Commit e5020801 authored by Tomáš Oberhuber's avatar Tomáš Oberhuber Committed by Tomáš Oberhuber
Browse files

Added check against cross device memory acces to Array and ArrayView.

parent 178431c5

src/TNL/Algorithms/CudaScanKernel.h

+1 −1
Original line number Diff line number Diff line
@@ -277,7 +277,7 @@ struct CudaScanKernelLauncher 
              elementsInBlock,

              &deviceInput[ gridOffset ],

              &deviceOutput[ gridOffset ],

              &blockSums[ gridIdx * maxGridSize() ] );

              &blockSums.getData()[ gridIdx * maxGridSize() ] );

      }



      // synchronize the null-stream after all grids

src/TNL/Algorithms/Scan.hpp

+5 −5
Original line number Diff line number Diff line
@@ -225,8 +225,8 @@ perform( Vector& v, 


   CudaScanKernelLauncher< Type, RealType, IndexType >::perform(

      end - begin,

      &v[ begin ],  // input

      &v[ begin ],  // output

      &v.getData()[ begin ],  // input

      &v.getData()[ begin ],  // output

      reduction,

      zero );

#else
@@ -251,8 +251,8 @@ performFirstPhase( Vector& v, 


   return CudaScanKernelLauncher< Type, RealType, IndexType >::performFirstPhase(

      end - begin,

      &v[ begin ],  // input

      &v[ begin ],  // output

      &v.getData()[ begin ],  // input

      &v.getData()[ begin ],  // output

      reduction,

      zero );

#else
@@ -279,7 +279,7 @@ performSecondPhase( Vector& v, 


   CudaScanKernelLauncher< Type, RealType, IndexType >::performSecondPhase(

      end - begin,

      &v[ begin ],  // output

      &v.getData()[ begin ],  // output

      blockShifts.getData(),

      reduction,

      shift );

src/TNL/Containers/Array.h

+8 −2
Original line number Diff line number Diff line
@@ -446,7 +446,10 @@ class Array 
       * to the memory space where the array was allocated. For example, if the

       * array was allocated in the host memory, it can be called only from

       * host, and if the array was allocated in the device memory, it can be

       * called only from device kernels.

       * called only from device kernels. If NDEBUG is not defined, assertions

       * inside this methods performs runtime checks for cross-device memory

       * accesses which lead to segmentation fault. If you need to do just a 

       * pointer arithmetics use \e getData instead.

       *

       * \param i The index of the element to be accessed.

       * \return Reference to the \e i-th element.
@@ -460,7 +463,10 @@ class Array 
       * to the memory space where the array was allocated. For example, if the

       * array was allocated in the host memory, it can be called only from

       * host, and if the array was allocated in the device memory, it can be

       * called only from device kernels.

       * called only from device kernels. If NDEBUG is not defined, assertions

       * inside this methods performs runtime checks for cross-device memory

       * accesses which lead to segmentation fault. If you need to do just a 

       * pointer arithmetics use \e getData instead.

       *

       * \param i The index of the element to be accessed.

       * \return Constant reference to the \e i-th element.

src/TNL/Containers/Array.hpp

+12 −0
Original line number Diff line number Diff line
@@ -509,6 +509,12 @@ Value& 
Array< Value, Device, Index, Allocator >::

operator[]( const Index& i )

{

#ifdef __CUDA_ARCH__

   TNL_ASSERT( ( std::is_same< Device, Devices::Cuda >::value ), "Attempt to access data not allocated on CUDA device from CUDA device." );

#else

   TNL_ASSERT( ( std::is_same< Device, Devices::Host >::value || std::is_same< Device, Devices::Sequential >::value ),

               "Attempt to access data not allocated on the host from the host." );

#endif

   TNL_ASSERT_GE( i, (Index) 0, "Element index must be non-negative." );

   TNL_ASSERT_LT( i, this->getSize(), "Element index is out of bounds." );

   return this->data[ i ];
@@ -523,6 +529,12 @@ const Value& 
Array< Value, Device, Index, Allocator >::

operator[]( const Index& i ) const

{

#ifdef __CUDA_ARCH__

   TNL_ASSERT( ( std::is_same< Device, Devices::Cuda >::value ), "Attempt to access data not allocated on CUDA device from CUDA device." );

#else

   TNL_ASSERT( ( std::is_same< Device, Devices::Host >::value || std::is_same< Device, Devices::Sequential >::value ),

               "Attempt to access data not allocated on the host from the host." );

#endif

   TNL_ASSERT_GE( i, (Index) 0, "Element index must be non-negative." );

   TNL_ASSERT_LT( i, this->getSize(), "Element index is out of bounds." );

   return this->data[ i ];

src/TNL/Containers/ArrayView.h

+8 −2
Original line number Diff line number Diff line
@@ -347,7 +347,10 @@ public: 
    * to the memory space where the data was allocated. For example, if the

    * data was allocated in the host memory, it can be called only from

    * host, and if the data was allocated in the device memory, it can be

    * called only from device kernels.

    * called only from device kernels. If NDEBUG is not defined, assertions

    * inside this methods performs runtime checks for cross-device memory

    * accesses which lead to segmentation fault. If you need to do just a 

    * pointer arithmetics use \e getData instead.

    *

    * \param i The index of the element to be accessed.

    * \return Reference to the \e i-th element.
@@ -362,7 +365,10 @@ public: 
    * to the memory space where the data was allocated. For example, if the

    * data was allocated in the host memory, it can be called only from

    * host, and if the data was allocated in the device memory, it can be

    * called only from device kernels.

    * called only from device kernels. If NDEBUG is not defined, assertions

    * inside this methods performs runtime checks for cross-device memory

    * accesses which lead to segmentation fault. If you need to do just a 

    * pointer arithmetics use \e getData instead.

    *

    * \param i The index of the element to be accessed.

    * \return Constant reference to the \e i-th element.