enabled most warnings on gcc, and fixed them

CMakeLists.txt

@@ -38,10 +38,10 @@ if(MSVC)
 
   set(CMAKE_STATIC_LINKER_FLAGS "${CMAKE_STATIC_LINKER_FLAGS} /LTCG:incremental")
 else()
-  set(CMAKE_CXX_FLAGS_RELEASE         " -std=c++11 -O3 -Ofast -m64 -pthread -march=${BUILD_ARCH} -msse4.1 -Wl,--no-as-needed -funroll-loops -ffinite-math-only -fPIC -Wno-unused-result -Wno-deprecated -Wno-deprecated-gpu-targets -Wsuggest-override -Werror")
+  set(CMAKE_CXX_FLAGS_RELEASE         " -std=c++11 -O3 -Ofast -m64 -pthread -march=${BUILD_ARCH} -msse4.1 -Wl,--no-as-needed -funroll-loops -ffinite-math-only -fPIC -Wno-unused-result -Wno-deprecated -Werror -Wno-pragmas")
   set(CMAKE_CXX_FLAGS_DEBUG           " -std=c++11 -g -rdynamic -O0 -pthread -fPIC -Wno-unused-result -Wno-deprecated -Wno-deprecated-gpu-targets -Wsuggest-override -Werror")
   set(CMAKE_CXX_FLAGS_ST              "${CMAKE_CXX_FLAGS_RELEASE} -DNDEBUG")
-  set(CMAKE_CXX_FLAGS_RELWITHDEBINFO  "${CMAKE_CXX_FLAGS_RELEASE} -pg -g -rdynamic")
+  set(CMAKE_CXX_FLAGS_RELWITHDEBINFO  "${CMAKE_CXX_FLAGS_RELEASE} -pg -g -rdynamic -Wall -Wextra -Wsuggest-override -Wno-unused-value -Wno-unknown-pragmas -Wno-sign-compare -Wno-missing-field-initializers -Wno-unused-variable")
   set(CMAKE_CXX_FLAGS_PROFILE         "${CMAKE_CXX_FLAGS_RELEASE} -pg -g -rdynamic")
   set(CMAKE_CXX_FLAGS_PROFGEN         "${CMAKE_CXX_FLAGS_RELEASE} -fprofile-generate -fprofile-correction")
   set(CMAKE_CXX_FLAGS_PROFUSE         "${CMAKE_CXX_FLAGS_RELEASE} -fprofile-use -fprofile-correction")

src/3rd_party/cnpy/cnpy.cpp

@@ -13,7 +13,8 @@
 
 char cnpy::BigEndianTest() {
     unsigned char x[] = {1,0};
-    short y = *(short*) x;
+    short y;
+    memcpy(&y, x, sizeof(y));
     return y == 1 ? '<' : '>';
 }
 

src/3rd_party/threadpool.h

@@ -106,7 +106,7 @@ class ThreadPool {
 
 // the constructor just launches some amount of workers
 inline ThreadPool::ThreadPool(size_t threads, size_t in_bound)
-  : stop(false), bound(in_bound) {
+  : bound(in_bound), stop(false) {
     reserve(threads);
 }
 

src/common/cli_wrapper.h

@@ -20,8 +20,8 @@ namespace cli {
 //
 // TODO: make use of it in the current CLI or remove. This is an old code used
 // for boost::program_options and might not be needed anymore.
-static uint16_t guess_terminal_width(uint16_t max_width = 0,
-                                     uint16_t default_width = 180);
+//static uint16_t guess_terminal_width(uint16_t max_width = 0,
+//                                     uint16_t default_width = 180);
 
