test_network.cpp 46.8 KB
Newer Older
1 2
/**
 * \file test/test_network.cpp
3
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
4
 *
5
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
6
 *
7 8 9
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
10 11 12 13 14 15 16 17
 */

#include "lite_build_config.h"

#if LITE_BUILD_WITH_MGE
#include "./test_common.h"
#include "megbrain/tensor.h"

M
Megvii Engine Team 已提交
18 19 20 21 22
#ifndef WIN32
#include <dirent.h>
#include <string.h>
#endif

23 24 25 26 27 28 29 30 31 32
#include <chrono>
#include <memory>
#include <random>
#include <unordered_map>
using namespace lite;

namespace {
class CheckAllocator : public lite::Allocator {
public:
    //! allocate memory of size in the given device with the given align
M
Megvii Engine Team 已提交
33
    void* allocate(LiteDeviceType device, int, size_t size, size_t align) override {
34 35 36 37 38 39 40 41 42 43
        LITE_ASSERT(device == LiteDeviceType::LITE_CPU);
        m_nr_left++;
        m_nr_allocated++;
#ifdef WIN32
        return _aligned_malloc(size, align);
#elif defined(__ANDROID__) || defined(ANDROID)
        return memalign(align, size);
#else
        void* ptr = nullptr;
        auto err = posix_memalign(&ptr, align, size);
M
Megvii Engine Team 已提交
44
        mgb_assert(!err, "failed to malloc %zubytes with align %zu", size, align);
45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97
        return ptr;
#endif
    };

    //! free the memory pointed by ptr in the given device
    void free(LiteDeviceType device, int, void* ptr) override {
        m_nr_left--;
        LITE_ASSERT(device == LiteDeviceType::LITE_CPU);
#ifdef WIN32
        _aligned_free(ptr);
#else
        ::free(ptr);
#endif
    };
    std::atomic_size_t m_nr_left{0};
    std::atomic_size_t m_nr_allocated{0};
};
}  // namespace

TEST(TestNetWork, Basic) {
    Config config;
    auto lite_tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    auto result_lite = mgelite_lar(model_path, config, "data", lite_tensor);
    auto result_mgb = mgb_lar(model_path, config, "data", lite_tensor);
    compare_lite_tensor<float>(result_lite, result_mgb);
}

TEST(TestNetWork, SetDeviceId) {
    Config config;
    auto lite_tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::shared_ptr<Network> network = std::make_shared<Network>(config);
    network->set_device_id(4);
    network->load_model(model_path);
    std::shared_ptr<Tensor> input_tensor = network->get_input_tensor(0);
    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);
    network->forward();
    network->wait();
    ASSERT_EQ(input_tensor->get_device_id(), 4);
    ASSERT_EQ(output_tensor->get_device_id(), 4);
}

TEST(TestNetWork, GetAllName) {
    Config config;
    auto lite_tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::shared_ptr<Network> network = std::make_shared<Network>(config);

    network->load_model(model_path);
    auto input_names = network->get_all_input_name();
    auto output_names = network->get_all_output_name();

98 99 100 101 102
    auto output_tensor = network->get_output_tensor(0);
    auto out_layout = output_tensor->get_layout();
    ASSERT_EQ(out_layout.ndim, 2);
    ASSERT_EQ(out_layout.shapes[0], 1);
    ASSERT_EQ(out_layout.shapes[1], 1000);
103 104 105
    ASSERT_EQ(input_names.size(), 1);
    ASSERT_EQ(output_names.size(), 1);
    ASSERT_TRUE(input_names[0] == "data");
M
Megvii Engine Team 已提交
106
    ASSERT_TRUE(output_names[0] == "TRUE_DIV(EXP[12065],reduce0[12067])[12077]");
107 108
}

109 110 111 112 113 114 115 116 117 118 119 120 121 122 123
TEST(TestNetWork, LoadFBSModel) {
    Config config;
    std::string model_path = "./ax.mge";
    std::shared_ptr<Network> network = std::make_shared<Network>(config);
    network->load_model(model_path);

    auto output_tensor = network->get_output_tensor(0);
    auto out_layout = output_tensor->get_layout();
    ASSERT_EQ(out_layout.ndim, 4);
    ASSERT_EQ(out_layout.shapes[0], 1);
    ASSERT_EQ(out_layout.shapes[1], 1);
    ASSERT_EQ(out_layout.shapes[2], 40);
    ASSERT_EQ(out_layout.shapes[3], 180);
}

124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189
TEST(TestNetWork, BasicInplaceAndSingleThreadAffinity) {
    Config config;
    auto lite_tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";

    auto result_mgb = mgb_lar(model_path, config, "data", lite_tensor);

    std::shared_ptr<Network> network = std::make_shared<Network>(config);
    Runtime::set_cpu_inplace_mode(network);

    network->load_model(model_path);
    std::shared_ptr<Tensor> input_tensor = network->get_input_tensor(0);

    int affinity_set = false;
    Runtime::set_runtime_thread_affinity(network, [&affinity_set](int id) {
        ASSERT_EQ(id, 0);
        affinity_set = true;
    });

    auto src_ptr = lite_tensor->get_memory_ptr();
    auto src_layout = lite_tensor->get_layout();
    input_tensor->reset(src_ptr, src_layout);

    //! inplace mode not support async mode
    ASSERT_THROW(network->set_async_callback([]() {}), std::exception);

    network->forward();
    network->wait();
    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);

    ASSERT_EQ(affinity_set, true);
    compare_lite_tensor<float>(output_tensor, result_mgb);
}

TEST(TestNetWork, NetworkShareWeights) {
    Config config;
    auto lite_tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";

    auto result_mgb = mgb_lar(model_path, config, "data", lite_tensor);

    std::shared_ptr<Network> network = std::make_shared<Network>(config);
    network->load_model(model_path);
    std::shared_ptr<Tensor> input_tensor = network->get_input_tensor(0);

    std::shared_ptr<Network> network2 = std::make_shared<Network>(config);
    Runtime::set_cpu_inplace_mode(network2);

