inference.cpp 89.9 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/**
 * \file src/gopt/impl/inference.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
 *
 * 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.
 */

#include "megbrain/gopt/inference.h"
#include "megbrain/gopt/gtrans.h"
#include "megbrain/gopt/basic_arith.h"
#include "megbrain/graph/event.h"
#include "megbrain/opr/dnn/batch_norm.h"
17
#include "megbrain/opr/dnn/local.h"
18 19 20 21 22 23 24 25 26 27 28
#include "megbrain/utils/shared_set.h"
#include "megbrain/serialization/opr_shallow_copy.h"
#include "megbrain/opr/basic_arith.h"
#include "megbrain/opr/dnn/convolution.h"
#include "megbrain/opr/blas.h"
#include "megbrain/opr/misc.h"
#include "megbrain/opr/utility.h"
#include "megbrain/opr/dnn/pooling.h"
#include "megbrain/opr/tensor_manip.h"
#include "megbrain/opr/imgproc.h"
#include "megbrain/opr/nn_int.h"
29
#include "megbrain/opr/tensor_gen.h"
30
#include "megbrain/utils/hash_ct.h"
31 32 33 34 35 36 37 38 39

#include "megdnn/tensor_format.h"

#if MGB_ENABLE_TENSOR_RT
#include "megbrain/tensorrt/tensorrt_opr.h"
#endif

#include "megbrain/gopt/misc.h"

40 41 42 43 44 45 46 47 48 49
#include "megbrain/utils/hash_ct.h"
#include "midout.h"

MIDOUT_DECL(megbrain_inference)
#define MIDOUT_B(tag) \
    MIDOUT_BEGIN(megbrain_inference, midout_iv(MGB_HASH_STR(tag))) {
#define MIDOUT_E \
    }            \
    MIDOUT_END();

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 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117
using namespace mgb;
using namespace gopt;

namespace {


template <typename SharedDeviceTensor, typename MultipleDeviceTensorHolder>
void param_merge(OptState& opt_state) {
    auto rewriter = opt_state.graph().make_rewriter();
    ThinHashMap<OperatorNodeBase*, size_t> opr2idx;
    std::vector<OperatorNodeBase*> all_oprs;
    typename MultipleDeviceTensorHolder::ValueArray all_values;

    auto cb_find_opr = [&](cg::OperatorNodeBase* opr) {
        if (opr->same_type<SharedDeviceTensor>()) {
            auto p = &opr->cast_final<SharedDeviceTensor>();
            // ShredD may be manu
            opr2idx[p] = all_values.size();
            all_values.push_back(p->dev_data());
            all_oprs.push_back(p);
        }
    };
    opt_state.graph().iter(cb_find_opr);
    SymbolVarArray new_vars;
    auto cb_replace = [&](cg::OperatorNodeBase* opr) {
        auto iter = opr2idx.find(opr);
        if (iter == opr2idx.end()) {
            rewriter.auto_replace_outputs(opr);
        } else {
            if (new_vars.empty()) {
                // new oprs must be created in iter callback; so we populate
                // new_vars lazily
                new_vars = MultipleDeviceTensorHolder::make(
                        *opt_state.graph().comp_graph(), std::move(all_values),
                        {ssprintf("merged%zu", all_values.size())});
                for (size_t i = 0; i < new_vars.size(); ++i) {
                    auto src = all_oprs[i]->output(0);
                    if (src->has_name_set()) {
                        new_vars[i].rename(src->name());
                    }
                }
            }
            rewriter.replace_var(
                    opr->output(0), new_vars.at(iter->second).node(),
                    mgb_cstr_log("replace multi SharedDeviceTensor(Format) to "
                                 "MultipleDeviceTensorHolder(Format)"));
        }
    };
    opt_state.graph().iter(cb_replace);

    rewriter.apply_inplace();
}

}

/* ================ global functions ================ */

SymbolVarArray gopt::optimize_for_inference(
        const SymbolVarArray& dest_vars,
        const OptimizeForInferenceOptions& opt) {
    return gopt::GraphOptimizer()
            .add_preset_passes(false, &opt,
                               &dest_vars[0].node()->owner_graph()->options())
            .apply({dest_vars})
            .endpoint_vars();
}

namespace {
118 119 120
void modify_conv_strategy(
        opr::mixin::Convolution& conv,
        opr::mixin::Convolution::ExecutionPolicy::Strategy strategy) {
121 122 123 124 125 126
    auto policy = conv.execution_policy_transient();
    policy.strategy = strategy;
    conv.set_execution_policy(policy);
}

template <typename Opr>
127 128 129 130
void inplace_conv_opr_modifier(
        OperatorNodeBase& opr,
        opr::mixin::Convolution::ExecutionPolicy::Strategy strategy) {
    modify_conv_strategy(
131
            opr.cast_final_safe<Opr>(),
132
            strategy);
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
}

void modify_conv_policy_workspace_limit(opr::mixin::Convolution& conv,
                                        size_t workspace_limit) {
    auto policy = conv.execution_policy_transient();
    policy.workspace_limit = workspace_limit;
    conv.set_execution_policy(policy);
}

template <typename Opr>
void inplace_conv_opr_workspace_limit_modifier(OperatorNodeBase& opr,
                                               size_t workspace_limit) {
    modify_conv_policy_workspace_limit(opr.cast_final_safe<Opr>(),
                                       workspace_limit);
}

}  // anonymous namespace

#define MGB_FOREACH_FASTRUN_OPR(cb)                                           \
    cb(ConvolutionForward), cb(ConvBiasForward), cb(ConvolutionBackwardData), \
            cb(ConvolutionBackwardFilter), cb(Convolution3DForward),          \
            cb(Convolution3DBackwardData), cb(Convolution3DBackwardFilter),   \
            cb(LocalShareForward), cb(LocalShareBackwardData),                \
            cb(LocalShareBackwardFilter), cb(DeformableConvForward),          \
            cb(DeformableConvBackwardFilter), cb(DeformableConvBackwardData), \
            cb(BatchConvBiasForward),

160 161 162 163 164 165 166 167 168 169 170 171 172 173
void gopt::modify_opr_algo_strategy_inplace(
        const VarNodeArrayView& dest_vars,
        opr::mixin::Convolution::ExecutionPolicy::Strategy strategy) {
#if !MGB_ENABLE_FASTRUN
    using S = opr::mixin::Convolution::ExecutionPolicy::Strategy;
    if (strategy == S::PROFILE || strategy == S::PROFILE_REPRODUCIBLE) {
        mgb_throw(MegBrainError, "fastrun is disabled at compile time");
    }
#endif
    const ThinHashMap<Typeinfo*, std::function<void(OperatorNodeBase&)>>
            modifiers = {
#define CONV(t)                                                       \
    {opr::t::typeinfo(), std::bind(inplace_conv_opr_modifier<opr::t>, \
                                   std::placeholders::_1, strategy)}
174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190
                    MGB_FOREACH_FASTRUN_OPR(CONV)
#undef CONV
            };

    auto on_opr = [&](OperatorNodeBase* opr) {
        auto iter = modifiers.find(opr->dyn_typeinfo());
        if (iter != modifiers.end()) {
            iter->second(*opr);
        }
    };

    cg::DepOprIter dep_iter{on_opr};
    for (auto i : dest_vars) {
        dep_iter.add(i);
    }
}

191
void gopt::enable_opr_algo_profiling_inplace(
192
        const VarNodeArrayView& dest_vars) {
193 194 195 196
    modify_opr_algo_strategy_inplace(dest_vars,
                                     opr::mixin::Convolution::ExecutionPolicy::
                                             Strategy::PROFILE);
}
197

198 199 200 201 202
void gopt::enable_opr_use_profiling_cache_inplace(
        const VarNodeArrayView& dest_vars) {
    modify_opr_algo_strategy_inplace(dest_vars,
                                     opr::mixin::Convolution::ExecutionPolicy::
                                             Strategy::PROFILE_HEURISTIC);
203 204
}

205

206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443
void gopt::set_opr_algo_workspace_limit_inplace(
        const VarNodeArrayView& dest_vars, size_t workspace_limit) {
    static const ThinHashMap<Typeinfo*, void (*)(OperatorNodeBase&, size_t)>
            modifiers = {
#define CONV(t) \
    {opr::t::typeinfo(), &inplace_conv_opr_workspace_limit_modifier<opr::t>}
                    MGB_FOREACH_FASTRUN_OPR(CONV)
#undef CONV
            };

    auto on_opr = [&](OperatorNodeBase* opr) {
        auto iter = modifiers.find(opr->dyn_typeinfo());
        if (iter != modifiers.end()) {
            iter->second(*opr, workspace_limit);
        }
    };

    cg::DepOprIter dep_iter{on_opr};
    for (auto i : dest_vars) {
        dep_iter.add(i);
    }
}
#undef MGB_FOREACH_FASTRUN_OPR

/* ================ ParamRedistributePass ================ */
const char* ParamRedistributePass::name() const {
    return mgb_cstr_log("param_redistribute");
}

class ParamRedistributePass::Impl final: public RecursiveSubGraphRewriteHelper {
    ConstVarPropogate m_cvprop;
    UniqReaderCheck m_uniq_reader_check;
    //! oprs already processed in try_distribute_then_reassociate() should be
    //! skipped in on_new_opr_check_should_process()
    ThinHashSet<OperatorNodeBase*> m_opr_blacklist;
    std::string m_distribute_reasso_log_msg;

    //! try applying BinaryTrans20::associtive
    GTransResult try_reassociate(OperatorNodeBase *opr);

    //! try applying BinaryTrans20::distributive_add
    GTransResult try_distribute_add(OperatorNodeBase *opr);

    //! try distribute MUL/DIV over ADD/SUB and then apply
    GTransResult try_distribute_then_reassociate(OperatorNodeBase *opr);

    GTransResult process_opr(VarNode *out_var) override;

    bool on_new_opr_check_should_process(
            OperatorNodeBase*opr, OperatorNodeBase *repl_opr) override {
        m_uniq_reader_check.update_on_opr_auto_replace(opr, repl_opr);
        auto ins = m_cvprop.add_opr(opr);
        return ins.has_const_inp && !ins.all_const_inp &&
            !m_opr_blacklist.count(opr);
    };

    void after_replace_var(VarNode *orig_var, VarNode* new_var) override {
        m_uniq_reader_check.update_on_opr_auto_replace(orig_var->owner_opr(),
                new_var->owner_opr());
    }

    /*!
     * \brief try to reorder opr inputs to a const one and a non-const one
     *
     * return true if it can be reformulated as f(nci, ci), where nci is
     * non-const and ci is const.
     */
    bool reorder_for_normconst(OperatorNodeBase *opr,
            bool &swap_inp, VarNode *&nci, VarNode *&ci);

    public:
        Impl(OptState &state);
};

GTransResult ParamRedistributePass::Impl::process_opr(VarNode *out_var) {
    auto opr = out_var->owner_opr();
    auto trans = try_reassociate(opr);

    if (!trans.valid()) {
        trans = try_distribute_add(opr);
        if (!trans.valid())
            trans = try_distribute_then_reassociate(opr);
    }

    return trans;
}

GTransResult ParamRedistributePass::Impl::try_reassociate(
        OperatorNodeBase *opr) {

