CN109522127A - A kind of fluid machinery simulated program isomery accelerated method based on GPU - Google Patents
A kind of fluid machinery simulated program isomery accelerated method based on GPU Download PDFInfo
- Publication number
- CN109522127A CN109522127A CN201811378843.XA CN201811378843A CN109522127A CN 109522127 A CN109522127 A CN 109522127A CN 201811378843 A CN201811378843 A CN 201811378843A CN 109522127 A CN109522127 A CN 109522127A
- Authority
- CN
- China
- Prior art keywords
- gpu
- program
- memory
- thread
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
- G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
- G06F9/5011—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resources being hardware resources other than CPUs, Servers and Terminals
- G06F9/5016—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resources being hardware resources other than CPUs, Servers and Terminals the resource being the memory
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
- G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
- G06F9/5027—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2209/00—Indexing scheme relating to G06F9/00
- G06F2209/50—Indexing scheme relating to G06F9/50
- G06F2209/5018—Thread allocation
Landscapes
- Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Devices For Executing Special Programs (AREA)
Abstract
The invention discloses a kind of fluid machinery simulated program isomery accelerated method based on GPU, step includes analysis of central issue, and finding has the subprogram for accelerating potentiality；The deduction and exemption of data transmission between host equipment；The merging of internal storage access and the utilization of multiple types memory promote valid memory utilization rate and calculate memory access ratio；Code refactoring exposes data parallelism, and the explicit global synchronization of GPU kernel damages data parallelism, and Ying Jinli is avoided；The GPU of serial algorithm is adapted to, by the way that serial algorithm is replaced with parallel algorithm with the same function；The adjustment of thread allocation of parameters, so that thread calculation delay is sufficiently hidden, is promoted to calculate and be handled up by adjusting thread distribution；If reaching ideal effect by above step, then accelerate to complete, otherwise a wheel iteration new since analysis of central issue is until reaching promising result.The present invention provides a kind of GPU accelerated method for fluid machinery simulated program characteristic, modified program can reach ideal acceleration effect.
Description
Technical field
The invention belongs to hydrodynamics and high-performance calculation crossing domain, in particular to a kind of fluid machinery based on GPU
Simulated program isomery accelerated method.
Background technique
Fluid Mechanics Computation is one of important technology of field of fluid mechanics, right in a computer by using numerical method
Hydromechanical governing equation is solved, thus the flowing in predictable flow field.It, can be with the promotion of computer computation ability
Construct increasingly finer fluid mechanic model.Meanwhile for the flow field change of more accurate Fluid Mechanics Computation, transport flow field
It moves more " true ", fluid mechanic model is also increasingly fine, and to computing capability, higher requirements are also raised for this.
As monokaryon chip transistor density and frequency reach bottleneck, multicore becomes the main side for nowadays promoting computing capability
Formula.The companies such as NVIDIA, AMD, Intel are reached wherein tall and handsome it is also proposed that a series of for handling the specialized hardwares of intensive calculations
GPU suffers from outstanding performance in terms of performance and energy consumption.Traditional fluid machinery program optimizes towards CPU architecture, and with
The difference of the rise of Heterogeneous Computing, coprocessor framework and CPU architecture is increasing, the volume of traditional fluid machinery simulated program
WriteMode all cannot be directly adapted to new hardware structure again with prioritization scheme.Whether fluid is write for new hardware platform
Machine emulated program, or the existing fluid machinery simulated program towards CPU is transplanted for GPU, it requires new
Guidance program.
In fluid machinery simulated program explicit solution device, mostly based on finite volume method to fluid carry out from
It dissipates, accounting equation continuous iteration on grid, until meeting the condition of convergence, in each iteration step, to the physics in grid
Attribute is updated, and such application belongs to compute-intensive applications.Writing new CPU/GPU isomery program or to existing
When serial program carries out GPU transplanting, the hardware structure characteristic new notice how matching is required.
