feat(imperative): merge common c++ code to megbrain

GitOrigin-RevId: d093778e10

CMakeLists.txt

@@ -213,6 +213,15 @@ if(MGE_WITH_TEST)
 endif()
 
 option(MGE_WITH_DISTRIBUTED "Build with distributed support" ON)
+option(MGE_BUILD_XXX "Build _xxx.so instead of mgb.so " OFF)
+if(MGE_BUILD_XXX)
+    set(CMAKE_CXX_STANDARD 17)
+endif()
+
+option(MGE_BUILD_SDK "Build load_and_run" ON)
+if(MGE_BUILD_XXX)
+    set(MGE_BUILD_SDK OFF)
+endif()
 
 if(NOT MGE_WITH_CUDA)
     message("-- Disable distributed support, as CUDA is not enabled.")
@@ -522,7 +531,7 @@ endif()
 
 set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${MARCH}")
 
-
+set(MGB_ENABLE_IMPERATIVE, ${MGE_BUILD_XXX})
 # Write out megbrain_build_config.h
 # It defines macros needed by both megbrain and dnn
 configure_file(src/megbrain_build_config.h.in ${CMAKE_CURRENT_BINARY_DIR}/genfiles/megbrain_build_config.h)
@@ -566,14 +575,23 @@ if(MGE_WITH_DISTRIBUTED)
 endif()
 
 add_subdirectory(src)
-add_subdirectory(sdk/load-and-run)
+
+if(MGE_BUILD_SDK)
+    add_subdirectory(sdk/load-and-run)
+endif()
 
 if(MGE_WITH_PYTHON_MODULE)
-    add_subdirectory(python_module)
+    if(MGE_BUILD_XXX)
+        add_subdirectory(imperative)
+    else()
+        add_subdirectory(python_module)
+    endif()
 endif()
 
 if(MGE_WITH_TEST AND MGE_ENABLE_RTTI)
-    add_subdirectory(test)
+    if(NOT MGE_BUILD_XXX)
+        add_subdirectory(test)
+    endif()
 endif()
 
 if(TARGET mgb)
@@ -597,6 +615,21 @@ if(TARGET mgb)
         DEPENDS mgb
         VERBATIM
     )
+elseif(TARGET _xxx)
+    add_custom_target(
+        develop
+        COMMAND ${CMAKE_COMMAND} -E create_symlink
+          ${CMAKE_CURRENT_BINARY_DIR}/imperative/python/${PACKAGE_NAME}/$<TARGET_FILE_NAME:${MODULE_NAME}>
+          ${CMAKE_CURRENT_SOURCE_DIR}/imperative/python/${PACKAGE_NAME}/$<TARGET_FILE_NAME:${MODULE_NAME}>
+        COMMAND ${CMAKE_COMMAND} -E create_symlink
+          ${CMAKE_CURRENT_BINARY_DIR}/imperative/python/${PACKAGE_NAME}/ops/_internal/generated_ops.py
+          ${CMAKE_CURRENT_SOURCE_DIR}/imperative/python/${PACKAGE_NAME}/ops/_internal/generated_ops.py
+        COMMAND ${CMAKE_COMMAND} -E create_symlink
+          ${CMAKE_CURRENT_BINARY_DIR}/imperative/python/${PACKAGE_NAME}/ops/_internal/param_defs.py
+          ${CMAKE_CURRENT_SOURCE_DIR}/imperative/python/${PACKAGE_NAME}/ops/_internal/param_defs.py
+        DEPENDS _xxx
+        VERBATIM
+    )
 endif()
 
 IF(APPLE)

dnn/CMakeLists.txt

@@ -59,7 +59,9 @@ install(TARGETS opr_param_defs EXPORT ${MGE_EXPORT_TARGETS})
 
 
 if(MGE_WITH_TEST)
-    add_subdirectory(test)
+    if(NOT MGE_BUILD_XXX)
+        add_subdirectory(test)
+    endif()
 endif()
 
 add_subdirectory(src)

dnn/scripts/gen_param_defs.py

@@ -298,6 +298,9 @@ class PyWriter(IndentWriterBase):
 
