Function sched_split_graph

ggml-backend.c:581–791

assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend TODO: merge passes

// assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend
// TODO: merge passes
static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
    // reset state
    size_t hash_size = sched->hash_set.size;
    memset(sched->hash_set.keys, 0, sizeof(sched->hash_set.keys[0]) * hash_size);
    memset(sched->node_talloc,   0, sizeof(sched->node_talloc[0])   * hash_size);
    memset(sched->node_copies,   0, sizeof(sched->node_copies[0])   * hash_size);
    sched->n_splits = 0;

    struct ggml_init_params params = {
        /*.mem_size =   */ sizeof(sched->context_buffer),
        /*.mem_buffer = */ sched->context_buffer,
        /*.no_alloc =   */ true
    };

    if (sched->ctx != NULL) {
        ggml_free(sched->ctx);
    }

    sched->ctx = ggml_init(params);

    // pass 1: assign backends to ops with allocated inputs
    for (int i = 0; i < graph->n_leafs; i++) {
        struct ggml_tensor * leaf = graph->leafs[i];
        if (node_allocr(leaf) != NULL) {
            // do not overwrite user assignments
            continue;
        }
        ggml_backend_t leaf_backend = ggml_get_backend(leaf);
        if (leaf_backend == NULL && leaf->view_src != NULL) {
            leaf_backend = ggml_get_backend(leaf->view_src);
        }
        if (leaf_backend != NULL) {
            node_allocr(leaf) = ggml_backend_sched_get_tallocr(sched, leaf_backend);
        }
    }

    for (int i = 0; i < graph->n_nodes; i++) {
        struct ggml_tensor * node = graph->nodes[i];
        if (node_allocr(node) != NULL) {
            // do not overwrite user assignments
            continue;
        }
        ggml_backend_t node_backend = sched_backend_from_cur(sched, node);
        if (node_backend != NULL) {
            node_allocr(node) = ggml_backend_sched_get_tallocr(sched, node_backend);
        }
    }
    //printf("PASS 1 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);

    // pass 2: assign backends to ops from current assignments
    // TODO:
    //  - reuse sched_backend_from_cur
    for (int i = 0; i < graph->n_nodes; i++) {
        struct ggml_tensor * node = graph->nodes[i];
        ggml_tallocr_t node_allocr = node_allocr(node);
        if (node_allocr == NULL) {
            int    cur_prio = INT_MAX;
            size_t cur_size = 0;