CudaReductionKernel.h

/***************************************************************************
                          CudaReductionKernel.h  -  description
                             -------------------
    begin                : Jun 17, 2015
    copyright            : (C) 2015 by Tomas Oberhuber et al.
    email                : tomas.oberhuber@fjfi.cvut.cz
 ***************************************************************************/

/* See Copyright Notice in tnl/Copyright */

#pragma once

#ifdef HAVE_CUDA
#include <cuda.h>
#endif

#include <TNL/Assert.h>
#include <TNL/Math.h>
#include <TNL/Devices/CudaDeviceInfo.h>
#include <TNL/Containers/Algorithms/CudaReductionBuffer.h>

namespace TNL {
namespace Containers {
namespace Algorithms {

#ifdef HAVE_CUDA
/****
 * The performance of this kernel is very sensitive to register usage.
 * Compile with --ptxas-options=-v and configure these constants for given
 * architecture so that there are no local memory spills.
 */
static constexpr int Reduction_maxThreadsPerBlock = 256;  // must be a power of 2
static constexpr int Reduction_registersPerThread = 32;   // empirically determined optimal value

// __CUDA_ARCH__ is defined only in device code!
#if (__CUDA_ARCH__ >= 300 )
   static constexpr int Reduction_minBlocksPerMultiprocessor = 8;
#else
   static constexpr int Reduction_minBlocksPerMultiprocessor = 4;
#endif

template< int blockSize,
   typename Result,
   typename DataFetcher,
   typename Reduction,
   typename VolatileReduction,
   typename Index >
__global__ void
__launch_bounds__( Reduction_maxThreadsPerBlock, Reduction_minBlocksPerMultiprocessor )
CudaReductionKernel( const Result zero,
                     const DataFetcher dataFetcher,
                     const Reduction reduction,
                     const VolatileReduction volatileReduction,
                     const Index size,
                     Result* output )
{
   using IndexType = Index;
   using ResultType = Result;

   ResultType* sdata = Devices::Cuda::getSharedMemory< ResultType >();

   /***
    * Get thread id (tid) and global thread id (gid).
    * gridSize is the number of element processed by all blocks at the
    * same time.
    */
   const IndexType tid = threadIdx.x;
         IndexType gid = blockIdx.x * blockDim. x + threadIdx.x;
   const IndexType gridSize = blockDim.x * gridDim.x;

   sdata[ tid ] = zero;

   /***
    * Read data into the shared memory. We start with the
    * sequential reduction.
    */
   while( gid + 4 * gridSize < size )
   {
      reduction( sdata[ tid ], dataFetcher( gid ) );
      reduction( sdata[ tid ], dataFetcher( gid + gridSize ) );
      reduction( sdata[ tid ], dataFetcher( gid + 2 * gridSize ) );
      reduction( sdata[ tid ], dataFetcher( gid + 3 * gridSize ) );
      gid += 4 * gridSize;
   }
   while( gid + 2 * gridSize < size )
   {
      reduction( sdata[ tid ], dataFetcher( gid ) );
      reduction( sdata[ tid ], dataFetcher( gid + gridSize ) );
      gid += 2 * gridSize;
   }
   while( gid < size )
   {
      reduction( sdata[ tid ], dataFetcher( gid ) );
      gid += gridSize;
   }
   __syncthreads();

   //printf( "1: tid %d data %f \n", tid, sdata[ tid ] );
   /***
    *  Perform the parallel reduction.
    */
   if( blockSize >= 1024 )
   {
      if( tid < 512 )
         reduction( sdata[ tid ], sdata[ tid + 512 ] );
      __syncthreads();
   }
   if( blockSize >= 512 )
   {
      if( tid < 256 )
         reduction( sdata[ tid ], sdata[ tid + 256 ] );
      __syncthreads();
   }
   if( blockSize >= 256 )
   {
      if( tid < 128 )
         reduction( sdata[ tid ], sdata[ tid + 128 ] );
      __syncthreads();
      //printf( "2: tid %d data %f \n", tid, sdata[ tid ] );
   }

   if( blockSize >= 128 )
   {
      if( tid <  64 )
         reduction( sdata[ tid ], sdata[ tid + 64 ] );
      __syncthreads();
      //printf( "3: tid %d data %f \n", tid, sdata[ tid ] );
   }

