elems op, rdiv/rsub, unary, set value

Makefile

@@ -12,9 +12,6 @@ ifndef DMLC_CORE
 	DMLC_CORE = dmlc-core
 endif
 
-ifndef RABIT
-	RABIT = rabit
-endif
 
 # use customized config file
 include $(config)
@@ -65,15 +62,14 @@ endif
 ENGINE=naive_engine.o
 BIN = tests/test_simple_engine
 OBJ = narray_function_cpu.o
-OBJCXX11 = batch_norm_cpu.o reshape_cpu.o narray.o c_api.o operator.o symbol.o storage.o fully_connected_cpu.o static_graph.o activation_cpu.o graph_executor.o softmax_cpu.o elementwise_sum_cpu.o pooling_cpu.o convolution_cpu.o io.o iter_mnist.o $(ENGINE)
-CUOBJ =
+OBJCXX11 = narray.o c_api.o operator.o symbol.o storage.o static_graph.o graph_executor.o io.o iter_mnist.o $(ENGINE)
+CUOBJ = narray_function_gpu.o
 SLIB = lib/libmxnet.so
 ALIB = lib/libmxnet.a
 LIB_DEP = $(DMLC_CORE)/libdmlc.a
-
-ifeq ($(USE_CUDA), 1)
-	CUOBJ += batch_norm_gpu.o reshape_gpu.o narray_function_gpu.o fully_connected_gpu.o activation_gpu.o elementwise_sum_gpu.o pooling_gpu.o softmax_gpu.o convolution_gpu.o
-endif
+ALL_DEP = $(OBJ) $(OBJCXX11) $(LIB_DEP)
+# common headers, change them will results in rebuild of all files
+COMMON_HEADERS=include/mxnet/*.h src/common/*.h
 
 .PHONY: clean all test lint doc
 
@@ -85,47 +81,47 @@ $(DMLC_CORE)/libdmlc.a:
 storage.o: src/storage/storage.cc
 naive_engine.o:  src/dag_engine/naive_engine.cc
 dag_engine.o: src/dag_engine/dag_engine.cc
-simple_engine.o: src/dag_engine/simple_engine.cc 
+simple_engine.o: src/dag_engine/simple_engine.cc
 narray.o: src/narray/narray.cc
 narray_function_cpu.o: src/narray/narray_function.cc src/narray/narray_function-inl.h
 narray_function_gpu.o: src/narray/narray_function.cu src/narray/narray_function-inl.h
-symbol.o: src/symbol/symbol.cc
-graph_executor.o: src/symbol/graph_executor.cc
-static_graph.o : src/symbol/static_graph.cc
+symbol.o: src/symbol/symbol.cc src/symbol/*.h
+graph_executor.o: src/symbol/graph_executor.cc src/symbol/*.h
+static_graph.o : src/symbol/static_graph.cc src/symbol/*.h
 operator.o: src/operator/operator.cc
 c_api.o: src/c_api.cc
-fully_connected_cpu.o: src/operator/fully_connected.cc
-fully_connected_gpu.o: src/operator/fully_connected.cu
-activation_cpu.o: src/operator/activation.cc
-activation_gpu.o: src/operator/activation.cu
-elementwise_sum_cpu.o: src/operator/elementwise_sum.cc
-elementwise_sum_gpu.o: src/operator/elementwise_sum.cu
-pooling_cpu.o: src/operator/pooling.cc
-pooling_gpu.o: src/operator/pooling.cu
-softmax_cpu.o: src/operator/softmax.cc
-softmax_gpu.o: src/operator/softmax.cu
-convolution_cpu.o: src/operator/convolution.cc
-convolution_gpu.o: src/operator/convolution.cu
-reshape_cpu.o: src/operator/reshape.cc
-reshape_gpu.o: src/operator/reshape.cu
-batch_norm_cpu.o: src/operator/batch_norm.cc
-batch_norm_gpu.o: src/operator/batch_norm.cu
 io.o: src/io/io.cc
-iter_mnist.o: src/io/iter_mnist.cc
+iter_mnist.o: src/io/iter_mnist.cc src/io/*.h
 
