Question

I'm writing a SSE code to 2D convolution but SSE documentation is very sparse. I'm calculating dot product with _mm_dp_ps and using _mm_extract_ps to get the dot product result, but _mm_extract_ps returns a hex that represents a float and I can't figure out how to convert this hex float to a regular float. I could use __builtin_ia32_vec_ext_v4sf that returns a float but I wanna keep compatibility with others compilers.

_mm_extract_ps (__m128 __X, const int __N)
{
  union { int i; float f; } __tmp;
  __tmp.f = __builtin_ia32_vec_ext_v4sf ((__v4sf)__X, __N);
  return __tmp.i;
}

What point I'm missing?

A little help will be appreciated, thanks.

OpenSUSE 11.2
GCC 4.4.1
C++
Compiler options
-fopenmp -Wall -O3 -msse4.1 -march=core2
Linker options
-lgomp -Wall -O3 -msse4.1 -march=core2

Was it helpful?

Solution

You should be able to use _MM_EXTRACT_FLOAT.

Incidentally it looks to me as if _mm_extract_ps and _MM_EXTRACT_FLOAT should be the other way around, i.e. _mm_extract_ps should return a float and _MM_EXTRACT_FLOAT should return the int representation, but what do I know.

OTHER TIPS

_mm_cvtss_f32(_mm_shuffle_ps(__X, __X, __N)) will do the job.

And just to exemplify all that has been mentioned so far:

main.c

#include <assert.h>

#include <x86intrin.h>

int main(void) {

    /* 32-bit. */
    {
        __m128 x = _mm_set_ps(1.5f, 2.5f, 3.5f, 4.5f);

        /* _MM_EXTRACT_FLOAT */
        float f;
        _MM_EXTRACT_FLOAT(f, x, 3);
        assert(f == 1.5f);
        _MM_EXTRACT_FLOAT(f, x, 2);
        assert(f == 2.5f);
        _MM_EXTRACT_FLOAT(f, x, 1);
        assert(f == 3.5f);
        _MM_EXTRACT_FLOAT(f, x, 0);
        assert(f == 4.5f);

        /* _mm_cvtss_f32 + _mm_shuffle_ps */
        assert(_mm_cvtss_f32(x) == 4.5f);
        assert(_mm_cvtss_f32(_mm_shuffle_ps(x, x, 1)) == 3.5f);
        assert(_mm_cvtss_f32(_mm_shuffle_ps(x, x, 2)) == 2.5f);
        assert(_mm_cvtss_f32(_mm_shuffle_ps(x, x, 3)) == 1.5f);
    }

    /* 64-bit. */
    {
        __m128d x = _mm_set_pd(1.5, 2.5);
        /* _mm_cvtsd_f64 + _mm_unpackhi_pd */
        assert(_mm_cvtsd_f64(x) == 2.5);
        assert(_mm_cvtsd_f64(_mm_unpackhi_pd(x, x)) == 1.5);
    }
}

GitHub upstream.

Compile and run:

gcc -ggdb3 -O0 -std=c99 -Wall -Wextra -pedantic -o main.out main.c
./main.out

Doubles mentioned at: _mm_cvtsd_f64 analogon for higher order floating point

Tested on Ubuntu 19.04 amd64.

extern void _mm_store_ss(float*, __m128);

See 'xmmintrin.h.'

Licensed under: CC-BY-SA with attribution
Not affiliated with StackOverflow
scroll top