refactor(opdef): split apply_on_physical_tensor into infer_output_mem_desc and execute

GitOrigin-RevId: 4d62b7cbbd

imperative/src/impl/interpreter/interpreter_impl.cpp

@@ -124,6 +124,7 @@ Handle ChannelImpl::put(const HostTensorND& value, bool no_cache) {
 TensorInfo* ChannelImpl::put_impl(const HostTensorND& value, bool no_cache) {
     auto info = alloc();
     init(info, {value.layout(), value.comp_node(), value.proxy_to_default_cpu()});
+    info->mem_desc.id = StorageIdentifier::make(++m_storage_id);
     info->h_value = value;
     m_buffer.enqueue(Put{info, value, no_cache});
     if (m_async_level == 0) {
@@ -141,6 +142,7 @@ Handle ChannelImpl::put(const DeviceTensorND& data) {
     auto info = alloc();
     RECORD_EVENT(TensorCommandEvent, info->id, TensorCommandEvent::Put);
     init(info, {data.layout(), data.comp_node()});
+    info->mem_desc.id = StorageIdentifier::make(++m_storage_id);
     info->ptr = Tensor::make(data);
     RECORD_EVENT(TensorProduceEvent, info->id, info->desc.layout, info->desc.comp_node, data.raw_ptr());
     info->status = TensorInfo::Produced;
@@ -487,6 +489,9 @@ void ChannelImpl::init(TensorInfo* info, LogicalTensorDesc desc) {
     RECORD_EVENT(TensorDeclareEvent, info->id, info->name);
     info->status = TensorInfo::Allocated;
     info->desc = std::move(desc);
+    info->mem_desc.layout = info->desc.layout;
+    info->mem_desc.cn = info->desc.comp_node;
+    info->mem_desc.offset = 0;
 }
 
 
@@ -605,6 +610,7 @@ void ChannelImpl::do_apply_op(const ApplyOp& cmd) {
     bool profiling_device = Profiler::is_profiling() && Profiler::get_option("profile_device", 0);
     uint64_t apply_id = cmd.id;
     SmallVector<TensorPtr> tensor_inputs;
+    SmallVector<MemoryDesc> input_memory_desc;
     if (state.options.enable_dtr_auto_drop) {
         m_dtr.pin(cmd.inputs); 
     }
@@ -618,8 +624,27 @@ void ChannelImpl::do_apply_op(const ApplyOp& cmd) {
     // refcnt == 1, owners: [TensorInfo::ptr]
     for (auto i : cmd.inputs) {
         mgb_assert(i->ptr, "Invalid input tensor ptr!");
+        mgb_assert(i->mem_desc.id, "Invalid input tensor mem desc!");
         // refcnt ++, owners: [i->ptr, tensor_inputs]
         tensor_inputs.push_back(i->ptr);
+        input_memory_desc.push_back(i->mem_desc);
+    }
+    // SmallVector<MemoryDesc> outputs_mem_desc;
+    // SmallVector<TensorPtr> tensor_outputs, workspaces;
+    auto [outputs_mem_desc, tensor_outputs, workspaces] = init_output_and_workspace(*cmd.op, tensor_inputs, input_memory_desc);
+    if (outputs_mem_desc.size()) {
+        for (size_t i = 0;i < outputs_mem_desc.size();i ++) {
+            if (cmd.outputs[i]) {
+                cmd.outputs[i]->mem_desc = outputs_mem_desc[i];
+            }
+        }
+    } else {
+        // fail to infer mem plan
+        for (auto && out : cmd.outputs) {
+            if (out) {
+                out->mem_desc.id = StorageIdentifier::make();  
+            }
+        }
     }
     RECORD_EVENT(OpExecuteEvent, apply_id);
     // Begin profiling operator
@@ -662,8 +687,13 @@ void ChannelImpl::do_apply_op(const ApplyOp& cmd) {
     }
     // Apply op
     // Here std::move is REQUIRED for removing duplicated references.
-    auto tensor_outputs = OpDef::apply_on_physical_tensor(
-        *cmd.op, std::move(tensor_inputs));
+    if (outputs_mem_desc.size()) {
+        OpDef::execute(
+            *cmd.op, std::move(tensor_inputs), tensor_outputs, std::move(workspaces));
+    } else {
+        tensor_outputs = OpDef::apply_on_physical_tensor(
+            *cmd.op, std::move(tensor_inputs));
+    }
     // After execute
     for (auto&& [device, kernel_id]: kernels) {
         RECORD_EVENT(KernelExecuteFinishEvent, apply_id, kernel_id, Timer::record_event(device));
@@ -705,7 +735,7 @@ void ChannelImpl::do_apply_op(const ApplyOp& cmd) {
     RECORD_EVENT(OpExecuteFinishEvent, apply_id);
     // End profiling operator
 }
-
+        
 void ChannelImpl::recompute(TensorInfo::ComputePath* path) {
     auto& state = get_worker_state();
     do_apply_op(ApplyOp{path->id, path->op, path->inputs, path->outputs, {}});
@@ -829,6 +859,47 @@ std::unordered_set<TensorInfo*> ChannelImpl::collect_valid_tensors() {
     return valid_tensors;
 }
 
+std::tuple<SmallVector<MemoryDesc>, SmallVector<TensorPtr>, SmallVector<TensorPtr>> ChannelImpl::init_output_and_workspace(
+        const OpDef& def,
+        SmallVector<TensorPtr> inputs,
+        SmallVector<MemoryDesc> inputs_mem_desc) {
+
+    auto [outputs_desc, workspaces_desc] = OpDef::infer_output_mem_desc(def, inputs, inputs_mem_desc);
+    if (!outputs_desc.size()) {
+        // failed to infer memplan
+        return {{}, {}, {}};
+    }
+    // refine storage id to make it unique
+    for (auto&& desc : outputs_desc) {
+        if (desc.id->is_sys_alloc()) {
+            // TODO: there may be some outputs sharing the same storage id
+            desc.id->id = ++ m_storage_id;
+        }
+    }
+    auto alloc_storage = [&](SmallVector<MemoryDesc>& desc) {
+        SmallVector<TensorPtr> tensors;
+        for (size_t i = 0; i < desc.size(); i ++) {
+            if (desc[i].id->is_sys_alloc()) {
+                tensors.push_back(Tensor::make(desc[i].layout, desc[i].cn));
+            } else if (desc[i].id->is_from_other()) {
+                for (size_t j = 0; j < inputs_mem_desc.size();j ++) {
+                    if (inputs_mem_desc[j].id->desc == desc[i].id->desc) {
+                        tensors.push_back(inputs[j]->sub(desc[i].offset, desc[i].layout));
+                        break;
+                    }
+                }
+            } else if (desc[i].id->is_device_ptr()) {
+                tensors.push_back(desc[i].id->ptr);
+            } else {
+                mgb_assert(0, "not implemented");
+            }
+        }
+        return tensors;
+    };
+    
+    return {outputs_desc, alloc_storage(outputs_desc), alloc_storage(workspaces_desc)};
+}
+
 void ChannelImpl::process_one_task(IdentifiedCommand& icmd) {
     using namespace ranges;
     using namespace ranges::views;

