Port multi-GPU to NCCL, add python support

Makefile

@@ -328,6 +328,12 @@ ifeq ($(USE_CUDNN), 1)
 	COMMON_FLAGS += -DUSE_CUDNN
 endif
 
+# NCCL acceleration configuration
+ifeq ($(USE_NCCL), 1)
+	LIBRARIES += nccl
+	COMMON_FLAGS += -DUSE_NCCL
+endif
+
 # configure IO libraries
 ifeq ($(USE_OPENCV), 1)
 	COMMON_FLAGS += -DUSE_OPENCV
@@ -446,9 +452,20 @@ endif
 	py mat py$(PROJECT) mat$(PROJECT) proto runtest \
 	superclean supercleanlist supercleanfiles warn everything
 
+ifeq ($(CPU_ONLY), 1)
+ifeq ($(USE_NCCL), 1)
+checks: $(error Cannot define USE_NCCL with CPU_ONLY)
+endif
+ifeq ($(USE_CUDNN), 1)
+checks: $(error Cannot define USE_CUDNN with CPU_ONLY)
+endif
+endif
+checks:
+
 all: lib tools examples
 
-lib: $(STATIC_NAME) $(DYNAMIC_NAME)
+lib:	checks \
+	$(STATIC_NAME) $(DYNAMIC_NAME)
 
 everything: $(EVERYTHING_TARGETS)
 
@@ -495,7 +512,8 @@ examples: $(EXAMPLE_BINS)
 
 py$(PROJECT): py
 
-py: $(PY$(PROJECT)_SO) $(PROTO_GEN_PY)
+py:	checks \
+	$(PY$(PROJECT)_SO) $(PROTO_GEN_PY)
 
 $(PY$(PROJECT)_SO): $(PY$(PROJECT)_SRC) $(PY$(PROJECT)_HXX) | $(DYNAMIC_NAME)
 	@ echo CXX/LD -o $@ $<

Makefile.config.example

@@ -94,6 +94,16 @@ LIBRARY_DIRS := $(PYTHON_LIB) /usr/local/lib /usr/lib
 # INCLUDE_DIRS += $(shell brew --prefix)/include
 # LIBRARY_DIRS += $(shell brew --prefix)/lib
 
+# NCCL acceleration switch (uncomment to build with NCCL)
+# E.g. setup:
+#   cd
+#   git clone https://github.com/NVIDIA/nccl
+#   cd nccl
+#   make -j
+# USE_NCCL := 1
+# INCLUDE_DIRS += $(HOME)/nccl/src
+# LIBRARY_DIRS += $(HOME)/nccl/build/lib
+
 # Uncomment to use `pkg-config` to specify OpenCV library paths.
 # (Usually not necessary -- OpenCV libraries are normally installed in one of the above $LIBRARY_DIRS.)
 # USE_PKG_CONFIG := 1

include/caffe/blob.hpp

@@ -220,6 +220,7 @@ class Blob {
   void set_cpu_data(Dtype* data);
   const int* gpu_shape() const;
   const Dtype* gpu_data() const;
+  void set_gpu_data(Dtype* data);
   const Dtype* cpu_diff() const;
   const Dtype* gpu_diff() const;
   Dtype* mutable_cpu_data();

include/caffe/common.hpp

@@ -158,11 +158,14 @@ class Caffe {
   // Search from start_id to the highest possible device ordinal,
   // return the ordinal of the first available device.
   static int FindDevice(const int start_id = 0);
-  // Parallel training info
+  // Parallel training
   inline static int solver_count() { return Get().solver_count_; }
   inline static void set_solver_count(int val) { Get().solver_count_ = val; }
-  inline static bool root_solver() { return Get().root_solver_; }
-  inline static void set_root_solver(bool val) { Get().root_solver_ = val; }
+  inline static int solver_rank() { return Get().solver_rank_; }
+  inline static void set_solver_rank(int val) { Get().solver_rank_ = val; }
+  inline static bool multi_process() { return Get().multi_process_; }
+  inline static void set_multi_process(bool val) { Get().multi_process_ = val; }
+  inline static bool root_solver() { return Get().solver_rank_ == 0; }
 
  protected:
 #ifndef CPU_ONLY
@@ -172,8 +175,11 @@ class Caffe {
   shared_ptr<RNG> random_generator_;
 