Summary of the invention
The purpose of the present invention is to provide a kind of fluid machinery simulated program isomery accelerated method based on GPU, to solve
The above problem.
To achieve the above object, the invention adopts the following technical scheme:
A kind of fluid machinery simulated program isomery accelerated method based on GPU, comprising the following steps:
Step 1, the mode combined by using static analysis and dynamic analysis carries out hot spot to the machine emulated program of fluid
Analysis；And preliminary parallelization realization is carried out to the hot spot chosen；
Step 2, for parallelisation procedure preliminary in step 1, data between host equipment are transmitted with intensive part, it will
Intermediate result is transplanted to the end GPU, and data are transmitted between reducing host equipment；
Step 3, for the program in step 2, by distributing the memory mapping of grid data, so that in the same thread beam
Thread read and update grid data adjacent in memory, and according to program characteristic using shared drive, constant memory and
Texture memory gives full play to concurrency；
Step 4, pass through code refactoring exposure data parallelism；
Step 5, the GPU adaptation of serial algorithm；For the program in step 4, if there is respective serial Riming time of algorithm
Accounting is big and cannot run at the end GPU, substitutes the elimination with iterative method when solving equation；
Step 6, the adjustment of thread allocation of parameters determines to maximize and hides memory access time delay, promotes the thread for calculating and handling up
Distribution；
Step 7, the acceleration effect that the program that step 1 arrives step 6 is completed in test then accelerates if speed-up ratio meets expection
It completes；Otherwise a wheel Accelerated iteration new since step 1；Until reaching satisfied acceleration effect.
Further, it in step 1, by the collection and analysis to hot spot partial operating time accounting, finds out in program and runs
The big part of time accounting carries out parallel optimization, the theoretical upper limit accelerated for the part.
Further, in step 2, will be accelerated by the way that entire multi grid is placed on the end GPU, by intermediate result whole
It is placed on the end GPU.
Further, in step 3, by adjusting layout of the data in global memory so that adjacent hardware thread is handled
Continuous data in memory, promote the effective bandwidth of internal storage access；And by shared drive, texture memory, constant memory
It utilizes, reduces thread to the access pressure of global memory's bandwidth, promote speed-up ratio.
Further, in step 4, after parallel version, the part of data contention can be generated, by calculating side for transplanting
Formula, the adjustment for calculating order avoid data contention；For the part that can not be avoided through the above way, it should current by terminating
GPU kernel, the mode for starting new GPU kernel introduce global synchronization.
Further, in step 5, for serial algorithm that can not be parallel in program, with identical function and with number
It is replaced according to the algorithm of concurrency, parallel algorithm is realized on GPU.
Further, in step 6, after the above step is finished, by experiment repeatedly, find out suitable thread grid,
Thread block allocation of parameters determines the best match of the program and hardware.
Compared with prior art, the present invention has following technical effect:
The present invention is by reasonable memory mapping, so that the effective bandwidth of GPU global access is got higher, gives full play to GPU meter
Calculation advantage；By replacement can not parallel serial algorithm make calling program can acceleration region expand, speed-up ratio promoted；By to complete
Office is synchronous to be avoided as possible, so that damage of the global synchronization to program parallelization is preferably minimized.
The present invention compares original CPU program, and GPU version has significant speed-up ratio；
Detailed description of the invention
Fig. 1 is the flow chart of method provided by the invention.
