feat(serialization): support the registry for new serialization format

GitOrigin-RevId: 8eacd5e77c

src/core/include/megbrain/utils/hash_ct.h

@@ -153,4 +153,6 @@ struct EnsureHashConstexpr {
 #define MGB_HASH_STR(v) \
     ::mgb::EnsureHashConstexpr<::mgb::XXHash64CT::hash(v, sizeof(v), 20160701)>::val
 
+#define MGB_HASH_RUNTIME(v) XXHash64CT::hash((v).c_str(), (v).size() + 1, 20160701)
+
 // vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/opr/impl/custom_opnode.sereg.h

@@ -52,6 +52,21 @@ mgb::cg::OperatorNodeBase* custom_loader(
     }                                                         \
     MGB_SEREG_OPR_INTL_CALL_ENTRY(cls, _OprReg##cls)
 
+#define CUSTOM_OP_SEREG_REG_V2(cls, _version_min, _version_max)                 \
+    namespace {                                                                 \
+    struct _OprRegV2##cls {                                                     \
+        static void entry() {                                                   \
+            MGB_SEREG_OPR_INTL_CALL_ADD_V2(                                     \
+                    cls, ::mgb::serialization::custom_dumper,                   \
+                    ::mgb::serialization::custom_loader, nullptr, _version_min, \
+                    _version_max);                                              \
+        }                                                                       \
+    };                                                                          \
+    }                                                                           \
+    MGB_SEREG_OPR_INTL_CALL_ENTRY_V2(cls, _OprRegV2##cls)
+
 using namespace mgb;
 using CustomOpNode = opr::CustomOpNode;
 CUSTOM_OP_SEREG_REG(CustomOpNode);
+
+CUSTOM_OP_SEREG_REG_V2(CustomOpNode, 2, CURRENT_VERSION);

