Algorithm CUDA中求解线性系统的QR分解
我正在GPU上写一个图像恢复算法,详细内容见 求解线性系统的QR分解法Algorithm CUDA中求解线性系统的QR分解,algorithm,cuda,gpgpu,gpu,Algorithm,Cuda,Gpgpu,Gpu,我正在GPU上写一个图像恢复算法,详细内容见 求解线性系统的QR分解法 Ax=b 工作如下 min||Ax-b|| ---> ||QRx-b|| ---> ||(Q^T)QRx-(Q^T)b|| ---> ||Rx-(Q^T)b|| 其中R是上三角矩阵。由此产生的上三角线性系统易于求解 我想使用CULA工具来实现这个方法。CULA例程GEQRF计算QR分解。手册上说: 退出时,数组对角线上及其上方的元素包含 min(M,N)-by-N上部梯形矩阵R(R为上部 三角形
Ax=b
工作如下
min||Ax-b|| ---> ||QRx-b|| ---> ||(Q^T)QRx-(Q^T)b|| ---> ||Rx-(Q^T)b||
其中R
是上三角矩阵。由此产生的上三角线性系统易于求解
我想使用CULA工具来实现这个方法。CULA例程GEQRF
计算QR分解。手册上说:
退出时,数组对角线上及其上方的元素包含
min(M,N)-by-N
上部梯形矩阵R
(R
为上部
三角形如果m>=n)
;对角线下方的元素,带有
数组TAU
,将正交/酉矩阵Q
表示为乘积
基本反射器的min(m,n)
我不知道Q
存储在哪里,而且算法对我来说太复杂了。你能给点建议吗
谢谢
void GEQRF(int M,int N,T* A,int LDA, T* TAU, T* WORK,int LWORK,int &INFO)
在GEQRF之后,R存储在A的上三角部分。然后,可以使用xORGQR生成Q,A和TAU作为输入
更多说明:自2015年2月起,CUDA 7.0(现为候选版本)提供了新的cuSOLVER库,包括计算矩阵的QR分解的可能性。这与库布拉斯图书馆一起,能够根据《库索尔弗用户指南》附录C中阐述的指南求解线性系统 您必须遵循以下三个步骤: 1)
geqrf
:它通过返回A
的上三角部分中的上三角矩阵R
,以及A
的下三角部分中以Householder向量形式存储的矩阵Q
来计算矩阵的QR分解,而户主向量的比例因子由TAU
参数返回
2) ormqr
:通过覆盖C
返回Q
和矩阵C
的乘积
3) trsm
:它求解一个上三角线性系统
下面,我将提供这些例程使用的完整示例
#include "cuda_runtime.h"
#include "device_launch_paraMeters.h"
#include<iostream>
#include<fstream>
#include<iomanip>
#include<stdlib.h>
#include<stdio.h>
#include<assert.h>
#include <cusolverDn.h>
#include <cublas_v2.h>
#include <cuda_runtime_api.h>
#include "Utilities.cuh"
#include "TimingGPU.cuh"
#define BLOCK_SIZE 32
#define prec_save 10
/***************/
/* COPY KERNEL */
/***************/
__global__ void copy_kernel(const double * __restrict d_in, double * __restrict d_out, const int M, const int N) {
const int i = blockIdx.x * blockDim.x + threadIdx.x;
const int j = blockIdx.y * blockDim.y + threadIdx.y;
if ((i < N) && (j < N)) d_out[j * N + i] = d_in[j * M + i];
}
/****************************************************/
/* LOAD INDIVIDUAL REAL MATRIX FROM txt FILE TO CPU */
/****************************************************/
// --- Load individual real matrix from txt file
template <class T>
void loadCPUrealtxt(T * __restrict h_out, const char *filename, const int M) {
std::ifstream infile;
infile.open(filename);
for (int i = 0; i < M; i++) {
double temp;
infile >> temp;
h_out[i] = (T)temp;
}
infile.close();
}
/************************************/
/* SAVE REAL ARRAY FROM GPU TO FILE */
/************************************/
template <class T>
void saveGPUrealtxt(const T * d_in, const char *filename, const int M) {
T *h_in = (T *)malloc(M * sizeof(T));
gpuErrchk(cudaMemcpy(h_in, d_in, M * sizeof(T), cudaMemcpyDeviceToHost));
std::ofstream outfile;
outfile.open(filename);
for (int i = 0; i < M; i++) outfile << std::setprecision(prec_save) << h_in[i] << "\n";