     _enum_member2num = None
 
+    def __init__(self, for_imperative=False):
+        self._imperative = for_imperative
+
     def __call__(self, fout, defs):
         super().__call__(fout)
         self._enum_member2num = []
@@ -339,19 +342,35 @@ class PyWriter(IndentWriterBase):
             '       return super()._missing_(value)\n'
             '\n'
         )
-        self._write(
-            'def _as_dtype_num(dtype):\n'
-            '   import megengine._internal.mgb as m\n'
-            '   return m._get_dtype_num(dtype)\n'
-            '\n'
-        )
-        self._write(
-            '''
-def _as_serialized_dtype(dtype):
-    import megengine._internal.mgb as m
-    return m._get_serialized_dtype(dtype)
-'''
-        )
+        if not self._imperative:
+            self._write(
+                'def _as_dtype_num(dtype):\n'
+                '    import megengine._internal.mgb as m\n'
+                '    return m._get_dtype_num(dtype)\n'
+                '\n'
+            )
+
+            self._write(
+                'def _as_serialized_dtype(dtype):\n'
+                '    import megengine._internal.mgb as m\n'
+                '    return m._get_serialized_dtype(dtype)\n'
+                '\n'
+            )
+        else:
+            self._write(
+                'def _as_dtype_num(dtype):\n'
+                '    import xxx._xxx.utils as m\n'
+                '    return m._get_dtype_num(dtype)\n'
+                '\n'
+            )
+
+            self._write(
+                'def _as_serialized_dtype(dtype):\n'
+                '    import xxx._xxx.utils as m\n'
+                '    return m._get_serialized_dtype(dtype)\n'
+                '\n'
+            )
+
         self._process(defs)
         self._write(
             '''
@@ -777,8 +796,12 @@ def main():
                         'cpp file')
     parser.add_argument('input')
     parser.add_argument('output')
+    parser.add_argument('--imperative', action='store_true',
+                        help='generate files for imperatvie ')
     args = parser.parse_args()
 
+    for_imperative = args.imperative
+
     with open(args.input) as fin:
         inputs = fin.read()
         exec(inputs, {'pdef': ParamDef, 'Doc': member_defs.Doc})
@@ -787,7 +810,7 @@ def main():
         input_hash = input_hash.hexdigest()
 
     if args.type == 'py':
-        writer = PyWriter()
+        writer = PyWriter(for_imperative=for_imperative)
     else:
         assert args.type == 'c++'
         if args.enumv:

src/CMakeLists.txt

@@ -151,27 +151,31 @@ if(ANDROID)
     target_link_libraries(megbrain PUBLIC log)
 endif()
 
-# Build as SHARED or STATIC depending on BUILD_SHARED_LIBS=ON/OFF
-add_library(megengine)
-target_link_libraries(megengine PUBLIC megbrain megdnn)
-if (UNIX AND NOT APPLE)
-    # TODO: Use target_link_options after upgrading to CMake 3.13
-    target_link_options(megengine PRIVATE -Wl,--no-undefined -Wl,--version-script=${PROJECT_SOURCE_DIR}/python_module/src/version.ld)
-endif()
-set_target_properties(megengine PROPERTIES CXX_VISIBILITY_PRESET default)
-set_target_properties(megengine PROPERTIES WINDOWS_EXPORT_ALL_SYMBOLS TRUE)
-if (MGE_WITH_DISTRIBUTED)
+if(NOT MGE_BUILD_XXX)
+    # Build as SHARED or STATIC depending on BUILD_SHARED_LIBS=ON/OFF
+    add_library(megengine)
+    target_link_libraries(megengine PUBLIC megbrain megdnn)
+    if (UNIX AND NOT APPLE)
+        # TODO: Use target_link_options after upgrading to CMake 3.13
+        # FIXME; Please use right directory for mgb or imperative
+        target_link_options(megengine PRIVATE -Wl,--no-undefined -Wl,--version-script=${PROJECT_SOURCE_DIR}/python_module/src/version.ld)
+    endif()
+    set_target_properties(megengine PROPERTIES CXX_VISIBILITY_PRESET default)
+    set_target_properties(megengine PROPERTIES WINDOWS_EXPORT_ALL_SYMBOLS TRUE)
     # Do not export targets if MGE_WITH_DISTRIBUTED is on. MegRay is not ready
     # for this.
     install(TARGETS megengine
             LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
             ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR})
-else()
-    install(TARGETS megengine megbrain
+endif()
+
+if (NOT MGE_WITH_DISTRIBUTED)
+    install(TARGETS megbrain
             EXPORT ${MGE_EXPORT_TARGETS}
             LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
             ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR})
 endif()
+
 foreach(_PATH ${MGB_INC})
     install(DIRECTORY ${_PATH}/megbrain DESTINATION ${CMAKE_INSTALL_INCLUDEDIR} FILES_MATCHING PATTERN "*.h")
 endforeach()