    // apply BinaryAssociative0 if opr is the form f(g(a, b), c) and b and c are
    // const

    bool swap_fop_inp = false, swap_gop_inp = false;
    VarNode *a, *b, *c, *ab;
    if (!reorder_for_normconst(opr, swap_fop_inp, ab, c))
        return None;

    if (!m_uniq_reader_check(ab))
        return None;

    if (!reorder_for_normconst(ab->owner_opr(), swap_gop_inp, a, b))
        return None;

    return BinaryTrans20::associtive().apply(opr, swap_fop_inp, swap_gop_inp);
}

GTransResult ParamRedistributePass::Impl::try_distribute_add(
        OperatorNodeBase *opr) {

    if (opr->same_type<opr::Elemwise>() || opr->input().size() != 2)
        return None;

    if (!m_cvprop.is_const(opr->input(1)))
        return None;

    auto ab = as_elem_opr(opr->input(0)->owner_opr(), opr::Elemwise::Mode::ADD);
    if (ab) {
        bool swap;
        VarNode *a, *b;
        if (reorder_for_normconst(ab, swap, a, b)) {
            return BinaryTrans20::distributive_add().apply(
                    opr, false, swap);
        }
    }
    return None;
}

GTransResult ParamRedistributePass::Impl::try_distribute_then_reassociate(
        OperatorNodeBase *opr) {
    if (!opr->same_type<opr::Elemwise>())
        return None;
    using Mode = opr::Elemwise::Mode;
    auto mode = opr->cast_final<opr::Elemwise>().param().mode;
    if (!(mode == Mode::MUL || mode == Mode::TRUE_DIV))
        return None;

    VarNode *a, *b;
    bool swap;
    if (!reorder_for_normconst(opr, swap, a, b))
        return None;

    auto chain_pred = [this](OperatorNodeBase *opr) {
        if (as_elem_opr(opr, Mode::ADD)) {
            auto var = opr->output(0);
            return m_uniq_reader_check(var) || m_cvprop.is_const(var);
        }
        return false;
    };
    auto chain = extract_opr_leaves(a, chain_pred);
    if (chain.size() <= 1)
        return None;
    std::unordered_map<VarNode*, VarNode*> repl_map;
    m_distribute_reasso_log_msg.clear();

    int nr_fail = 0, nr_succ = 0;
    for (auto &&var: chain) {
        {
            auto iter = repl_map.find(var);
            if (iter != repl_map.end()) {
                var = iter->second;
                continue;
            }
        }

        auto vnew = (SymbolVar{var} * b).node();
        m_opr_blacklist.insert(vnew->owner_opr());
        if (!m_cvprop.is_const(var)) {
            auto trans = try_reassociate(vnew->owner_opr());
            if (!trans.valid()) {
                // allow at most one failed redistribution
                if (nr_fail)
                    return None;
                ++ nr_fail;
            } else {
                ++ nr_succ;
                vnew = trans->result;
                if (!m_distribute_reasso_log_msg.empty()) {
                    m_distribute_reasso_log_msg.append(mgb_cstr_log(";"));
                }
                m_distribute_reasso_log_msg.append(trans->msg);
            }
        }

        repl_map[var] = vnew;
        var = vnew;
    }
    if (nr_succ) {
        m_distribute_reasso_log_msg.insert(0,
                mgb_cstr_log("distribute_mul("));
        m_distribute_reasso_log_msg.append(mgb_cstr_log(")"));
        return GTransResultItem{
                elemwise_reduce_var_list(chain, Mode::ADD),
                m_distribute_reasso_log_msg.c_str(),
                {}};
    }
    return None;
}

bool ParamRedistributePass::Impl::reorder_for_normconst(
        OperatorNodeBase *opr, bool &swap_inp, VarNode *&nci, VarNode *&ci) {
    if (opr->input().size() != 2)
        return false;

    nci = opr->input(0);
    ci = opr->input(1);
    if (!m_cvprop.is_const(ci)) {
        if (!is_commutable_binary(opr) || !m_cvprop.is_const(nci))
            return false;
        swap_inp = true;
        std::swap(nci, ci);
    } else {
        if (m_cvprop.is_const(nci))
            return false;
        swap_inp = false;
    }

    return true;
}

ParamRedistributePass::Impl::Impl(OptState &state):
    RecursiveSubGraphRewriteHelper{state},
    m_cvprop{ConstVarType::IMMUTABLE_AND_PARAM},
    m_uniq_reader_check{state.graph()}
{
    auto cg = state.graph().comp_graph();
    auto on_new_opr = [this](const cg::event::OprInserted &ev) {
        if (!ev.is_dedup && !ev.exc) {
            // call add_opr eagerly to avoid deep recursion
            m_cvprop.add_opr(ev.opr);
        }
    };
    auto hdl = cg->event().register_receiver
        <cg::event::OprInserted>(on_new_opr);
    apply();
}

void ParamRedistributePass::apply(OptState &state) const {
444
    MIDOUT_B("ParamRedistributePass::apply")
445
    Impl{state};
446
    MIDOUT_E
447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527
}

/* ================ ParamFusePass ================ */

/*!
 * \brief get name for new param
 */
class ParamFusePass::VarNamer {
#if MGB_BUILD_SLIM_SERVING
    public:
        const std::string& name(VarNode*) {
            static std::string ret("fuse");
            return ret;
        }
#else
    using SrcSet = SharedSet<OperatorNodeBase*>;
    //! map from var to source SharedDeviceTensor/MultiSharedDeviceHolder oprs
    //! that it depends on
    ThinHashMap<OperatorNodeBase*, SrcSet> m_opr2srcs;
    std::string m_name_cache;
    std::vector<const char*> m_cur_name;

    SrcSet& get_src_set(OperatorNodeBase* opr) {
        auto opr_typeinfo = opr->dyn_typeinfo();

        auto iter = m_opr2srcs.find(opr);
        if (iter != m_opr2srcs.end()) {
            return iter->second;
        }
        auto &&ret = m_opr2srcs[opr];
        if (opr->input().empty()) {
            if (opr_typeinfo == opr::SharedDeviceTensor::typeinfo() ||
                opr_typeinfo == opr::MultipleDeviceTensorHolder::typeinfo()) {
                ret.insert(opr);
            } else {
                mgb_assert(opr_typeinfo == opr::ImmutableTensor::typeinfo());
            }
            return ret;
        }

        for (auto i: opr->input()) {
            ret.merge_from(get_src_set(i->owner_opr()));
        }
        return ret;
    }

    public:

        const std::string& name(VarNode *var) {
            m_cur_name.clear();
            for (auto i : get_src_set(var->owner_opr())) {
                m_cur_name.push_back(i->cname());
            }

            auto cmp = [](const char *x, const char *y) {
                return strcmp(x, y) < 0;
            };
            std::sort(m_cur_name.begin(), m_cur_name.end(), cmp);
            m_name_cache.clear();
            m_name_cache.append(mgb_cstr_log("fuse("));
            bool first = true;
            for (auto i: m_cur_name) {
                if (first) {
                    first = false;
                } else {
                    m_name_cache.push_back(',');
                }
                m_name_cache.append(i);
            }
            m_name_cache.append(mgb_cstr_log(
                        ssprintf("):%s@%zu", var->cname(), var->id())));
            return m_name_cache;
        }
#endif
};

const char* ParamFusePass::name() const {
    return mgb_cstr_log("param_fuse");
}

void ParamFusePass::apply(OptState &state) const {
528
    MIDOUT_B("ParamFusePass::apply")
529 530
    auto rewriter = state.graph().make_rewriter();
    auto cg = state.graph().comp_graph();
531 532 533 534 535 536 537

    ConstVarPropogate cvprop{ConstVarType::IMMUTABLE_AND_PARAM};
    state.graph().iter([&cvprop](OperatorNodeBase *opr) {
        cvprop.add_opr(opr);
    });


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
    ThinHashSet<VarNode*> processed_var;
    VarNamer var_namer;
    // reader: null if used as endvar
    auto replace_single_var = [&](VarNode *var, OperatorNodeBase *reader) {
        if (!processed_var.insert(var).second)
            return;

        auto inferred_val = std::make_shared<DeviceTensorND>(
                var->comp_node(), var->dtype());
        auto cb = [&](DeviceTensorND& val) {
            // retain format of val
            mgb_assert(val.format() == var->format());
            inferred_val->format(val.format())
                    .resize(val.shape())
                    .copy_from_fixlayout(val);
        };

        {
            auto orig_level = cg->options().log_level;
            cg->options().log_level = 0;
            MGB_TRY {
                cg->compile({{var, cb}})->execute();
            } MGB_FINALLY(cg->options().log_level = orig_level);
        }

        SymbolVar new_var;
564
        bool is_default_format = var->format().is_default();
565 566 567 568 569 570 571 572
        if (cg::is_static_var_value(var) && is_default_format) {
            // use ImmutableTensor for inferable vars
            HostTensorND hv;
            hv.copy_from(*inferred_val).sync();
            new_var = opr::ImmutableTensor::make(
                    *var->owner_graph(), hv, var_namer.name(var));
        } else {
            if (is_default_format) {
573
                new_var = opr::SharedDeviceTensor::make_const(
574 575
                        *var->owner_graph(), inferred_val, var_namer.name(var));
            } else {
576
                new_var = opr::SharedDeviceTensorWithFormat::make_const(
577 578 579 580 581 582 583 584 585 586 587 588 589 590
                        *var->owner_graph(), inferred_val, var_namer.name(var));
            }
        }
        std::string log;
        if (reader) {
            log = mgb_ssprintf_log(
                    "due to read by %s{%s}",
                    reader->cname(), reader->dyn_typeinfo()->name);
        } else {
            log = mgb_cstr_log("as endpoint");
        }
        rewriter.replace_var(var, new_var.node(), log.c_str());
    };

591 592
    auto replace_opr = [&](OperatorNodeBase* opr) {
        auto add_ret = cvprop.opr_rst(opr);
593 594 595 596 597 598 599 600 601
        if (!add_ret.all_const_inp && add_ret.has_midconst_inp) {
            for (auto i: opr->input()) {
                if (cvprop.is_midconst(i)) {
                    state.call_with_opr(i->owner_opr(),
                        [&]{replace_single_var(i, opr);});
                }
            }
        }
        rewriter.auto_replace_outputs(opr);
602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625

        //! we should deal with midconst var after auto_replace_outputs, as
        //! on_midconst_opr will replace the endpoint output which may cause
        //! double replace.
        if (add_ret.all_const_inp) {
            for (auto var : opr->output()) {
                if (var->contain_flag(VarNode::Flag::VOLATILE_CONTENT))
                    continue;

                auto osize = ConstVarPropogate::var_mem_size(var);
                if (osize >= cvprop.max_size(opr) &&
                    osize - cvprop.max_size(opr) > m_param_grow_limit) {
                    return;
                }

                // const oprs should be evaluated when output is used by another
                // non-const opr or output is needed by the user
                if (state.graph().endpoint_contain(var)) {
                    replace_single_var(var, nullptr);
                }

            }

        }
626 627
    };

628
    state.graph().iter(replace_opr);
629
    rewriter.apply_inplace();
630
    MIDOUT_E
631 632 633 634 635 636 637 638
}

/* ================ One2OneOprReplacePass ================ */
const char* ConvertF32ToF16Pass::name() const {
    return mgb_cstr_log("convert_f32_to_f16");
}

void ConvertF32ToF16Pass::apply(OptState& state) const {
639
    MIDOUT_B("ConvertF32ToF16Pass::apply")
640 641 642 643
    state.set_var_replace_check_flag(m_var_replace_check_flag);
    auto rewriter = state.graph().make_rewriter();
    VarNodeArray new_inp_cache;