outfile.close();
}
/********/
/* MAIN */
/********/
int main(){
// --- Extension of Appendix C.1 of cuSOLVER library User's Guide
// --- See also http://www.netlib.org/lapack/lug/node40.html
// --- ASSUMPTION Nrows >= Ncols
const int Nrows = 500;
const int Ncols = 500;
TimingGPU timerGPU;
double timingQR, timingSolve;
// --- cuSOLVE input/output parameters/arrays
int work_size = 0;
int *devInfo; gpuErrchk(cudaMalloc(&devInfo, sizeof(int)));
// --- CUDA solver initialization
cusolverDnHandle_t solver_handle;
cusolveSafeCall(cusolverDnCreate(&solver_handle));
// --- CUBLAS initialization
cublasHandle_t cublas_handle;
cublasSafeCall(cublasCreate(&cublas_handle));
/***********************/
/* SETTING THE PROBLEM */
/***********************/
// --- Setting the host, Nrows x Ncols matrix
double *h_A = (double *)malloc(Nrows * Ncols * sizeof(double));
loadCPUrealtxt(h_A, "D:\\Project\\solveNonSquareLinearSystemQRCUDA\\solveNonSquareLinearSystemQRCUDA\\testMatrix.txt", Nrows * Ncols);
// --- Setting the device matrix and moving the host matrix to the device
double *d_A; gpuErrchk(cudaMalloc(&d_A, Nrows * Ncols * sizeof(double)));
gpuErrchk(cudaMemcpy(d_A, h_A, Nrows * Ncols * sizeof(double), cudaMemcpyHostToDevice));
// --- Initializing the data matrix C (Of course, this step could be done by a kernel function directly on the device).
// --- Notice that, in this case, only the first column of C contains actual data, the others being empty (zeroed). However, cuBLAS trsm
// has the capability of solving triangular linear systems with multiple right hand sides.
double *h_C = (double *)calloc(Nrows * Nrows, sizeof(double));
loadCPUrealtxt(h_C, "D:\\Project\\solveNonSquareLinearSystemQRCUDA\\solveNonSquareLinearSystemQRCUDA\\testVector.txt", Nrows);
double *d_C; gpuErrchk(cudaMalloc(&d_C, Nrows * Nrows * sizeof(double)));
gpuErrchk(cudaMemcpy(d_C, h_C, Nrows * Nrows * sizeof(double), cudaMemcpyHostToDevice));
/**********************************/
/* COMPUTING THE QR DECOMPOSITION */
/**********************************/
timerGPU.StartCounter();
// --- CUDA QR GEQRF preliminary operations
double *d_TAU; gpuErrchk(cudaMalloc((void**)&d_TAU, min(Nrows, Ncols) * sizeof(double)));
cusolveSafeCall(cusolverDnDgeqrf_bufferSize(solver_handle, Nrows, Ncols, d_A, Nrows, &work_size));
double *work; gpuErrchk(cudaMalloc(&work, work_size * sizeof(double)));
// --- CUDA GEQRF execution: The matrix R is overwritten in upper triangular part of A, including diagonal
// elements. The matrix Q is not formed explicitly, instead, a sequence of householder vectors are
// stored in lower triangular part of A.
