opr_impl.cpp 27.8 KB
Newer Older
1
/**
2
 g * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
3
 *
4
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
5 6 7
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
8 9
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
10 11 12 13 14 15
 */
#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/conv_bias/algos.h"
16
#include "src/fallback/conv_bias/conv1x1/algos.h"
17
#include "src/fallback/conv_bias/conv1x1/algos_conv1x1_gemv.h"
18 19
#include "src/fallback/conv_bias/im2col/algos.h"
#include "src/fallback/conv_bias/opr_impl.h"
20
#include "src/fallback/convolution/opr_impl.h"
21 22 23
#include "src/naive/convolution/algorithms.h"
#include "src/naive/handle.h"

24 25 26 27 28 29 30 31
#if MEGDNN_X86
#include "src/x86/conv_bias/opr_impl.h"
#elif MEGDNN_AARCH64
#include "src/aarch64/conv_bias/opr_impl.h"
#elif MEGDNN_ARMV7
#include "src/armv7/conv_bias/opr_impl.h"
#endif

32 33 34 35 36
#include <cstring>

using namespace megdnn;
using namespace fallback;

37
size_t megdnn::fallback::pack_size(param::ConvBias::Format format) {
38
    switch (format) {
39
        case param::ConvBias::Format::NCHW44:
40
        case param::ConvBias::Format::NCHW44_DOT:
41 42 43 44 45 46 47 48 49
        case param::ConvBias::Format::NCHW4:
            return 4_z;
        case param::ConvBias::Format::NCHW88:
            return 8_z;
        default:
            return 1_z;
    }
}

50 51 52 53 54 55 56 57
namespace {
template <typename T>
void incr_ptr(T*& dst, ptrdiff_t delta) {
    dst = reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(dst) + delta);
}

}  // namespace

58 59 60 61 62 63 64 65 66 67 68 69 70 71
#if MEGDNN_X86
#define SKIP_GEMV()
//! As we haven't direct conv for int8x8x16 yet, if we disable gemv here, it may
//! fallback to naive implementation, which may cause performance very low, so
//! here we just enable im2col for gemv in x86 backend.
//! FIXME: remove it when we add direct conv support for int8x8x16
#else
#define SKIP_GEMV()                                                            \
    if (algo->algoset() == MatrixMulImpl::AlgoBase::AlgoSet::ALGO_TYPE_GEMV) { \
        continue;                                                              \
    }
#endif


72 73 74
class ConvBiasImpl::AlgoPack : NonCopyableObj {
    AlgoNaive algo_naive;
    SmallVector<std::unique_ptr<AlgoBase>> refhold;
75 76
    SmallVector<AlgoBase*> m_all_algos;
    AlgoBase::Mapper m_all_algos_map;
77 78

public:
79

80
    AlgoPack() {
81
        refhold.emplace_back(new AlgoConv1x1Gemv());
82
        m_all_algos.emplace_back(refhold.back().get());
83

84 85
        static CpuOprDelegationStorage<> storage;
        auto matmul_opr = storage.get<MatrixMul>();
86 87
        auto&& matmul_algos = static_cast<fallback::MatrixMulImpl*>(matmul_opr)
                                      ->get_all_packed_algo();
88
        for (auto&& algo : matmul_algos) {
89 90 91 92 93 94
#if MEGDNN_X86
//! As we haven't direct conv for int8x8x16 yet, if we disable gemv here, it may
//! fallback to naive implementation, which may cause performance very low, so
//! here we just enable im2col for gemv in x86 backend.
//! FIXME: remove it when we add direct conv support for int8x8x16
#else
95 96 97 98
            if (algo->algoset() ==
                MatrixMulImpl::AlgoBase::AlgoSet::ALGO_TYPE_GEMV) {
                continue;
            }
99 100
#endif

101 102 103 104
//! As we haven't riscv64 postprocess yet, im2col and conv1x1 can not pass ci
//! test. so we just disable all im2col and conv1x1 in riscv64
//! FIXME: remove it when impl postprocess for riscv64
#if !MEGDNN_RISCV64
105 106 107 108
            for (size_t ohw_tile_size : {192, 384, 96, 48, 24}) {
                refhold.emplace_back(new AlgoIm2col(
                        static_cast<MatrixMulImpl::AlgoBase*>(algo),
                        ohw_tile_size));
109
                m_all_algos.emplace_back(refhold.back().get());
110
            }
111
            for (size_t oc_tile_size : {48, 24}) {
112
                refhold.emplace_back(new AlgoConv1x1(
113 114
                        static_cast<MatrixMulImpl::AlgoBase*>(algo),
                        oc_tile_size));
115
                m_all_algos.emplace_back(refhold.back().get());
116
            }
117 118
#endif

