Simple batch normalization example

example_bnrm_fwd.cpp

#include <stdio.h>
#include <math.h>

#include "mkldnn.hpp"

using namespace mkldnn;

void init_data(float *dat, int size) {
    for (int i = 0; i < size; ++i)
        dat[i] = 1.f + 2.f * sinf(0.2 * i);
}

void exec_bnrm_fwd() {
    engine eng(engine::cpu, 0);

    const int N = 4, C = 4, H = 27, W = 27;
    const int sz = N * C * H * W;

    float *src = (float *)malloc(sizeof(float) * sz);
    init_data(src, sz); // ideally mean ~= 1.f, var ~= 2.f

    float *dst = (float *)malloc(sizeof(float) * sz);
    float *mean = (float *)malloc(sizeof(float) * C);
    float *var = (float *)malloc(sizeof(float) * C);

    memory::desc data_desc({N, C, H, W}, memory::f32, memory::nchw);
    memory src_mem({data_desc, eng}, src);
    memory dst_mem({data_desc, eng}, dst);

    memory::desc stat_desc({C}, memory::f32, memory::x);
    memory mean_mem({stat_desc, eng}, mean);
    memory var_mem({stat_desc, eng}, var);

    unsigned flags = 0;
    // set flags for different flavors (use | to combine flags)
    //      use_global_stats -- do not compute mean and variance in the primitive, user has to provide them
    //      use_scale_shift  -- in addition to batch norm also scale and shift the result

    batch_normalization_forward::desc bnrm_fwd_d(
            prop_kind::forward_training, // might be forward_inference, backward, backward_data
            data_desc,  // data descriptor (i.e. sizes, data type, and layout)
            0.001f,     // eps
            flags);
    batch_normalization_forward::primitive_desc bnrm_fwd_pd(bnrm_fwd_d, eng);

    batch_normalization_forward bnrm_fwd(bnrm_fwd_pd,
            src_mem, dst_mem, mean_mem, var_mem);
    stream(stream::kind::eager).submit({bnrm_fwd}).wait(); // execute bnrm

    for (int c = 0; c < C; ++c) {
        printf("[%d] mean:%f var:%f\n", c, mean[c], var[c]);
    }
}

int main() {
    exec_bnrm_fwd();
    return 0;
}

// $ ( g++ example_bnrm_fwd.cpp -lmkldnn -lm && ./a.out )
// [0] mean:0.994635 var:1.997512
// [1] mean:1.000734 var:2.003489
// [2] mean:1.005777 var:1.996543
// [3] mean:1.002504 var:2.002276