src/opr/impl/dnn/dnn.sereg.v2.h

@@ -0,0 +1,228 @@
+#include "megbrain/graph/symbol_var.h"
+#include "megdnn/oprs/general.h"
+#if MGB_ENABLE_FBS_SERIALIZATION
+#include "megbrain/opr/basic_arith.h"
+#include "megbrain/opr/dnn/softmax.h"
+#include "megbrain/serialization/oss_opr_load_dump.h"
+#include "megbrain/serialization/sereg.h"
+#include "megdnn/opr_param_defs.h"
+#include "megdnn/oprs/nn.h"
+
+namespace mgb {
+namespace serialization {
+
+template <>
+struct OprLoadDumpImplV2<opr::Softmax, 1> {
+    using Opr = opr::Softmax;
+    using PersisParam = opr::Softmax::Param;
+    static void dump(OprDumpContext& ctx, const cg::OperatorNodeBase& opr) {
+        ctx.write_param<PersisParam>(opr.cast_final_safe<Opr>().param());
+    }
+
+    static cg::OperatorNodeBase* replace_opr(
+            cg::OperatorNodeBase* opr, const VarNodeArray& inputs) {
+        int32_t axis = opr->cast_final_safe<Opr>().param().axis;
+        auto input_var = inputs[0];
+        auto max_reduce_out =
+                opr::Reduce::make(input_var, {megdnn::Reduce::Mode::MAX, axis});
+        auto elemwise_sub_out = opr::Elemwise::make(
+                {input_var, max_reduce_out}, {megdnn::Elemwise::Mode::SUB});
+        auto elemwise_exp_out =
+                opr::Elemwise::make({elemwise_sub_out}, {megdnn::Elemwise::Mode::EXP});
+        auto sum_reduce_out =
+                opr::Reduce::make(elemwise_exp_out, {megdnn::Reduce::Mode::SUM, axis});
+        auto out = opr::Elemwise::make(
+                {elemwise_exp_out, sum_reduce_out}, {megdnn::Elemwise::Mode::TRUE_DIV});
+        return out.node()->owner_opr();
+    }
+
+    static cg::OperatorNodeBase* load(
+            OprLoadContext& ctx, const cg::VarNodeArray& inputs,
+            const OperatorNodeConfig& config) {
+        auto& fbs_ctx = CAST_TO_FBS_V2_CTX(ctx);
+        auto param = fbs_ctx.read_param<PersisParam>(0);
+        return Opr::make(inputs[0], param, config).node()->owner_opr();
+    }
+};
+
+template <
+        class Opr, class Maker0, class MegDNNConv,
+        class Maker1 = MakeConvCallerEmpty<MegDNNConv>,
+        class Maker2 = MakeConvCallerEmpty<MegDNNConv>,
+        typename ConvParam = megdnn::param::Convolution>
+struct WithPolicyOprLoadDumpImpl {
+    static void dump(OprDumpContext& ctx, const cg::OperatorNodeBase& opr_) {
+        auto&& opr = opr_.cast_final_safe<Opr>();
+        ctx.write_param<ConvParam>(opr.param());
+        ctx.write_param<megdnn::param::ExecutionPolicy>(
+                opr.execution_policy_transient());
+    }
+    static VarNode* make(
+            const cg::VarNodeArray& inputs, const ConvParam& param,
+            const megdnn::param::ExecutionPolicy& execution_policy,
+            const OperatorNodeConfig& config) {
+        VarNode* ret =
+                Maker0::template make<Opr>(inputs, param, execution_policy, config);
+        if (!ret) {
+            ret = Maker1::template make<Opr>(inputs, param, execution_policy, config);
+        }
+        if (!ret) {
+            ret = Maker2::template make<Opr>(inputs, param, execution_policy, config);
+        }
+        mgb_assert(ret);
+        return ret;
+    }
+
+    static cg::OperatorNodeBase* load(
+            OprLoadContext& ctx, const cg::VarNodeArray& inputs,
+            const OperatorNodeConfig& config) {
+        auto& fbs_ctx = CAST_TO_FBS_V2_CTX(ctx);
+        auto fopr = reinterpret_cast<const fbs::v2::Operator*>(
+                fbs_ctx.get_current_opr_data());
+        auto conv_param = fbs_ctx.read_param<ConvParam>(0);
+        megdnn::param::ExecutionPolicy policy;
+        if (fopr->additional_params() && fopr->additional_params()->size()) {
+            policy = fbs_ctx.read_param<megdnn::param::ExecutionPolicy>(1);
+        }
+        return make(inputs, conv_param, policy, config)->owner_opr();
+    }
+};
+
+template <>
+struct OprLoadDumpImplV2<opr::Convolution, 0>
+        : public WithPolicyOprLoadDumpImpl<
+                  opr::Convolution, MakeConvCaller2<megdnn::Convolution>,
+                  megdnn::Convolution> {};
+template <>
+struct OprLoadDumpImplV2<opr::ConvolutionBackwardData, 0>
+        : public WithPolicyOprLoadDumpImpl<
+                  opr::ConvolutionBackwardData, MakeConvCaller2<megdnn::Convolution>,
+                  megdnn::Convolution, MakeConvCaller3<megdnn::Convolution>> {};
+template <>
+struct OprLoadDumpImplV2<opr::ConvolutionBackwardFilter, 0>
+        : public WithPolicyOprLoadDumpImpl<
+                  opr::ConvolutionBackwardFilter, MakeConvCaller3<megdnn::Convolution>,
+                  megdnn::Convolution> {};
+
+template <>
+struct OprLoadDumpImplV2<opr::Convolution3D, 0>
+        : public WithPolicyOprLoadDumpImpl<
+                  opr::Convolution3D, MakeConvCaller2<megdnn::Convolution3D>,
+                  megdnn::Convolution3D, MakeConvCallerEmpty<megdnn::Convolution3D>,
+                  MakeConvCallerEmpty<megdnn::Convolution3D>,
+                  megdnn::param::Convolution3D> {};
+template <>
+struct OprLoadDumpImplV2<opr::Convolution3DBackwardData, 0>
+        : public WithPolicyOprLoadDumpImpl<