Specific embodiment
Below in conjunction with attached drawing, the present invention is further described:
Referring to Fig. 1, a kind of fluid machinery simulated program isomery accelerated method based on GPU, comprising the following steps:
Step 1, the mode combined by using static analysis and dynamic analysis carries out hot spot to the machine emulated program of fluid
Analysis；And preliminary parallelization realization is carried out to the hot spot chosen；
Step 2, for parallelisation procedure preliminary in step 1, data between host equipment are transmitted with intensive part, it will
Intermediate result is transplanted to the end GPU, and data are transmitted between reducing host equipment；
Step 3, for the program in step 2, by distributing the memory mapping of grid data, so that in the same thread beam
Thread read and update grid data adjacent in memory, and according to program characteristic using shared drive, constant memory and
Texture memory gives full play to concurrency；
Step 4, pass through code refactoring exposure data parallelism；
Step 5, the GPU adaptation of serial algorithm；For the program in step 4, if there is respective serial Riming time of algorithm
Accounting is big and cannot run at the end GPU, substitutes the elimination with iterative method when solving equation；
Step 6, the adjustment of thread allocation of parameters determines to maximize and hides memory access time delay, promotes the thread for calculating and handling up
Distribution；
Step 7, the acceleration effect that the program that step 1 arrives step 6 is completed in test then accelerates if speed-up ratio meets expection
It completes；Otherwise a wheel Accelerated iteration new since step 1；Until reaching satisfied acceleration effect.
In step 1, by the collection and analysis to hot spot partial operating time accounting, it is big to find out runing time accounting in program
Part, for the part carry out parallel optimization, the theoretical upper limit accelerated.
In step 2, it will be accelerated by the way that entire multi grid is placed on the end GPU, intermediate result is all placed on GPU
End.
In step 3, by adjusting layout of the data in global memory so that adjacent hardware thread handles in memory and connects
Continuous data promote the effective bandwidth of internal storage access；And by shared drive, texture memory, the utilization of constant memory, reduction
Thread promotes speed-up ratio to the access pressure of global memory's bandwidth.
In step 4, after parallel version, the part of data contention can be generated for transplanting, pass through calculation, calculating time
The adjustment of sequence avoids data contention；For the part that can not be avoided through the above way, it should by terminating current GPU kernel,
The mode for starting new GPU kernel introduces global synchronization.
In step 5, for serial algorithm that can not be parallel in program, with identical function and with data parallelism
Algorithm replacement, by parallel algorithm realize on GPU.
In step 6, after the above step is finished, by experiment repeatedly, suitable thread grid, thread block distribution are found out
Parameter determines the best match of the program and hardware.
Analysis of central issue.Amdahl's law shows to carry out accelerating to can be obtained speed-up ratio to program in parallel computation
The upper limit is limited to the percentage that accelerating part accounts for former serial program.According to the law, the first step of parallel optimization should be found
Hot spot only finds out the part that runing time accounting is big in program, carries out parallel optimization for the part, could obtain ideal
Acceleration effect.If having selected non-hot is to accelerate object, the effort of parallelization necessarily produces little effect.Analysis hot-spots generally have
Static and dynamic two ways.Static mode is to find out the hot spot of program by analyzing program code instructions.Dynamic part is logical
It crosses using the tool as gprof etc, collects the information such as each function call the time it takes in the program operation phase, according to
These information find out the hot spot of program.Fluid machinery simulated program due to model characteristic, from static analysis as can be seen that hot spot
Part generally is solved in multi-grid iteration, GPU acceleration generally is carried out to the part.
Reduce the data transmission between host equipment.The computing capability powerful compared to GPU, the data between CPU transmit effect
Rate is extremely low, can greatly drag down accelerating ability, it should avoid as far as possible.In multi grid solution procedure, if GPU calculation part
Divide and fail that iterative cycles part is completely covered, then in every single-step iteration of circulation, it is necessary to which transmission calculates knot between GPU
Fruit, and the data transmission between CPU is extremely time-consuming and inefficient compared to the calculated performance of GPU.If in every single-step iteration mistake
It requires to carry out data transmission in journey, then therefore GPU accelerating ability can have a greatly reduced quality, therefore by putting entire multi grid
Accelerated at the end GPU, intermediate result is all placed on to the end GPU, is followed to convert the data transmission in every single-step iteration to
The transmission of data twice of ring iterative before and after, so that the time of data transmission consumption becomes a constant, without
It is proportional to the number of iterations.And iterating to calculate front and back data and transmitting the occupied time can neglect compared with the time of iterative calculation
Slightly disregard, therefore the influence to accelerating ability can also be ignored.For not having the subprogram of concurrency, such as difference individually extremely
Process cannot be placed on the end GPU and be calculated, can be only placed at the end CPU, at this time the data transmission in iterative cycles between GPU and CPU
Unavoidably.Although only needing to transmit the calculated result of these serial programs, data volume is not very big.But due to this part number
To cycle-index be according to the number of transmissions it is proportional, also have certain influence to the performance of program, it is also desirable to take certain optimization
Measure.Fixed memory can be taken to optimize the transmission of this partial data.Number under default situations, between cpu process and GPU
It needs first to make primary copy in operating system nucleus space buffer area according to transmission, is then copied to destination from interior and buffer area.