119
#if 0
120 121 122 123 124
        //! As these algos maybe very slow, it will make fastrun search slow, so
        //! we disable it, but for the test of strategyhelper, we just keep it.
        //! FIXME: I do not know a better way to do it.
            refhold.emplace_back(new AlgoWinogradF32(
                    static_cast<MatrixMulImpl::AlgoBase*>(algo)));
125
            m_all_algos.emplace_back(refhold.back().get());
126 127
            refhold.emplace_back(new AlgoWinogradF32_4x4(
                    static_cast<MatrixMulImpl::AlgoBase*>(algo)));
128
            m_all_algos.emplace_back(refhold.back().get());
129 130
            refhold.emplace_back(new AlgoWinogradQS8(
                    static_cast<MatrixMulImpl::AlgoBase*>(algo)));
131
            m_all_algos.emplace_back(refhold.back().get());
132 133
            refhold.emplace_back(new AlgoWinogradQS8_8x8(
                    static_cast<MatrixMulImpl::AlgoBase*>(algo)));
134
            m_all_algos.emplace_back(refhold.back().get());
135 136
#endif
        }
137 138 139 140 141
        m_all_algos.emplace_back(&algo_naive);

        for (auto&& algo : m_all_algos) {
            m_all_algos_map.emplace(algo->info().desc, algo);
        }
142
    }
143 144
    const SmallVector<AlgoBase*>& all_algos() const { return m_all_algos; }
    const AlgoBase::Mapper& all_algos_map() const { return m_all_algos_map; }
145 146
};

147 148 149 150 151 152 153
const ConvBiasImpl::AlgoPack& ConvBiasImpl::algo_pack() {
    static AlgoPack algo_pack;
    return algo_pack;
}

SmallVector<ConvBiasImpl::AlgoBase*> ConvBiasImpl::get_all_packed_algo() {
    return algo_pack().all_algos();
154
}
155 156 157 158 159 160

SmallVector<ConvBiasImpl::AlgoBase*> ConvBiasImpl::select_algo_type(
        ConvAlgoTypePack target_type) {
    megdnn_assert(nr_type_contain(target_type.data_type),
                  "ConvBias algo selection only support one type");
    SmallVector<ConvBiasImpl::AlgoBase*> algos;
161
    for (auto&& algo : get_all_packed_algo()) {
162 163 164 165 166 167 168 169 170
        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;
}

171 172 173
bool ConvBiasImpl::is_naive_algo(ConvBiasImpl::Algorithm* algo) {
    return algo == nullptr || strcmp(algo->name(), "DEFAULT") == 0;
}
174 175

#define NCB_ALGO_FUNC(name, algo, param) \
176
    static_cast<AlgoBase*>(algo)->name(param)
177

178 179
void ConvBiasImpl::exec(_megdnn_tensor_in src, _megdnn_tensor_in filter,
                        _megdnn_tensor_in bias, _megdnn_tensor_in z,
180 181 182
                        _megdnn_tensor_out dst,
                        const PreprocessedFilter* preprocessed_filter,
                        _megdnn_workspace workspace) {
183
    check_exec(src.layout, filter.layout, bias.layout, z.layout, dst.layout,
184 185 186
               workspace.size, preprocessed_filter);
    auto fparam = make_ncb_kern_param(src, filter, bias, dst, workspace,
                                      preprocessed_filter);
187
    auto&& algo = get_algorithm(fparam, workspace.size);
188
    if (!is_naive_algo(algo) &&
189
        NCB_ALGO_FUNC(get_workspace, algo, fparam) <= workspace.size) {
190 191
        exec_with_ncb_kern(fparam, algo);
    } else {
192 193
        naive::ConvBiasForwardImpl::exec(src, filter, bias, z, dst,
                                         preprocessed_filter, workspace);
194 195 196
    }
}

