Commit 5ac2ba24 authored by Tomáš Oberhuber's avatar Tomáš Oberhuber Committed by Jakub Klinkovský
Browse files

Fixing CSR Light kernel.

parent b9603592

src/Benchmarks/scripts/run-tnl-benchmark-spmv

+1 −1
Original line number Diff line number Diff line
@@ -13,7 +13,7 @@ BENCHMARK_DBG="tnl-benchmark-spmv-dbg --with-legacy-matrices no" 
export CUDA_PROFILE_CONFIG="$IWD/cuda-profiler.conf"

PROCESS_CUDA_PROFILE="$IWD/process-cuda-profile.pl"

source matrix-market

M_MATRICES=""

#MM_MATRICES=""

source florida-matrix-market

#FLORIDA_MM_MATRICES=""

src/Benchmarks/scripts/tnl-spmv-benchmark-make-tables-json.py

+1 −1
Original line number Diff line number Diff line
@@ -254,7 +254,7 @@ formats = get_formats( input_df ) 
multicolumns, df_data = get_multiindex( input_df, formats )



print( "Converting data..." )

result = convert_data_frame( input_df, multicolumns, df_data, 200 )

result = convert_data_frame( input_df, multicolumns, df_data, 20000 )

compute_speedup( result, formats )



print( "Writting to HTML file..." )

src/TNL/Algorithms/Segments/Kernels/CSRLightKernel.hpp

+8 −17
Original line number Diff line number Diff line
@@ -275,12 +275,8 @@ struct CSRLightKernelreduceSegmentsDispatcher< Index, Device, Fetch, Reduce, Kee 
#ifdef HAVE_CUDA

      const int threads = 128;

      Index blocks, groupSize;

      //if (KernelType == CSRLightWithoutAtomic)

      int  neededThreads = threadsPerSegment * ( last - first );

      //else

      //   neededThreads = rows * (threadsPerSegment > 32 ? 32 : threadsPerSegment);



      /* Execute kernels on device */

      for (Index grid = 0; neededThreads != 0; ++grid)

      {

         if( TNL::Cuda::getMaxGridXSize() * threads >= neededThreads)
@@ -334,30 +330,25 @@ void 
CSRLightKernel< Index, Device >::

init( const Offsets& offsets )

{

   //const Index elementsInSegment = std::ceil( ( double ) offsets.getElement( segmentsCount ) / ( double ) segmentsCount );

   //this->threadsPerSegment = TNL::min( std::pow( 2, std::ceil( std::log2( elementsInSegment ) ) ) ); //TNL::Cuda::getWarpSize() );



   const Index segmentsCount = offsets.getSize() - 1;

   //const Index threads = 128; // !!!!!!!!!!!!!!!!!!!!!! block size

   size_t neededThreads = segmentsCount * 32;//warpSize;

   Index blocks, threadsPerSegment;



   const Index elementsInSegment = roundUpDivision( offsets.getElement( segmentsCount ), segmentsCount ); // non zeroes per row

   if( elementsInSegment <= 2 )

      threadsPerSegment = 2;

      this->threadsPerSegment = 2;

   else if( elementsInSegment <= 4 )

      threadsPerSegment = 4;

      this->threadsPerSegment = 4;

   else if( elementsInSegment <= 8 )

      threadsPerSegment = 8;

      this->threadsPerSegment = 8;

   else if( elementsInSegment <= 16 )

      threadsPerSegment = 16;

      this->threadsPerSegment = 16;

   else //if (nnz <= 2 * matrix.MAX_ELEMENTS_PER_WARP)

      threadsPerSegment = 32; // CSR Vector

      this->threadsPerSegment = 32; // CSR Vector

   //else

   //   threadsPerSegment = roundUpDivision(nnz, matrix.MAX_ELEMENTS_PER_WARP) * 32; // CSR MultiVector



   TNL_ASSERT_GE( threadsPerSegment, 0, "" );

   TNL_ASSERT_LE( threadsPerSegment, 33, "" );

   TNL_ASSERT_GE( this->threadsPerSegment, 0, "" );

   TNL_ASSERT_LE( this->threadsPerSegment, 33, "" );



}