Commit 37aab443 authored by Jakub Klinkovský's avatar Jakub Klinkovský
Browse files

Refactored Algorithms/Atomic.h and added atomicAdd for long int via atomicCAS on CUDA

parent bd59c376

src/TNL/Algorithms/AtomicOperations.h

+6 −42
Original line number Original line Diff line number Diff line
@@ -8,19 +8,13 @@ 




#pragma once

#pragma once





#ifdef HAVE_CUDA

#include <TNL/Atomic.h>

   #include <cuda.h>

#endif

#include <TNL/Devices/Sequential.h>

#include <TNL/Devices/Host.h>

#include <TNL/Devices/Cuda.h>





namespace TNL {

namespace TNL {

namespace Algorithms {

namespace Algorithms {





template< typename Device >

template< typename Device >

struct AtomicOperations

struct AtomicOperations;

{};





template<>

template<>

struct AtomicOperations< Devices::Host >

struct AtomicOperations< Devices::Host >
@@ -72,40 +66,9 @@ struct AtomicOperations< Devices::Cuda > 
   {

   {

#ifdef HAVE_CUDA

#ifdef HAVE_CUDA

      return atomicAdd( &v, a );

      return atomicAdd( &v, a );

#endif

#else

   }



#ifdef HAVE_CUDA

   __device__

   static double

   add( double& v, const double& a )

   {

   #if __CUDA_ARCH__ < 600

      unsigned long long int* v_as_ull = (unsigned long long int*) &v;

      unsigned long long int old = *v_as_ull, assumed;



      do {

         assumed = old;

         old = atomicCAS( v_as_ull, assumed, __double_as_longlong( a + __longlong_as_double( assumed ) ) );



         // Note: uses integer comparison to avoid hang in case of NaN (since NaN != NaN)

      } while( assumed != old );



      return old;

   #else   // __CUDA_ARCH__ < 600

      return atomicAdd( &v, a );

   #endif  //__CUDA_ARCH__ < 600

   }

#endif  // HAVE_CUDA



   __cuda_callable__

   static long int

   add( long int& v, const long int& a )

   {

#ifdef HAVE_CUDA

      TNL_ASSERT_TRUE( false, "Atomic add for long int is not supported on CUDA." );

#endif  // HAVE_CUDA

      return 0;

      return 0;

#endif

   }

   }





   __cuda_callable__

   __cuda_callable__
@@ -114,9 +77,10 @@ struct AtomicOperations< Devices::Cuda > 
   {

   {

#ifdef HAVE_CUDA

#ifdef HAVE_CUDA

      TNL_ASSERT_TRUE( false, "Atomic add for short int is not supported on CUDA." );

      TNL_ASSERT_TRUE( false, "Atomic add for short int is not supported on CUDA." );

#endif  // HAVE_CUDA

#endif

      return 0;

      return 0;

   }

   }

};

};



}  // namespace Algorithms

}  // namespace Algorithms

}  // namespace TNL
 
}  // namespace TNL

src/TNL/Atomic.h

+25 −4
Original line number Original line Diff line number Diff line
@@ -14,11 +14,12 @@ 
#include <TNL/Devices/Sequential.h>

#include <TNL/Devices/Sequential.h>

#include <TNL/Devices/Cuda.h>

#include <TNL/Devices/Cuda.h>





// double-precision atomicAdd function for Maxwell and older GPUs

// copied from: https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#atomic-functions

#ifdef HAVE_CUDA

#ifdef HAVE_CUDA

   #if __CUDA_ARCH__ < 600

namespace {

namespace {



// double-precision atomicAdd function for Maxwell and older GPUs

// copied from: https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#atomic-functions

   #if defined( __CUDA_ARCH__ ) && __CUDA_ARCH__ < 600

__device__

__device__

double

double

atomicAdd( double* address, double val )

atomicAdd( double* address, double val )
@@ -35,8 +36,28 @@ atomicAdd( double* address, double val ) 




   return __longlong_as_double( old );

   return __longlong_as_double( old );

}

}

}  // namespace

   #endif

   #endif



__device__

long int

atomicAdd( long int* address, long int val )

{

   unsigned long long int* address_as_unsigned = reinterpret_cast< unsigned long long int* >( address );

   long int old = *address;

   long int assumed;



   do {

      assumed = old;

      long int sum = val + assumed;

      old = atomicCAS( address_as_unsigned,

                       *reinterpret_cast< unsigned long long int* >( &assumed ),

                       *reinterpret_cast< unsigned long long int* >( &sum ) );

   } while( assumed != old );



   return old;

}



}  // namespace

#endif

#endif





namespace TNL {

namespace TNL {