197 198
void ConvBiasImpl::exec_preprocess(const TensorLayout& src_layout,
                                   _megdnn_tensor_in filter,
199
                                   _megdnn_tensor_in bias,
200 201 202 203
                                   const TensorLayout& z_layout,
                                   const TensorLayout& dst_layout,
                                   PreprocessedFilter* preprocessed_filter,
                                   _megdnn_workspace workspace) {
204 205 206
    //! exec_preprocess currently only support preprocess weights and bias
    //! before exec, src/dst/z will be ignored, just set to nullptr
    TensorND src{nullptr, src_layout}, dst{nullptr, dst_layout};
207 208
    auto fparam = make_ncb_kern_param(src, filter, bias, dst, workspace,
                                      preprocessed_filter);
209
    //! should not pass workspace_size limit otherwise can not find match algo
210 211 212 213
    auto&& algo = get_algorithm(fparam);
    if (!is_naive_algo(algo) &&
        NCB_ALGO_FUNC(get_preprocess_workspace, algo, fparam) <=
                workspace.size) {
214 215 216
        exec_preprocess_with_ncb_kern(fparam, algo);
    } else {
        naive::ConvBiasForwardImpl::exec_preprocess(
217
                src_layout, filter, bias, z_layout, dst_layout,
218 219 220 221
                preprocessed_filter, workspace);
    }
}

222 223 224 225 226
size_t ConvBiasImpl::get_workspace_in_bytes(
        const TensorLayout& src, const TensorLayout& filter,
        const TensorLayout& bias, const TensorLayout& z,
        const TensorLayout& dst,
        const PreprocessedFilter* preprocessed_filter) {
227 228
    auto fparam = make_ncb_kern_size_param(src, filter, bias, dst,
                                           preprocessed_filter);
229
    auto&& algo = get_algorithm(fparam);
230
    if (is_naive_algo(algo)) {
231 232
        return naive::ConvBiasForwardImpl::get_workspace_in_bytes(
                src, filter, bias, z, dst, preprocessed_filter);
233
    } else {
234 235 236 237 238 239 240 241 242
        return NCB_ALGO_FUNC(get_workspace, algo, fparam);
    }
}

size_t ConvBiasImpl::get_preprocess_workspace_in_bytes(
        const TensorLayout& src, const TensorLayout& filter,
        const TensorLayout& bias, const TensorLayout& z,
        const TensorLayout& dst) {
    auto fparam = make_ncb_kern_size_param(src, filter, bias, dst, nullptr);
243
    auto&& algo = get_algorithm(fparam);
244 245 246 247 248 249 250 251 252 253 254 255 256
    if (is_naive_algo(algo)) {
        return naive::ConvBiasForwardImpl::get_preprocess_workspace_in_bytes(
                src, filter, bias, z, dst);
    } else {
        return NCB_ALGO_FUNC(get_preprocess_workspace, algo, fparam);
    }
}

SmallVector<TensorLayout> ConvBiasImpl::deduce_preprocessed_filter_layout(
        const TensorLayout& src, const TensorLayout& filter,
        const TensorLayout& bias, const TensorLayout& z,
        const TensorLayout& dst) {
    auto fparam = make_ncb_kern_size_param(src, filter, bias, dst, nullptr);
257
    auto&& algo = get_algorithm(fparam);
258 259 260 261 262
    if (is_naive_algo(algo)) {
        return naive::ConvBiasForwardImpl::deduce_preprocessed_filter_layout(
                src, filter, bias, z, dst);
    } else {
        return NCB_ALGO_FUNC(deduce_preprocessed_filter_layout, algo, fparam);
263 264 265 266 267 268 269
    }
}

std::vector<ConvBiasImpl::Algorithm*> ConvBiasImpl::get_all_algorithms(
        const TensorLayout& src, const TensorLayout& filter,
        const TensorLayout& bias, const TensorLayout& z,
        const TensorLayout& dst) {
270
    auto fparam = make_ncb_kern_size_param(src, filter, bias, dst, nullptr);
271 272 273 274 275 276 277 278 279 280 281 282
    auto ret = get_all_algorithms_with_ncb(fparam);
    if (ret.empty()) {
        return naive::ConvBiasForwardImpl::get_all_algorithms(src, filter, bias,
                                                              z, dst);
    }
    return ret;
}

