Function ggml_backend_sched_new

external/ggml/src/ggml-backend.cpp:1727–1794

}

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,
        bool op_offload) {
    GGML_ASSERT(n_backends > 0);
    GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
    GGML_ASSERT(ggml_backend_dev_type(ggml_backend_get_device(backends[n_backends - 1])) == GGML_BACKEND_DEVICE_TYPE_CPU);

    struct ggml_backend_sched * sched = (ggml_backend_sched *) calloc(1, sizeof(struct ggml_backend_sched));

    const char * GGML_SCHED_DEBUG = getenv("GGML_SCHED_DEBUG");
    sched->debug = GGML_SCHED_DEBUG ? atoi(GGML_SCHED_DEBUG) : 0;

    sched->debug_realloc = 0;
#ifdef GGML_SCHED_NO_REALLOC
    sched->debug_realloc = 1;
#endif
    const char * GGML_SCHED_DEBUG_REALLOC = getenv("GGML_SCHED_DEBUG_REALLOC");
    sched->debug_realloc = GGML_SCHED_DEBUG_REALLOC ? atoi(GGML_SCHED_DEBUG_REALLOC) : sched->debug_realloc;

    sched->n_backends = n_backends;
    sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;

    // initialize hash table
    // FIXME: needs to be size*2 to account for leafs (do it in graph_split instead)
    sched->hash_set    = ggml_hash_set_new(graph_size);
    sched->hv_tensor_backend_ids = (int *) malloc(sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0]));
    sched->hv_tensor_copies      = (ggml_tensor **) malloc(sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *));

    const size_t ggml_sched_max_splits = graph_size; // at most there is one split for each node in the graph
    const size_t nodes_size = graph_size + ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2;
    sched->node_backend_ids = (int *) calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
    sched->leaf_backend_ids = (int *) calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
    sched->prev_node_backend_ids = (int *) calloc(nodes_size, sizeof(sched->prev_node_backend_ids[0]));
    sched->prev_leaf_backend_ids = (int *) calloc(nodes_size, sizeof(sched->prev_leaf_backend_ids[0]));

    sched->debug_graph_size = 0;
    sched->debug_prev_graph_size = 0;

    sched->context_buffer_size = ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + ggml_graph_overhead_custom(graph_size, false);
    sched->context_buffer = (char *) malloc(sched->context_buffer_size);

    const int initial_splits_capacity = 16;
    sched->splits = (ggml_backend_sched_split *) calloc(initial_splits_capacity, sizeof(sched->splits[0]));
    sched->splits_capacity = initial_splits_capacity;

    for (int b = 0; b < n_backends; b++) {
        sched->backends[b] = backends[b];
        sched->bufts[b] = bufts ? bufts[b] : ggml_backend_get_default_buffer_type(backends[b]);
        GGML_ASSERT(ggml_backend_supports_buft(backends[b], sched->bufts[b]));

        if (sched->n_copies > 1) {
            for (int c = 0; c < sched->n_copies; c++) {
                sched->events[b][c] = ggml_backend_event_new(backends[b]->device);
            }