 // TODO: use validators in ConfigParser
 namespace validators {

src/common/file_stream.h

@@ -19,6 +19,7 @@
 #ifdef __GNUC__
 #pragma GCC diagnostic push
 #pragma GCC diagnostic ignored "-Wsuggest-override"
+#pragma GCC diagnostic ignored "-Wunused-variable"
 #endif
 #include <boost/iostreams/filtering_stream.hpp>
 #ifdef __GNUC__

src/common/filesystem.h

@@ -7,6 +7,7 @@
 
 #ifdef __GNUC__
 #pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wunused-variable"
 #pragma GCC diagnostic ignored "-Wsuggest-override"
 #endif
 #include <boost/filesystem.hpp>

src/common/logging.cpp

@@ -133,9 +133,9 @@ static void setErrorHandlers() {
   struct sigaction sa = { 0 };
   sigemptyset(&sa.sa_mask);
   sa.sa_flags = SA_SIGINFO;
-  sa.sa_sigaction = [](int signal, siginfo_t *si, void *arg) { ABORT("Segmentation fault"); };
+  sa.sa_sigaction = [](int /*signal*/, siginfo_t*, void*) { ABORT("Segmentation fault"); };
   sigaction(SIGSEGV, &sa, NULL);
-  sa.sa_sigaction = [](int signal, siginfo_t *si, void *arg) { ABORT("Floating-point exception"); };
+  sa.sa_sigaction = [](int /*signal*/, siginfo_t*, void*) { ABORT("Floating-point exception"); };
   sigaction(SIGFPE, &sa, NULL);
 #endif
 }

src/data/sentencepiece_vocab.cpp

@@ -133,7 +133,7 @@ Ptr<VocabBase> createSentencePieceVocab(const std::string& vocabPath, Ptr<Option
 #ifdef USE_SENTENCEPIECE
     return New<SentencePieceVocab>(options, batchIndex);
 #else
-    batchIndex;
+    batchIndex; options;
     ABORT("*.spm suffix in path {} reserved for SentencePiece models, "
           "but support for SentencePiece is not compiled into Marian. "
           "Try to recompile after `cmake .. -DUSE_SENTENCEPIECE=on [...]`",

src/data/shortlist.h

@@ -288,7 +288,7 @@ public:
     }
   }
 
-  Ptr<Shortlist> generate(Ptr<data::CorpusBatch> batch) override {
+  Ptr<Shortlist> generate(Ptr<data::CorpusBatch> /*batch*/) override {
     std::vector<Word> tmp;
     return New<Shortlist>(idx_, tmp, reverseIdx_);
   }

src/examples/mnist/dataset.h

@@ -135,7 +135,7 @@ private:
 
 public:
   MNISTData(std::vector<std::string> paths,
-            std::vector<Ptr<Vocab>> vocabs = {},
+            std::vector<Ptr<Vocab>> /*vocabs*/ = {},
             Ptr<Config> options = nullptr)
       : Dataset(paths, options), IMAGE_MAGIC_NUMBER(2051), LABEL_MAGIC_NUMBER(2049) {
     loadData();
@@ -154,7 +154,7 @@ public:
     }
   }
 
-  Example next() override {}
+  Example next() override { return{ }; } //@TODO: this return was added to fix a warning. Is it correct?
 
 private:
   typedef unsigned char uchar;

src/examples/mnist/model.h

@@ -58,24 +58,24 @@ public:
 
   virtual Expr build(Ptr<ExpressionGraph> graph,
                      Ptr<data::Batch> batch,
-                     bool clean = false) override {
+                     bool /*clean*/ = false) override {
     return construct(graph, batch, inference_);
   }
 
-  void load(Ptr<ExpressionGraph> graph, const std::string& name, bool) override {
+  void load(Ptr<ExpressionGraph> /*graph*/, const std::string& /*name*/, bool) override {
     LOG(critical, "Loading MNIST model is not supported");
   }
 
-  void save(Ptr<ExpressionGraph> graph, const std::string& name, bool) override {
+  void save(Ptr<ExpressionGraph> /*graph*/, const std::string& /*name*/, bool) override {
     LOG(critical, "Saving MNIST model is not supported");
   }
 