644 645 646 647 648 649 650 651
    // record original output dtype
    const SymbolVarArray& vars = state.graph().endpoint_vars();
    std::vector<DType> dtypes;
    for (size_t i = 0; i < vars.size(); i++) {
        dtypes.push_back(vars[i].node()->dtype());
    }

    auto on_opr = [this, &rewriter, &new_inp_cache](OperatorNodeBase* opr) {
652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667
        auto it = m_opr_replace_func.find(opr->dyn_typeinfo());
        if (it != m_opr_replace_func.end()) {
            auto&& new_inp = new_inp_cache;
            new_inp.clear();
            new_inp.reserve(opr->input().size());
            for (auto i: opr->input()) {
                new_inp.push_back(rewriter.get_var(i));
            }
            auto new_opr = (it->second)(opr, new_inp);

            auto &&origin_out = opr->output(), &&cur_out = new_opr->output();
            mgb_assert(origin_out.size() == cur_out.size(),
                       "bad opr replace: src=%s{%s} dst=%s{%s}", opr->cname(),
                       opr->dyn_typeinfo()->name, new_opr->cname(),
                       new_opr->dyn_typeinfo()->name);
            for (size_t i = 0; i < origin_out.size(); i++) {
668
                rewriter.replace_var(origin_out[i], cur_out[i], nullptr);
669 670
            }
        } else {
671
            rewriter.auto_replace_outputs(opr);
672 673 674 675
        }
    };
    state.graph().iter(on_opr);
    rewriter.apply_inplace();
676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692

    // recover output dtype
    rewriter = state.graph().make_rewriter();
    const SymbolVarArray& endpoints = state.graph().endpoint_vars();
    auto replace_output = [&]() {
        for (size_t i = 0; i < endpoints.size(); i++) {
            VarNode* var = endpoints[i].node();
            if (var->dtype().enumv() != dtypes[i].enumv()) {
                auto new_var = opr::TypeCvt::make(var, dtypes[i]).node();
                rewriter.replace_var(var, new_var, nullptr);
            }
        }
    };
    mgb_assert(endpoints.size() > 0);
    auto opr = endpoints[0].node()->owner_opr();
    state.call_with_opr(opr, replace_output, OprPropertyFlag::NONE);
    rewriter.apply_inplace();
693
    MIDOUT_E
694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737
}

std::unique_ptr<ConvertF32ToF16Pass> ConvertF32ToF16Pass::make(
        bool use_f32_comp) {
#if MEGDNN_DISABLE_FLOAT16
    mgb_throw(SystemError, "float16 disabled at compile time.");
#else
    auto replace_h2d_opr = [](OperatorNodeBase* opr,
                              const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size());
        auto& h2d_opr = opr->cast_final_safe<opr::Host2DeviceCopy>();
        if (h2d_opr.output(0)->dtype() == dtype::Float32()) {
            auto cvt_var =
                    opr::TypeCvt::make(h2d_opr.output(0), dtype::Float16(), {});
            return cvt_var.node()->owner_opr();
        }
        return opr;
    };

    auto replace_sdt_opr = [](OperatorNodeBase* opr,
                              const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size());
        auto& sdt_opr = opr->cast_final_safe<opr::SharedDeviceTensor>();
        if (sdt_opr.output(0)->dtype() == dtype::Float32()) {
            auto cvt_var =
                    opr::TypeCvt::make(sdt_opr.output(0), dtype::Float16(), {});
            return cvt_var.node()->owner_opr();
        }
        return opr;
    };

    auto replace_imt_opr = [](OperatorNodeBase* opr,
                              const VarNodeArray& new_inp) {
        mgb_assert(opr->same_type<opr::ImmutableTensor>());
        mgb_assert(opr->input().size() == new_inp.size());
        auto& imt_opr = opr->cast_final_safe<opr::ImmutableTensor>();
        if (imt_opr.output(0)->dtype() == dtype::Float32()) {
            auto cvt_var =
                    opr::TypeCvt::make(imt_opr.output(0), dtype::Float16(), {});
            return cvt_var.node()->owner_opr();
        }
        return opr;
    };

738 739 740 741 742 743 744 745 746 747 748 749 750
    auto replace_lsp_opr = [](OperatorNodeBase* opr,
                              const VarNodeArray& new_inp) {
        mgb_assert(opr->same_type<opr::Linspace>());
        mgb_assert(opr->input().size() == new_inp.size());
        auto& lsp_opr = opr->cast_final_safe<opr::Linspace>();
        if (lsp_opr.output(0)->dtype() != dtype::Float16()) {
            auto cvt_var =
                    opr::TypeCvt::make(lsp_opr.output(0), dtype::Float16(), {});
            return cvt_var.node()->owner_opr();
        }
        return opr;
    };

751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773
    auto replace_conv_opr = [use_f32_comp](OperatorNodeBase* opr,
                                           const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size());
        auto& conv_opr = opr->cast_final_safe<opr::ConvolutionForward>();
        auto new_param = conv_opr.param();
        if (use_f32_comp) {
            new_param.compute_mode =
                    megdnn::param::Convolution::ComputeMode::FLOAT32;
        }
        mgb_assert(new_inp[0]->dtype() == dtype::Float16(),
                   "inp %s:%s, owner_opr:%s", new_inp[0]->dtype().name(),
                   new_inp[0]->name().c_str(),
                   new_inp[0]->owner_opr()->name().c_str());
        mgb_assert(new_inp[1]->dtype() == dtype::Float16(),
                   "inp %s:%s, owner_opr:%s", new_inp[1]->dtype().name(),
                   new_inp[1]->name().c_str(),
                   new_inp[1]->owner_opr()->name().c_str());
        auto new_conv_opr = opr::Convolution::make(
                new_inp[0], new_inp[1], new_param, conv_opr.execution_policy(),
                conv_opr.config());
        return new_conv_opr.node()->owner_opr();
    };

774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796
    auto replace_deconv_opr = [use_f32_comp](OperatorNodeBase* opr,
                                           const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size());
        auto& deconv_opr = opr->cast_final_safe<opr::ConvolutionBackwardData>();
        auto new_param = deconv_opr.param();
        if (use_f32_comp) {
            new_param.compute_mode =
                    megdnn::param::Convolution::ComputeMode::FLOAT32;
        }
        mgb_assert(new_inp[0]->dtype() == dtype::Float16(),
                   "inp %s:%s, owner_opr:%s", new_inp[0]->dtype().name(),
                   new_inp[0]->name().c_str(),
                   new_inp[0]->owner_opr()->name().c_str());
        mgb_assert(new_inp[1]->dtype() == dtype::Float16(),
                   "inp %s:%s, owner_opr:%s", new_inp[1]->dtype().name(),
                   new_inp[1]->name().c_str(),
                   new_inp[1]->owner_opr()->name().c_str());
        auto new_deconv_opr = opr::ConvolutionBackwardData::make(
                new_inp[0], new_inp[1], new_param, deconv_opr.execution_policy(),
                deconv_opr.config());
        return new_deconv_opr.node()->owner_opr();
    };

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
    auto replace_convbias_opr = [use_f32_comp](OperatorNodeBase* opr,
                                               const VarNodeArray& new_inp) {
        auto& convbias_opr = opr->cast_final_safe<opr::ConvBiasForward>();
        auto new_param = convbias_opr.param();
        if (use_f32_comp) {
            new_param.compute_mode =
                    megdnn::param::ConvBias::ComputeMode::FLOAT32;
        }
        mgb_assert(new_inp[0]->dtype() == dtype::Float16(),
                   "inp %s:%s, owner_opr:%s", new_inp[0]->dtype().name(),
                   new_inp[0]->name().c_str(),
                   new_inp[0]->owner_opr()->name().c_str());
        mgb_assert(new_inp[1]->dtype() == dtype::Float16(),
                   "inp %s:%s, owner_opr:%s", new_inp[1]->dtype().name(),
                   new_inp[1]->name().c_str(),
                   new_inp[1]->owner_opr()->name().c_str());
        if(opr->input().size() == 2) {
            auto new_conv_opr = opr::ConvBias::make(
                    new_inp[0], new_inp[1], new_param, convbias_opr.execution_policy(),
                    convbias_opr.config());
            return new_conv_opr.node()->owner_opr();
        } else if(opr->input().size() == 3) {
            auto new_conv_opr = opr::ConvBias::make(
                    new_inp[0], new_inp[1], new_inp[2], new_param, convbias_opr.execution_policy(),
                    convbias_opr.config());
            return new_conv_opr.node()->owner_opr();
        } else {
            mgb_assert(opr->input().size() == 4, "invalid input size %zu",
                       opr->input().size());
            auto new_conv_opr = opr::ConvBias::make(
                    new_inp[0], new_inp[1], new_inp[2], new_inp[3], new_param, convbias_opr.execution_policy(),
                    convbias_opr.config());
            return new_conv_opr.node()->owner_opr();
        }
    };

833 834 835 836 837 838 839 840 841 842 843 844 845 846
    auto replace_matmul_opr = [use_f32_comp](OperatorNodeBase* opr,
                                             const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size());
        auto& matmul_opr = opr->cast_final_safe<opr::MatrixMul>();
        auto new_param = matmul_opr.param();
        if (use_f32_comp) {
            new_param.compute_mode =
                    megdnn::param::MatrixMul::ComputeMode::FLOAT32;
        }
        auto new_matmul_opr = opr::MatrixMul::make(
                new_inp[0], new_inp[1], new_param, matmul_opr.config());
        return new_matmul_opr.node()->owner_opr();
    };

847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869
    auto replace_batched_matmul_opr = [use_f32_comp](
                                              OperatorNodeBase* opr,
                                              const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size());
        auto& matmul_opr = opr->cast_final_safe<opr::BatchedMatrixMul>();
        auto new_param = matmul_opr.param();
        if (use_f32_comp) {
            new_param.compute_mode =
                    megdnn::param::MatrixMul::ComputeMode::FLOAT32;
        }
        mgb_assert(new_inp[0]->dtype() == dtype::Float16(),
                   "inp %s:%s, owner_opr:%s", new_inp[0]->dtype().name(),
                   new_inp[0]->name().c_str(),
                   new_inp[0]->owner_opr()->name().c_str());
        mgb_assert(new_inp[1]->dtype() == dtype::Float16(),
                   "inp %s:%s, owner_opr:%s", new_inp[1]->dtype().name(),
                   new_inp[1]->name().c_str(),
                   new_inp[1]->owner_opr()->name().c_str());
        auto new_matmul_opr = opr::BatchedMatrixMul::make(
                new_inp[0], new_inp[1], new_param, matmul_opr.config());
        return new_matmul_opr.node()->owner_opr();
    };

870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933
    auto replace_reduce_opr = [use_f32_comp](OperatorNodeBase* opr,
                                             const VarNodeArray& new_inp) {
        auto& reduce_opr = opr->cast_final_safe<opr::Reduce>();
        auto new_param = reduce_opr.param();
        if (use_f32_comp) {
            new_param.data_type =
                    megdnn::param::Reduce::DataType::FLOAT_O16xC32;
        }
        if (opr->input().size() == 1) {
            auto new_matmul_opr = opr::Reduce::make(new_inp[0], new_param, {},
                                                    reduce_opr.config());
            return new_matmul_opr.node()->owner_opr();
        } else {
            mgb_assert(opr->input().size() == 2, "invalid input size %zu",
                       opr->input().size());
            auto new_matmul_opr = opr::Reduce::make(
                    new_inp[0], new_param, new_inp[1], reduce_opr.config());
            return new_matmul_opr.node()->owner_opr();
        }
    };

    auto replace_cvt_opr = [](OperatorNodeBase* opr,
                              const VarNodeArray& new_inp) {
        auto& cvt_opr = opr->cast_final_safe<opr::TypeCvt>();
        SymbolVar new_cvt;
        if (cvt_opr.output(0)->dtype() == dtype::Float32()) {
            new_cvt = opr::TypeCvt::make(new_inp[0], dtype::Float16(),
                                              cvt_opr.config());
        } else {
            new_cvt = opr::TypeCvt::make(
                    new_inp[0], cvt_opr.output()[0]->dtype(), cvt_opr.config());
        }
        return new_cvt.node()->owner_opr();
    };

    auto replace_warp_opr = [](OperatorNodeBase* opr,
                               const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size() &&
                   (new_inp.size() == 3 || new_inp.size() == 4));
        auto& warp_opr = opr->cast_final<opr::WarpPerspective>();
        // mat tensor must be float32
        auto new_mat = new_inp[1];
        if (new_inp[1]->dtype() != dtype::Float32()) {
            if (try_cast_as_op<opr::TypeCvt>(new_mat->owner_opr()) &&
                new_mat->owner_opr()->input(0)->dtype() == dtype::Float32())
                new_mat = new_mat->owner_opr()->input(0);
            else
                new_mat =
                        opr::TypeCvt::make(new_inp[1], dtype::Float32(), {}).node();
        }
        SymbolVar new_warp;
        if (new_inp.size() == 3) {
            new_warp = opr::WarpPerspective::make(new_inp[0], new_mat,
                                                  new_inp[2], warp_opr.param(),
                                                  warp_opr.config());
        } else {
            mgb_assert(new_inp.size() == 4);
            new_warp = opr::WarpPerspective::make(
                    new_inp[0], new_mat, new_inp[2], new_inp[3],
                    warp_opr.param(), warp_opr.config());
        }
        return new_warp.node()->owner_opr();
    };

