aboutsummaryrefslogtreecommitdiff
path: root/liblangdet/impl.h
blob: f2165185d9267d43148c8c658912ff64d560823c (plain) (blame)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
/* Inline vector operations keep attention rows in registers across reductions. */

#ifndef KERNELS_IMPL_H
#define KERNELS_IMPL_H

#include <stdint.h>

float dot_f32_generic(const float *a, const float *b, int n);
void vec_scale_inplace_generic(float *dst, float scale, int n);
void vec_axpy_inplace_generic(float *dst, const float *src, float alpha, int n);
void vec_scale_add_generic(float *dst, const float *src, float correction, int n);

#ifdef __ARM_NEON

#    include <arm_neon.h>

static inline float dot_f32_impl(const float *a, const float *b, int n) {
    int i = 0;
    float32x4_t acc0 = vdupq_n_f32(0.0f);
    float32x4_t acc1 = vdupq_n_f32(0.0f);
    for (; i + 8 <= n; i += 8) {
        float32x4_t a0 = vld1q_f32(a + i);
        float32x4_t b0 = vld1q_f32(b + i);
        float32x4_t a1 = vld1q_f32(a + i + 4);
        float32x4_t b1 = vld1q_f32(b + i + 4);
        acc0 = vfmaq_f32(acc0, a0, b0);
        acc1 = vfmaq_f32(acc1, a1, b1);
    }
    float sum = vaddvq_f32(vaddq_f32(acc0, acc1));
    for (; i < n; i++)
        sum += a[i] * b[i];
    return sum;
}

static inline void vec_scale_inplace_impl(float *dst, float scale, int n) {
    int i = 0;
    float32x4_t s = vdupq_n_f32(scale);
    for (; i + 8 <= n; i += 8) {
        float32x4_t d0 = vld1q_f32(dst + i);
        float32x4_t d1 = vld1q_f32(dst + i + 4);
        vst1q_f32(dst + i, vfmaq_f32(vdupq_n_f32(0.0f), d0, s));
        vst1q_f32(dst + i + 4, vfmaq_f32(vdupq_n_f32(0.0f), d1, s));
    }
    for (; i < n; i++)
        dst[i] *= scale;
}

static inline void vec_axpy_inplace_impl(float *dst, const float *src, float alpha, int n) {
    int i = 0;
    float32x4_t a = vdupq_n_f32(alpha);
    for (; i + 8 <= n; i += 8) {
        float32x4_t d0 = vld1q_f32(dst + i);
        float32x4_t s0 = vld1q_f32(src + i);
        float32x4_t d1 = vld1q_f32(dst + i + 4);
        float32x4_t s1 = vld1q_f32(src + i + 4);
        vst1q_f32(dst + i, vfmaq_f32(d0, s0, a));
        vst1q_f32(dst + i + 4, vfmaq_f32(d1, s1, a));
    }
    for (; i < n; i++)
        dst[i] += alpha * src[i];
}

static inline void vec_scale_add_impl(float *dst, const float *src, float correction, int n) {
    int i = 0;
    float32x4_t c = vdupq_n_f32(correction);
    for (; i + 8 <= n; i += 8) {
        float32x4_t d0 = vld1q_f32(dst + i);
        float32x4_t s0 = vld1q_f32(src + i);
        float32x4_t d1 = vld1q_f32(dst + i + 4);
        float32x4_t s1 = vld1q_f32(src + i + 4);
        vst1q_f32(dst + i, vfmaq_f32(s0, d0, c));
        vst1q_f32(dst + i + 4, vfmaq_f32(s1, d1, c));
    }
    for (; i < n; i++)
        dst[i] = dst[i] * correction + src[i];
}

#elif defined(__AVX2__) && defined(__FMA__)