imperative/src/impl/interpreter/interpreter_impl.h

@@ -101,6 +101,11 @@ private:
     void regenerate(TensorInfo* dest);
     void recompute(TensorInfo::ComputePath* path);
     void do_apply_op(const ApplyOp& cmd);
+    
+    std::tuple<SmallVector<MemoryDesc>, SmallVector<TensorPtr>, SmallVector<TensorPtr>> init_output_and_workspace(
+        const OpDef& def,
+        SmallVector<TensorPtr> inputs,
+        SmallVector<MemoryDesc> inputs_mem_desc);
 
     void dispatch_default_cpu(
         std::shared_ptr<OpDef> op,
@@ -139,6 +144,7 @@ private:
     uint64_t m_waitee_id = 0;
     std::exception_ptr m_worker_exc;
     std::function<void(std::string, std::string)> m_profile_dump_callback;
+    size_t m_storage_id = 0;
 
     bool m_closed = false;
 

imperative/src/impl/interpreter/tensor_info.h

@@ -58,6 +58,7 @@ struct TensorInfo {
     // Lock interpreter when visiting `ptr`.
     TensorPtr ptr;
     LogicalTensorDesc desc;
+    MemoryDesc mem_desc;
 
     double compute_time;
     size_t memory;

imperative/src/impl/op_def.cpp

@@ -45,6 +45,21 @@ SmallVector<TensorPtr> OpDef::apply_on_physical_tensor(
     return def.trait()->apply_on_physical_tensor(def, std::move(inputs));
 }
 
+std::tuple<SmallVector<MemoryDesc>, SmallVector<MemoryDesc>> OpDef::infer_output_mem_desc(
+    const OpDef& def,
+    const SmallVector<TensorPtr>& inputs_tensors,
+    const SmallVector<MemoryDesc>& inputs_mems) {
+    return def.trait()->infer_output_mem_desc(def, inputs_tensors, inputs_mems);
+}
+
+void OpDef::execute(
+    const OpDef& def,
+    SmallVector<TensorPtr> inputs,
+    SmallVector<TensorPtr> outputs,
+    SmallVector<TensorPtr> workspace) {
+    def.trait()->execute(def, std::move(inputs), outputs, std::move(workspace));
+}
+
 void OpDef::apply_on_device_tensornd(
     const OpDef& def,
     const SmallVector<DeviceTensorND>& inputs,

imperative/src/impl/op_trait.cpp

@@ -79,6 +79,13 @@ OpTraitRegistry& OpTraitRegistry::fallback() {
             trait->apply_on_physical_tensor =
                     proxy_graph_detail::apply_on_physical_tensor;
         }
+        if (!trait->execute) {
+            trait->execute = proxy_graph_detail::execute;
+        }
+        if (!trait->infer_output_mem_desc) {
+            trait->infer_output_mem_desc =
+                    proxy_graph_detail::infer_output_mem_desc;
+        }
         if (!trait->infer_output_attrs_fallible) {
             trait->infer_output_attrs_fallible =
                     proxy_graph_detail::infer_output_attrs_fallible;

imperative/src/impl/op_trait.h

@@ -64,6 +64,10 @@ using DecideDispatchMode = detail::OpMeth<
         decltype(OpDef::decide_dispatch_mode)>;
 using ApplyOnPhysicalTensor = detail::OpMeth<
         decltype(OpDef::apply_on_physical_tensor)>;
+using InferOutputMemDesc = detail::OpMeth<
+        decltype(OpDef::infer_output_mem_desc)>;
+using Execute = detail::OpMeth<
+        decltype(OpDef::execute)>;
 using ApplyOnDeviceTensorND = detail::OpMeth<
         decltype(OpDef::apply_on_device_tensornd)>;
 using ApplyOnVarNode = detail::OpMeth<
@@ -82,6 +86,8 @@ struct OpTrait {
     OpDefMaker make_from_op_node;
     DecideDispatchMode decide_dispatch_mode;
     ApplyOnPhysicalTensor apply_on_physical_tensor;
+    InferOutputMemDesc infer_output_mem_desc;
+    Execute execute;
     ApplyOnDeviceTensorND apply_on_device_tensornd;
     ApplyOnVarNode apply_on_var_node;
     InferOutputAttrsFallible infer_output_attrs_fallible;
@@ -100,6 +106,8 @@ struct OpTrait {
     cb(make_from_op_node) \
     cb(decide_dispatch_mode) \
     cb(apply_on_physical_tensor) \
+    cb(infer_output_mem_desc) \
+    cb(execute) \
     cb(apply_on_device_tensornd) \
     cb(apply_on_var_node) \
     cb(infer_output_attrs_fallible) \

imperative/src/impl/ops/cond_take.cpp

@@ -79,10 +79,24 @@ std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(
     }, false};
 }
 
+std::tuple<SmallVector<MemoryDesc>, SmallVector<MemoryDesc>> infer_output_mem_desc(
+        const OpDef& def,
+        const SmallVector<TensorPtr>& inputs,
+        const SmallVector<MemoryDesc>& inputs_mems) {
+    return {{}, {}};
+}
+
+void execute(const OpDef& def, const SmallVector<TensorPtr>& inputs,
+          const SmallVector<TensorPtr>& outputs, const SmallVector<TensorPtr>& workspace) {
+    mgb_assert(0);
+}
+
 OP_TRAIT_REG(CondTake, CondTake, opr::CondTake)
     .apply_on_var_node(apply_on_var_node)
     .apply_on_physical_tensor(apply_on_physical_tensor)
     .infer_output_attrs_fallible(infer_output_attrs_fallible)
+    .infer_output_mem_desc(infer_output_mem_desc)
+    .execute(execute)
     .fallback();
 
 } // namespace

imperative/src/impl/ops/elemwise.cpp

@@ -118,6 +118,35 @@ void apply_on_device_tensornd(
     opr::Elemwise::perform(op_def.mode, (*outputs)[0], inputs, dnn_opr);
 }
 
