Using __m256d registers

bobobobo picture bobobobo · Oct 13, 2012 · Viewed 12.5k times · Source

How do you use __m256d?

Say I want to use the Intel AVX instruction _mm256_add_pd on a simple Vector3 class with 3-64 bit double precision components (x, y, and z). What is the correct way to use this?

Since x, y and z are members of the Vector3 class, _can I declare them in union with an __m256d variable?

union Vector3
{
  struct { double x,y,z ; } ;
  __m256d _register ;  // the Intel register?
} ;

Then can I go:

Vector3 add( const Vector3& o )
{
  Vector3 result;
  result._register = _mm256_add_pd( _register, o._register ) ; // add 'em
  return result; 
}

Is that going to work? Or do I need to declare temporaries,

Vector3 add( const Vector3& o )
{
  __m256d d1 = *(__m256d*)(&x) ; // ? Cast to __m256d?
  __m256d d2 = *(__m256d*)(&o.x) ; // ? Cast to __m256d?
  __m256d result = _mm256_add_pd( d1, d2 ) ; // add 'em
  return Vector3( result ) ; // make a ctor that accepts __m256d?
}

Edit

I came up with this example,

#include <stdio.h>
#include <intrin.h>

int main()
{
  __m256d a, b, res;

  for( int i = 0; i < sizeof(__m256d)/sizeof(double); i++ )
  {
    a.m256d_f64[i] = i ;
    b.m256d_f64[i] = 2*i ;
  }

  // Perform __4__ adds.
  res = _mm256_add_pd(a, b);

  for( int i = 0; i < sizeof(__m256d)/sizeof(double); i++ )
  {
    printf("%f + %f = %f\n", a.m256d_f64[i], b.m256d_f64[i], res.m256d_f64[i]);
  }
  puts("");
}

I guess the question is now, does _mm256_add_pd do load operations automatically, or will something get messed up if I don't declare my __m256d registers as locals close to where they are used? (I'm afraid of a hotel room / deskdrawer type problem)

Edit 2:

I tried adding an __m256 register to my rather large project, and I got a whole bunch of

error C2719: 'value': formal parameter with __declspec(align('32')) won't be aligned

Errors, it leads me to believe that you can't keep __m256 registers inside a class, instead they should be declared as locals?

Answer

Dietrich Epp picture Dietrich Epp · Oct 14, 2012

First I'd like to clear up a little confusion. __m256d isn't a type of register, it's a data type that can be loaded into an AVX register. A __m256d is no more a register than an int is a register. There are a few ways to get data in and out of an __m256d (or any other vector type):

Using a union: Yes, the union trick works. It works very well, since the union will generally have the correct alignment (although malloc might not, use posix_memalign or _aligned_malloc).

class Vector3 {
public:
    Vector3(double xx, double yy, double zz);
    Vector3(__m256d vvec);


    Vector3 operator+(const Vector3 &other) const
    {
        return Vector3(_mm256_add_pd(vec, other.vec));
    }

    union {
        struct {
            double x, y, z;
        };
        __m256d vec; // a data field, maybe a register, maybe not
    };
};

Using intrinsics: Inside a function, it's usually easier to use intrinsics to get data in and out of a vector type.

__m256d vec = ...;
double x, y, z;
vec = _mm256_add_pd(vec, _mm256_set_pd(x, y, z, 0.0));

Using pointer casts: Casting pointers is the last resort for a couple of reasons.

  1. The pointer might not be aligned correctly.

  2. Casting pointers can sometimes mess with the compiler's aliasing analysis.

  3. Pointer casting bypasses a number of safety guarantees.

So I'd only use pointer casting to plow through a big array of data.