File size: 12,033 Bytes
b664585
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
255
256
257
#pragma once

// ggml-backend internal header

#include "ggml-backend.h"

#ifdef  __cplusplus
extern "C" {
#endif

    #define GGML_BACKEND_API_VERSION 1

    //
    // Backend buffer type
    //

    struct ggml_backend_buffer_type_i {
        const char *          (*get_name)      (ggml_backend_buffer_type_t buft);
        // allocate a buffer of this type
        ggml_backend_buffer_t (*alloc_buffer)  (ggml_backend_buffer_type_t buft, size_t size);
        // tensor alignment
        size_t                (*get_alignment) (ggml_backend_buffer_type_t buft);
        // (optional) max buffer size that can be allocated (defaults to SIZE_MAX)
        size_t                (*get_max_size)  (ggml_backend_buffer_type_t buft);
        // (optional) data size needed to allocate the tensor, including padding (defaults to ggml_nbytes)
        size_t                (*get_alloc_size)(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
        // (optional) check if tensor data is in host memory and uses standard ggml tensor layout (defaults to false)
        bool                  (*is_host)       (ggml_backend_buffer_type_t buft);
    };

    struct ggml_backend_buffer_type {
        struct ggml_backend_buffer_type_i  iface;
        ggml_backend_dev_t device;
        void * context;
    };

    //
    // Backend buffer
    //

    struct ggml_backend_buffer_i {
        // (optional) free the buffer
        void         (*free_buffer)  (ggml_backend_buffer_t buffer);
        // base address of the buffer
        void *       (*get_base)     (ggml_backend_buffer_t buffer);
        // (optional) initialize a tensor in the buffer (eg. add tensor extras)
        void         (*init_tensor)  (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
        // tensor data access
        void         (*memset_tensor)(ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);
        void         (*set_tensor)   (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
        void         (*get_tensor)   (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
        // (optional) tensor copy: dst is in the buffer, src may be in any buffer, including buffers from a different backend (return false if not supported)
        bool         (*cpy_tensor)   (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst);
        // clear the entire buffer
        void         (*clear)        (ggml_backend_buffer_t buffer, uint8_t value);
        // (optional) reset any internal state due to tensor initialization, such as tensor extras
        void         (*reset)        (ggml_backend_buffer_t buffer);
    };

    struct ggml_backend_buffer {
        struct ggml_backend_buffer_i  iface;
        ggml_backend_buffer_type_t    buft;
        void * context;
        size_t size;
        enum ggml_backend_buffer_usage usage;
    };

    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
                   ggml_backend_buffer_type_t buft,
            struct ggml_backend_buffer_i      iface,
                   void *                     context,
                   size_t                     size);

    // do not use directly, use ggml_backend_tensor_copy instead
    GGML_API bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);

    // multi-buffer
    // buffer that contains a collection of buffers
    GGML_API ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
    GGML_API bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
    GGML_API void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);

    //
    // Backend (stream)
    //

    struct ggml_backend_i {
        const char * (*get_name)(ggml_backend_t backend);

        void (*free)(ggml_backend_t backend);

        // (optional) asynchronous tensor data access
        void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
        void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
        bool (*cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);

        // (optional) complete all pending operations (required if the backend supports async operations)
        void (*synchronize)(ggml_backend_t backend);

        // (optional) graph plans (not used currently)
        // compute graph with a plan
        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
        void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
        // update the plan with a new graph - this should be faster than creating a new plan when the graph has the same topology
        void                      (*graph_plan_update) (ggml_backend_t backend, ggml_backend_graph_plan_t plan, const struct ggml_cgraph * cgraph);
        // compute the graph with the plan
        enum ggml_status          (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);

        // compute graph (always async if supported by the backend)
        enum ggml_status          (*graph_compute)     (ggml_backend_t backend, struct ggml_cgraph * cgraph);

        // (optional) event synchronization
        // record an event on this stream
        void (*event_record)(ggml_backend_t backend, ggml_backend_event_t event);
        // wait for an event on on a different stream
        void (*event_wait)  (ggml_backend_t backend, ggml_backend_event_t event);
    };

    struct ggml_backend {
        ggml_guid_t guid;
        struct ggml_backend_i iface;
        ggml_backend_dev_t device;
        void * context;
    };

    struct ggml_backend_event {
        struct ggml_backend_device * device;
        void * context;
    };

    //
    // Backend device
    //

    // Note: if additional properties are needed, we should add a struct with all of them
    //       the current functions to obtain the properties can remain, since they are more convenient for often used properties
    struct ggml_backend_device_i {
        // device name: short identifier for this device, such as "CPU" or "CUDA0"
        const char * (*get_name)(ggml_backend_dev_t dev);