cusolveSafeCall(cusolverDnDgeqrf(solver_handle, Nrows, Ncols, d_A, Nrows, d_TAU, work, work_size, devInfo));
int devInfo_h = 0; gpuErrchk(cudaMemcpy(&devInfo_h, devInfo, sizeof(int), cudaMemcpyDeviceToHost));
if (devInfo_h != 0) std::cout << "Unsuccessful gerf execution\n\n";
timingQR = timerGPU.GetCounter();
printf("Timing for QR calculation = %f [ms]\n", timingQR);
/*****************************/
/* SOLVING THE LINEAR SYSTEM */
/*****************************/
timerGPU.StartCounter();
// --- CUDA ORMQR execution: Computes the multiplication Q^T * C and stores it in d_C
cusolveSafeCall(cusolverDnDormqr(solver_handle, CUBLAS_SIDE_LEFT, CUBLAS_OP_T, Nrows, Ncols, min(Nrows, Ncols), d_A, Nrows, d_TAU, d_C, Nrows, work, work_size, devInfo));
// --- Reducing the linear system size
double *d_R; gpuErrchk(cudaMalloc(&d_R, Ncols * Ncols * sizeof(double)));
double *d_B; gpuErrchk(cudaMalloc(&d_B, Ncols * sizeof(double)));
dim3 Grid(iDivUp(Ncols, BLOCK_SIZE), iDivUp(Ncols, BLOCK_SIZE));
dim3 Block(BLOCK_SIZE, BLOCK_SIZE);
copy_kernel << <Grid, Block >> >(d_A, d_R, Nrows, Ncols);
gpuErrchk(cudaMemcpy(d_B, d_C, Ncols * sizeof(double), cudaMemcpyDeviceToDevice));
// --- Solving an upper triangular linear system - compute x = R \ Q^T * B
const double alpha = 1.;
cublasSafeCall(cublasDtrsm(cublas_handle, CUBLAS_SIDE_LEFT, CUBLAS_FILL_MODE_UPPER, CUBLAS_OP_N,
CUBLAS_DIAG_NON_UNIT, Ncols, 1, &alpha, d_R, Ncols, d_B, Ncols));
timingSolve = timerGPU.GetCounter();
printf("Timing for solution of the linear system = %f [ms]\n", timingSolve);
printf("Overall timing = %f [ms]\n", timingQR + timingSolve);
/************************/
/* CHECKING THE RESULTS */
/************************/
// --- The upper triangular part of A contains the elements of R. Showing this.
saveGPUrealtxt(d_A, "D:\\Project\\solveNonSquareLinearSystemQRCUDA\\solveNonSquareLinearSystemQRCUDA\\d_R.txt", Nrows * Ncols);
// --- The first Nrows elements of d_C contain the result of Q^T * C
saveGPUrealtxt(d_C, "D:\\Project\\solveNonSquareLinearSystemQRCUDA\\solveNonSquareLinearSystemQRCUDA\\d_QTC.txt", Nrows);
// --- Initializing the output Q matrix (Of course, this step could be done by a kernel function directly on the device)
double *h_Q = (double *)malloc(Nrows * Nrows * sizeof(double));
for (int j = 0; j < Nrows; j++)
for (int i = 0; i < Nrows; i++)
if (j == i) h_Q[j + i*Nrows] = 1.;
else h_Q[j + i*Nrows] = 0.;
double *d_Q; gpuErrchk(cudaMalloc(&d_Q, Nrows * Nrows * sizeof(double)));
gpuErrchk(cudaMemcpy(d_Q, h_Q, Nrows * Nrows * sizeof(double), cudaMemcpyHostToDevice));
// --- Calculation of the Q matrix
cusolveSafeCall(cusolverDnDormqr(solver_handle, CUBLAS_SIDE_LEFT, CUBLAS_OP_N, Nrows, Ncols, min(Nrows, Ncols), d_A, Nrows, d_TAU, d_Q, Nrows, work, work_size, devInfo));
// --- d_Q contains the elements of Q. Showing this.
saveGPUrealtxt(d_Q, "D:\\Project\\solveNonSquareLinearSystemQRCUDA\\solveNonSquareLinearSystemQRCUDA\\d_Q.txt", Nrows * Nrows);
// --- At this point, d_C contains the elements of Q^T * C, where C is the data vector. Showing this.
