llama-cpp-sys-4 0.7.0

Low Level Bindings to llama.cpp
Documentation 
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

#ifndef HVX_SCALE_H
#define HVX_SCALE_H

#include <assert.h>
#include <stddef.h>
#include <stdint.h>

#include "hvx-base.h"

#define hvx_scale_f32_loop_body(dst_type, src_type, vec_store)                       \
    do {                                                                             \
        dst_type * restrict vdst = (dst_type *) dst;                                 \
        src_type * restrict vsrc = (src_type *) src;                                 \
                                                                                     \
        HVX_Vector vs = hvx_vec_splat_f32(scale);                                    \
                                                                                     \
        const uint32_t elem_size = sizeof(float);                                    \
        const uint32_t epv = 128 / elem_size;                                        \
        const uint32_t nvec = n / epv;                                               \
        const uint32_t nloe = n % epv;                                               \
                                                                                     \
        uint32_t i = 0;                                                              \
                                                                                     \
        _Pragma("unroll(4)")                                                         \
        for (; i < nvec; ++i) {                                                      \
            HVX_Vector v = Q6_Vqf32_vmpy_VsfVsf(vsrc[i], vs);                        \
            vdst[i]      = Q6_Vsf_equals_Vqf32(v);                                   \
        }                                                                            \
        if (nloe) {                                                                  \
            HVX_Vector v = Q6_Vqf32_vmpy_VsfVsf(vsrc[i], vs);                        \
            vec_store((void *) &vdst[i], nloe * elem_size, Q6_Vsf_equals_Vqf32(v));  \
        }                                                                            \
    } while(0)

static inline void hvx_scale_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale) {
    assert((size_t) dst % 128 == 0);
    assert((size_t) src % 128 == 0);
    hvx_scale_f32_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
}

static inline void hvx_scale_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale) {
    assert((size_t) dst % 128 == 0);
    hvx_scale_f32_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
}

static inline void hvx_scale_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale) {
    assert((size_t) src % 128 == 0);
    hvx_scale_f32_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
}

static inline void hvx_scale_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale) {
    hvx_scale_f32_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
}

static inline void hvx_scale_f32(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale) {
    if (((size_t) dst & 127) == 0) {
        if (((size_t) src & 127) == 0) {
            hvx_scale_f32_aa(dst, src, n, scale);
        } else {
            hvx_scale_f32_au(dst, src, n, scale);
        }
    } else {
        if (((size_t) src & 127) == 0) {
            hvx_scale_f32_ua(dst, src, n, scale);
        } else {
            hvx_scale_f32_uu(dst, src, n, scale);
        }
    }
}

#define hvx_scale_offset_f32_loop_body(dst_type, src_type, vec_store)                \
    do {                                                                             \
        dst_type * restrict vdst = (dst_type *) dst;                                 \
        src_type * restrict vsrc = (src_type *) src;                                 \
                                                                                     \
        HVX_Vector vs = hvx_vec_splat_f32(scale);                                    \
        HVX_Vector vo = hvx_vec_splat_f32(offset);                                   \
                                                                                     \
        const uint32_t elem_size = sizeof(float);                                    \
        const uint32_t epv = 128 / elem_size;                                        \
        const uint32_t nvec = n / epv;                                               \
        const uint32_t nloe = n % epv;                                               \
                                                                                     \
        uint32_t i = 0;                                                              \
                                                                                     \
        _Pragma("unroll(4)")                                                         \
        for (; i < nvec; ++i) {                                                      \
            HVX_Vector v = Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vmpy_VsfVsf(vsrc[i], vs), vo); \
            vdst[i] = Q6_Vsf_equals_Vqf32(v);                                        \
        }                                                                            \
        if (nloe) {                                                                  \
            HVX_Vector v = Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vmpy_VsfVsf(vsrc[i], vs), vo); \
            vec_store((void *) &vdst[i], nloe * elem_size, Q6_Vsf_equals_Vqf32(v));  \
        }                                                                            \
    } while(0)

