opr_impl.cpp 30.4 KB
Newer Older
1 2 3 4
/**
 * \file dnn/src/fallback/convolution/opr_impl.cpp
 * 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
 */

#include "src/common/algo_chooser.h"
#include "src/common/metahelper.h"
#include "src/common/opr_delegate.h"
#include "src/common/utils.h"
#include "src/fallback/convolution/algos.h"
18
#include "src/fallback/convolution/opr_impl.h"
19 20 21 22 23 24
#include "src/fallback/convolution/run_conv.h"
#include "src/naive/convolution/helper.h"
#include "src/naive/handle.h"

#include "midout.h"

25 26 27 28
#if MEGDNN_AARCH64 || MEGDNN_ARMV7
#include "src/arm_common/convolution/opr_impl.h"
#endif

29
#include <cstring>
30
#include <unordered_map>
31 32 33 34 35 36 37 38 39 40 41

MIDOUT_DECL(megdnn_fb_convbwd_float)

using namespace megdnn;
using namespace fallback;

namespace {
template <typename T>
void incr_ptr(T*& dst, ptrdiff_t delta) {
    dst = reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(dst) + delta);
}
42

43 44 45 46 47 48
}  // namespace

class ConvolutionImpl::AlgoPack : NonCopyableObj {
    AlgoFallback algo_fallback;
    AlgoNaive algo_naive;
    SmallVector<std::unique_ptr<AlgoBase>> refhold;
49 50
    SmallVector<AlgoBase*> m_all_algos;
    AlgoBase::Mapper m_all_algos_map;
51 52 53 54 55
public:
    AlgoPack() {
        static CpuOprDelegationStorage<1> storage;
        auto conv_bias_opr = storage.get<ConvBias, 0>();
        auto&& conv_bias_algo =
56
                static_cast<ConvBiasImpl*>(conv_bias_opr)->get_all_packed_algo();
57 58
        for (auto&& algorithm : conv_bias_algo) {
            // fallback algo
59
            refhold.emplace_back(new AlgoDefault(algorithm));
60
            m_all_algos.emplace_back(refhold.back().get());
61 62
        }

63 64 65 66 67 68
        m_all_algos.emplace_back(&algo_fallback);
        m_all_algos.emplace_back(&algo_naive);

        for (auto&& algo : m_all_algos) {
            m_all_algos_map.emplace(algo->info().desc, algo);
        }
69
    }
70 71 72

    const SmallVector<AlgoBase*>& all_algos() const { return m_all_algos; }
    const AlgoBase::Mapper& all_algos_map() const { return m_all_algos_map; }
73 74
};

75 76 77 78 79 80 81
const ConvolutionImpl::AlgoPack& ConvolutionImpl::algo_pack() {
    static AlgoPack algo_pack;
    return algo_pack;
}

SmallVector<ConvolutionImpl::AlgoBase*> ConvolutionImpl::get_all_packed_algo() {
    return algo_pack().all_algos();
82
}
83 84 85 86 87 88

SmallVector<ConvolutionImpl::AlgoBase*> ConvolutionImpl::select_algo_type(
        ConvAlgoTypePack target_type) {
    megdnn_assert(nr_type_contain(target_type.data_type),
                  "ConvBias algo selection only support one type");
    SmallVector<ConvolutionImpl::AlgoBase*> algos;
89
    for (auto&& algo : get_all_packed_algo()) {
90 91 92 93 94 95 96 97 98
        auto algo_type = algo->get_algo_type();
        if (contain_data_type(algo_type.data_type, target_type.data_type) &&
            algo_type.algo_category == target_type.algo_category) {
            algos.push_back(algo);
        }
    }
    return algos;
}

99 100 101
bool ConvolutionImpl::is_naive_algo(ConvolutionImpl::Algorithm* algo) {
    return algo == nullptr || strcmp(algo->name(), "DEFAULT") == 0;
}
102 103

#define NCB_ALGO_FUNC(name, algo, param) \
104
    static_cast<AlgoBase*>(algo)->name(param)
105

106 107
void ConvolutionImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_in filter,
                           _megdnn_tensor_out dst,
108
                           const PreprocessedFilter* preprocessed_filter,
109
                           _megdnn_workspace workspace) {
110 111
    auto fparam = make_ncb_kern_param(src, filter, dst, preprocessed_filter,
                                      workspace);
112
    auto&& algo = get_algorithm(fparam, workspace.size);
113
    if (!is_naive_algo(algo) &&
114
        NCB_ALGO_FUNC(get_workspace, algo, fparam) <= workspace.size) {
115 116
        exec_with_ncb_kern(fparam, algo);
    } else {
117 118
        naive::ConvolutionForwardImpl::exec(src, filter, dst,
                                            preprocessed_filter, workspace);
119 120 121
    }
}

122
void ConvolutionImpl::exec_preprocess(const TensorLayout& src_layout,
123 124 125 126
                                      _megdnn_tensor_in filter,
                                      const TensorLayout& dst_layout,
                                      PreprocessedFilter* preprocessed_filter,
                                      _megdnn_workspace workspace) {
127 128 129 130 131
    //! exec_preprocess currently only support preprocess weights before exec,
    //! src/dst will be ignored, just set to nullptr
    TensorND src{nullptr, src_layout}, dst{nullptr, dst_layout};
    auto fparam = make_ncb_kern_param(src, filter, dst, preprocessed_filter,
                                      workspace);
132 133

    //! should not pass workspace_size limit otherwise can not find match algo
134 135 136 137
    auto&& algo = get_algorithm(fparam);
    if (!is_naive_algo(algo) &&
        NCB_ALGO_FUNC(get_preprocess_workspace, algo, fparam) <=
                workspace.size) {
138 139 140 141 142 143 144
        exec_preprocess_with_ncb_kern(fparam, algo);
    } else {
        naive::ConvolutionForwardImpl::exec_preprocess(
                src_layout, filter, dst_layout, preprocessed_filter, workspace);
    }
}

145 146 147 148
size_t ConvolutionImpl::get_workspace_in_bytes(
        const TensorLayout& src, const TensorLayout& filter,
        const TensorLayout& dst,
        const PreprocessedFilter* preprocessed_filter) {
149 150
    auto fparam =
            make_ncb_kern_size_param(src, filter, dst, preprocessed_filter);
151
    auto&& algo = get_algorithm(fparam);
152 153
    if (is_naive_algo(algo)) {
        return naive::ConvolutionForwardImpl::get_workspace_in_bytes(
154
                src, filter, dst, preprocessed_filter);
155
    } else {
156
        return NCB_ALGO_FUNC(get_workspace, algo, fparam);
157 158 159 160 161 162 163
    }
}

size_t ConvolutionImpl::get_preprocess_workspace_in_bytes(
        const TensorLayout& src, const TensorLayout& filter,
        const TensorLayout& dst) {
    auto fparam = make_ncb_kern_size_param(src, filter, dst, nullptr);
164
    auto&& algo = get_algorithm(fparam);
165 166 167 168
    if (is_naive_algo(algo)) {
        return naive::ConvolutionForwardImpl::get_preprocess_workspace_in_bytes(
                src, filter, dst);
    } else {
169
        return NCB_ALGO_FUNC(get_preprocess_workspace, algo, fparam);
170 171 172 173 174
    }
}

SmallVector<TensorLayout> ConvolutionImpl::deduce_preprocessed_filter_layout(
        const TensorLayout& src, const TensorLayout& filter,
175
        const TensorLayout& dst) {
176
    auto fparam = make_ncb_kern_size_param(src, filter, dst, nullptr);
177
    auto&& algo = get_algorithm(fparam);
178 179 180 181
    if (is_naive_algo(algo)) {
        return naive::ConvolutionForwardImpl::deduce_preprocessed_filter_layout(
                src, filter, dst);
    } else {
182
        return NCB_ALGO_FUNC(deduce_preprocessed_filter_layout, algo, fparam);
183 184 185 186 187 188
    }
}

std::vector<ConvolutionImpl::Algorithm*> ConvolutionImpl::get_all_algorithms(
        const TensorLayout& src, const TensorLayout& filter,
        const TensorLayout& dst) {
189
    auto fparam = make_ncb_kern_size_param(src, filter, dst, nullptr);
190 191 192 193 194 195 196 197 198 199 200
    auto ret = get_all_algorithms_with_ncb(fparam);
    if (ret.empty()) {
        return naive::ConvolutionForwardImpl::get_all_algorithms(src, filter,
                                                                 dst);
    }
    return ret;
}

ConvolutionImpl::Algorithm* ConvolutionImpl::get_algorithm_heuristic(
        const TensorLayout& src, const TensorLayout& filter,
        const TensorLayout& dst, size_t workspace_limit_in_bytes,
201
        const AlgoAttribute& attr) {
202
    auto fparam = make_ncb_kern_size_param(src, filter, dst, nullptr);
203
    auto result = get_algorithm_heuristic_with_ncb(
204
            fparam, workspace_limit_in_bytes, attr);
205 206
    if (result == nullptr) {
        result = naive::ConvolutionForwardImpl::get_algorithm_heuristic(
207
                src, filter, dst, workspace_limit_in_bytes, attr);
208 209 210 211 212 213
    }
    return result;
}

ConvolutionImpl::NCBKernSizeParam ConvolutionImpl::make_ncb_kern_size_param(
        const TensorLayout& src, const TensorLayout& filter,
214 215
        const TensorLayout& dst,
        const PreprocessedFilter* preprocessed_filter) {
216 217 218 219 220 221 222 223
    auto safe_u32 = [](size_t v) -> uint32_t {
        megdnn_assert(v <= std::numeric_limits<uint32_t>::max(),
                      "value too large: %zu", v);
        return v;
    };
    size_t spatial_pos;
    if (param().format == Param::Format::NCHW88 ||
        param().format == Param::Format::NCHW8 ||
224
        param().format == Param::Format::NCHW4 ||
225
        param().format == Param::Format::NCHW44_DOT ||
226
        param().format == Param::Format::NCHW44) {
227
        spatial_pos = 2;
228
    } else if (param().format == Param::Format::NCHW) {
229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251
        spatial_pos = 2;
    } else if (param().format == Param::Format::NHWC) {
        spatial_pos = 1;
    } else {
        megdnn_assert(0, "invalid conv format %d",
                      static_cast<int>(param().format));
    }
    size_t nr_threads = static_cast<naive::HandleImpl*>(handle())
                                ->megcore_dispatcher()
                                ->nr_threads();

    return {safe_u32(src[0]),
            {{safe_u32(src[spatial_pos]), safe_u32(src[spatial_pos + 1])}},
            {{safe_u32(dst[spatial_pos]), safe_u32(dst[spatial_pos + 1])}},
            check_layout_fwd(src, filter, dst),
            src.dtype,
            filter.dtype,
            dst.dtype,
            src.stride[0],
            dst.stride[0],
            {src.stride[0], src.stride[1], src.stride[2], src.stride[3]},
            {dst.stride[0], dst.stride[1], dst.stride[2], dst.stride[3]},
            param().compute_mode,
252 253
            nr_threads,
            preprocessed_filter};
254 255 256 257
}

ConvolutionImpl::NCBKernParam ConvolutionImpl::make_ncb_kern_param(
        _megdnn_tensor_in src, _megdnn_tensor_in filter, _megdnn_tensor_out dst,
258
        const PreprocessedFilter* preprocessed_filter,
259 260
        _megdnn_workspace workspace) {
    NCBKernParam ret;
261 262
    static_cast<NCBKernSizeParam&>(ret) = make_ncb_kern_size_param(
            src.layout, filter.layout, dst.layout, preprocessed_filter);
263 264 265 266 267 268 269 270
    ret.src_ptr = src.raw_ptr;
    ret.filter_ptr = filter.raw_ptr;
    ret.dst_ptr = dst.raw_ptr;
    ret.workspace_ptr = workspace.raw_ptr;
    ret.workspace_size = workspace.size;
    return ret;
}

271 272
void ConvolutionImpl::exec_preprocess_with_ncb_kern(const NCBKernParam& param,
                                                    Algorithm* algo) {
273 274 275
    auto&& kerns = NCB_ALGO_FUNC(dispatch_preprocess_kern, algo, param);
    auto&& fallback_handle = handle();
    for (auto&& kernel : kerns) {
276 277 278 279
        megdnn_assert(
                param.filter_meta.format == Param::Format::NCHW ||
                        param.filter_meta.format == Param::Format::NHWC ||
                        param.filter_meta.format == Param::Format::NCHW88 ||
280 281
                        param.filter_meta.format == Param::Format::NCHW44 ||
                        param.filter_meta.format == Param::Format::NCHW44_DOT,
282 283 284 285 286 287 288 289 290 291
                "invalid conv format");
        auto run = [param, kernel](size_t index, size_t thread_id) {
            CpuNDRange ndrange_id(kernel.global_size, index);
            kernel.kern(param, {thread_id, ndrange_id});
        };
        static_cast<naive::HandleImpl*>(fallback_handle)
                ->dispatch_kern(run, kernel.global_size.total_size());
    }
}

292 293
void ConvolutionImpl::exec_with_ncb_kern(const NCBKernParam& param,
                                         Algorithm* algo) {
294 295 296
    auto&& kerns = NCB_ALGO_FUNC(dispatch_kern, algo, param);
    auto&& fallback_handle = handle();
    for (auto&& kernel : kerns) {
297 298 299 300
        megdnn_assert(
                param.filter_meta.format == Param::Format::NCHW ||
                        param.filter_meta.format == Param::Format::NHWC ||
                        param.filter_meta.format == Param::Format::NCHW88 ||
301 302
                        param.filter_meta.format == Param::Format::NCHW44 ||
                        param.filter_meta.format == Param::Format::NCHW44_DOT,
303
                "invalid conv format");
304
        auto run = [param, kernel](size_t index, size_t thread_id) {
305
            CpuNDRange ndrange_id(kernel.global_size, index);
306
            kernel.kern(param, {thread_id, ndrange_id});
307 308 309 310 311 312 313 314
        };
        static_cast<naive::HandleImpl*>(fallback_handle)
                ->dispatch_kern(run, kernel.global_size.total_size());
    }
}

ConvolutionImpl::Algorithm* ConvolutionImpl::get_algorithm_heuristic_with_ncb(
        const NCBKernSizeParam& param, size_t workspace_limit_in_bytes,
315
        const AlgoAttribute& attr) {
316 317 318 319 320 321
    auto algo_data_type = param.deduce_algo_data_type();
    auto suggest_category_order = suggest_algo_category_order(param);
    for (auto category : suggest_category_order) {
        auto&& origin_algos = select_algo_type({algo_data_type, category});
        ConvolutionImpl::Algorithm* heuristic_algo = nullptr;
        for (auto i : origin_algos) {
322 323 324
            bool usable_attribute = static_cast<AlgoBase*>(i)->usable_attribute(
                    param, AlgoSelectionStrategy::HEURISTIC, attr);
            if (usable_attribute &&
325 326 327 328 329 330 331 332 333 334 335 336 337 338 339
                static_cast<AlgoBase*>(i)->get_workspace(param) <=
                        workspace_limit_in_bytes) {
                //! store the first usable algo if no prefer algo, choose it as
                //! the target algo
                if (!heuristic_algo) {
                    heuristic_algo = i;
                }
                //! choose the first prefer algo
                if (i->is_preferred(param)) {
                    return i;
                }
            }
        }
        if (heuristic_algo) {
            return heuristic_algo;
340 341 342 343 344 345 346 347 348
        }
    }
    return nullptr;
}

std::vector<ConvolutionImpl::Algorithm*>
ConvolutionImpl::get_all_algorithms_with_ncb(const NCBKernSizeParam& param) {
    std::vector<Algorithm*> ret;
    std::vector<Algorithm*> prefer_algos;
349
    for (auto&& i : get_all_packed_algo()) {
350 351
        if (i->usable(param, AlgoSelectionStrategy::FULL_RUN)) {
            if (i->is_preferred(param)) {
352 353 354 355 356 357 358 359 360 361
                prefer_algos.push_back(i);
            } else {
                ret.push_back(i);
            }
        }
    }
    ret.insert(ret.begin(), prefer_algos.begin(), prefer_algos.end());
    return ret;
}

362 363
ConvolutionImpl::Algorithm* ConvolutionImpl::get_algorithm_from_desc(
        const AlgorithmDesc& desc) {
364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385
    if (!desc.valid()) {
        return nullptr;
    } else {
        switch (desc.handle_type) {
            case Handle::HandleType::FALLBACK: {
                const auto& map = algo_pack().all_algos_map();
                megdnn_assert(map.find(desc) != map.end());
                return map.at(desc);
            }
            case Handle::HandleType::NAIVE: {
                auto algo = static_cast<naive::HandleImpl*>(handle())
                                    ->default_conv_fwd_algo();
                megdnn_assert(algo->info().desc == desc);
                return algo;
            }
            default:
                megdnn_throw("Unknown handle type");
                return nullptr;
        }
    }
}

386 387
ConvolutionImpl::Algorithm* ConvolutionImpl::get_algorithm(
        const NCBKernSizeParam& param, size_t workspace_size) {
388
    if (auto algo = get_algorithm_from_desc(execution_policy().algo)) {
389
        return algo;
390 391 392
    }
    if (!m_prev_selected_algo ||
        memcmp(&m_prev_selected_algo_sizep, &param, sizeof(NCBKernSizeParam))) {
393 394
        m_prev_selected_algo = get_algorithm_heuristic_with_ncb(
                param, workspace_size, AlgoAttribute::DEFAULT);
395 396 397 398 399
        m_prev_selected_algo_sizep = param;
    }
    return m_prev_selected_algo;
}

400 401 402 403 404 405 406 407 408 409
SmallVector<AlgoCategory> ConvolutionImpl::suggest_algo_category_order(
        const NCBKernSizeParam& param) const {
    static CpuOprDelegationStorage<1> storage;
    auto conv_bias_opr = storage.get<ConvBias, 0>();
    auto conv_bias_param =
            ConvolutionImpl::AlgoDefault::init_conv_bias_param(param);
    return static_cast<ConvBiasImpl*>(conv_bias_opr)
            ->suggest_algo_category_order(conv_bias_param);
}

410 411 412 413 414
const char* ConvolutionImpl::get_algorithm_set_name() const {
    // fallback version 0
    return "F0";
}

415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432
ConvolutionImpl::AlgoDataType
ConvolutionImpl::NCBKernSizeParam::deduce_algo_data_type() const {
    if (src_type.enumv() == DTypeEnum::Float32) {
        return ConvolutionImpl::AlgoDataType::FLOAT32;
#if !MEGDNN_DISABLE_FLOAT16
    } else if (src_type.enumv() == DTypeEnum::Float16) {
        return ConvolutionImpl::AlgoDataType::FLOAT16;
#endif
    } else if (src_type.enumv() == DTypeEnum::Int8 ||
               src_type.enumv() == DTypeEnum::QuantizedS8) {
        if (dst_type.enumv() == DTypeEnum::Int16) {
            return ConvolutionImpl::AlgoDataType::INT8X8X16;
        } else {
            return ConvolutionImpl::AlgoDataType::QINT8X8X32;
        }
    } else if (src_type.enumv() == DTypeEnum::Quantized8Asymm) {
        return ConvolutionImpl::AlgoDataType::QUINT8X8X32;
    } else {
M
Megvii Engine Team 已提交
433
        megdnn_throw(ssprintf("not support data type of %s * %s -> %s\n",
434 435 436 437 438
                              src_type.name(), filter_type.name(),
                              dst_type.name()));
    }
}

439 440
/* ===================== ConvolutionBackwardData ===================== */

441 442 443 444
class ConvolutionBackwardDataImpl::AlgoPack : NonCopyableObj {
    AlgoNaive algo_naive;
    AlgoDirect algo_direct;
    AlgoMatrixMul algo_matmul;
445 446
    SmallVector<AlgoBase*> m_all_algos;
    AlgoBase::Mapper m_all_algos_map;
447 448 449

public:
    AlgoPack() {
450 451 452 453 454 455 456
        m_all_algos.emplace_back(&algo_matmul);
        m_all_algos.emplace_back(&algo_direct);
        m_all_algos.emplace_back(&algo_naive);

        for (auto&& algo : m_all_algos) {
            m_all_algos_map.emplace(algo->info().desc, algo);
        }
457
    }
458 459
    const SmallVector<AlgoBase*>& all_algos() const { return m_all_algos; }
    const AlgoBase::Mapper& all_algos_map() const { return m_all_algos_map; }
460
};
461 462 463 464 465
const ConvolutionBackwardDataImpl::AlgoPack&
ConvolutionBackwardDataImpl::algo_pack() {
    static AlgoPack algo_pack;
    return algo_pack;
}
466 467

SmallVector<ConvolutionBackwardDataImpl::AlgoBase*>
468 469
ConvolutionBackwardDataImpl::get_all_packed_algo() {
    return algo_pack().all_algos();
470
}
471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515

void ConvolutionBackwardDataImpl::exec(_megdnn_tensor_in filter,
                                       _megdnn_tensor_in diff,
                                       _megdnn_tensor_out grad,
                                       _megdnn_workspace workspace) {
    if (param().format == param::Convolution::Format::NHWCD4 ||
        param().format == param::Convolution::Format::NCHW4) {
        return naive::ConvolutionBackwardDataImpl::exec(filter, diff, grad,
                                                        workspace);
    }
    auto fparam = make_ncb_kern_param(filter, diff, grad, workspace);
    return exec_with_ncb_kern(fparam);
}

size_t ConvolutionBackwardDataImpl::get_workspace_in_bytes(
        const TensorLayout& filter, const TensorLayout& diff,
        const TensorLayout& grad) {
    if (param().format == param::Convolution::Format::NHWCD4 ||
        param().format == param::Convolution::Format::NCHW4) {
        return naive::ConvolutionBackwardDataImpl::get_workspace_in_bytes(
                filter, diff, grad);
    }
    auto fparam = make_ncb_kern_size_param(filter, diff, grad);
    return get_workspace_with_ncb(fparam);
}

std::vector<ConvolutionBackwardDataImpl::Algorithm*>
ConvolutionBackwardDataImpl::get_all_algorithms(const TensorLayout& filter,
                                                const TensorLayout& diff,
                                                const TensorLayout& grad) {
    if (param().format == param::Convolution::Format::NHWCD4 ||
        param().format == param::Convolution::Format::NCHW4) {
        return naive::ConvolutionBackwardDataImpl::get_all_algorithms(
                filter, diff, grad);
    }
    auto fparam = make_ncb_kern_size_param(filter, diff, grad);
    auto ret = get_all_algorithms_with_ncb(fparam);
    megdnn_assert(!ret.empty(), "no usable conv fwd algorithm");
    return ret;
}

ConvolutionBackwardDataImpl::Algorithm*
ConvolutionBackwardDataImpl::get_algorithm_heuristic(
        const TensorLayout& filter, const TensorLayout& diff,
        const TensorLayout& grad, size_t workspace_limit_in_bytes,
516
        const AlgoAttribute& attr) {
517 518 519
    if (param().format == param::Convolution::Format::NHWCD4 ||
        param().format == param::Convolution::Format::NCHW4) {
        return naive::ConvolutionBackwardDataImpl::get_algorithm_heuristic(
520
                filter, diff, grad, workspace_limit_in_bytes, attr);
521 522 523
    }
    auto fparam = make_ncb_kern_size_param(filter, diff, grad);
    return get_algorithm_heuristic_with_ncb(fparam, workspace_limit_in_bytes,
524
                                            attr);
525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595
}

ConvolutionBackwardDataImpl::NCBKernSizeParam
ConvolutionBackwardDataImpl::make_ncb_kern_size_param(
        const TensorLayout& filter, const TensorLayout& diff,
        const TensorLayout& grad) {
    auto safe_u32 = [](size_t v) -> uint32_t {
        megdnn_assert(v <= std::numeric_limits<uint32_t>::max(),
                      "value too large: %zu", v);
        return v;
    };
    size_t spatial_pos;
    if (param().format == Param::Format::NCHW) {
        spatial_pos = 2;
    } else {
        megdnn_assert(param().format == Param::Format::NHWC,
                      "invalid conv format");
        spatial_pos = 1;
    }
    auto grad_fwd = grad;
    auto filter_fwd = filter;
    auto diff_fwd = diff;

    std::swap(grad_fwd.dtype, diff_fwd.dtype);

    return {
            safe_u32(diff[0]),
            {{safe_u32(diff[spatial_pos]), safe_u32(diff[spatial_pos + 1])}},
            {{safe_u32(grad[spatial_pos]), safe_u32(grad[spatial_pos + 1])}},
            check_layout_fwd(grad_fwd, filter_fwd, diff_fwd),
            diff.dtype,
            filter.dtype,
            grad.dtype,
            diff,
            filter,
            grad,
            diff.stride[0],
            grad.stride[0],
            0,
            0,
            0,
            param().compute_mode,
    };
}

ConvolutionBackwardDataImpl::NCBKernParam
ConvolutionBackwardDataImpl::make_ncb_kern_param(_megdnn_tensor_in filter,
                                                 _megdnn_tensor_in diff,
                                                 _megdnn_tensor_out grad,
                                                 _megdnn_workspace workspace) {
    NCBKernParam ret;
    static_cast<NCBKernSizeParam&>(ret) =
            make_ncb_kern_size_param(filter.layout, diff.layout, grad.layout);

    auto required_workspace_in_bytes = get_workspace_with_ncb(ret);
    megdnn_assert(workspace.size >= required_workspace_in_bytes,
                  "required workspace: %zu; provided workspace: %zu",
                  required_workspace_in_bytes, workspace.size);
    ret.filter_ptr = filter.raw_ptr;
    ret.diff_ptr = diff.raw_ptr;
    ret.grad_ptr = grad.raw_ptr;
    ret.workspace_ptr = workspace.raw_ptr;
    ret.workspace_size = workspace.size;
    return ret;
}

void ConvolutionBackwardDataImpl::exec_with_ncb_kern(
        const NCBKernParam& param) {
    auto p1g = param;
    auto group = p1g.filter_meta.group;
    p1g.filter_meta.group = 1;
596
    auto&& algo = get_algorithm(p1g);
597
    auto kptr = ncb_1g_dispatch_kern(algo, p1g);
598
    if (group == 1 || static_cast<AlgoBase*>(algo)->is_naive()) {
599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647
        auto run = [kptr, param]() { kptr(param); };
        static_cast<naive::HandleImpl*>(handle())->dispatch_kern(run);
    } else {
        megdnn_assert(p1g.filter_meta.format == Param::Format::NCHW ||
                              p1g.filter_meta.format == Param::Format::NHWC,
                      "invalid conv format");
        auto run = [kptr, p1g_orig = p1g, group]() {
            auto p1g = p1g_orig;
            ptrdiff_t istrd, fstrd, ostrd;
            fstrd = p1g.filter_meta.icpg * p1g.filter_meta.ocpg *
                    p1g.filter_meta.spatial[0] * p1g.filter_meta.spatial[1] *
                    p1g.filter_type.size();
            istrd = p1g.filter_meta.ocpg * p1g.diff_type.size();
            ostrd = p1g.filter_meta.icpg * p1g.grad_type.size();
            p1g.diff_extra_mem_size =
                    (group - 1) * p1g.filter_meta.ocpg * p1g.diff_type.size();
            p1g.filter_extra_mem_size =
                    (group - 1) * p1g.filter_meta.icpg * p1g.filter_meta.ocpg *
                    p1g.filter_meta.spatial[0] * p1g.filter_meta.spatial[1] *
                    p1g.filter_type.size();
            p1g.grad_extra_mem_size =
                    (group - 1) * p1g.filter_meta.icpg * p1g.grad_type.size();
            if (p1g.filter_meta.format == Param::Format::NCHW) {
                istrd *= p1g.isz[0] * p1g.isz[1];
                ostrd *= p1g.osz[0] * p1g.osz[1];
                p1g.diff_extra_mem_size *= p1g.isz[0] * p1g.isz[1];
                p1g.grad_extra_mem_size *= p1g.osz[0] * p1g.osz[1];
            } else {
                // must be NHWC. No action performed.
            }
            for (size_t i = 0; i < group; ++i) {
                kptr(p1g);
                incr_ptr(p1g.diff_ptr, istrd);
                incr_ptr(p1g.filter_ptr, fstrd);
                incr_ptr(p1g.grad_ptr, ostrd);
                p1g.diff_extra_mem_size -= istrd;
                p1g.filter_extra_mem_size -= fstrd;
                p1g.grad_extra_mem_size -= ostrd;
            }
        };
        static_cast<naive::HandleImpl*>(handle())->dispatch_kern(run);
    }
}

size_t ConvolutionBackwardDataImpl::get_workspace_with_ncb(
        const NCBKernSizeParam& param) {
    if (param.filter_meta.group != 1) {
        auto p1g = param;
        p1g.filter_meta.group = 1;
648 649
        auto algo = get_algorithm(p1g);
        return ncb_1g_get_workspace(algo, p1g);
650
    }
651 652
    auto algo = get_algorithm(param);
    return ncb_1g_get_workspace(algo, param);
653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668
}

std::vector<ConvolutionBackwardDataImpl::Algorithm*>
ConvolutionBackwardDataImpl::get_all_algorithms_with_ncb(
        const NCBKernSizeParam& param) {
    if (param.filter_meta.group != 1) {
        auto p1g = param;
        p1g.filter_meta.group = 1;
        return ncb_1g_get_all_algorithms(p1g);
    }
    return ncb_1g_get_all_algorithms(param);
}

ConvolutionBackwardDataImpl::Algorithm*
ConvolutionBackwardDataImpl::get_algorithm_heuristic_with_ncb(
        const NCBKernSizeParam& param, size_t workspace_limit_in_bytes,
669
        const AlgoAttribute& attr) {
670 671 672 673
    if (param.filter_meta.group != 1) {
        auto p1g = param;
        p1g.filter_meta.group = 1;
        return ncb_1g_get_algorithm_heuristic(p1g, workspace_limit_in_bytes,
674
                                              attr);
675 676
    }
    return ncb_1g_get_algorithm_heuristic(param, workspace_limit_in_bytes,
677
                                          attr);
678 679 680 681 682
}

size_t ConvolutionBackwardDataImpl::ncb_1g_get_workspace(
        Algorithm* algo, const NCBKernSizeParam& param) {
    megdnn_assert(param.filter_meta.group == 1);
683
    if (algo->handle_type() == Handle::HandleType::FALLBACK) {
684 685 686 687 688 689 690 691 692 693
        return static_cast<AlgoBase*>(algo)->get_workspace(this, param);
    }
    return 0;
}

