llama.cpp/ggml-backend-impl.h

#pragma once

// ggml-backend internal header

#include "ggml-backend.h"

#ifdef  __cplusplus
extern "C" {
#endif

    //
    // Backend buffer
    //

    typedef void * ggml_backend_buffer_context_t;

    struct ggml_backend_buffer_i {
        void   (*free_buffer)   (ggml_backend_buffer_t buffer);
        void * (*get_base)      (ggml_backend_buffer_t buffer); // get base pointer
        size_t (*get_alloc_size)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-allocation callback
        void   (*init_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // post-allocation callback
        void   (*free_tensor)   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor); // pre-free callback
    };

    struct ggml_backend_buffer {
        struct ggml_backend_buffer_i iface;

        ggml_backend_t                backend;
        ggml_backend_buffer_context_t context;

        size_t size;
    };

    GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
            struct ggml_backend                  * backend,
            struct ggml_backend_buffer_i           iface,
                   ggml_backend_buffer_context_t   context,
                   size_t                          size);

    //
    // Backend
    //

    typedef void * ggml_backend_context_t;

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

        void (*free)(ggml_backend_t backend);

        // buffer allocation
        ggml_backend_buffer_t (*alloc_buffer)(ggml_backend_t backend, size_t size);

        // get buffer alignment
        size_t (*get_alignment)(ggml_backend_t backend);

        // tensor data access
        // these functions can be asynchronous, helper functions are provided for synchronous access that automatically call synchronize
        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);
        void (*synchronize)     (ggml_backend_t backend);

        // (optional) copy tensor between different backends, allow for single-copy tranfers
        void (*cpy_tensor_from)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
        void (*cpy_tensor_to)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);

        // compute graph with a plan
        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
        void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
        void                      (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);

        // compute graph without a plan
        void (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);

        // check if the backend supports an operation
        bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
    };

    struct ggml_backend {
        struct ggml_backend_i iface;

        ggml_backend_context_t context;
    };

#ifdef  __cplusplus
}
#endif