ConvBiasImpl::Algorithm* ConvBiasImpl::get_algorithm_heuristic(
        const TensorLayout& src, const TensorLayout& filter,
        const TensorLayout& bias, const TensorLayout& z,
        const TensorLayout& dst, size_t workspace_limit_in_bytes,
283 284
        const AlgoAttribute& positive_attr,
        const AlgoAttribute& negative_attr) {
285
    auto fparam = make_ncb_kern_size_param(src, filter, bias, dst, nullptr);
286
    auto result = get_algorithm_heuristic_with_ncb(
287
            fparam, workspace_limit_in_bytes, positive_attr, negative_attr);
288 289
    if (result == nullptr) {
        result = naive::ConvBiasForwardImpl::get_algorithm_heuristic(
290 291
                src, filter, bias, z, dst, workspace_limit_in_bytes,
                positive_attr, negative_attr);
292 293 294 295
    }
    return result;
}

296 297
ConvBiasImpl::Algorithm* ConvBiasImpl::get_algorithm_heuristic_with_ncb(
        const NCBKernSizeParam& param, size_t workspace_limit_in_bytes,
298 299
        const AlgoAttribute& positive_attr,
        const AlgoAttribute& negative_attr) {
300 301 302
    if (ConvBiasImpl::param().format == Param::Format::NHWCD4) {
        return nullptr;
    }
303 304 305 306 307 308
    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});
        ConvBiasImpl::Algorithm* heuristic_algo = nullptr;
        for (auto i : origin_algos) {
309
            bool usable_attribute = static_cast<AlgoBase*>(i)->usable_attribute(
310 311
                    param, AlgoSelectionStrategy::HEURISTIC, positive_attr,
                    negative_attr);
312
            if (usable_attribute &&
313 314 315 316 317 318 319 320 321 322 323 324 325 326 327
                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;
328 329 330 331 332
        }
    }
    return nullptr;
}

333 334
ConvBiasImpl::NCBKernSizeParam ConvBiasImpl::make_ncb_kern_size_param(
        const TensorLayout& src, const TensorLayout& filter,
335 336
        const TensorLayout& bias, const TensorLayout& dst,
        const PreprocessedFilter* preprocessed_filter) {
337 338 339 340 341 342 343 344 345
    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 ||
        param().format == Param::Format::NCHW4 ||
346
        param().format == Param::Format::NCHW44 ||
347
        param().format == Param::Format::NCHW44_DOT ||
348 349
        param().format == Param::Format::NCHW ||
        param().format == Param::Format::NCHW32 ||
350
        param().format == Param::Format::NCHW64) {
351
        spatial_pos = 2;
352
    } else if (param().format == Param::Format::NHWC || param().format == Param::Format::NHWCD4) {
353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372
        spatial_pos = 1;
    } else {
        megdnn_assert(0, "invalid conv format %d",
                      static_cast<int>(param().format));
    }
    BiasMode bias_mode;
    if (bias.ndim == 0) {
        bias_mode = BiasMode::NO_BIAS;
    } else if (bias.eq_shape(dst)) {
        bias_mode = BiasMode::BIAS;
    } else {
        //! just check the ndim, the detail shape check is in check_exec
        megdnn_assert(bias.ndim == dst.ndim);
        bias_mode = BiasMode::BROADCAST_CHANNEL_BIAS;
    }

    static_assert(sizeof(CanonizedFilterMeta) ==
                          sizeof(ConvolutionImpl::CanonizedFilterMeta),
                  "sizeof CanonizedFilterMeta in convolution and conv_bias "
                  "should be equal");
373 374
    auto&& fm = check_layout_fwd(src, filter, dst);
    auto& conv_fm = reinterpret_cast<ConvolutionImpl::CanonizedFilterMeta&>(fm);
375

376 377 378 379 380 381 382 383 384 385 386 387 388 389 390
    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])}},
             conv_fm,
             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,
391 392 393
             nr_threads,
             reinterpret_cast<const ConvolutionForward::PreprocessedFilter*>(
                     preprocessed_filter)},
394 395 396 397 398 399 400 401
            bias.dtype,
            bias.stride[0],
            bias_mode,
            param().nonlineMode};
}

ConvBiasImpl::NCBKernParam ConvBiasImpl::make_ncb_kern_param(
        _megdnn_tensor_in src, _megdnn_tensor_in filter, _megdnn_tensor_in bias,
402 403
        _megdnn_tensor_out dst, _megdnn_workspace workspace,
        const PreprocessedFilter* preprocessed_filter) {
404
    NCBKernParam ret;
405 406 407
    static_cast<NCBKernSizeParam&>(ret) =
            make_ncb_kern_size_param(src.layout, filter.layout, bias.layout,
                                     dst.layout, preprocessed_filter);
408 409 410 411 412 413 414 415 416 417 418
    ret.src_ptr = src.raw_ptr;
    ret.filter_ptr = filter.raw_ptr;
    ret.bias_ptr = bias.raw_ptr;
    ret.dst_ptr = dst.raw_ptr;
    ret.workspace_ptr = workspace.raw_ptr;
    ret.workspace_size = workspace.size;
    return ret;
}

void ConvBiasImpl::exec_with_ncb_kern(const NCBKernParam& param,
                                      ConvBiasImpl::Algorithm* algo) {
419
    auto&& ncb_kerns = NCB_ALGO_FUNC(dispatch_kerns, algo, param);
420
    for (auto&& kernel : ncb_kerns) {
421
        auto run = [kernel, param](size_t index, size_t thread_id) {
422
            CpuNDRange ndrange_id(kernel.global_size, index);
423
            kernel.kern(param, {thread_id, ndrange_id});
424 425 426 427 428 429
        };
        static_cast<naive::HandleImpl*>(handle())->dispatch_kern(
                run, kernel.global_size.total_size());
    }
}

430 431
void ConvBiasImpl::exec_preprocess_with_ncb_kern(
        const NCBKernParam& param, ConvBiasImpl::Algorithm* algo) {
432
    auto&& ncb_kerns = NCB_ALGO_FUNC(dispatch_preprocess_kerns, algo, param);
433 434 435 436 437 438 439 440
    for (auto&& kernel : ncb_kerns) {
        auto run = [kernel, param](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*>(handle())->dispatch_kern(
                run, kernel.global_size.total_size());
    }
441 442 443 444 445 446 447
}

std::vector<ConvBiasImpl::Algorithm*> ConvBiasImpl::get_all_algorithms_with_ncb(
        const NCBKernSizeParam& param) {
    MEGDNN_MARK_USED_VAR(param);
    std::vector<Algorithm*> algos;
    std::vector<Algorithm*> prefer_algos;
448
    for (auto&& algo : get_all_packed_algo()) {
449 450
        if (algo->usable(param, AlgoSelectionStrategy::FULL_RUN)) {
            if (algo->is_preferred(param)) {
451 452 453 454 455 456 457 458 459 460 461
                prefer_algos.push_back(algo);
            } else {
                algos.push_back(algo);
            }
        }
    }
    //! Prefer algo inserted from begin
    algos.insert(algos.begin(), prefer_algos.begin(), prefer_algos.end());
    return algos;
}

462 463
ConvBiasImpl::Algorithm* ConvBiasImpl::get_algorithm_from_desc(
        const AlgorithmDesc& desc) {
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
    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_X86
            case Handle::HandleType::X86:
                return x86::ConvBiasImpl::get_algo_from_desc(desc);
#elif MEGDNN_AARCH64 || MEGDNN_ARMV7
            case Handle::HandleType::ARM_COMMON:
                return arm_common::ConvBiasImpl::get_algo_from_desc(desc);
#if MEGDNN_AARCH64
            case Handle::HandleType::AARCH64:
                return aarch64::ConvBiasImpl::get_algo_from_desc(desc);
#else
            case Handle::HandleType::ARMV7:
                return armv7::ConvBiasImpl::get_algo_from_desc(desc);
#endif
#endif
            case Handle::HandleType::NAIVE: {
                auto algo = static_cast<naive::HandleImpl*>(handle())
                                    ->default_conv_bias_fwd_algo();
                megdnn_assert(algo->info().desc == desc);
                return algo;
            }
            default:
                megdnn_throw("Unknown handle type");
                return nullptr;
        }
    }
}

501 502
ConvBiasImpl::Algorithm* ConvBiasImpl::get_algorithm(
        const NCBKernSizeParam& param, size_t workspace_size) {
503 504 505
    if (ConvBiasImpl::param().format == Param::Format::NHWCD4) {
        return nullptr;
    }
506
    if (auto algo = get_algorithm_from_desc(execution_policy().algo)) {
507
        return algo;
508 509 510
    }
    if (!m_prev_selected_algo ||
        memcmp(&m_prev_selected_algo_sizep, &param, sizeof(NCBKernSizeParam))) {
511
        m_prev_selected_algo = get_algorithm_heuristic_with_ncb(
512 513
                param, workspace_size, AlgoAttribute::DEFAULT,
                AlgoAttribute::DEFAULT);
514 515 516 517 518
        m_prev_selected_algo_sizep = param;
    }
    return m_prev_selected_algo;
}