   /***
    * This runs in one warp so it is synchronized implicitly.
    */
   if( tid < 32 )
   {
      volatile ResultType* vsdata = sdata;
      if( blockSize >= 64 )
      {
         volatileReduction( vsdata[ tid ], vsdata[ tid + 32 ] );
         //printf( "4: tid %d data %f \n", tid, sdata[ tid ] );
      }
      // TODO: If blocksize == 32, the following does not work
      // We do not check if tid < 16. Fix it!!!
      if( blockSize >= 32 )
      {
         volatileReduction( vsdata[ tid ], vsdata[ tid + 16 ] );
         //printf( "5: tid %d data %f \n", tid, sdata[ tid ] );
      }
      if( blockSize >= 16 )
      {
         volatileReduction( vsdata[ tid ], vsdata[ tid + 8 ] );
         //printf( "6: tid %d data %f \n", tid, sdata[ tid ] );
      }
      if( blockSize >=  8 )
      {
         volatileReduction( vsdata[ tid ], vsdata[ tid + 4 ] );
         //printf( "7: tid %d data %f \n", tid, sdata[ tid ] );
      }
      if( blockSize >=  4 )
      {
         volatileReduction( vsdata[ tid ], vsdata[ tid + 2 ] );
         //printf( "8: tid %d data %f \n", tid, sdata[ tid ] );
      }
      if( blockSize >=  2 )
      {
         volatileReduction( vsdata[ tid ], vsdata[ tid + 1 ] );
         //printf( "9: tid %d data %f \n", tid, sdata[ tid ] );
      }
   }

   /***
    * Store the result back in the global memory.
    */
   if( tid == 0 )
   {
      //printf( "Block %d result = %f \n", blockIdx.x, sdata[ 0 ] );
      output[ blockIdx.x ] = sdata[ 0 ];
   }

}

template< typename Index,
          typename Result >
struct CudaReductionKernelLauncher
{
   using IndexType = Index;
   using ResultType = Result;

   ////
   // The number of blocks should be a multiple of the number of multiprocessors
   // to ensure optimum balancing of the load. This is very important, because
   // we run the kernel with a fixed number of blocks, so the amount of work per
   // block increases with enlarging the problem, so even small imbalance can
   // cost us dearly.
   // Therefore,  desGridSize = blocksPerMultiprocessor * numberOfMultiprocessors
   // where blocksPerMultiprocessor is determined according to the number of
   // available registers on the multiprocessor.
   // On Tesla K40c, desGridSize = 8 * 15 = 120.
   //
   // Update:
   // It seems to be better to map only one CUDA block per one multiprocessor or maybe
   // just slightly more. Therefore we omit blocksdPerMultiprocessor in the following.
   CudaReductionKernelLauncher( const Index size )
   : activeDevice( Devices::CudaDeviceInfo::getActiveDevice() ),
     blocksdPerMultiprocessor( Devices::CudaDeviceInfo::getRegistersPerMultiprocessor( activeDevice )
                               / ( Reduction_maxThreadsPerBlock * Reduction_registersPerThread ) ),
     //desGridSize( blocksdPerMultiprocessor * Devices::CudaDeviceInfo::getCudaMultiprocessors( activeDevice ) ),
     desGridSize( Devices::CudaDeviceInfo::getCudaMultiprocessors( activeDevice ) ),
     originalSize( size )
   {
   }

   template< typename DataFetcher,
             typename Reduction,
             typename VolatileReduction >
   int start( const Reduction& reduction,
              const VolatileReduction& volatileReduction,
              const DataFetcher& dataFetcher,
              const Result& zero,
              ResultType*& output )
   {
      ////
      // create reference to the reduction buffer singleton and set size
      const size_t buf_size = 2 * desGridSize * sizeof( ResultType );
      CudaReductionBuffer& cudaReductionBuffer = CudaReductionBuffer::getInstance();
      cudaReductionBuffer.setSize( buf_size );
      output = cudaReductionBuffer.template getData< ResultType >();

      this-> reducedSize = this->launch( originalSize, reduction, volatileReduction, dataFetcher, zero, output );
      return this->reducedSize;
   }

   template< typename Reduction,
             typename VolatileReduction >
   Result finish( const Reduction& reduction,
                  const VolatileReduction& volatileReduction,
                  const Result& zero )
   {
      ////
      // Input is the first half of the buffer, output is the second half
      const size_t buf_size = desGridSize * sizeof( ResultType );
      CudaReductionBuffer& cudaReductionBuffer = CudaReductionBuffer::getInstance();
      ResultType* input = cudaReductionBuffer.template getData< ResultType >();
      ResultType* output = &input[ buf_size ];

      auto copyFetch = [=] __cuda_callable__ ( IndexType i ) { return input[ i ]; };
      while( this->reducedSize > 1 )
      {
         this-> reducedSize = this->launch( this->reducedSize, reduction, volatileReduction, copyFetch, zero, output );
         std::swap( input, output );
      }

      ////
      // Copy result on CPU
      ResultType result;
      ArrayOperations< Devices::Host, Devices::Cuda >::copyMemory( &result, output, 1 );
      return result;
   }

   protected:
      template< typename DataFetcher,
                typename Reduction,
                typename VolatileReduction >
      int launch( const Index size,
                  const Reduction& reduction,
                  const VolatileReduction& volatileReduction,
                  const DataFetcher& dataFetcher,
                  const Result& zero,
                  Result* output )
      {
         dim3 blockSize, gridSize;
         blockSize.x = Reduction_maxThreadsPerBlock;
         gridSize.x = TNL::min( Devices::Cuda::getNumberOfBlocks( size, blockSize.x ), desGridSize );

         ////
         // when there is only one warp per blockSize.x, we need to allocate two warps
         // worth of shared memory so that we don't index shared memory out of bounds
         const IndexType shmem = (blockSize.x <= 32)
                  ? 2 * blockSize.x * sizeof( ResultType )
                  : blockSize.x * sizeof( ResultType );

         /***
          * Depending on the blockSize we generate appropriate template instance.
          */
         switch( blockSize.x )
         {
            case 512:
               CudaReductionKernel< 512 >
               <<< gridSize, blockSize, shmem >>>( zero, dataFetcher, reduction, volatileReduction, size, output);
               break;
            case 256:
               cudaFuncSetCacheConfig(CudaReductionKernel< 256, Result, DataFetcher, Reduction, VolatileReduction, Index >, cudaFuncCachePreferShared);

               CudaReductionKernel< 256 >
               <<< gridSize, blockSize, shmem >>>( zero, dataFetcher, reduction, volatileReduction, size, output);
               break;
            case 128:
               cudaFuncSetCacheConfig(CudaReductionKernel< 128, Result, DataFetcher, Reduction, VolatileReduction, Index >, cudaFuncCachePreferShared);

               CudaReductionKernel< 128 >
               <<< gridSize, blockSize, shmem >>>( zero, dataFetcher, reduction, volatileReduction, size, output);
               break;
            case  64:
               cudaFuncSetCacheConfig(CudaReductionKernel<  64, Result, DataFetcher, Reduction, VolatileReduction, Index >, cudaFuncCachePreferShared);

               CudaReductionKernel<  64 >
               <<< gridSize, blockSize, shmem >>>( zero, dataFetcher, reduction, volatileReduction, size, output);
               break;
            case  32:
               cudaFuncSetCacheConfig(CudaReductionKernel<  32, Result, DataFetcher, Reduction, VolatileReduction, Index >, cudaFuncCachePreferShared);

               CudaReductionKernel<  32 >
               <<< gridSize, blockSize, shmem >>>( zero, dataFetcher, reduction, volatileReduction, size, output);
               break;
            case  16:
               cudaFuncSetCacheConfig(CudaReductionKernel<  16, Result, DataFetcher, Reduction, VolatileReduction, Index >, cudaFuncCachePreferShared);

               CudaReductionKernel<  16 >
               <<< gridSize, blockSize, shmem >>>( zero, dataFetcher, reduction, volatileReduction, size, output);
               break;
            case   8:
               cudaFuncSetCacheConfig(CudaReductionKernel<   8, Result, DataFetcher, Reduction, VolatileReduction, Index >, cudaFuncCachePreferShared);

               CudaReductionKernel<   8 >
               <<< gridSize, blockSize, shmem >>>( zero, dataFetcher, reduction, volatileReduction, size, output);
               break;
            case   4:
               cudaFuncSetCacheConfig(CudaReductionKernel<   4, Result, DataFetcher, Reduction, VolatileReduction, Index >, cudaFuncCachePreferShared);

               CudaReductionKernel<   4 >
               <<< gridSize, blockSize, shmem >>>( zero, dataFetcher, reduction, volatileReduction, size, output);
               break;
            case   2:
               cudaFuncSetCacheConfig(CudaReductionKernel<   2, Result, DataFetcher, Reduction, VolatileReduction, Index >, cudaFuncCachePreferShared);

               CudaReductionKernel<   2 >
               <<< gridSize, blockSize, shmem >>>( zero, dataFetcher, reduction, volatileReduction, size, output);
               break;
            case   1:
               throw std::logic_error( "blockSize should not be 1." );
            default:
               throw std::logic_error( "Block size is " + std::to_string(blockSize.x) + " which is none of 1, 2, 4, 8, 16, 32, 64, 128, 256 or 512." );
         }
         TNL_CHECK_CUDA_DEVICE;

         ////
         // return the size of the output array on the CUDA device
         return gridSize.x;
      }

      const int activeDevice;
      const int blocksdPerMultiprocessor;
      const int desGridSize;
      const IndexType originalSize;
      IndexType reducedSize;
};
#endif

} // namespace Algorithms
} // namespace Containers
} // namespace TNL