        // device description: short informative description of the device, could be the model name
        const char * (*get_description)(ggml_backend_dev_t dev);

        // device memory in bytes
        void         (*get_memory)(ggml_backend_dev_t dev, size_t * free, size_t * total);

        // device type
        enum ggml_backend_dev_type (*get_type)(ggml_backend_dev_t dev);

        // device properties
        void (*get_props)(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props);

        // backend (stream) initialization
        ggml_backend_t (*init_backend)(ggml_backend_dev_t dev, const char * params);

        // preferred buffer type
        ggml_backend_buffer_type_t (*get_buffer_type)(ggml_backend_dev_t dev);

        // (optional) host buffer type (in system memory, typically this is a pinned memory buffer for faster transfers between host and device)
        ggml_backend_buffer_type_t (*get_host_buffer_type)(ggml_backend_dev_t dev);

        // (optional) buffer from pointer: create a buffer from a host pointer (useful for memory mapped models and importing data from other libraries)
        ggml_backend_buffer_t (*buffer_from_host_ptr)(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size);

        // check if the backend can compute an operation
        bool (*supports_op)(ggml_backend_dev_t dev, const struct ggml_tensor * op);

        // check if the backend can use tensors allocated in a buffer type
        bool (*supports_buft)(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft);

        // (optional) check if the backend wants to run an operation, even if the weights are allocated in an incompatible buffer
        // these should be expensive operations that may benefit from running on this backend instead of the CPU backend
        bool (*offload_op)(ggml_backend_dev_t dev, const struct ggml_tensor * op);

        // (optional) event synchronization
        ggml_backend_event_t (*event_new)         (ggml_backend_dev_t dev);
        void                 (*event_free)        (ggml_backend_dev_t dev, ggml_backend_event_t event);
        void                 (*event_synchronize) (ggml_backend_dev_t dev, ggml_backend_event_t event);
    };

    struct ggml_backend_device {
        struct ggml_backend_device_i iface;
        ggml_backend_reg_t reg;
        void * context;
    };

    //
    // Backend (reg)
    //

    struct ggml_backend_reg_i {
        const char * (*get_name)(ggml_backend_reg_t reg);

        // enumerate available devices
        size_t             (*get_device_count)(ggml_backend_reg_t reg);
        ggml_backend_dev_t (*get_device)(ggml_backend_reg_t reg, size_t index);

        // (optional) get a pointer to a function in the backend
        // backends can add custom functions that are not part of the standard ggml-backend interface
        void * (*get_proc_address)(ggml_backend_reg_t reg, const char * name);
    };

    struct ggml_backend_reg {
        int api_version; // initialize to GGML_BACKEND_API_VERSION
        struct ggml_backend_reg_i iface;
        void * context;
    };

    // Internal backend registry API
    GGML_API void ggml_backend_register(ggml_backend_reg_t reg);
    GGML_API void ggml_backend_device_register(ggml_backend_dev_t device);

    // Add backend dynamic loading support to the backend

    // Initialize the backend
    typedef ggml_backend_reg_t (*ggml_backend_init_t)(void);
    // Optional: obtain a score for the backend based on the system configuration
    // Higher scores are preferred, 0 means the backend is not supported in the current system
    typedef int                (*ggml_backend_score_t)(void);

#ifdef GGML_BACKEND_DL
#    ifdef __cplusplus
#        define GGML_BACKEND_DL_IMPL(reg_fn)                             \
            extern "C" {                                                 \
            GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
            }                                                            \
            ggml_backend_reg_t ggml_backend_init(void) {                 \
                return reg_fn();                                         \
            }
#        define GGML_BACKEND_DL_SCORE_IMPL(score_fn)       \
            extern "C" {                                   \
            GGML_BACKEND_API int ggml_backend_score(void); \
            }                                              \
            int ggml_backend_score(void) {                 \
                return score_fn();                         \
            }
#    else
#        define GGML_BACKEND_DL_IMPL(reg_fn)                              \
            GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void);  \
            ggml_backend_reg_t                  ggml_backend_init(void) { \
                return reg_fn();                                          \
            }
#        define GGML_BACKEND_DL_SCORE_IMPL(score_fn)        \
            GGML_BACKEND_API int ggml_backend_score(void);  \
            int                  ggml_backend_score(void) { \
                return score_fn();                          \
            }
#    endif
#else
#    define GGML_BACKEND_DL_IMPL(reg_fn)
#    define GGML_BACKEND_DL_SCORE_IMPL(score_fn)
#endif

#ifdef  __cplusplus
}
#endif