From 3c9dbc5fcb1ba55678462035e802ac61c72721a0 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Tom=C3=A1=C5=A1=20Oberhuber?= <oberhuber.tomas@gmail.com>
Date: Wed, 27 Jan 2021 14:54:43 +0100
Subject: [PATCH] Added new CSR segments kernels to SpMV benchmark.
---
src/Benchmarks/SpMV/spmv-legacy.h | 54 +++++++++++--------
.../CSRAdaptiveKernelBlockDescriptor.h | 2 +-
2 files changed, 34 insertions(+), 22 deletions(-)
diff --git a/src/Benchmarks/SpMV/spmv-legacy.h b/src/Benchmarks/SpMV/spmv-legacy.h
index 3416ad3eff..fed37410cf 100644
--- a/src/Benchmarks/SpMV/spmv-legacy.h
+++ b/src/Benchmarks/SpMV/spmv-legacy.h
@@ -49,7 +49,16 @@ using SlicedEllpackAlias = Matrices::Legacy::SlicedEllpack< Real, Device, Index
// Segments based sparse matrix aliases
template< typename Real, typename Device, typename Index >
-using SparseMatrix_CSR = Matrices::SparseMatrix< Real, Device, Index, Matrices::GeneralMatrix, Algorithms::Segments::CSRDefault >;
+using SparseMatrix_CSR_Scalar = Matrices::SparseMatrix< Real, Device, Index, Matrices::GeneralMatrix, Algorithms::Segments::CSRScalar >;
+
+template< typename Real, typename Device, typename Index >
+using SparseMatrix_CSR_Vector = Matrices::SparseMatrix< Real, Device, Index, Matrices::GeneralMatrix, Algorithms::Segments::CSRVector >;
+
+template< typename Real, typename Device, typename Index >
+using SparseMatrix_CSR_Hybrid = Matrices::SparseMatrix< Real, Device, Index, Matrices::GeneralMatrix, Algorithms::Segments::CSRHybrid >;
+
+template< typename Real, typename Device, typename Index >
+using SparseMatrix_CSR_Adaptive = Matrices::SparseMatrix< Real, Device, Index, Matrices::GeneralMatrix, Algorithms::Segments::CSRAdaptive >;
template< typename Device, typename Index, typename IndexAllocator >
using EllpackSegments = Algorithms::Segments::Ellpack< Device, Index, IndexAllocator >;
@@ -309,26 +318,29 @@ benchmarkSpmvSynthetic( Benchmark& benchmark,
benchmark.time< Devices::Cuda >( resetCusparseVectors, "GPU", spmvCusparse, cusparseBenchmarkResults );
#endif
- benchmarkSpMV< Real, SparseMatrixLegacy_CSR_Scalar >( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, SparseMatrixLegacy_CSR_Vector >( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, SparseMatrixLegacy_CSR_Light >( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, SparseMatrixLegacy_CSR_Light2 >( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, SparseMatrixLegacy_CSR_Light3 >( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, SparseMatrixLegacy_CSR_Light4 >( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, SparseMatrixLegacy_CSR_Light5 >( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, SparseMatrixLegacy_CSR_Light6 >( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, SparseMatrixLegacy_CSR_Adaptive >( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, SparseMatrixLegacy_CSR_MultiVector>( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, SparseMatrixLegacy_CSR_LightWithoutAtomic>( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, SparseMatrix_CSR >( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, Matrices::Legacy::Ellpack >( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, SparseMatrix_Ellpack >( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, SlicedEllpackAlias >( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, SparseMatrix_SlicedEllpack >( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, Matrices::Legacy::ChunkedEllpack >( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, SparseMatrix_ChunkedEllpack >( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, Matrices::Legacy::BiEllpack >( benchmark, hostOutVector, inputFileName, verboseMR );
- benchmarkSpMV< Real, SparseMatrix_BiEllpack >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrixLegacy_CSR_Scalar >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrixLegacy_CSR_Vector >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrixLegacy_CSR_Light >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrixLegacy_CSR_Light2 >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrixLegacy_CSR_Light3 >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrixLegacy_CSR_Light4 >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrixLegacy_CSR_Light5 >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrixLegacy_CSR_Light6 >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrixLegacy_CSR_Adaptive >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrixLegacy_CSR_MultiVector >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrixLegacy_CSR_LightWithoutAtomic >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrix_CSR_Scalar >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrix_CSR_Vector >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrix_CSR_Hybrid >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrix_CSR_Adaptive >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, Matrices::Legacy::Ellpack >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrix_Ellpack >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SlicedEllpackAlias >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrix_SlicedEllpack >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, Matrices::Legacy::ChunkedEllpack >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrix_ChunkedEllpack >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, Matrices::Legacy::BiEllpack >( benchmark, hostOutVector, inputFileName, verboseMR );
+ benchmarkSpMV< Real, SparseMatrix_BiEllpack >( benchmark, hostOutVector, inputFileName, verboseMR );
/* AdEllpack is broken
benchmarkSpMV< Real, Matrices::AdEllpack >( benchmark, hostOutVector, inputFileName, verboseMR );
*/
diff --git a/src/TNL/Algorithms/Segments/details/CSRAdaptiveKernelBlockDescriptor.h b/src/TNL/Algorithms/Segments/details/CSRAdaptiveKernelBlockDescriptor.h
index 90f8a7bfcc..96f1899b26 100644
--- a/src/TNL/Algorithms/Segments/details/CSRAdaptiveKernelBlockDescriptor.h
+++ b/src/TNL/Algorithms/Segments/details/CSRAdaptiveKernelBlockDescriptor.h
@@ -185,7 +185,7 @@ struct CSRAdaptiveKernelBlockDescriptor
//return ( twobytes[ sizeof( Index ) == 4 ? 3 : 5 ] & 0x3FFF );
}
- __cuda_callable__ const uint8_t getWarpIdx() const
+ __cuda_callable__ uint8_t getWarpIdx() const
{
return this->warpIdx;
}
--
GitLab