    Runtime::shared_weight_with_network(network2, network);

    std::shared_ptr<Tensor> input_tensor2 = network2->get_input_tensor(0);

    auto src_ptr = lite_tensor->get_memory_ptr();
    auto src_layout = lite_tensor->get_layout();
    input_tensor->reset(src_ptr, src_layout);
    input_tensor2->reset(src_ptr, src_layout);
    ASSERT_NE(input_tensor, input_tensor2);

    network->forward();
    network->wait();

    network2->forward();
    network2->wait();
    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);
    std::shared_ptr<Tensor> output_tensor2 = network2->get_output_tensor(0);

M
Megvii Engine Team 已提交
190
    ASSERT_NE(output_tensor->get_memory_ptr(), output_tensor2->get_memory_ptr());
191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273
    compare_lite_tensor<float>(output_tensor, result_mgb);
    compare_lite_tensor<float>(output_tensor2, result_mgb);
}

TEST(TestNetWork, SharedRuntimeMem) {
    Config config;
    auto lite_tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";

    auto result_mgb = mgb_lar(model_path, config, "data", lite_tensor);

    std::shared_ptr<Network> network_src = std::make_shared<Network>(config);
    std::shared_ptr<Network> network_dst = std::make_shared<Network>(config);
    Runtime::share_runtime_memory_with(network_dst, network_src);
    network_src->load_model(model_path);
    network_dst->load_model(model_path);
}

TEST(TestNetWork, UserAllocator) {
    auto allocator = std::make_shared<CheckAllocator>();
    {
        Config config;
        auto lite_tensor = get_input_data("./input_data.npy");
        std::string model_path = "./shufflenet.mge";

        auto result_mgb = mgb_lar(model_path, config, "data", lite_tensor);
        std::shared_ptr<Network> network = std::make_shared<Network>(config);

        Runtime::set_memory_allocator(network, allocator);

        network->load_model(model_path);
        std::shared_ptr<Tensor> input_tensor = network->get_input_tensor(0);

        auto src_ptr = lite_tensor->get_memory_ptr();
        auto src_layout = lite_tensor->get_layout();
        input_tensor->reset(src_ptr, src_layout);

        network->forward();
        network->wait();

        ASSERT_GE(allocator->m_nr_allocated, 1);
        std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);

        compare_lite_tensor<float>(output_tensor, result_mgb);
    }
    ASSERT_EQ(allocator->m_nr_left, 0);
}

TEST(TestNetWork, BasicMultiThread) {
    Config config;
    auto lite_tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";

    auto result_mgb = mgb_lar(model_path, config, "data", lite_tensor);

    std::shared_ptr<Network> network = std::make_shared<Network>(config);
    Runtime::set_cpu_threads_number(network, 2);

    network->load_model(model_path);
    std::shared_ptr<Tensor> input_tensor = network->get_input_tensor(0);

    auto src_ptr = lite_tensor->get_memory_ptr();
    auto src_layout = lite_tensor->get_layout();
    input_tensor->reset(src_ptr, src_layout);

    network->forward();
    network->wait();
    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);

    compare_lite_tensor<float>(output_tensor, result_mgb);
}

TEST(TestNetWork, ThreadAffinity) {
    size_t nr_threads = 4;
    Config config;
    auto lite_tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";

    auto result_mgb = mgb_lar(model_path, config, "data", lite_tensor);

    std::shared_ptr<Network> network = std::make_shared<Network>(config);
    Runtime::set_cpu_threads_number(network, nr_threads);

M
Megvii Engine Team 已提交
274 275
    ASSERT_THROW(
            Runtime::set_runtime_thread_affinity(network, [](int) {}), std::exception);
276 277
    network->load_model(model_path);
    std::vector<std::thread::id> thread_ids(nr_threads);
M
Megvii Engine Team 已提交
278
    auto affinity = [&](int id) { thread_ids[id] = std::this_thread::get_id(); };
279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305
    Runtime::set_runtime_thread_affinity(network, affinity);

    std::shared_ptr<Tensor> input_tensor = network->get_input_tensor(0);
    auto src_ptr = lite_tensor->get_memory_ptr();
    auto src_layout = lite_tensor->get_layout();
    input_tensor->reset(src_ptr, src_layout);

    network->forward();
    network->wait();

    for (size_t i = 0; i < nr_threads; i++) {
        for (size_t j = i + 1; j < nr_threads; j++) {
            ASSERT_NE(thread_ids[i], thread_ids[j]);
        }
    }

    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);
    compare_lite_tensor<float>(output_tensor, result_mgb);
}

TEST(TestNetWork, BasicCryptAes) {
    Config config;
    auto lite_tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::string model_crypt_path = "./shufflenet_crypt_aes.mge";
    auto result_mgb = mgb_lar(model_path, config, "data", lite_tensor);
    config.bare_model_cryption_name = "AES_default";
M
Megvii Engine Team 已提交
306
    auto result_lite = mgelite_lar(model_crypt_path, config, "data", lite_tensor);
307 308 309 310 311 312 313 314 315 316
    compare_lite_tensor<float>(result_lite, result_mgb);
}

TEST(TestNetWork, BasicCryptRc4) {
    Config config;
    auto lite_tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::string model_crypt_path = "./shufflenet_crypt_rc4.mge";
    auto result_mgb = mgb_lar(model_path, config, "data", lite_tensor);
    config.bare_model_cryption_name = "RC4_default";
M
Megvii Engine Team 已提交
317
    auto result_lite = mgelite_lar(model_crypt_path, config, "data", lite_tensor);
318 319 320 321 322 323 324 325 326
    compare_lite_tensor<float>(result_lite, result_mgb);
}

TEST(TestNetWork, PackedCryptRc4) {
    Config config;
    auto lite_tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::string model_crypt_path = "./test_packed_model_rc4.lite";
    auto result_mgb = mgb_lar(model_path, config, "data", lite_tensor);
M
Megvii Engine Team 已提交
327
    auto result_lite = mgelite_lar(model_crypt_path, config, "data", lite_tensor);
328 329 330 331 332 333 334 335 336 337
    compare_lite_tensor<float>(result_lite, result_mgb);
}