#    include <immintrin.h>

static inline float dot_f32_impl(const float *a, const float *b, int n) {
#    if defined(__AVX512F__)
    int i = 0;
    __m512 acc0 = _mm512_setzero_ps();
    __m512 acc1 = _mm512_setzero_ps();
    __m512 acc2 = _mm512_setzero_ps();
    __m512 acc3 = _mm512_setzero_ps();
    for (; i + 64 <= n; i += 64) {
        acc0 = _mm512_fmadd_ps(_mm512_loadu_ps(a + i), _mm512_loadu_ps(b + i), acc0);
        acc1 = _mm512_fmadd_ps(_mm512_loadu_ps(a + i + 16), _mm512_loadu_ps(b + i + 16), acc1);
        acc2 = _mm512_fmadd_ps(_mm512_loadu_ps(a + i + 32), _mm512_loadu_ps(b + i + 32), acc2);
        acc3 = _mm512_fmadd_ps(_mm512_loadu_ps(a + i + 48), _mm512_loadu_ps(b + i + 48), acc3);
    }
    __m512 acc = _mm512_add_ps(_mm512_add_ps(acc0, acc1), _mm512_add_ps(acc2, acc3));
    for (; i + 16 <= n; i += 16) {
        acc = _mm512_fmadd_ps(_mm512_loadu_ps(a + i), _mm512_loadu_ps(b + i), acc);
    }
    float sum = _mm512_reduce_add_ps(acc);
    for (; i < n; i++)
        sum += a[i] * b[i];
    return sum;
#    else
    int i = 0;
    __m256 acc0 = _mm256_setzero_ps();
    __m256 acc1 = _mm256_setzero_ps();
    __m256 acc2 = _mm256_setzero_ps();
    __m256 acc3 = _mm256_setzero_ps();
    for (; i + 32 <= n; i += 32) {
        acc0 = _mm256_fmadd_ps(_mm256_loadu_ps(a + i), _mm256_loadu_ps(b + i), acc0);
        acc1 = _mm256_fmadd_ps(_mm256_loadu_ps(a + i + 8), _mm256_loadu_ps(b + i + 8), acc1);
        acc2 = _mm256_fmadd_ps(_mm256_loadu_ps(a + i + 16), _mm256_loadu_ps(b + i + 16), acc2);
        acc3 = _mm256_fmadd_ps(_mm256_loadu_ps(a + i + 24), _mm256_loadu_ps(b + i + 24), acc3);
    }
    acc0 = _mm256_add_ps(_mm256_add_ps(acc0, acc1), _mm256_add_ps(acc2, acc3));
    for (; i + 8 <= n; i += 8) {
        acc0 = _mm256_fmadd_ps(_mm256_loadu_ps(a + i), _mm256_loadu_ps(b + i), acc0);
    }
    __m128 r = _mm_add_ps(_mm256_castps256_ps128(acc0), _mm256_extractf128_ps(acc0, 1));
    r = _mm_hadd_ps(r, r);
    r = _mm_hadd_ps(r, r);
    float sum = _mm_cvtss_f32(r);
    for (; i < n; i++)
        sum += a[i] * b[i];
    return sum;
#    endif
}

static inline void vec_scale_inplace_impl(float *dst, float scale, int n) {
#    if defined(__AVX512F__)
    int i = 0;
    __m512 s = _mm512_set1_ps(scale);
    for (; i + 64 <= n; i += 64) {
        _mm512_storeu_ps(dst + i, _mm512_mul_ps(_mm512_loadu_ps(dst + i), s));
        _mm512_storeu_ps(dst + i + 16, _mm512_mul_ps(_mm512_loadu_ps(dst + i + 16), s));
        _mm512_storeu_ps(dst + i + 32, _mm512_mul_ps(_mm512_loadu_ps(dst + i + 32), s));
        _mm512_storeu_ps(dst + i + 48, _mm512_mul_ps(_mm512_loadu_ps(dst + i + 48), s));
    }
    for (; i + 16 <= n; i += 16) {
        _mm512_storeu_ps(dst + i, _mm512_mul_ps(_mm512_loadu_ps(dst + i), s));
    }
    for (; i < n; i++)
        dst[i] *= scale;
#    else
    int i = 0;
    __m256 s = _mm256_set1_ps(scale);
    for (; i + 32 <= n; i += 32) {
        _mm256_storeu_ps(dst + i, _mm256_mul_ps(_mm256_loadu_ps(dst + i), s));
        _mm256_storeu_ps(dst + i + 8, _mm256_mul_ps(_mm256_loadu_ps(dst + i + 8), s));
        _mm256_storeu_ps(dst + i + 16, _mm256_mul_ps(_mm256_loadu_ps(dst + i + 16), s));
        _mm256_storeu_ps(dst + i + 24, _mm256_mul_ps(_mm256_loadu_ps(dst + i + 24), s));
    }
    for (; i + 8 <= n; i += 8) {
        _mm256_storeu_ps(dst + i, _mm256_mul_ps(_mm256_loadu_ps(dst + i), s));
    }
    for (; i < n; i++)
        dst[i] *= scale;
#    endif
}