934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957
    auto replace_remap_opr = [](OperatorNodeBase* opr,
                                const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size() &&
                   (new_inp.size() == 2));
        auto& remap_opr = opr->cast_final<opr::Remap>();
        // map tensor must be float32
        auto new_map = new_inp[1];
        if (new_inp[1]->dtype() != dtype::Float32()) {
            if (try_cast_as_op<opr::TypeCvt>(new_map->owner_opr()) &&
                new_map->owner_opr()->input(0)->dtype() == dtype::Float32())
                new_map = new_map->owner_opr()->input(0);
            else
                new_map =
                        opr::TypeCvt::make(new_inp[1], dtype::Float32(), {}).node();
        }
        SymbolVar new_remap;

        new_remap = opr::Remap::make(new_inp[0], new_map,
                                               remap_opr.param(),
                                               remap_opr.config());
        return new_remap.node()->owner_opr();
    };


958 959 960 961 962
    auto ret = std::make_unique<ConvertF32ToF16Pass>();
    // don't check dtype
    ret->set_var_replace_check_flag(VarReplaceCheckFlag::CHECK_ALL ^
                                    VarReplaceCheckFlag::CHECK_DTYPE);
    auto&& replace_func = ret->m_opr_replace_func;
963
    replace_func[opr::Linspace::typeinfo()] = replace_lsp_opr;
964 965 966
    replace_func[opr::Host2DeviceCopy::typeinfo()] = replace_h2d_opr;
    replace_func[opr::SharedDeviceTensor::typeinfo()] = replace_sdt_opr;
    replace_func[opr::Convolution::typeinfo()] = replace_conv_opr;
967
    replace_func[opr::ConvolutionBackwardData::typeinfo()] = replace_deconv_opr;
968
    replace_func[opr::ConvBias::typeinfo()] = replace_convbias_opr;
969 970 971 972 973
    replace_func[opr::MatrixMul::typeinfo()] = replace_matmul_opr;
    replace_func[opr::Reduce::typeinfo()] = replace_reduce_opr;
    replace_func[opr::ImmutableTensor::typeinfo()] = replace_imt_opr;
    replace_func[opr::TypeCvt::typeinfo()] = replace_cvt_opr;
    replace_func[opr::WarpPerspective::typeinfo()] = replace_warp_opr;
974
    replace_func[opr::Remap::typeinfo()] = replace_remap_opr;
975 976
    replace_func[opr::BatchedMatrixMul::typeinfo()] =
            replace_batched_matmul_opr;
977 978 979 980 981 982 983
    return ret;
#endif
}

/* ================ ConvertFormatPass ================ */

void ConvertFormatPass::apply(OptState& state) const {
984
    MIDOUT_B("ConvertFormatPass::apply")
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
    state.set_var_replace_check_flag(m_var_replace_check_flag);
    auto rewriter = state.graph().make_rewriter();
    VarNodeArray new_inp_cache;
    auto on_opr = [this, &state, &rewriter,
                   &new_inp_cache](OperatorNodeBase* opr) {
        auto it = m_opr_replace_func.find(opr->dyn_typeinfo());
        if (it != m_opr_replace_func.end()) {
            auto&& new_inp = new_inp_cache;
            new_inp.clear();
            new_inp.reserve(opr->input().size());
            for (auto i : opr->input()) {
                new_inp.push_back(rewriter.get_var(i));
            }
            auto new_opr = (it->second)(opr, new_inp);
            auto &&out0 = opr->output(), &&out1 = new_opr->output();
            mgb_assert(out0.size() == out1.size(),
                       "bad opr replace: src=%s{%s} dst=%s{%s}, src.size=%zu "
                       "dst.size=%zu",
                       opr->cname(), opr->dyn_typeinfo()->name,
                       new_opr->cname(), new_opr->dyn_typeinfo()->name,
                       out0.size(), out1.size());
            for (size_t i = 0; i < out0.size(); i++) {
                if (!out0[i]->contain_flag(VarNode::Flag::VOLATILE_CONTENT)) {
                    mgb_assert(!out1[i]->contain_flag(
                            VarNode::Flag::VOLATILE_CONTENT));
                    auto src = out0[i];
                    auto dst = out1[i];
                    auto dst_is_image = dst->format().type() ==
                                        TensorFormat::Type::IMAGE2D_PACK4;
                    if (!dst_is_image &&
                        !src->owner_opr()->same_type<opr::ImmutableTensor>()) {
                        mgb_log_warn(
                                "convert NHWCD4 replaced to non-img format: "
                                "dst_opr=%s{%s} format=%s",
                                dst->owner_opr()->cname(),
                                dst->owner_opr()->dyn_typeinfo()->name,
                                dst->format().to_string().c_str());
                    }
                    if (state.graph().endpoint_contain(src) && dst_is_image) {
                        // relayout back to NCHW for output vars
                        dst = opr::RelayoutFormat::make(
                                      dst, {opr::RelayoutFormat::Param::Mode::
                                                    NHWCD4I_NCHW})
                                      .node();
                    }
                    rewriter.replace_var(src, dst, nullptr);
                }
            }
        } else {
            rewriter.auto_replace_outputs(opr);
        }
    };
    state.graph().iter(on_opr);
    rewriter.apply_inplace();
1039
    MIDOUT_E
1040 1041 1042
}

std::unique_ptr<ConvertFormatPass> ConvertFormatPass::make_nhwcd4_converter() {
1043
    MIDOUT_B("ConvertFormatPass::make")
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
    auto filter_mode =
            [](const megdnn::param::Convolution::Sparse conv_mode,
               const VarNode* filter) -> megdnn::param::RelayoutFormat::Mode {
        bool use_dot = false;
        if (filter->dtype().enumv() == megdnn::DTypeEnum::QuantizedS8 ||
            filter->dtype().enumv() == megdnn::DTypeEnum::Quantized8Asymm)
            use_dot = true;
        if (conv_mode == megdnn::param::Convolution::Sparse::DENSE) {
            if (use_dot)
                return megdnn::param::RelayoutFormat::Mode::
                        INTER_WEIGHT_DENSEI_DOT;
            return megdnn::param::RelayoutFormat::Mode::INTER_WEIGHT_DENSEI;
        } else {
            mgb_assert(conv_mode == megdnn::param::Convolution::Sparse::GROUP);
            if (filter->shape()[1] == 1 && filter->shape()[2] == 1) {
                return megdnn::param::RelayoutFormat::Mode::INTER_WEIGHT_CHANI;
            } else {
                if (use_dot)
                    return megdnn::param::RelayoutFormat::Mode::
                            INTER_WEIGHT_GROUPI_DOT;
                return megdnn::param::RelayoutFormat::Mode::INTER_WEIGHT_GROUPI;
            }
        }
    };

1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089
    auto size_one_conv_to_dense_conv =
            [](VarNode* origin_filter_input,
               megdnn::param::Convolution::Sparse sparse) {
                VarNode* reshaped_filter = origin_filter_input;
                bool is_size_one_group_conv = false;
                if (sparse == megdnn::param::Convolution::Sparse::GROUP &&
                    origin_filter_input->shape()[0] == 1) {
                    is_size_one_group_conv = true;
                    auto new_shape = origin_filter_input->shape();
                    new_shape.ndim = 4;
                    for (int i = 0; i < 4; i++) {
                        new_shape[i] = origin_filter_input->shape()[i + 1];
                    }
                    SymbolVar new_var(origin_filter_input);
                    reshaped_filter = new_var.reshape(new_shape).node();
                }
                return std::make_tuple(reshaped_filter, is_size_one_group_conv);
            };

    auto replace_conv_opr =
        [&filter_mode, &size_one_conv_to_dense_conv](OperatorNodeBase* opr,
1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 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 1147 1148 1149 1150 1151 1152 1153
                               const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size());
        auto& conv_opr = opr->cast_final_safe<opr::ConvolutionForward>();
        mgb_assert(conv_opr.param().format ==
                           megdnn::param::Convolution::Format::NCHW,
                   "ConvertFormat Pass only support converting NCHW to NHWCD4");
        VarNode *conv_src = nullptr, *conv_weights = nullptr;
        if (new_inp[0]->shape().ndim == 4) {
            // new input src is NCHW
            size_t group, icpg, ocpg;
            if (conv_opr.param().sparse ==
                megdnn::param::Convolution::Sparse::DENSE) {
                group = 1;
                icpg = new_inp[1]->shape()[1];
                ocpg = new_inp[1]->shape()[0];
            } else {
                mgb_assert(conv_opr.param().sparse ==
                           megdnn::param::Convolution::Sparse::GROUP);
                group = new_inp[1]->shape()[0];
                icpg = new_inp[1]->shape()[2];
                ocpg = new_inp[1]->shape()[1];
            }
            if (ocpg % 4 == 0 && (icpg % 4 == 0 || group == 1)) {
                auto param = megdnn::param::RelayoutFormat();
                param.mode = megdnn::param::RelayoutFormat::Mode::NCHW_NHWCD4I;
                auto rf = opr::RelayoutFormat::make(new_inp[0], param);
                conv_src = rf.node();
            } else {
                // can not convert to hwcd4
                return serialization::copy_opr_shallow(*opr, new_inp,
                                                       opr->config());
            }
        } else {
            size_t ocpg;
            bool is_channel_wise = false;
            if (conv_opr.param().sparse ==
                megdnn::param::Convolution::Sparse::DENSE) {
                ocpg = new_inp[1]->shape()[0];
            } else {
                mgb_assert(conv_opr.param().sparse ==
                           megdnn::param::Convolution::Sparse::GROUP);
                size_t icpg = new_inp[1]->shape()[2];
                ocpg = new_inp[1]->shape()[1];
                if (icpg == 1 && ocpg == 1) {
                   is_channel_wise = true;
                }
            }
            if (ocpg % 4 != 0 && !is_channel_wise) {
                VarNodeArray t_inp = new_inp;
                auto param = megdnn::param::RelayoutFormat();
                param.mode = megdnn::param::RelayoutFormat::Mode::NHWCD4I_NCHW;
                auto rf = opr::RelayoutFormat::make(new_inp[0], param);
                t_inp[0] = rf.node();
                auto new_opr = serialization::copy_opr_shallow(*opr, t_inp,
                                                               opr->config());
                return new_opr;
            }
            // new input src is NHWCD4
            auto&& fmt = new_inp[0]
                                 ->format()
                                 .as_impl<megdnn::Image2DPack4TensorFormat>();
            mgb_assert(new_inp[0]->shape().ndim == 5 && fmt.align_axis() == 2);
            conv_src = new_inp[0];
        }
1154 1155 1156 1157 1158 1159 1160 1161 1162
        VarNode* reshaped_filter;
        bool is_size_one_group_conv;
        std::tie(reshaped_filter, is_size_one_group_conv) =
                size_one_conv_to_dense_conv(new_inp[1],
                                            conv_opr.param().sparse);
        auto new_conv_param = conv_opr.param();
        if (is_size_one_group_conv) {
            new_conv_param.sparse = megdnn::param::Convolution::Sparse::DENSE;
        }
1163 1164 1165
        mgb_assert(new_inp[1]->format().type() !=
                   TensorFormat::Type::IMAGE2D_PACK4);
        auto param = megdnn::param::RelayoutFormat();
1166 1167
        param.mode = filter_mode(new_conv_param.sparse, reshaped_filter);
        auto relayout_weight = opr::RelayoutFormat::make(reshaped_filter, param);
1168
        conv_weights = relayout_weight.node();
1169
        new_conv_param.format = megdnn::param::Convolution::Format::NHWCD4;
1170 1171 1172 1173
        mgb_assert(conv_src->shape().ndim == 5 &&
                   conv_src->format().type() ==
                           TensorFormat::Type::IMAGE2D_PACK4);
        auto new_conv_opr = opr::Convolution::make(
1174
                conv_src, conv_weights, new_conv_param, conv_opr.execution_policy(),
1175 1176 1177 1178 1179 1180 1181 1182
                conv_opr.config());
        OperatorNodeBase* ret = new_conv_opr.node()->owner_opr();
        mgb_assert(new_conv_opr.shape().ndim == 5 &&
                   new_conv_opr.format().type() ==
                           TensorFormat::Type::IMAGE2D_PACK4);
        return ret;
    };

