测试CUDA 11 cusolverDnDSgels

分类：编程问答

如何解决测试CUDA 11 cusolverDnDSgels

试图理解cusolverDnDSgels功能。如果我像在文档中那样使用简单的3x3示例运行它，那么它可以工作，但是当我用我的数据运行它时，d_info返回-1，正如文档所说，如果d_info = -i则第i个参数无效。

下面我发布了3×3和4×3矩阵的代码，前者起作用而第二个不起作用。

作为参考，我使用了该网站计算器https://adrianstoll.com/linear-algebra/least-squares.html 

#include <stdio.h>
#include <stdlib.h>
#include <assert.h>

#include <cuda_runtime.h>
#include <cusolverDn.h>


void printMatrix(int m,int n,const double* A,int lda,const char* name)
{
    for (int row = 0; row < m; row++) {
        for (int col = 0; col < n; col++) {
            double Areg = A[row + col * lda];
            printf("%s(%d,%d) = %f\n",name,row + 1,col + 1,Areg);
        }
    }
}

int main(int argc,char*argv[])
{
    // 3x3 example works fine
    int m = 3;
    int n = 3;
    double A[9] = { 1.0,4.0,2.0,5.0,1.0,3.0,6.0,1.0 };
    double B[3] = { 6.0,15.0,4.0 };
    
    // 4x3 example d_info/info_gpu returns -1
    //int m = 4;
    //int n = 3;
    //double A[12] = { 1.0,2.0 };
    //double B[4] = { 6.0,5.0 };
    
    double X[3];
    
    int lda = m;
    int ldb = m;
    int ldx = n;
    int nrhs = 1;
    int niter = 0;
    int info_gpu = 0;
    size_t lwork = 0;
    
    double *d_A = NULL;
    double *d_B = NULL;
    double *d_X = NULL;
    double *d_work = NULL;
    int* d_info = NULL;
    
    cusolverDnHandle_t cusolverH = NULL;
    cudaError_t cudaStat = cudaSuccess;
    cusolverStatus_t cusolver_status = CUSOLVER_STATUS_SUCCESS;
    
    cusolver_status = cusolverDnCreate(&cusolverH);
    assert(CUSOLVER_STATUS_SUCCESS == cusolver_status);
    
    // Allocate space in the GPU
    cudaStat = cudaMalloc((void**)&d_A,sizeof(double) * m * n);
    assert(cudaSuccess == cudaStat);
    
    cudaStat = cudaMalloc((void**)&d_B,sizeof(double) * m * nrhs);
    assert(cudaSuccess == cudaStat);
    
    cudaStat = cudaMalloc((void**)&d_X,sizeof(double) * n * nrhs);
    assert(cudaSuccess == cudaStat);
    
    cudaStat = cudaMalloc((void**)&d_info,sizeof(int));
    assert(cudaSuccess == cudaStat);
    
    // Copy matrices into GPU space
    cudaStat = cudaMemcpy(d_A,A,sizeof(double) * m * n,cudaMemcpyHostToDevice);
    assert(cudaSuccess == cudaStat);
    cudaStat = cudaMemcpy(d_B,B,sizeof(double) * m * nrhs,cudaMemcpyHostToDevice);
    assert(cudaSuccess == cudaStat);
    
    // Get work buffer size
    cusolver_status = cusolverDnDSgels_bufferSize(cusolverH,m,n,nrhs,d_A,lda,d_B,ldb,d_X,ldx,d_work,&lwork);
    assert(CUSOLVER_STATUS_SUCCESS == cusolver_status);
    
    // Allocate workspace
    cudaStat = cudaMalloc((void**)&d_work,sizeof(float) * lwork);
    assert(cudaSuccess == cudaStat);
    
    // Run solver
    cusolver_status = cusolverDnDSgels(cusolverH,lwork,&niter,d_info);
    
    // Sync threads
    cudaStat = cudaDeviceSynchronize();
    assert(cudaSuccess == cudaStat);
    
    // Copy GPU info
    cudaStat = cudaMemcpy(&info_gpu,d_info,sizeof(int),cudaMemcpyDeviceToHost);
    assert(cudaSuccess == cudaStat);
    
    // Get solved data
    cudaStat = cudaMemcpy(X,sizeof(double) * n * nrhs,cudaMemcpyDeviceToHost);
    assert(cudaSuccess == cudaStat);
    
    printf("after DDgels: info_gpu = %d\n",info_gpu);
    printMatrix(n,X,"X");
    
    assert(CUSOLVER_STATUS_SUCCESS == cusolver_status);
    
    if (d_A) cudaFree(d_A);
    if (d_B) cudaFree(d_B);
    if (d_X) cudaFree(d_X);
    if (d_info) cudaFree(d_info);
    if (d_work) cudaFree(d_work);
    if (cusolverH) cusolverDnDestroy(cusolverH);
    cudaDeviceReset();
    return 0;
}

