Question

我在尝试并行化k-means算法的以下功能时遇到了麻烦。原始代码就是这个：

void PointsToCentroides(int *Punts, int N, int D, int C, double *Centroides, int *PC, int *Sep){
int i, j, k, m;
int dist[C];    
for(i=0; i<C; i++)Sep[i]=0;     
for(i=0; i<N; i++){
    for(j=0; j<C; j++){
        dist[j] = 0;
        for(k=0; k<D; k++){
            dist[j] += (Punts[i*D+k] - Centroides[j*D+k]) * (Punts[i*D+k] - Centroides[j*D+k]);
        }
    }
    m = min(dist, C);
    PC[m*N+Sep[m]] = i;
    Sep[m]+=1;
}}

我试图将它转换为最后两条指令的关键段并行，但它使程序运行得慢得多。代码的并行如下：

void PointsToCentroides(int *Punts, int N, int D, int C, double *Centroides, int *PC, int *Sep){
int i, j, k;
for(i=0; i<C; i++)Sep[i]=0; 

#pragma omp parallel for shared(PC, Sep)
for(i=0; i<N; i++){
    int dist[C], m;
    for(j=0; j<C; j++){
        dist[j] = 0;
        for(k=0; k<D; k++){
            dist[j] += (Punts[i*D+k] - Centroides[j*D+k]) * (Punts[i*D+k] - Centroides[j*D+k]);
        }
    }
    m = min(dist, C);

    #pragma omp critical
    {
        PC[m*N+Sep[m]] = i;
        Sep[m]+=1;
    }

}}

我还尝试创建一个结构，因此每个线程都有自己的数据集来保存它使用的不同点，但是编译器给出了内部错误消息。代码如下：

void PointsToCentroides(int *Punts, int N, int D, int C, double *Centroides, int *PC, int *Sep){
int i;

for(i=0; i<C; i++){
    Sep[i] = 0;
}

struct centroids{
    int p[N];
    int count;
};

struct kt{
    struct centroids cent[C];
} ptc[omp_get_max_threads()];

#pragma omp parallel private(i)
{
    for(i=0; i<C; i++) ptc[omp_get_thread_num()].cent[i].count = 0;
}

#pragma omp parallel for //shared(PC, Sep)
for(i=0; i<N; i++){
    int dist[C], m, j, k;
    for(j=0; j<C; j++){
        dist[j] = 0;
        for(k=0; k<D; k++){
            dist[j] += (Punts[i*D+k] - Centroides[j*D+k]) * (Punts[i*D+k] - Centroides[j*D+k]);
        }
    }
    m = min(dist, C);
    //PC[m*N+Sep[m]] = i;
    //Sep[m]+=1;
    ptc[omp_get_thread_num()].cent[m].p[ptc[omp_get_thread_num()].cent[m].count]=i;
    ptc[omp_get_thread_num()].cent[m].count ++;
}

#pragma omp parallel private(i)
{
    int p = 0;
    for(i=0; i<C; i++){
        while(p<ptc[omp_get_thread_num()].cent[i].count){
            #pragma omp critical
            {
                PC[i*N+Sep[i]] = ptc[omp_get_thread_num()].cent[i].p[ptc[omp_get_thread_num()].cent[i].count];
                Sep[i]++;
            }
        }
    }
}}

编译器提供的错误消息如下：

user@pc:~/Descargas/HG$ gcc -O3 -g hg_k-means.c -fopenmp
hg_k-means.c: En la función ‘PointsToCentroides._omp_fn.0’:
hg_k-means.c:92:10: error interno del compilador: en emit_move_insn, en expr.c:3698
  #pragma omp parallel private(i)
          ^~~
0x75201a emit_move_insn(rtx_def*, rtx_def*)
        ../.././gcc/expr.c:3697
0x7409f3 extract_bit_field_1
        ../.././gcc/expmed.c:1748
0x7411f5 extract_bit_field(rtx_def*, unsigned long, unsigned long, int, rtx_def*, machine_mode, machine_mode, bool)
        ../.././gcc/expmed.c:1932
0x74e433 expand_expr_real_1(tree_node*, rtx_def*, machine_mode, expand_modifier, rtx_def**, bool)
        ../.././gcc/expr.c:10698
0x758f7a store_expr_with_bounds(tree_node*, rtx_def*, int, bool, bool, tree_node*)
        ../.././gcc/expr.c:5552
0x759bce expand_assignment(tree_node*, tree_node*, bool)
        ../.././gcc/expr.c:5321
0x66d1d0 expand_gimple_stmt_1
        ../.././gcc/cfgexpand.c:3639
0x66d1d0 expand_gimple_stmt
        ../.././gcc/cfgexpand.c:3737
0x66e79f expand_gimple_basic_block
        ../.././gcc/cfgexpand.c:5744
0x673906 execute
        ../.././gcc/cfgexpand.c:6357
Por favor, envíe un informe completo de errores,
con el código preprocesado si es apropiado.
Please include the complete backtrace with any bug report.
Véase <https://gcc.gnu.org/bugs/> para instrucciones.

我对OpenMP并行化方法有点新，所以欢迎任何帮助。

K-means openmp并行化内部编译器错误

0 个答案: