对cuMemFreeHost的OpenACC调用返回错误

Question

我正在使用PGI C编译器--pgcc v16.10.0 64位 - 了解如何使用OpenACC进行编程

这是我模拟粒子传输过程的代码

typedef struct {
    double position;
    double direction;
    double weight;
    int cell;
    int group;
    int alive;
    } Particle;

int size = 100000;             // number of particles to be simulated
int tot = (int) (1.3 * size);  // this variable limits the maximum of next generation particles
int capacity = 0;              // this variable indicates the actual number of next generation particles

/* particles to be simulated */
Particle *par = (Particle *) malloc(size * sizeof(Particle));

/* next generation particles produced */
particle *next = (Particle *) malloc(tot * sizeof(Particle));

/* initialization */
for (int i = 0; i < size; i++){
    par[i].position = rand1() * 100.0;  // random number between 0.0~1.0
    par[i].direction = rand2();         // random number between -1.0~1.0
    par[i].weight = 1.0;
    par[i].cell = 2;
    par[i].group = rand1() > 0.5 ? 1 : 2;
    par[i].alive = 1;
}

/* some parameters used in simulation */
double keff = 1.0;
double tracklength, collision, absorption;

/* start simulating */
int generation;
for (generation = 1; generation <= 100; generation++){
    int CellID, MatID, GroupID;
    int k;   // k-th particle to be simulated
    #pragma acc parallel copy(capacity) copyin(par[0:size],size, keff) copyout(next[0:tot])
    #pragma acc loop reduction(+:tracklength, collision, absorption)
    for (k = 0; k < size; k++){
        /* do some calculating with par[k] */
        /* secondary particle produced under certain circumstances */
        if (condition){
            next[capacity].position = par[k].position;
            next[capacity].direction = rand2();
            next[capacity].weight = 1.0;
            next[capacity].cell = par[k].cell;
            next[capacity].group = rand1() < 0.9 ? 1 : 2;
            next[capacity].alive = 1;
            capacity++;
        }
    }
    /* after simulation of current generation, update the parameters */
    keff = ........          // one formula to update keff

    size = capacity;
    capacity = 0;
    tot = (int) (1.3 * size);
    free(par);
    par = next;
    next = (Particle *) malloc(tot * sizeof(Particle));
}

free(par);
free(next);

我用

编译了代码

pgcc -acc -Minfo=accel -ta=tesla:cc30,time -O0 main.c -o test

并获取以下信息：

Loop carried dependence of par->alive, par->cell, par->direction, par->position, par->weight, par->group prevents parallelization Loop carried dependence of par->direction, par->group, par->position prevents vectorization Loop carried reuse of next->position prevents parallelization

然后运行可执行文件./test

发生错误

call to cuMemFreeHost returned error 700: Illegal address during kernel execution

我不知道如何解决这个问题[SAD]。

BTW代码运行良好，并在gcc忽略#pragma

编译时返回正确的结果