解决方法

不幸的是，cuSolver设置存在不一致之处，从而造成了此问题。 有一种方法可以避免这种问题，方法是调用专家API“ cusolverDnIRSXgels”“ cusolverDnIRSXgels_bufferSize”，从而为用户提供更多控制权。

因此在您的代码中替换

    cusolver_status = cusolverDnDDgels_bufferSize(cusolverH,m,n,nrhs,d_A,lda,d_B,ldb,d_X,ldx,d_work,&lwork);
    assert(CUSOLVER_STATUS_SUCCESS == cusolver_status);

    // Allocate workspace
    cudaStat = cudaMalloc((void**)&d_work,lwork);
    assert(cudaSuccess == cudaStat);

    // Run solver
    cusolver_status = cusolverDnDDgels(cusolverH,lwork,&niter,d_info);
    printf("gels status: %d\n",int(cusolver_status));

作者

    // create the params and info structure for the expert interface
    cusolverDnIRSParams_t gels_irs_params;
    cusolverDnIRSParamsCreate( &gels_irs_params );
    cusolverDnIRSInfos_t gels_irs_infos;
    cusolverDnIRSInfosCreate( &gels_irs_infos );

    // Set the main and the low precision of the solver DSgels 
    // D is for double S for single precision thus 
    // main_precision is CUSOLVER_R_FP64,low_precision is CUSOLVER_R_FP32
    cusolverDnIRSParamsSetSolverPrecisions( gels_irs_params,CUSOLVER_R_64F,CUSOLVER_R_32F );
    // Set the refinement solver.
    cusolverDnIRSParamsSetRefinementSolver( gels_irs_params,CUSOLVER_IRS_REFINE_CLASSICAL );
    // Get work buffer size
    cusolver_status = cusolverDnIRSXgels_bufferSize(cusolverH,gels_irs_params,&lwork);
    assert(CUSOLVER_STATUS_SUCCESS == cusolver_status);
    // Allocate workspace
    cudaStat = cudaMalloc((void**)&d_work,lwork);
    assert(cudaSuccess == cudaStat);
    // Run solver
    cusolver_status = cusolverDnIRSXgels(cusolverH,gels_irs_infos,(void *)d_A,(void *)d_B,(void *)d_X,int(cusolver_status));

还请注意，当m> n是一个超额认购的方程组时，您就不能选择RHS然后找到SO，因此，最好选择SOL，生成RHS = A * SOL，然后使用RHS和与SOL比较。

还要注意，LDX应该> = max（m，n）

我通过以下方式修改了您的代码：

#include <stdio.h>
#include <stdlib.h>
#include <assert.h>

#include <cuda_runtime.h>
#include <cusolverDn.h>


#define USE_BUG
typedef double mt;

#ifndef max
#define max(a,b) ((a) > (b) ? (a) : (b))
#endif

void matvec(int m,int n,int nrhs,const mt* A,int lda,mt *X,int ldx,mt *B,int ldb)
{
    mt sum[nrhs];

    for (int row = 0; row < m; row++) {
        for (int r = 0; r < nrhs; r++) sum[r] = 0.0;
        for (int col = 0; col < n; col++) {
            for (int r = 0; r < nrhs; r++){
                sum[r] += A[row + col * lda] * X[col + r*ldx];
            }
        }
        for (int r = 0; r < nrhs; r++) B[row + r*ldb] = sum[r];
    }
}

mt check_solution(int n,mt *ref,int ldr,int ldx)
{
    mt error=0.0;
    for (int r = 0; r < nrhs; r++){
        for (int i = 0; i < n; i++) {
            error = max(error,abs(ref[i+r*ldr] - X[i+r*ldr]));
        }
    }
    return error;
}


void printMatrix(int m,const char* name)
{
    for (int row = 0; row < m; row++) {
        for (int col = 0; col < n; col++) {
            mt Areg = A[row + col * lda];
            printf("%s(%d,%d) = %f\n",name,row + 1,col + 1,Areg);
        }
    }
}





int main(int argc,char*argv[])
{
#ifndef USE_BUG
        // 3x3 example works fine
    const int m = 3;
    const int n = 3;
    mt A[m*n] = { 1.0,4.0,2.0,5.0,1.0,3.0,6.0,1.0 };
    mt sol[n] = { 6.0,15.0,4.0 };
#else
    // 4x3 example d_info/info_gpu returns -1
    const int m = 4;
    const int n = 3;
    mt A[m*n] = { 1.0,2.0 };
    mt sol[n] =   { 6.0,4.0 };
#endif
    mt X[n];
    mt B[m];

    int lda = m;
    int ldb = max(m,n);
    int ldx = max(m,n);
    int nrhs = 1;
    int niter = 0;
    int info_gpu = 0;
    size_t lwork = 0;

    mt *d_A = NULL;
    mt *d_B = NULL;
    mt *d_X = NULL;
    mt *d_work = NULL;
    int* d_info = NULL;

