1
// g++ -g -std=c++20 -msse4.1 main.cpp && ./a.out
2
#pragma GCC diagnostic ignored "-Wignored-attributes"
3
#include <iomanip>
4
#include <iostream>
5
// 下面头文件从下到上为包含关系
6
// #include <mmmintrin.h> // MMX
7
#include <xmmintrin.h> // SSE
8
// #include <emmintrin.h> // SSE2
9
// #include <pmmintrin.h> // SSE3
10
// #include <tmmintrin.h> // SSSE3
11
// #include <smmintrin.h> // SSE4.1
12
// #include <nmmintrin.h> // SSE4.2
13
// #include <immintrin.h> // AVX/AVX2/... (建议)
14
// #include <x86intrin.h> // auto(GCC)
15

16
using namespace std;
17

18
#define IS_SIMD_TYPE(T, SIMD_TYPE) (is_same_v<remove_cv_t<remove_reference_t<T>>, SIMD_TYPE>)
19
template <class T>
20
void print_m128_32(T v, int line = 0) {
21
    if constexpr (IS_SIMD_TYPE(T, __m128)) {
22
        const float* f = reinterpret_cast<const float*>(&v);
23
        cout << fixed << setprecision(6) << "[" << f[0] << ", " << f[1] << ", " << f[2] << ", " << f[3] << "]\n";
24
    } else if constexpr (IS_SIMD_TYPE(T, __m128i)) {
25
        const int* i = reinterpret_cast<const int*>(&v);
26
        cout << "[" << i[0] << ", " << i[1] << ", " << i[2] << ", " << i[3] << "]\n";
27
    } else {
28
        const char* file = __FILE__;
29
        printf("unk vec %s:%d\n", file, line);
30
    }
31
}
32
template <class T>
33
void print_m128_16(T v, int line = 0) {
34
    if constexpr (IS_SIMD_TYPE(T, __m128i)) {
35
        const short* i = reinterpret_cast<const short*>(&v);
36
        cout << "[" << i[0] << ", " << i[1] << ", " << i[2] << ", " << i[3] //
37
             << ", " << i[4] << ", " << i[5] << ", " << i[6] << ", " << i[7] << "]\n";
38
    } else {
39
        const char* file = __FILE__;
40
        printf("unk vec %s:%d\n", file, line);
41
    }
42
}
43
template <class T>
44
void print_m128_64(T v, int line = 0) {
45
    if constexpr (IS_SIMD_TYPE(T, __m128d)) {
46
        const double* d = reinterpret_cast<const double*>(&v);
47
        cout << fixed << setprecision(6) << "[" << d[0] << ", " << d[1] << "]\n";
48
    } else {
49
        const char* file = __FILE__;
50
        printf("unk vec %s:%d\n", file, line);
51
    }
52
}
53

54
#define PRINT32(v) print_m128_32(v, __LINE__)
55
#define PRINT16(v) print_m128_16(v, __LINE__)
56
#define PRINT64(v) print_m128_64(v, __LINE__)
57

