point4_sse.cpp

#include <iostream>
#include <algorithm>
#include <vector>
#include <assert.h>
#include <time.h>

#include "vectorclass/vectorf128.h"

#include "alignedallocator.hpp"

using namespace std;

struct Point4f {
	Point4f(){ sse = _mm_set_ps1(0); }
	Point4f(float x, float y, float z, float w) {
		sse = _mm_set_ps(w, z, y, x);
	}
	Point4f(float v){ sse = _mm_set_ps1(v); }
	Point4f(__m128 v) { sse = v; }
	
	Point4f& operator=(const Point4f& p) {
		sse = p.sse;
		return *this;
	}

	Point4f operator+(const Point4f& p) {
		return _mm_add_ps(sse, p.sse);
	}
	Point4f operator-(const Point4f& p) {
		return _mm_sub_ps(sse, p.sse);
	}
	Point4f operator*(const Point4f& p) {
		return _mm_mul_ps(sse, p.sse);
	}
	Point4f operator/(const Point4f& p) {
		return _mm_div_ps(sse, p.sse);
	}

	__m128 operator==(const Point4f& p) {
		return _mm_cmpeq_ps(sse, p.sse);
	}

	__m128 operator!=(const Point4f& p) {
		return _mm_cmpneq_ps(sse, p.sse);
	}

	operator __m128() {
		return sse;
	}

	float& operator[](int idx) {
		assert(idx >= 0 && idx < 4);
		return (&x)[idx];
	}

	float operator[](int idx) const {
		assert(idx >= 0 && idx < 4);
		return (&x)[idx];
	}

	friend ostream& operator << (ostream& os, const Point4f& p);
	friend istream& operator >> (istream& is, Point4f& p);

	union {
		__m128 sse;
		struct { float x, y, z, w; };
	};
};

ostream& operator << (ostream& os, const Point4f& p) {
	os << p.x << ' ' << p.y << ' ' << p.z << ' ' << p.w << endl;
	return os;
}

istream& operator >>(istream& is, Point4f& p) {
	is >> p.x >> p.y >> p.z >> p.w;
	return is;
}

int main() {
	__m128 m1 = _mm_set_ps1(7.0);
	__m128 mref = _mm_set_ps1(sqrtf(7.0));
	float res_val[4] = { 0 };

	int runs = 1;
	int n = 64 * 1024 * 1024;
	vector<Point4f, aligned_allocator<Point4f, 16>> vec(n, m1);

	clock_t start, stop;
	start = clock();
	for (int j = 0; j < runs;j++)
	for (int i = 0; i < n;i++) {
		vec[i] = _mm_sqrt_ps(vec[i]);
	}
	stop = clock();
	cout << "time cost with SIMD = " << (stop - start) / (float)CLOCKS_PER_SEC * 1000.0 << "ms" << endl;

	// verify
	bool flag = false;
	for (int i = 0; i < n; i++) {
		Vec4fb res = (vec[i] != mref);
		if (res[0] || res[1] || res[2] || res[3]) {
			cout << vec[i] << '\t' << mref << endl;
			flag = true;
			break;
		}
	}
	if (flag) cout << "Failed." << endl;
	else cout << "Succeeded." << endl;

	start = clock();
	for (int j = 0; j < runs; j++)
	for (int i = 0; i < n; i++) {
		Point4f& p = vec[i];
		p[0] = sqrtf(7.0);
		p[1] = sqrtf(7.0);
		p[2] = sqrtf(7.0);
		p[3] = sqrtf(7.0);
	}
	stop = clock();
	cout << "time cost with SIMD = " << (stop - start) / (float)CLOCKS_PER_SEC * 1000.0 << "ms" << endl;
	// verify
	flag = false;
	for (int i = 0; i < n; i++) {
		Vec4fb res = (vec[i] != mref);
		if (res[0] || res[1] || res[2] || res[3]) {
			flag = true;
			break;
		}
	}
	if (flag) cout << "Failed." << endl;
	else cout << "Succeeded." << endl;

	Point4f zero;
	cout << zero << endl;
	zero = zero + Point4f(1, 2, 3, 4);
	cout << zero << endl;
	zero = zero * zero;
	cout << zero << endl;
	zero = zero / zero;
	cout << zero << endl;
	return 0;
}