-  void save(Ptr<ExpressionGraph> graph, const std::string& name) {
+  void save(Ptr<ExpressionGraph> /*graph*/, const std::string& /*name*/) {
     LOG(critical, "Saving MNIST model is not supported");
   }
 
-  Ptr<data::BatchStats> collectStats(Ptr<ExpressionGraph> graph,
-                                     size_t multiplier) {
+  Ptr<data::BatchStats> collectStats(Ptr<ExpressionGraph> /*graph*/,
+                                     size_t /*multiplier*/) {
     LOG(critical, "Collecting stats in MNIST model is not supported");
     return nullptr;
   }
@@ -98,7 +98,7 @@ protected:
    */
   virtual Expr construct(Ptr<ExpressionGraph> g,
                          Ptr<data::Batch> batch,
-                         bool inference = false) {
+                         bool /*inference*/ = false) {
     const std::vector<int> dims = {784, 2048, 2048, 10};
 
     // Start with an empty expression graph

src/graph/expression_graph.cpp

@@ -25,7 +25,7 @@ Expr ExpressionGraph::dropout(float prob, const Shape& shape) {
 }
 
 void ExpressionGraph::checkNan(Tensor t) {
-  ABORT_IF(throwNaN_, "Not implemented");
+  ABORT_IF(throwNaN_, "Not implemented"); t;
   // ABORT_IF(throwNaN_ && IsNan(t), "Tensor has NaN");
 }
 

src/graph/node_initializers.h

@@ -35,7 +35,7 @@ NodeInitializer dropout(float prob);
 
 void gumbel(Tensor t);
 
-static inline void dummy(Tensor t) {}
+static inline void dummy(Tensor) {}
 
 NodeInitializer from_vector(const std::vector<float>& v);
 NodeInitializer from_vector(const std::vector<IndexType>& v);

src/layers/weight.h

@@ -14,9 +14,9 @@ public:
   virtual Expr getWeights(Ptr<ExpressionGraph> graph,
                           Ptr<data::CorpusBatch> batch)
       = 0;
-  virtual void debugWeighting(std::vector<float> weightedMask,
-                              std::vector<float> freqMask,
-                              Ptr<data::CorpusBatch> batch){};
+  virtual void debugWeighting(std::vector<float> /*weightedMask*/,
+                              std::vector<float> /*freqMask*/,
+                              Ptr<data::CorpusBatch> /*batch*/){};
 };
 
 class DataWeighting : public WeightingBase {

src/models/states.h

@@ -97,6 +97,6 @@ public:
   // Set current target token position in state when decoding
   void setPosition(size_t position) { position_ = position; }
 
-  virtual void blacklist(Expr totalCosts, Ptr<data::CorpusBatch> batch) {}
+  virtual void blacklist(Expr /*totalCosts*/, Ptr<data::CorpusBatch> /*batch*/) {}
 };
 }  // namespace marian

src/models/transformer.h

@@ -200,7 +200,7 @@ public:
 
   // determine the multiplicative-attention probability and performs the associative lookup as well
   // q, k, and v have already been split into multiple heads, undergone any desired linear transform.
-  Expr Attention(std::string prefix,
+  Expr Attention(std::string /*prefix*/,
                  Expr q,              // [-4: beam depth * batch size, -3: num heads, -2: max tgt length, -1: split vector dim]
                  Expr k,              // [-4: batch size, -3: num heads, -2: max src length, -1: split vector dim]
                  Expr v,              // [-4: batch size, -3: num heads, -2: max src length, -1: split vector dim]
@@ -444,7 +444,7 @@ public:
                        const rnn::State& prevDecoderState,
                        std::string prefix,
                        Expr input,
-                       Expr selfMask,
+                       Expr /*selfMask*/,
                        int /*startPos*/) const {
     float dropoutRnn = inference_ ? 0.f : opt<float>("dropout-rnn");
 
@@ -609,7 +609,7 @@ private:
   Ptr<mlp::MLP> output_;
 
 private:
-  void LazyCreateOutputLayer(std::string prefix)
+  void LazyCreateOutputLayer()
   {
     if(output_) // create it lazily
       return;
@@ -667,7 +667,7 @@ public:
   virtual Ptr<DecoderState> step(Ptr<ExpressionGraph> graph,
                                  Ptr<DecoderState> state) override {
     ABORT_IF(graph != graph_, "An inconsistent graph parameter was passed to step().");
-    LazyCreateOutputLayer(prefix_ + "_ff_logit_out");
+    LazyCreateOutputLayer();
     return step(state);
   }
 

src/rnn/constructors.h

@@ -92,7 +92,7 @@ public:
   }
 
   virtual void add_input(Expr input) {
-    inputs_.push_back([input](Ptr<rnn::RNN> rnn) { return input; });
+    inputs_.push_back([input](Ptr<rnn::RNN> /*rnn*/) { return input; });
   }
 };
 

src/tensors/cpu/prod.cpp

@@ -92,6 +92,7 @@ void Prod(marian::Tensor C,
         C->data(),
         ldc);
 #else
+  C; A; B; transA; transB; beta; scalar;
   ABORT("You need to compile with MKL in order to use the CPU version");
 #endif
 }
@@ -148,6 +149,7 @@ void ProdBatched(marian::Tensor C,
           (int)ldc);
   }
 #else
+  C; allocator; A; B; transA; transB; beta; scalar;
   ABORT("You need to compile with MKL in order to use the CPU version");
 #endif
 }

src/training/communicator.cpp

@@ -148,16 +148,21 @@ public:
 #pragma warning(disable: 4100) // unreferenced formal parameter
   // most functions are no-ops when applied to a single process
   virtual void barrier(MPI_Comm comm) const override {
+    comm;
   }
   virtual void bCast(void* buf, size_t count, MPI_Datatype datatype, size_t rootRank, MPI_Comm comm) const override {
+    buf; count; datatype; rootRank; comm;
   }
   virtual void sSend(void* buf, size_t count, MPI_Datatype datatype, size_t destRank, int tag, MPI_Comm comm) const override {
+    buf; count; datatype; destRank; tag; comm;
   }
   virtual void recv(void* buf, size_t count, MPI_Datatype datatype, size_t sourceRank, int tag, MPI_Comm comm, MPI_Status* status) const override {
+    buf; count; datatype; sourceRank; tag; comm;
     // @TODO: fill in status
     ABORT_IF(status != MPI_STATUS_IGNORE, "FakeMPIWrapper::recv() does not yet implement returning a status object");
   }
   virtual void allReduce(const void* sendbuf, void* recvbuf, size_t count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) const override {
+    count; datatype; op; comm;
     // @TODO: There is only one place where this is called with sendbuf != recvbuf, which is sync multi-node.
     //        I think that can be changed to use the same buffer. Then we should change this API
     //        to only accept one parameter, and remove this error check can be removed.

src/training/communicator_nccl.h

@@ -126,8 +126,8 @@ public:
         comms_(graphs.size()),
         streams_(graphs.size()),
         devices_(graphs.size()),
-        threadPool_(graphs.size(), graphs.size()), threadResults_(graphs.size()),
-        mpi_(mpi) {
+        mpi_(mpi),
+        threadPool_(graphs.size(), graphs.size()), threadResults_(graphs.size()) {
     mpiBarrier(); // barrier to group the multiple log messages from MPI processes
     LOG(info, "[comm] Using NCCL {} {}for GPU communication",
         ncclVersionString(),

src/training/gradient_dropping/sparse_tensor.h

@@ -34,7 +34,7 @@ class SparseTensorBase : public std::enable_shared_from_this<SparseTensorBase> {
 
 public:
   SparseTensorBase(int capacity, Ptr<Backend> backend)
-      : backend_(backend), capacity_(capacity) {
+      : capacity_(capacity), backend_(backend) {
     data_ = newData<float>(capacity, backend);
     indices_ = newData<int>(capacity, backend);
   }

src/training/graph_group_async_drop.cpp

@@ -106,8 +106,10 @@ void AsyncGraphGroupDrop::init(Ptr<data::Batch> batch) {
 
       // prepare droppers
       std::vector<GradientDrop> tmpDropper;
-      for(auto device : devices_)
+      for(auto device : devices_) {
+        device; // @TODO: 'device' is not used. Is this nested loop correct? 
         tmpDropper.push_back(PrepareGradientDrop(graphs_[i]->getDeviceId()));
+      }
       droppers_.push_back(tmpDropper);
 
       // sparsetensor to store sparsified gradients per-device per-shard

src/training/graph_group_async_drop.h

@@ -32,8 +32,8 @@ protected:
 public:
   AsyncGraphGroupDrop(Ptr<Config> options)
       : AsyncGraphGroup(options),
+        dropping_warmup{options->get<size_t>("grad-dropping-warmup")},
         droping_rate{options->get<float>("grad-dropping-rate")},
-        dropping_momentum{options->get<float>("grad-dropping-momentum")},
-        dropping_warmup{options->get<size_t>("grad-dropping-warmup")} {}
+        dropping_momentum{options->get<float>("grad-dropping-momentum")} {}
 };
 }  // namespace marian

src/training/graph_group_multinode.cpp

@@ -269,7 +269,7 @@ void MultiNodeGraphGroup::launchServerThread() {
         size_t size = shardSizes_[gpu];
 
         threads.emplace_back(std::thread(
-            [=](int gpu, size_t offset, size_t size, size_t batchWords) {
+            [=](int gpu, size_t offset, size_t size) {
               std::lock_guard<std::mutex> guard(shardMutex_[gpu]);
 
               // Copy grads to appropriate GPU
@@ -292,8 +292,7 @@ void MultiNodeGraphGroup::launchServerThread() {
             },
             gpu,
             offset,
-            size,
-            messageInfo[MSG_INFO_BATCHWORDS_]));
+            size));
 
         offset += size;
       }

src/training/graph_group_multinode_sync.h

@@ -134,9 +134,9 @@ public:
   MultiNodeGraphGroupSync(Ptr<Config> options)
       : Base(options),
         tau_{options_->get<size_t>("optimizer-delay")},
+        syncOptimizer_{Optimizer(options_)},
         movingAvg_{options_->get<float>("exponential-smoothing") > 0},
-        mvDecay_{options_->get<float>("exponential-smoothing")},
-        syncOptimizer_{Optimizer(options_)} {
+        mvDecay_{options_->get<float>("exponential-smoothing")} {
   }
 
   /**

src/training/validator.h

@@ -390,8 +390,8 @@ private:
 public:
   BleuValidator(std::vector<Ptr<Vocab>> vocabs, Ptr<Config> options, bool detok = false)
       : Validator(vocabs, options, false),
-        quiet_(options_->get<bool>("quiet-translation")),
-        detok_(detok) {
+        detok_(detok),
+        quiet_(options_->get<bool>("quiet-translation")) {
     Ptr<Options> opts = New<Options>();
     opts->merge(options);
     opts->set("inference", true);

src/translator/scorers.h

@@ -13,7 +13,7 @@ public:
 
   virtual float breakDown(size_t i) { return getLogProbs()->val()->get(i); }
 
-  virtual void blacklist(Expr totalCosts, Ptr<data::CorpusBatch> batch){};
+  virtual void blacklist(Expr /*totalCosts*/, Ptr<data::CorpusBatch> /*batch*/){};
 };
 
 class Scorer {
@@ -40,10 +40,9 @@ public:
                                 int beamSize)
       = 0;
 
-  virtual void init(Ptr<ExpressionGraph> graph) {}
+  virtual void init(Ptr<ExpressionGraph>) {}
 
-  virtual void setShortlistGenerator(
-      Ptr<data::ShortlistGenerator> shortlistGenerator){};
+  virtual void setShortlistGenerator( Ptr<data::ShortlistGenerator> /*shortlistGenerator*/){};
   virtual Ptr<data::Shortlist> getShortlist() { return nullptr; };
   virtual std::vector<float> getAlignment() { return {}; };
 };

vs/Marian.vcxproj.filters

@@ -268,6 +268,15 @@
     <ClCompile Include="..\src\common\version.cpp">
       <Filter>common</Filter>
     </ClCompile>
+    <ClCompile Include="..\src\examples\iris\helper.cpp">
+      <Filter>examples\iris</Filter>
+    </ClCompile>
+    <ClCompile Include="..\src\examples\iris\iris.cpp">
+      <Filter>examples\iris</Filter>
+    </ClCompile>
+    <ClCompile Include="..\src\examples\mnist\mnist_ffnn.cpp">
+      <Filter>examples\mnist</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="..\src\marian.h" />
@@ -1093,6 +1102,21 @@
     <ClInclude Include="..\src\common\version.h">
       <Filter>common</Filter>
     </ClInclude>
+    <ClInclude Include="..\src\examples\mnist\dataset.h">
+      <Filter>examples\mnist</Filter>
+    </ClInclude>
+    <ClInclude Include="..\src\examples\mnist\model.h">
+      <Filter>examples\mnist</Filter>
+    </ClInclude>
+    <ClInclude Include="..\src\examples\mnist\model_lenet.h">
+      <Filter>examples\mnist</Filter>
+    </ClInclude>
+    <ClInclude Include="..\src\examples\mnist\training.h">
+      <Filter>examples\mnist</Filter>
+    </ClInclude>
+    <ClInclude Include="..\src\examples\mnist\validator.h">
+      <Filter>examples\mnist</Filter>
+    </ClInclude>
   </ItemGroup>
   <ItemGroup>
     <Filter Include="3rd_party">
@@ -1212,11 +1236,40 @@
     <Filter Include="microsoft">
       <UniqueIdentifier>{bf291eae-3de9-4294-b5aa-07ab631951e1}</UniqueIdentifier>
     </Filter>
+    <Filter Include="examples">
+      <UniqueIdentifier>{3dbe1ce2-4645-42d0-99b1-0989e42cd7df}</UniqueIdentifier>
+    </Filter>
+    <Filter Include="examples\iris">
+      <UniqueIdentifier>{c1ebcdfa-3994-4430-aad0-0d9d54cf10e3}</UniqueIdentifier>
+    </Filter>
+    <Filter Include="examples\mnist">
+      <UniqueIdentifier>{55dd3952-a6fd-41d8-b99c-306b9e6c348d}</UniqueIdentifier>
+    </Filter>
   </ItemGroup>
   <ItemGroup>
     <None Include=".editorConfig" />
     <None Include="..\src\training\communicator_nccl.h">
       <Filter>training</Filter>
     </None>
+    <None Include="..\src\examples\cmake_install.cmake">
+      <Filter>examples</Filter>
+    </None>
+    <None Include="..\src\examples\Makefile">
+      <Filter>examples</Filter>
+    </None>
+    <None Include="..\src\examples\README.md">
+      <Filter>examples</Filter>
+    </None>
+    <None Include="..\src\examples\iris\iris.data">
+      <Filter>examples\iris</Filter>
+    </None>
+    <None Include="..\src\examples\mnist\download.sh">
+      <Filter>examples\mnist</Filter>
+    </None>
+  </ItemGroup>
+  <ItemGroup>
+    <Text Include="..\src\examples\CMakeLists.txt">
+      <Filter>examples</Filter>
+    </Text>
   </ItemGroup>
 </Project>