58
// 历史遗留问题：MMX 第一个 SIMD 指令集只支持整数运算。故 SSE(Streaming SIMD Extensions)的叫epi
59
// mm: MultiMedia Extension
60
// ps: Packed Single-precision floating-point, pd: Packed Double-precision floating-point
61
// si: Signed Integer
62
// epu: Extended Packed Unsigned integer
63
int main() {
64
    {
65
        // 基础运算与数据类型 and 类型转换与舍入模式
66
        float x = 1.5, y = 2.5, z = 2.6, w = -3.5;
67
        __m128 m = _mm_setr_ps(x, y, z, w);
68
        __m128 one = _mm_set1_ps(1.0f);
69
        m = _mm_add_ps(m, one);
70
        m = _mm_sub_ps(m, one);
71
        m = _mm_mul_ps(m, one);
72
        PRINT32(m);
73
        __m128i mi = _mm_castps_si128(m); // 无生成指令，按位读取
74
        PRINT32(mi);
75
        _MM_SET_ROUNDING_MODE(_MM_ROUND_NEAREST); // 默认，四舍五往偶数取整六入（考虑对全.5的数据不那么影响期望）
76
        // _MM_SET_ROUNDING_MODE(_MM_ROUND_DOWN);
77
        // _MM_SET_ROUNDING_MODE(_MM_ROUND_UP);
78
        // _MM_SET_ROUNDING_MODE(_MM_ROUND_TOWARD_ZERO);
79
        mi = _mm_cvtps_epi32(m);
80
        PRINT32(mi);
81
        mi = _mm_cvttps_epi32(m); // convert truncate: _MM_ROUND_TOWARD_ZERO
82
        PRINT32(mi);
83
    }
84
    {
85
        __m128i m = _mm_setr_epi16(1, 2, 3, 4, 5, 6, 7, 8);
86
        __m128i two = _mm_set1_epi16(2);
87
        __m128i tmp;
88
        tmp = _mm_mulhi_epi16(m, two);
89
        PRINT16(tmp);
90
        tmp = _mm_mullo_epi16(m, two);
91
        PRINT16(tmp);
92
    }
93
    {
94
        // 广播 与 Shuffle
95
        __m128i mi = _mm_setr_epi32(1, 2, 3, 4);
96
        mi = _mm_shuffle_epi32(mi, _MM_SHUFFLE(3, 2, 1, 0)); // 要反着看（
97
        PRINT32(mi);
98
        mi = _mm_shuffle_epi32(mi, _MM_SHUFFLE(2, 2, 0, 1)); // 这个参数是立即数，只能填常量
99
        PRINT32(mi);
100
        mi = _mm_shuffle_epi32(mi, 0x55); // 广播第二个元素(索引为1)
101
        PRINT32(mi);
102

103
        __m128 m1 = _mm_setr_ps(1, 2, 3, 4);
104
        __m128 m2 = _mm_setr_ps(5, 6, 7, 8);
105
        __m128 m = _mm_shuffle_ps(m1, m2, _MM_SHUFFLE(3, 2, 1, 0));
106
        PRINT32(m);
107
        m = _mm_shuffle_ps(m1, m2, _MM_SHUFFLE(3, 1, 2, 0));
108
        PRINT32(m);
109
        printf("\n");
110
        // 向量水平求和: m1.sum()
111
        auto horizontal_sum = [](__m128 v) -> float {
112
            __m128 tmp = _mm_shuffle_ps(v, v, _MM_SHUFFLE(2, 3, 0, 1));
113
            __m128 sums = _mm_add_ps(v, tmp);
114
            tmp = _mm_shuffle_ps(sums, sums, _MM_SHUFFLE(1, 0, 3, 2));
115
            return _mm_cvtss_f32(_mm_add_ss(sums, tmp));
116
        };
117
        __m128 tmp = _mm_shuffle_ps(m1, m1, _MM_SHUFFLE(2, 3, 0, 1));
118
        PRINT32(tmp);
119
        tmp = _mm_add_ps(m1, tmp);
120
        PRINT32(tmp);
121
        m1 = _mm_shuffle_ps(tmp, tmp, _MM_SHUFFLE(1, 0, 3, 2));
122
        m1 = _mm_add_ps(m1, tmp);
123
        PRINT32(m1);
124
        float sum = _mm_cvtss_f32(m1);
125
        printf("m1 sum = %f\n", sum);
126
        printf("m2 sum = %f\n\n", horizontal_sum(m2));
127
    }
128
    {
129
        // Load and Store
130
        // 高效构造 f[]
131
        int _space = 0;
132
        float f[4] = {1, 2, 3, 4};
133
        alignas(32) __m128 m = _mm_loadu_ps(f); // _mm_load_ps(), _mm_load_ss(), _mm_load1_ps()
134
        PRINT32(m);
135
        float a[4];
136
        _mm_storeu_ps(a, m);
137
        for (float f : a) {
138
            printf(" %f ", f);
139
        }
140
    }
141
}