-lib/libmxnet.a: $(OBJ) $(OBJCXX11) $(CUOBJ) $(LIB_DEP)
-lib/libmxnet.so: $(OBJ) $(OBJCXX11) $(CUOBJ) $(LIB_DEP)
+# Rules for operators
+OPERATOR_HDR=$(wildcard src/operator/*-inl.h)
+OPERATOR_OBJ=$(patsubst %-inl.h, %_cpu.o, $(OPERATOR_HDR))
+OPERATOR_CUOBJ=$(patsubst %-inl.h, %_gpu.o, $(OPERATOR_HDR))
+
+ALL_DEP += $(OPERATOR_OBJ)
+ifeq ($(USE_CUDA), 1)
+	ALL_DEP += $(OPERATOR_CUOBJ) $(CUOBJ)
+endif
+
+src/operator/%_cpu.o : src/operator/%.cc src/operator/%-inl.h src/operator/mshadow_op.h src/operator/operator_common.h $(COMMON_HEADERS)
+	$(CXX) -std=c++0x -c $(CFLAGS) -o $@  $(filter %.cpp %.c %.cc, $^)
+
+src/operator/%_gpu.o : src/operator/%.cu src/operator/%-inl.h src/operator/operator_common.h src/operator/mshadow_op.h $(COMMON_HEADERS)
+	$(NVCC) -c -o $@ $(NVCCFLAGS) -Xcompiler "$(CFLAGS)" $(filter %.cu, $^)
+
+lib/libmxnet.a:  $(ALL_DEP)
+lib/libmxnet.so: $(ALL_DEP)
 
 tests/test_storage: tests/test_storage.cc lib/libmxnet.a
 tests/test_simple_engine: tests/test_simple_engine.cc lib/libmxnet.a
 
 $(BIN) :
 	$(CXX) $(CFLAGS) -std=c++0x -o $@ $(filter %.cpp %.o %.c %.a %.cc, $^) $(LDFLAGS)
 
-$(OBJ) :
+$(OBJ) : $(COMMON_HEADERS)
 	$(CXX) -c $(CFLAGS) -o $@ $(filter %.cpp %.c %.cc, $^)
 
-$(OBJCXX11) :
+$(OBJCXX11) : $(COMMON_HEADERS)
 	$(CXX) -std=c++0x -c $(CFLAGS) -o $@  $(filter %.cpp %.c %.cc, $^)
 
 $(SLIB) :
@@ -134,19 +130,18 @@ $(SLIB) :
 $(ALIB): $(OBJ) $(OBJCXX11)
 	ar cr $@ $+
 
-$(CUOBJ) :
+$(CUOBJ) :$(COMMON_HEADERS)
 	$(NVCC) -c -o $@ $(NVCCFLAGS) -Xcompiler "$(CFLAGS)" $(filter %.cu, $^)
 
 $(CUBIN) :
 	$(NVCC) -o $@ $(NVCCFLAGS) -Xcompiler "$(CFLAGS)" -Xlinker "$(LDFLAGS)" $(filter %.cu %.cpp %.o, $^)
 
-
 lint:
 	python dmlc-core/scripts/lint.py mxnet ${LINT_LANG} include src scripts test python
 
 doxygen:
 	doxygen doc/Doxyfile
 
 clean:
-	$(RM) $(OBJ) $(OBJCXX11) $(BIN) $(CUBIN) $(CUOBJ) $(SLIB) $(ALIB) *~ */*~ */*/*~ */*/*/*~
+	$(RM) $(ALL_DEP) $(SLIB) $(ALIB) *~ */*~ */*/*~ */*/*/*~
 	cd $(DMLC_CORE); make clean; cd -

include/mxnet/narray.h

@@ -269,8 +269,9 @@ class NArray {
   friend void BinaryOp(const NArray &lhs, const NArray &rhs, NArray *out);
   template<typename OP>
   friend void UnaryOp(const NArray &lhs, const NArray &rhs, NArray *out);
-  template<typename OP>
+  template<typename OP, bool reverse>
   friend void ScalarOp(const NArray &lhs, const real_t &rhs, NArray *out);
+  friend void SetValueOp(const real_t &rhs, NArray *out);
 };
 
