Function run_graph

tests/test_metal_overhead.cpp:18–48

#include <vector>

static double run_graph(ggml_backend_t backend, ggml_context* ctx, ggml_cgraph* graph,
                        const std::vector<ggml_tensor*>& inputs, int warmup, int repeats) {
    auto* galloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(backend));
    if (!ggml_gallocr_reserve(galloc, graph) || !ggml_gallocr_alloc_graph(galloc, graph)) {
        fprintf(stderr, "FAIL: alloc\n");
        ggml_gallocr_free(galloc);
        return -1;
    }
    for (auto* t : inputs) {
        int64_t nel = ggml_nelements(t);
        if (t->type == GGML_TYPE_F32) {
            std::vector<float> d(nel, 0.1f);
            ggml_backend_tensor_set(t, d.data(), 0, nel * sizeof(float));
        } else if (t->type == GGML_TYPE_F16) {
            std::vector<uint16_t> d(nel);
            for (int64_t j = 0; j < nel; j++) d[j] = ggml_fp32_to_fp16(0.1f);
            ggml_backend_tensor_set(t, d.data(), 0, nel * sizeof(uint16_t));
        }
    }
    if (ggml_backend_is_cpu(backend))
        ggml_backend_cpu_set_n_threads(backend, 8);

    for (int i = 0; i < warmup; i++)
        ggml_backend_graph_compute(backend, graph);
    auto t0 = std::chrono::high_resolution_clock::now();
    for (int i = 0; i < repeats; i++)
        ggml_backend_graph_compute(backend, graph);
    auto t1 = std::chrono::high_resolution_clock::now();
    ggml_gallocr_free(galloc);
    return std::chrono::duration<double, std::milli>(t1 - t0).count() / repeats;
}

// ═══════════════════════════════════════════════════════════════════════════════
//  TEST 1: Pure dispatch overhead — chain of N trivial add ops