ConvolutionBackwardDataImpl::ncb_kern_t
ConvolutionBackwardDataImpl::ncb_1g_dispatch_kern(
        Algorithm* algo, const NCBKernSizeParam& param) {
    megdnn_assert(param.filter_meta.group == 1);

694
    if (algo->handle_type() == Handle::HandleType::FALLBACK) {
695 696 697
        return static_cast<AlgoBase*>(algo)->dispatch_kern(this, param);
    }

M
Megvii Engine Team 已提交
698
    megdnn_throw("no suitable ConvolutionBackwardData algorithm");
699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714
}

bool ConvolutionBackwardDataImpl::is_matrix_mul_preferred(
        const NCBKernSizeParam& param) {
    auto&& fm = param.filter_meta;
    auto OC = fm.ocpg, IC = fm.icpg;

    return (OC * IC >= 32) ||
           (fm.spatial[0] == 1 && fm.spatial[1] == 1 && fm.padding[0] == 0 &&
            fm.padding[1] == 0 && fm.stride[0] == 1 && fm.stride[1] == 1);
}

std::vector<ConvolutionBackwardDataImpl::Algorithm*>
ConvolutionBackwardDataImpl::ncb_1g_get_all_algorithms(
        const NCBKernSizeParam& param) {
    std::vector<Algorithm*> ret;
715
    std::vector<Algorithm*> prefer_algos;
716
    for (auto&& i : get_all_packed_algo()) {
717 718 719
        if (i->usable(this, param)) {
            if (i->is_preferred(param)) {
                prefer_algos.push_back(i);
720
            } else {
721
                ret.push_back(i);
722 723 724
            }
        }
    }
725
    ret.insert(ret.begin(), prefer_algos.begin(), prefer_algos.end());
726 727 728 729 730 731
    return ret;
}

ConvolutionBackwardDataImpl::Algorithm*
ConvolutionBackwardDataImpl::ncb_1g_get_algorithm_heuristic(
        const NCBKernSizeParam& param, size_t workspace_limit_in_bytes,
732
        const AlgoAttribute& attr) {
733 734
    for (auto i : ncb_1g_get_all_algorithms(param)) {
        if (ncb_1g_get_workspace(i, param) <= workspace_limit_in_bytes) {
735
            if (i->contain_attribute(attr)) {
736 737 738 739 740 741 742 743
                return i;
            }
        }
    }
    megdnn_assert(0,
                  "no suitable algorithm found within given workspace limit");
}

744
ConvolutionBackwardDataImpl::Algorithm*
745 746
ConvolutionBackwardDataImpl::get_algorithm_from_desc(
        const AlgorithmDesc& desc) {
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
    if (!desc.valid()) {
        return nullptr;
    } else {
        switch (desc.handle_type) {
            case Handle::HandleType::FALLBACK: {
                const auto& map = algo_pack().all_algos_map();
                megdnn_assert(map.find(desc) != map.end());
                return map.at(desc);
            }
#if MEGDNN_AARCH64 || MEGDNN_ARMV7
            case Handle::HandleType::ARM_COMMON:
            case Handle::HandleType::AARCH64:
            case Handle::HandleType::ARMV7:
                return arm_common::ConvolutionBackwardDataImpl::
                        get_algo_from_desc(desc);
#endif
            case Handle::HandleType::NAIVE: {
                auto algo = static_cast<naive::HandleImpl*>(handle())
                                    ->default_conv_bwd_data_algo();
                megdnn_assert(algo->info().desc == desc);
                return algo;
            }
            default:
                megdnn_throw("Unknown handle type");
                return nullptr;
        }
    }
}


777 778
ConvolutionBackwardDataImpl::Algorithm*
ConvolutionBackwardDataImpl::get_algorithm(const NCBKernSizeParam& param) {
779
    if (auto algo = get_algorithm_from_desc(execution_policy().algo)) {
780
        return algo;
781 782 783 784
    }
    if (!m_prev_selected_algo ||
        memcmp(&m_prev_selected_algo_sizep, &param, sizeof(NCBKernSizeParam))) {
        m_prev_selected_algo = ncb_1g_get_algorithm_heuristic(
785 786
                param, std::numeric_limits<size_t>::max(),
                AlgoAttribute::DEFAULT);
787 788 789 790 791 792 793 794 795 796 797
        m_prev_selected_algo_sizep = param;
    }
    return m_prev_selected_algo;
}

const char* ConvolutionBackwardDataImpl::get_algorithm_set_name() const {
    // fallback version 0
    return "FALLBACK_CONVOLUTION_BACKWARD_DATA_IMPL0";
}

// vim: syntax=cpp.doxygen