    // compute B = A*sol
    matvec(m,A,sol,B,ldb);

    cusolverDnHandle_t cusolverH = NULL;
    cudaError_t cudaStat = cudaSuccess;
    cusolverStatus_t cusolver_status = CUSOLVER_STATUS_SUCCESS;

    cusolver_status = cusolverDnCreate(&cusolverH);
    assert(CUSOLVER_STATUS_SUCCESS == cusolver_status);

    // Allocate space in the GPU
    cudaStat = cudaMalloc((void**)&d_A,sizeof(mt) * m * n);
    assert(cudaSuccess == cudaStat);

    cudaStat = cudaMalloc((void**)&d_B,sizeof(mt) * m * nrhs);
    assert(cudaSuccess == cudaStat);

    cudaStat = cudaMalloc((void**)&d_X,sizeof(mt) * n * nrhs);
    assert(cudaSuccess == cudaStat);

    cudaStat = cudaMalloc((void**)&d_info,sizeof(int));
    assert(cudaSuccess == cudaStat);

    // Copy matrices into GPU space
    cudaStat = cudaMemcpy(d_A,sizeof(mt) * m * n,cudaMemcpyHostToDevice);
    assert(cudaSuccess == cudaStat);
    cudaStat = cudaMemcpy(d_B,sizeof(mt) * m * nrhs,cudaMemcpyHostToDevice);
    assert(cudaSuccess == cudaStat);

    #if 1
    // =======================================================
    // create the params and info structure for the expert interface
    cusolverDnIRSParams_t gels_irs_params;
    cusolverDnIRSParamsCreate( &gels_irs_params );
    cusolverDnIRSInfos_t gels_irs_infos;
    cusolverDnIRSInfosCreate( &gels_irs_infos );

    // Set the main and the low precision of the solver DSgels 
    // D is for double S for single precision thus 
    // main_precision is CUSOLVER_R_FP64,int(cusolver_status));
    #else

    // Get work buffer size
    cusolver_status = cusolverDnDDgels_bufferSize(cusolverH,int(cusolver_status));
    #endif

    // Sync threads
    cudaStat = cudaDeviceSynchronize();
    assert(cudaSuccess == cudaStat);

    // Copy GPU info
    cudaStat = cudaMemcpy(&info_gpu,d_info,sizeof(int),cudaMemcpyDeviceToHost);
    assert(cudaSuccess == cudaStat);

    // Get solved data
    cudaStat = cudaMemcpy(X,sizeof(mt) * n * nrhs,cudaMemcpyDeviceToHost);
    assert(cudaSuccess == cudaStat);

    printf("after gels: info_gpu = %d\n",info_gpu);
    printf("after gels: niter    = %d\n",niter);
    printf("after gels: error    = %e\n",check_solution(n,X,ldx));
    printMatrix(3,"X");


    if (d_A) cudaFree(d_A);
    if (d_B) cudaFree(d_B);
    if (d_X) cudaFree(d_X);
    if (d_info) cudaFree(d_info);
    if (d_work) cudaFree(d_work);
    if (cusolverH) cusolverDnDestroy(cusolverH);
    cudaDeviceReset();
    return 0;
}

使用nvcc -o test test.cu -lcusolver编译

版权声明：本文内容由互联网用户自发贡献，该文观点与技术仅代表作者本人。本站仅提供信息存储空间服务，不拥有所有权，不承担相关法律责任。如发现本站有涉嫌侵权/违法违规的内容， 请发送邮件至 dio@foxmail.com 举报，一经查实，本站将立刻删除。