+                  opr::Convolution3DBackwardData,
+                  MakeConvCaller2<megdnn::Convolution3D>, megdnn::Convolution3D,
+                  MakeConvCaller3<megdnn::Convolution3D>,
+                  MakeConvCallerEmpty<megdnn::Convolution3D>,
+                  megdnn::param::Convolution3D> {};
+template <>
+struct OprLoadDumpImplV2<opr::Convolution3DBackwardFilter, 0>
+        : public WithPolicyOprLoadDumpImpl<
+                  opr::Convolution3DBackwardFilter,
+                  MakeConvCaller3<megdnn::Convolution3D>, megdnn::Convolution3D,
+                  MakeConvCallerEmpty<megdnn::Convolution3D>,
+                  MakeConvCallerEmpty<megdnn::Convolution3D>,
+                  megdnn::param::Convolution3D> {};
+template <>
+struct OprLoadDumpImplV2<opr::ConvBiasForward, 0>
+        : public WithPolicyOprLoadDumpImpl<
+                  opr::ConvBiasForward, MakeConvCaller2<megdnn::ConvBiasForward>,
+                  megdnn::ConvBiasForward, MakeConvCaller3<megdnn::ConvBiasForward>,
+                  MakeConvCaller4<megdnn::ConvBiasForward>, megdnn::param::ConvBias> {};
+template <>
+struct OprLoadDumpImplV2<opr::BatchConvBiasForward, 0>
+        : public WithPolicyOprLoadDumpImpl<
+                  opr::BatchConvBiasForward,
+                  MakeConvCaller2<megdnn::BatchConvBiasForward>,
+                  megdnn::BatchConvBiasForward,
+                  MakeConvCaller3<megdnn::BatchConvBiasForward>,
+                  MakeConvCaller4<megdnn::BatchConvBiasForward>,
+                  megdnn::param::BatchConvBias> {};
+
+template <>
+struct OprLoadDumpImplV2<opr::LocalShare, 0>
+        : public WithPolicyOprLoadDumpImpl<
+                  opr::LocalShare, MakeLocalShareCaller2<megdnn::LocalShare>,
+                  megdnn::LocalShare, MakeLocalShareCallerEmpty<megdnn::LocalShare>,
+                  MakeLocalShareCallerEmpty<megdnn::LocalShare>,
+                  megdnn::param::LocalShare> {};
+template <>
+struct OprLoadDumpImplV2<opr::LocalShareBackwardData, 0>
+        : public WithPolicyOprLoadDumpImpl<
+                  opr::LocalShareBackwardData,
+                  MakeLocalShareCaller3<megdnn::LocalShare>, megdnn::LocalShare,
+                  MakeLocalShareCallerEmpty<megdnn::LocalShare>,
+                  MakeLocalShareCallerEmpty<megdnn::LocalShare>,
+                  megdnn::param::LocalShare> {};
+template <>
+struct OprLoadDumpImplV2<opr::LocalShareBackwardFilter, 0>
+        : public WithPolicyOprLoadDumpImpl<
+                  opr::LocalShareBackwardFilter,
+                  MakeLocalShareCaller3<megdnn::LocalShare>, megdnn::LocalShare,
+                  MakeLocalShareCallerEmpty<megdnn::LocalShare>,
+                  MakeLocalShareCallerEmpty<megdnn::LocalShare>,
+                  megdnn::param::LocalShare> {};
+template <>
+struct OprLoadDumpImplV2<opr::DeformableConvForward, 0>
+        : public WithPolicyOprLoadDumpImpl<
+                  opr::DeformableConvForward,
+                  MakeConvCaller4<megdnn::DeformableConvForward>, megdnn::Convolution> {
+};
+template <>
+struct OprLoadDumpImplV2<opr::DeformableConvBackwardData, 0>
+        : public WithPolicyOprLoadDumpImpl<
+                  opr::DeformableConvBackwardData,
+                  MakeConvCaller5<megdnn::DeformableConvBackwardData>,
+                  megdnn::Convolution> {};
+template <>
+struct OprLoadDumpImplV2<opr::DeformableConvBackwardFilter, 0>
+        : public WithPolicyOprLoadDumpImpl<
+                  opr::DeformableConvBackwardFilter,
+                  MakeConvCaller5<megdnn::DeformableConvBackwardFilter>,
+                  megdnn::Convolution> {};
+
+}  // namespace serialization
+
+namespace opr {
+#define SERGE_OPR_V2_CONVERTER(_cls, _arity, _converter) \
+    MGB_SEREG_OPR_V2(_cls, _arity, _converter, VERSION_2, CURRENT_VERSION);
+
+#define SERGE_OPR_V2_NO_CONVERTER(_cls, _arity) \
+    MGB_SEREG_OPR_V2(_cls, _arity, nullptr, VERSION_2, CURRENT_VERSION);
+
+SERGE_OPR_V2_CONVERTER(
+        Softmax, 1,
+        (mgb::serialization::OprLoadDumpImplV2<opr::Softmax, 1>::replace_opr));
+
+SERGE_OPR_V2_NO_CONVERTER(ConvBiasForward, 0)
+SERGE_OPR_V2_NO_CONVERTER(BatchConvBiasForward, 0);
+
+SERGE_OPR_V2_NO_CONVERTER(Convolution, 0)
+SERGE_OPR_V2_NO_CONVERTER(ConvolutionBackwardData, 0)
+SERGE_OPR_V2_NO_CONVERTER(ConvolutionBackwardFilter, 0)
+
+SERGE_OPR_V2_NO_CONVERTER(Convolution3D, 0);
+SERGE_OPR_V2_NO_CONVERTER(Convolution3DBackwardData, 0);
+SERGE_OPR_V2_NO_CONVERTER(Convolution3DBackwardFilter, 0);
+
+SERGE_OPR_V2_NO_CONVERTER(LocalShareForward, 0);
+SERGE_OPR_V2_NO_CONVERTER(LocalShareBackwardData, 0);
+SERGE_OPR_V2_NO_CONVERTER(LocalShareBackwardFilter, 0);
+
+SERGE_OPR_V2_NO_CONVERTER(DeformableConvForward, 0);
+SERGE_OPR_V2_NO_CONVERTER(DeformableConvBackwardData, 0);
+SERGE_OPR_V2_NO_CONVERTER(DeformableConvBackwardFilter, 0);
+
+#undef SERGE_OPR_V2_CONVERTER
+#undef SERGE_OPR_V2_NO_CONVERTER
+}  // namespace opr
+
+}  // namespace mgb
+
+#endif
+
+// vim: ft=cpp syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/opr/impl/io.sereg.v2.h

