Test coding styles

a

Method 1:

1a Base code

for (size_t i = 0; i < n; i++)
{
    c1[i] = a[i] * b[i];
    for (size_t j = 0; j < 4096; j++)
        c1[i] *= (c1[i] + 1.0) * (a[i] + 1.0) * (b[i] + 1.0);
}

1b With OpenMP directives

#pragma omp parallel for simd for (size_t i = 0; i < n; i++) { c1[i] = a[i] * b[i]; #pragma omp simd for (size_t j = 0; j < nn; j++) c1[i] *= (c1[i] + 1.0) * (a[i] + 1.0) * (b[i] + 1.0); }

1c With AVX2

for (size_t i = 0; i < n; i += VEC_WIDTH)
{
    va = _mm256_loadu_ps(&a[i]);
    vb = _mm256_loadu_ps(&b[i]);
    vc = _mm256_mul_ps(va, vb);
    for (size_t j = 1; j < nn; j++)
    {
        v1 = _mm256_add_ps(vc, one_vec);
        v1 = _mm256_mul_ps(v1, _mm256_add_ps(va, one_vec));
        v1 = _mm256_mul_ps(v1, _mm256_add_ps(vb, one_vec));
        vc = _mm256_mul_ps(vc, v1);
    }
    _mm256_storeu_ps(&c2[i], vc);
}

1d With OpenCL

#define nn 4096
__kernel void method1(__global const float4 *a, __global const float4 *b,
                    __global float4 *c) {
size_t i = get_global_id(0);
float4 va = a[i];
float4 vb = b[i];
float4 vc = va * vb;

float4 one_vec = (float4)(1.0f);
for (size_t j = 0; j < nn; j += 4) {
    float4 v1 = vc + one_vec;
    v1 *= (va + one_vec) * (vb + one_vec);
    vc *= v1;

    v1 = vc + one_vec;
    v1 *= (va + one_vec) * (vb + one_vec);
    vc *= v1;

    v1 = vc + one_vec;
    v1 *= (va + one_vec) * (vb + one_vec);
    vc *= v1;

    v1 = vc + one_vec;
    v1 *= (va + one_vec) * (vb + one_vec);
    vc *= v1;
}

c[i] = vc;
}

Method 2:

2a Base code

for (size_t i = 0; i < n; i++)
    c2[i] = a[i] * b[i];
for (size_t j = 0; j < nn; j++)
    for (size_t i = 0; i < n; i++)
        c2[i] *= (c2[i] + 1.0) * (a[i] + 1.0) * (b[i] + 1.0);

2b With OpenMP directives

#pragma omp parallel for
for (size_t i = 0; i < n; i++)
    c2[i] = a[i] * b[i];
#pragma omp barrier
for (size_t j = 0; j < 4096; j++)
#pragma omp parallel for
    for (size_t i = 0; i < n; i++)
        c2[i] *= (c2[i] + 1.0) * (a[i] + 1.0) * (b[i] + 1.0);
#pragma omp barrier

2c With AVX2

2d With OpenCL

__kernel void multiply_arrays(__global float *a, __global float *b,__global float *c) {
  size_t i = get_global_id(0);
  c[i] = a[i] * b[i];
  for (size_t j = 0; j < 4096; ++j)
    c[i] *= (c[i] + 1.0) * (a[i] + 1.0) * (b[i] + 1.0); 
}

Method 3:

3a Base code

for (size_t j = 0; j < nn; j++)
{
    for (size_t i = 0; i < n; i++)
    {
        temp = a[i] * b[i];
        temp *= (temp + 1.0) * (a[i] + 1.0) * (b[i] + 1.0);
        c2[i] += temp;
    }
}

3b With OpenMP directives

3c With AVX2

3d With OpenCL

results for n=1024*1024;