 /*!
@@ -385,6 +386,23 @@ struct NArrayFunctionReg
         num_mutate_vars(0),
         num_scalars(0),
         type_mask(0) {}
+  /*!
+   * \brief set the function body to a NArray setvalue function
+   *  this will also auto set the parameters correctly
+   * \param fsetvalue function body to set
+   * \return ref to the registered entry, used to set properties
+   */
+  inline NArrayFunctionReg &set_function(void fsetvalue(const real_t &rhs,
+                                                        NArray *out)) {
+    body = [fsetvalue] (NArray **used_vars,
+                       real_t *s, NArray **mutate_vars) {
+      fsetvalue(s[0], mutate_vars[0]);
+    };
+    num_mutate_vars = 1; num_scalars = 1;
+    // type_mask = kNArrayArgBeforeScalar;
+    this->add_argument("rhs", "real_t", "Right operand to the function.");
+    return *this;
+  }
   /*!
    * \brief set the function body to a binary NArray function
    *  this will also auto set the parameters correctly

include/mxnet/operator.h

@@ -80,7 +80,7 @@ class Operator {
    * \param req the request types of saving operation, can only be kWriteTo or kWriteInplace.
    * \param out_data array of output data, pointer is used to indicate that this is holder
    *        the space of TBlob in out_data must be pre-allocated with InferShape
-   * \param aux_states Auxiliary states of operator. Normally operator doesn't 
+   * \param aux_states Auxiliary states of operator. Normally operator doesn't
    *        need, epecial case like Batch Norm requires.
    * \sa OpReqType, OpContext
    */
@@ -411,11 +411,12 @@ struct OperatorPropertyReg
  * \endcode
  */
 #define MXNET_REGISTER_OP_PROPERTY(name, OperatorPropertyType)          \
-  static ::mxnet::OperatorProperty* __create__ ## OperatorPropertyType ## __() { \
+  static ::mxnet::OperatorProperty* __create__ ## OperatorProperty ## name ## __() { \
     return new OperatorPropertyType;                                    \
   }                                                                     \
   DMLC_REGISTRY_REGISTER(::mxnet::OperatorPropertyReg, OperatorPropertyReg, name) \
-  .set_body(__create__ ## OperatorPropertyType ## __)
+  .set_body(__create__ ## OperatorProperty ## name ## __)
+
 #endif  // DMLC_USE_CXX11
 }  // namespace mxnet
 #endif  // MXNET_OPERATOR_H_

python/mxnet/base.py

@@ -15,11 +15,13 @@
 #----------------------------
 if sys.version_info[0] == 3:
     string_types = str,
+    numeric_types = (float, int)
     # this function is needed for python3
     # to convert ctypes.char_p .value back to python str
     py_str = lambda x: x.decode('utf-8')
 else:
     string_types = basestring,
+    numeric_types = (float, int, long)
     py_str = lambda x: x
 
 

python/mxnet/narray.py

@@ -3,8 +3,9 @@
 from __future__ import absolute_import
 
 import ctypes
+import warnings
 import sys
-from .base import _LIB, string_types
+from .base import _LIB, string_types, numeric_types
 from .base import c_array, py_str, c_str
 from .base import mx_uint, mx_float, NArrayHandle, FunctionHandle
 from .base import ctypes2numpy_shared, ctypes2buffer
@@ -66,15 +67,15 @@ def __del__(self):
     def __add__(self, other):
         if isinstance(other, NArray):
             return NArray._plus(self, other)
-        elif isinstance(other, float) or isinstance(other, int):
+        elif isinstance(other, numeric_types):
             return NArray._plus_scalar(self, float(other))
         else:
             raise TypeError('type %s not supported' % str(type(other)))
 
