// ***************** prog.c ******************************************
//                 Dsp program sample calculates a*x+y
//
// (c) multicore.ru
//
// *******************************************************************
typedef __attribute__((__vector_size__(4 *sizeof(float)))) float _v4f32;

float x[200];
float y[200];
float z0[200];
float z1[200];

//
// @description: initialize input vectors with some values
//
void SetupVectors(float *in1, float* in2, int length)
{
	int i, j;
	for (i = 0; i < length; i++)
	{
		in1[i]= i * 2.3 + 1;
	}
	for (j = length; j >= 0; j--)
	{
		in2[j]= (length-j) * 1.6;
	}
}

//
// @description: calculate const*vector1+vector2
//
void Func0(float *in1, float* in2, float k, float* out, int length)
{
	for (int i = 0; i < length; i++)
	{
		out[i]= k * in1[i] + in2[i];
	}
}

//
// @description: calculate const*vector1+vector2 by using elcore builtin
//				 functions
//
void Func1(float *in1, float* in2, float k, float* out, int length)
{
	_v4f32 *x1 = (_v4f32*)in1;
	_v4f32 *y1 = (_v4f32*)in2;
	_v4f32 *z1 = (_v4f32*)out;

	for (int i = 0; i < length/4; i++)
	{
		_v4f32 mul =__builtin_cv4f32_mul(k, x1[i]);
		z1[i] =__builtin_v4f32_add(mul, y1[i]);
	}
}

int main()
{
	SetupVectors(x, y, 200);
	Func0(x, y, 4.0f, z0, 200);
	Func1(x, y, 4.0f, z1, 200);
	return 100;
}