static inline void hvx_scale_offset_f32_aa(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
    assert((size_t) dst % 128 == 0);
    assert((size_t) src % 128 == 0);
    hvx_scale_offset_f32_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
}

static inline void hvx_scale_offset_f32_au(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
    assert((size_t) dst % 128 == 0);
    hvx_scale_offset_f32_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
}

static inline void hvx_scale_offset_f32_ua(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
    assert((size_t) src % 128 == 0);
    hvx_scale_offset_f32_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
}

static inline void hvx_scale_offset_f32_uu(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
    hvx_scale_offset_f32_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
}

static inline void hvx_scale_offset_f32(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
    if (((size_t) dst & 127) == 0) {
        if (((size_t) src & 127) == 0) {
            hvx_scale_offset_f32_aa(dst, src, n, scale, offset);
        } else {
            hvx_scale_offset_f32_au(dst, src, n, scale, offset);
        }
    } else {
        if (((size_t) src & 127) == 0) {
            hvx_scale_offset_f32_ua(dst, src, n, scale, offset);
        } else {
            hvx_scale_offset_f32_uu(dst, src, n, scale, offset);
        }
    }
}

// Scale+offset computed by promoting f16 -> f32, then narrowing the result back to f16.
#define hvx_scale_offset_f16_loop_body(dst_type, src_type, vec_store)                                                \
    do {                                                                                                             \
        dst_type * restrict vdst = (dst_type *) dst;                                                                 \
        src_type * restrict vsrc = (src_type *) src;                                                                 \
                                                                                                                     \
        HVX_Vector vs = hvx_vec_splat_f32(scale);                                                                    \
        HVX_Vector vo = hvx_vec_splat_f32(offset);                                                                   \
                                                                                                                     \
        const uint32_t nvec = n / VLEN_FP16;                                                                         \
        const uint32_t nloe = n % VLEN_FP16;                                                                         \
                                                                                                                     \
        uint32_t i = 0;                                                                                              \
                                                                                                                     \
        _Pragma("unroll(4)")                                                                                         \
        for (; i < nvec; ++i) {                                                                                      \
            HVX_VectorPair p = hvx_vec_f16_to_f32(vsrc[i]);                                                          \
            HVX_Vector r0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(p), vs), vo)); \
            HVX_Vector r1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(p), vs), vo)); \
            vdst[i] = hvx_vec_f32_to_f16(r0, r1);                                                                    \
        }                                                                                                            \
        if (nloe) {                                                                                                  \
            HVX_VectorPair p = hvx_vec_f16_to_f32(vsrc[i]);                                                          \
            HVX_Vector r0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(p), vs), vo)); \
            HVX_Vector r1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(p), vs), vo)); \
            HVX_Vector v = hvx_vec_f32_to_f16(r0, r1);                                                               \
            vec_store((void *) &vdst[i], nloe * SIZEOF_FP16, v);                                                     \
        }                                                                                                            \
    } while(0)

static inline void hvx_scale_offset_f16_aa(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
    assert((size_t) dst % 128 == 0);
    assert((size_t) src % 128 == 0);
    hvx_scale_offset_f16_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
}

static inline void hvx_scale_offset_f16_au(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
    assert((size_t) dst % 128 == 0);
    hvx_scale_offset_f16_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
}

static inline void hvx_scale_offset_f16_ua(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
    assert((size_t) src % 128 == 0);
    hvx_scale_offset_f16_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
}

static inline void hvx_scale_offset_f16_uu(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
    hvx_scale_offset_f16_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
}

static inline void hvx_scale_offset_f16(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
    if (((size_t) dst & 127) == 0) {
        if (((size_t) src & 127) == 0) {
            hvx_scale_offset_f16_aa(dst, src, n, scale, offset);
        } else {
            hvx_scale_offset_f16_au(dst, src, n, scale, offset);
        }
    } else {
        if (((size_t) src & 127) == 0) {
            hvx_scale_offset_f16_ua(dst, src, n, scale, offset);
        } else {
            hvx_scale_offset_f16_uu(dst, src, n, scale, offset);
        }
    }
}

#endif // HVX_SCALE_H