     def __iadd__(self, other):
         if isinstance(other, NArray):
             return NArray._plus(self, other, out=self)
-        elif isinstance(other, float) or isinstance(other, int):
+        elif isinstance(other, numeric_types):
             return NArray._plus_scalar(self, float(other), out=self)
         else:
             raise TypeError('type %s not supported' % str(type(other)))
@@ -85,31 +86,40 @@ def __radd__(self, other):
     def __sub__(self, other):
         if isinstance(other, NArray):
             return NArray._minus(self, other)
-        elif isinstance(other, float) or isinstance(other, int):
+        elif isinstance(other, numeric_types):
             return NArray._minus_scalar(self, float(other))
         else:
             raise TypeError('type %s not supported' % str(type(other)))
 
     def __isub__(self, other):
         if isinstance(other, NArray):
             return NArray._minus(self, other, out=self)
-        elif isinstance(other, float) or isinstance(other, int):
+        elif isinstance(other, numeric_types):
             return NArray._minus_scalar(self, float(other), out=self)
         else:
             raise TypeError('type %s not supported' % str(type(other)))
 
+    def __rsub__(self, other):
+        if isinstance(other, numeric_types):
+            return NArray._rminus_scalar(self, float(other))
+        else:
+            raise TypeError('type %s not supported' % str(type(other)))
+
     def __mul__(self, other):
         if isinstance(other, NArray):
             return NArray._mul(self, other)
-        elif isinstance(other, float) or isinstance(other, int):
+        elif isinstance(other, numeric_types):
             return NArray._mul_scalar(self, float(other))
         else:
             raise TypeError('type %s not supported' % str(type(other)))
 
+    def __neg__(self):
+        return NArray._mul_scalar(self, -1.0, out=self)
+
     def __imul__(self, other):
         if isinstance(other, NArray):
             return NArray._mul(self, other, out=self)
-        elif isinstance(other, float) or isinstance(other, int):
+        elif isinstance(other, numeric_types):
             return NArray._mul_scalar(self, float(other), out=self)
         else:
             raise TypeError('type %s not supported' % str(type(other)))
@@ -120,15 +130,21 @@ def __rmul__(self, other):
     def __div__(self, other):
         if isinstance(other, NArray):
             return NArray._div(self, other)
-        elif isinstance(other, float) or isinstance(other, int):
+        elif isinstance(other, numeric_types):
             return NArray._div_scalar(self, float(other))
         else:
             raise TypeError('type %s not supported' % str(type(other)))
 
+    def __rdiv__(self, other):
+        if isinstance(other, numeric_types):
+            return NArray._rdiv_scalar(self, float(other))
+        else:
+            raise TypeError('type %s not supported' % str(type(other)))
+
     def __idiv__(self, other):
         if isinstance(other, NArray):
             return NArray._div(self, other, out=self)
-        elif isinstance(other, float) or isinstance(other, int):
+        elif isinstance(other, numeric_types):
             return NArray._div_scalar(self, float(other), out=self)
         else:
             raise TypeError('type %s not supported' % str(type(other)))
@@ -163,9 +179,13 @@ def __setitem__(self, in_slice, value):
         """Set narray value"""
         if in_slice.step != None:
             raise Exception("Set NArray should use empty index array[:] = target_array")
-        if isinstance(value, NArray) == False:
+        if isinstance(value, NArray):
+            if value.handle is not self.handle:
+                value.copyto(self)
+        elif isinstance(value, numeric_types):
+            return NArray._set_value(float(value), out=self)
+        else:
             raise TypeError('type %s not supported' % str(type(value)))
-        value.copyto(self)
 
     def __getitem__(self, in_slice):
         """Get narray"""
@@ -238,6 +258,10 @@ def copyto(self, other):
             The copy target NArray
         """
         if isinstance(other, NArray):
+            if other.handle is self.handle:
+                warnings.warn('copy an array to itself, is it intended?',
+                              RuntimeWarning)
+                return
             return NArray._copyto(self, out=other)
         elif isinstance(other, Context):
             hret = NArray(_new_alloc_handle(self.shape, other, True))