+void execute(
+        const OpDef& def,
+        SmallVector<TensorPtr> inputs,
+        SmallVector<TensorPtr> outputs,
+        SmallVector<TensorPtr> workspace) {
+    mgb_assert(outputs.size() == 1);
+    SmallVector<DeviceTensorND> inp_tensornds(inputs.size());
+    for (size_t i = 0;i < inputs.size(); ++i) {
+        inp_tensornds[i] = inputs[i]->dev_tensor();
+    }
+    SmallVector<DeviceTensorND> out_tensornds = {outputs[0]->dev_tensor()};
+    apply_on_device_tensornd(def, inp_tensornds, &out_tensornds);
+}
+
+std::tuple<SmallVector<MemoryDesc>, SmallVector<MemoryDesc>> infer_output_mem_desc(
+        const OpDef& def,
+        const SmallVector<TensorPtr>& inputs_tensors,
+        const SmallVector<MemoryDesc>& inputs_mems) {
+    auto&& op_def = def.cast_final_safe<Elemwise>();
+    TensorShapeArray inp_shapes(inputs_tensors.size());
+    for (size_t i = 0;i < inputs_tensors.size(); ++i) {
+        inp_shapes[i] = inputs_tensors[i]->layout();
+    }
+    TensorShape shape = opr::Elemwise::get_output_var_shape(op_def.mode, inp_shapes);
+    SmallVector<MemoryDesc> outputs = {{{shape, inputs_tensors[0]->dtype()}, 0, inputs_tensors[0]->comp_node(), StorageIdentifier::make(1)}};
+    return {outputs, {}};
+}
+
+
 SmallVector<TensorPtr> apply_on_physical_tensor(
         const OpDef& def,
         const SmallVector<TensorPtr>& inputs) {
@@ -224,7 +253,7 @@ cg::OperatorNodeBase* apply_inplace_add_on_var_node(
 SmallVector<TensorPtr> apply_inplace_add_on_physical_tensor(
         const OpDef& def,
         const SmallVector<TensorPtr>& inputs){
-    mgb_assert(inputs[0]->blob().unique() && inputs[0]->blob()->storage().unique(),
+    mgb_assert(inputs[0]->blob().use_count() == 2 && inputs[0]->blob()->storage().unique(),
             "This inplace modification may change the elements of other tensors. "
             "Please set MEGENGINE_INPLACE_UPDATE to 0 to ensure the program runs correctly.");
     auto dest = inputs[0], delta = inputs[1],
@@ -238,6 +267,23 @@ SmallVector<TensorPtr> apply_inplace_add_on_physical_tensor(
     return { std::make_shared<Tensor>(dest->blob(), dest->offset(), dest->layout()) };
 }
 
+void execute_inplace(
+        const OpDef& def,
+        SmallVector<TensorPtr> inputs,
+        SmallVector<TensorPtr> outputs,
+        SmallVector<TensorPtr> workspace) {
+    apply_inplace_add_on_physical_tensor(def, inputs);
+}
+
+std::tuple<SmallVector<MemoryDesc>, SmallVector<MemoryDesc>> infer_inplace_output_mem_desc(
+        const OpDef& def,
+        const SmallVector<TensorPtr>& inputs_tensors,
+        const SmallVector<MemoryDesc>& inputs_mems) {
+    auto dest = inputs_tensors[0];
+    SmallVector<MemoryDesc> outputs = {{dest->layout(), 0, dest->comp_node(), StorageIdentifier::make(&inputs_mems[0])}};
+    return {outputs, {}};
+}
+
 std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_inplace_add_output_attrs_fallible(
         const OpDef& def,
         const SmallVector<LogicalTensorDesc>& inputs) {
@@ -271,12 +317,16 @@ OP_TRAIT_REG(Elemwise, Elemwise, opr::Elemwise)
     .infer_output_attrs_fallible(infer_output_attrs_fallible)
     .apply_on_device_tensornd(apply_on_device_tensornd)
     .apply_on_physical_tensor(apply_on_physical_tensor)
+    .infer_output_mem_desc(infer_output_mem_desc)
+    .execute(execute)
     .fallback();
 
 OP_TRAIT_REG(InplaceAdd, InplaceAdd, opr::AddUpdate)
     .apply_on_var_node(apply_inplace_add_on_var_node)
     .apply_on_physical_tensor(apply_inplace_add_on_physical_tensor)
     .infer_output_attrs_fallible(infer_inplace_add_output_attrs_fallible)
+    .infer_output_mem_desc(infer_inplace_output_mem_desc)
+    .execute(execute_inplace)
     .fallback();
 } // anonymous namespace
 

imperative/src/impl/ops/rng.cpp

@@ -331,6 +331,7 @@ struct _RNGOprInvoker<DNN_NR_INPUTS> {
     }                                                                                                  \
 };
 
+
 #define _INST_RNG_MAKER(MGB_NR_INPUTS)                                                                 \
 template<>                                                                                             \
 struct _RNGOprMaker<MGB_NR_INPUTS> {                                                                   \
@@ -366,7 +367,7 @@ _INST_RNG_MAKER(2)
 
 template <typename Op>
 void exec(const OpDef& op, const SmallVector<TensorPtr>& inputs,
-          const SmallVector<TensorPtr>& outputs) {
+          const SmallVector<TensorPtr>& outputs, const SmallVector<TensorPtr>& workspace) {
     auto&& rng = op.cast_final_safe<Op>();
  
     auto dest = outputs[0];
@@ -419,6 +420,18 @@ SmallVector<LogicalTensorDesc> infer_output_attrs(
 }
 
 template <typename Op>
+std::tuple<SmallVector<MemoryDesc>, SmallVector<MemoryDesc>> infer_output_mem_desc(
+        const OpDef& def,
+        const SmallVector<TensorPtr>& inputs_tensors,
+        const SmallVector<MemoryDesc>& inputs_mems) {
+    auto &&dest = infer_output_attrs<Op>(def, inputs_tensors);
+    SmallVector<MemoryDesc> outputs = {{dest[0].layout, 0, dest[0].comp_node, StorageIdentifier::make(1)}};
+    
+    return {outputs, {}};    
+}
+
+
+template <typename Op>
 SmallVector<TensorPtr> apply_on_physical_tensor(
         const OpDef& def, const SmallVector<TensorPtr>& inputs) {
     SmallVector<TensorPtr> outputs;
@@ -427,10 +440,19 @@ SmallVector<TensorPtr> apply_on_physical_tensor(
     for (auto&& i : desc) {
         outputs.push_back(Tensor::make(i.layout, i.comp_node));
     }
-    exec<Op>(def, inputs, outputs);
+    exec<Op>(def, inputs, outputs, {});
     return outputs;
 }
 
