VC11: Auto-Vectorizer and concurrency::graphics

Question

Hi,

The following functions do not seem to auto-vectorize (even under /fp:fast) on an x64 target:

#include "stdafx.h"

#include <amp_graphics.h>

using namespace concurrency::graphics;

float_4 myfun(float_4 a, float_4 b)
{
	return a + b;
}

float_4 myfunref(const float_4& a, const float_4& b)
{
	return a + b;
}

The generated code is basically a sequence of movss and addss, instead of movups and addps instructions.

Should auto-vectorization work in that case ? Is auto-vectorization a loop-only feature ? Am I missing something ?

Best regards, Arnaud.

Answer 1

Hi Arnaud,

Auto-vectorization works only on loops.  That's to say, it looks at operations that would execute in different iterations of a loop, and joins them together (safely) into a single operation.

It is possible to vectorize your examples too (called "statement vectorization"), but the auto-vectorizer doesn't tackle that problem (in this release).

Thanks,

Jim

---

Jim