相关推荐

matplotlib报错：AttributeError: module 'backend_interagg' has no attribute 'FigureCanvas'. Did you mean: 'FigureCanvasAgg'?
使用本地python环境可以成功执行 import pandas as pd import matplotlib.pyplot as plt # 设置字体 plt.rcParams[&#39;font.sans-serif&#39;] = [&#39;SimHei&#39;] # 能正确显示负号 p
gitlab登录失败，报错：This challenge page was accidentally cached by an intermediary and is no longer available.
设置时间 控制面板
后端开发常见错误
错误1：Request method ‘DELETE‘ not supported 错误还原：controller层有一个接口，访问该接口时报错：Request method ‘DELETE‘ not supported 错误原因：没有接收到前端传入的参数，修改为如下 参考 错误2：cannot r
docker常见错误
错误1：启动docker镜像时报错：Error response from daemon: driver failed programming external connectivity on endpoint quirky_allen 解决方法：重启docker -&gt; systemctl r
idea常见错误
错误1：private field ‘xxx‘ is never assigned 按Altʾnter快捷键，选择第2项 参考：https://blog.csdn.net/shi_hong_fei_hei/article/details/88814070 错误2：启动时报错，不能找到主启动类 #
pip安装依赖失败
报错如下，通过源不能下载，最后警告pip需升级版本 Requirement already satisfied: pip in c:\users\ychen\appdata\local\programs\python\python310\lib\site-packages (22.0.4) Coll
maven常见错误
错误1：maven打包报错 错误还原：使用maven打包项目时报错如下 [ERROR] Failed to execute goal org.apache.maven.plugins:maven-resources-plugin:3.2.0:resources (default-resources)
cloud项目中常见错误
错误1：服务调用时报错 服务消费者模块assess通过openFeign调用服务提供者模块hires 如下为服务提供者模块hires的控制层接口 @RestController @RequestMapping(&quot;/hires&quot;) public class FeignControl
springboot常见错误
错误1：运行项目后报如下错误 解决方案 报错2：Failed to execute goal org.apache.maven.plugins:maven-compiler-plugin:3.8.1:compile (default-compile) on project sb 解决方案：在pom.
springmvc拦截器中使用redisTemplate报空指针异常
参考 错误原因 过滤器或拦截器在生效时，redisTemplate还没有注入 解决方案：在注入容器时就生效 @Component //项目运行时就注入Spring容器 public class RedisBean { @Resource private RedisTemplate&lt;String
vite报错：npm ERR! command C:WINDOWSsystem32cmd.exe /d /s /c C:UsersychenAppDataLocalTempnpx-dde1c848.cmd
使用vite构建项目报错 C:\Users\ychen\work&gt;npm init @vitejs/app @vitejs/create-app is deprecated, use npm init vite instead C:\Users\ychen\AppData\Local\npm-
VM VirtualBox中安装centos时报错：设置基础软件仓库时出错
参考1 参考2 解决方案 # 点击安装源 协议选择 http:// 路径填写 mirrors.aliyun.com/centos/8.3.2011/BaseOS/x86_64/os URL类型 软件库URL 其他路径 # 版本 7 mirrors.aliyun.com/centos/7/os/x86
kafka常见问题
报错1 [root@slave1 data_mocker]# kafka-console-consumer.sh --bootstrap-server slave1:9092 --topic topic_db [2023-12-19 18:31:12,770] WARN [Consumer clie
hive常见问题
错误1 # 重写数据 hive (edu)&gt; insert overwrite table dwd_trade_cart_add_inc &gt; select data.id, &gt; data.user_id, &gt; data.course_id, &gt; date_format(
装载数据时报错：Failed to execute spark task, with exception 'org.apache.hadoop.hive.ql.metadata.HiveException(Failed to create spark client.)'
错误1 hive (edu)&gt; insert into huanhuan values(1,&#39;haoge&#39;); Query ID = root_20240110071417_fe1517ad-3607-41f4-bdcf-d00b98ac443e Total jobs = 1
flume常见问题
报错1：执行到如下就不执行了，没有显示Successfully registered new MBean. [root@slave1 bin]# /usr/local/software/flume-1.9.0/bin/flume-ng agent -n a1 -c /usr/local/softwa
hadoop集群常见错误
虚拟及没有启动任何服务器查看jps会显示jps，如果没有显示任何东西 [root@slave2 ~]# jps 9647 Jps 解决方案 # 进入/tmp查看 [root@slave1 dfs]# cd /tmp [root@slave1 tmp]# ll 总用量 48 drwxr-xr-x. 2
hive常见错误
报错1 hive&gt; show databases; OK Failed with exception java.io.IOException:java.lang.RuntimeException: Error in configuring object Time taken: 0.474 se
Linux常见错误
报错1 [root@localhost ~]# vim -bash: vim: 未找到命令 安装vim yum -y install vim* # 查看是否安装成功 [root@hadoop01 hadoop]# rpm -qa |grep vim vim-X11-7.4.629-8.el7_9.x
使用sqoop一直卡在：mapreduce.Job: Running job: job_1703173956074_0001
修改hadoop配置 vi /usr/local/software/hadoop-2.9.2/etc/hadoop/yarn-site.xml # 添加如下 &lt;configuration&gt; &lt;property&gt; &lt;name&gt;yarn.nodemanager.res
pythonJavaScriptjavaHTMLPHPreactjsC#AndroidCSSNode.jssqlrpython-3.xMysqLjQueryc++pandasFlutterangularIOSdjangolinuxswifttypescript路由器JSON路由器设置无线路由器h3c华三华三路由器设置华三路由器电脑软件教程arraysdocker软件图文教程Cvue.jslaravelspring-boot