+template <typename Op>
+void execute(
+        const OpDef& def,
+        SmallVector<TensorPtr> inputs,
+        SmallVector<TensorPtr> outputs,
+        SmallVector<TensorPtr> workspace) {
+    exec<Op>(def, inputs, outputs, {});
+}
+
 template<typename Op>
 SymbolVar apply_on_var_node(
         const OpDef& def,
@@ -492,6 +514,8 @@ OP_TRAIT_REG(NAME, NAME, OpMeth<NAME>::OpNode) \
     .apply_on_var_node(apply_on_var_node<NAME>) \
     .apply_on_physical_tensor(apply_on_physical_tensor<NAME>) \
     .infer_output_attrs_fallible(infer_output_attrs_fallible<NAME>) \
+    .infer_output_mem_desc(infer_output_mem_desc<NAME>) \
+    .execute(execute<NAME>) \
     .fallback(); \
 } \
 

imperative/src/impl/ops/tensor_manip.cpp

@@ -86,22 +86,22 @@ void apply_on_device_tensornd(
     (*outputs)[0] = DeviceTensorND::make_proxy(hv);
 }
 
-SmallVector<TensorPtr> apply_on_physical_tensor(
-        const OpDef& def,
-        const SmallVector<TensorPtr>& inputs) {
+HostTensorND get_var_shape_host_tensor(const OpDef& def, const SmallVector<TensorPtr>& inputs) {
     SmallVector<DeviceTensorND> input_tensornds;
     input_tensornds.reserve(inputs.size());
     for (auto&& inp : inputs) {
         input_tensornds.push_back(inp->dev_tensor());
     }
     SmallVector<DeviceTensorND> output_tensornds = {{CompNode::default_cpu(), dtype::Int32()}};
-
     apply_on_device_tensornd(def, input_tensornds, &output_tensornds);
-
     // restore to input comp_node
-    HostTensorND host_tensornd = HostTensorND::make_proxy(output_tensornds[0])
-        .proxy_to_comp_node(inputs[0]->comp_node());
-    return {Tensor::make(std::move(host_tensornd))};
+    return HostTensorND::make_proxy(output_tensornds[0]).proxy_to_comp_node(inputs[0]->comp_node());
+}
+
+SmallVector<TensorPtr> apply_on_physical_tensor(
+        const OpDef& def,
+        const SmallVector<TensorPtr>& inputs) {
+    return {Tensor::make(std::move(get_var_shape_host_tensor(def, inputs)))};
 }
 
 std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(
@@ -142,6 +142,33 @@ std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(
     return {{{value.layout(), desc.comp_node, std::move(value)}}, true};
 }
 
+std::tuple<SmallVector<MemoryDesc>, SmallVector<MemoryDesc>> infer_output_mem_desc(
+        const OpDef& def,
+        const SmallVector<TensorPtr>& inputs,
+        const SmallVector<MemoryDesc>& inputs_mems) {
+    HostTensorND tensor = get_var_shape_host_tensor(def, inputs);
+    SmallVector<MemoryDesc> ret;
+    auto&& blob = MultiCNConstTensorCache::inst().lookup(tensor);
+    if (blob) {
+        ret.push_back({tensor.layout(), 0, inputs[0]->comp_node(),
+                StorageIdentifier::make(Tensor::make(std::forward<decltype(blob)>(blob), tensor.layout(), tensor))});
+    } else {
+        ret.push_back({tensor.layout(), 0, inputs[0]->comp_node(), StorageIdentifier::make(1)});
+    }
+    return {ret, {}};
+}
+
+void execute(const OpDef& def, const SmallVector<TensorPtr>& inputs,
+          const SmallVector<TensorPtr>& outputs, const SmallVector<TensorPtr>& workspace) {
+    HostTensorND tensor = get_var_shape_host_tensor(def, inputs);
+    SmallVector<MemoryDesc> ret;
+    auto&& blob = MultiCNConstTensorCache::inst().lookup(tensor);
+    if (!blob || blob->storage() != outputs[0]->blob()->storage()) {
+        outputs[0]->dev_tensor().copy_from_fixlayout(tensor);
+        AsyncReleaser::inst()->add(tensor);
+    }
+}
+
 std::shared_ptr<OpDef> make_from_op_node(cg::OperatorNodeBase* node_) {
     auto* node = &node_->cast_final_safe<opr::GetVarShape>();
     return GetVarShape::make(node->param());
@@ -154,6 +181,8 @@ OP_TRAIT_REG(GetVarShape, GetVarShape, opr::GetVarShape)
     .apply_on_var_node(apply_on_var_node)
     .apply_on_device_tensornd(apply_on_device_tensornd)
     .apply_on_physical_tensor(apply_on_physical_tensor)
+    .infer_output_mem_desc(infer_output_mem_desc)
+    .execute(execute)
     .fallback();
 } // get_var_shape
 
@@ -181,6 +210,31 @@ cg::OperatorNodeBase* param_pack_split_apply_on_var_node(
     return opr;
 }
 
+std::tuple<SmallVector<MemoryDesc>, SmallVector<MemoryDesc>> param_pack_split_infer_output_mem_desc(
+        const OpDef& def,
+        const SmallVector<TensorPtr>& inputs,
+        const SmallVector<MemoryDesc>& inputs_mems) {
+    auto&& param = def.cast_final_safe<ParamPackSplit>();
+    mgb_assert(inputs.size() == 1, "ParamPackSplit take 1 input, got %lu", inputs.size());
+    auto&& inp = inputs[0];
+    auto&& shp = inp->layout();
+    mgb_assert(shp.ndim == 1, "ParamPackSplit input shape invalid, ndim should be 1");
+    mgb_assert(param.shapes.size() * 2 == param.offsets.size());
+    SmallVector<MemoryDesc> ret;
+    auto&& shapes = get_shapes(param.shapes);
+    size_t dtype_size = inputs[0]->layout().dtype.size();
+    for (size_t i = 0; i < shapes.size(); ++i) {
+        // memory forward
+        ret.push_back({{shapes[i], inputs[0]->dtype()}, param.offsets[i * 2] * dtype_size, inp->comp_node(), StorageIdentifier::make(&inputs_mems[0])});
+    }
+    return {ret, {}};
+}
+
+void param_pack_split_execute(const OpDef& def, const SmallVector<TensorPtr>& inputs,
+          const SmallVector<TensorPtr>& outputs, const SmallVector<TensorPtr>& workspace) {
+    // do nothing
+}
+
 SmallVector<TensorPtr> param_pack_split_apply_on_physical_tensor(
         const OpDef& def,
         const SmallVector<TensorPtr>& inputs) {
@@ -203,6 +257,8 @@ SmallVector<TensorPtr> param_pack_split_apply_on_physical_tensor(
 
 OP_TRAIT_REG(ParamPackSplit, ParamPackSplit, mgb::opr::ParamPackSplit)
         .apply_on_var_node(param_pack_split_apply_on_var_node)
+        .infer_output_mem_desc(param_pack_split_infer_output_mem_desc)
+        .execute(param_pack_split_execute)
         .apply_on_physical_tensor(param_pack_split_apply_on_physical_tensor)
         .fallback();
 