// --- According to the above, only the first column of d_C makes sense.
//gpuErrchk(cudaMemcpy(h_C, d_C, Nrows * Nrows * sizeof(double), cudaMemcpyDeviceToHost));
//printf("\n\n");
//for (int j = 0; j < Nrows; j++)
// for (int i = 0; i < Nrows; i++)
// printf("C[%i, %i] = %f\n", j, i, h_C[j + i*Nrows]);
// --- Check final result
saveGPUrealtxt(d_B, "D:\\Project\\solveNonSquareLinearSystemQRCUDA\\solveNonSquareLinearSystemQRCUDA\\d_B.txt", Ncols);
cusolveSafeCall(cusolverDnDestroy(solver_handle));
return 0;
}
请根据需要对行进行注释/取消注释
定时
计时(毫秒)(在GTX960卡上执行的测试,抄送5.2):
以下代码是Jackolanten对一般M-by-K输入RHS矩阵b的回答的略微扩展。基本上你需要复制R和中间b的上矩阵,这样矩阵就有了正确的步幅
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
#include <iostream>
#include "cuda_runtime.h"
#include "cublas_v2.h"
#include "cusolverDn.h"
#include "cublas_test.h"
#include "Eigen/Dense"
#include "gpu_util.h"
//##############################################################################
template<typename T>
void PrintEMatrix(const T &mat, const char *name) {
std::cout << name << " =\n";
std::cout << mat << std::endl;
}
//##############################################################################
template<typename T>
__global__
void Ker_CopyUpperSubmatrix(const T *__restrict d_in,
T *__restrict d_ou,
const int M, const int N, const int subM) {
const int i = threadIdx.x + blockIdx.x*blockDim.x;
const int j = threadIdx.y + blockIdx.y*blockDim.y;
if (i>=subM || j>=N)
return;
d_ou[j*subM+i] = d_in[j*M+i];
}
//##############################################################################
int TestQR() {
typedef double T; // NOTE: don't change this. blas has different func name
typedef Eigen::Matrix<T,Eigen::Dynamic,Eigen::Dynamic> MatrixXd;
typedef Eigen::Matrix<T,Eigen::Dynamic,1> VectorXd;
// define handles
cusolverDnHandle_t cusolverH = NULL;
cublasHandle_t cublasH = NULL;
const int M = 3;
const int N = 2;
const int K = 5;
MatrixXd A;
A = MatrixXd::Random(M,N);
MatrixXd x_ref, x_sol;
x_sol.resize(N,K);
x_ref = MatrixXd::Random(N,K);
MatrixXd b = A*x_ref;
PrintEMatrix(A, "A");
PrintEMatrix(b, "b");
PrintEMatrix(x_ref, "x_ref");
#define CUSOLVER_ERRCHK(x) \
assert(x == CUSOLVER_STATUS_SUCCESS && "cusolver failed");
#define CUBLAS_ERRCHK(x) \
assert(x == CUBLAS_STATUS_SUCCESS && "cublas failed");
CUSOLVER_ERRCHK(cusolverDnCreate(&cusolverH));
CUBLAS_ERRCHK(cublasCreate(&cublasH));
T *d_A, *d_b, *d_work, *d_work2, *d_tau;
int *d_devInfo, devInfo;