src/core/impl/graph/cg_impl.cpp

@@ -271,6 +271,23 @@ OperatorNodeBase* ComputingGraphImpl::insert_opr(
         std::unique_ptr<OperatorNodeBase> opr_uniqp) {
     auto opr = opr_uniqp.get();
 
+    if (options().imperative_proxy_graph) {
+        if (!opr->inserted_in_graph()) {
+            m_opr_refkeeper.emplace_back(std::move(opr_uniqp));
+            opr->set_inserted_in_graph();
+            opr->init_output_comp_node();
+            opr->init_output_dtype();
+            opr->init_output_format();
+            // register static infer
+            {
+                auto&& mgr = static_infer_manager_impl();
+                auto old = mgr.set_register_allowed_opr(opr);
+                opr->init_output_static_infer_desc();
+                mgr.set_register_allowed_opr(old);
+            }
+        }
+        return opr;
+    }
     if (opr->inserted_in_graph()) {
         // FIXME: it's just a trick used for re-evaluation in eager evaluation
         // mode. Since comp_graph has already taken an ownership of the opr,

src/core/impl/graph/grad_manager.cpp

@@ -133,6 +133,15 @@ void cg::register_grad_func(Typeinfo *opr_type, OprGradFunc grad) {
             opr_type->name);
 }
 
+OprGradFunc* cg::lookup_grad_func(Typeinfo *opr_type) {
+    auto giter = static_data().grad_func_registry.find(opr_type);
+    if (giter != static_data().grad_func_registry.end()) {
+        return &giter->second;
+    } else {
+        return nullptr;
+    }
+}
+
 class GradManager::StreamStrongPropInfer {
     DepOprIter m_opr_iter;
     ThinHashSet<OperatorNodeBase*> m_strong_oprs;

src/core/impl/graph/operator_node.cpp

@@ -101,6 +101,11 @@ OperatorNodeBase::~OperatorNodeBase() noexcept {
 }
 
 void OperatorNodeBase::execute(ExecEnv &env) {
+    if (owner_graph()->options().imperative_proxy_graph) {
+        do_execute(env);
+        return;
+    }
+
     owner_graph()->event().signal_inplace<event::OprExecStart>(this, &env);
 
     // dispatch waiting commands

src/core/impl/graph/var_node.cpp

@@ -230,6 +230,9 @@ VarNode& VarNode::format(TensorFormat format) {
 
 bool VarNode::set_fwd_in2out_readonly(
         VarNode *input, const SubTensorSpec &sub) {
+    if (owner_graph()->options().imperative_proxy_graph) {
+        return false;
+    }
     return static_cast<ComputingGraphImpl*>(owner_graph())
         ->var_node_mem_manager().fwd_in2out_readonly(input, sub, this);
 }
@@ -242,6 +245,7 @@ VarNode& VarNode::set_fwd_in2out_writable(VarNode *input) {
 
 
 VarNode& VarNode::set_fwd_in2out_writable_force(VarNode *input) {
+    mgb_assert(!owner_graph()->options().imperative_proxy_graph);
     static_cast<ComputingGraphImpl*>(owner_graph())
         ->var_node_mem_manager().fwd_in2out_writable_force(input, this);
     return *this;

src/core/include/megbrain/graph/cg.h

@@ -440,6 +440,8 @@ class ComputingGraph : public std::enable_shared_from_this<ComputingGraph>,
             bool eager_evaluation = false;
 #endif
 
+            bool imperative_proxy_graph = false;
+
             //! add extra deps for the comp seq if a specific var is dependent
             ThinHashMap<VarNode*, VarNodeArray> extra_vardeps;
 

src/core/include/megbrain/graph/grad_impl.h

@@ -73,6 +73,11 @@ namespace cg {
      */
     void register_grad_func(Typeinfo *opr_type, OprGradFunc grad);
 