   Brew mode_;
+
+  // Parallel training
   int solver_count_;
-  bool root_solver_;
+  int solver_rank_;
+  bool multi_process_;
 
  private:
   // The private constructor to avoid duplicate instantiation.

include/caffe/internal_thread.hpp

@@ -42,8 +42,8 @@ class InternalThread {
   bool must_stop();
 
  private:
-  void entry(int device, Caffe::Brew mode, int rand_seed, int solver_count,
-      bool root_solver);
+  void entry(int device, Caffe::Brew mode, int rand_seed,
+      int solver_count, int solver_rank, bool multi_process);
 
   shared_ptr<boost::thread> thread_;
 };

include/caffe/layer.hpp

@@ -38,7 +38,7 @@ class Layer {
    * layer.
    */
   explicit Layer(const LayerParameter& param)
-    : layer_param_(param), is_shared_(false) {
+    : layer_param_(param) {
       // Set phase and copy blobs (if there are any).
       phase_ = param.phase();
       if (layer_param_.blobs_size() > 0) {
@@ -66,7 +66,6 @@ class Layer {
    */
   void SetUp(const vector<Blob<Dtype>*>& bottom,
       const vector<Blob<Dtype>*>& top) {
-    InitMutex();
     CheckBlobCounts(bottom, top);
     LayerSetUp(bottom, top);
     Reshape(bottom, top);
@@ -92,30 +91,6 @@ class Layer {
   virtual void LayerSetUp(const vector<Blob<Dtype>*>& bottom,
       const vector<Blob<Dtype>*>& top) {}
 
-  /**
-   * @brief Whether a layer should be shared by multiple nets during data
-   *        parallelism. By default, all layers except for data layers should
-   *        not be shared. data layers should be shared to ensure each worker
-   *        solver access data sequentially during data parallelism.
-   */
-  virtual inline bool ShareInParallel() const { return false; }
-
-  /** @brief Return whether this layer is actually shared by other nets.
-   *         If ShareInParallel() is true and using more than one GPU and the
-   *         net has TRAIN phase, then this function is expected return true.
-   */
-  inline bool IsShared() const { return is_shared_; }
-
-  /** @brief Set whether this layer is actually shared by other nets
-   *         If ShareInParallel() is true and using more than one GPU and the
-   *         net has TRAIN phase, then is_shared should be set true.
-   */
-  inline void SetShared(bool is_shared) {
-    CHECK(ShareInParallel() || !is_shared)
-        << type() << "Layer does not support sharing.";
-    is_shared_ = is_shared;
-  }
-
   /**
    * @brief Adjust the shapes of top blobs and internal buffers to accommodate
    *        the shapes of the bottom blobs.
@@ -428,19 +403,6 @@ class Layer {
   }
 
  private:
-  /** Whether this layer is actually shared by other nets*/
-  bool is_shared_;
-
-  /** The mutex for sequential forward if this layer is shared */
-  shared_ptr<boost::mutex> forward_mutex_;
-
-  /** Initialize forward_mutex_ */
-  void InitMutex();
-  /** Lock forward_mutex_ if this layer is shared */
-  void Lock();
-  /** Unlock forward_mutex_ if this layer is shared */
-  void Unlock();
-
   DISABLE_COPY_AND_ASSIGN(Layer);
 };  // class Layer
 
@@ -450,8 +412,6 @@ class Layer {
 template <typename Dtype>
 inline Dtype Layer<Dtype>::Forward(const vector<Blob<Dtype>*>& bottom,
     const vector<Blob<Dtype>*>& top) {
-  // Lock during forward to ensure sequential forward
-  Lock();
   Dtype loss = 0;
   Reshape(bottom, top);
   switch (Caffe::mode()) {
@@ -482,7 +442,6 @@ inline Dtype Layer<Dtype>::Forward(const vector<Blob<Dtype>*>& bottom,
   default:
     LOG(FATAL) << "Unknown caffe mode.";
   }
-  Unlock();
   return loss;
 }
 

include/caffe/layers/base_data_layer.hpp

@@ -68,15 +68,16 @@ class BasePrefetchingDataLayer :
       const vector<Blob<Dtype>*>& top);
 
   // Prefetches batches (asynchronously if to GPU memory)
-  static const int PREFETCH_COUNT = 3;
+  static const int PREFETCH_COUNT = 4;  // same as proto
 
  protected:
   virtual void InternalThreadEntry();
   virtual void load_batch(Batch<Dtype>* batch) = 0;
 
