILE+j*VLEN, snum+i*TILE+j*VLEN+VLEN), and the sum is reduced.

[0055] S4. Each core performs vectorized iterative calculations on the remaining data after iteration.

[0056] S4.1. If lenPerT cannot be divided by TILE, the remaining data segments need to be processed. The remaining data segments are represented by drm, then drm=lenPerT%VLEN, and the corresponding data segment is [snum+lenPerT-drm, snum+lenPerT). Use the method of S3 to perform vectorized iterative calculation on the data segment.

[0057] S4.2. If drm cannot be divided by VLEN, the remaining number needs to be processed. Use drrm to represent the remainder, then drrm=drm%VLEN, and the corresponding data segment is [snum+lenPerT-drrm, snum+lenPerT). Load the data into the vector VM, and then perform the reduction and summation, VS_i[0:VLEN] = VS_i[0:VLEN] +VM[0:VLEN].

[0058] S5. All cores store data to off-chip global storage.

[0059] S5.1. Allocate an array space VA of size NCORE*VLEN in the global storage, and each core copies the reduction result VS to VA.

[0060] S5.2. Use the multi-core synchronization module to perform core synchronization operations.

[0061] S6. Use the single-core reduction module to complete the single-core reduction processing of the vectorized reduction results of each core.

[0062] S6.1. If taskId is 0, the data in array VA with a length of NCORE*VLEN is reduced using the method in step 3. VS is also used to save the reduction result.

[0063] S6.2. Accumulate the vector VS to obtain the final reduction result sum, sum = vs[0] + vs[1] + … + vs[VLEN-1].

[0064] S7. The platform adaptive tuning parameters TILE and VLEN improve the operator platform compatibility.

[0065] Data access and data parallelism are the main factors affecting algorithm performance. The TILE parameter determines the amount of data accessed in each outer loop, and the VLEN parameter determines the amount of data calculated in each inner loop. The optimal TILE and VLEN combination is different for different computing platforms.

[0066] S7.1. Obtain a protocol program that can run on the current platform according to S1-S6;

[0067] S7.2, VLEN can be an integer exponent multiple of 2, the integer exponent ranges from 4 to 14, VLEN ∈[2^4, 2^5, …, 2^14];

[0068] S7.3, TILE can be an integer multiple of VLEN, TILE ∈ [1, 2, 3, 4, 5, 20]*VLEN;

[0069] S7.4, the total amount of data N calculated by the protocol is an integer exponential multiple of 2, the integer exponent ranges from 10 to 26, N∈[2^10, 2^11,…, 2^26];

[0070] S7.5. For each combination of VLEN and TILE, run the protocol program separately, test the running time under different data amounts, and obtain the total time for all data amount conditions;

[0071] S7.6. Select the VLEN and TILE combination with the shortest total time as the TILE and VLEN parameters for this platform.