+    /*!
+     * \brief lookup grad func for an operator type
+     */
+    OprGradFunc* lookup_grad_func(Typeinfo *opr_type);
+
     /*!
      * \brief add a callback to be invoked when grad of given var is computed
      *

src/core/include/megbrain/graph/operator_node.h

@@ -69,6 +69,10 @@ class OperatorNodeConfig final: public Hashable {
             return *this;
         }
 
+        const Maybe<std::string>& name() const {
+            return m_name;
+        }
+
         /*!
          * \brief update instance ID
          *

src/core/include/megbrain/graph/var_node.h

@@ -22,6 +22,10 @@
 #include <mutex>
 
 namespace mgb {
+namespace imperative {
+    class ProxyGraph;
+} // namespace imperative
+
 namespace cg {
 namespace static_infer {
     class StaticInferManagerImpl;
@@ -576,6 +580,7 @@ class VarNode final: public GraphNodeBase {
         friend class VarDevMemDefragmenter;
         friend class EagerEvalManager;
         friend class MemAllocPlan;
+        friend class imperative::ProxyGraph;
 };
 
 enum class VarNode::Flag: uint32_t {

src/megbrain_build_config.h.in

@@ -29,6 +29,8 @@
 #cmakedefine01 MGB_ENABLE_FBS_SERIALIZATION
 #cmakedefine01 MGB_IS_DEV
 
+#cmakedefine01 MGB_ENABLE_IMPERATIVE
+
 // DNN related flags
 // Platform macro's
 #cmakedefine01 MEGDNN_WITH_CUDA

src/opr/impl/dnn/batch_norm.cpp

@@ -40,29 +40,37 @@ BatchNormForward::BatchNormForward(VarNode *x,
     Super{x->owner_graph(), config, "batch_norm",
           {x, scale, bias, mean, variance}}
 {
-    auto check_dest = [&](VarNode* dest) {
-        auto dest_opr = dest->owner_opr();
-        mgb_throw_if(!(dest_opr->same_type<SharedDeviceTensor>() ||
-                       dest_opr->same_type<VolatileSharedDeviceTensor>()),
-                     GraphError,
-                     "mean&variance in BatchNorm must be SharedDeviceTensor/VolatileSharedDeviceTensor; "
-                     "got %s{%s} actually",
-                     dest_opr->cname(), dest_opr->dyn_typeinfo()->name);
-    };
-    check_dest(mean);
-    check_dest(variance);
+    if(owner_graph()->options().imperative_proxy_graph) {
+        m_force_inplace = false;
+    }
+
+    if (m_force_inplace) {
+        auto check_dest = [&](VarNode* dest) {
+            auto dest_opr = dest->owner_opr();
+            mgb_throw_if(!(dest_opr->same_type<SharedDeviceTensor>() ||
+                    dest_opr->same_type<VolatileSharedDeviceTensor>()),
+                    GraphError,
+                    "mean and variance in BatchNorm must be SharedDeviceTensor "
+                    "or VolatileSharedDeviceTensor; got %s{%s} actually",
+                    dest_opr->cname(), dest_opr->dyn_typeinfo()->name);
+        };
+        check_dest(mean);
+        check_dest(variance);
+    }
 
     init_megdnn_opr(*this, param);
 
     add_input({x, scale, bias, mean, variance});
 
-    output(0)->
-        set_fwd_in2out_writable_force(input(3)).
-        add_flag(VarNode::Flag::NO_MEM_RECLAIM);
+    if (m_force_inplace) {
+        output(0)->
+            set_fwd_in2out_writable_force(input(3)).
+            add_flag(VarNode::Flag::NO_MEM_RECLAIM);
 
-    output(1)->
-        set_fwd_in2out_writable_force(input(4)).
-        add_flag(VarNode::Flag::NO_MEM_RECLAIM);
+        output(1)->
+            set_fwd_in2out_writable_force(input(4)).
+            add_flag(VarNode::Flag::NO_MEM_RECLAIM);
+    }
 }
 