It, can be to avoid operating system nucleus buffer area after setting fixed memory for the page where the array for needing to transmit
Copy, is transmitted directly between cpu process and GPU.The data transmission of reduction is after successive ignition circulation, to the property of program
It can be promoted more obvious.
Internal storage access merges and a variety of memories utilize.Fluid machinery simulated program is related to Multidimensional numerical operation, by array
In each Elemental partition to thread when, it should do between the minimum peacekeeping thread most low-dimensional of array and map one by one, so that continuously
Thread is updated adjacent data.As unit of warp, a warp includes 32 hardware threads for the execution of GPU thread.It is interior
As unit of a cache line, a cache line is generally continuous and alignment 32 bytes or 128 for the reading deposited
Byte.If the memory of the thread accesses in a warp is continuous, GPU bandwidth can be fully utilized.If the line in a warp
The memory of journey access is discontinuous, and it will cause bandwidth wastes, and effective bandwidth is caused to reduce.For example, if 32 bytes
Cache line data transmission in, only 4 byte datas are utilized, then handling capacity has just been reduced 8 times, GPU is powerful
Computing capability is difficult to be played, and can significantly reduce calculating speed.Internal storage access is incorporated as promoting memory effective bandwidth
Basic means, when optimization should pay attention to first.When carrying out GPU acceleration to three-dimensional grid, it should make continuous line as far as possible
Journey accesses data continuously distributed in memory.If taking other mapping schemes, the thread in a warp can be made
Reading data becomes the reading that strides from continuous data reading, so that memory effective bandwidth is lower, such case should be avoided.Simultaneously
In the conceived case also should using the type memories such as share memory, the memory access time delay and size of share memory and
Cache is suitable, but the data stored can greatly promote acceleration when data access has AD HOC by software management
Performance.
Pass through code refactoring exposure data parallelism.The structure of multi-grid iteration solver is that a number of iterations is huge
Big circulation, in the circulating cycle flowing of the analog physical amount between three dimension adjacent mesh.When some physical quantity is in adjacent net
When compartment flows, the numerical value of two adjacent mesh can be changed simultaneously.If same thread is responsible for a certain array element and three dimensions
The physical quantity inflow and outflow of adjacent thread element on degree, then each thread require to change the data element that the thread possesses with
And three data adjacent with the data in three dimensions, while other threads can also modify this four elements.Serial situation
Under this mode it is not problematic, if but untreated in parallel computation read/write conflict can be caused.In order to guarantee that program is correct
Property, it needs to restart a new kernel by terminating current kernel to introduce global synchronization.If every dimension more
A global synchronization is all introduced when new, then script a subprogram will because of extra global synchronization introducing and be split
At smaller 6 GPU kernel function.Concurrency can be greatly lost in the introducing of global synchronization, and brings additional GPU
Kernel Start-up costs.So that acceleration effect substantially reduces.This defect can avoid counting by rearranging calculating order
According to competition, stream that per thread can be allowed to be responsible for updating the grid that itself is possessed to three dimension forward direction adjacent mesh physical quantitys
Inflow with three dimension negative sense adjacent mesh to itself grid physical amount out.Arrangement in this way makes per thread more
The data of new owned grid, and then data contention is avoided, so that calling program is obtained ideal accelerating ability.
