Method eval_perf

smallthinker/tests/test-backend-ops.cpp:596–747

    }

    bool eval_perf(ggml_backend_t backend, const char * op_name) {
        mode = MODE_PERF;

        static const size_t graph_nodes = 8192;

        ggml_init_params params = {
            /* .mem_size = */ ggml_tensor_overhead()*128 + ggml_graph_overhead_custom(graph_nodes, false),
            /* .mem_base = */ NULL,
            /* .no_alloc = */ true,
        };
        ggml_context_ptr ctx(ggml_init(params)); // smart ptr
        GGML_ASSERT(ctx);

        ggml_tensor * out = build_graph(ctx.get());

        if (op_name != nullptr && op_desc(out) != op_name) {
            //printf("  %s: skipping\n", op_desc(out).c_str());
            return true;
        }

        int len = printf("  %s(%s): ", op_desc(out).c_str(), vars().c_str());
        fflush(stdout);

        // check if backends support op
        if (!ggml_backend_supports_op(backend, out)) {
            printf("not supported\n");
            return true;
        }

        // align while also leaving some margin for variations in parameters
        int align = 8;
        int last = (len + align - 1) / align * align;
        if (last - len < 5) {
            last += align;
        }
        printf("%*s", last - len, "");

        // allocate
        ggml_backend_buffer_ptr buf(ggml_backend_alloc_ctx_tensors(ctx.get(), backend)); // smart ptr

        if (buf == NULL) {
            printf("failed to allocate tensors\n");
            return false;
        }

        // randomize tensors
        initialize_tensors(ctx.get());

        // build graph
        ggml_cgraph * gf = ggml_new_graph_custom(ctx.get(), graph_nodes, false);
        ggml_build_forward_expand(gf, out);

        // warmup run
        ggml_status status = ggml_backend_graph_compute(backend, gf);
        if (status != GGML_STATUS_SUCCESS) {
            fprintf(stderr, "%s: ggml_backend_graph_compute failed. status=%s \n", __func__, ggml_status_to_string(status));
            return false;
        }