In order to use my GPU with OpenACC and cublas in an application, which is comiled with g++, I setup a small testexample. Therefor I created the files:
- main.cpp
- pgiCudaCode.h
- pgiCudaCode.cpp
My testsystem ist an Ubuntu 18.04 linux with g++ version 7.5.0 and the pgc++ version 19.10-0 with a Nvidia GTX1070 card.
The file pgiCudaCode.cpp got some impementations for a general matrix-verctor multiply with openACC and cublas. This file is complied with the PGI compiler and the command:
pgc++ -fast -Minfo=opt -ta:tesla:cc60,managed,nordc -Mcudalib=cublas -Minfo=accel -fPIC pgiCudaCode.cpp -c pgiCudaCode.o
I do use the option nordc in order to operate with g++.
The main-file is compiled and linked with g++:
g++ -fPIC pgiCudaCode.o -L/opt/pgi/linux86-64/19.10/lib/ -laccapi -laccg -laccn -laccg2 -lpgiman -ldl -lcudadevice -lcudapgi -lomp -lnuma -lpthread -lnspgc -lpgc -lm -lgcc -lc -lgcc -lpgmath -lblas -lpgatm -lpgkomp -L/opt/pgi/linux86-64/2019/cuda/10.1/lib64/ -lcublas -lcublasLt -lcudart main.cpp -o mainGCC
After setting up these exports on my Ubuntu 18.04
export LD_LIBRARY_PATH="/opt/pgi/linux86-64/19.10/lib/:$LD_LIBRARY_PATH"
export LD_LIBRARY_PATH="/opt/pgi/linux86-64/2019/cuda/10.1/lib64/:$LD_LIBRARY_PATH"
I can run the Executable mainGCC an get the following output:
./mainGCC
Vector 1:
1
1
Matrix:
1 3
2 4
matrix*vec pure openACC:
4
6
matrix*vec cublas with internal allocation:
4
6
matrix*vec cublas without internal allocation:
Failing in Thread:1
call to cuMemcpyDtoHAsync returned error 700: Illegal address during kernel execution
Failing in Thread:1
call to cuMemFreeHost returned error 700: Illegal address during kernel execution
I do not get this error, when linking and compiling main.cpp with the pgi compiler:
pgc++ -fast -Minfo=opt -ta:tesla:cc60,managed,nordc -Mcudalib=cublas -Minfo=accel -fPIC pgiCudaCode.o main.cpp -o mainPGI
Here the output of mainPGI is correct:
Vector 1:
1
1
Matrix:
1 3
2 4
matrix*vec pure openACC:
4
6
matrix*vec cublas with internal allocation:
4
6
matrix*vec cublas without internal allocation:
4
6
So the interesting parts are:
- When using g++ allocated memory in the function matmulPureOpenACC it does not cause any problem, while the function matmul using cublas has an error.
- When you use pgi allocated memory in the function matmul_internAlloc, cublas can operate with g++.
This brings me to my question:
How can i prevent this error, using g++ allocated memory in the function matmul?
Here are the corresponding .cpp and .h files.
main.cpp:
#include <iostream>
#include "pgiCudaCode.h"
void printVec(int N, double* vec)
{
for(int i = 0; i < N; i++)
{
std::cout << vec[i] << std::endl;
}
}
void printMatrix(int N, double* matr)
{
for(int i = 0; i < N; i++)
{
for(int j = 0; j < N; j++)
{
std::cout << '\t' << matr[i + j * N];
}
std::cout << std::endl;
}
}
int main()
{
int N = 2;
double* vec1 = new double[N];
vec1[0] = 1.0;
vec1[1] = 1.0;
double* vec2 = new double[N];
vec2[0] = 0.0;
vec2[1] = 0.0;
double* matr = new double[N*N];
matr[0] = 1.0;
matr[1] = 2.0;
matr[2] = 3.0;
matr[3] = 4.0;
std::cout << "Vector 1:" << std::endl;
printVec(N, vec1);
std::cout << std::endl;
std::cout << "Matrix:" << std::endl;
printMatrix(N, matr);
std::cout << std::endl;
std::cout << "matrix*vec pure openACC:" << std::endl;
matmulPureOpenACC(N, matr, vec1, vec2);
printVec(N, vec2);
std::cout << std::endl;
vec2[0] = 0.0;
vec2[1] = 0.0;
std::cout << "matrix*vec cublas with internal allocation:" << std::endl;
matmul_internAlloc(N, matr, vec1, vec2);
printVec(N, vec2);
std::cout << std::endl;
vec2[0] = 0.0;
vec2[1] = 0.0;
std::cout << "matrix*vec cublas without internal allocation:" << std::endl;
matmul(N, matr, vec1, vec2);
printVec(N, vec2);
std::cout << std::endl;
delete [] vec1;
delete [] vec2;
delete [] matr;
return 0;
}
pgiCudaCode.h:
#ifndef PGICUDACODE_H
#define PGICUDACODE_H
bool matmul(int n, const double* matr, const double* b, double* c);
bool matmul_internAlloc(int n, const double* matr, const double* b, double* c);
bool matmulPureOpenACC(int n, const double* matr, const double* b, double* c);
#endif
pgiCudaCode.cpp:
#include <iostream>
#include <cublas_v2.h>
void matmul(int n, const double* matr, const double* b, double* c)
{
#pragma acc data pcopyin(n , matr[0:n*n], b[0:n]) pcopy(c[0:n])
{
#pragma acc host_data use_device(matr, b, c)
{
cublasHandle_t handle;
cublasStatus_t stat = cublasCreate(&handle);
if ( CUBLAS_STATUS_SUCCESS != stat ) {
std::cerr<<"CUBLAS initialization failed"<<std::endl;
}
if ( CUBLAS_STATUS_SUCCESS == stat )
{
const double alpha = 1.0;
const double beta = 1.0;
stat = cublasDgemv_v2(handle, CUBLAS_OP_N, n,n, &alpha, matr, n, b, 1, &beta, c, 1);
if (stat != CUBLAS_STATUS_SUCCESS) {
std::cerr<<"cublasDgemm failed"<<std::endl;
}
}
cublasDestroy(handle);
}
}
}
void matmul_internAlloc(int n2, const double* matr2, const double* b2, double* c2)
{
int n = n2;
double* matr = new double[n*n];
double* b = new double[n];
double* c = new double[n];
std::copy(&matr2[0], &matr2[n*n], &matr[0]);
std::copy(&b2[0], &b2[n], &b[0]);
std::copy(&c2[0], &c2[n], &c[0]);
#pragma acc data pcopyin(n , matr[0:n*n], b[0:n]) pcopy(c[0:n])
{
#pragma acc host_data use_device(matr, b, c)
{
cublasHandle_t handle;
cublasStatus_t stat = cublasCreate(&handle);
if ( CUBLAS_STATUS_SUCCESS != stat ) {
std::cerr<<"CUBLAS initialization failed"<<std::endl;
}
if ( CUBLAS_STATUS_SUCCESS == stat )
{
const double alpha = 1.0;
const double beta = 1.0;
stat = cublasDgemv_v2(handle, CUBLAS_OP_N, n,n, &alpha, matr, n, b, 1, &beta, c, 1);
if (stat != CUBLAS_STATUS_SUCCESS) {
std::cerr<<"cublasDgemm failed"<<std::endl;
}
}
cublasDestroy(handle);
}
}
std::copy(&c[0], &c[n], &c2[0]);
delete [] matr;
delete [] b;
delete [] c;
}
void matmulPureOpenACC(int n, const double* matr, const double* b, double* c)
{
#pragma acc data pcopyin(n, matr[0:n*n], b[0:n]) pcopy(c[0:n])
{
#pragma acc parallel loop
for(int i = 0; i < n; i++)
{
#pragma acc loop seq
for(int j = 0; j < n; j++)
{
c[i] += matr[i + j*n]*b[j];
}
}
}
}