The GPU of serial algorithm is adapted to.For specific algorithm, such as commonly used three Diagonal Equation in equation solution
Solution procedure, data parallelism are obvious unlike other its algorithms.Due to the serial property of elimination solve system of equation, tradition
CPU solution procedure can not be grafted directly to GPU, be transplanted to before GPU and need to do corresponding algorithm adaptation.Classical GPU is asked
It solves three Diagonal Equations and realizes that algorithm has PCR (Parallel Cyclic Reduction) etc., current tall and handsome high ranking official side cusparse
Library provides the solver of one group of three Diagonal Equation based on above-mentioned algorithm, can use the library and solves progress to three Diagonal Equations
GPU parallelization.This kind of algorithm can also be optimized in conjunction with program self character simultaneously.Such as in explicit solution device, this
The parameter matrix of three Diagonal Equation of class is in an iterative process and unchanged, and the number of iterations is huge.It, can be with according to above-mentioned characteristic
It inverts before iteration starts to parameter matrix, the solution of three Diagonal Equations is converted into the multiplication of parameter inverse matrix and vector.
Matrix multiple has good data parallelism, can give full play to the performance of GPU.It is above-mentioned list a kind of pair of serial algorithm into
The thinking of row parallel adaptation.It for the serial algorithm run at the end CPU, can use with equivalent functions, and have good
The parallel algorithm of data parallelism is replaced.Such as the member that disappears of equation group solves, and can be replaced with iterative method.Pass through tune
With the good algorithm of data parallelism, the architected features of GPU are preferably matched, promote accelerating ability.
The adjustment of thread allocation of parameters determines to maximize and hides memory access time delay, promotes the thread configuration for calculating and handling up, lead to
It crosses and attempts different thread blocks, thread lattice specification.Suitable specific program is tested out, specific scale can maximize performance hardware and add
The kernel parameter configuration of fast performance.
Claims (7)
1. a kind of fluid machinery simulated program isomery accelerated method based on GPU, which comprises the following steps:
Step 1, the mode combined by using static analysis and dynamic analysis carries out hot spot point to the machine emulated program of fluid
Analysis；And preliminary parallelization realization is carried out to the hot spot chosen；
Step 2, for parallelisation procedure preliminary in step 1, data between host equipment are transmitted with intensive part, it will be intermediate
As a result it is transplanted to the end GPU, data are transmitted between reducing host equipment；
Step 3, for the program in step 2, by distributing the memory mapping of grid data, so that the line in the same thread beam
Journey reads and updates grid data adjacent in memory, and utilizes shared drive, constant memory and texture according to program characteristic
Memory gives full play to concurrency；
Step 4, pass through code refactoring exposure data parallelism；
Step 5, the GPU adaptation of serial algorithm；For the program in step 4, if there is respective serial Riming time of algorithm accounting
It cannot run greatly and at the end GPU, substitute the elimination with iterative method when solving equation；
Step 6, the adjustment of thread allocation of parameters determines to maximize and hides memory access time delay, promotes the thread distribution for calculating and handling up；
Step 7, the acceleration effect that the program that step 1 arrives step 6 is completed in test then accelerates to complete if speed-up ratio meets expection；
Otherwise a wheel Accelerated iteration new since step 1；Until reaching satisfied acceleration effect.
2. a kind of fluid machinery simulated program isomery accelerated method based on GPU according to claim 1, feature exist
In by the collection and analysis to hot spot partial operating time accounting, finding out the portion that runing time accounting is big in program in step 1
Point, parallel optimization, the theoretical upper limit accelerated are carried out for the part.
3. a kind of fluid machinery simulated program isomery accelerated method based on GPU according to claim 1, feature exist
In will be accelerated by the way that entire multi grid is placed on the end GPU, intermediate result is all placed on to the end GPU in step 2.
