mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2024-11-27 06:30:28 +08:00
Code Issues Packages Projects Releases Wiki Activity

remove the 1D and 2D parallelizer, keep only the GEMM specialized one

Browse Source
This commit is contained in:
Gael Guennebaud 2010-03-05 10:04:17 +01:00
parent 24ef5fedcd
commit d13b877014
2 changed files with 5 additions and 73 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
Eigen/src/Core/products/GeneralMatrixMatrix.h
View File

@ -273,11 +273,9 @@ class GeneralProduct<Lhs, Rhs, GemmProduct>
(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor>,
_ActualLhsType,
_ActualRhsType,
Dest> Functor;
Dest> GemmFunctor;
// ei_run_parallel_1d<true>(Functor(lhs, rhs, dst, actualAlpha), this->rows());
// ei_run_parallel_2d<true>(Functor(lhs, rhs, dst, actualAlpha), this->rows(), this->cols());
ei_run_parallel_gemm<true>(Functor(lhs, rhs, dst, actualAlpha), this->rows(), this->cols());
ei_parallelize_gemm<Dest::MaxRowsAtCompileTime>32>(GemmFunctor(lhs, rhs, dst, actualAlpha), this->rows(), this->cols());
}
};

72
Eigen/src/Core/products/Parallelizer.h
View File

@ -25,71 +25,6 @@
#ifndef EIGEN_PARALLELIZER_H
#define EIGEN_PARALLELIZER_H
template<bool Parallelize,typename Functor>
void ei_run_parallel_1d(const Functor& func, int size)
{
#ifndef EIGEN_HAS_OPENMP
func(0,size);
#else
if(!Parallelize)
return func(0,size);
int threads = omp_get_num_procs();
int blockSize = size / threads;
#pragma omp parallel for schedule(static,1)
for(int i=0; i<threads; ++i)
{
int blockStart = i*blockSize;
int actualBlockSize = std::min(blockSize, size - blockStart);
func(blockStart, actualBlockSize);
}
#endif
}
template<bool Parallelize,typename Functor>
void ei_run_parallel_2d(const Functor& func, int size1, int size2)
{
#ifndef EIGEN_HAS_OPENMP
func(0,size1, 0,size2);
#else
int threads = omp_get_max_threads();
if((!Parallelize)||(threads==1))
return func(0,size1, 0,size2);
// 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16
static const int divide1[17] = { 0, 1, 2, 3, 2, 5, 3, 7, 4, 3, 5, 1, 4, 1, 7, 5, 4};
static const int divide2[17] = { 0, 1, 1, 1, 2, 1, 2, 1, 2, 3, 2, 11, 3, 13, 2, 3, 4};
ei_assert(threads<=16 && "too many threads !");
int blockSize1 = size1 / divide1[threads];
int blockSize2 = size2 / divide2[threads];
Matrix<int,4,Dynamic> ranges(4,threads);
int k = 0;
for(int i1=0; i1<divide1[threads]; ++i1)
{
int blockStart1 = i1*blockSize1;
int actualBlockSize1 = std::min(blockSize1, size1 - blockStart1);
for(int i2=0; i2<divide2[threads]; ++i2)
{
int blockStart2 = i2*blockSize2;
int actualBlockSize2 = std::min(blockSize2, size2 - blockStart2);
ranges.col(k++) << blockStart1, actualBlockSize1, blockStart2, actualBlockSize2;
}
}
#pragma omp parallel for schedule(static,1)
for(int i=0; i<threads; ++i)
{
func(ranges.col(i)[0],ranges.col(i)[1],ranges.col(i)[2],ranges.col(i)[3]);
}
#endif
}
struct GemmParallelInfo
{
GemmParallelInfo() : sync(-1), users(0) {}
@ -102,18 +37,17 @@ struct GemmParallelInfo
float* blockB;
};
template<bool Parallelize,typename Functor>
void ei_run_parallel_gemm(const Functor& func, int rows, int cols)
template<bool Condition,typename Functor>
void ei_parallelize_gemm(const Functor& func, int rows, int cols)
{
#ifndef EIGEN_HAS_OPENMP
func(0,rows, 0,cols);
#else
int threads = omp_get_max_threads();
if((!Parallelize)||(threads==1))
if((!Condition)||(threads==1))
return func(0,rows, 0,cols);
int blockCols = (cols / threads) & ~0x3;
int blockRows = (rows / threads) & ~0x7;