@@ -0,0 +1,197 @@
+#if MGB_ENABLE_FBS_SERIALIZATION
+#include "megbrain/comp_node_env.h"
+#include "megbrain/opr/dnn/softmax.h"
+#include "megbrain/opr/io.h"
+#include "megbrain/serialization/oss_opr_load_dump.h"
+#include "megbrain/serialization/sereg.h"
+
+#include "megbrain/serialization/internal/mgb_cpp_opr_generated.h"
+#include "megbrain/serialization/internal/schema_v2_generated.h"
+
+namespace mgb {
+namespace serialization {
+
+template <>
+struct OprLoadDumpImplV2<opr::ImmutableTensor, 0> {
+    using Opr = opr::ImmutableTensor;
+
+    static void dump(OprDumpContext& ctx, const cg::OperatorNodeBase& opr_) {
+        using Meth = OprDumpContext::TensorWriteMethod;
+        auto&& opr = opr_.cast_final_safe<Opr>();
+        ctx.dump_tensor(
+                {}, HostTensorND{}.copy_from(opr.value()).sync(),
+                Meth::VALUE_ANONYMOUS);
+    }
+
+    static cg::OperatorNodeBase* load(
+            OprLoadContext& ctx, const cg::VarNodeArray& inputs,
+            const OperatorNodeConfig& config) {
+        auto& fbs_ctx = CAST_TO_FBS_V2_CTX(ctx);
+        mgb_assert(inputs.empty());
+        auto fopr = reinterpret_cast<const fbs::v2::Operator*>(
+                fbs_ctx.get_current_opr_data());
+        if (fopr->tensors() && fopr->tensors()->size() > 0) {
+            auto val = fbs_ctx.load_tensor();
+            return Opr::make(fbs_ctx.graph(), *val, config).node()->owner_opr();
+        } else {
+            mgb_throw(SerializationError, "ImmutableTensor load with no tensor data.");
+        }
+    }
+};
+
+template <>
+struct OprLoadDumpImplV2<opr::Host2DeviceCopy, 0> {
+    using Opr = opr::Host2DeviceCopy;
+    static void dump(OprDumpContext& ctx, const cg::OperatorNodeBase& opr_) {
+        auto&& opr = opr_.cast_final_safe<Opr>();
+        ctx.write_param(opr.param());
+
+        using Meth = OprDumpContext::TensorWriteMethod;
+        ctx.dump_tensor(
+                opr.name(), *opr.host_data(),
+                opr.param().dump_default_value ? Meth::VALUE_INPUT : Meth::META_INPUT);
+    }
+    static cg::OperatorNodeBase* load(
+            OprLoadContext& ctx, const cg::VarNodeArray& inputs,
+            const OperatorNodeConfig& config) {
+        mgb_assert(inputs.empty());
+        auto& fbs_ctx = CAST_TO_FBS_V2_CTX(ctx);
+        auto param = fbs_ctx.read_param<Opr::Param>(0);
+        auto tensor = fbs_ctx.load_tensor();
+        return Opr::make(fbs_ctx.graph(), tensor, param, config).node()->owner_opr();
+    }
+};
+
+template <>
+struct OprLoadDumpImplV2<opr::SharedDeviceTensorWithFormat, 0> {
+    static void dump(OprDumpContext& ctx, const cg::OperatorNodeBase& opr_) {
+        using Meth = OprDumpContext::TensorWriteMethod;
+        auto&& opr = opr_.cast_final_safe<opr::SharedDeviceTensorWithFormat>();
+        HostTensorND val;
+        val.copy_from(opr.get_dev_tensor()).sync();
+        ctx.dump_tensor({}, val, Meth::VALUE_ANONYMOUS);
+    }
+
+    static cg::OperatorNodeBase* load(
+            OprLoadContext& ctx, const cg::VarNodeArray& inputs,
+            const OperatorNodeConfig& config) {
+        mgb_assert(inputs.empty());
+        auto val = ctx.load_tensor();
+        auto dev_val =
+                std::make_shared<DeviceTensorND>(val->comp_node(), val->layout());
+        dev_val->copy_from_fixlayout(*val);
+        auto out_var =
+                opr::SharedDeviceTensorWithFormat::make(ctx.graph(), dev_val, config);
+        dev_val->sync();
+        return out_var.node()->owner_opr();
+    }
+};
+
+template <>
+struct OprLoadDumpImplV2<opr::MultipleDeviceTensorHolder, 0> {
+    using Opr = opr::MultipleDeviceTensorHolder;
+
+    static void dump(OprDumpContext& ctx, const cg::OperatorNodeBase& opr_) {
+        using Meth = OprDumpContext::TensorWriteMethod;
+        auto&& opr = opr_.cast_final_safe<Opr>();
+        uint32_t nr_val = opr.values().size();
+        for (uint32_t i = 0; i < nr_val; ++i) {
+            HostTensorND val;