1183 1184
    auto replace_conv_bias_opr =
        [&filter_mode, &size_one_conv_to_dense_conv](OperatorNodeBase* opr,
1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 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 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252
                               const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size());
        auto& conv_bias_opr = opr->cast_final_safe<opr::ConvBiasForward>();
        mgb_assert(conv_bias_opr.param().format ==
                           megdnn::param::ConvBias::Format::NCHW,
                   "ConvertFormat Pass only support converting NCHW to NHWCD4");
        VarNode *conv_bias_src = nullptr, *conv_bias_weights = nullptr,
                *conv_bias_bias = nullptr;
        if (new_inp[0]->shape().ndim == 4) {
            // new input src is NCHW
            size_t group, icpg, ocpg;
            if (conv_bias_opr.param().sparse ==
                megdnn::param::ConvBias::Sparse::DENSE) {
                group = 1;
                icpg = new_inp[1]->shape()[1];
                ocpg = new_inp[1]->shape()[0];
            } else {
                mgb_assert(conv_bias_opr.param().sparse ==
                           megdnn::param::ConvBias::Sparse::GROUP);
                group = new_inp[1]->shape()[0];
                icpg = new_inp[1]->shape()[2];
                ocpg = new_inp[1]->shape()[1];
            }
            if (ocpg % 4 == 0 && (icpg % 4 == 0 || group == 1)) {
                auto param = megdnn::param::RelayoutFormat();
                param.mode = megdnn::param::RelayoutFormat::Mode::NCHW_NHWCD4I;
                auto rf = opr::RelayoutFormat::make(new_inp[0], param);
                conv_bias_src = rf.node();
            } else {
                // can not convert to hwcd4
                return serialization::copy_opr_shallow(*opr, new_inp,
                                                       opr->config());
            }
        } else {
            size_t ocpg;
            bool is_channel_wise = false;
            if (conv_bias_opr.param().sparse ==
                megdnn::param::ConvBias::Sparse::DENSE) {
                ocpg = new_inp[1]->shape()[0];
            } else {
                mgb_assert(conv_bias_opr.param().sparse ==
                           megdnn::param::ConvBias::Sparse::GROUP);
                size_t icpg = new_inp[1]->shape()[2];
                ocpg = new_inp[1]->shape()[1];
                if (icpg == 1 && ocpg == 1) {
                   is_channel_wise = true;
                }
            }
            if (ocpg % 4 != 0 && !is_channel_wise) {
                VarNodeArray t_inp = new_inp;
                auto param = megdnn::param::RelayoutFormat();
                param.mode = megdnn::param::RelayoutFormat::Mode::NHWCD4I_NCHW;
                auto rf = opr::RelayoutFormat::make(new_inp[0], param);
                t_inp[0] = rf.node();
                auto new_opr = serialization::copy_opr_shallow(*opr, t_inp,
                                                               opr->config());
                return new_opr;
            }
            // new input src is NHWCD4
            auto&& fmt = new_inp[0]
                                 ->format()
                                 .as_impl<megdnn::Image2DPack4TensorFormat>();
            mgb_assert(new_inp[0]->shape().ndim == 5 && fmt.align_axis() == 2);
            conv_bias_src = new_inp[0];
        }
        mgb_assert(new_inp[1]->format().type() !=
                   TensorFormat::Type::IMAGE2D_PACK4);

1253 1254 1255 1256 1257 1258 1259 1260 1261
        VarNode* reshaped_filter;
        bool is_size_one_group_conv;
        std::tie(reshaped_filter, is_size_one_group_conv) =
                size_one_conv_to_dense_conv(new_inp[1],
                                            conv_bias_opr.param().sparse);
        auto new_conv_param = conv_bias_opr.param();
        if (is_size_one_group_conv) {
            new_conv_param.sparse = megdnn::param::Convolution::Sparse::DENSE;
        }
1262
        auto param = megdnn::param::RelayoutFormat();
1263 1264
        param.mode = filter_mode(new_conv_param.sparse, reshaped_filter);
        auto relayout_weight = opr::RelayoutFormat::make(reshaped_filter, param);
1265 1266
        conv_bias_weights = relayout_weight.node();

1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277
        mgb_assert(new_inp.size() < 4,
                   "ConvertFormat pass does not support fuse Z");
        bool has_bias = new_inp.size() > 2;
        if (has_bias &&
            new_inp[2]->format().type() == TensorFormat::Type::DEFAULT) {
            param.mode = megdnn::param::RelayoutFormat::Mode::NCHW_NHWCD4I;
            auto relayout_bias = opr::RelayoutFormat::make(new_inp[2], param);
            conv_bias_bias = relayout_bias.node();
        } else if (has_bias) {
            conv_bias_bias = new_inp[2];
        }
1278

1279
        new_conv_param.format = megdnn::param::ConvBias::Format::NHWCD4;
1280 1281 1282
        mgb_assert(conv_bias_src->shape().ndim == 5 &&
                   conv_bias_src->format().type() ==
                           TensorFormat::Type::IMAGE2D_PACK4);
1283 1284 1285
        SymbolVar new_conv_bias_opr;
        if (has_bias) {
            new_conv_bias_opr = opr::ConvBias::make(
1286
                    conv_bias_src, conv_bias_weights, conv_bias_bias, new_conv_param,
1287 1288 1289
                    conv_bias_opr.execution_policy(), conv_bias_opr.config());
        } else {
            new_conv_bias_opr = opr::ConvBias::make(
1290
                    conv_bias_src, conv_bias_weights, new_conv_param,
1291 1292
                    conv_bias_opr.execution_policy(), conv_bias_opr.config());
        }
1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 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 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382
        OperatorNodeBase* ret = new_conv_bias_opr.node()->owner_opr();
        mgb_assert(new_conv_bias_opr.shape().ndim == 5 &&
                   new_conv_bias_opr.format().type() ==
                           TensorFormat::Type::IMAGE2D_PACK4);
        return ret;
    };


    auto replace_deconv_opr = [&filter_mode](OperatorNodeBase* opr,
                               const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size());
        auto& deconv_opr = opr->cast_final_safe<opr::ConvolutionBackwardData>();
        mgb_assert(deconv_opr.param().format ==
                           megdnn::param::Convolution::Format::NCHW,
                   "ConvertFormat Pass only support converting NCHW to NHWCD4");
        VarNode *deconv_src = nullptr, *deconv_weights = nullptr;
        if (new_inp[1]->shape().ndim == 4) {
            // new input src is NCHW
            size_t group, icpg, ocpg;
            if (deconv_opr.param().sparse ==
                megdnn::param::Convolution::Sparse::DENSE) {
                group = 1;
                icpg = new_inp[0]->shape()[0];
                ocpg = new_inp[0]->shape()[1];
            } else {
                mgb_assert(deconv_opr.param().sparse ==
                           megdnn::param::Convolution::Sparse::GROUP);
                group = new_inp[0]->shape()[0];
                icpg = new_inp[0]->shape()[1];
                ocpg = new_inp[0]->shape()[2];
            }
            if (ocpg % 4 == 0 && (icpg % 4 == 0 || group == 1)) {
                auto param = megdnn::param::RelayoutFormat();
                param.mode = megdnn::param::RelayoutFormat::Mode::NCHW_NHWCD4I;
                auto rf = opr::RelayoutFormat::make(new_inp[1], param);
                deconv_src = rf.node();
            } else {
                // can not convert to hwcd4
                return serialization::copy_opr_shallow(*opr, new_inp,
                                                       opr->config());
            }
        } else {
            //! XXXX, fix me, check filter size
            size_t ocpg;
            if (deconv_opr.param().sparse ==
                megdnn::param::Convolution::Sparse::DENSE) {
                ocpg = new_inp[0]->shape()[1];
            } else {
                mgb_assert(deconv_opr.param().sparse ==
                           megdnn::param::Convolution::Sparse::GROUP);

                ocpg = new_inp[0]->shape()[2];
            }
            if (ocpg % 4 != 0) {
                VarNodeArray t_inp = new_inp;
                auto param = megdnn::param::RelayoutFormat();
                param.mode = megdnn::param::RelayoutFormat::Mode::NHWCD4I_NCHW;
                auto rf = opr::RelayoutFormat::make(new_inp[1], param);
                t_inp[1] = rf.node();
                auto new_opr = serialization::copy_opr_shallow(*opr, t_inp,
                                                               opr->config());
                return new_opr;
            }
            // new input src is NHWCD4
            auto&& fmt = new_inp[1]
                                 ->format()
                                 .as_impl<megdnn::Image2DPack4TensorFormat>();
            mgb_assert(new_inp[1]->shape().ndim == 5 && fmt.align_axis() == 2);
            deconv_src = new_inp[1];
        }
        mgb_assert(new_inp[0]->format().type() !=
                   TensorFormat::Type::IMAGE2D_PACK4);
        auto param = megdnn::param::RelayoutFormat();
        param.mode = filter_mode(deconv_opr.param().sparse, new_inp[0]);
        auto relayout_weight = opr::RelayoutFormat::make(new_inp[0], param);
        deconv_weights = relayout_weight.node();
        auto new_param = deconv_opr.param();
        new_param.format = megdnn::param::Convolution::Format::NHWCD4;
        mgb_assert(deconv_src->shape().ndim == 5 &&
                   deconv_src->format().type() ==
                           TensorFormat::Type::IMAGE2D_PACK4);
        auto new_deconv_opr = opr::ConvolutionBackwardData::make(
                deconv_weights, deconv_src, new_param,
                deconv_opr.execution_policy(), deconv_opr.config());
        OperatorNodeBase* ret = new_deconv_opr.node()->owner_opr();
        mgb_assert(new_deconv_opr.shape().ndim == 5 &&
                   new_deconv_opr.format().type() ==
                           TensorFormat::Type::IMAGE2D_PACK4);
        return ret;
    };
1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398
    /* This helper function guarantees the format convert pass won't change
     * output var's channel. Changing output's channel will cause channel
     * mismatch problem for replacing conv/conv_bias operator.
     */
    auto replace_helper = [](OperatorNodeBase* opr,
                             const VarNodeArray& new_inp) -> OperatorNodeBase* {
        auto&& new_shp = new_inp[0]->shape();
        size_t inp_channel = new_shp[1];
        if (new_shp.eq_shape(opr->input(0)->shape())&& inp_channel % 4 != 0) {
            auto new_opr = serialization::copy_opr_shallow(*opr, new_inp,
                                                           opr->config());
            return new_opr;
        }
        return nullptr;
    };
    auto replace_resize_opr = [replace_helper](OperatorNodeBase* opr,
1399 1400
                                 const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size());
1401 1402 1403
        if (auto opr_shallow_copy = replace_helper(opr, new_inp)) {
            return opr_shallow_copy;
        }
1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428
        auto& resize_opr = opr->cast_final_safe<opr::ResizeForward>();
        mgb_assert(resize_opr.param().format ==
                           megdnn::param::Resize::Format::NCHW,
                   "ConvertFormat Pass only support converting NCHW to NHWCD4");
        VarNode* inp = nullptr;
        if (new_inp[0]->shape().ndim == 4) {
            auto param = megdnn::param::RelayoutFormat();
            param.mode = megdnn::param::RelayoutFormat::Mode::NCHW_NHWCD4I;
            auto rf = opr::RelayoutFormat::make(new_inp[0], param);
            inp = rf.node();
        } else {
            // new input src is NHWCD
            auto&& fmt = new_inp[0]
                                 ->format()
                                 .as_impl<megdnn::Image2DPack4TensorFormat>();
            mgb_assert(new_inp[0]->shape().ndim == 5 && fmt.align_axis() == 2);
            inp = new_inp[0];
        }
        auto new_param = resize_opr.param();
        new_param.format = megdnn::param::Resize::Format::NHWCD4;
        auto new_resize_opr = opr::ResizeForward::make(
                inp, new_inp[1], new_param, opr->config());
        return new_resize_opr.node()->owner_opr();
    };

