Commit eb3ed5b2 authored by Tomáš Oberhuber's avatar Tomáš Oberhuber
Browse files

Fixing SPMV benchmark.

parent ea47d656

src/matrices/tnlCSRMatrix_impl.h

+1 −1
Original line number Diff line number Diff line
@@ -672,7 +672,7 @@ void tnlCSRMatrix< Real, Device, Index >::spmvCudaVectorized( const InVector& in 
         if( inWarpIdx < 2 ) aux[ threadIdx.x ] += aux[ threadIdx.x + 2 ];

      if( warpSize >= 2 )

         if( inWarpIdx < 1 ) aux[ threadIdx.x ] += aux[ threadIdx.x + 1 ];

      __syncthreads(); // TODO: I am not sure why

      __syncthreads(); // TODO: I am not sure why - aux must be volatile



      if( inWarpIdx == 0 )

         outVector[ row ] = aux[ threadIdx.x ];

tests/benchmarks/tnl-benchmark-spmv.h

+2 −2
Original line number Diff line number Diff line
@@ -393,7 +393,7 @@ bool setupBenchmark( const tnlParameterContainer& parameters ) 
         cusparseDestroy( cusparseHandle );



         cout << " done.   \r";

         cudaCSRMatrix.setCudaKernelType( CSRMatrixCudaType::scalar );

         /*cudaCSRMatrix.setCudaKernelType( CSRMatrixCudaType::scalar );

         benchmarkMatrix( cudaCSRMatrix,

                          cudaX,

                          cudaB,
@@ -524,7 +524,7 @@ bool setupBenchmark( const tnlParameterContainer& parameters ) 
                          stopTime,

                          baseline,

                          verbose,

                          logFile );

                          logFile );*/

      }

      cudaCSRMatrix.reset();

#endif