@@ -219,6 +275,64 @@ cg::OperatorNodeBase* param_pack_concat_apply_on_var_node(
     return opr;
 }
 
+
+std::tuple<SmallVector<MemoryDesc>, SmallVector<MemoryDesc>> param_pack_concat_infer_output_mem_desc(
+        const OpDef& def,
+        const SmallVector<TensorPtr>& inputs,
+        const SmallVector<MemoryDesc>& inputs_mems) {
+    def.cast_final_safe<ParamPackConcat>();
+    mgb_assert(inputs.size() > 1, "param_pack should have at least one input");
+    auto comp_node = inputs.front()->comp_node();
+    auto dtype = inputs.front()->dtype();
+    size_t nr_inputs = inputs.size() - 1;
+    size_t nr_elems = 0;
+    for (size_t i = 0; i < nr_inputs; ++i) {
+        auto& input = inputs[i];
+        mgb_assert(comp_node == input->comp_node(), "inputs for param_pack_concat must in same comp_node");
+        mgb_assert(dtype == input->dtype(), "inputs for param_pack_concat must have same dtype");
+        nr_elems += input->layout().total_nr_elems();
+    }
+    auto dest_layout = TensorLayout({nr_elems}, dtype);
+    auto caller = DnnOprCaller<megdnn::ParamPackConcat>(comp_node);
+    size_t ws_size;
+    {
+        TensorShapeArray src_shapes;
+        for (size_t i = 0; i < nr_inputs; ++i) {
+            src_shapes.push_back(inputs[i]->shape());
+        }
+        ws_size = caller.op->get_workspace_in_bytes(src_shapes, inputs.back()->shape(), TensorShape{});
+    }
+
+    SmallVector<MemoryDesc> outputs = {{dest_layout, 0, comp_node, StorageIdentifier::make(1)}};
+    MemoryDesc workspace = {{{ws_size}, dtype::Byte()}, 0, comp_node, StorageIdentifier::make(2)};
+
+    return {outputs, {workspace}};
+}
+
+void param_pack_concat_execute(const OpDef& def, const SmallVector<TensorPtr>& inputs,
+          const SmallVector<TensorPtr>& outputs, const SmallVector<TensorPtr>& workspace) {
+    def.cast_final_safe<ParamPackConcat>();
+    mgb_assert(inputs.size() > 1, "param_pack should have at least one input");
+    auto comp_node = inputs.front()->comp_node();
+    size_t nr_inputs = inputs.size() - 1;
+    auto caller = DnnOprCaller<megdnn::ParamPackConcat>(comp_node);
+    size_t srcs_size = sizeof(void*)*nr_inputs;
+    void** srcs_raw_ptr = (void**)comp_node.alloc_host(srcs_size);
+    std::shared_ptr<dt_byte> srcs_ptr = {(dt_byte*)srcs_raw_ptr, [comp_node](dt_byte* ptr){
+        comp_node.free_host(ptr);
+    }};
+    TensorLayout srcs_layout = TensorLayout{{nr_inputs}, dtype::Int32()};
+    for (size_t i = 0; i < nr_inputs; ++i) {
+        srcs_raw_ptr[i] = inputs[i]->dev_tensor().as_megdnn().raw_ptr;
+    }
+    HostTensorStorage srcs_storage;
+    srcs_storage.reset(comp_node, srcs_size, srcs_ptr);
+    megdnn::Workspace dnn_wk(workspace[0]->blob()->storage().get(), workspace[0]->blob()->size());
+    caller.op->exec({srcs_raw_ptr, srcs_layout}, inputs.back()->dev_tensor().as_megdnn(), outputs[0]->dev_tensor().as_megdnn(),
+            dnn_wk);
+    AsyncReleaser::inst()->add(HostTensorND{comp_node, srcs_layout}.storage(srcs_storage));
+}
+
 SmallVector<TensorPtr> param_pack_concat_apply_on_physical_tensor(
         const OpDef& def,
         const SmallVector<TensorPtr>& inputs) {
@@ -264,6 +378,8 @@ SmallVector<TensorPtr> param_pack_concat_apply_on_physical_tensor(
 
 OP_TRAIT_REG(ParamPackConcat, ParamPackConcat, mgb::opr::ParamPackConcat)
         .apply_on_var_node(param_pack_concat_apply_on_var_node)
+        .infer_output_mem_desc(param_pack_concat_infer_output_mem_desc)
+        .execute(param_pack_concat_execute)
         .apply_on_physical_tensor(param_pack_concat_apply_on_physical_tensor)
         .fallback();
 } // param_pack

imperative/src/impl/physical_tensor.cpp

@@ -77,149 +77,6 @@ public:
 bool CompNodeSyncManager::is_into_atexit = false;
 #endif
 