1429 1430 1431
    auto replace_warp_perspective_opr = [replace_helper](
                                                OperatorNodeBase* opr,
                                                const VarNodeArray& new_inp) {
1432
        mgb_assert(opr->input().size() == new_inp.size());
1433 1434 1435
        if (auto opr_shallow_copy = replace_helper(opr, new_inp)) {
            return opr_shallow_copy;
        }
1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470
        auto& warp_opr = opr->cast_final_safe<opr::WarpPerspectiveForward>();
        mgb_assert(warp_opr.param().format ==
                           megdnn::param::WarpPerspective::Format::NCHW,
                   "ConvertFormat Pass only support converting NCHW to NHWCD4");
        VarNode* inp = nullptr;
        if (new_inp[0]->shape().ndim == 4) {
            // new input src is NCHW
            auto param = megdnn::param::RelayoutFormat();
            param.mode = megdnn::param::RelayoutFormat::Mode::NCHW_NHWCD4I;
            auto rf = opr::RelayoutFormat::make(new_inp[0], param);
            inp = rf.node();
        } else {
            // new input src is NHWCD
            auto&& fmt = new_inp[0]
                                 ->format()
                                 .as_impl<megdnn::Image2DPack4TensorFormat>();
            mgb_assert(new_inp[0]->shape().ndim == 5 && fmt.align_axis() == 2);
            inp = new_inp[0];
        }
        auto new_param = warp_opr.param();
        new_param.format = megdnn::param::WarpPerspective::Format::NHWCD4;
        SymbolVar new_warp_opr;
        if (new_inp.size() == 3) {
            new_warp_opr = opr::WarpPerspectiveForward::make(
                    inp, new_inp[1], nullptr, new_inp[2], new_param,
                    opr->config());
        } else {
            mgb_assert(new_inp.size() == 4);
            new_warp_opr = opr::WarpPerspectiveForward::make(
                    inp, new_inp[1], new_inp[2], new_inp[3], new_param,
                    opr->config());
        }
        return new_warp_opr.node()->owner_opr();
    };

1471
    auto replace_warp_affine_opr = [replace_helper](OperatorNodeBase* opr,
1472 1473
                                      const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size());
1474 1475 1476
        if (auto opr_shallow_copy = replace_helper(opr, new_inp)) {
            return opr_shallow_copy;
        }
1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503
        auto& warp_opr = opr->cast_final_safe<opr::WarpAffineForward>();
        mgb_assert(warp_opr.param().format ==
                           megdnn::param::WarpAffine::Format::NCHW,
                   "ConvertFormat Pass only support converting NCHW to NHWCD4");
        VarNode* inp = nullptr;
        if (new_inp[0]->shape().ndim == 4) {
            // new input src is NCHW
            auto param = megdnn::param::RelayoutFormat();
            param.mode = megdnn::param::RelayoutFormat::Mode::NCHW_NHWCD4I;
            auto rf = opr::RelayoutFormat::make(new_inp[0], param);
            inp = rf.node();
        } else {
            // new input src is NHWCD
            auto&& fmt = new_inp[0]
                                 ->format()
                                 .as_impl<megdnn::Image2DPack4TensorFormat>();
            mgb_assert(new_inp[0]->shape().ndim == 5 && fmt.align_axis() == 2);
            inp = new_inp[0];
        }
        auto new_param = warp_opr.param();
        new_param.format = megdnn::param::WarpAffine::Format::NHWCD4;
        SymbolVar new_warp_opr;
        new_warp_opr = opr::WarpAffineForward::make(inp, new_inp[1], new_inp[2],
                                                    new_param, opr->config());
        return new_warp_opr.node()->owner_opr();
    };

1504
    auto replace_pooling_opr = [replace_helper](OperatorNodeBase* opr,
1505 1506
                                  const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size());
1507 1508 1509
        if (auto opr_shallow_copy = replace_helper(opr, new_inp)) {
            return opr_shallow_copy;
        }
1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535
        auto& pooling_opr = opr->cast_final_safe<opr::PoolingForward>();
        mgb_assert(pooling_opr.param().format ==
                           megdnn::param::Pooling::Format::NCHW,
                   "ConvertFormat Pass only support converting NCHW to NHWCD4");
        VarNode* inp = nullptr;
        if (new_inp[0]->shape().ndim == 4) {
            // new input src is NCHW
            auto param = megdnn::param::RelayoutFormat();
            param.mode = megdnn::param::RelayoutFormat::Mode::NCHW_NHWCD4I;
            auto rf = opr::RelayoutFormat::make(new_inp[0], param);
            inp = rf.node();
        } else {
            // new input src is NHWCD
            auto&& fmt = new_inp[0]
                                 ->format()
                                 .as_impl<megdnn::Image2DPack4TensorFormat>();
            mgb_assert(new_inp[0]->shape().ndim == 5 && fmt.align_axis() == 2);
            inp = new_inp[0];
        }
        auto new_param = pooling_opr.param();
        new_param.format = megdnn::param::Pooling::Format::NHWCD4;
        auto new_pooling_opr =
                opr::PoolingForward::make(inp, new_param, opr->config());
        return new_pooling_opr.node()->owner_opr();
    };

1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552
    auto var_to_chw = [](VarNode* inp, VarNode* new_inp) {
        if (!inp->shape().eq_shape(new_inp->shape())) {
            mgb_assert(inp->shape().ndim == 4 &&
                       inp->format().type() !=
                               TensorFormat::Type::IMAGE2D_PACK4);
            mgb_assert(new_inp->shape().ndim == 5 &&
                       new_inp->format().type() ==
                               TensorFormat::Type::IMAGE2D_PACK4);
            auto param = megdnn::param::RelayoutFormat();
            param.mode = megdnn::param::RelayoutFormat::Mode::NHWCD4I_NCHW;
            auto rf = opr::RelayoutFormat::make(new_inp, param);
            return rf.node();
        }
        return new_inp;
    };

    auto relayout_inp_to_chw = [var_to_chw](OperatorNodeBase* opr,
1553 1554 1555 1556
                                  const VarNodeArray& new_inp) {
        mgb_assert(opr->input().size() == new_inp.size());
        VarNodeArray t_inp = new_inp;
        for (size_t i = 0; i < opr->input().size(); i++) {
1557
            t_inp[i] = var_to_chw(opr->input(i), new_inp[i]);
1558 1559 1560 1561 1562 1563
        }
        auto new_opr =
                serialization::copy_opr_shallow(*opr, t_inp, opr->config());
        return new_opr;
    };

1564 1565 1566
    auto replace_elemwise_opr = [&relayout_inp_to_chw](
                                        OperatorNodeBase* opr,
                                        const VarNodeArray& new_inp) {
1567 1568
        mgb_assert(opr->input().size() == new_inp.size());
        bool has_inp_changed = false;
1569
        bool can_exec_cd4 = true;
1570 1571 1572
        for (size_t i = 0; i < opr->input().size(); i++) {
            if (!new_inp[i]->format().is_default()) {
                has_inp_changed = true;
1573 1574 1575 1576 1577 1578 1579
            } else if (new_inp[i]->shape().ndim == 4) {
                if (new_inp[i]->shape()[1] % 4 != 0) {
                    can_exec_cd4 = false;
                }
                //! cd4 elemwise with scaler is supported
            } else if (!new_inp[i]->shape().is_scalar()) {
                can_exec_cd4 = false;
1580 1581
            }
        }
1582 1583 1584
        if (!can_exec_cd4) {
            return relayout_inp_to_chw(opr, new_inp);
        }
1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608
        if (has_inp_changed) {
            // assumption: all inputs are changed from nchw to nhwcd4
            auto t_inp = new_inp;
            for (size_t i = 0; i < opr->input().size(); i++) {
                if (new_inp[i]->shape().ndim == 4) {
                    auto param = megdnn::param::RelayoutFormat();
                    param.mode =
                            megdnn::param::RelayoutFormat::Mode::NCHW_NHWCD4I;
                    auto rf = opr::RelayoutFormat::make(new_inp[i], param);
                    t_inp[i] = rf.node();
                } else {
                    mgb_assert((new_inp[i]->shape().ndim == 5 &&
                                new_inp[i]->format().type() ==
                                        TensorFormat::Type::IMAGE2D_PACK4) ||
                               new_inp[i]->shape().is_scalar());
                }
            }
            return serialization::copy_opr_shallow(*opr, t_inp, opr->config());
        } else {
            return serialization::copy_opr_shallow(*opr, new_inp,
                                                   opr->config());
        }
    };

1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620
    /* This helper function converts the first input to the NCHW format to
     * handle operations that do not support NHWCD4 format
     */
    auto relayout_first_inp_to_chw =
            [var_to_chw](OperatorNodeBase* opr,
               const VarNodeArray& new_inp) -> OperatorNodeBase* {
        mgb_assert(opr->input().size() == new_inp.size());
        VarNodeArray t_inp = new_inp;
        t_inp[0] = var_to_chw(opr->input(0), new_inp[0]);
        return serialization::copy_opr_shallow(*opr, t_inp, opr->config());
    };

