c - 信息 C5012 : loop not parallelized due to reason ‘1007’

Question

我正在 x86_64 上试用 Visual Studio 2013 的自动矢量化器模式，我对以下内容感到有些惊讶。考虑原始代码:

static void rescale( double * __restrict out, unsigned short * __restrict in, size_t n, const double intercept, const double slope )
{
    for( size_t i = 0; i < n; ++i )
        out[i] = slope * in[i] + intercept;
}

Visual Studio 返回它在这样一个天真的例子中失败了:

--- Analyzing function: rescale
c:\users\malat\autovec\vec.c(18) : info C5012: loop not parallelized due to reason '1007'

编译行在哪里(我现在只对SSE2感兴趣):

cl vec.c /O2 /Qpar /Qpar-report:2

查看文档:

• https://msdn.microsoft.com/en-us/library/hh872235(v=vs.120).aspx

导致:

• https://msdn.microsoft.com/en-us/library/jj658585(v=vs.120).aspx#BKMK_ReasonCode100x

内容如下:

The loop induction variable or the loop bounds are not signed 32-bit numbers (int or long). Resolve this by changing the type of the induction variable.

有没有办法重写这个循环，以便正确触发自动矢量化器模式？

我未能使用简单的方法重写代码:

static void rescale( double * __restrict out, unsigned short * __restrict in, size_t n, const double intercept, const double slope )
{
  const long first = (long)n;
  const long secnd = n > LONG_MAX ? n - LONG_MAX : 0;
  for( long i = 0; i < first; ++i )
    out[i] = slope * in[i] + intercept;
  for( long i = 0; i < secnd; ++i )
    out[LONG_MAX+i] = slope * in[LONG_MAX+i] + intercept;
}

在上述情况下，Visual Studio 现在报告:

--- Analyzing function: rescale
c:\users\malat\autovec\vec.c(21) : info C5012: loop not parallelized due to reason '1000'
c:\users\malat\autovec\vec.c(23) : info C5012: loop not parallelized due to reason '1000'

这意味着:

The compiler detected a data dependency in the loop body.

在我的第二种情况下，我看不出哪里会存在数据依赖性。

我应该如何重写我的初始代码以满足 Visual Studio 2013 的自动矢量化器模式？

Answer 1

SSE2 及其前身 SSE 都没有将 uint16_t-s 转换为 double-s 的正确指令。

将 in 转换为 double*。