+            val.copy_from(*opr.values()[i]).sync();
+            ctx.dump_tensor(opr.output(i)->name(), val, Meth::VALUE_SHARED);
+        }
+    }
+
+    static cg::OperatorNodeBase* load(
+            OprLoadContext& ctx, const cg::VarNodeArray& inputs,
+            const OperatorNodeConfig& config) {
+        mgb_assert(inputs.empty());
+        auto& fbs_ctx = CAST_TO_FBS_V2_CTX(ctx);
+        auto fopr = reinterpret_cast<const fbs::v2::Operator*>(
+                fbs_ctx.get_current_opr_data());
+        uint32_t nr = 0;
+        if (fopr && fopr->tensors()) {
+            nr = fopr->tensors()->size();
+        }
+        Opr::ValueArray values(nr);
+        for (auto&& i : values) {
+            i = ctx.load_tensor_shared();
+        }
+        return Opr::make(ctx.graph(), std::move(values), config)[0].node()->owner_opr();
+    }
+};
+
+template <>
+struct OprLoadDumpImplV2<opr::MultipleDeviceTensorWithFormatHolder, 0> {
+    using Opr = opr::MultipleDeviceTensorWithFormatHolder;
+
+    static void dump(OprDumpContext& ctx, const cg::OperatorNodeBase& opr_) {
+        using Meth = OprDumpContext::TensorWriteMethod;
+        auto&& opr = opr_.cast_final_safe<Opr>();
+        uint32_t nr_val = opr.values().size();
+        for (uint32_t i = 0; i < nr_val; ++i) {
+            HostTensorND val;
+            auto value = *opr.values()[i];
+            val.copy_from(value).sync();
+            ctx.dump_tensor(opr.output(i)->name(), val, Meth::VALUE_SHARED);
+        }
+    }
+
+    static cg::OperatorNodeBase* load(
+            OprLoadContext& ctx, const cg::VarNodeArray& inputs,
+            const OperatorNodeConfig& config) {
+        mgb_assert(inputs.empty());
+        auto& fbs_ctx = CAST_TO_FBS_V2_CTX(ctx);
+        auto fopr = reinterpret_cast<const fbs::v2::Operator*>(
+                fbs_ctx.get_current_opr_data());
+        uint32_t nr = 0;
+        if (fopr && fopr->tensors()) {
+            nr = fopr->tensors()->size();
+        }
+        Opr::ValueArray values(nr);
+        for (auto&& i : values) {
+            i = ctx.load_tensor_shared();
+            //! set tensor format
+            TensorLayout layout_with_format = i->layout();
+
+            if (i->storage().comp_node().mem_node() ==
+                CompNode::default_cpu().mem_node()) {
+                mgb_assert(
+                        i->storage().ptr(),
+                        "storage should not be nullptr if mem_node is "
+                        "default_cpu");
+                HostTensorND src{i->storage().comp_node(), layout_with_format};
+                src.copy_from_fixlayout(*i).sync();
+                *i = DeviceTensorND::make_proxy(src);
+            } else {
+                //! actually only layout of this tensor will be used later, see
+                //! src/serialization/impl/batched_device_value_loader.cpp:49. But we
+                //! have no way to reset layout only, so just construct a invalid
+                //! storage instead
+                auto size = layout_with_format.span().dist_byte();
+                DeviceTensorStorage storage;
+                storage.reset(i->comp_node(), size, nullptr);
+                i->reset(storage, layout_with_format);
+            }
+        }
+        return Opr::make(ctx.graph(), std::move(values), config)[0].node()->owner_opr();
+    }
+};
+
+}  // namespace serialization
+
+namespace opr {
+#define SERGE_OPR_V2_NO_CONVERTER(_cls, _arity) \
+    MGB_SEREG_OPR_V2(_cls, _arity, nullptr, VERSION_2, CURRENT_VERSION);
+
+SERGE_OPR_V2_NO_CONVERTER(ImmutableTensor, 0);
+SERGE_OPR_V2_NO_CONVERTER(Host2DeviceCopy, 0);
+SERGE_OPR_V2_NO_CONVERTER(SharedDeviceTensorWithFormat, 0);
+SERGE_OPR_V2_NO_CONVERTER(MultipleDeviceTensorWithFormatHolder, 0);
+SERGE_OPR_V2_NO_CONVERTER(MultipleDeviceTensorHolder, 0);
+}  // namespace opr
+}  // namespace mgb
+
+#endif
+
+// vim: ft=cpp syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/opr/impl/loop/forward_sereg.cpp

