CS:APP 第五章 习题

5.13

A

关键路径%xmm0 add。

B

浮点数加法的延迟界限。

C

整数加法的延迟界限。

D

由于乘法并未不是循环迭代间的关键路径，下一循环乘法开始时不需要上一循环的乘法结束，所以乘法操作可以一直处于流水线中并可以提前进行，乘法操作完毕后将值赋给当前循环的sum即可；且由于乘法操作的吞吐量大于该程序的运行速度，所以该程序瓶颈点并非乘法。

5.14

void inner4(vec_ptr u, vec_ptr v, data_t* dest) {
	long i;
	long len = vec_length(u);
	data_t* udata = get_vec_start(u);
	data_t* vdata = get_vec_start(v);
	data_t sum = 0;

	for (i = 0; i < len - 5; i++) {
		sum += udata[0] * vdata[0] +
			udata[1] * vdata[1] +
			udata[2] * vdata[2] +
			udata[3] * vdata[3] +
			udata[4] * vdata[4] +
			udata[5] * vdata[5];
	}
	for (; i < len; i++) {
		sum += udata[i] * vdata[i];
	}
	*dest = sum;
}

A

每个时钟周期只能从内存中加载两个值。

B

浮点数加法限制了速度。

5.15

void inner4(vec_ptr u, vec_ptr v, data_t* dest) {
	long i;
	long len = vec_length(u);
	data_t* udata = get_vec_start(u);
	data_t* vdata = get_vec_start(v);
	data_t sum0 = 0;
	data_t sum1 = 0;
	data_t sum2 = 0;
	data_t sum3 = 0;
	data_t sum4 = 0;
	data_t sum5 = 0;

	for (i = 0; i < len - 5; i++) {
		sum0 += udata[0] * vdata[0];
		sum1 += udata[1] * vdata[1];
		sum2 += udata[2] * vdata[2];
		sum3 += udata[3] * vdata[3];
		sum4 += udata[4] * vdata[4];
		sum5 += udata[5] * vdata[5];
	}
	for (; i < len; i++) {
		sum0 += udata[i] * vdata[i];
	}
	sum0 += sum1 + sum2 + sum3 + sum4 + sum5;
	*dest = sum0;
}

1.00只是理论的下界，有多种其他因素限制了实际速度。

5.16

void inner4(vec_ptr u, vec_ptr v, data_t* dest) {
	long i;
	long len = vec_length(u);
	data_t* udata = get_vec_start(u);
	data_t* vdata = get_vec_start(v);
	data_t sum = 0;

	for (i = 0; i < len - 5; i++) {
		sum = sum +
			(udata[0] * vdata[0] +
				udata[1] * vdata[1] +
				udata[2] * vdata[2] +
				udata[3] * vdata[3] +
				udata[4] * vdata[4] +
				udata[5] * vdata[5]);
	}
	for (; i < len; i++) {
		sum += udata[i] * vdata[i];
	}
	*dest = sum;
}

5.17

void* basic_memset(void* s, int c, size_t n) {
	unsigned long qword = (unsigned char)c;
	unsigned long* sqword = (unsigned long*)s;
	unsigned char* schar = (unsigned char*)s;
	qword &= qword << 8;
	qword &= qword << 16;
	qword &= qword << 32;
	int cnt = 0;
	int k = sizeof(unsigned long);
	int limit = n - k + 1;
	while (cnt < n && (unsigned int)sqword % k) {
		*(schar++) = (unsigned char)c;
		cnt++;
	}
	sqword = (unsigned long*)schar;
	while (cnt < limit) {
		*sqword = qword;
		sqword += 8;
		cnt += 8;
	}
	schar = (unsigned char*)sqword;
	while (cnt < n) {
		(*schar++) = (unsigned char)c;
		++cnt;
	}
	return s;
}

5.18

double mul() {
	double result0 = 0;
	double result1 = 0;
	double result2 = 0;
	double result3 = 0;
	double result4 = 0;
	double xpwr = 1;
	double xmul = x * x * x * x * x;
	double xpwr1 = xmul;
	int limit = N - 9;
	int i = 0;
	for (i = 0; i < limit; i+=10) {
		result0 = result0 + (xpwr * a[i] + xpwr1 * a[i + 5]);
		result1 = result1 + (xpwr * a[i + 1] + xpwr1 * a[i + 6]);
		result2 = result2 + (xpwr * a[i + 2] + xpwr1 * a[i + 7]);
		result3 = result3 + (xpwr * a[i + 3] + xpwr1 * a[i + 8]);
		result4 = result4 + (xpwr * a[i + 4] + xpwr1 * a[i + 9]);

		xpwr *= xmul;
		xpwr1 = xpwr * xmul;
	}
	for (; i < N; i++) {
		result0 += xpwr * a[i];
		xpwr *= x;
	}
	result0 += x * result1 +
		x * x * result2 +
		x * x * x * result3 +
		x * x * x * x * result4;
	return result0;
}

5.19

两次展开：

void psum(float a[], float p[], long n) {
	long i;
	long limit = n - 1;
	float last_val, val0, val1;
	last_val = p[0] = a[0];
	for (i = 1; i < limit; i+=2) {
		val0 = last_val + a[i];
		val1 = last_val + a[i] + a[i + 1];
		p[i] = val0;
		p[i + 1] = val1;
		last_val = val1;
	}
	for (; i < n; i++) {
		val0 = last_val + a[i];
	}
}