[0072] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-core vectorized computing platform adaptive protocol calculation method for heterogeneous chips, characterized in that: The specific steps include: Step 1, the data to be processed by the multi-core vectorized computing platform is the first data, and the total data volume is N; the data volume of the first-layer data segment that each core needs to process is lenPerT, the data volume that needs to be transmitted each time is TILE, the element number range in the first-layer data segment is [snum, snum+lenPerT), and snum is the data segment benchmark; Step 2: determine the second-layer data segment that each core needs to process for iterative transmission, where the element number range in the second-layer data segment is [snum+i*TILE, snum+i*TILE+TILE), where i is the number of iterative transmissions; each core performs iterative transmission, and after each iterative transmission is executed, the initial value of the corresponding result vector VS_i is set to zero; Step 3, each core performs iterative vector operations on the second-layer data segment, and the total number of iterative vector operations is inloop, the value of inloop is TILE / VLEN, VLEN is the amount of data for each iterative vector operation; the data processed each time is the third-layer data segment V_j, the element number range in the third-layer data segment is [snum+i*TILE+j*VLEN, snum+i*TILE+j*VLEN+VLEN), and iterative vector operation reduction calculation is performed on VS_i and V_j, and VS_i is updated with the operation result until j is equal to inloop, and j is the number of iterative vector operations; Let i be incremented by 1. If i is less than loop, execute step 2. Otherwise, perform reduction calculation on all VS_i to obtain the first calculation result VS of the current core. loop is the total number of iterative transmissions of each core, and its value is lenPerT / TILE. Step 4, if lenPerT is not divisible by TILE, then perform iterative vector operations on the remaining data segment drm with an element number range of [snum+lenPerT-drm,snum+lenPerT); if drm is not divisible by VLEN, then load the data segment with an element number range of [snum+lenPerT-drrm,snum+lenPerT) into [0:drrm] of the vector VM, set the element at [drrm:VLEN] of VM to 0, and perform reduction calculations on VS and VM; Step 5: The core stores VS in an array space VA of size NCORE*VLEN allocated in the off-chip global storage, performs a reduction calculation on VA with NCORE*VLEN as the unit length, and saves the result as VSL; performs a reduction calculation on the elements in the vector VSL to obtain the final reduction result; NCORE is the total number of cores; Step 6, determine the combination of VLEN and TILE values, traverse the total data volume N to execute steps 1 to 5 to obtain the execution time corresponding to all the total data volumes, sum these execution times to obtain the total execution time under the current combination of VLEN and TILE values, traverse the value range of VLEN and TILE to obtain the total execution time of all combinations of VLEN and TILE values; select the combination of VLEN and TILE with the shortest total execution time as the optimal parameter of the multi-core vectorized computing platform.

2. The method for adaptive reduction calculation of a multi-core vectorized computing platform according to claim 1, characterized in that: The first layer data segment is determined as follows: Let the element number of the element in the first data be dataId, and the value of dataId be 0, 1, 2, ..., (N-1); let the total number of cores be NCORE, the core number be taskId, and the value of taskId be 0, 1, 2, ..., (NCORE-1); let the value of the numbering base rmn be the remainder of N divided by NCORE; If the core's taskId value is less than rmn, lenPerT is N / (NCORE+1) and snum is lenPerT*taskId; otherwise, lenPerT is N / NCORE and snum is lenPerT*taskId+rmn.

3. The method for adaptive reduction calculation of a multi-core vectorized computing platform according to claim 1, characterized in that: The value of TILE is 4*BW / DSIZE, where BW is the data transmission bandwidth of the multi-core vectorized computing platform, and DSIZE is the number of bytes corresponding to the data type.

4. The method for adaptive reduction calculation of a multi-core vectorized computing platform according to claim 1, characterized in that: The value of VLEN is the bit width of the vector register.

5. The method for adaptive reduction calculation of a multi-core vectorized computing platform according to claim 1, characterized in that: Use zero padding to set TILE to an integer multiple of VLEN.

6. The method for adaptive reduction calculation of a multi-core vectorized computing platform according to claim 1, characterized in that: The reduced calculation is a reduced sum.

7. The method for adaptive reduction calculation of a multi-core vectorized computing platform according to claim 1, characterized in that: When determining the value combination of VLEN and TILE in step 6, the value of VLEN is an integer multiple of 2, and the value range of the integer exponent is [4, 14]; the value of TILE is an integer multiple of VLEN; the value of N is an integer multiple of 2, and the value range of the integer exponent is [10, 26].

Citation Information

Patent Citations

  • Multi-dimensional vector specification method, electronic equipment and storage medium

    CN114677549A

  • Data reduction method and device based on vector processor, equipment and storage medium

    CN116627896A

  • Time delay constraint scheduling method for heterogeneous computing power resources in computing power network

    CN116916393A

  • Task processing method and device, electronic equipment and storage medium

    CN116934571A

  • Multicore Processor And Method Of Use That Configures Core Functions Based On Executing Instructions

    US20100153700A1