代码

// fileName: test_O2_Transpose_Tilling_35.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 si = 0; si < N; si += 35)
		for (int sj = 0; sj < N; sj += 35)
			for (int i = si; i < si + 35; ++i)
				for (int j = sj; j < sj + 35; ++j)
					for (int k = 0; 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