gcc自动矢量化(未处理的数据参考)

Question

我不明白为什么这样的代码没有用gcc 4.4.6进行矢量化

int MyFunc(const float *pfTab, float *pfResult, int iSize, int iIndex)
{
  for (int i = 0; i < iSize; i++)
     pfResult[i] = pfResult[i] + pfTab[iIndex];
}

 note: not vectorized: unhandled data-ref

但是,如果我写下面的代码

   int MyFunc(const float *pfTab, float *pfResult, int iSize, int iIndex)
{
  float fTab =  pfTab[iIndex];
  for (int i = 0; i < iSize; i++)
     pfResult[i] = pfResult[i] + fTab;
}

gcc成功自动向量化此循环

如果我添加omp指令

   int MyFunc(const float *pfTab, float *pfResult, int iSize, int iIndex)
{
  float fTab =  pfTab[iIndex];
  #pragma omp parallel for
  for (int i = 0; i < iSize; i++)
     pfResult[i] = pfResult[i] + fTab;
}

我有以下错误没有矢量化:未处理的数据 - 参考

你能帮我解释为什么第一个代码和第三个代码没有自动矢量化吗？

第二个问题:数学操作数似乎没有矢量化(exp,log等等),例如这个代码

for (int i = 0; i < iSize; i++)
         pfResult[i] = exp(pfResult[i]);

没有矢量化.这是由于我的gcc版本？

编辑:新版本的gcc 4.8.1和openMP 2011(echo | cpp -fopenmp -dM | grep -i open)我对所有类型的循环都有以下错误,甚至基本上

   for (iGID = 0; iGID < iSize; iGID++)
        {
             pfResult[iGID] = fValue;
        }


note: not consecutive access *_144 = 5.0e-1;
note: Failed to SLP the basic block.
note: not vectorized: failed to find SLP opportunities in basic block.

EDIT2:

#include<stdio.h>
#include<sys/time.h>
#include <string.h>
#include <math.h>
#include <stdlib.h>
#include <omp.h>

int main()
{
        int szGlobalWorkSize = 131072;
        int iGID = 0;
        int j = 0;
        omp_set_dynamic(0);
        // warmup
        #if WARMUP
        #pragma omp parallel
        {
        #pragma omp master
        {
        printf("%d threads\n", omp_get_num_threads());
        }
        }
        #endif
        printf("Pagesize=%d\n", getpagesize());
        float *pfResult = (float *)malloc(szGlobalWorkSize * 100* sizeof(float));
        float fValue = 0.5f;
        struct timeval tim;
        gettimeofday(&tim, NULL);
        double tLaunch1=tim.tv_sec+(tim.tv_usec/1000000.0);
        double time = omp_get_wtime();
        int iChunk = getpagesize();
        int iSize = ((int)szGlobalWorkSize * 100) / iChunk;
        //#pragma omp parallel for
        for (iGID = 0; iGID < iSize; iGID++)
        {
             pfResult[iGID] = fValue;
        }
        time = omp_get_wtime() - time;
        gettimeofday(&tim, NULL);
        double tLaunch2=tim.tv_sec+(tim.tv_usec/1000000.0);
        printf("%.6lf Time1\n", tLaunch2-tLaunch1);
        printf("%.6lf Time2\n", time);
}

结果

#define _OPENMP 201107
gcc (GCC) 4.8.2 20140120 (Red Hat 4.8.2-15)

gcc -march=native -fopenmp -O3 -ftree-vectorizer-verbose=2 test.c -lm

大量的

note: Failed to SLP the basic block.
note: not vectorized: failed to find SLP opportunities in basic block.
and note: not consecutive access *_144 = 5.0e-1;

谢谢

Answer 1

GCC无法向你的循环的第一个版本进行矢量化,因为它无法证明它pfTab[iIndex]不包含在跨越的内存中pfResult[0] ... pfResult[iSize-1](指针别名).实际上,如果pfTab[iIndex]在某个内存中,那么它的值必须被循环体中的赋值覆盖,并且必须在迭代中使用新值.你应该使用restrict关键字提示编译器,这可能永远不会发生,然后它应该愉快地矢量化你的代码:

$ cat foo.c
int MyFunc(const float *restrict pfTab, float *restrict pfResult,
           int iSize, int iIndex)
{
   for (int i = 0; i < iSize; i++)
     pfResult[i] = pfResult[i] + pfTab[iIndex];
}
$ gcc -v
...
gcc version 4.6.1 (GCC)
$ gcc -std=c99 -O3 -march=native -ftree-vectorizer-verbose=2 -c foo.c
foo.c:3: note: LOOP VECTORIZED.
foo.c:1: note: vectorized 1 loops in function.

第二个版本会向量化,因为值会传输到具有自动存储持续时间的变量.这里的一般假设是pfResult不跨越存储的堆栈存储器fTab(通过C99语言规范粗略读取并不清楚如果该假设很弱或标准中的某些内容允许它).

由于OpenMP在GCC中的实现方式,OpenMP版本无法进行矢量化.它使用并行区域的代码大纲.

int MyFunc(const float *pfTab, float *pfResult, int iSize, int iIndex)
{
  float fTab =  pfTab[iIndex];
  #pragma omp parallel for
  for (int i = 0; i < iSize; i++)
     pfResult[i] = pfResult[i] + fTab;
}

有效地变成:

struct omp_data_s
{
  float *pfResult;
  int iSize;
  float *fTab;
};

int MyFunc(const float *pfTab, float *pfResult, int iSize, int iIndex)
{
  float fTab =  pfTab[iIndex];
  struct omp_data_s omp_data_o;

  omp_data_o.pfResult = pfResult;
  omp_data_o.iSize = iSize;
  omp_data_o.fTab = fTab;

  GOMP_parallel_start (MyFunc_omp_fn0, &omp_data_o, 0);
  MyFunc._omp_fn.0 (&omp_data_o);
  GOMP_parallel_end ();
  pfResult = omp_data_o.pfResult;
  iSize = omp_data_o.iSize;
  fTab = omp_data_o.fTab;
}

void MyFunc_omp_fn0 (struct omp_data_s *omp_data_i)
{
  int start = ...; // compute starting iteration for current thread
  int end = ...; // compute ending iteration for current thread

  for (int i = start; i < end; i++)
    omp_data_i->pfResult[i] = omp_data_i->pfResult[i] + omp_data_i->fTab;
}

MyFunc_omp_fn0包含概述的功能代码.编译器无法证明omp_data_i->pfResult不指向别名的内存,omp_data_i特别是其成员fTab.

为了矢量化该循环,你必须做fTab firstprivate.这将在概述的代码中将其转换为自动变量,这将等同于您的第二种情况:

$ cat foo.c
int MyFunc(const float *pfTab, float *pfResult, int iSize, int iIndex)
{
   float fTab = pfTab[iIndex];
   #pragma omp parallel for firstprivate(fTab)
   for (int i = 0; i < iSize; i++)
     pfResult[i] = pfResult[i] + fTab;
}
$ gcc -std=c99 -fopenmp -O3 -march=native -ftree-vectorizer-verbose=2 -c foo.c
foo.c:6: note: LOOP VECTORIZED.
foo.c:4: note: vectorized 1 loops in function.