@@ -135,6 +135,16 @@ void LoopSerializer::reg_all() {
     MGB_SEREG_OPR_INTL_CALL_ADD(InputMaker, dump_input_maker, load_input_maker);
     MGB_SEREG_OPR_INTL_CALL_ADD(
             CounterProvider, dump_counter_provider, load_counter_provider);
+
+    MGB_SEREG_OPR_INTL_CALL_ADD_V2(
+            opr::Loop, dump_loop, load_loop, nullptr, 2,
+            CURRENT_VERSION);
+    MGB_SEREG_OPR_INTL_CALL_ADD_V2(
+            InputMaker, dump_input_maker, load_input_maker, nullptr, 2,
+            CURRENT_VERSION);
+    MGB_SEREG_OPR_INTL_CALL_ADD_V2(
+            CounterProvider, dump_counter_provider, load_counter_provider, nullptr, 2,
+            CURRENT_VERSION);
 }
 
 void LoopSerializer::dump_loop(OprDumpContext& ctx, const cg::OperatorNodeBase& opr) {

src/serialization/impl/opr_registry.cpp

@@ -20,6 +20,11 @@ struct StaticData {
     ThinHashMap<Typeinfo*, OprRegistry*> type2reg;
     std::unordered_map<std::string, OprRegistry*> name2reg;
     ThinHashMap<size_t, OprRegistry*> unversioned_id2reg;
+
+    //! versioned OprRegistryV2, version_id_reg_map is used for Operator
+    //! load/shallow copy and version_type_reg_map is used for Operator dump
+    ThinHashMap<uint8_t, ThinHashMap<size_t, OprRegistryV2>> version_id_reg_map;
+    ThinHashMap<uint8_t, ThinHashMap<Typeinfo*, OprRegistryV2*>> version_type_reg_map;
 };
 
 StaticData& static_data() {
@@ -47,6 +52,20 @@ const OprRegistry* dynamic_registry() {
     return ret;
 }
 
+const OprRegistryV2* dynamic_registry_v2() {
+    static const OprRegistryV2* ret = nullptr;
+    if (ret)
+        return ret;
+
+    auto id = MGB_HASH_STR("dynamic");
+    OprRegistryV2::versioned_add(
+            {nullptr, id, {}, {}, dynamic_loader, {}}, CURRENT_VERSION,
+            CURRENT_VERSION);
+    ret = OprRegistryV2::versioned_find_by_id(id, CURRENT_VERSION);
+    mgb_assert(ret);
+    return ret;
+}
+
 class _Init {
 public:
     _Init() {
@@ -63,8 +82,7 @@ void OprRegistry::add(const OprRegistry& record) {
     auto registry_ins = sd.id2reg.emplace(persist_id, record);
     mgb_assert(
             registry_ins.second || persist_id == dynamic_registry()->persist_type_id,
-            "duplicated operator persist_type_id: %s",
-            std::to_string(persist_id).c_str());
+            "duplicated operator name : %s", record.name.c_str());
 
     OprRegistry* persis_record_ptr;
     if (registry_ins.second) {
@@ -129,6 +147,73 @@ const OprRegistry* OprRegistry::find_by_unversioned_id(size_t unversioned_id) {
     return iter == uid2reg.end() ? nullptr : iter->second;
 }
 
+//! find the registry equal to the giving version
+const OprRegistryV2* OprRegistryV2::versioned_find_by_id(
+        const size_t id, uint8_t version) {
+    auto&& id_reg_map = static_data().version_id_reg_map;
+    auto iter_version = id_reg_map.find(version);
+    if (iter_version != id_reg_map.end()) {
+        auto iter = iter_version->second.find(id);
+        return iter == iter_version->second.end() ? nullptr : &iter->second;
+    }
+    return nullptr;
+}
+//! find the registry equal or below the giving version
+const OprRegistryV2* OprRegistryV2::versioned_find_by_typeinfo(
+        Typeinfo* type, uint8_t version) {
+    const auto& type_reg_map = static_data().version_type_reg_map;
+    for (int version_id = version; version_id > 0; version_id--) {
+        auto iter_version = type_reg_map.find(version_id);
+        if (iter_version != type_reg_map.end()) {
+            auto iter = iter_version->second.find(type);
+            if (iter == iter_version->second.end()) {
+                continue;
+            } else {
+                return iter->second;
+            }
+        }
+    }
+    return nullptr;
+}
+
+void OprRegistryV2::versioned_add(
+        const OprRegistryV2& record, uint8_t min_version, uint8_t max_version) {
+    mgb_assert(max_version >= min_version);
+
+    auto&& sd = static_data();
+    auto id = record.type_id;
+    uint64_t type_id = id;
+    //! record.type->name is nullptr when MGB_VERBOSE_TYPEINFO_NAME==0
+    if (record.type && record.type->name) {
+        type_id = MGB_HASH_RUNTIME(std::string(record.type->name));
+    }
+    for (uint8_t version = min_version; version <= max_version; version++) {
+        auto&& registry_map = sd.version_id_reg_map[version];
+        auto versioned_record = record;
+        versioned_record.version = version;
+        mgb_assert(
+                registry_map.find(id) == registry_map.end() ||
+                        id == dynamic_registry_v2()->type_id,
+                "dduplicated OprRegistryV2 of %s\n", record.name.c_str());
+        auto registry_ins = registry_map.emplace(id, versioned_record);
+        if (!registry_ins.second) {
+            //! the registry is dynamic
+            mgb_assert(!record.converter);
+            registry_map[id] = versioned_record;
+        }
+        //! sometimes the register id and the hash typeinfo is not same, just as
+        //! dynamic Operator
+        if (id != type_id) {
+            mgb_assert(
+                    registry_map.find(type_id) == registry_map.end(),
+                    "dduplicated OprRegistryV2 of %s\n", record.name.c_str());
+            registry_map.emplace(type_id, versioned_record);
+        }
+        auto&& registry_type_map = sd.version_type_reg_map[version];
+        registry_type_map.emplace(record.type, &registry_map[id]);
+    }
+}
+
 void OprRegistry::add_using_dynamic_loader(
         Typeinfo* type, const std::string& name, const OprDumper& dumper) {
     // dynamic oprs are implemented by mapping different opr types to the same
@@ -140,6 +225,11 @@ void OprRegistry::add_using_dynamic_loader(
          {},
          {},
          dynamic_registry()->unversioned_type_id});
+    mgb_assert(type, "type must be not nullptr");
+    OprRegistryV2::versioned_add(
+            {type, dynamic_registry_v2()->type_id, type->name, dumper,
+             dynamic_registry_v2()->loader, nullptr},
+            CURRENT_VERSION, CURRENT_VERSION);
 }
 