-// Cache for small blobs
-// 1. A blob has to be seen twice (within a window) to be eligible for cache
-// 2. Cache eviction occurs when cache size reaches a threshold, in least frequently used order
-class ConstTensorCache {
-public:
-    struct Entry {
-        size_t hitcnt = 0;
-        std::unique_ptr<dt_byte[]> data;
-        size_t size;
-        BlobPtr blob;
-
-        Entry() = default;
-        Entry(const dt_byte* ptr, size_t size_, BlobPtr blob_)
-                : data(new dt_byte[size_]), size(size_), blob(blob_) {
-            memcpy(data.get(), ptr, size);
-        }
-
-        // does not check input
-        bool match(const HostTensorND& hv) {
-            return 0 == memcmp(data.get(), hv.raw_ptr(), hv.layout().span().high_byte);
-        }
-    };
-
-    using KV = std::pair<uint64_t, Entry>;
-
-    bool check(const HostTensorND& hv) {
-        auto&& layout = hv.layout();
-        auto&& span = layout.span();
-        return hv.format().is_default() && !hv.empty() &&
-            layout.is_contiguous() && span.low_byte == 0 &&
-            span.high_byte <= max_bytes;
-    }
-
-    // hash storage; does not check input
-    static uint64_t hash(const HostTensorND& hv) {
-        auto&& span = hv.layout().span();
-        return XXHash{}
-            .update(hv.raw_ptr(), span.high_byte)
-            .digest();
-    }
-
-    BlobPtr lookup(const HostTensorND& hv) {
-        if (!check(hv)) {
-            return {};
-        }
-        auto h = hash(hv);
-        MGB_LOCK_GUARD(mtx);
-        // lookup in g1
-        auto it = g1.find(h);
-        if (it != g1.end()) {
-            if (!it->second.match(hv)) {
-                mgb_log_warn("hash collision in const tensor cache");
-                return {};
-            }
-            it->second.hitcnt += 1;
-            return it->second.blob;
-        }
-        // lookup in g0
-        if (!g0.extract(h) && !g0b.extract(h)) {
-            maybe_collect_g0();
-            g0.emplace(h);
-            return {};
-        }
-        // add new entry to g1
-        maybe_collect_g1();
-        Entry entry(hv.raw_ptr(), hv.layout().span().high_byte, Tensor(hv).blob());
-        it = g1.emplace_hint(it, h, std::move(entry));
-        it->second.hitcnt += 1;
-        return it->second.blob;
-    }
-
-    void clear() {
-        MGB_LOCK_GUARD(mtx);
-        g0.clear();
-        g0b.clear();
-        g1.clear();
-    }
-
-    std::mutex mtx;
-    const size_t hwm = 1024, lwm = 512, max_bytes = TensorShape::MAX_NDIM * 8, window = 65536;
-
-private:
-    void maybe_collect_g0() {
-        if (g0.size() > window) {
-            std::swap(g0, g0b);
-            g0.clear();
-        }
-    }
-    void maybe_collect_g1() {
-        if (g1.size() < hwm) return;
-
-        tmp.clear();
-        for (auto&& kv : g1) {
-            tmp.emplace_back(kv.first, std::move(kv.second));
-        }
-        std::nth_element(tmp.begin(), tmp.begin() + lwm, tmp.end(), [](const KV& lhs, const KV& rhs) {
-                return lhs.second.hitcnt > rhs.second.hitcnt;
-            });
-        tmp.resize(lwm);
-        g1.clear();
-        for (auto&& kv : tmp) {
-            kv.second.hitcnt = 0;
-            g1.emplace(std::move(kv));
-        }
-    }
-
-    // g0: records blobs which have been seen at least once (within a window)
-    // g0b: backup of g0
-    // g1: records the most frequently used blobs which have been seen at least
-    // twice. When `g1.size() == hwm`, it will be refreshed and only the top
-    // `lhw` frequently used blobs will be kept.
-    std::unordered_set<uint64_t> g0, g0b;
-    std::unordered_map<uint64_t, Entry> g1;
-    std::vector<KV> tmp;
-
-public:
-    ConstTensorCache() {
-        g0.reserve(window), g0b.reserve(window);
-        g1.reserve(hwm), tmp.reserve(hwm);
-    }
-};
-
-struct MultiCNConstTensorCache : CompNodeDepedentObject {
-    std::mutex mtx;
-    CompNode::UnorderedMap<ConstTensorCache> cn2cache;
-
-    std::shared_ptr<void> on_comp_node_finalize() {
-        MGB_LOCK_GUARD(mtx);
-        cn2cache.clear();
-        return {};
-    }
-
-    BlobPtr lookup(const HostTensorND& hv) {
-        MGB_LOCK_GUARD(mtx);
-        return cn2cache[hv.comp_node()].lookup(hv);
-    }
-
-    static MultiCNConstTensorCache& inst() {
-        static MultiCNConstTensorCache sl_inst;
-        return sl_inst;
-    }
-};
-
 }  // namespace
 
 void EventDeleter::operator()(CompNode::Event* event) {

imperative/src/impl/proxy_graph.cpp

@@ -522,9 +522,10 @@ SmallVector<LogicalTensorDesc> ProxyGraph::infer_output_attrs(
 
 void ProxyGraph::invoke_op(const OpDef& opdef,
         const SmallVector<Tensor*>& inputs,
-        const SmallVector<Tensor*>& outputs) {
+        const SmallVector<Tensor*>& outputs,
+        const SmallVector<Tensor*>& workspaces) {
     CUR_OPR_GUARD(get_proxy_opr(opdef, inputs));
-    init_output_tensor(outputs);
+    init_output_tensor(outputs, workspaces);
     for (auto oup : m_cur_opr->output()) {
         m_graph->add_used_comp_node(oup->comp_node());
     }
@@ -544,19 +545,30 @@ void ProxyGraph::cleanup() {
     m_cur_opr = nullptr;
 }
 
-void ProxyGraph::init_output_tensor(const SmallVector<Tensor*>& outputs) {
+void ProxyGraph::init_output_tensor(const SmallVector<Tensor*>& outputs, const SmallVector<Tensor*>& workspaces) {
     // get proxy opr
     auto proxy = m_cur_opr;
 
     do_shape_infer(true);
 
     size_t j = 0;
+    size_t k = 0;
     for (auto&& var : proxy->output()) {
         auto &&chk = var->m_mem_plan.reset_from_owner_var().chunk();
         if (var->contain_flag(VarNode::Flag::VOLATILE_CONTENT)) {
-            // alloc workspace
-            TensorLayout layout{var->shape(), var->dtype(), var->format()};
-            var->m_dev_tensor = BlobManager::inst()->alloc_workspace_with_defrag(var->comp_node(), layout);
+            // workspace
+            if (workspaces.size()) {
+                mgb_assert(k < workspaces.size());
+                auto && layout = workspaces[k]->layout();
+                mgb_assert(var->comp_node() == workspaces[k]->comp_node() &&
+                            var->shape().eq_shape(layout) &&
+                            var->dtype() == layout.dtype);
+                var->m_dev_tensor = workspaces[k]->dev_tensor();
+                ++ k;
+            } else {
+                TensorLayout layout{var->shape(), var->dtype(), var->format()};
+                var->m_dev_tensor = BlobManager::inst()->alloc_workspace_with_defrag(var->comp_node(), layout);
+            }
         } else {
             mgb_assert(j < outputs.size());
             auto &&tensor = outputs[j];
@@ -570,6 +582,7 @@ void ProxyGraph::init_output_tensor(const SmallVector<Tensor*>& outputs) {
         chk.mem_alloc_status.set_from_owner_var();
     }
     mgb_assert(j == outputs.size());
+    mgb_assert(k == workspaces.size());
 
     // Memory forwarding was bypassed in megbrain with graph option
     // imerative_proxy_graph on, here we call mem_plan_fwd_in2out_readonly
@@ -623,6 +636,26 @@ std::tuple<SmallVector<LogicalTensorDesc>, bool> ProxyGraph::infer_output_attrs_
     return {outputs, validated && !need_check};
 }
 
