代码

// fileName: test_O2_Transpose_Unroll_loops_13_type_2.cpp
#pragma GCC optimize("O2")

#include <random>
#include <chrono>

std::mt19937 rng(std::chrono::steady_clock::now().time_since_epoch().count());

typedef unsigned int UI;
typedef unsigned long long ULL;

const UI Mod = 998244353;
const int N = 1400;

UI A[N][N], B[N][N], C[N][N];

void Mul() {
	static UI B[N][N];
	for (int i = 0; i < N; ++i)
		for (int j = 0; j < N; ++j)
			B[i][j] = ::B[j][i];
	for (int i = 0; i < N; ++i)
		for (int j = 0; j < N; ++j)
			C[i][j] = 0;
	for (int i = 0; i < N; ++i)
		for (int j = 0; j < N; ++j) {
			int k = 0;
			for ( ; k + 12 < N; k += 13)
				C[i][j] = (C[i][j] + (ULL)A[i][k] * B[j][k] +
				(ULL)A[i][k + 1] * B[j][k + 1] +
				(ULL)A[i][k + 2] * B[j][k + 2] +
				(ULL)A[i][k + 3] * B[j][k + 3] +
				(ULL)A[i][k + 4] * B[j][k + 4] +
				(ULL)A[i][k + 5] * B[j][k + 5] +
				(ULL)A[i][k + 6] * B[j][k + 6] +
				(ULL)A[i][k + 7] * B[j][k + 7] +
				(ULL)A[i][k + 8] * B[j][k + 8] +
				(ULL)A[i][k + 9] * B[j][k + 9] +
				(ULL)A[i][k + 10] * B[j][k + 10] +
				(ULL)A[i][k + 11] * B[j][k + 11] +
				(ULL)A[i][k + 12] * B[j][k + 12]) % Mod;
			for ( ; k < N; ++k)
				C[i][j] = (C[i][j] + (ULL)A[i][k] * B[j][k]) % Mod;
		}
}

int Check() {
	static UI V[N], W1[N], W2[N], Tmp[N];
	for (int i = 0; i < N; ++i) {
		V[i] = rng() % Mod;
		W1[i] = 0;
		W2[i] = 0;
		Tmp[i] = 0;
	}
	for (int i = 0; i < N; ++i)
		for (int j = 0; j < N; ++j)
			Tmp[i] = (Tmp[i] + (ULL)B[i][j] * V[j]) % Mod;
	for (int i = 0; i < N; ++i)
		for (int j = 0; j < N; ++j)
			W1[i] = (W1[i] + (ULL)A[i][j] * Tmp[j]) % Mod;
	for (int i = 0; i < N; ++i)
		for (int j = 0; j < N; ++j)
			W2[i] = (W2[i] + (ULL)C[i][j] * V[j]) % Mod;
	for (int i = 0; i < N; ++i)
		if (W1[i] != W2[i]) return 1;
	return 0;
}

int main() {
	for (int i = 0; i < N; ++i) {
		for (int j = 0; j < N; ++j) {
			A[i][j] = rng() % Mod;
			B[i][j] = rng() % Mod;
		}
	}
	Mul();
	return Check();
}

评测结果