 #if MGB_ENABLE_DEBUG_UTIL

src/serialization/impl/sereg_caller.cpp

@@ -9,10 +9,12 @@ void call_sereg() {}
 #include "../../opr/impl/blas.sereg.h"
 #include "../../opr/impl/cond.sereg.h"
 #include "../../opr/impl/dnn/dnn.sereg.h"
+#include "../../opr/impl/dnn/dnn.sereg.v2.h"
 #include "./extern_c_opr.sereg.h"
 #include "../../opr/impl/imgproc.sereg.h"
 #include "../../opr/impl/indexing.sereg.h"
 #include "../../opr/impl/io.sereg.h"
+#include "../../opr/impl/io.sereg.v2.h"
 #include "../../opr/impl/loop/forward.sereg.h"
 #include "../../opr/impl/loop/grad.sereg.h"
 #include "../../opr/impl/misc.sereg.h"

src/serialization/include/megbrain/serialization/opr_registry.h

@@ -53,7 +53,6 @@ struct OprRegistry {
     uint64_t unversioned_type_id;
 
     MGE_WIN_DECLSPEC_FUC static void add(const OprRegistry& record);
-
     /*!
      * \brief register an operator to use dynamic loader
      *
@@ -89,6 +88,39 @@ struct OprRegistry {
 #endif
 };
 
+//! Convert some modified Opr to compatible Opr
+using OprConvertToCompatible = thin_function<cg::OperatorNodeBase*(
+        cg::OperatorNodeBase*, const VarNodeArray&)>;
+
+//! record of a single operator
+struct OprRegistryV2 {
+    Typeinfo* type;
+    uint64_t type_id;
+    std::string name;
+    OprDumper dumper;
+    OprLoaderWrapper loader;
+    OprConvertToCompatible converter;
+    uint8_t version = 2;
+
+    MGE_WIN_DECLSPEC_FUC uint8_t get_version() const { return version; }
+
+    //! register opr load/dump to version2regmap
+    MGE_WIN_DECLSPEC_FUC static void versioned_add(
+            const OprRegistryV2& record, uint8_t min_version, uint8_t max_version);
+
+    MGE_WIN_DECLSPEC_FUC static const OprRegistryV2* versioned_find_by_id(
+            const size_t id, uint8_t version);
+
+    MGE_WIN_DECLSPEC_FUC static const OprRegistryV2* versioned_find_by_typeinfo(
+            Typeinfo* type, uint8_t version);
+
+#if MGB_ENABLE_DEBUG_UTIL
+    //! dump registered oprs
+    MGE_WIN_DECLSPEC_FUC static std::vector<std::pair<size_t, std::string>>
+    dump_registries();
+#endif
+};
+
 }  // namespace serialization
 }  // namespace mgb
 

src/serialization/include/megbrain/serialization/sereg.h

@@ -3,6 +3,7 @@
 #include "megbrain/serialization/opr_load_dump.h"
 #include "megbrain/serialization/opr_registry.h"
 #include "megbrain/serialization/opr_shallow_copy.h"
+#include "megbrain/serialization/oss_opr_load_dump.h"
 #include "megbrain/utils/hash_ct.h"
 
 namespace mgb {
@@ -66,6 +67,9 @@ struct OprLoadDumpImpl {
     }
 };
 
+template <class Opr, size_t arity>
+struct OprLoadDumpImplV2 : public OprLoadDumpImpl<Opr, arity> {};
+
 #define IMPL_OPR_MAKER(_arity, _args...)                                              \
     template <class Opr>                                                              \
     struct OprMaker<Opr, _arity> {                                                    \
@@ -124,6 +128,12 @@ struct OprRegistryCaller : public OprRegistryCallerDefaultImpl<Callee> {};
             __caller_OprReg##_cls##_ins;                                 \
     }
 