static inline void vec_axpy_inplace_impl(float *dst, const float *src, float alpha, int n) {
#    if defined(__AVX512F__)
    int i = 0;
    __m512 a = _mm512_set1_ps(alpha);
    for (; i + 64 <= n; i += 64) {
        _mm512_storeu_ps(dst + i,
                         _mm512_fmadd_ps(_mm512_loadu_ps(src + i), a, _mm512_loadu_ps(dst + i)));
        _mm512_storeu_ps(dst + i + 16, _mm512_fmadd_ps(_mm512_loadu_ps(src + i + 16), a,
                                                       _mm512_loadu_ps(dst + i + 16)));
        _mm512_storeu_ps(dst + i + 32, _mm512_fmadd_ps(_mm512_loadu_ps(src + i + 32), a,
                                                       _mm512_loadu_ps(dst + i + 32)));
        _mm512_storeu_ps(dst + i + 48, _mm512_fmadd_ps(_mm512_loadu_ps(src + i + 48), a,
                                                       _mm512_loadu_ps(dst + i + 48)));
    }
    for (; i + 16 <= n; i += 16) {
        _mm512_storeu_ps(dst + i,
                         _mm512_fmadd_ps(_mm512_loadu_ps(src + i), a, _mm512_loadu_ps(dst + i)));
    }
    for (; i < n; i++)
        dst[i] += alpha * src[i];
#    else
    int i = 0;
    __m256 a = _mm256_set1_ps(alpha);
    for (; i + 32 <= n; i += 32) {
        _mm256_storeu_ps(dst + i,
                         _mm256_fmadd_ps(_mm256_loadu_ps(src + i), a, _mm256_loadu_ps(dst + i)));
        _mm256_storeu_ps(dst + i + 8, _mm256_fmadd_ps(_mm256_loadu_ps(src + i + 8), a,
                                                      _mm256_loadu_ps(dst + i + 8)));
        _mm256_storeu_ps(dst + i + 16, _mm256_fmadd_ps(_mm256_loadu_ps(src + i + 16), a,
                                                       _mm256_loadu_ps(dst + i + 16)));
        _mm256_storeu_ps(dst + i + 24, _mm256_fmadd_ps(_mm256_loadu_ps(src + i + 24), a,
                                                       _mm256_loadu_ps(dst + i + 24)));
    }
    for (; i + 8 <= n; i += 8) {
        _mm256_storeu_ps(dst + i,
                         _mm256_fmadd_ps(_mm256_loadu_ps(src + i), a, _mm256_loadu_ps(dst + i)));
    }
    for (; i < n; i++)
        dst[i] += alpha * src[i];
#    endif
}

static inline void vec_scale_add_impl(float *dst, const float *src, float correction, int n) {
#    if defined(__AVX512F__)
    int i = 0;
    __m512 c = _mm512_set1_ps(correction);
    for (; i + 64 <= n; i += 64) {
        _mm512_storeu_ps(dst + i,
                         _mm512_fmadd_ps(_mm512_loadu_ps(dst + i), c, _mm512_loadu_ps(src + i)));
        _mm512_storeu_ps(dst + i + 16, _mm512_fmadd_ps(_mm512_loadu_ps(dst + i + 16), c,
                                                       _mm512_loadu_ps(src + i + 16)));
        _mm512_storeu_ps(dst + i + 32, _mm512_fmadd_ps(_mm512_loadu_ps(dst + i + 32), c,
                                                       _mm512_loadu_ps(src + i + 32)));
        _mm512_storeu_ps(dst + i + 48, _mm512_fmadd_ps(_mm512_loadu_ps(dst + i + 48), c,
                                                       _mm512_loadu_ps(src + i + 48)));
    }
    for (; i + 16 <= n; i += 16) {
        _mm512_storeu_ps(dst + i,
                         _mm512_fmadd_ps(_mm512_loadu_ps(dst + i), c, _mm512_loadu_ps(src + i)));
    }
    for (; i < n; i++)
        dst[i] = dst[i] * correction + src[i];
#    else
    int i = 0;
    __m256 c = _mm256_set1_ps(correction);
    for (; i + 32 <= n; i += 32) {
        _mm256_storeu_ps(dst + i,
                         _mm256_fmadd_ps(_mm256_loadu_ps(dst + i), c, _mm256_loadu_ps(src + i)));
        _mm256_storeu_ps(dst + i + 8, _mm256_fmadd_ps(_mm256_loadu_ps(dst + i + 8), c,
                                                      _mm256_loadu_ps(src + i + 8)));
        _mm256_storeu_ps(dst + i + 16, _mm256_fmadd_ps(_mm256_loadu_ps(dst + i + 16), c,
                                                       _mm256_loadu_ps(src + i + 16)));
        _mm256_storeu_ps(dst + i + 24, _mm256_fmadd_ps(_mm256_loadu_ps(dst + i + 24), c,
                                                       _mm256_loadu_ps(src + i + 24)));
    }
    for (; i + 8 <= n; i += 8) {
        _mm256_storeu_ps(dst + i,
                         _mm256_fmadd_ps(_mm256_loadu_ps(dst + i), c, _mm256_loadu_ps(src + i)));
    }
    for (; i < n; i++)
        dst[i] = dst[i] * correction + src[i];
#    endif
}

#else /* no SIMD path: dispatch to the scalar references */

#    define dot_f32_impl           dot_f32_generic
#    define vec_scale_inplace_impl vec_scale_inplace_generic
#    define vec_axpy_inplace_impl  vec_axpy_inplace_generic
#    define vec_scale_add_impl     vec_scale_add_generic

#endif

#endif /* KERNELS_IMPL_H */