1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641
    auto ret = std::make_unique<ConvertFormatPass>();
    ret->set_var_replace_check_flag(VarReplaceCheckFlag::NOCHECK);
    auto&& replace_func = ret->m_opr_replace_func;
    replace_func[opr::Convolution::typeinfo()] = replace_conv_opr;
    replace_func[opr::ConvBias::typeinfo()] = replace_conv_bias_opr;
    replace_func[opr::ConvolutionBackwardData::typeinfo()] = replace_deconv_opr;
    replace_func[opr::PoolingForward::typeinfo()] = replace_pooling_opr;
    replace_func[opr::Elemwise::typeinfo()] = replace_elemwise_opr;
    replace_func[opr::Concat::typeinfo()] = relayout_inp_to_chw;
    replace_func[opr::Reshape::typeinfo()] = relayout_inp_to_chw;
    replace_func[opr::GetVarShape::typeinfo()] = relayout_inp_to_chw;
    replace_func[opr::Dimshuffle::typeinfo()] = relayout_inp_to_chw;
    replace_func[opr::Reduce::typeinfo()] = relayout_inp_to_chw;
    replace_func[opr::AssertEqual::typeinfo()] = relayout_inp_to_chw;
    replace_func[opr::Subtensor::typeinfo()] = relayout_inp_to_chw;
    replace_func[opr::Broadcast::typeinfo()] = relayout_inp_to_chw;
    replace_func[opr::IncrSubtensor::typeinfo()] = relayout_inp_to_chw;
    replace_func[opr::ResizeForward::typeinfo()] = replace_resize_opr;
    replace_func[opr::WarpPerspectiveForward::typeinfo()] =
            replace_warp_perspective_opr;
    replace_func[opr::WarpAffineForward::typeinfo()] = replace_warp_affine_opr;
1642 1643 1644
    replace_func[opr::LocalForward::typeinfo()] = relayout_first_inp_to_chw;
    replace_func[opr::GroupLocalForward::typeinfo()] =
            relayout_first_inp_to_chw;
1645
    return ret;
1646
    MIDOUT_E
1647 1648 1649 1650 1651 1652 1653 1654
}

/* ================ ConvertBatchNormPass ================ */
const char* ConvertBatchNormToElemwisePass::name() const {
    return "convert_batch_norm";
}

void ConvertBatchNormToElemwisePass::apply(OptState& state) const {
1655
    MIDOUT_B("ConvertBatchNormToElemwisePass::apply")
1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666
    auto rewriter = state.graph().make_rewriter();
    auto on_opr = [&](OperatorNodeBase* opr) {
        if (auto bn = try_cast_as_op<opr::BatchNorm>(opr)) {
            if (bn->input().size() == 5) {
                mgb_assert(bn->param().fwd_mode ==
                           opr::BatchNorm::Param::FwdMode::INFERENCE);
                SymbolVar x = {rewriter.get_var(bn->input(0))};
                SymbolVar scale = {rewriter.get_var(bn->input(1))};
                SymbolVar bias = {rewriter.get_var(bn->input(2))};
                SymbolVar mean = {rewriter.get_var(bn->input(3))};
                SymbolVar variance = {rewriter.get_var(bn->input(4))};
1667 1668
                SymbolVar invsqrt_variance = opr::PowC::make(variance
                        + variance.make_scalar_dt(float(bn->param().epsilon)), {-0.5});
1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683
                auto res = scale * (x - mean) * invsqrt_variance + bias;
                rewriter.replace_var(
                        opr->output(4), res.node(),
                        mgb_cstr_log(
                                "replace batch_norm(x, scale, bias, mean, "
                                "varience) "
                                "-> (sclae * (x - mean) / sqrt(variance)) + b)"));
                return;
            }
        }
        rewriter.auto_replace_outputs(opr);
    };
    state.graph().iter(on_opr);

    rewriter.apply_inplace();
1684
    MIDOUT_E
1685 1686 1687 1688 1689 1690 1691 1692
}

/* ================ FuseConvBiasNonlinPass ================ */
const char* FuseConvBiasNonlinPass::name() const {
    return "combine_conv_bias_and_relu";
}

void FuseConvBiasNonlinPass::apply(OptState& state) const {
1693
    MIDOUT_B("FuseConvBiasNonlinPass::apply")
1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758
    std::unordered_map<VarNode*, std::vector<OperatorNodeBase*>> m_deps;
    state.graph().iter([&m_deps](OperatorNodeBase* opr) {
        for (auto& inp : opr->input()) {
            m_deps[inp].push_back(opr);
        }
    });

    auto rewriter = state.graph().make_rewriter();
    using Mode = opr::Elemwise::Param::Mode;
    using NonlineMode = opr::ConvBiasForward::Param::NonlineMode;

    auto get_nonlinearity_mode = [&](opr::Elemwise* elem) -> NonlineMode {
        if (elem->param().mode == Mode::FUSE_ADD_RELU ||
            elem->param().mode == Mode::RELU) {
            return NonlineMode::RELU;
        } else if (elem->param().mode == Mode::FUSE_ADD_SIGMOID ||
                   elem->param().mode == Mode::SIGMOID) {
            return NonlineMode::SIGMOID;
        } else {
            return NonlineMode::IDENTITY;
        }
    };

    auto try_fuse_bias_nonlinearity = [&](opr::Elemwise* elem) -> bool {

        bool can_be_fused = true;
        can_be_fused &= (elem->input().size() == 2);
        can_be_fused &= (elem->param().mode == Mode::FUSE_ADD_RELU) ||
                        (elem->param().mode == Mode::FUSE_ADD_TANH) ||
                        (elem->param().mode == Mode::FUSE_ADD_SIGMOID);

        return can_be_fused;
    };

    auto try_fuse_bias = [&](opr::Elemwise* elem) -> bool {

        bool can_be_fused = true;
        can_be_fused &= (elem->input().size() == 2);
        can_be_fused &= (elem->param().mode == Mode::ADD);
        return can_be_fused;
    };

    auto try_fuse_nonlinearity = [&](opr::Elemwise* elem) -> bool {

        bool can_be_fused = true;
        can_be_fused &= (elem->input().size() == 1);
        can_be_fused &= (elem->param().mode == Mode::RELU) ||
                        (elem->param().mode == Mode::TANH) ||
                        (elem->param().mode == Mode::SIGMOID);

        return can_be_fused;
    };

    auto convert_to_conv_bias_param = [&](const opr::Convolution::Param& param)
            -> opr::ConvBiasForward::Param {
        using Param = opr::ConvBiasForward::Param;
        return opr::ConvBiasForward::Param{Param::NonlineMode::IDENTITY,
                                           param.mode,
                                           param.sparse,
                                           param.format,
                                           param.pad_h,
                                           param.pad_w,
                                           param.stride_h,
                                           param.stride_w,
                                           param.dilate_h,
1759 1760 1761
                                           param.dilate_w,
                                           0,
                                           param.compute_mode};
1762 1763 1764 1765 1766 1767 1768 1769 1770
    };

    auto check_bias_shape = [&](opr::Convolution* conv, VarNode* bias) -> bool {
        bool valid_bias_shape = true;
        using Format = opr::Convolution::Param::Format;
        using Sparse = opr::Convolution::Param::Sparse;
        auto dst_shape = conv->output(0)->shape();
        auto filter_shape = conv->input(1)->shape();
        auto bias_shape = bias->shape();
1771 1772 1773 1774 1775 1776

        //! pay attention: make sure bias node is not const provider when
        //! batch > 1 cause shape assert problem in convbias
        //! if you resize the input shape, can not update the bias shape too.
        //! so do not fuse conv bias in this situation
        if (dst_shape.eq_shape(bias_shape) && !cg::is_const_var_shape(bias)) {
1777 1778
            return valid_bias_shape;
        }
1779

1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813
        size_t OC = filter_shape[0];
        if (conv->param().sparse == Sparse::GROUP) {
            OC *= filter_shape[1];
        }
        if (conv->param().format == Format::NCHW) {
            valid_bias_shape &=
                    ((bias_shape.ndim == 4) && (bias_shape[0] == 1) &&
                     (bias_shape[1] == OC) && (bias_shape[2] == 1) &&
                     (bias_shape[3] == 1));
        } else if (conv->param().format == Format::NCHW4) {
            valid_bias_shape &=
                    ((bias_shape.ndim == 5) && (bias_shape[0] == 1) &&
                     (bias_shape[1] == OC / 4) && (bias_shape[2] == 1) &&
                     (bias_shape[3] == 1) && bias_shape[4] == 4);
        } else if (conv->param().format == Format::NHWC) {
            valid_bias_shape &= ((bias_shape.ndim == 4) &&
                                 (bias_shape[0] == 1) && (bias_shape[1] == 1) &&
                                 (bias_shape[2] == 1) && (bias_shape[3] == OC));
        } else {
            valid_bias_shape &=
                    ((bias_shape.ndim == 5) && (bias_shape[0] == 1) &&
                     (bias_shape[1] == 1) && (bias_shape[2] == OC) &&
                     (bias_shape[3] == 1) && (bias_shape[4] == 4));
            mgb_assert(conv->param().format == Format::NHWCD4);
        }
        return valid_bias_shape;
    };

    auto try_fuse_typecvt = [&](opr::TypeCvt* typecvt) -> OperatorNodeBase* {
        mgb_assert(typecvt->input().size() == 1);
        auto conv_bias = try_cast_as_op<opr::ConvBias>(
                rewriter.get_var(typecvt->input(0))->owner_opr());
        if (!conv_bias || m_deps.count(typecvt->input(0)) != 1 ||
            typecvt->output(0)->dtype().enumv() !=
1814 1815 1816
                    DTypeTrait<dtype::QuantizedS8>::enumv ||
            typecvt->input(0)->dtype().enumv() !=
                    DTypeTrait<dtype::QuantizedS32>::enumv)
1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948
            return nullptr;

        auto config = conv_bias->config();
        config.output_dtype(typecvt->output(0)->dtype());
        if (conv_bias->input().size() == 3) {
            // conv + bias
            return opr::ConvBias::make(conv_bias->input(0), conv_bias->input(1),
                                       conv_bias->input(2), conv_bias->param(),
                                       conv_bias->execution_policy(), config)
                    .node()
                    ->owner_opr();
        } else {
            // conv without bias
            return opr::ConvBias::make(conv_bias->input(0), conv_bias->input(1),
                                       conv_bias->param(),
                                       conv_bias->execution_policy(), config)
                    .node()
                    ->owner_opr();
        }
    };
    auto on_opr = [&](OperatorNodeBase* opr) {
        auto check_conv = [](opr::Convolution* conv) -> bool {
            return conv->param().format ==
                           megdnn::param::Convolution::Format::NHWCD4 ||
                   conv->param().format ==
                           megdnn::param::Convolution::Format::NHWC ||
                   conv->param().format ==
                           megdnn::param::Convolution::Format::NCHW ||
                   conv->param().format ==
                           megdnn::param::Convolution::Format::NCHW4
                   ;
        };
        if (auto elem = try_cast_as_op<opr::Elemwise>(opr)) {
            if (try_fuse_bias_nonlinearity(elem) || try_fuse_bias(elem)) {
                auto inp1 = rewriter.get_var(elem->input(0));
                auto inp2 = rewriter.get_var(elem->input(1));
                opr::Convolution* conv = nullptr;
                size_t bias_idx = 0;
                if (inp1->owner_opr()->same_type<opr::Convolution>() &&
                    m_deps[elem->input(0)].size() == 1) {
                    conv = try_cast_as_op<opr::Convolution>(inp1->owner_opr());
                    bias_idx = 1;
                } else if (inp2->owner_opr()->same_type<opr::Convolution>() &&
                           m_deps[elem->input(1)].size() == 1) {
                    conv = try_cast_as_op<opr::Convolution>(inp2->owner_opr());
                    bias_idx = 0;
                }
                auto bias_inp = rewriter.get_var(elem->input(bias_idx));
                if (conv && check_conv(conv) &&
                    check_bias_shape(conv, bias_inp)) {
                    opr::ConvBiasForward::Param param =
                            convert_to_conv_bias_param(conv->param());
                    param.nonlineMode = get_nonlinearity_mode(elem);
                    auto new_var =
                            opr::ConvBiasForward::make(
                                    conv->input(0), conv->input(1), bias_inp,
                                    param, conv->execution_policy(),
                                    conv->config())
                                    .node();
                    rewriter.replace_var(
                            opr->output(0), new_var,
                            mgb_cstr_log("replace nonlinearity(conv(x, w) + b) "
                                         "-> conv_bias(x, w, b)"));
                    return;
                }
            } else if (try_fuse_nonlinearity(elem)) {
                auto inp = rewriter.get_var(elem->input(0));
                {
                    auto conv =
                            try_cast_as_op<opr::Convolution>(inp->owner_opr());
                    if (conv && check_conv(conv) &&
                        m_deps[elem->input(0)].size() == 1) {
                        opr::ConvBiasForward::Param param =
                                convert_to_conv_bias_param(conv->param());
                        param.nonlineMode = get_nonlinearity_mode(elem);
                        auto new_var = opr::ConvBiasForward::make(
                                               conv->input(0), conv->input(1),
                                               param, conv->execution_policy(),
                                               conv->config())
                                               .node();
                        rewriter.replace_var(
                                opr->output(0), new_var,
                                mgb_cstr_log("replace nonlinearity(conv(x, w)) "
                                             "-> conv_bias(x, w)"));
                        return;
                    }
                }
                {
                    auto conv = try_cast_as_op<opr::ConvBias>(inp->owner_opr());
                    auto check_conv_bias = [&](opr::ConvBias* opr) {
                        return opr->param().format ==
                                       opr::ConvBias::Param::Format::NHWC ||
                               opr->param().format ==
                                       opr::ConvBias::Param::Format::NCHW ||
                               opr->param().format ==
                                       opr::ConvBias::Param::Format::NCHW4
                               ;
                    };
                    if (conv && check_conv_bias(conv) &&
                        m_deps[elem->input(0)].size() == 1) {
                        auto param = conv->param();
                        param.nonlineMode = get_nonlinearity_mode(elem);
                        auto new_var = opr::ConvBiasForward::make(
                                               conv->input(0), conv->input(1),
                                               conv->input(2), param,
                                               conv->execution_policy(),
                                               conv->config())
                                               .node();
                        rewriter.replace_var(
                                opr->output(0), new_var,
                                mgb_cstr_log("replace nonlinearity(conv(x, w)) "
                                             "-> conv_bias(x, w)"));
                        return;
                    }
                }
            }
        } else if (auto typecvt = try_cast_as_op<opr::TypeCvt>(opr)) {
            auto new_opr = try_fuse_typecvt(typecvt);
            if (new_opr) {
                rewriter.replace_var(
                        opr->output(0), new_opr->output(0),
                        mgb_cstr_log("replace typecvt(conv_bias(x, w, b)) -> "
                                     "conv_bias(x, w, b)"));
                return;
            }
        }
        rewriter.auto_replace_outputs(opr);