+#define MGB_SEREG_OPR_INTL_CALL_ENTRY_V2(_cls, _impl)                    \
+    namespace {                                                          \
+    [[gnu::unused]] ::mgb::serialization::OprRegistryCaller<_cls, _impl> \
+            __caller_V2_OprReg##_cls##_ins;                              \
+    }
+
 // Trim the terminating null character and a "V0" like suffix from the string
 // then hash it.
 // TODO: Get rid of this.
@@ -138,17 +148,35 @@ struct OprRegistryCaller : public OprRegistryCallerDefaultImpl<Callee> {};
                              : 0),                               \
             20160701)>::val
 
-//! call OprRegistry::add
-#define MGB_SEREG_OPR_INTL_CALL_ADD(_cls, _dump, _load)            \
-    do {                                                           \
-        ::mgb::serialization::OprRegistry::add(                    \
-                {_cls::typeinfo(),                                 \
-                 MGB_HASH_STR(#_cls),                              \
-                 _MGB_SEREG_OPR_NAME_FROM_CLS(_cls),               \
-                 _dump,                                            \
-                 _load,                                            \
-                 {},                                               \
-                 MGB_HASH_STR_WITHOUT_TAIL_0_AND_VERSION(#_cls)}); \
+//! call OprRegistry::add for old serialization
+//! call OprRegistryV2::versioned_add for new serialization which is compatiable
+//! with old serialization, convert is nullptr, this registry is just only for
+//! varsion 1
+#define MGB_SEREG_OPR_INTL_CALL_ADD(_cls, _dump, _load)                            \
+    do {                                                                           \
+        ::mgb::serialization::OprRegistry::add(                                    \
+                {_cls::typeinfo(),                                                 \
+                 MGB_HASH_STR(#_cls),                                              \
+                 _MGB_SEREG_OPR_NAME_FROM_CLS(_cls),                               \
+                 _dump,                                                            \
+                 _load,                                                            \
+                 {},                                                               \
+                 MGB_HASH_STR_WITHOUT_TAIL_0_AND_VERSION(#_cls)});                 \
+        ::mgb::serialization::OprRegistryV2::versioned_add(                        \
+                {_cls::typeinfo(), MGB_HASH_STR_WITHOUT_TAIL_0_AND_VERSION(#_cls), \
+                 _MGB_SEREG_OPR_NAME_FROM_CLS(_cls), _dump, _load, nullptr},       \
+                ::mgb::VERSION_1, ::mgb::VERSION_1);                               \
+    } while (0)
+
+//! call OprRegistryV2::versioned_add for new serialization, in which convert the
+//! function converter the Operator to the compatiable
+#define MGB_SEREG_OPR_INTL_CALL_ADD_V2(                                       \
+        _cls, _dump, _load, _convert, _version_min, _version_max)             \
+    do {                                                                      \
+        ::mgb::serialization::OprRegistryV2::versioned_add(                   \
+                {_cls::typeinfo(), MGB_HASH_STR(#_cls),                       \
+                 _MGB_SEREG_OPR_NAME_FROM_CLS(_cls), _dump, _load, _convert}, \
+                _version_min, _version_max);                                  \
     } while (0)
 
 /*!
@@ -171,6 +199,27 @@ struct OprRegistryCaller : public OprRegistryCallerDefaultImpl<Callee> {};
     }                                                                           \
     MGB_SEREG_OPR_INTL_CALL_ENTRY(_cls, _OprReg##_cls)
 
+//! new dump/load function should implement in OprLoadDumpImplV2, _converter is
+//! optional , if not implement pass nullptr
+#define MGB_SEREG_OPR_V2(_cls, _arity, _converter, _version_min, _version_max)  \
+    namespace {                                                                 \
+    namespace ser = ::mgb::serialization;                                       \
+    struct _OprRegV2##_cls {                                                    \
+        using Impl = ser::OprLoadDumpImplV2<_cls, _arity>;                      \
+        static ser::OprWithOutputAccessor wrap_loader(                          \
+                ser::OprLoadContext& ctx, const mgb::cg::VarNodeArray& inputs,  \
+                const mgb::cg::OperatorNodeConfig& config) {                    \
+            return ser::OprWithOutputAccessor(Impl::load(ctx, inputs, config)); \
+        }                                                                       \
+        static void entry() {                                                   \
+            MGB_SEREG_OPR_INTL_CALL_ADD_V2(                                     \
+                    _cls, Impl::dump, wrap_loader, _converter, _version_min,    \
+                    _version_max);                                              \
+        }                                                                       \
+    };                                                                          \
+    }                                                                           \
+    MGB_SEREG_OPR_INTL_CALL_ENTRY_V2(_cls, _OprRegV2##_cls)
+
 //! use to check type is complete or not, midout need a complete type
 template <class T, class = void>
 struct IsComplete : std::false_type {};