llama.cpp学习笔记：后端加载

单例

struct ggml_backend_registry {std::vector<ggml_backend_reg_entry> backends;std::vector<ggml_backend_dev_t> devices;// ...
}struct ggml_backend_reg_entry {ggml_backend_reg_t reg;dl_handle_ptr handle;
};typedef struct ggml_backend_reg * ggml_backend_reg_t;typedef struct ggml_backend_device * ggml_backend_dev_t;
struct ggml_backend_device {struct ggml_backend_device_i iface;ggml_backend_reg_t reg;void * context;
};

每种后端一个单例：cuda、cpu等

版本号、接口和上下文信息

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

后端对应的设备：接口、所属后端和上下文信息

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

cuda后端

cuda后端iface

static const ggml_backend_reg_i ggml_backend_cuda_reg_interface = {/* .get_name          = */ ggml_backend_cuda_reg_get_name,/* .get_device_count  = */ ggml_backend_cuda_reg_get_device_count,/* .get_device        = */ ggml_backend_cuda_reg_get_device,/* .get_proc_address  = */ ggml_backend_cuda_reg_get_proc_address,
};

cuda后端context，存储多个cuda设备（GPU）的接口、所属后端和上下文等信息

struct ggml_backend_cuda_reg_context {std::vector<ggml_backend_dev_t> devices;
};

 cuda设备GPU的接口iface

static const ggml_backend_device_i ggml_backend_cuda_device_interface = {/* .get_name                = */ ggml_backend_cuda_device_get_name,/* .get_description         = */ ggml_backend_cuda_device_get_description,/* .get_memory              = */ ggml_backend_cuda_device_get_memory,/* .get_type                = */ ggml_backend_cuda_device_get_type,/* .get_props               = */ ggml_backend_cuda_device_get_props,/* .init_backend            = */ ggml_backend_cuda_device_init_backend,/* .get_buffer_type         = */ ggml_backend_cuda_device_get_buffer_type,/* .get_host_buffer_type    = */ ggml_backend_cuda_device_get_host_buffer_type,/* .buffer_from_host_ptr    = */ NULL,/* .supports_op             = */ ggml_backend_cuda_device_supports_op,/* .supports_buft           = */ ggml_backend_cuda_device_supports_buft,/* .offload_op              = */ ggml_backend_cuda_device_offload_op,/* .event_new               = */ ggml_backend_cuda_device_event_new,/* .event_free              = */ ggml_backend_cuda_device_event_free,/* .event_synchronize       = */ ggml_backend_cuda_device_event_synchronize,
};

cuda设备GPU的context

struct ggml_backend_cuda_device_context {int device;std::string name;std::string description;
};

 

CPU后端

CPU后端iface

static const struct ggml_backend_reg_i ggml_backend_cpu_reg_i = {/* .get_name         = */ ggml_backend_cpu_reg_get_name,/* .get_device_count = */ ggml_backend_cpu_reg_get_device_count,/* .get_device       = */ ggml_backend_cpu_reg_get_device,/* .get_proc_address = */ ggml_backend_cpu_get_proc_address,
};

CPU后端无context

CPU设备接口iface

static const struct ggml_backend_device_i ggml_backend_cpu_device_i = {/* .get_name             = */ ggml_backend_cpu_device_get_name,/* .get_description      = */ ggml_backend_cpu_device_get_description,/* .get_memory           = */ ggml_backend_cpu_device_get_memory,/* .get_type             = */ ggml_backend_cpu_device_get_type,/* .get_props            = */ ggml_backend_cpu_device_get_props,/* .init_backend         = */ ggml_backend_cpu_device_init_backend,/* .get_buffer_type      = */ ggml_backend_cpu_device_get_buffer_type,/* .get_host_buffer_type = */ NULL,/* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_host_ptr,/* .supports_op          = */ ggml_backend_cpu_device_supports_op,/* .supports_buft        = */ ggml_backend_cpu_device_supports_buft,/* .offload_op           = */ NULL,/* .event_new            = */ NULL,/* .event_free           = */ NULL,/* .event_synchronize    = */ NULL,
};

CPU设备CPU的context

struct ggml_backend_cpu_device_context {std::string description = "CPU";// ...
}