test.c

void bm64_baseisa(const uint64_t in[64], uint64_t out[64])
{
	for (int i2 = 0; i2 < 64; i2 += 8) // in[] rows, word index
        for (int j = 0; j < 8; j++) // in[] columns, bit index
        {
            int j_shift = j*8;
            uint64_t res = 0;

            for (int i=0; i<8; i++)
                res |= ((in[i2+i] >> j_shift) & 0xFF) << (i*8);

            out[j + i2] = res;
        }
}

void bm64_baseisa2(const uint64_t in[64], uint64_t out[64])
{
	for (int i2 = 0; i2 < 64; i2 += 8) // in[] rows, word index
        for (int j = 0; j < 8; j++) // in[] columns, bit index
        {
            int j_shift = j*8;
            uint64_t res = 0;
            
            res |= ((in[i2+0] >> j_shift) & 0xFF);
            res |= ((in[i2+1] >> j_shift) & 0xFF) <<  8;
            res |= ((in[i2+2] >> j_shift) & 0xFF) << 16;
            res |= ((in[i2+3] >> j_shift) & 0xFF) << 24;
            res |= ((in[i2+4] >> j_shift) & 0xFF) << 32;
            res |= ((in[i2+5] >> j_shift) & 0xFF) << 40;
            res |= ((in[i2+6] >> j_shift) & 0xFF) << 48;
            res |= ((in[i2+7] >> j_shift) & 0xFF) << 56;

            out[j + i2] = res;
        }
}


void bm64t_baseisa(const uint64_t in[64], uint64_t out[64])
{
    for (int j = 0; j < 64; j++)
    {
        uint64_t res = 0;
	    for (int i = 0; i < 64; i++)
            res |= ((in[i] >> j) & 1) << i;
        out[j] = res;
    }
}

void bm64t_baseisa2(const uint64_t in[64], uint64_t out[64])
{
    for (int j = 0; j < 64; j+=8)
    {
        uint64_t res0 = 0;
        uint64_t res1 = 0;
        uint64_t res2 = 0;
        uint64_t res3 = 0;
        uint64_t res4 = 0;
        uint64_t res5 = 0;
        uint64_t res6 = 0;
        uint64_t res7 = 0;

	    for (int i = 0; i < 64; i++)
        {
            uint64_t in_tmp = in[i];
            res0 |= ((in_tmp >> (j+0)) & 1) << i;
            res1 |= ((in_tmp >> (j+1)) & 1) << i;
            res2 |= ((in_tmp >> (j+2)) & 1) << i;
            res3 |= ((in_tmp >> (j+3)) & 1) << i;
            res4 |= ((in_tmp >> (j+4)) & 1) << i;
            res5 |= ((in_tmp >> (j+5)) & 1) << i;
            res6 |= ((in_tmp >> (j+6)) & 1) << i;
            res7 |= ((in_tmp >> (j+7)) & 1) << i;
        }
        
        out[j+0] = res0;
        out[j+1] = res1;
        out[j+2] = res2;
        out[j+3] = res3;
        out[j+4] = res4;
        out[j+5] = res5;
        out[j+6] = res6;
        out[j+7] = res7;
    }
}


static uint64_t hxor(uint64_t inp)
{
    inp ^= (inp>>32);
    inp ^= (inp>>16);
    inp ^= (inp>>8);
    inp ^= (inp>>4);
    inp ^= (inp>>2);
    inp ^= (inp>>1);
    return inp;
}

void bm64m_baseisa(const uint64_t in1[64], const uint64_t in2[8], uint64_t out[8])
{
    for (int j = 0; j < 8; j++) // in2[] columns, word index
    {
        uint64_t res = 0;
        uint64_t in_j = in2[j];

        for (int i = 0; i < 64; i++) // in1[] rows, word index
            res |= (hxor(in1[i] & in_j) & 1) << i;
        
        out[j] = res;
    }
}

void bm64m_baseisa2(const uint64_t in1[64], const uint64_t in2[8], uint64_t out[8])
{
    uint64_t res0 = 0;
    uint64_t res1 = 0;
    uint64_t res2 = 0;
    uint64_t res3 = 0;
    uint64_t res4 = 0;
    uint64_t res5 = 0;
    uint64_t res6 = 0;
    uint64_t res7 = 0;

    uint64_t in20 = in2[0];
    uint64_t in21 = in2[1];
    uint64_t in22 = in2[2];
    uint64_t in23 = in2[3];
    uint64_t in24 = in2[4];
    uint64_t in25 = in2[5];
    uint64_t in26 = in2[6];
    uint64_t in27 = in2[7];

    for (int i = 0; i < 64; i++) // in1[] rows, word index
    {
        uint64_t in_i = in1[i];

        res0 |= (hxor(in_i & in20) & 1) << i;
        res1 |= (hxor(in_i & in21) & 1) << i;
        res2 |= (hxor(in_i & in22) & 1) << i;
        res3 |= (hxor(in_i & in23) & 1) << i;
        res4 |= (hxor(in_i & in24) & 1) << i;
        res5 |= (hxor(in_i & in25) & 1) << i;
        res6 |= (hxor(in_i & in26) & 1) << i;
        res7 |= (hxor(in_i & in27) & 1) << i;
    }
    
    out[0] = res0;
    out[1] = res1;
    out[2] = res2;
    out[3] = res3;
    out[4] = res4;
    out[5] = res5;
    out[6] = res6;
    out[7] = res7;
}