I have played around with SSE. I am using VS2012, and I wanted to know what was the fastest way to calculate the length of a vector.
The code looks very bad.

float Magnitude() const
{
#if SSE && SSE_ASM
	float result;
	//Optimized magnitude calculation with SSE and Assembly
	__asm
	{
	MOV EAX, this								//Move [this] to EAX.
	MOVAPS XMM2, [EAX]							//Copy data EAX to XMM2 register
	MULPS  XMM2, XMM2                            //Square the XMM2 register.
	MOVAPS XMM1, XMM2                            //Make a copy
	SHUFPS XMM2, XMM1, _MM_SHUFFLE(1, 0, 3, 2)   //Shuffle so that we can add together the elements.
	ADDPS  XMM2, XMM1			                //Add the elements.
	MOVAPS XMM1, XMM2			                //Make a copy
	SHUFPS XMM1, XMM1, _MM_SHUFFLE(0, 1, 0, 1)   //Second addition of elements using shuffle
	ADDPS  XMM2, XMM1
	SQRTPS XMM2, XMM2			                //Get the square root
	MOVSS [result], XMM2                         //Store the result in the float.
	}
	return result;
#elif SSE
	__m128 tmp = _mm_mul_ps(components, components); 
	tmp = _mm_add_ps(_mm_shuffle_ps(tmp, tmp, _MM_SHUFFLE(1, 0, 3, 2)), tmp);
	tmp = _mm_sqrt_ps(_mm_add_ps(tmp, _mm_shuffle_ps(tmp, tmp, _MM_SHUFFLE(0, 1, 0, 1))));
		
	float result;
	_mm_store_ss(&result, tmp);
		
	return result;
#endif

#if !SSE && !SSE_ASM
        return sqrtf(__x * __x + __y * __y + __z * __z + __w * __w);
#endif
}

Guess what, the normal method was as fast as the SSE-Intrinsics, while my assembly code was acctually slower.
So yeah, you can't beat the compiler :)

You lose some performance by moving value from SSE register to normal register.

If you're fine with using SSE4 there's instruction to calculate dot product right away: http://msdn.microsoft.com/en-us/library/bb514054(v=vs.100).aspx. This way you won't need shuffles and such.

If you're alright with restricting yourself to SSE3+, you can use the _mm_hadd_ps intrinsic to save several steps in your code.

I find that SIMD is not very useful for your standard coordinate vectors (i.e. position), but are far more useful when you are dealing with many independent parallel operations (which vector magnitude is not, hence the horizontal add) and can manage to store your data aligned and in a SIMD-friendly format. For instance, you could write a much more efficient function to compute the magnitude of 4 different vectors at once, returning a 128 bit value containing the result of all 4, no shuffling required.

For this reason, my math library contains a specialized SIMDScalar<Type,N> template class that represents an N-wide SIMD value of the given type (so that you can support doubles, ints, or whatever else). I then have a SIMDVector3<Type,N> class which implements an N-wide standard 3D coordinate vector using a SIMDScalar for each of the x, y, and z components. You can then use a SIMDVector3 to perform parallel operations on many 3D vectors at once, stored in structure-of-arrays format. The advantage of doing it this way is that you avoid all of the packing/unpacking that would be necessary (and slow) if you were using an __m128 to make a 4-component coordinate vector class as you are trying to do.

You shouldn't use SSE to calculate the length of a vector, you should use it to calculate the length of 4 vectors at a time.

You shouldn't use SSE to calculate the length of a vector, you should use it to calculate the length of 4 vectors at a time.

That is generally the correct approach for SIMD instructions.

The best performance improvements do not come from doing a single computation in one part, but doing a series of many operations as a block. Better for the cache, better for the predictive engines, better for the OOO core, less overhead from function calls.

Here the better approach would be a function that takes four __m128 values, squares and sums them, then returns the square root of all four in an __m128. All of it done with intrinsics, no asm statements. Maybe also have another function signature that takes four arrays.

If you're only computing a single vector's magnitude the SIMD approach is probably not worth the effort. The optimizer will probably recognize it and do something smart, it isn't worth attempting to hand-optimize.

In what way? Obviously there'd be the extra instruction to move the value out of the register, but what other performance do you lose? It probably will create a dependency, but OTOH if you used the original value there'd be a dependency anyway. Maybe I'm missing something?

You will trigger a very, very costly Load-Hit-Store, which is especially expensive on consoles, but is also far from trivial on desktop CPUs. The same happens for conversion from int to float too, and to put it short in case of a TL;DR; from the article: You mess up your CPUs pipelining by doing so, which is way worse than an additional move-instruction.