-  Batch<Dtype> prefetch_[PREFETCH_COUNT];
+  vector<shared_ptr<Batch<Dtype> > > prefetch_;
   BlockingQueue<Batch<Dtype>*> prefetch_free_;
   BlockingQueue<Batch<Dtype>*> prefetch_full_;
+  Batch<Dtype>* prefetch_current_;
 
   Blob<Dtype> transformed_data_;
 };

include/caffe/layers/data_layer.hpp

@@ -4,7 +4,6 @@
 #include <vector>
 
 #include "caffe/blob.hpp"
-#include "caffe/data_reader.hpp"
 #include "caffe/data_transformer.hpp"
 #include "caffe/internal_thread.hpp"
 #include "caffe/layer.hpp"
@@ -29,9 +28,13 @@ class DataLayer : public BasePrefetchingDataLayer<Dtype> {
   virtual inline int MaxTopBlobs() const { return 2; }
 
  protected:
+  void Next();
+  bool Skip();
   virtual void load_batch(Batch<Dtype>* batch);
 
-  DataReader reader_;
+  shared_ptr<db::DB> db_;
+  shared_ptr<db::Cursor> cursor_;
+  uint64_t skip_counter_;
 };
 
 }  // namespace caffe

include/caffe/layers/python_layer.hpp

@@ -21,8 +21,8 @@ class PythonLayer : public Layer<Dtype> {
     // Disallow PythonLayer in MultiGPU training stage, due to GIL issues
     // Details: https://github.com/BVLC/caffe/issues/2936
     if (this->phase_ == TRAIN && Caffe::solver_count() > 1
-        && !ShareInParallel()) {
-      LOG(FATAL) << "PythonLayer is not implemented in Multi-GPU training";
+        && !Caffe::root_solver() && !Caffe::multi_process()) {
+      LOG(FATAL) << "PythonLayer does not support CLI Multi-GPU, use train.py";
     }
     self_.attr("param_str") = bp::str(
         this->layer_param_.python_param().param_str());

include/caffe/net.hpp

@@ -23,10 +23,9 @@ namespace caffe {
 template <typename Dtype>
 class Net {
  public:
-  explicit Net(const NetParameter& param, const Net* root_net = NULL);
+  explicit Net(const NetParameter& param);
   explicit Net(const string& param_file, Phase phase,
-      const int level = 0, const vector<string>* stages = NULL,
-      const Net* root_net = NULL);
+      const int level = 0, const vector<string>* stages = NULL);
   virtual ~Net() {}
 
   /// @brief Initialize a network with a NetParameter.
@@ -228,6 +227,31 @@ class Net {
   static bool StateMeetsRule(const NetState& state, const NetStateRule& rule,
       const string& layer_name);
 
+  // Invoked at specific points during an iteration
+  class Callback {
+   protected:
+    virtual void run(int layer) = 0;
+
+    template <typename T>
+    friend class Net;
+  };
+  const vector<Callback*>& before_forward() const { return before_forward_; }
+  void add_before_forward(Callback* value) {
+    before_forward_.push_back(value);
+  }
+  const vector<Callback*>& after_forward() const { return after_forward_; }
+  void add_after_forward(Callback* value) {
+    after_forward_.push_back(value);
+  }
+  const vector<Callback*>& before_backward() const { return before_backward_; }
+  void add_before_backward(Callback* value) {
+    before_backward_.push_back(value);
+  }
+  const vector<Callback*>& after_backward() const { return after_backward_; }
+  void add_after_backward(Callback* value) {
+    after_backward_.push_back(value);
+  }
+
  protected:
   // Helpers for Init.
   /// @brief Append a new top blob to the net.
@@ -306,9 +330,13 @@ class Net {
   size_t memory_used_;
   /// Whether to compute and display debug info for the net.
   bool debug_info_;
-  /// The root net that actually holds the shared layers in data parallelism
-  const Net* const root_net_;
-  DISABLE_COPY_AND_ASSIGN(Net);
+  // Callbacks
+  vector<Callback*> before_forward_;
+  vector<Callback*> after_forward_;
+  vector<Callback*> before_backward_;
+  vector<Callback*> after_backward_;
+
+DISABLE_COPY_AND_ASSIGN(Net);
 };