Commit 4196f915 authored by Illia Kolesnik's avatar Illia Kolesnik Committed by Tomáš Oberhuber
Browse files

Fixed block sizes for CSR Light, other improvements

parent 382eb38f

src/TNL/Matrices/Legacy/CSR.h

+6 −6
Original line number Diff line number Diff line
@@ -110,16 +110,16 @@ public: 
   /* Configuration of CSR SpMV kernels ----------------------------------------- */



   /* Block sizes */

   static constexpr Index THREADS_ADAPTIVE = 1024;

   static constexpr Index THREADS_SCALAR = 1024;

   static constexpr Index THREADS_VECTOR = 1024;

   static constexpr Index THREADS_LIGHT = 1024;

   static constexpr Index THREADS_ADAPTIVE = sizeof(Index) == 8 ? 128 : 256;

   static constexpr Index THREADS_SCALAR = 128;

   static constexpr Index THREADS_VECTOR = 128;

   static constexpr Index THREADS_LIGHT = 128;



   /* Max length of row to process one warp */

   static constexpr Index MAX_ELEMENTS_PER_WARP = 1024;

   static constexpr Index MAX_ELEMENTS_PER_WARP = 512;



   /* How many shared memory use per block in CSR Adaptive kernel */

   static constexpr Index SHARED_PER_BLOCK = 49152;

   static constexpr Index SHARED_PER_BLOCK = 24576;



   /* Number of elements in shared memory */

   static constexpr Index SHARED = SHARED_PER_BLOCK/sizeof(Real);

src/TNL/Matrices/Legacy/CSR_impl.h

+4 −4
Original line number Diff line number Diff line
@@ -134,7 +134,7 @@ Index findLimit(const Index start, 
            type = Type::STREAM;

            return current;

         } else {                  // one long row

            if (sum <= matrix.MAX_ELEMENTS_PER_WARP)

            if (sum <= 2 * matrix.MAX_ELEMENTS_PER_WARP)

               type = Type::VECTOR;

            else

               type = Type::LONG;
@@ -1407,7 +1407,7 @@ template< typename Real, 
void SpMVCSRLightPrepare( const Real *inVector,

                          Real* outVector,

                          const CSR< Real, Device, Index, KernelType >& matrix) {

   const Index threads = matrix.THREADS_LIGHT; // max block size

   const Index threads = 1024; // max block size

   const Index rows = matrix.getRowPointers().getSize() - 1;

   /* Copy rowCnt to GPU */

   unsigned rowCnt = 0;
@@ -1554,7 +1554,7 @@ void SpMVCSRLightWithoutAtomicPrepare( const Real *inVector, 
      groupSize = 8;

   else if (nnz <= 16)

      groupSize = 16;

   else if (nnz <= matrix.MAX_ELEMENTS_PER_WARP)

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

      groupSize = 32; // CSR Vector

   else

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