gpuErrchk(cudaMalloc((void**)&d_A, sizeof(T)*M*N));
gpuErrchk(cudaMalloc((void**)&d_b, sizeof(T)*M*K));
gpuErrchk(cudaMalloc((void**)&d_tau, sizeof(T)*M));
gpuErrchk(cudaMalloc((void**)&d_devInfo, sizeof(int)));
gpuErrchk(cudaMemcpy(d_A, A.data(), sizeof(T)*M*N, cudaMemcpyHostToDevice));
gpuErrchk(cudaMemcpy(d_b, b.data(), sizeof(T)*M*K, cudaMemcpyHostToDevice));
int bufSize,bufSize2;
// in-place A = QR
CUSOLVER_ERRCHK(
cusolverDnDgeqrf_bufferSize(
cusolverH,
M,
N,
d_A,
M,
&bufSize
)
);
gpuErrchk(cudaMalloc((void**)&d_work, sizeof(T)*bufSize));
CUSOLVER_ERRCHK(
cusolverDnDgeqrf(
cusolverH,
M,
N,
d_A,
M,
d_tau,
d_work,
bufSize,
d_devInfo
)
);
gpuErrchk(cudaMemcpy(&devInfo, d_devInfo, sizeof(int),
cudaMemcpyDeviceToHost));
assert(0 == devInfo && "QR factorization failed");
// Q^T*b
CUSOLVER_ERRCHK(
cusolverDnDormqr_bufferSize(
cusolverH,
CUBLAS_SIDE_LEFT,
CUBLAS_OP_T,
M,
K,
N,
d_A,
M,
d_tau,
d_b,
M,
&bufSize2
)
);
gpuErrchk(cudaMalloc((void**)&d_work2, sizeof(T)*bufSize2));
CUSOLVER_ERRCHK(
cusolverDnDormqr(
cusolverH,
CUBLAS_SIDE_LEFT,
CUBLAS_OP_T,
M,
K,
min(M,N),
d_A,
M,
d_tau,
d_b,
M,
d_work2,
bufSize2,
d_devInfo
)
);
gpuErrchk(cudaDeviceSynchronize());
gpuErrchk(cudaMemcpy(&devInfo, d_devInfo, sizeof(int),
cudaMemcpyDeviceToHost));
assert(0 == devInfo && "Q^T b failed");
// need to explicitly copy submatrix for the triangular solve
T *d_R, *d_b_;
gpuErrchk(cudaMalloc((void**)&d_R, sizeof(T)*N*N));
gpuErrchk(cudaMalloc((void**)&d_b_,sizeof(T)*N*K));
dim3 thd_size(32,32);
dim3 blk_size((N+thd_size.x-1)/thd_size.x,(N+thd_size.y-1)/thd_size.y);
Ker_CopyUpperSubmatrix<T><<<blk_size,thd_size>>>(d_A, d_R, M, N, N);
blk_size = dim3((N+thd_size.x-1)/thd_size.x,(K+thd_size.y-1)/thd_size.y);
Ker_CopyUpperSubmatrix<T><<<blk_size,thd_size>>>(d_b, d_b_, M, K, N);
// solve x = R \ (Q^T*B)
const double one = 1.0;
CUBLAS_ERRCHK(
cublasDtrsm(
cublasH,
CUBLAS_SIDE_LEFT,
CUBLAS_FILL_MODE_UPPER,
CUBLAS_OP_N,
CUBLAS_DIAG_NON_UNIT,
N,
K,
&one,
d_R,
N,
d_b_,
N
)
);
gpuErrchk(cudaDeviceSynchronize());
gpuErrchk(cudaMemcpy(x_sol.data(), d_b_, sizeof(T)*N*K,
cudaMemcpyDeviceToHost));
PrintEMatrix(x_ref, "x_ref");
PrintEMatrix(x_sol, "x_sol");
std::cout << "solution l2 error = " << (x_ref-x_sol).norm()
<< std::endl;
exit(0);
return 0;
}
//##############################################################################
#包括
#包括
#包括
#包括
#包括“cuda_runtime.h”
#包括“cublas_v2.h”
#包括“cusolverDn.h”
#包括“cublas_test.h”
#包括“本征/密集”
#包括“gpu_util.h”
//##############################################################################
模板
无效打印矩阵(常量T&mat,常量字符*名称){
std::cout Thank现在我知道如何获得Q,使用更新的QR方法可以提供一些建议吗?culaDeviceSgels可以获得Q?那里没有TAU输入?是的,如果你愿意,伪代码是不够的,我真的不明白你打算做什么,因为你想解线性系统(Ax=b)使用QR分解,所以你的输入就是你的b和矩阵A,或者如果你想独立于问题计算Q和R,在这种情况下,你只需要矩阵A作为输入,这就是矩阵A的分解。我正在尝试解决一个称为OMP的图像恢复算法,这意味着Ax=b将被计算至少500在每次计算中,b保持不变,A将添加一列,因此我想使用上一个Q的结果,计算更新列的系数以节省运行时间。2019年,实用程序。cuh
已从您引用的项目中删除。您可以升级吗吃了你的解决方案?(这就是为什么引用可能发生变化的外部回购可能是危险的)。“TimingGPU.cuh”似乎也不见了!double*h_C=(double*)calloc(Nrows*Nrows,sizeof(double))
-Nrows*Nrows
是错误还是实际需要?我希望您的解决方案能够适应我的问题,因为我的问题矩阵(6列,但有数百万行)。这会杀了我。报告的只是一个例子,说明C
可以有多个列,但我认为cusolverndormqr
也可以用于单列矩阵C
。您可以在此处找到正在搜索的文件。
Size QR decomposition Solving system Overall
100x100 0.89 1.41 2.30
200x200 5.97 3.23 9.20
500x500 17.08 21.6 38.7
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
#include <iostream>
#include "cuda_runtime.h"
#include "cublas_v2.h"
#include "cusolverDn.h"
#include "cublas_test.h"
#include "Eigen/Dense"
#include "gpu_util.h"
//##############################################################################
template<typename T>
void PrintEMatrix(const T &mat, const char *name) {
std::cout << name << " =\n";
std::cout << mat << std::endl;
}
//##############################################################################
template<typename T>
__global__
void Ker_CopyUpperSubmatrix(const T *__restrict d_in,
T *__restrict d_ou,
const int M, const int N, const int subM) {
const int i = threadIdx.x + blockIdx.x*blockDim.x;
const int j = threadIdx.y + blockIdx.y*blockDim.y;
if (i>=subM || j>=N)
return;
d_ou[j*subM+i] = d_in[j*M+i];
}
//##############################################################################
int TestQR() {
typedef double T; // NOTE: don't change this. blas has different func name
typedef Eigen::Matrix<T,Eigen::Dynamic,Eigen::Dynamic> MatrixXd;
typedef Eigen::Matrix<T,Eigen::Dynamic,1> VectorXd;
// define handles
cusolverDnHandle_t cusolverH = NULL;
cublasHandle_t cublasH = NULL;
const int M = 3;
const int N = 2;
const int K = 5;
MatrixXd A;
A = MatrixXd::Random(M,N);
MatrixXd x_ref, x_sol;
x_sol.resize(N,K);
x_ref = MatrixXd::Random(N,K);
MatrixXd b = A*x_ref;
PrintEMatrix(A, "A");
PrintEMatrix(b, "b");
PrintEMatrix(x_ref, "x_ref");
#define CUSOLVER_ERRCHK(x) \
assert(x == CUSOLVER_STATUS_SUCCESS && "cusolver failed");
#define CUBLAS_ERRCHK(x) \
assert(x == CUBLAS_STATUS_SUCCESS && "cublas failed");
CUSOLVER_ERRCHK(cusolverDnCreate(&cusolverH));
