Merge pull request apache#17 from winstywang/dev

Dev
stefanhenneking · Dec 16, 2014 · f966f11 · f966f11
2 parents f2eb4c2 + c874225
commit f966f11
Show file tree

Hide file tree

Showing 5 changed files with 52 additions and 6 deletions.
diff --git a/example/Makefile b/example/Makefile
@@ -2,7 +2,7 @@
 export CC  = gcc
 export CXX = g++
 export NVCC =nvcc
-export CFLAGS = -Wall -O3 -msse3 -Wno-unknown-pragmas -funroll-loops -I../
+export CFLAGS = -Wall -O3 -msse3 -Wno-unknown-pragmas -funroll-loops -I../ -I/opt/intel/mkl/include -I/usr/local/cuda/include/ -L/opt/intel/mkl/lib/intel64 -L/opt/intel/lib/intel64 -L/usr/local/cuda/lib64
 export LDFLAGS= -lm -lcudart -lcublas -lmkl_core -lmkl_intel_lp64 -lmkl_intel_thread -liomp5 -lpthread 
 export NVCCFLAGS = -O3 --use_fast_math -ccbin $(CXX)
 

diff --git a/example/basic.cpp b/example/basic.cpp
@@ -35,6 +35,22 @@ int main( void ){
         }
         printf("\n");
     }
+
+    // create a tensor without explictly allocating spaces.
+    Tensor<cpu, 2> mat3 = NewTensor<cpu>(Shape2(2, 5), 0.0f);
+    Tensor<cpu, 2> mat4 = NewTensor<cpu>(Shape2(2, 5), 1.0f);
+    // transpose, and then add mat4.
+    mat3 = mat.T() + mat4;
+
+    // index the shape using size(), this is more natural for MATLAB/numpy user.
+    printf("%u X %u matrix\n", mat3.size(0), mat3.size(1) );
+    // print out matrix
+    for( index_t i = 0; i < mat3.size(0); i ++ ){
+        for( index_t j = 0; j < mat3.size(1); j ++ ){
+            printf("%.2f ", mat3[i][j]);
+        }
+        printf("\n");
+    }
     // shutdown tensor enigne after usage
     ShutdownTensorEngine();
     return 0;

diff --git a/mshadow/tensor.h b/mshadow/tensor.h
@@ -230,7 +230,7 @@ namespace mshadow {
          * \return the size
          */
         MSHADOW_XINLINE index_t size(index_t i) const {
-            return shape_[dimension - 1 - i];
+            return shape[dimension - 1 - i];
         }
         /*!
          * \brief flatten the tensor to 2 dimension, collapse the higher dimensions together
@@ -303,7 +303,7 @@ namespace mshadow {
                                      (dptr) + begin, s);
         }
         MSHADOW_XINLINE index_t size(index_t i) const {
-            return shape_[0];
+            return shape[0];
         }        
         MSHADOW_XINLINE real_t &operator[](index_t idx) { return dptr[ idx ]; }
         MSHADOW_XINLINE const real_t &operator[](index_t idx)const { return dptr[ idx ]; }
@@ -470,12 +470,12 @@ namespace mshadow{
 
 // execution implementation of expression evaluations
 #include "tensor_expr_engine-inl.hpp"
-// extension of expressions
-#include "tensor_expr_ext.h"
 // cpu implementation of functions
 #include "tensor_cpu-inl.hpp"
 // gpu implementation of functions
 #include "tensor_gpu-inl.hpp"
+// extension of expressions
+#include "tensor_expr_ext.h"
 // io 
 #include "tensor_io.h"
 // container

diff --git a/mshadow/tensor_expr.h b/mshadow/tensor_expr.h
@@ -68,7 +68,7 @@ namespace mshadow{
 
         /*! \brief represent a transpose expression of a container */
         template<typename EType>
-        struct TransposeExp: public Exp< TransposeExp<EType>, type::kComplex >{
+        struct TransposeExp: public Exp< TransposeExp<EType>, type::kMapper >{
         public:
             /*! \brief expression to be transposed */
             const EType &exp;

diff --git a/mshadow/tensor_expr_engine-inl.hpp b/mshadow/tensor_expr_engine-inl.hpp
@@ -111,6 +111,17 @@ namespace mshadow{
             Plan<SubType> src_;  
         };
 
+        template<typename EType>
+        class Plan< TransposeExp<EType> >{
+        public:
+            Plan( const Plan<EType> &src ):src_(src){}
+            MSHADOW_XINLINE real_t Eval( index_t y, index_t x ) const{
+                return src_.Eval( x, y );
+            }
+        private:
+            Plan<EType> src_;
+        };
+
         // allow UnaryMap see the plan
         template<typename OP, typename TA, typename TB, int etype>
         inline Plan< BinaryMapExp<OP,TA,TB,etype> > MakePlan( const BinaryMapExp<OP,TA,TB,etype> &e );
@@ -125,6 +136,11 @@ namespace mshadow{
             return Plan<T>( e.self() );
         }
 
+        template<typename T>
+        inline Plan<TransposeExp<T> > MakePlan( const TransposeExp<T> &e ){
+            return Plan<TransposeExp<T> >( MakePlan(e.exp) );
+        }
+
         template<typename T, typename SrcExp, int dim>
         inline Plan< T > MakePlan( const MakeTensorExp<T,SrcExp,dim> &e ){
             return Plan< T >( e.real_self() );
@@ -159,6 +175,11 @@ namespace mshadow{
             const static int kDim = 0;
             const static int kDevMask = 0xffff;
         };
+        template<typename E>
+        struct ExpInfo<TransposeExp< E > >{
+            const static int kDim = ExpInfo<E>::kDim;
+            const static int kDevMask = ExpInfo<E>::kDevMask;
+        };
         template<typename Device, int dim>
         struct ExpInfo< Tensor<Device,dim> >{
             const static int kDim = dim;
@@ -224,6 +245,15 @@ namespace mshadow{
                 return shape;
             }
         };
+        template<int dim, typename E>
+        struct ShapeCheck<dim, TransposeExp< E > >{
+            inline static Shape<dim> Check( const TransposeExp< E > &e ){
+                // swap the lowest two dimensions
+                Shape<dim> s = ShapeCheck<dim,E>::Check( e.exp ); 
+                std::swap(s[0], s[1]);
+                return s;
+            }
+        };
         template<int dim,typename Device>
         struct ShapeCheck<dim,Tensor<Device,dim> >{
             inline static Shape<dim> Check( const Tensor<Device,dim> &t ){