519 520 521 522 523 524 525
SmallVector<AlgoCategory> ConvBiasImpl::suggest_algo_category_order(
        const NCBKernSizeParam& param) const {
    auto IC = param.filter_meta.icpg;
    auto OC = param.filter_meta.ocpg;
    auto FH = param.filter_meta.spatial[0];
    auto FW = param.filter_meta.spatial[1];
    //! TODO: now winograd only support in fast-run
526

527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543
    //! im2col + matmul
    bool im2col_prefer = (IC >= 32 || OC >= 32);
    //! quantized algo use matmul when direct algo is unusable
    if (param.src_type.category() == DTypeCategory::QUANTIZED) {
        im2col_prefer = is_matmul_quantized_prefer(param);
    }
    //! conv1x1
    im2col_prefer |= (FH == 1 && FW == 1);
    if (im2col_prefer) {
        return {AlgoCategory::IM2COL, AlgoCategory::DIRECT,
                AlgoCategory::NAIVE};
    } else {
        return {AlgoCategory::DIRECT, AlgoCategory::IM2COL,
                AlgoCategory::NAIVE};
    }
}

544 545 546 547 548
const char* ConvBiasImpl::get_algorithm_set_name() const {
    // fallback version 0
    return "F0";
}

549
namespace megdnn {
550
namespace fallback {
551

552
template <typename T>
553 554 555 556
const T* ConvBiasImpl::NCBKernParam::src(size_t batch_id, size_t group_pack_id,
                                         size_t channel_pack_id,
                                         size_t group_pack_size,
                                         size_t channel_pack_size) const {
557
    size_t batch_offset = batch_id * inp_bs * src_type.size();
558
    size_t group_offset = group_pack_size * group_pack_id * filter_meta.icpg *
559
                          isz[0] * isz[1] * src_type.size();
560 561
    size_t channel_offset = channel_pack_size * channel_pack_id * isz[0] *
                            isz[1] * src_type.size();
562
    return reinterpret_cast<T*>(reinterpret_cast<ptrdiff_t>(src_ptr) +
563
                                batch_offset + group_offset + channel_offset);
564 565 566
}

template <typename T>
567
const T* ConvBiasImpl::NCBKernParam::filter(size_t group_pack_id,
568 569 570 571
                                            size_t pack_group_size) const {
    size_t group_offset = 0_z;
    switch (filter_meta.format) {
        case Param::Format::NCHW: {
572
            group_offset = pack_group_size * group_pack_id * filter_meta.icpg *
573 574 575 576 577 578 579 580 581
                           filter_meta.ocpg * filter_meta.spatial[0] *
                           filter_meta.spatial[1] * filter_type.size();
            break;
        }
        case Param::Format::NCHW88: {
            size_t group = filter_meta.group;
            size_t icpg = filter_meta.icpg;
            size_t ocpg = filter_meta.ocpg;
            //! four format of weight layout
582 583 584
            //! 1. {oc/8, ic/8, fh, fw, 8, 8},
            //! 2. {g, oc/8, ic/8, fh, fw, 8, 8},
            //! 3. {g/8, fh, fw, 1, 1, 8}, 4. {oc/8, fh, fw, ic, 8}
585 586 587 588 589
            megdnn_assert((icpg % 8 == 0 && ocpg % 8 == 0) ||
                                  (group % 8 == 0 && icpg == 1 && ocpg == 1 &&
                                   pack_group_size > 1) ||
                                  (group == 1 && ocpg % 8 == 0),
                          "The filter shepe is not right of nchw88");
590 591 592 593 594 595
            group_offset = pack_group_size * group_pack_id * filter_meta.icpg *
                           filter_meta.ocpg * filter_meta.spatial[0] *
                           filter_meta.spatial[1] * filter_type.size();