+std::tuple<SmallVector<MemoryDesc>, SmallVector<MemoryDesc>> ProxyGraph::infer_output_mem_desc(
+        const OpDef& def,
+        const SmallVector<Tensor*>& inputs_tensors,
+        const SmallVector<MemoryDesc>& inputs_mems) {
+    auto opr = get_proxy_opr(def, inputs_tensors);
+    CUR_OPR_GUARD(opr);
+    do_shape_infer(true);
+    SmallVector<MemoryDesc> outputs;
+    SmallVector<MemoryDesc> workspaces;
+    size_t cur_id = 0;
+    for (auto&& i : opr->output()) {
+        if (i->contain_flag(VarNode::Flag::VOLATILE_CONTENT)) {
+            workspaces.push_back({{i->shape(), i->dtype(), i->format()}, 0, i->comp_node(), StorageIdentifier::make(++ cur_id)});
+        } else {
+            outputs.push_back({{i->shape(), i->dtype()}, 0, i->comp_node(), StorageIdentifier::make(++ cur_id)});
+        }
+    }
+    return {outputs, workspaces};
+}
+
 struct ProxyGraph::GradGraph {
     cg::VarNodeArray inputs;
     cg::VarNodeArray outputs;

imperative/src/impl/proxy_graph.h

@@ -37,7 +37,8 @@ public:
     void invoke_op(
             const OpDef& opdef,
             const SmallVector<Tensor*>& inputs,
-            const SmallVector<Tensor*>& outputs);
+            const SmallVector<Tensor*>& outputs,
+            const SmallVector<Tensor*>& workspace);
 
     BackwardGraphResult make_backward_graph(
             const OpDef& opdef,
@@ -45,6 +46,11 @@ public:
             const SmallVector<bool>& input_requires_grad,
             const SmallVector<bool>& output_has_grad);
 
+    std::tuple<SmallVector<MemoryDesc>, SmallVector<MemoryDesc>> infer_output_mem_desc(
+            const OpDef& def,
+            const SmallVector<Tensor*>& inputs_tensors,
+            const SmallVector<MemoryDesc>& inputs_mems);
+
     /********************** Logical Tensor API **********************/
 
     size_t get_opr_output_size(
@@ -74,7 +80,8 @@ private:
     void cleanup();
 
     void init_output_tensor(
-            const SmallVector<Tensor*>& outputs);
+            const SmallVector<Tensor*>& outputs,
+            const SmallVector<Tensor*>& workspace);
 
     cg::OperatorNodeBase* get_proxy_opr(
             const OpDef& opdef,

imperative/src/impl/proxy_graph_detail.cpp

@@ -43,10 +43,12 @@ infer_output_attrs(const OpDef& def,
 
 void exec(const OpDef& def,
         const SmallVector<TensorPtr>& inputs,
-        const SmallVector<TensorPtr>& outputs) {
+        const SmallVector<TensorPtr>& outputs,
+        const SmallVector<TensorPtr>& workspaces) {
     auto&& graph = ProxyGraph::get_default_graph();
     auto raw_inputs = to_raw_ptr_array(inputs),
-         raw_outputs = to_raw_ptr_array(outputs);
+         raw_outputs = to_raw_ptr_array(outputs),
+         raw_workspaces = to_raw_ptr_array(workspaces);
     CompNode::UnorderedSet used_cns;
     for (auto&& out: raw_outputs) {
         auto cn = out->comp_node();
@@ -59,7 +61,7 @@ void exec(const OpDef& def,
             }
         }
     }
-    graph->invoke_op(def, raw_inputs, raw_outputs);
+    graph->invoke_op(def, raw_inputs, raw_outputs, raw_workspaces);
     for (auto&& cn: used_cns) {
         for (auto&& in: inputs) {
             if (in->comp_node() != cn) {
@@ -77,7 +79,7 @@ apply_on_physical_tensor(const OpDef& def,
     for (size_t i = 0; i < outputs.size(); i++) {
         outputs[i] = Tensor::make(output_descs[i].layout, output_descs[i].comp_node);
     }
-    exec(def, inputs, outputs);
+    exec(def, inputs, outputs, {});
     auto async_error = ProxyGraph::get_async_error();
     if (async_error) {
         throw *async_error;
@@ -85,6 +87,26 @@ apply_on_physical_tensor(const OpDef& def,
     return outputs;
 }
 
+std::tuple<SmallVector<MemoryDesc>, SmallVector<MemoryDesc>> infer_output_mem_desc(
+    const OpDef& def,
+    const SmallVector<TensorPtr>& inputs_tensors,
+    const SmallVector<MemoryDesc>& inputs_mems) {
+    auto&& graph = ProxyGraph::get_default_graph();
+    return graph->infer_output_mem_desc(def, to_raw_ptr_array(inputs_tensors), inputs_mems);
+}
+
+void execute(const OpDef& def,
+        SmallVector<TensorPtr> inputs,
+        SmallVector<TensorPtr> outputs,
+        SmallVector<TensorPtr> workspace) {
+    exec(def, inputs, outputs, workspace);
+    auto async_error = ProxyGraph::get_async_error();
+    if (async_error) {
+        throw *async_error;
+    }
+    return;
+}
+
 // std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(const OpDef& def,
 //         const SmallVector<LogicalTensorDesc>& inputs) {
 //     auto&& graph = ProxyGraph::get_default_graph();

imperative/src/include/megbrain/imperative/op_def.h

@@ -108,6 +108,13 @@ public:
     static SmallVector<TensorPtr> apply_on_physical_tensor(
         const OpDef& def,
         SmallVector<TensorPtr> inputs);
+    
+    static void execute(
+        const OpDef& def,
+        SmallVector<TensorPtr> inputs,
+        SmallVector<TensorPtr> outputs,
+        SmallVector<TensorPtr> workspace);
+
 
     /*!
      * \brief Call the corresponding dnn op to calculate results. Output
@@ -126,6 +133,11 @@ public:
         const OpDef& def,
         const SmallVector<LogicalTensorDesc>& inputs);
 
+    static std::tuple<SmallVector<MemoryDesc>, SmallVector<MemoryDesc>> infer_output_mem_desc(
+        const OpDef& def,
+        const SmallVector<TensorPtr>& inputs_tensors,
+        const SmallVector<MemoryDesc>& inputs_mems);
+
     static BackwardGraphResult make_backward_graph(
         const OpDef& def,
         const SmallVector<LogicalTensorDesc>& inputs,

imperative/src/include/megbrain/imperative/physical_tensor.h

@@ -150,12 +150,192 @@ private:
     EventPtr m_value_ready = nullptr;
 };
 