TEST(TestNetWork, BasicCryptSfRc4) {
    Config config;
    auto lite_tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::string model_crypt_path = "./shufflenet_crypt_sfrc4.mge";
    auto result_mgb = mgb_lar(model_path, config, "data", lite_tensor);
    config.bare_model_cryption_name = "SIMPLE_FAST_RC4_default";
M
Megvii Engine Team 已提交
338
    auto result_lite = mgelite_lar(model_crypt_path, config, "data", lite_tensor);
339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408
    compare_lite_tensor<float>(result_lite, result_mgb);
}

TEST(TestNetWork, ResetInput) {
    Config config;
    auto tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";
    auto result_mgb = mgb_lar(model_path, config, input_name, tensor);

    std::shared_ptr<Network> network = std::make_shared<Network>(config);

    network->load_model(model_path);
    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);

    auto src_ptr = tensor->get_memory_ptr();
    auto src_layout = tensor->get_layout();
    input_tensor->reset(src_ptr, src_layout);

    network->forward();
    network->wait();
    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);

    compare_lite_tensor<float>(output_tensor, result_mgb);
}

TEST(TestNetWork, ChangeInputShape) {
    Config config;
    auto tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";
    auto result_mgb = mgb_lar(model_path, config, input_name, tensor);

    std::shared_ptr<Network> network = std::make_shared<Network>(config);

    network->load_model(model_path);
    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);

    auto src_layout = Layout{{2, 3, 200, 200}, 4, LiteDataType::LITE_FLOAT};
    input_tensor->set_layout(src_layout);
    std::shared_ptr<Tensor> input_tensor2 = network->get_io_tensor(input_name);
    //! Check memory is equal
    ASSERT_EQ(input_tensor->get_memory_ptr(), input_tensor2->get_memory_ptr());

    network->forward();
    network->wait();
    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);
    auto output_layout = output_tensor->get_layout();
    ASSERT_EQ(output_layout.shapes[0], 2);
    ASSERT_EQ(output_layout.shapes[1], 1000);
}

TEST(TestNetWork, ResetOutput) {
    Config config;
    auto tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";
    auto result_mgb = mgb_lar(model_path, config, input_name, tensor);

    std::shared_ptr<Network> network = std::make_shared<Network>(config);

    network->load_model(model_path);
    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);

    auto src_ptr = tensor->get_memory_ptr();
    auto src_layout = tensor->get_layout();
    input_tensor->reset(src_ptr, src_layout);

    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);
    auto result_tensor = std::make_shared<Tensor>(
M
Megvii Engine Team 已提交
409
            LiteDeviceType::LITE_CPU, Layout{{1, 1000}, 2, LiteDataType::LITE_FLOAT});
410 411 412 413 414 415 416 417 418 419

    void* out_data = result_tensor->get_memory_ptr();
    output_tensor->reset(out_data, result_tensor->get_layout());

    network->forward();
    network->wait();

    compare_lite_tensor<float>(output_tensor, result_mgb);
}

420 421 422
namespace {

void test_output_no_copy(int record) {
423 424
    Config config;
    config.options.force_output_use_user_specified_memory = true;
425
    config.options.comp_node_seq_record_level = record;
426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463
    auto tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";
    auto result_mgb = mgb_lar(model_path, config, input_name, tensor);

    std::shared_ptr<Network> network = std::make_shared<Network>(config);

    network->load_model(model_path);
    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);

    auto src_ptr = tensor->get_memory_ptr();
    auto src_layout = tensor->get_layout();
    input_tensor->reset(src_ptr, src_layout);

    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);
    size_t times = 5;
    std::vector<std::shared_ptr<Tensor>> result_tensors;
    for (size_t i = 0; i < times; i++) {
        auto tmp = std::make_shared<Tensor>(
                LiteDeviceType::LITE_CPU,
                Layout{{1, 1000}, 2, LiteDataType::LITE_FLOAT});
        result_tensors.push_back(tmp);
    }

    for (size_t i = 0; i < times; i++) {
        void* out_data = result_tensors[i]->get_memory_ptr();
        output_tensor->reset(out_data, result_tensors[i]->get_layout());

        network->forward();
        network->wait();
        ASSERT_EQ(output_tensor->get_memory_ptr(), out_data);
        compare_lite_tensor<float>(output_tensor, result_mgb);
    }
    for (size_t i = 0; i < times; i++) {
        compare_lite_tensor<float>(result_tensors[i], result_mgb);
    }
}

464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504
void test_input_no_copy(int record) {
    Config config;
    config.options.force_output_use_user_specified_memory = true;
    config.options.comp_node_seq_record_level = record;
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";

    Layout layout_in{{1, 3, 224, 224}, 4};
    std::vector<std::shared_ptr<Tensor>> inputs;
    std::vector<std::shared_ptr<Tensor>> outputs;
    for (int i = 0; i < 3; i++) {
        auto tmp_in = std::make_shared<Tensor>(LiteDeviceType::LITE_CPU, layout_in);

        auto ptr = static_cast<float*>(tmp_in->get_memory_ptr());
        for (size_t id = 0; id < 2 * 224 * 224; id++) {
            ptr[id] = i + 1;
        }
        inputs.push_back(tmp_in);
        outputs.push_back(mgb_lar(model_path, config, input_name, tmp_in));
    }

    std::shared_ptr<Network> network = std::make_shared<Network>(config);

    network->load_model(model_path);
    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);
    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);

    for (int i = 0; i < 3; i++) {
        auto ptr = inputs[i]->get_memory_ptr();
        input_tensor->reset(ptr, layout_in);

        auto tmp_out = std::make_shared<Tensor>(
                LiteDeviceType::LITE_CPU,
                Layout{{1, 1000}, 2, LiteDataType::LITE_FLOAT});
        output_tensor->reset(tmp_out->get_memory_ptr(), output_tensor->get_layout());

        network->forward();
        network->wait();
        compare_lite_tensor<float>(output_tensor, outputs[i]);
    }
}
M
Megvii Engine Team 已提交
505 506 507 508 509 510 511 512

void test_io_no_copy_ax(std::string model_name, int record = 1) {
    std::string model_path = model_name;
    std::vector<std::string> input_names, output_names;

    std::vector<std::vector<std::shared_ptr<Tensor>>> inputs;
    std::vector<std::vector<std::shared_ptr<Tensor>>> outputs;

513 514 515 516
    Config config;

    config.options.graph_opt_level = 0;
    std::shared_ptr<Network> network = std::make_shared<Network>(config);
M
Megvii Engine Team 已提交
517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567
    network->load_model(model_path);

    input_names = network->get_all_input_name();
    output_names = network->get_all_output_name();

    // prepare test data
    for (int i = 0; i < 3; i++) {
        std::vector<std::shared_ptr<Tensor>> net_inputs;
        std::vector<std::shared_ptr<Tensor>> net_outputs;

        for (size_t j = 0; j < input_names.size(); j++) {
            auto in_tesnor = network->get_io_tensor(input_names[j]);
            auto in_layout = in_tesnor->get_layout();
            auto tmp_in = std::make_shared<Tensor>(LiteDeviceType::LITE_CPU, in_layout);

            auto size = in_tesnor->get_tensor_total_size_in_byte() /
                        in_layout.get_elem_size();
            if (in_layout.data_type == LiteDataType::LITE_INT16) {
                auto ptr = static_cast<short*>(tmp_in->get_memory_ptr());
                for (size_t id = 0; id < size; id++) {
                    ptr[id] = i + 1;
                }
            } else if (in_layout.data_type == LiteDataType::LITE_UINT8) {
                auto ptr = static_cast<uint8_t*>(tmp_in->get_memory_ptr());
                for (size_t id = 0; id < size; id++) {
                    ptr[id] = i + 1;
                }
            }
            net_inputs.push_back(tmp_in);
            in_tesnor->copy_from(*tmp_in);
        }

        inputs.push_back(net_inputs);
        network->forward();
        network->wait();

        for (size_t j = 0; j < output_names.size(); j++) {
            auto out_tesnor = network->get_io_tensor(output_names[j]);
            auto out_layout = out_tesnor->get_layout();
            auto tmp_out =
                    std::make_shared<Tensor>(LiteDeviceType::LITE_CPU, out_layout);

            tmp_out->copy_from(*out_tesnor);
            net_outputs.push_back(tmp_out);
        }
        outputs.push_back(net_outputs);
    }

    config.options.force_output_use_user_specified_memory = true;
    config.options.comp_node_seq_record_level = record;
    config.options.const_shape = true;
568
    config.options.graph_opt_level = 2;
M
Megvii Engine Team 已提交
569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616

    std::shared_ptr<Network> network_record = std::make_shared<Network>(config);

    network_record->load_model(model_path);

    for (int i = 0; i < 3; i++) {
        for (size_t j = 0; j < inputs[i].size(); j++) {
            auto input_tensor = network_record->get_io_tensor(input_names[j]);
            input_tensor->reset(
                    inputs[i][j]->get_memory_ptr(), inputs[i][j]->get_layout());
        }

        std::vector<std::shared_ptr<Tensor>> net_outputs;

        for (size_t j = 0; j < outputs[i].size(); j++) {
            auto output_tensor = network_record->get_io_tensor(output_names[j]);
            auto tmp_out = std::make_shared<Tensor>(
                    LiteDeviceType::LITE_CPU, output_tensor->get_layout());
            output_tensor->reset(
                    tmp_out->get_memory_ptr(), output_tensor->get_layout());
            net_outputs.push_back(tmp_out);
        }

        network_record->forward();
        network_record->wait();

        for (size_t j = 0; j < outputs[i].size(); j++) {
            auto output_tensor = network_record->get_io_tensor(output_names[j]);
            compare_lite_tensor<float>(output_tensor, outputs[i][j]);
        }
    }
    printf("profile the model %s run\n", model_path.c_str());
    std::vector<std::shared_ptr<Tensor>> net_outputs;
    for (size_t j = 0; j < outputs[0].size(); j++) {
        auto output_tensor = network_record->get_io_tensor(output_names[j]);
        auto tmp_out = std::make_shared<Tensor>(
                LiteDeviceType::LITE_CPU, output_tensor->get_layout());
        output_tensor->reset(tmp_out->get_memory_ptr(), output_tensor->get_layout());
        net_outputs.push_back(tmp_out);
    }
    lite::Timer timer("profile");
    for (int i = 0; i < 10; i++) {
        network_record->forward();
        network_record->wait();
    }
    auto sum_time = timer.get_used_time();
    printf("model %s used time average %f ms\n", model_path.c_str(), sum_time / 10);
}
617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634
}  // namespace

TEST(TestNetWork, OutputNoCopy) {
    test_output_no_copy(0);
}

TEST(TestNetWork, OutputNoCopyRecord) {
    test_output_no_copy(1);
}

TEST(TestNetWork, IONoCopy) {
    test_input_no_copy(0);
}

TEST(TestNetWork, IONoCopyRecord) {
    test_input_no_copy(1);
}

M
Megvii Engine Team 已提交
635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656
TEST(TestNetWork, IONoCopyRecordAx) {
    std::vector<std::string> file_names;
#ifndef WIN32
    DIR* dirptr = NULL;
    struct dirent* dirp;
    std::string model_dir = "./ax_models";
    dirptr = opendir(model_dir.c_str());
    while (dirptr != NULL && (dirp = readdir(dirptr)) != NULL) {
        std::string file_name(dirp->d_name);
        if (file_name.find(".axe", 0) != std::string::npos) {
            file_names.push_back(model_dir + "/" + file_name);
        }
    }
    closedir(dirptr);
#endif

    for (auto file_name : file_names) {
        printf("test model: %s\n", file_name.c_str());
        test_io_no_copy_ax(file_name);
    }
}