 BatchNormForward::BatchNormForward(VarNode *x,
@@ -129,17 +137,40 @@ BatchNormForward::do_make_node_prop() const {
 
 void BatchNormForward::scn_do_execute() {
     auto &&x = input(0)->dev_tensor();
+    auto &&y = output(4)->dev_tensor();
+    mgb_assert(x.layout().is_contiguous() &&
+               y.layout().is_contiguous());
+#if MGB_ENABLE_IMPERATIVE
+    if (input().size() == 5) { // need running mean/variance
+        auto &&o0 = output(0)->dev_tensor(),
+             &&o1 = output(1)->dev_tensor(),
+             &&i0 = input(3)->dev_tensor(),
+             &&i1 = input(4)->dev_tensor();
+        mgb_assert(o0.raw_ptr() && o1.raw_ptr()); // non-empty tensor
+        mgb_assert(o0.comp_node() == i0.comp_node() &&
+                   o1.comp_node() == i1.comp_node() &&
+                   o0.layout().eq_layout(i0.layout()) &&
+                   o1.layout().eq_layout(i1.layout()));
+        if (!m_force_inplace) {
+            if (o0.raw_ptr() != i0.raw_ptr()) {
+                o0.copy_from_fixlayout(i0);
+            }
+            if (o1.raw_ptr() != i1.raw_ptr()) {
+                o1.copy_from_fixlayout(i1);
+            }
+        } else {
+            mgb_assert(o0.raw_ptr() == i0.raw_ptr()
+                    && o1.raw_ptr() == i1.raw_ptr());
+        }
+    }
+#endif
     auto scale = input(1)->dev_tensor().as_megdnn();
     auto bias = input(2)->dev_tensor().as_megdnn();
     auto mean = output(0)->dev_tensor().as_megdnn();
     auto variance = output(1)->dev_tensor().as_megdnn();
     auto save_mean = output(2)->dev_tensor().as_megdnn();
     auto save_variance = output(3)->dev_tensor().as_megdnn();
-    auto &&y = output(4)->dev_tensor();
-    auto workspace = intl::get_megdnn_workspace_from_var(
-        output().back());
-    mgb_assert(x.layout().is_contiguous() &&
-               y.layout().is_contiguous());
+    auto workspace = intl::get_megdnn_workspace_from_var(output().back());
     megdnn_opr()->exec(x.as_megdnn(), scale, bias, mean, variance,
         save_mean, save_variance, y.as_megdnn(), workspace);
 }
@@ -191,6 +222,14 @@ void BatchNormForward::init_output_dtype() {
     }
 }
 
+void BatchNormForward::mem_plan_fwd_in2out_writable() {
+    if (!m_force_inplace && input().size() == 5) {
+        // TODO: testing
+        output(0)->set_fwd_in2out_writable(input(3));
+        output(1)->set_fwd_in2out_writable(input(4));
+    }
+}
+
 MGB_IMPL_OPR_GRAD(BatchNormForward) {
     mgb_assert(wrt_idx < 5);
     if (wrt_idx < 3) {

src/opr/impl/internal/megdnn_opr_wrapper.cpp

@@ -271,17 +271,26 @@ WorkspaceLimitGetter::get_impl(ComputingGraph *graph) {
 
 size_t WorkspaceLimitGetter::get_workspace_limit(
         ComputingGraph *graph, CompNode cn, size_t old_limit) {
+    if (graph->options().imperative_proxy_graph) {
+        return old_limit;
+    }
     if (!graph->options().seq_opt.enable_mem_reuse_alloc)
         return old_limit;
     return get_impl(graph)->get_workspace_limit(cn, old_limit);
 }
 
 bool WorkspaceLimitGetter::is_prealloc_run(ComputingGraph* graph) {
+    if (graph->options().imperative_proxy_graph) {
+        return false;
+    }
     return graph->options().seq_opt.enable_mem_reuse_alloc &&
            get_impl(graph)->is_prealloc_run();
 }
 
 VarNode* WorkspaceLimitGetter::register_to_graph(ComputingGraph *graph) {
+    if (graph->options().imperative_proxy_graph) {
+        return nullptr;
+    }
     auto maker = [graph](){
         return std::make_shared<Impl>(graph);
     };

src/opr/include/megbrain/opr/dnn/batch_norm.h

@@ -75,6 +75,10 @@ MGB_DEFINE_OPR_CLASS(BatchNormForward,
             const TensorShapeArray &output_shapes) const override;
         void init_output_static_infer_desc() override;
         void init_output_dtype() override;
+        void mem_plan_fwd_in2out_writable() override;
+
+        // if set to True, running mean/variance will be updated inplace
+        bool m_force_inplace = true;
 };
 
 using BatchNorm = BatchNormForward;