Vectorizing a loop

A compiler that automatically merges scalar data into a parallel-packed SIMD data structure is called an auto-vectorizing compiler. Such compilers must handle all the high-level language constructs, and therefore do not always produce optimal code.

A simple example of vectorizing a loop is shown below. The original loop multiplies two arrays, term by term. The arrays are assumed to remain scalar outside of the subroutine vmult.

/* Scalar version */
int mult(float *array1, float *array2, float *out, int arraySize) {
	int i;
	for (i = 0; i < arraySize; i++) {
		out[i] = array1[i] * array2[i];
	}
	return 0;
}

/* Vectorized version */
int vmult(float *array1, float *array2, float *out, int arraySize) {
	/* This code assumes that the arrays are quadword-aligned. */
	/* This code assumes that the arraySize is divisible by 4. */

	int i, arraySizebyfour;
	arraySizebyfour = arraySize >> 2; 			 /* arraySize/4 vectors */
	vector float *varray1 = (vector float *) (array1);
	vector float *varray2 = (vector float *) (array2);
	vector float *vout = (vector float *) (out);

	for (i = 0; i < arraySizebyfour; i++) {
		/*spu_mul is an intrinsic that multiplies vectors */
		vout[i] = spu_mul(varray1[i], varray2[i]);		 
	}

	return 0;
}