657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682
TEST(TestNetWork, OutputDynamicAlloc) {
    Config config;
    config.options.force_output_dynamic_alloc = true;
    auto tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";
    auto result_mgb = mgb_lar(model_path, config, input_name, tensor);

    std::shared_ptr<Network> network = std::make_shared<Network>(config);

    network->load_model(model_path);
    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);

    auto src_ptr = tensor->get_memory_ptr();
    auto src_layout = tensor->get_layout();
    input_tensor->reset(src_ptr, src_layout);

    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);
    size_t times = 5;
    for (size_t i = 0; i < times; i++) {
        network->forward();
        network->wait();
        compare_lite_tensor<float>(output_tensor, result_mgb);
    }
}

683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702
TEST(TestNetWork, AsyncExec) {
    Config config;
    config.options.var_sanity_check_first_run = false;
    auto tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";
    auto result_mgb = mgb_lar(model_path, config, input_name, tensor);

    std::shared_ptr<Network> network = std::make_shared<Network>(config);

    network->load_model(model_path);

    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);

    auto src_ptr = tensor->get_memory_ptr();
    auto src_layout = tensor->get_layout();
    input_tensor->reset(src_ptr, src_layout);

    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);
    auto result_tensor = std::make_shared<Tensor>(
M
Megvii Engine Team 已提交
703
            LiteDeviceType::LITE_CPU, Layout{{1, 1000}, 2, LiteDataType::LITE_FLOAT});
704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863

    void* out_data = result_tensor->get_memory_ptr();
    output_tensor->reset(out_data, result_tensor->get_layout());

    //! set async mode and callback
    volatile bool finished = false;
    network->set_async_callback([&finished]() { finished = true; });

    network->forward();
    size_t count = 0;
    while (finished == false) {
        count++;
    }
    ASSERT_GT(count, 0);
    compare_lite_tensor<float>(output_tensor, result_mgb);
}

TEST(TestNetWork, CPUDeviceInput) {
    auto tensor = get_input_data("./input_data.npy");
    Layout layout{{1, 3, 224, 224}, 4, LiteDataType::LITE_FLOAT};
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";
    auto result_mgb = mgb_lar(model_path, {}, input_name, tensor);

    NetworkIO IO;
    bool is_host = false;
    IO.inputs.push_back({input_name, is_host});
    std::shared_ptr<Network> network = std::make_shared<Network>(IO);

    network->load_model(model_path);

    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);

    auto src_ptr = tensor->get_memory_ptr();
    input_tensor->reset(src_ptr, layout);

    network->forward();
    network->wait();

    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);
    compare_lite_tensor<float>(output_tensor, result_mgb);
}

TEST(TestNetWork, ShareTensorWith) {
    auto tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";
    auto result_mgb = mgb_lar(model_path, {}, input_name, tensor);

    std::shared_ptr<Network> network = std::make_shared<Network>();
    network->load_model(model_path);

    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);

    input_tensor->share_memory_with(*tensor);

    network->forward();
    network->wait();

    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);
    compare_lite_tensor<float>(output_tensor, result_mgb);
}

TEST(TestNetWork, InputCallBack) {
    auto tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";
    auto result_mgb = mgb_lar(model_path, {}, input_name, tensor);

    NetworkIO ios;
    bool is_host = false;
    ios.inputs.push_back({input_name, is_host});
    std::shared_ptr<Network> network = std::make_shared<Network>(ios);
    network->load_model(model_path);

    volatile bool finised_check_input = false;
    auto input_callback =
            [&tensor, &finised_check_input,
             input_name](const std::unordered_map<
                         std::string, std::pair<IO, std::shared_ptr<Tensor>>>&
                                 input_map) {
                ASSERT_EQ(input_map.size(), 1);
                auto tensor_input = input_map.at(input_name).second;
                compare_lite_tensor<float>(tensor_input, tensor);
                finised_check_input = true;
            };

    network->set_start_callback(input_callback);

    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);

    input_tensor->share_memory_with(*tensor);

    network->forward();
    network->wait();

    ASSERT_TRUE(finised_check_input);
    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);
    compare_lite_tensor<float>(output_tensor, result_mgb);
}

TEST(TestNetWork, OutputCallBack) {
    auto tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";
    auto result_mgb = mgb_lar(model_path, {}, input_name, tensor);

    std::shared_ptr<Network> network = std::make_shared<Network>();
    network->load_model(model_path);
    auto output_name = network->get_output_name(0);

    volatile bool finised_check_output = false;
    auto output_callback =
            [&result_mgb, &finised_check_output,
             output_name](const std::unordered_map<
                          std::string, std::pair<IO, std::shared_ptr<Tensor>>>&
                                  output_map) {
                ASSERT_EQ(output_map.size(), 1);
                auto tensor_output = output_map.at(output_name).second;
                compare_lite_tensor<float>(tensor_output, result_mgb);
                finised_check_output = true;
            };

    network->set_finish_callback(output_callback);

    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);

    input_tensor->share_memory_with(*tensor);

    network->forward();
    network->wait();

    ASSERT_TRUE(finised_check_output);
    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);
    compare_lite_tensor<float>(output_tensor, result_mgb);
}

TEST(TestNetWork, OutputShapeOnly) {
    auto tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";
    std::string output_name = "TRUE_DIV(EXP[12065],reduce0[12067])[12077]";

    NetworkIO IO;
    bool is_host = true;
    IO.outputs.push_back({output_name, is_host, LiteIOType::LITE_IO_SHAPE});
    Config config;
    std::shared_ptr<Network> network = std::make_shared<Network>(config, IO);

    network->load_model(model_path);

    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);
    std::shared_ptr<Tensor> output_tensor = network->get_io_tensor(output_name);

    auto src_ptr = tensor->get_memory_ptr();
    auto src_layout = tensor->get_layout();
    input_tensor->reset(src_ptr, src_layout);

    network->forward();
    network->wait();
M
Megvii Engine Team 已提交
864
    ASSERT_EQ(output_tensor->get_tensor_total_size_in_byte() / sizeof(float), 1000);
865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911
}

TEST(TestNetWork, ProfileIOdump) {
    auto tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";

    NetworkIO IO;
    Config config;
    std::shared_ptr<Network> network = std::make_shared<Network>(config, IO);
    network->enable_profile_performance("./profile.json");
    network->load_model(model_path);
    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);

    auto src_ptr = tensor->get_memory_ptr();
    auto src_layout = tensor->get_layout();
    input_tensor->reset(src_ptr, src_layout);

    network->forward();
    network->wait();
    ASSERT_TRUE(fopen("./profile.json", "r"));

    Runtime::enable_io_txt_dump(network, "./io_txt_dump.txt");
    network->forward();
    network->wait();
    ASSERT_TRUE(fopen("./io_txt_dump.txt", "r"));
}

TEST(TestNetWork, LoadPackedModel) {
    auto tensor = get_input_data("./input_data.npy");
    std::string model_path = "./test_packed_model.lite";
    std::string input_name = "data";

    NetworkIO IO;
    Config config;
    std::shared_ptr<Network> network = std::make_shared<Network>(config, IO);
    network->load_model(model_path);
    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);

    auto src_ptr = tensor->get_memory_ptr();
    auto src_layout = tensor->get_layout();
    input_tensor->reset(src_ptr, src_layout);

    network->forward();
    network->wait();
}

912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932
TEST(TestNetWork, GlabalLayoutTransform) {
    auto tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";
    std::string dump_model_name = "./shufflenet_after_trans.mge";

    NetworkIO IO;
    Config config;
    std::shared_ptr<Network> network = std::make_shared<Network>(config, IO);
    Runtime::enable_global_layout_transform(network);
    network->load_model(model_path);

    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);
    auto src_ptr = tensor->get_memory_ptr();
    auto src_layout = tensor->get_layout();
    input_tensor->reset(src_ptr, src_layout);

    Runtime::dump_layout_transform_model(network, dump_model_name);
    network->forward();
    network->wait();
    ASSERT_TRUE(fopen(dump_model_name.c_str(), "r"));
933
    remove(dump_model_name.c_str());
934 935
}

936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955
TEST(TestNetWork, GetDeviceType) {
    auto tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";

    Config config;
    std::shared_ptr<Network> network = std::make_shared<Network>(config);
    network->load_model(model_path);
    ASSERT_TRUE(network->get_device_type() == LiteDeviceType::LITE_CPU);
}

TEST(TestNetWork, GetModelExtraInfo) {
    std::string model_path = "./track_640_320_pack_model_rc4_with_info.lite";
    Config config;
    std::shared_ptr<Network> network = std::make_shared<Network>(config);
    network->load_model(model_path);
    auto& extra_info = network->get_model_extra_info();
    ASSERT_TRUE(extra_info.size() > 0);
    printf("extra_info %s \n", extra_info.c_str());
}

956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984
#ifndef __IN_TEE_ENV__
#if MGB_ENABLE_JSON
TEST(TestNetWork, GetMemoryInfo) {
    Config config;
    auto lite_tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";

    auto result_mgb = mgb_lar(model_path, config, "data", lite_tensor);

    std::shared_ptr<Network> network = std::make_shared<Network>(config);
    Runtime::set_cpu_threads_number(network, 2);

    network->load_model(model_path);
    network->get_static_memory_alloc_info();
    std::shared_ptr<Tensor> input_tensor = network->get_input_tensor(0);

    auto src_ptr = lite_tensor->get_memory_ptr();
    auto src_layout = lite_tensor->get_layout();
    input_tensor->reset(src_ptr, src_layout);

    network->forward();
    network->wait();
    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);

    compare_lite_tensor<float>(output_tensor, result_mgb);
}
#endif
#endif

985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071
#if LITE_WITH_CUDA

TEST(TestNetWork, BasicDevice) {
    auto lite_tensor = get_input_data("./input_data.npy");
    Config config;
    config.device_type = LiteDeviceType::LITE_CUDA;
    std::string model_path = "./shufflenet.mge";
    auto result_lite = mgelite_lar(model_path, config, "data", lite_tensor);
    auto result_mgb = mgb_lar(model_path, config, "data", lite_tensor);
    compare_lite_tensor<float>(result_lite, result_mgb);
}

TEST(TestNetWork, DeviceInput) {
    auto tensor = get_input_data("./input_data.npy");
    Layout layout{{1, 3, 224, 224}, 4, LiteDataType::LITE_FLOAT};
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";
    auto result_mgb = mgb_lar(model_path, {}, input_name, tensor);

    NetworkIO IO;
    bool is_host = false;
    IO.inputs.push_back({input_name, is_host});
    Config config;
    config.device_type = LiteDeviceType::LITE_CUDA;
    std::shared_ptr<Network> network = std::make_shared<Network>(config, IO);

    network->load_model(model_path);

    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);

    auto tensor_cuda = Tensor(LiteDeviceType::LITE_CUDA, layout);
    tensor_cuda.copy_from(*tensor);

    auto src_ptr = tensor_cuda.get_memory_ptr();
    input_tensor->reset(src_ptr, layout);

    network->forward();
    network->wait();

    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);
    compare_lite_tensor<float>(output_tensor, result_mgb);
}

TEST(TestNetWork, ChangeInputShapeDevice) {
    Config config;
    auto tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";
    auto result_mgb = mgb_lar(model_path, config, input_name, tensor);

    config.device_type = LiteDeviceType::LITE_CUDA;
    std::shared_ptr<Network> network = std::make_shared<Network>(config);

    network->load_model(model_path);
    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);

    auto src_layout = Layout{{2, 3, 200, 200}, 4, LiteDataType::LITE_FLOAT};
    input_tensor->set_layout(src_layout);
    std::shared_ptr<Tensor> input_tensor2 = network->get_io_tensor(input_name);
    //! Check memory is equal
    ASSERT_EQ(input_tensor->get_memory_ptr(), input_tensor2->get_memory_ptr());

    network->forward();
    network->wait();
    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);
    auto output_layout = output_tensor->get_layout();
    ASSERT_EQ(output_layout.shapes[0], 2);
    ASSERT_EQ(output_layout.shapes[1], 1000);
}