    };
    state.graph().iter(on_opr);

    rewriter.apply_inplace();
1949
    MIDOUT_E
1950 1951 1952 1953 1954 1955 1956 1957
}

/* ================ FuseConvBiasZPass ================ */
const char* FuseConvBiasZPass::name() const {
    return "combine_conv_bias_and_z";
}

void FuseConvBiasZPass::apply(OptState& state) const {
1958
    MIDOUT_B("FuseConvBiasZPass::apply")
1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
    UniqReaderCheck uniq_reader_check{state.graph()};

    auto rewriter = state.graph().make_rewriter();
    using Mode = opr::Elemwise::Param::Mode;
    using MultiMode = opr::ElemwiseMultiType::Param::Mode;
    using NonlineMode = opr::ConvBiasForward::Param::NonlineMode;

    auto check_conv_bias = [](opr::ConvBias* conv_bias) -> bool {
        return conv_bias->param().format ==
                       megdnn::param::ConvBias::Format::NHWC ||
               conv_bias->param().format ==
                       megdnn::param::ConvBias::Format::NCHW ||
               conv_bias->param().format ==
                       megdnn::param::ConvBias::Format::NCHW4
               ;
    };
    auto check_fuse_shape = [&](opr::ConvBias* conv_bias, VarNode* z) -> bool {
        bool valid_fuse_shape = true;
        auto z_shape = z->shape();
        auto bias_shape = conv_bias->input(2)->shape();
        auto conv_bias_shape = conv_bias->output(0)->shape();

        valid_fuse_shape &= (!conv_bias_shape.eq_shape(bias_shape));
        valid_fuse_shape &= conv_bias_shape.eq_shape(z_shape);

        return valid_fuse_shape;
    };
    auto check_fuse_dtype = [&](opr::ConvBias* conv_bias, VarNode* z) -> bool {
        return conv_bias->output(0)->dtype().enumv() == z->dtype().enumv();
    };
    auto get_convbias_nonline_mode = [&](OperatorNodeBase* opr) -> NonlineMode {
        if (opr->same_type<opr::Elemwise>()) {
            auto elem = try_cast_as_op<opr::Elemwise>(opr);
            if (elem->param().mode == Mode::FUSE_ADD_RELU)
                return NonlineMode::RELU;
        }

        if (opr->same_type<opr::ElemwiseMultiType>()) {
            auto elem = try_cast_as_op<opr::ElemwiseMultiType>(opr);
            if (elem->param().mode == MultiMode::QFUSE_ADD_RELU)
                return NonlineMode::RELU;
2000 2001
            else if (elem->param().mode == MultiMode::QFUSE_ADD_H_SWISH)
                return NonlineMode::H_SWISH;
2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067
        }
        return NonlineMode::IDENTITY;
    };
    auto try_replace_var_node = [&](OperatorNodeBase* opr) {
        opr::ConvBias* conv_bias = nullptr;
        size_t z_idx = 0;
        size_t nr_inps = opr->input().size();
        for (size_t i = 0; i < nr_inps; i++) {
            auto inp = rewriter.get_var(opr->input(i));
            if (inp->owner_opr()->same_type<opr::ConvBias>()) {
                auto cb = try_cast_as_op<opr::ConvBias>(inp->owner_opr());
                if (cb->input().size() == 3 &&
                    cb->param().nonlineMode ==
                            opr::ConvBias::Param::NonlineMode::IDENTITY &&
                    uniq_reader_check(opr->input(i))) {
                    conv_bias = cb;
                    z_idx = nr_inps - i - 1;
                    break;
                }
            }
        }
        auto z_inp = rewriter.get_var(opr->input(z_idx));

        if (conv_bias && check_conv_bias(conv_bias) &&
            check_fuse_shape(conv_bias, z_inp) &&
            check_fuse_dtype(conv_bias, z_inp)) {
            auto param = conv_bias->param();
            param.nonlineMode = get_convbias_nonline_mode(opr);
            auto config = conv_bias->config();

            auto new_var = opr::ConvBiasForward::make(
                                   conv_bias->input(0), conv_bias->input(1),
                                   conv_bias->input(2), z_inp, param,
                                   conv_bias->execution_policy(),
                                   config.output_dtype(opr->output(0)->dtype()))
                                   .node();
            rewriter.replace_var(
                    opr->output(0), new_var,
                    mgb_cstr_log("replace "
                                 "nonlinearity(conv_bias(x,w,b) + z) "
                                 "-> conv_bias(x, w, b, z)"));
            uniq_reader_check.update_on_opr_auto_replace(opr,
                                                         new_var->owner_opr());
            return true;
        }
        return false;
    };
    auto try_fuse_elemwise = [&](OperatorNodeBase* opr) {
        if (!opr->same_type<opr::Elemwise>())
            return false;
        auto elem = try_cast_as_op<opr::Elemwise>(opr);
        if (elem->input().size() != 2)
            return false;
        if (elem->param().mode != Mode::ADD &&
            elem->param().mode != Mode::FUSE_ADD_RELU)
            return false;
        return try_replace_var_node(opr);
    };

    auto try_fuse_elemwise_multi_type = [&](OperatorNodeBase* opr) {
        if (!opr->same_type<opr::ElemwiseMultiType>())
            return false;
        auto elem = try_cast_as_op<opr::ElemwiseMultiType>(opr);
        if (elem->input().size() != 2)
            return false;
        if (elem->param().mode != MultiMode::QADD &&
2068 2069
            elem->param().mode != MultiMode::QFUSE_ADD_RELU &&
            elem->param().mode != MultiMode::QFUSE_ADD_H_SWISH)
2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084
            return false;
        return try_replace_var_node(opr);
    };

    auto on_opr = [&](OperatorNodeBase* opr) {
        if (try_fuse_elemwise(opr))
            return;
        if (try_fuse_elemwise_multi_type(opr))
            return;
        auto new_opr = rewriter.auto_replace_outputs(opr);
        uniq_reader_check.update_on_opr_auto_replace(opr, new_opr);
    };
    state.graph().iter(on_opr);

    rewriter.apply_inplace();
2085
    MIDOUT_E
2086 2087 2088 2089 2090 2091 2092 2093 2094
}

/* ================ FuseDeconvCvtPass ================ */
const char* FuseDeconvCvtPass::name() const {
    return "combine_deconv_and_typecvt";
}


void FuseDeconvCvtPass::apply(OptState& state) const {
2095
    MIDOUT_B("FuseDeconvCvtPass::apply")
2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145
    std::unordered_map<VarNode*, std::vector<OperatorNodeBase*>> m_deps;
    state.graph().iter([&m_deps](OperatorNodeBase* opr) {
        for (auto& inp : opr->input()) {
            m_deps[inp].push_back(opr);
        }
    });

    UniqReaderCheck uniq_reader_check{state.graph()};
    auto rewriter = state.graph().make_rewriter();
    auto try_fuse_deconv_typecvt =
            [&](opr::TypeCvt* typecvt) -> OperatorNodeBase* {
        mgb_assert(typecvt->input().size() == 1);
        auto deconv = try_cast_as_op<opr::ConvolutionBackwardData>(
                rewriter.get_var(typecvt->input(0))->owner_opr());
        if (!deconv
                || m_deps.count(typecvt->input(0)) != 1 ||
            typecvt->output(0)->dtype().enumv() !=
                    DTypeTrait<dtype::QuantizedS8>::enumv) {
            return nullptr;
        }
        if (!uniq_reader_check(deconv->output(0)))
            return nullptr;

        auto config = deconv->config();
        config.output_dtype(typecvt->output(0)->dtype());
        return opr::ConvolutionBackwardData::make(
                       deconv->input(0), deconv->input(1), deconv->param(),
                       deconv->execution_policy(), config)
                .node()
                ->owner_opr();
    };

    auto on_opr = [&](OperatorNodeBase* opr) {
        if (auto typecvt = try_cast_as_op<opr::TypeCvt>(opr)) {
            if (auto deconv_new = try_fuse_deconv_typecvt(typecvt)) {
                rewriter.replace_var(
                        opr->output(0), deconv_new->output(0),
                        mgb_cstr_log("replace typecvt(deconv(x, w)) -> "
                                     "deconv(x, w)"));
                uniq_reader_check.update_on_opr_auto_replace(opr, deconv_new);
                return;
            }
        }
        auto new_opr = rewriter.auto_replace_outputs(opr);
        uniq_reader_check.update_on_opr_auto_replace(
                opr, new_opr);
    };
    state.graph().iter(on_opr);

    rewriter.apply_inplace();
2146
    MIDOUT_E
2147 2148 2149 2150 2151 2152 2153 2154
}

/* ================ ParamMergePass ================ */
const char* ParamMergePass::name() const {
    return mgb_cstr_log("param_merge");
}

void ParamMergePass::apply(OptState& opt_state) const {
2155
    MIDOUT_B("ParamMergePass::apply")
2156 2157 2158 2159
    param_merge<opr::SharedDeviceTensor, opr::MultipleDeviceTensorHolder>(
            opt_state);
    param_merge<opr::SharedDeviceTensorWithFormat,
                opr::MultipleDeviceTensorWithFormatHolder>(opt_state);
2160
    MIDOUT_E
2161 2162 2163
}

// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}