4. a kind of fluid machinery simulated program isomery accelerated method based on GPU according to claim 1, feature exist
In in step 3, by adjusting layout of the data in global memory so that adjacent hardware thread is handled in memory and continuously counted
According to promoting the effective bandwidth of internal storage access；And by shared drive, texture memory, the utilization of constant memory, reduction thread pair
The access pressure of global memory's bandwidth promotes speed-up ratio.
5. a kind of fluid machinery simulated program isomery accelerated method based on GPU according to claim 1, feature exist
In after being parallel version for transplanting, generating the part of data contention, by calculation, calculate order in step 4
Adjustment avoids data contention；For the part that can not be avoided through the above way, it should by terminating current GPU kernel, starting
The mode of new GPU kernel introduces global synchronization.
6. a kind of fluid machinery simulated program isomery accelerated method based on GPU according to claim 1, feature exist
In in step 5, for serial algorithm that can not be parallel in program, with identical function and with the calculation of data parallelism
Method replacement, parallel algorithm is realized on GPU.
7. a kind of fluid machinery simulated program isomery accelerated method based on GPU according to claim 1, feature exist
In, in step 6, after the above step is finished, by experiment repeatedly, suitable thread grid, thread block allocation of parameters are found out,
Determine the best match of the program and hardware.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811378843.XA CN109522127B (en) | 2018-11-19 | 2018-11-19 | Fluid machinery simulation program heterogeneous acceleration method based on GPU |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811378843.XA CN109522127B (en) | 2018-11-19 | 2018-11-19 | Fluid machinery simulation program heterogeneous acceleration method based on GPU |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109522127A true CN109522127A (en) | 2019-03-26 |
| CN109522127B CN109522127B (en) | 2021-01-19 |
Family
ID=65776286
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811378843.XA Active CN109522127B (en) | 2018-11-19 | 2018-11-19 | Fluid machinery simulation program heterogeneous acceleration method based on GPU |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109522127B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112148361A (en) * | 2020-08-27 | 2020-12-29 | 中国海洋大学 | Method and system for transplanting encryption algorithm of processor |
| CN112612476A (en) * | 2020-12-28 | 2021-04-06 | 吉林大学 | SLAM control method, equipment and storage medium based on GPU |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103729180A (en) * | 2013-12-25 | 2014-04-16 | 浪潮电子信息产业股份有限公司 | Method for quickly developing CUDA (compute unified device architecture) parallel programs |
| CN104035781A (en) * | 2014-06-27 | 2014-09-10 | 北京航空航天大学 | Method for quickly developing heterogeneous parallel program |
| CN104408019A (en) * | 2014-10-29 | 2015-03-11 | 浪潮电子信息产业股份有限公司 | Method for realizing GMRES (generalized minimum residual) algorithm parallel acceleration on basis of MIC (many integrated cores) platform |
| CN106250349A (en) * | 2016-08-08 | 2016-12-21 | 浪潮(北京)电子信息产业有限公司 | A kind of high energy efficiency heterogeneous computing system |
| CN108388537A (en) * | 2018-03-06 | 2018-08-10 | 上海熠知电子科技有限公司 | A kind of convolutional neural networks accelerator and method |
-
2018
- 2018-11-19 CN CN201811378843.XA patent/CN109522127B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103729180A (en) * | 2013-12-25 | 2014-04-16 | 浪潮电子信息产业股份有限公司 | Method for quickly developing CUDA (compute unified device architecture) parallel programs |
| CN104035781A (en) * | 2014-06-27 | 2014-09-10 | 北京航空航天大学 | Method for quickly developing heterogeneous parallel program |
