checker.h 20.9 KB
Newer Older
1 2 3 4
/**
 * \file dnn/test/common/checker.h
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
5
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
6 7 8
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
9 10
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26
 */

#pragma once

#include "megdnn/basic_types.h"
#include "megdnn/tensor_iter.h"
#include "test/common/opr_algo_proxy.h"
#include "test/common/opr_proxy.h"
#include "test/common/rng.h"

#include <gtest/gtest.h>

#include <memory>
#include <regex>
#include <unordered_map>

27
// clang-format off
28
#if defined(__has_feature)
29 30 31 32 33
    #if __has_feature(address_sanitizer)
        #define MEGDNN_TEST_ASAN 1
    #else
        #define MEGDNN_TEST_ASAN 0
    #endif
34
#elif defined(__SANITIZE_ADDRESS__)
35
    #define MEGDNN_TEST_ASAN 1
36
#else
37 38 39 40
    #define MEGDNN_TEST_ASAN 0
#endif
// clang-format on

41 42 43 44 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
namespace megdnn {
namespace test {

class CheckerHelper {
    // TensorLayoutArray and TensorValueArray should be protected in theory;
    // but g++-4.9 bugs handle access privilege wrongfully, so we change it
    // to public.
public:
    using TensorValueArray = TensorNDArray;
    using TensorsConstriant = std::function<void(TensorValueArray& tensors)>;
    using ExtraOprImpl = std::function<void(const TensorNDArray&)>;
    using OutputCanonizer = std::function<void(const TensorValueArray&)>;
    static std::shared_ptr<TensorValueArray> alloc_tensors(
            Handle* handle, const TensorLayoutArray& layouts, size_t offset);

    Handle* handle() const { return m_handle_cur; }

protected:
    //! whether to use physically contiguous (i.e. default layout) for naive
    //! impl
    bool m_enable_contig_naive = false;

    bool m_prev_succ = true;
    const char* m_input_tensors_fpath = nullptr;
    thin_function<void()> m_expect_exec_fail;
    std::unique_ptr<Handle> m_handle_naive;
    Handle* m_handle_cur;
    std::unique_ptr<RNG> m_default_rng;
    std::unordered_map<size_t, RNG*> m_rng;
    std::unordered_map<size_t, DType> m_dtype;
    std::unordered_map<size_t, TensorFormat> m_fmt;
    float_t m_epsilon = 1e-3, m_max_avg_error = 1e-3,
            m_max_avg_biased_error = 1e-3;
    float_t m_perf_check_threshold = -1;
    bool m_perf_check = false;
    ExtraOprImpl m_extra_opr_impl;
    OutputCanonizer m_output_canonizer;
    TensorsConstriant m_tensor_constraint;
M
Megvii Engine Team 已提交
79
    bool no_naive_and_check = false;
80 81 82 83 84 85 86 87 88 89
    /**
     * the offset from the start of malloc memory
     *
     * \note alloc \p m_offset more memory when alloc memory for a tensor,
     * the start of tensor just begin at \p m_offset.
     * \warning current only used for opencl
     */
    size_t m_offset = 0;

    CheckerHelper(Handle* handle, bool check_dispatch = true);
90

91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108
    ~CheckerHelper() noexcept;

    using OprExec = std::function<void(const TensorValueArray&)>;

    void do_exec_with_testcases(const TensorValueArray& testcase_in,
                                const TensorValueArray& testcase_out,
                                const OprExec& exec_opr);

    void do_exec(const TensorLayoutArray& user_layouts,
                 const TensorLayoutArray& deduced_layouts,
                 const OprExec& exec_naive, const OprExec& exec_opr);

    void enable_contig_naive() { m_enable_contig_naive = true; }

    void copy_tensors_to_device(const TensorValueArray& dest,
                                const TensorValueArray& src);
    void copy_tensors_from_device(const TensorValueArray& dest,
                                  const TensorValueArray& src);
109 110 111 112 113

private:
    std::shared_ptr<TensorValueArray> m_tensors_naive;

    void init_naive_values();
114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131
    void check_tensors(const TensorValueArray& expected,
                       const TensorValueArray& computed);
};

template <typename Opr, typename Proxy = OprProxy<Opr>>
class Checker : public CheckerHelper {
public:
    using Param = typename Opr::Param;
    using BeforeExecCallback =
            std::function<void(Opr*, const TensorValueArray&)>;
    Checker(Handle* handle, bool check_dispatch = true)
            : CheckerHelper(handle, check_dispatch), m_param(Param()) {}

    TensorLayoutArray make_layouts(const TensorShapeArray& shapes) {
        TensorLayoutArray layouts(shapes.size());
        for (size_t i = 0; i < shapes.size(); ++i) {
            DType dt = (m_dtype.find(i) != m_dtype.end() ? m_dtype[i]
                                                         : dtype::Float32());
132 133 134 135 136
            if (m_fmt.find(i) == m_fmt.end()) {
                layouts[i] = TensorLayout(shapes[i], dt);
                layouts[i].init_contiguous_stride();
            } else
                layouts[i] = TensorLayout(shapes[i], dt, m_fmt[i]);
137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250
        }
        return layouts;
    }

    /*!
     * \brief execute opr on current param/dtype/rng config
     * \param shapes input/output shapes, which would be passed as
     *      arguments to Opr::deduce_layout
     *
     * Checker would construct TensorLayout vectors from shapes and dtypes,
     * and call exec(TensorLayoutArray &).
     */
    Checker& exec(const TensorShapeArray& shapes) {
        exec(make_layouts(shapes));
        return *this;
    }

    void exec(TensorLayoutArray layouts);

    //! explicitly require argument to be TensorShape
    Checker& execs(const TensorShapeArray& shapes) { return exec(shapes); }

    //! explicitly require argument to be TensorLayout
    Checker& execl(const TensorLayoutArray& layouts) {
        exec(layouts);
        return *this;
    }

    Checker& exect(const TensorValueArray& testcase_in,
                   const TensorValueArray& testcase_out);

    Checker& set_param(Param param) {
        m_param = param;
        opr()->param() = param;
        return *this;
    }
    Checker& set_dtype(size_t idx, DType dtype) {
        m_dtype[idx] = dtype;
        return *this;
    }
    Checker& set_fmt(size_t idx, TensorFormat fmt) {
        m_fmt[idx] = fmt;
        return *this;
    }
    Checker& set_rng(size_t idx, RNG* rng) {
        m_rng[idx] = rng;
        return *this;
    }
    //! max error of a single element
    Checker& set_epsilon(dt_float32 epsilon) {
        m_epsilon = epsilon;
        m_max_avg_error = epsilon;
        m_max_avg_biased_error = epsilon;
        return *this;
    }
    //! max average error; defaults to epsilon
    Checker& set_max_avg_error(dt_float32 error) {
        m_max_avg_error = error;
        return *this;
    }
    //! max average biased error; defaults to epsilon
    Checker& set_max_avg_biased_error(dt_float32 error) {
        m_max_avg_biased_error = error;
        return *this;
    }
    Checker& set_offset(size_t offset) {
        m_offset = offset;
        return *this;
    }

    Checker& set_proxy(const Proxy& proxy) {
        m_naive_proxy = proxy;
        m_cur_proxy = proxy;
        return *this;
    }

    //! set_perf_check and set_perf_check_threshold control the
    //! performance checking behavior.
    //!
    //! If perf_check is on (default to off), the running time of the
    //! current operator and the naive operator would be measured and
    //! checked when calling exec.
    //! The accelerating ratio should be larger than perf_check_threshold,
    //! otherwise errors would be reported.
    //! perf_check_threshold must be set in advance since the default value
    //! (which is negative) is invalid.
    Checker& set_perf_check(bool perf_check) {
        m_perf_check = perf_check;
        return *this;
    }

    Checker& set_perf_check_threshold(float perf_check_threshold) {
        m_perf_check_threshold = perf_check_threshold;
        return *this;
    }

    //! load input tensors from file for next run
    Checker& load_input_tensors(const char* fpath) {
        m_input_tensors_fpath = fpath;
        return *this;
    }

    //! add another checker to ensure naive implementation is correct
    Checker& set_extra_opr_impl(const ExtraOprImpl& chk) {
        m_extra_opr_impl = chk;
        return *this;
    }

    //! set a callback to be invoked before executing the operator
    Checker& set_before_exec_callback(const BeforeExecCallback& cb) {
        m_before_exec_callback = cb;
        return *this;
    }

251 252 253 254 255
    Checker& reset_before_exec_callback() {
        m_before_exec_callback = nullptr;
        return *this;
    }

256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405
    //! set a tensors constraints function, for the purpose of manipulating
    //! tensors when testing.
    Checker& set_tensors_constraint(
            const TensorsConstriant& tensor_constraint) {
        m_tensor_constraint = tensor_constraint;
        return *this;
    }

    /*!
     * \brief set that exec() on opr should fail, so naive is not called and
     * exec() returns directly after opr is called.
     *
     * This is only valid for next exec() call. It is usually used for
     * testing megcore::AsyncErrorInfo.
     *
     * \param cb callback to be invoked after opr exec (so error would not
     *           be passed to destructor)
     */
    Checker& set_expect_exec_fail(const thin_function<void()>& cb) {
        m_expect_exec_fail = cb;
        return *this;
    }

    /*!
     * \brief set a function to canonize the outputs
     *
     * For some oprs maybe multiple outputs can be accepted; we can use a
     * function to transform them into a canonized form before comparing.
     *
     * The arguments are tensors on CPU and should be modified in-place.
     */
    Checker& set_output_canonizer(OutputCanonizer canonizer) {
        m_output_canonizer = std::move(canonizer);
        return *this;
    }

    //! get the opr impl so setting other than param() can be modified
    Opr* opr() {
        if (!m_opr_cur) {
            m_opr_cur = m_handle_cur->create_operator<Opr>();
        }
        return m_opr_cur.get();
    }

    //! whether previous exec succeeds
    bool prev_succ() const { return m_prev_succ; }

private:
    BeforeExecCallback m_before_exec_callback;
    Param m_param;
    Proxy m_naive_proxy, m_cur_proxy;
    std::unique_ptr<Opr> m_opr_cur;
};

::testing::AssertionResult __assert_tensor_eq(
        const char* expr0, const char* expr1, const char* expr_maxerr,
        const char* expr_maxerr_avg, const char* expr_maxerr_avg_biased,
        const TensorND& v0, const TensorND& v1, float maxerr, float maxerr_avg,
        float maxerr_avg_biased);

#define MEGDNN_ASSERT_TENSOR_EQ_EPS_AVG(v0, v1, maxerr, maxerr_avg,         \
                                        maxerr_avg_biased)                  \
    ASSERT_PRED_FORMAT5(::megdnn::test::__assert_tensor_eq, v0, v1, maxerr, \
                        maxerr_avg, maxerr_avg_biased)

#define MEGDNN_ASSERT_TENSOR_EQ_EPS(v0, v1, maxerr) \
    MEGDNN_ASSERT_TENSOR_EQ_EPS_AVG(v0, v1, maxerr, maxerr, maxerr)

#define MEGDNN_ASSERT_TENSOR_EQ(v0, v1) \
    MEGDNN_ASSERT_TENSOR_EQ_EPS(v0, v1, 1e-3)

template <typename Opr, typename Proxy>
void Checker<Opr, Proxy>::exec(TensorLayoutArray layouts) {
    auto opr_naive = m_handle_naive->create_operator<Opr>();
    auto opr_relayout = m_handle_naive->create_operator<RelayoutForward>();

    auto opr_cur = this->opr();
    opr_naive->param() = m_param;
    opr_cur->param() = m_param;
    m_naive_proxy.deduce_layout(opr_naive.get(), layouts);
    auto exec_naive = [this, &opr_naive, &layouts,
                       &opr_relayout](const TensorValueArray& values) {
        TensorValueArray contig_values = values;
        TensorValueArray real_values = values;
        std::shared_ptr<TensorValueArray> tensors_naive_contig_storage;
        if (m_enable_contig_naive) {
            TensorLayoutArray contig_layouts;
            for (auto&& layout : layouts) {
                contig_layouts.emplace_back(TensorLayout{
                        static_cast<const TensorShape&>(layout), layout.dtype});
            }
            m_naive_proxy.deduce_layout(opr_naive.get(), contig_layouts);
            tensors_naive_contig_storage = alloc_tensors(
                    m_handle_naive.get(), contig_layouts, m_offset);
            contig_values = *tensors_naive_contig_storage;
            //! relayout value to the contig_values
            for (size_t i = 0; i < contig_values.size(); ++i) {
                if (real_values[i].layout.ndim == 0)
                    continue;
                real_values[i].layout.format = {};
                opr_relayout->exec(real_values[i], contig_values[i],
                                   m_handle_naive.get());
            }
        }

        m_naive_proxy.exec(opr_naive.get(), contig_values);

        if (m_enable_contig_naive) {
            //! relayout to the values
            for (size_t i = 0; i < contig_values.size(); ++i) {
                if (real_values[i].layout.ndim == 0)
                    continue;
                opr_relayout->exec(contig_values[i], real_values[i],
                                   m_handle_naive.get());
            }
        }
    };
    auto exec_opr = [this, opr_cur](const TensorValueArray& values) {
        if (m_before_exec_callback) {
            m_before_exec_callback(opr_cur, values);
        }
        m_cur_proxy.exec(opr_cur, values);
    };
    auto user_layouts = layouts;
    do_exec(user_layouts, layouts, exec_naive, exec_opr);
}

template <typename Opr, typename Proxy>
Checker<Opr, Proxy>& Checker<Opr, Proxy>::exect(
        const TensorValueArray& testcase_in,
        const TensorValueArray& testcase_out) {
    auto opr_cur = this->opr();
    opr_cur->param() = m_param;
    auto exec_opr = [this, opr_cur](const TensorValueArray& values) {
        if (m_before_exec_callback) {
            m_before_exec_callback(opr_cur, values);
        }
        m_cur_proxy.exec(opr_cur, values);
    };
    do_exec_with_testcases(testcase_in, testcase_out, exec_opr);
    return *this;
}

template <typename T, typename U>
TensorND TensorValue(const TensorShape& shape, T dtype,
                     std::initializer_list<U> values) {
    TensorND tensor;
    tensor.layout = {shape, dtype};
    tensor.raw_ptr =
            static_cast<dt_byte*>(malloc(tensor.layout.span().dist_byte()));
406 407
    megdnn_assert(values.size() == tensor.layout.total_nr_elems(), "%zu == %zu",
                  values.size(), tensor.layout.total_nr_elems());
408 409 410 411 412 413 414 415 416
    auto ptr = tensor.ptr<typename DTypeTrait<T>::ctype>();
    for (const auto& v : values) {
        *ptr++ = typename DTypeTrait<T>::ctype(v);
    }
    return tensor;
}

template <typename T, typename U>
TensorND TensorValueLowbit4(const TensorShape& shape, T dtype,
417
                                std::vector<U> values) {
418 419 420 421 422
    TensorND tensor;
    tensor.layout = {shape, dtype};
    tensor.raw_ptr =
            static_cast<dt_byte*>(malloc(tensor.layout.span().dist_byte()));
    megdnn_assert(values.size() == tensor.layout.total_nr_elems());
423
    auto ptr = tensor.ptr<typename DTypeTrait<T>::ctype>();
424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440
    auto layout = tensor.layout;
    auto dim_in = shape[layout.ndim - 1];
    auto elems = tensor.layout.total_nr_elems();
    auto dim_out = elems / dim_in;
    auto stride_out = div_ceil(dim_in, 2_z);
    size_t in_offset = 0;
    for (size_t i = 0; i < dim_out; ++i) {
        for (size_t j = 0; j < dim_in; j += 2) {
            U a = values[in_offset + j];
            U b = 0;
            if (j + 1 < dim_in)
                b = values[in_offset + j + 1];
            megdnn_assert(a >= DTypeTrait<T>::min());
            megdnn_assert(a <= DTypeTrait<T>::max());
            megdnn_assert(b >= DTypeTrait<T>::min());
            megdnn_assert(b <= DTypeTrait<T>::max());
            ptr[j / 2] = (a & 0xF) | (b << 4);
441
        }
442 443
        in_offset += dim_in;
        ptr += stride_out;
444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463
    }
    return tensor;
}

class Testcase : public SmallVector<TensorND> {
public:
    using SmallVector<TensorND>::SmallVector;
    ~Testcase() {
        // Suicide
        for (const auto& tensor : *this) {
            if (tensor.raw_ptr) {
                free(tensor.raw_ptr);
            }
        }
    }

    Testcase(const Testcase&) = delete;
    Testcase operator=(const Testcase&) = delete;
};

464 465 466 467 468 469 470 471 472 473 474
struct ExecutionPolicyAlgoName {
    std::string name;
    std::vector<ExecutionPolicyAlgoName> sub_policy_names;

    ExecutionPolicyAlgoName(const char* name) : name{name} {}

    ExecutionPolicyAlgoName(
            const char* name,
            const std::vector<ExecutionPolicyAlgoName>& sub_policy)
            : name{name}, sub_policy_names{sub_policy} {}
};
475 476 477 478 479 480 481 482 483 484
/*!
 * \brief a callable to check that given algorithm is used for heuristic
 * \param require_algo if its value is true, then requires
 *      get_algorithm_heuristic() to return the expected algo; otherwise the
 *      expected algo must exist in get_all_algorithms() and it would be set to
 *      be used
 */
template <class Opr, typename OprAlgoProxy = OprAlgoProxy<Opr>>
class AlgoChecker {
public:
485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503
    AlgoChecker(ExecutionPolicyAlgoName name, bool* require_algo = nullptr)
            : m_policy_name{name}, m_require_algo{require_algo} {}

    AlgoChecker(ExecutionPolicy policy, bool* require_algo = nullptr)
            : m_policy{policy}, m_require_algo{require_algo} {}

    static ExecutionPolicy construct_execution_policy_from_name(
            const ExecutionPolicyAlgoName& policy_name,
            const TensorLayoutArray& layouts, const std::string& param,
            Handle* handle) {
        ExecutionPolicy ret;
        megdnn_assert(layouts.size() == OprTrait<Opr>::arity);
        auto opr = handle->create_operator<Opr>();
        opr->param() =
                Algorithm::deserialize_read_pod<typename Opr::Param>(param);
        for (auto algo_info :
             AlgoProxy<Opr, OprTrait<Opr>::arity>::get_all_algorithms_info(
                     opr.get(), layouts)) {
            if (std::regex_match(
504
                        algo_info.desc.name,
505 506 507 508 509 510 511 512 513
                        std::regex("(" + policy_name.name + ")(.*)"))) {
                ret.algo = algo_info.desc;
            } else {
                continue;
            }

            Algorithm* algo = opr->get_algorithm_from_desc(algo_info.desc);
            std::vector<Algorithm::SearchItem>&& sub_items =
                    algo->get_subopr_list(layouts, opr.get());
514 515 516
            if (sub_items.size() != policy_name.sub_policy_names.size()) {
                printf("Invalid sub_policy_names in %s, expected %zu but got "
                       "%zu\n",
517
                       algo_info.desc.name.c_str(), sub_items.size(),
518 519 520
                       policy_name.sub_policy_names.size());
                return {};
            }
521 522 523 524 525 526 527 528 529 530 531
            FOREACH_OPR_TYPE_DISPATCH(sub_items, {
                ExecutionPolicy policy =
                        AlgoChecker<_Opr>::construct_execution_policy_from_name(
                                policy_name.sub_policy_names[_item_idx],
                                _item.layouts, _item.param, handle);
                ret.sub_policy.push_back(policy);
            });
            return ret;
        }
        return ret;
    }
532 533 534 535 536 537

    void operator()(Opr* opr, const CheckerHelper::TensorValueArray& arr) {
        TensorLayoutArray layouts;
        for (auto&& val : arr) {
            layouts.push_back(val.layout);
        }
538 539 540 541 542 543 544 545
        if (!m_policy_name.name.empty()) {
            std::string param_str;
            Algorithm::serialize_write_pod(opr->param(), param_str);
            m_policy = construct_execution_policy_from_name(
                    m_policy_name, layouts, param_str, opr->handle());
            ASSERT_TRUE(m_policy.algo.valid())
                    << "algorithm " << m_policy_name.name << " not found";
        }
546
        if (m_require_algo && *m_require_algo) {
547 548
            auto algo =
                    OprAlgoProxy::get_algorithm_info_heuristic(opr, layouts);
549
            ASSERT_STREQ(opr->get_algorithm_from_desc(m_policy.algo)->name(),
550
                         algo.desc.name.c_str());
551
        } else {
552
            opr->execution_policy() = m_policy;
553 554
        }
    }
555 556 557 558 559

private:
    ExecutionPolicyAlgoName m_policy_name;
    ExecutionPolicy m_policy;
    bool* m_require_algo;
560 561
};

562 563 564 565 566 567 568 569 570 571
template <typename Opr>
void construct_sub_execution_policy_heuristic(ExecutionPolicy& policy,
                                              const TensorLayoutArray& layouts,
                                              const std::string& param,
                                              Handle* handle) {
    megdnn_assert(layouts.size() == OprTrait<Opr>::arity);
    auto opr = handle->create_operator<Opr>();
    opr->param() = Algorithm::deserialize_read_pod<typename Opr::Param>(param);
    if (!policy.algo.valid()) {
        policy.algo = AlgoProxy<Opr, OprTrait<Opr>::arity>::
572 573
                              get_algorithm_info_heuristic(opr.get(), layouts)
                                      .desc;
574 575 576 577 578 579 580 581
    }

    Algorithm* algo = opr->get_algorithm_from_desc(policy.algo);
    std::vector<Algorithm::SearchItem>&& sub_items =
            algo->get_subopr_list(layouts, opr.get());
    FOREACH_OPR_TYPE_DISPATCH(sub_items, {
        policy.sub_policy.push_back(ExecutionPolicy{});
        construct_sub_execution_policy_heuristic<_Opr>(
582
                policy.sub_policy.back(), _item.layouts, _item.param, handle);
583 584 585
    });
}

586 587 588 589
}  // namespace test
}  // namespace megdnn

// vim: syntax=cpp.doxygen