bitset<N> B1[N]; // матрица B1
bitset<N> dp[1 << k];

int up = 0;
for (int mask=1; mask<(1<<n); mask++) {
	if (mask == (1 << (up+1)))	
		up++;
	dp[mask] = dp[mask^(1<<up)] ^ B1[up]; // N/w
}
// (2^k) * (N/w)

// 0 <= A1[i] < 2^k
// N * (N/w)
forn(i, N)
	C1[i] = dp[A1[i]] // N/w

// Time(A1*B1) = (2^k + N) * (N/w)
// Time(A*B) = (N/k) * (2^k + N) * (N/w)

k = logN
Time = N^3 / logN / w