Michael S (already5chosen.delete@this.yahoo.com) on August 4, 2018 2:33 pm wrote:
> Travis (travis.downs.delete@this.gmail.com) on August 4, 2018 1:50 pm wrote:
> > I forgot one big advantage of auto-vectorization: it is way more future-proof than
> > assembly or intrinsics. As new ISAs become available, compilers will just be able to
> > use them on your original source, so things should only get faster over time.
> >
>
> The alternative to auto-vectorization that is discussed here is not asm/intrinsics/OS threads,
> but tools like CUDA, ispc, OpenCL etc... The later two more
> suitable for CPUs, but unfortunately pretty close to dead.
> I would also include more explicitly parallel tools, like operations
> on arrays in Modern Fortran, Matlab/Octave, Julia.
> May be, the next version of C++ will include something of either first or second sort.


Definitely - this is only an advantage with respect to intrinsics/assembly. I meant to write that many of the other competitors have this same advantage. I didn't know the thread only allowed discussion/comparison with respect to the those other alternatives, sorry.

BTW, I think CUDA and auto-vectorization are rarely competitors on a practical level, even if they are at some kind of macro level.

Using CUDA means you are probably using some very different and specific hardware. So the decision either goes (A) "what is the best hardware for us to run this problem on" and then you choose the implementation language based on that (so CUDA if you chose GPUs or whatever or (B) "what is the hardware we already have/already rent" at which point you again choose the implementation based on what that is.

So yeah, at a high level you have CUDA and GPUs/other GPU-like CPU struggling against SIMD options like x86 chips, but at an individual or small team decision-making level, you the hardware is already approximately fixed.

That's x86 SIMDs biggest advantage, I think: it's already there on every EC2 instance, on every private cloud, on every developer's PC and to access it is a smooth ramp from turning on auto-vectorization, to using intrinsics (or OpenCL or whatever) or assembly. You don't may more for it, you don't have convince operations and purchasing to give you a rack of specialized hardware, you don't have to have considered CUDA when you built your dev box, you don't have to use specialized instances on your cloud provider: it's just "there", available, built-in to your CPU, a couple of cycles away, in coherent memory and on every core and without any cross-processing sharing issues.

If it wasn't for that, you'd have expect GPU/APU to have already crushed CPU/SIMD for many more use cases than it already has.