TEST(TestNetWork, DeviceOutput) {
    auto tensor = get_input_data("./input_data.npy");
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";
    std::string output_name = "TRUE_DIV(EXP[12065],reduce0[12067])[12077]";
    auto result_mgb = mgb_lar(model_path, {}, input_name, tensor);

    NetworkIO IO;
    bool is_host = false;
    IO.outputs.push_back({output_name, is_host});
    Config config;
    config.device_type = LiteDeviceType::LITE_CUDA;
    std::shared_ptr<Network> network = std::make_shared<Network>(config, IO);

    network->load_model(model_path);

    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);
M
Megvii Engine Team 已提交
1072
    std::shared_ptr<Tensor> output_tensor_cuda = network->get_io_tensor(output_name);
1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092

    auto src_ptr = tensor->get_memory_ptr();
    auto src_layout = tensor->get_layout();
    input_tensor->reset(src_ptr, src_layout);

    network->forward();
    network->wait();
    auto output_tensor = std::make_shared<Tensor>();
    output_tensor->copy_from(*output_tensor_cuda);

    compare_lite_tensor<float>(output_tensor, result_mgb);
}

TEST(TestNetWork, WrongIONameDevice) {
    auto tensor = get_input_data("./input_data.npy");
    Layout layout{{1, 3, 224, 224}, 4, LiteDataType::LITE_FLOAT};
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";
    std::string input_name_wrong = "data0";
    std::string output_name = "TRUE_DIV(EXP[12065],reduce0[12067])[12077]";
M
Megvii Engine Team 已提交
1093
    std::string output_name_wrong = "w_TRUE_DIV(EXP[12065],reduce0[12067])[12077]";
1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113
    auto result_mgb = mgb_lar(model_path, {}, input_name, tensor);

    NetworkIO IO;
    bool is_host = false;
    IO.inputs.push_back({input_name, is_host});
    IO.outputs.push_back({output_name, is_host});
    IO.outputs.push_back({output_name_wrong, is_host});
    Config config;
    config.device_type = LiteDeviceType::LITE_CUDA;
    std::shared_ptr<Network> network = std::make_shared<Network>(config, IO);

    network->load_model(model_path);

    auto tensor_cuda = Tensor(LiteDeviceType::LITE_CUDA, layout);
    tensor_cuda.copy_from(*tensor);
    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);
    auto src_ptr = tensor_cuda.get_memory_ptr();
    auto src_layout = tensor_cuda.get_layout();
    input_tensor->reset(src_ptr, src_layout);

M
Megvii Engine Team 已提交
1114
    std::shared_ptr<Tensor> output_tensor_cuda = network->get_io_tensor(output_name);
1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146

    network->forward();
    network->wait();
    auto output_tensor = std::make_shared<Tensor>();
    output_tensor->copy_from(*output_tensor_cuda);

    compare_lite_tensor<float>(output_tensor, result_mgb);
}

TEST(TestNetWork, ConfigIONameDevice) {
    std::string model_path = "./model.mgb";

    NetworkIO IO;
    bool is_host = false;
    IO.outputs.push_back({"clsfy", is_host});
    Config config;
    config.device_type = LiteDeviceType::LITE_CUDA;
    std::shared_ptr<Network> network = std::make_shared<Network>(config, IO);
    network->compute_only_configured_output();
    network->load_model(model_path);

    ASSERT_EQ(network->get_all_output_name().size(), 1);
    ASSERT_EQ(network->get_all_output_name()[0], "clsfy");

    std::shared_ptr<Network> network2 = std::make_shared<Network>(config, IO);
    network2->load_model(model_path);

    ASSERT_EQ(network2->get_all_output_name().size(), 2);
}

TEST(TestNetWork, SetDeviceIdDeviceTest) {
#if LITE_WITH_CUDA
M
Megvii Engine Team 已提交
1147
    if (get_device_count(LITE_CUDA) <= 1)
1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171
        return;
#endif
    std::string model_path = "./model.mgb";

    NetworkIO IO;
    bool is_host = false;
    IO.inputs.push_back({"data", is_host});
    IO.outputs.push_back({"clsfy", is_host});
    Config config;
    config.device_type = LiteDeviceType::LITE_CUDA;
    std::shared_ptr<Network> network = std::make_shared<Network>(config, IO);
    network->set_device_id(1);
    network->load_model(model_path);
    auto inputs_names = network->get_all_input_name();
    for (auto name : inputs_names) {
        auto tensor = network->get_io_tensor(name);
        ASSERT_EQ(tensor->get_device_id(), 1);
        if (name == "idx") {
            int* index_ptr = static_cast<int*>(tensor->get_memory_ptr());
            for (int i = 0; i < 23; i++) {
                index_ptr[i] = i % 3;
            }
        }
        if (name == "landmark") {
M
Megvii Engine Team 已提交
1172
            float* landmakrk_ptr = static_cast<float*>(tensor->get_memory_ptr());
1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208
            for (int i = 0; i < 23 * 18 * 2; i++) {
                landmakrk_ptr[i] = 0.1f;
            }
        }
    }
    auto outputs_names = network->get_all_output_name();
    for (auto name : outputs_names) {
        auto tensor = network->get_io_tensor(name);
        ASSERT_EQ(tensor->get_device_id(), 1);
    }
    network->forward();
    network->wait();
}

TEST(TestNetWork, SetStreamIdDeviceTest) {
    std::string model_path = "./model.mgb";

    NetworkIO IO;
    bool is_host = false;
    IO.inputs.push_back({"data", is_host});
    IO.outputs.push_back({"clsfy", is_host});
    Config config;
    config.device_type = LiteDeviceType::LITE_CUDA;
    std::shared_ptr<Network> network = std::make_shared<Network>(config, IO);
    network->set_stream_id(1);
    network->load_model(model_path);
    auto inputs_names = network->get_all_input_name();
    for (auto name : inputs_names) {
        auto tensor = network->get_io_tensor(name);
        if (name == "idx") {
            int* index_ptr = static_cast<int*>(tensor->get_memory_ptr());
            for (int i = 0; i < 23; i++) {
                index_ptr[i] = i % 3;
            }
        }
        if (name == "landmark") {
M
Megvii Engine Team 已提交
1209
            float* landmakrk_ptr = static_cast<float*>(tensor->get_memory_ptr());
1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240
            for (int i = 0; i < 23 * 18 * 2; i++) {
                landmakrk_ptr[i] = 0.1f;
            }
        }
    }
    network->forward();
    network->wait();
}

#if CUDART_VERSION >= 10000
TEST(TestNetWork, DeviceAsyncExec) {
    auto tensor = get_input_data("./input_data.npy");
    Config config;
    config.device_type = LiteDeviceType::LITE_CUDA;
    config.options.var_sanity_check_first_run = false;
    std::string model_path = "./shufflenet.mge";
    std::string input_name = "data";
    auto result_mgb = mgb_lar(model_path, config, input_name, tensor);

    std::shared_ptr<Network> network = std::make_shared<Network>(config);

    network->load_model(model_path);

    std::shared_ptr<Tensor> input_tensor = network->get_io_tensor(input_name);

    auto src_ptr = tensor->get_memory_ptr();
    auto src_layout = tensor->get_layout();
    input_tensor->reset(src_ptr, src_layout);

    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);
    auto result_tensor = std::make_shared<Tensor>(
M
Megvii Engine Team 已提交
1241
            LiteDeviceType::LITE_CPU, Layout{{1, 1000}, 2, LiteDataType::LITE_FLOAT});
1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261

    void* out_data = result_tensor->get_memory_ptr();
    output_tensor->reset(out_data, result_tensor->get_layout());

    //! set async mode and callback
    volatile bool finished = false;
    network->set_async_callback([&finished]() { finished = true; });

    network->forward();
    size_t count = 0;
    while (finished == false) {
        count++;
    }

    ASSERT_GT(count, 0);
    compare_lite_tensor<float>(output_tensor, result_mgb);
}

#endif
#endif
1262 1263 1264 1265

#if MGB_ATLAS || MGB_CAMBRICON
namespace {
void load_no_device(LiteDeviceType device_type, const std::string& model_path) {
1266
    lite::Config config;
1267
    config.device_type = device_type;
1268
    auto network = std::make_shared<lite::Network>(config);
1269
    network->load_model(model_path);
1270 1271 1272 1273
    network->forward();
    network->wait();
}

1274 1275 1276
void load_device_input(
        LiteDeviceType device_type, const std::string& model_path,
        const std::vector<std::string>& inputs) {
1277 1278
    lite::NetworkIO networkio;
    lite::IO input_data_io = {};
1279
    input_data_io.name = inputs[0];
1280 1281 1282
    input_data_io.is_host = false;
    networkio.inputs.emplace_back(input_data_io);
    lite::IO input_input0_io = {};
1283
    input_input0_io.name = inputs[1];
1284 1285 1286
    input_input0_io.is_host = false;
    networkio.inputs.emplace_back(input_input0_io);
    lite::Config config;
1287
    config.device_type = device_type;
1288
    auto network = std::make_shared<lite::Network>(config, networkio);
1289
    network->load_model(model_path);
1290 1291 1292 1293
    network->forward();
    network->wait();
}

1294 1295
void load_device_id(
        LiteDeviceType device_type, int device_id, const std::string& model_path) {
1296
    lite::Config config;
1297
    config.device_type = device_type;
1298
    auto network = std::make_shared<lite::Network>(config);
1299 1300 1301 1302
    network->set_device_id(device_id);
    network->load_model(model_path);
    std::shared_ptr<Tensor> input_tensor = network->get_input_tensor(0);
    std::shared_ptr<Tensor> output_tensor = network->get_output_tensor(0);
1303 1304
    network->forward();
    network->wait();
1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322
    ASSERT_EQ(output_tensor->get_device_id(), device_id);
}
}  // namespace
#endif

#if MGB_ATLAS
TEST(TestNetWork, AtlasLoadNoDevice) {
    load_no_device(LiteDeviceType::LITE_DEVICE_DEFAULT, "./model_atlas.mgb");
}

TEST(TestNetWork, AtlasLoadDeviceInput) {
    load_device_input(
            LiteDeviceType::LITE_DEVICE_DEFAULT, "./model_atlas.mgb",
            {"data", "input0"});
}

TEST(TestNetWork, AtlasLoadAtlas) {
    load_no_device(LiteDeviceType::LITE_ATLAS, "./model_atlas.mgb");
1323 1324 1325
}

TEST(TestNetWork, AtlasLoadAtlasDeviceInput) {
1326 1327
    load_device_input(
            LiteDeviceType::LITE_ATLAS, "./model_atlas.mgb", {"data", "input0"});
1328 1329 1330
}

TEST(TestNetWork, AtlasDeviceID) {
1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357
    load_device_id(LiteDeviceType::LITE_ATLAS, 1, "./model_atlas.mgb");
}
#endif

#if MGB_CAMBRICON
TEST(TestNetWork, CambriconLoadNoDevice) {
    load_no_device(LiteDeviceType::LITE_DEVICE_DEFAULT, "./model_magicmind.mgb");
}

TEST(TestNetWork, CambriconLoadDeviceInput) {
    load_device_input(
            LiteDeviceType::LITE_DEVICE_DEFAULT, "./model_magicmind.mgb",
            {"data", "input0"});
}

TEST(TestNetWork, CambriconLoadCambricon) {
    load_no_device(LiteDeviceType::LITE_CAMBRICON, "./model_magicmind.mgb");
}

TEST(TestNetWork, CambriconLoadCambriconDeviceInput) {
    load_device_input(
            LiteDeviceType::LITE_CAMBRICON, "./model_magicmind.mgb",
            {"data", "input0"});
}

TEST(TestNetWork, CambriconDeviceID) {
    load_device_id(LiteDeviceType::LITE_CAMBRICON, 0, "./model_magicmind.mgb");
1358 1359
}
#endif
1360

1361 1362
#endif
// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}