Function load_model

external/ggml/tests/test-conv2d.cpp:38–138

};

void load_model(test_model & model, bool use_gpu = false) {
    // create data
    int KW = 3, KH = 3, IC = 10, OC = 10;
    int IW = 8, IH = 6, N = 1;

    // Initialize adata
    std::vector<float> adata(KW * KH * IC * OC);
    for (int i = 0; i < KW * KH * IC * OC; i++) {
        adata[i] = 2.5f;
    }

    // Convert adata to fp16 format
    std::vector<ggml_fp16_t> hadata(KW * KH * IC * OC);
    ggml_fp32_to_fp16_row(adata.data(), hadata.data(), KW * KH * IC * OC);

    // Initialize bdata
    std::vector<float> bdata(IW * IH * IC * N);
    for (int i = 0; i < IW * IH * IC * N; i++) {
        bdata[i] = 1.5f;
    }

    size_t buffer_size = 0;
    {
        buffer_size += KW * KH * IC * OC * ggml_type_size(GGML_TYPE_F16); // tensor a
        buffer_size += IW * IH * IC * N  * ggml_type_size(GGML_TYPE_F32); // tensor b
        buffer_size += 1024; // overhead
    }

    printf("%s: ggml tensor size    = %d bytes\n", __func__, (int) sizeof(ggml_tensor));
    printf("%s: backend buffer size = %0.2f MB\n", __func__, (buffer_size/ 1024.f/ 1024.f));

    int num_tensors = 2;
    struct ggml_init_params params {
            /*.mem_size   =*/ ggml_tensor_overhead() * num_tensors,
            /*.mem_buffer =*/ NULL,
            /*.no_alloc   =*/ true,
    };

    ggml_log_set(ggml_log_callback_default, nullptr);

    // initialize the backend
#ifdef GGML_USE_CUDA
    if (use_gpu) {
        fprintf(stderr, "%s: using CUDA backend\n", __func__);
        model.backend = ggml_backend_cuda_init(0);
        if (!model.backend) {
            fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
        }
    }
#endif

#ifdef GGML_USE_METAL
    if (use_gpu) {
        fprintf(stderr, "%s: using Metal backend\n", __func__);
        model.backend = ggml_backend_metal_init();
        if (!model.backend) {
            fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
        }