+// Cache for small blobs
+// 1. A blob has to be seen twice (within a window) to be eligible for cache
+// 2. Cache eviction occurs when cache size reaches a threshold, in least frequently used order
+class ConstTensorCache {
+public:
+    struct Entry {
+        size_t hitcnt = 0;
+        std::unique_ptr<dt_byte[]> data;
+        size_t size;
+        BlobPtr blob;
+
+        Entry() = default;
+        Entry(const dt_byte* ptr, size_t size_, BlobPtr blob_)
+                : data(new dt_byte[size_]), size(size_), blob(blob_) {
+            memcpy(data.get(), ptr, size);
+        }
+
+        // does not check input
+        bool match(const HostTensorND& hv) {
+            return 0 == memcmp(data.get(), hv.raw_ptr(), hv.layout().span().high_byte);
+        }
+    };
+
+    using KV = std::pair<uint64_t, Entry>;
+
+    bool check(const HostTensorND& hv) {
+        auto&& layout = hv.layout();
+        auto&& span = layout.span();
+        return hv.format().is_default() && !hv.empty() &&
+            layout.is_contiguous() && span.low_byte == 0 &&
+            span.high_byte <= max_bytes;
+    }
+
+    // hash storage; does not check input
+    static uint64_t hash(const HostTensorND& hv) {
+        auto&& span = hv.layout().span();
+        return XXHash{}
+            .update(hv.raw_ptr(), span.high_byte)
+            .digest();
+    }
+
+    BlobPtr lookup(const HostTensorND& hv) {
+        if (!check(hv)) {
+            return {};
+        }
+        auto h = hash(hv);
+        MGB_LOCK_GUARD(mtx);
+        // lookup in g1
+        auto it = g1.find(h);
+        if (it != g1.end()) {
+            if (!it->second.match(hv)) {
+                mgb_log_warn("hash collision in const tensor cache");
+                return {};
+            }
+            it->second.hitcnt += 1;
+            return it->second.blob;
+        }
+        // lookup in g0
+        if (!g0.extract(h) && !g0b.extract(h)) {
+            maybe_collect_g0();
+            g0.emplace(h);
+            return {};
+        }
+        // add new entry to g1
+        maybe_collect_g1();
+        Entry entry(hv.raw_ptr(), hv.layout().span().high_byte, Tensor(hv).blob());
+        it = g1.emplace_hint(it, h, std::move(entry));
+        it->second.hitcnt += 1;
+        return it->second.blob;
+    }
+
+    void clear() {
+        MGB_LOCK_GUARD(mtx);
+        g0.clear();
+        g0b.clear();
+        g1.clear();
+    }
+
+    std::mutex mtx;
+    const size_t hwm = 1024, lwm = 512, max_bytes = TensorShape::MAX_NDIM * 8, window = 65536;
+
+private:
+    void maybe_collect_g0() {
+        if (g0.size() > window) {
+            std::swap(g0, g0b);
+            g0.clear();
+        }
+    }
+    void maybe_collect_g1() {
+        if (g1.size() < hwm) return;
+
+        tmp.clear();
+        for (auto&& kv : g1) {
+            tmp.emplace_back(kv.first, std::move(kv.second));
+        }
+        std::nth_element(tmp.begin(), tmp.begin() + lwm, tmp.end(), [](const KV& lhs, const KV& rhs) {
+                return lhs.second.hitcnt > rhs.second.hitcnt;
+            });
+        tmp.resize(lwm);
+        g1.clear();
+        for (auto&& kv : tmp) {
+            kv.second.hitcnt = 0;
+            g1.emplace(std::move(kv));
+        }
+    }
+
+    // g0: records blobs which have been seen at least once (within a window)
+    // g0b: backup of g0
+    // g1: records the most frequently used blobs which have been seen at least
+    // twice. When `g1.size() == hwm`, it will be refreshed and only the top
+    // `lhw` frequently used blobs will be kept.
+    std::unordered_set<uint64_t> g0, g0b;
+    std::unordered_map<uint64_t, Entry> g1;
+    std::vector<KV> tmp;
+
+public:
+    ConstTensorCache() {
+        g0.reserve(window), g0b.reserve(window);
+        g1.reserve(hwm), tmp.reserve(hwm);
+    }
+};
+
+struct MultiCNConstTensorCache : CompNodeDepedentObject {
+    std::mutex mtx;
+    CompNode::UnorderedMap<ConstTensorCache> cn2cache;
+
+    std::shared_ptr<void> on_comp_node_finalize() {
+        MGB_LOCK_GUARD(mtx);
+        cn2cache.clear();
+        return {};
+    }
+
+    BlobPtr lookup(const HostTensorND& hv) {
+        MGB_LOCK_GUARD(mtx);
+        return cn2cache[hv.comp_node()].lookup(hv);
+    }
+
+    static MultiCNConstTensorCache& inst() {
+        static MultiCNConstTensorCache sl_inst;
+        return sl_inst;
+    }
+};
+
 struct LogicalTensorDesc {
     TensorLayout layout;
     CompNode comp_node;
     DeviceTensorND value; // cpu:default
 };
 
+struct StorageIdentifier;
+struct MemoryDesc {
+    TensorLayout layout;
+    size_t offset;
+    CompNode cn;
+    std::shared_ptr<StorageIdentifier> id;
+};
+
+struct StorageIdentifier {
+    enum { INVALID, SYS_ALLOC, FROM_OTHER, DEVICE_PTR } tag;
+    union {
+        size_t id;
+        MemoryDesc* desc;
+    };
+    TensorPtr ptr;
+    StorageIdentifier() = default;
+    StorageIdentifier(size_t id): tag(SYS_ALLOC), id(id) {}
+    StorageIdentifier(const MemoryDesc* desc): tag(FROM_OTHER), desc(desc->id->desc) {}
+    StorageIdentifier(TensorPtr dev_ptr): tag(DEVICE_PTR), ptr(dev_ptr) {}
+
+    template<typename ...Args>
+    static std::shared_ptr<StorageIdentifier> make(Args&& ...args) {
+        return std::make_shared<StorageIdentifier>(std::forward<Args>(args)...);
+    }
+    bool is_sys_alloc() {
+        return tag == SYS_ALLOC;
+    }
+    bool is_from_other() {
+        return tag == FROM_OTHER;
+    }
+    bool is_device_ptr() {
+        return tag == DEVICE_PTR;
+    }
+    bool is_invalid() {
+        return tag == INVALID;
+    }
+};
 } // namespace imperative
 } // namespace mgb
 

imperative/src/include/megbrain/imperative/proxy_graph_detail.h

@@ -21,9 +21,19 @@ SmallVector<TensorPtr>
 apply_on_physical_tensor(const OpDef& def,
         SmallVector<TensorPtr> inputs);
 
+void execute(const OpDef& def,
+        SmallVector<TensorPtr> inputs,
+        SmallVector<TensorPtr> outputs,
+        SmallVector<TensorPtr> workspace);
+
 std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(const OpDef& def,
         const SmallVector<LogicalTensorDesc>& inputs);
 
+std::tuple<SmallVector<MemoryDesc>, SmallVector<MemoryDesc>> infer_output_mem_desc(
+        const OpDef& def,
+        const SmallVector<TensorPtr>& inputs_tensors,
+        const SmallVector<MemoryDesc>& inputs_mems);
+
 void exec(const OpDef& def,
         const SmallVector<TensorPtr>& inputs,
         const SmallVector<TensorPtr>& outputs);