            break;
        }
596
        case Param::Format::NCHW44_DOT:
597 598 599 600 601 602 603
        case Param::Format::NCHW44: {
            size_t group = filter_meta.group;
            size_t icpg = filter_meta.icpg;
            size_t ocpg = filter_meta.ocpg;
            //! four format of weight layout
            //! 1. {oc/4, ic/4, fh, fw, 4, 4},
            //! 2. {g, oc/4, ic/4, fh, fw, 4, 4},
604 605
            //! 3. {g/4, fh, fw, 1, 1, 4},
            //! 4. {oc/4, fh, fw, ic, 4}
606 607 608 609 610
            megdnn_assert((icpg % 4 == 0 && ocpg % 4 == 0) ||
                                  (group % 4 == 0 && icpg == 1 && ocpg == 1 &&
                                   pack_group_size > 1) ||
                                  (group == 1 && ocpg % 4 == 0),
                          "The filter shepe is not right of nchw44");
611
            group_offset = pack_group_size * group_pack_id * filter_meta.icpg *
612 613 614 615 616 617
                           filter_meta.ocpg * filter_meta.spatial[0] *
                           filter_meta.spatial[1] * filter_type.size();

            break;
        }
        default:
618
            megdnn_assert(0, "other filter format is not support yet");
619 620 621 622 623 624
    }
    return reinterpret_cast<T*>(reinterpret_cast<ptrdiff_t>(filter_ptr) +
                                group_offset);
}

template <typename T>
625 626 627 628
const T* ConvBiasImpl::NCBKernParam::bias(size_t batch_id, size_t group_pack_id,
                                          size_t channel_pack_id,
                                          size_t group_pack_size,
                                          size_t channel_pack_size) const {
629 630
    size_t batch_offset = 0_z;
    size_t group_offset = 0_z;
631
    size_t channel_offset = 0_z;
632 633
    if (bias_mode == BiasMode::BIAS) {
        batch_offset = batch_id * bias_bs * bias_type.size();
634 635 636 637
        group_offset = group_pack_size * group_pack_id * filter_meta.ocpg *
                       osz[0] * osz[1] * bias_type.size();
        channel_offset = channel_pack_size * channel_pack_id * osz[0] * osz[1] *
                         bias_type.size();
638
    } else if (bias_mode == BiasMode::BROADCAST_CHANNEL_BIAS) {
639
        group_offset = group_pack_size * group_pack_id * filter_meta.ocpg *
640
                       bias_type.size();
641
        channel_offset = channel_pack_size * channel_pack_id * bias_type.size();
642 643
    }
    return reinterpret_cast<T*>(reinterpret_cast<ptrdiff_t>(bias_ptr) +
644
                                batch_offset + group_offset + channel_offset);
645 646 647
}

template <typename T>
648 649 650 651
T* ConvBiasImpl::NCBKernParam::dst(size_t batch_id, size_t group_pack_id,
                                   size_t channel_pack_id,
                                   size_t group_pack_size,
                                   size_t channel_pack_size) const {
652
    size_t batch_offset = batch_id * out_bs * dst_type.size();
653
    size_t group_offset = group_pack_size * group_pack_id * filter_meta.ocpg *
654
                          osz[0] * osz[1] * dst_type.size();
655 656
    size_t channel_offset = channel_pack_size * channel_pack_id * osz[0] *
                            osz[1] * dst_type.size();
657
    return reinterpret_cast<T*>(reinterpret_cast<ptrdiff_t>(dst_ptr) +
658
                                batch_offset + group_offset + channel_offset);
659 660
}

661 662 663 664 665 666 667 668 669 670 671 672
#define INST(T)                                                      \
    template const T* ConvBiasImpl::NCBKernParam::src<T>(            \
            size_t batch_id, size_t group_id, size_t channel_id,     \
            size_t group_pack_size, size_t channel_pack_size) const; \
    template const T* ConvBiasImpl::NCBKernParam::bias<T>(           \
            size_t batch_id, size_t group_id, size_t channel_id,     \
            size_t group_pack_size, size_t channel_pack_size) const; \
    template const T* ConvBiasImpl::NCBKernParam::filter<T>(         \
            size_t group_id, size_t group_pack_size) const;          \
    template T* ConvBiasImpl::NCBKernParam::dst<T>(                  \
            size_t batch_id, size_t group_id, size_t channel_id,     \
            size_t group_pack_size, size_t channel_pack_size) const;
673 674 675 676

#define INST_DT(d) INST(DTypeTrait<d>::ctype)

MEGDNN_FOREACH_COMPUTING_DTYPE(INST_DT)
677
INST(void)
678 679 680 681 682
#undef INST
#undef INST_DT
}  // namespace fallback
}  // namespace megdnn

683
// vim: syntax=cpp.doxygen