| CN104408019A (en) * | 2014-10-29 | 2015-03-11 | 浪潮电子信息产业股份有限公司 | Method for realizing GMRES (generalized minimum residual) algorithm parallel acceleration on basis of MIC (many integrated cores) platform |
| CN106250349A (en) * | 2016-08-08 | 2016-12-21 | 浪潮(北京)电子信息产业有限公司 | A kind of high energy efficiency heterogeneous computing system |
| CN108388537A (en) * | 2018-03-06 | 2018-08-10 | 上海熠知电子科技有限公司 | A kind of convolutional neural networks accelerator and method |
Non-Patent Citations (3)
| Title |
|---|
| YINGRUI WANG: "GPU Acceleration Smoothed Particle Hydrodynamics for the Navier-Stokes Equations", 《2016 24TH EUROMICRO INTERNATIONAL CONFERENCE ON PARALLEL,DISTRIBUTED AND NETWORK-BASED PROCESSING》 * |
| 李铭: "基于GPU的张量分解及重构方法研究及应用", 《中国优秀硕士学位论文全文数据库 基础科学辑》 * |
| 魏洪昌: "面向 CPU-GPU源到源编译***的渐近拟合优化方法", 《计算机工程与应用》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112148361A (en) * | 2020-08-27 | 2020-12-29 | 中国海洋大学 | Method and system for transplanting encryption algorithm of processor |
| CN112148361B (en) * | 2020-08-27 | 2022-03-04 | 中国海洋大学 | Method and system for transplanting encryption algorithm of processor |
| CN112612476A (en) * | 2020-12-28 | 2021-04-06 | 吉林大学 | SLAM control method, equipment and storage medium based on GPU |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109522127B (en) | 2021-01-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Besta et al. | Slimsell: A vectorizable graph representation for breadth-first search | |
| Li et al. | GPU-accelerated preconditioned iterative linear solvers | |
| CN103336758A (en) | Sparse matrix storage method CSRL (Compressed Sparse Row with Local Information) and SpMV (Sparse Matrix Vector Multiplication) realization method based on same | |
| CN110516316B (en) | GPU acceleration method for solving Euler equation by interrupted Galerkin method | |
| Prasad et al. | GPU-based Parallel R-tree Construction and Querying | |
| Cao et al. | Implementing sparse matrix-vector multiplication using CUDA based on a hybrid sparse matrix format | |
| CN109002659A (en) | A kind of fluid machinery simulated program optimization method based on supercomputer | |
| CN104765589A (en) | Grid parallel preprocessing method based on MPI | |
| Shata et al. | Optimized implementation of OpenCL kernels on FPGAs | |
| Bernaschi et al. | A factored sparse approximate inverse preconditioned conjugate gradient solver on graphics processing units | |
| CN109522127A (en) | A kind of fluid machinery simulated program isomery accelerated method based on GPU | |
| Vartziotis et al. | Improved GETMe by adaptive mesh smoothing | |
| Mostafazadeh Davani et al. | Unsteady Navier-Stokes computations on GPU architectures | |
| CN106484532B (en) | GPGPU parallel calculating method towards SPH fluid simulation | |
| Zhu et al. | GPU acceleration of an iterative scheme for gas-kinetic model equations with memory reduction techniques | |
| CN102722472B (en) | Complex matrix optimizing method | |
| Bandyopadhyay et al. | GRS—GPU radix sort for multifield records | |
| Liu et al. | Parallel reconstruction of neighbor-joining trees for large multiple sequence alignments using CUDA | |
| Li et al. | A speculative HMMER search implementation on GPU | |
| Diamantopoulos et al. | A system-level transprecision FPGA accelerator for BLSTM using on-chip memory reshaping | |
| Charif et al. | Detailed and highly parallelizable cycle-accurate network-on-chip simulation on GPGPU | |
| CN109670001A (en) | Polygonal gird GPU parallel calculating method based on CUDA | |
| CN114116208A (en) | Short wave radiation transmission mode three-dimensional acceleration method based on GPU | |
| Zhang et al. | Implementing sparse matrix-vector multiplication with QCSR on GPU | |
| CN106648886A (en) | Realization method and apparatus for distributed storage system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |