Backport #16798, #16836 and #16838 to 1.6

CMakeLists.txt

@@ -633,14 +633,17 @@ if(USE_CUDA)
  else()
  list(APPEND CUDA_INCLUDE_DIRS ${INCLUDE_DIRECTORIES})
  # define preprocessor macro so that we will not include the generated forcelink header
+ if(ENABLE_CUDA_RTC)
+ add_definitions(-DMXNET_ENABLE_CUDA_RTC=1)
+ endif()
+ # Create '.cmake' files for cuda compiles given definitions added thus far
  mshadow_cuda_compile(cuda_objs ${CUDA})
  if(MSVC)
  if(ENABLE_CUDA_RTC)
  FIND_LIBRARY(CUDA_nvrtc_LIBRARY nvrtc "${CUDA_TOOLKIT_ROOT_DIR}/lib/x64" "${CUDA_TOOLKIT_ROOT_DIR}/lib/win32")
  list(APPEND mxnet_LINKER_LIBS ${CUDA_nvrtc_LIBRARY})
  set(CUDA_cuda_LIBRARY "${CUDA_nvrtc_LIBRARY}/../cuda.lib")
  list(APPEND mxnet_LINKER_LIBS ${CUDA_cuda_LIBRARY})
- add_definitions(-DMXNET_ENABLE_CUDA_RTC=1)
  endif()
  FIND_LIBRARY(CUDA_cufft_LIBRARY nvrtc "${CUDA_TOOLKIT_ROOT_DIR}/lib/x64" "${CUDA_TOOLKIT_ROOT_DIR}/lib/win32")
  list(APPEND mxnet_LINKER_LIBS "${CUDA_cufft_LIBRARY}/../cufft.lib") # For fft operator
@@ -652,7 +655,6 @@ if(USE_CUDA)
  list(APPEND mxnet_LINKER_LIBS cufft cusolver)
  if(ENABLE_CUDA_RTC)
  list(APPEND mxnet_LINKER_LIBS nvrtc cuda)
- add_definitions(-DMXNET_ENABLE_CUDA_RTC=1)
  endif()
  link_directories("${CUDA_TOOLKIT_ROOT_DIR}/lib64")
  endif()

appveyor.yml

@@ -69,7 +69,7 @@ before_build:
 
  set OpenCV_DIR=%APPVEYOR_BUILD_FOLDER%/%MXNET_OPENCV_DIR%/build
 
- cmake .. -DOPENCV_DIR=%OpenCV_DIR% -DUSE_PROFILER=1 -DUSE_CUDA=0 -DUSE_CUDNN=0 -DUSE_NVRTC=0 -DUSE_OPENCV=1 -DUSE_OPENMP=1 -DUSE_BLAS=open -DUSE_LAPACK=1 -DUSE_DIST_KVSTORE=0 -G "Visual Studio 12 2013 Win64"
+ cmake .. -DOPENCV_DIR=%OpenCV_DIR% -DUSE_PROFILER=1 -DUSE_CUDA=0 -DUSE_CUDNN=0 -DENABLE_CUDA_RTC=0 -DUSE_OPENCV=1 -DUSE_OPENMP=1 -DUSE_BLAS=open -DUSE_LAPACK=1 -DUSE_DIST_KVSTORE=0 -G "Visual Studio 12 2013 Win64"
 
 build_script:
  - cmd: >-

ci/build_windows.py

@@ -54,7 +54,7 @@ class BuildFlavour(Enum):
  'WIN_CPU': (
  '-DUSE_CUDA=OFF '
  '-DUSE_CUDNN=OFF '
- '-DUSE_NVRTC=OFF '
+ '-DENABLE_CUDA_RTC=OFF '
  '-DUSE_OPENCV=ON '
  '-DUSE_OPENMP=ON '
  '-DUSE_BLAS=open '
@@ -67,7 +67,7 @@ class BuildFlavour(Enum):
  , 'WIN_CPU_MKLDNN': (
  '-DUSE_CUDA=OFF '
  '-DUSE_CUDNN=OFF '
- '-DUSE_NVRTC=OFF '
+ '-DENABLE_CUDA_RTC=OFF '
  '-DUSE_OPENCV=ON '
  '-DUSE_OPENMP=ON '
  '-DUSE_BLAS=open '
@@ -80,7 +80,7 @@ class BuildFlavour(Enum):
  , 'WIN_CPU_MKLDNN_MKL': (
  '-DUSE_CUDA=OFF '
  '-DUSE_CUDNN=OFF '
- '-DUSE_NVRTC=OFF '
+ '-DENABLE_CUDA_RTC=OFF '
  '-DUSE_OPENCV=ON '
  '-DUSE_OPENMP=ON '
  '-DUSE_BLAS=mkl '
@@ -93,7 +93,7 @@ class BuildFlavour(Enum):
  , 'WIN_CPU_MKL': (
  '-DUSE_CUDA=OFF '
  '-DUSE_CUDNN=OFF '
- '-DUSE_NVRTC=OFF '
+ '-DENABLE_CUDA_RTC=OFF '
  '-DUSE_OPENCV=ON '
  '-DUSE_OPENMP=ON '
  '-DUSE_BLAS=mkl '
@@ -106,7 +106,7 @@ class BuildFlavour(Enum):
  , 'WIN_GPU': (
  '-DUSE_CUDA=ON '
  '-DUSE_CUDNN=ON '
- '-DUSE_NVRTC=ON '
+ '-DENABLE_CUDA_RTC=ON '
  '-DUSE_OPENCV=ON '
  '-DUSE_OPENMP=ON '
  '-DUSE_BLAS=open '
@@ -122,7 +122,7 @@ class BuildFlavour(Enum):
  , 'WIN_GPU_MKLDNN': (
  '-DUSE_CUDA=ON '
  '-DUSE_CUDNN=ON '
- '-DUSE_NVRTC=ON '
+ '-DENABLE_CUDA_RTC=ON '
  '-DUSE_OPENCV=ON '
  '-DUSE_OPENMP=ON '
  '-DUSE_BLAS=open '

make/maven/maven_darwin_mkl.mk

@@ -77,7 +77,7 @@ USE_CUDNN = 0
 # CUDA_ARCH :=
 
 # whether use cuda runtime compiling for writing kernels in native language (i.e. Python)
-USE_NVRTC = 0
+ENABLE_CUDA_RTC = 0
 
 # use openmp for parallelization
 USE_OPENMP = 0

make/maven/maven_linux_cu90mkl.mk

@@ -80,7 +80,7 @@ USE_NCCL = 1
 
 # whether use cuda runtime compiling for writing kernels in native language (i.e. Python)
 USE_NVTX=1
-USE_NVRTC = 1
+ENABLE_CUDA_RTC = 1
 
 # use openmp for parallelization
 USE_OPENMP = 1

make/maven/maven_linux_cu92mkl.mk

@@ -80,7 +80,7 @@ USE_NCCL = 1
 
 # whether use cuda runtime compiling for writing kernels in native language (i.e. Python)
 USE_NVTX=1
-USE_NVRTC = 1
+ENABLE_CUDA_RTC = 1
 
 # use openmp for parallelization
 USE_OPENMP = 1

make/maven/maven_linux_mkl.mk

@@ -76,7 +76,7 @@ USE_CUDNN = 0
 # CUDA_ARCH :=
 
 # whether use cuda runtime compiling for writing kernels in native language (i.e. Python)
-USE_NVRTC = 0
+ENABLE_CUDA_RTC = 0
 
 # use openmp for parallelization
 USE_OPENMP = 1

src/executor/exec_pass.h

@@ -221,6 +221,11 @@ Graph FusePointwiseForward(Graph&& g);
  */
 Graph FusePointwiseBackward(Graph&& g);
 
+/*!
+ * \brief Issue a one-time warning that fusion is not possible for this platform or build.
+ */
+void WarnFusionNotSupported();
+
 /*!
  * \brief Infer shapes in the graph given the information.
  * \param graph The input graph.

src/executor/graph_executor.cc

@@ -50,7 +50,7 @@ static const std::string GetDefaultSubgraphBackend() {
 #endif
 }
 
-GraphExecutor::GraphExecutor() {
+GraphExecutor::GraphExecutor(const nnvm::Symbol& symbol) {
  log_verbose_ = dmlc::GetEnv("MXNET_EXEC_VERBOSE_LOGGING", false);
  need_grad_ = false;
  is_dynamic_ = false;
@@ -60,6 +60,7 @@ GraphExecutor::GraphExecutor() {
  LOG(INFO) << "MXNET_SUBGRAPH_BACKEND=NONE is detected, subgraph backend is not in use";
  }
  engine_ref_ = Engine::_GetSharedRef();
+ symbol_ = symbol.Copy();
 }
 
 GraphExecutor::~GraphExecutor() {
@@ -890,10 +891,9 @@ Executor* GraphExecutor::Reshape(const bool partial_shaping,
  std::vector<NDArray>* arg_grads,
  std::vector<NDArray>* aux_states) {
  nnvm::Graph g;
- g.outputs = std::vector<nnvm::NodeEntry>(graph_.outputs.begin(),
- graph_.outputs.begin() + num_forward_outputs_);
  nnvm::Symbol symbol;
- symbol.outputs = g.outputs;
+ symbol.outputs = symbol_.outputs;
+ g.outputs = symbol_.outputs;
  const nnvm::IndexedGraph& idx = g.indexed_graph();
  mxnet::ShapeVector arg_shapes(idx.input_nodes().size(), mxnet::TShape());
  for (size_t i = 0; i < num_forward_inputs_; ++i) {
@@ -977,8 +977,8 @@ Executor* GraphExecutor::Reshape(const bool partial_shaping,
  }
  }
  }
- auto exec = new GraphExecutor();
- exec->Init(symbol, default_ctx, ctx_map,
+ auto exec = new GraphExecutor(symbol);
+ exec->Init(symbol.Copy(), default_ctx, ctx_map,
  *in_args, *arg_grads, grad_req_types, *aux_states,
  this);
  return exec;
@@ -1001,7 +1001,7 @@ Graph GraphExecutor::InitGraph(nnvm::Symbol symbol,
  // setup gradient
  nnvm::Graph g = InitFullGraph(symbol, grad_req_types);
 
-#if MXNET_USE_CUDA && !defined(_WIN32)
+#if MXNET_USE_CUDA && MXNET_ENABLE_CUDA_RTC && !defined(_WIN32)
  if (default_ctx.dev_mask() == Context::kGPU && dmlc::GetEnv("MXNET_USE_FUSION", true)) {
  nnvm::Graph unoptimized_graph;
  common::CopyGraph(&unoptimized_graph, g, false);
@@ -1034,7 +1034,12 @@ Graph GraphExecutor::InitGraph(nnvm::Symbol symbol,
  << "Graph contains duplicate names for some of its inputs - fusion is NOT enabled!";
  }
  }
-#endif // MXNET_USE_CUDA
+#else
+ // Only warn user if MXNET_USE_FUSION env var is explicitly set
+ if (default_ctx.dev_mask() == Context::kGPU && dmlc::GetEnv("MXNET_USE_FUSION", false)) {
+ WarnFusionNotSupported();
+ }
+#endif // MXNET_USE_CUDA && MXNET_ENABLE_CUDA_RTC && !defined(_WIN32)
 
  // create "device" and "context" attrs for the graph
  g = AssignContext(g, default_ctx, ctx_map,
@@ -1969,7 +1974,7 @@ Executor *Executor::SimpleBind(nnvm::Symbol symbol,
  std::vector<NDArray>* aux_states,
  std::unordered_map<std::string, NDArray>* shared_buffer,
  Executor* shared_exec) {
- auto exec = new exec::GraphExecutor();
+ auto exec = new exec::GraphExecutor(symbol);
  bool init = false;
  if (!exec->subgraph_property().empty()) {
  static int verbose = dmlc::GetEnv("MXNET_SUBGRAPH_VERBOSE", 1);
@@ -1989,6 +1994,8 @@ Executor *Executor::SimpleBind(nnvm::Symbol symbol,
  symbol = exec::BuildSubgraph(symbol, backend, arg_shape_map, arg_dtype_map, arg_stype_map,
  default_ctx, group2ctx, &tmp_in_arg_ctxes, &tmp_arg_grad_ctxes,
  &tmp_grad_req_types, &tmp_aux_state_ctxes, verbose);
+ // Subgraph cannot be recreated from unoptimized symbol
+ exec = new exec::GraphExecutor(symbol);
  exec->Init(symbol.Copy(), default_ctx, group2ctx, tmp_in_arg_ctxes, tmp_arg_grad_ctxes,
  tmp_aux_state_ctxes, arg_shape_map, arg_dtype_map, arg_stype_map,
  tmp_grad_req_types, shared_arg_names, &tmp_in_args, &tmp_arg_grads,
@@ -2043,7 +2050,7 @@ Executor *Executor::Bind(nnvm::Symbol symbol,
  const std::vector<OpReqType> &grad_req_type,
  const std::vector<NDArray> &aux_states,
  Executor* shared_exec) {
- auto exec = new exec::GraphExecutor();
+ auto exec = new exec::GraphExecutor(symbol);
  static int verbose = dmlc::GetEnv("MXNET_SUBGRAPH_VERBOSE", 1);
  std::vector<NDArray> tmp_in_args = in_args;
  std::vector<NDArray> tmp_arg_grad_store = arg_grad_store;
@@ -2058,6 +2065,8 @@ Executor *Executor::Bind(nnvm::Symbol symbol,
  symbol = exec::BuildSubgraph(symbol, backend, default_ctx, group2ctx, &tmp_in_args,
  &tmp_arg_grad_store, &tmp_grad_req_type, &tmp_aux_states,
  verbose);
+ // Subgraph cannot be recreated from unoptimized symbol
+ exec = new exec::GraphExecutor(symbol);
  }
  }
  exec->Init(symbol.Copy(), default_ctx, group2ctx, tmp_in_args, tmp_arg_grad_store,

src/executor/graph_executor.h

@@ -58,7 +58,7 @@ class GraphExecutor : public Executor {
  public:
  using Executor::MonitorCallback;
 
- GraphExecutor();
+ explicit GraphExecutor(const nnvm::Symbol& symbol);
  virtual ~GraphExecutor();
  void Forward(bool is_train) override;
  void PartialForward(bool is_train, int step, int *step_left) override;
@@ -267,6 +267,9 @@ class GraphExecutor : public Executor {
  std::string subgraph_property_;
  // ref of engine
  std::shared_ptr<Engine> engine_ref_;
+ // Unoptimized copy of the symbol for sharing with
+ // child executors
+ nnvm::Symbol symbol_;
 };
 
 } // namespace exec