OpenMP atomic _mm_add_pd

问题:

I'm trying to use OpenMP for parallelization of an already vectorized code with intrinsics, but the problem is that I'm using one XMM register as an outside 'variable' that I increment each loop. For now I'm using the shared clause

__m128d xmm0 = _mm_setzero_pd();
__declspec(align(16)) double res[2];

#pragma omp parallel for shared(xmm0)
for (int i = 0; i < len; i++)
{
    __m128d xmm7 = ... result of some operations

    xmm0 = _mm_add_pd(xmm0, xmm7);
}

_mm_store_pd(res, xmm0);
double final_result = res[0] + res[1];

because the atomic operation is not supported (in VS2010)

__m128d xmm0 = _mm_setzero_pd();
__declspec(align(16)) double res[2];

#pragma omp parallel for
for (int i = 0; i < len; i++)
{
    __m128d xmm7 = ... result of some operations

    #pragma omp atomic
    xmm0 = _mm_add_pd(xmm0, xmm7);
}

_mm_store_pd(res, xmm0);
double final_result = res[0] + res[1];

Does anyone know a clever work-around?

---

EDIT: I've also tried it using the Parallel Patterns Library just now:

__declspec(align(16)) double res[2];
combinable<__m128d> xmm0_comb([](){return _mm_setzero_pd();});

parallel_for(0, len, 1, [&xmm0_comb, ...](int i)
{
    __m128d xmm7 = ... result of some operations

    __m128d& xmm0 = xmm0_comb.local();
    xmm0 = _mm_add_pd(xmm0, xmm7);
});

__m128d xmm0 = xmm0_comb.combine([](__m128d a, __m128d b){return _mm_add_pd(a, b);});
_mm_store_pd(res, xmm0);
double final_result = res[0] + res[1];

but the result was disappointing.

回答1:

You're solving the problem the wrong way. You should be using a reduction instead of atomic operations:

This is a better approach:

double sum = 0;

#pragma omp parallel for reduction(+:sum)
for (int i = 0; i < len; i++)
{
    __m128d xmm7;// = ... result of some operations

    //  Collapse to a "double".
    _declspec(align(16)) double res[2];
    _mm_store_pd(res, xmm7);

    //  Add to reduction variable.
    sum += res[0] + res[1];
}

double final_result = sum;

A reduction is essentially an operation that collapses "reduces" everything to a single variable using an associative operation such as +.

If you're doing a reduction, always try to use an actual reduction approach. Don't try to cheat it with atomic operations or critical sections.

The reason for this is that atomic/critical section approaches are inherently not scalable as they maintain a long critical path data-dependency. A proper reduction approach reduces this critical path to log(# of threads).

The only downside of course is that it breaks floating-point associativity. If that matters, then you're basically stuck with sequentially summing up each iteration.

回答2:

What you're looking for is a reduction. You can do that as an omp reduction if your compiler supports it (gcc does), or you can roll one yourself by summing into a private xmm for each thread. Below is a simple version doing both:

#include <emmintrin.h>
#include <omp.h>
#include <stdio.h>


int main(int argc, char **argv) {

    const int NTHREADS=8;
    const int len=100;

    __m128d xmm0[NTHREADS];
    __m128d xmmreduction = _mm_setzero_pd();
    #pragma omp parallel for num_threads(NTHREADS)
    for (int i=0; i<NTHREADS; i++)
        xmm0[i]= _mm_setzero_pd();

    __attribute((aligned(16))) double res[2];

    #pragma omp parallel num_threads(NTHREADS) reduction(+:xmmreduction)
    {
        int tid = omp_get_thread_num();
        #pragma omp for
        for (int i = 0; i < len; i++)
        {
            double d = (double)i;
            __m128d xmm7 = _mm_set_pd( d, 2.*d );

            xmm0[tid] = _mm_add_pd(xmm0[tid], xmm7);
            xmmreduction = _mm_add_pd(xmmreduction, xmm7);
        }
    }

    for (int i=1; i<NTHREADS; i++)
        xmm0[0] = _mm_add_pd(xmm0[0], xmm0[i]);

    _mm_store_pd(res, xmm0[0]);
    double final_result = res[0] + res[1];

    printf("Expected result   = %f\n", 3.0*(len-1)*(len)/2);
    printf("Calculated result = %lf\n", final_result);

    _mm_store_pd(res, xmmreduction);
    final_result = res[0] + res[1];

    printf("Calculated result (reduction) = %lf\n", final_result);

    return 0;
}

回答3:

With great help from the people who answered my question I've come up with this:

double final_result = 0.0;

#pragma omp parallel reduction(+:final_result)
{
    __declspec(align(16)) double r[2];
    __m128d xmm0 = _mm_setzero_pd();

    #pragma omp for
    for (int i = 0; i < len; i++)
    {
        __m128d xmm7 = ... result of some operations

        xmm0 = _mm_add_pd(xmm0, xmm7);
    }
    _mm_store_pd(r, xmm0);
    final_result += r[0] + r[1];
}

It basically separates the collapse and reduction, performs very well.

Many thanks to all who have helped me!

回答4:

I guess you can't add your own intrinsics to the compiler, and MS compilers decided to skip inline assembler. Not sure there is an easy solution at all.