CUBLAS_ERRCHK(cublasCreate(&cublasH));
T *d_A, *d_b, *d_work, *d_work2, *d_tau;
int *d_devInfo, devInfo;
gpuErrchk(cudaMalloc((void**)&d_A, sizeof(T)*M*N));
gpuErrchk(cudaMalloc((void**)&d_b, sizeof(T)*M*K));
gpuErrchk(cudaMalloc((void**)&d_tau, sizeof(T)*M));
gpuErrchk(cudaMalloc((void**)&d_devInfo, sizeof(int)));
gpuErrchk(cudaMemcpy(d_A, A.data(), sizeof(T)*M*N, cudaMemcpyHostToDevice));
gpuErrchk(cudaMemcpy(d_b, b.data(), sizeof(T)*M*K, cudaMemcpyHostToDevice));
int bufSize,bufSize2;
// in-place A = QR
CUSOLVER_ERRCHK(
cusolverDnDgeqrf_bufferSize(
cusolverH,
M,
N,
d_A,
M,
&bufSize
)
);
gpuErrchk(cudaMalloc((void**)&d_work, sizeof(T)*bufSize));
CUSOLVER_ERRCHK(
cusolverDnDgeqrf(
cusolverH,
M,
N,
d_A,
M,
d_tau,
d_work,
bufSize,
d_devInfo
)
);
gpuErrchk(cudaMemcpy(&devInfo, d_devInfo, sizeof(int),
cudaMemcpyDeviceToHost));
assert(0 == devInfo && "QR factorization failed");
// Q^T*b
CUSOLVER_ERRCHK(
cusolverDnDormqr_bufferSize(
cusolverH,
CUBLAS_SIDE_LEFT,
CUBLAS_OP_T,
M,
K,
N,
d_A,
M,
d_tau,
d_b,
M,
&bufSize2
)
);
gpuErrchk(cudaMalloc((void**)&d_work2, sizeof(T)*bufSize2));
CUSOLVER_ERRCHK(
cusolverDnDormqr(
cusolverH,
CUBLAS_SIDE_LEFT,
CUBLAS_OP_T,
M,
K,
min(M,N),
d_A,
M,
d_tau,
d_b,
M,
d_work2,
bufSize2,
d_devInfo
)
);
gpuErrchk(cudaDeviceSynchronize());
gpuErrchk(cudaMemcpy(&devInfo, d_devInfo, sizeof(int),
cudaMemcpyDeviceToHost));
assert(0 == devInfo && "Q^T b failed");
// need to explicitly copy submatrix for the triangular solve
T *d_R, *d_b_;
gpuErrchk(cudaMalloc((void**)&d_R, sizeof(T)*N*N));
gpuErrchk(cudaMalloc((void**)&d_b_,sizeof(T)*N*K));
dim3 thd_size(32,32);
dim3 blk_size((N+thd_size.x-1)/thd_size.x,(N+thd_size.y-1)/thd_size.y);
Ker_CopyUpperSubmatrix<T><<<blk_size,thd_size>>>(d_A, d_R, M, N, N);
blk_size = dim3((N+thd_size.x-1)/thd_size.x,(K+thd_size.y-1)/thd_size.y);
Ker_CopyUpperSubmatrix<T><<<blk_size,thd_size>>>(d_b, d_b_, M, K, N);
// solve x = R \ (Q^T*B)
const double one = 1.0;
CUBLAS_ERRCHK(
cublasDtrsm(
cublasH,
CUBLAS_SIDE_LEFT,
CUBLAS_FILL_MODE_UPPER,
CUBLAS_OP_N,
CUBLAS_DIAG_NON_UNIT,
N,
K,
&one,
d_R,
N,
d_b_,
N
)
);
gpuErrchk(cudaDeviceSynchronize());
gpuErrchk(cudaMemcpy(x_sol.data(), d_b_, sizeof(T)*N*K,
cudaMemcpyDeviceToHost));
PrintEMatrix(x_ref, "x_ref");
PrintEMatrix(x_sol, "x_sol");
std::cout << "solution l2 error = " << (x_ref-x_sol).norm()
<< std::endl;
exit(0);
return 0;
}
//##############################################################################