File size: 13,346 Bytes
4bdb245
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#pragma once

#include "ggml.h"
#include "ggml-alloc.h"

#ifdef  __cplusplus
extern "C" {
#endif

    typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
    typedef struct ggml_backend_event * ggml_backend_event_t;
    typedef struct ggml_backend * ggml_backend_t;
    typedef void * ggml_backend_graph_plan_t;

    //
    // Backend buffer
    //

    // buffer type
    GGML_API           const char *          ggml_backend_buft_name            (ggml_backend_buffer_type_t buft);
    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_buft_alloc_buffer    (ggml_backend_buffer_type_t buft, size_t size);
    GGML_API           size_t                ggml_backend_buft_get_alignment   (ggml_backend_buffer_type_t buft);
    GGML_API           size_t                ggml_backend_buft_get_max_size    (ggml_backend_buffer_type_t buft);
    GGML_API GGML_CALL size_t                ggml_backend_buft_get_alloc_size  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
    GGML_API           bool                  ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
    GGML_API           bool                  ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);

    // buffer
    enum ggml_backend_buffer_usage {
        GGML_BACKEND_BUFFER_USAGE_ANY = 0,
        GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
    };

    GGML_API           const char *               ggml_backend_buffer_name          (ggml_backend_buffer_t buffer);
    GGML_API           void                       ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
    GGML_API           void *                     ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
    GGML_API           size_t                     ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
    GGML_API GGML_CALL void                       ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
    GGML_API           size_t                     ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
    GGML_API           size_t                     ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
    GGML_API           size_t                     ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
    GGML_API           void                       ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
    GGML_API           bool                       ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
    GGML_API           void                       ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
    GGML_API           ggml_backend_buffer_type_t ggml_backend_buffer_get_type      (ggml_backend_buffer_t buffer);
    GGML_API           void                       ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);

    //
    // Backend
    //

    GGML_API ggml_guid_t  ggml_backend_guid(ggml_backend_t backend);
    GGML_API const char * ggml_backend_name(ggml_backend_t backend);
    GGML_API void         ggml_backend_free(ggml_backend_t backend);

    GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
    GGML_API ggml_backend_buffer_t      ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
    GGML_API size_t                     ggml_backend_get_alignment(ggml_backend_t backend);
    GGML_API size_t                     ggml_backend_get_max_size(ggml_backend_t backend);

    GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
    GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);

    GGML_API GGML_CALL void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
    GGML_API GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);

    GGML_API void ggml_backend_synchronize(ggml_backend_t backend);

    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph);
    GGML_API void                      ggml_backend_graph_plan_free  (ggml_backend_t backend, ggml_backend_graph_plan_t plan);

    GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
    GGML_API enum ggml_status ggml_backend_graph_compute      (ggml_backend_t backend, struct ggml_cgraph * cgraph);
    GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
    GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
    GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op);

    // tensor copy between different backends
    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);

    // asynchronous copy
    // the copy is performed after all the currently queued operations in backend_src
    // backend_dst will wait for the copy to complete before performing other operations
    // automatic fallback to sync copy if async is not supported
    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst);

    // events
    GGML_API ggml_backend_event_t   ggml_backend_event_new        (ggml_backend_t backend);
    GGML_API void                   ggml_backend_event_free       (ggml_backend_event_t event);
    GGML_API void                   ggml_backend_event_record     (ggml_backend_event_t event);
    GGML_API void                   ggml_backend_event_synchronize(ggml_backend_event_t event);
    GGML_API void                   ggml_backend_event_wait       (ggml_backend_t backend, ggml_backend_event_t event); // wait async on event

    //
    // CPU backend
    //

    GGML_API ggml_backend_t ggml_backend_cpu_init(void);

    GGML_API GGML_CALL bool ggml_backend_is_cpu                (ggml_backend_t backend);
    GGML_API           void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
    GGML_API           void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);

    // Create a backend buffer from an existing pointer
    GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);

    GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);

#ifdef GGML_USE_CPU_HBM
    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
#endif

    //
    // Backend registry
    //

    // The backend registry is a registry of all the available backends, and allows initializing backends in a generic way

    GGML_API size_t                     ggml_backend_reg_get_count(void);
    GGML_API size_t                     ggml_backend_reg_find_by_name(const char * name);
    GGML_API ggml_backend_t             ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is name[:params]
    GGML_API const char *               ggml_backend_reg_get_name(size_t i);
    GGML_API ggml_backend_t             ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
    GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
    GGML_API ggml_backend_buffer_t      ggml_backend_reg_alloc_buffer(size_t i, size_t size);

    //
    // Backend scheduler
    //

    // The backend scheduler allows for multiple backends to be used together
    // Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends
    // The backends are selected based on:
    // - the backend that supports the operation
    // - the location of the pre-allocated tensors (e.g. the weights)
    /*
      Example usage:

        // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be assigned
        // preferrably to run on the same backend as the buffer
        ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);

        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false);

        // initialize buffers from a max size graph (optional)
        reserve_graph = build_graph(sched, max_batch_size);

        // manually assign nodes to a backend (optional, should not be needed in most cases)
        struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
        ggml_backend_sched_set_tensor_backend(sched, node, backend_gpu);

        ggml_backend_sched_reserve(sched, reserve_graph);

        // compute
        graph = build_graph(sched);
        ggml_backend_sched_graph_compute(sched, graph);

        // if there are graph inputs:
        ggml_backend_sched_reset(sched);
        ggml_backend_sched_alloc_graph(sched, graph);
        ggml_backend_tensor_set(input_tensor, ...);
        ggml_backend_sched_graph_compute(sched, graph);
    }
    */

    struct ggml_backend_sched;
    typedef struct ggml_backend_sched * ggml_backend_sched_t;

    // when ask == true, the scheduler wants to know if the user wants to observe this node
    // this allows the scheduler to batch nodes together in order to evaluate them in a single call
    //
    // when ask == false, the scheduler is passing the node tensor to the user for observation
    // if the user returns false, the scheduler will cancel the graph compute
    //
    typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);

    // Initialize a backend scheduler
    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
    GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);

    // Initialize backend buffers from a measure graph
    GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);

    // Get the number of splits of the last graph
    GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
    GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);

    GGML_API size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);

    GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
    GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);

    // Allocate and compute graph on the backend scheduler
    GGML_API bool                 ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
    GGML_API enum ggml_status     ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
    GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
    GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);

    // Reset all assignments and allocators - must be called before changing the node backends
    GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);

    // Set a callback to be called for each resulting node during graph compute
    GGML_API void                 ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);

    //
    // Utils
    //

    struct ggml_backend_graph_copy {
        ggml_backend_buffer_t buffer;
        struct ggml_context * ctx_allocated;
        struct ggml_context * ctx_unallocated;
        struct ggml_cgraph * graph;
    };

    // Copy a graph to a different backend
    GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
    GGML_API void                           ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);

    typedef bool (*GGML_CALL ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);

    // Compare the output of two backends
    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);

    // Tensor initialization
    GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
    GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);


#ifdef  __cplusplus
}
#endif