Function ggml_graph_compute_thread

ggml.c:16857–16977

}

static thread_ret_t ggml_graph_compute_thread(void * data) {
    struct ggml_compute_state * state = (struct ggml_compute_state *) data;

    const struct ggml_cgraph * cgraph = state->shared->cgraph;
    const struct ggml_cplan  * cplan  = state->shared->cplan;

    const int   n_threads   = state->shared->n_threads;

    set_numa_thread_affinity(state->ith, n_threads);

    int node_n = -1;

    while (true) {
        if (cplan->abort_callback && cplan->abort_callback(cplan->abort_callback_data)) {
            state->shared->node_n += 1;
            return (thread_ret_t) GGML_EXIT_ABORTED;
        }
        if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) {
            // all other threads are finished and spinning
            // do finalize and init here so we don't have synchronize again
            struct ggml_compute_params params = {
                /*.type  =*/ GGML_TASK_FINALIZE,
                /*.ith   =*/ 0,
                /*.nth   =*/ 0,
                /*.wsize =*/ cplan->work_size,
                /*.wdata =*/ cplan->work_data,
                /*.aic   =*/ &state->shared->aic,
            };

            if (node_n != -1) {
                /* FINALIZE */
                struct ggml_tensor * node = cgraph->nodes[node_n];
                if (GGML_OP_HAS_FINALIZE[node->op]) {
                    params.nth = ggml_get_n_tasks(node, n_threads);
                    ggml_compute_forward(&params, node);
                }
                ggml_graph_compute_perf_stats_node(node, state->shared);
            }

            // distribute new work or execute it direct if 1T
            while (++node_n < cgraph->n_nodes) {
                GGML_PRINT_DEBUG_5("%s: %d/%d\n", __func__, node_n, cgraph->n_nodes);

                struct ggml_tensor * node = cgraph->nodes[node_n];
                const int n_tasks = ggml_get_n_tasks(node, n_threads);

                state->shared->perf_node_start_cycles  = ggml_perf_cycles();
                state->shared->perf_node_start_time_us = ggml_perf_time_us();

                params.nth = n_tasks;

                /* INIT */
                if (GGML_OP_HAS_INIT[node->op]) {
                    params.type = GGML_TASK_INIT;
                    ggml_compute_forward(&params, node);
                }

                if (n_tasks == 1) {