Auto-vectorization of loop containing comparisons

问题

I'm trying to use the Visual C++ 2013 auto-vectorizer to make the following loop vectorized (/arch:AVX2) but the compiler refuses and gives the following message:

info C5002: loop not vectorized due to reason '1100'

This reason code means

Loop contains control flow—for example, "if" or "?".

I have tried to split the comparisons and the final assignment into a separate loop but that seems inefficient when there are intrinsics available for performing comparisons on floating point values.

Why should the compiler treat comparisons as flow control, and what can I change in the implementation so that the compiler will vectorize this function?

void triplets_positive(
    const std::uint64_t count,
    double * const a,
    double * const b,
    double * const c,
    std::uint64_t * const all_positive)
{
    for (std::uint64_t i = 0; i < count; ++i)
    {
        // These >= operations make the loop not vectorisable because
        // they introduce control flow.
        std::uint64_t a_pos = (a[i] >= 0.0);
        std::uint64_t b_pos = (b[i] >= 0.0);
        std::uint64_t c_pos = (c[i] >= 0.0);

        all_positive[i] = a_pos & b_pos & c_pos;
    }
}

回答1:

Unfortunately, this appears to be either a bug or limitation in the Visual C++ 2013 compiler. Other compilers make use either of the CMPPD instruction (AVX/AVX2) or CMP*PD instructions (SSE2).

Compilers that successfully vectorise this loop include:

Visual C++ 2017
Visual C++ 2015
Clang + LLVM (Apple LLVM version 8.1.0 (clang-802.0.42))

While it's theoretically possible to write the comparison as bitwise operations, that's counterproductive and the best option is to upgrade to another compiler.

来源：https://stackoverflow.com/questions/45339998/auto-vectorization-of-loop-containing-comparisons

标签

c++

visual-c++

vectorization