mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-01-18 14:34:17 +08:00
Code Issues Packages Projects Releases Wiki Activity

* makes all product use the new API to set the blocking sizes

Browse Source 
* fix an issue preventing multithreading (now Dynamic = -1 ...)
This commit is contained in:
Gael Guennebaud 2010-06-22 16:08:35 +02:00
parent fd9a9fa0ae
commit b4fe53f561
7 changed files with 45 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
Eigen/src/Core/products/GeneralMatrixMatrix.h
View File

@ -284,7 +284,7 @@ class GeneralProduct<Lhs, Rhs, GemmProduct>
_ActualRhsType, _ActualRhsType,
Dest> GemmFunctor; Dest> GemmFunctor;
ei_parallelize_gemm<(Dest::MaxRowsAtCompileTime>32)>(GemmFunctor(lhs, rhs, dst, actualAlpha), this->rows(), this->cols()); ei_parallelize_gemm<(Dest::MaxRowsAtCompileTime>32 || Dest::MaxRowsAtCompileTime==Dynamic)>(GemmFunctor(lhs, rhs, dst, actualAlpha), this->rows(), this->cols());
} }
}; };

2
Eigen/src/Core/products/Parallelizer.h
View File

@ -29,7 +29,7 @@
inline void ei_manage_multi_threading(Action action, int* v) inline void ei_manage_multi_threading(Action action, int* v)
{ {
static int m_maxThreads = -1; static int m_maxThreads = -1;
if(action==SetAction) if(action==SetAction)
{ {
ei_internal_assert(v!=0); ei_internal_assert(v!=0);

12
Eigen/src/Core/products/SelfadjointMatrixMatrix.h
View File

@ -258,8 +258,10 @@ struct ei_product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,true,Conjugate
typedef ei_product_blocking_traits<Scalar> Blocking; typedef ei_product_blocking_traits<Scalar> Blocking;
Index kc = std::min<Index>(Blocking::Max_kc,size); // cache block size along the K direction Index kc = size; // cache block size along the K direction
Index mc = std::min<Index>(Blocking::Max_mc,rows); // cache block size along the M direction Index mc = rows; // cache block size along the M direction
Index nc = cols; // cache block size along the N direction
computeProductBlockingSizes<Scalar,Scalar>(kc, mc, nc);
Scalar* blockA = ei_aligned_stack_new(Scalar, kc*mc); Scalar* blockA = ei_aligned_stack_new(Scalar, kc*mc);
std::size_t sizeB = kc*Blocking::PacketSize*Blocking::nr + kc*cols; std::size_t sizeB = kc*Blocking::PacketSize*Blocking::nr + kc*cols;
@ -339,8 +341,10 @@ struct ei_product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,false,Conjugat
typedef ei_product_blocking_traits<Scalar> Blocking; typedef ei_product_blocking_traits<Scalar> Blocking;
Index kc = std::min<Index>(Blocking::Max_kc,size); // cache block size along the K direction Index kc = size; // cache block size along the K direction
Index mc = std::min<Index>(Blocking::Max_mc,rows); // cache block size along the M direction Index mc = rows; // cache block size along the M direction
Index nc = cols; // cache block size along the N direction
computeProductBlockingSizes<Scalar,Scalar>(kc, mc, nc);
Scalar* blockA = ei_aligned_stack_new(Scalar, kc*mc); Scalar* blockA = ei_aligned_stack_new(Scalar, kc*mc);
std::size_t sizeB = kc*Blocking::PacketSize*Blocking::nr + kc*cols; std::size_t sizeB = kc*Blocking::PacketSize*Blocking::nr + kc*cols;

8
Eigen/src/Core/products/SelfadjointProduct.h
View File

@ -70,14 +70,16 @@ struct ei_selfadjoint_product<Scalar, Index, MatStorageOrder, ColMajor, AAT, UpL
typedef ei_product_blocking_traits<Scalar> Blocking; typedef ei_product_blocking_traits<Scalar> Blocking;
Index kc = std::min<Index>(Blocking::Max_kc,depth); // cache block size along the K direction Index kc = depth; // cache block size along the K direction
Index mc = std::min<Index>(Blocking::Max_mc,size); // cache block size along the M direction Index mc = size; // cache block size along the M direction
Index nc = size; // cache block size along the N direction
computeProductBlockingSizes<Scalar,Scalar>(kc, mc, nc);
Scalar* blockA = ei_aligned_stack_new(Scalar, kc*mc); Scalar* blockA = ei_aligned_stack_new(Scalar, kc*mc);
std::size_t sizeB = kc*Blocking::PacketSize*Blocking::nr + kc*size; std::size_t sizeB = kc*Blocking::PacketSize*Blocking::nr + kc*size;
Scalar* allocatedBlockB = ei_aligned_stack_new(Scalar, sizeB); Scalar* allocatedBlockB = ei_aligned_stack_new(Scalar, sizeB);
Scalar* blockB = allocatedBlockB + kc*Blocking::PacketSize*Blocking::nr; Scalar* blockB = allocatedBlockB + kc*Blocking::PacketSize*Blocking::nr;
// note that the actual rhs is the transpose/adjoint of mat // note that the actual rhs is the transpose/adjoint of mat
typedef ei_conj_helper<NumTraits<Scalar>::IsComplex && !AAT, NumTraits<Scalar>::IsComplex && AAT> Conj; typedef ei_conj_helper<NumTraits<Scalar>::IsComplex && !AAT, NumTraits<Scalar>::IsComplex && AAT> Conj;

18
Eigen/src/Core/products/TriangularMatrixMatrix.h
View File

@ -114,8 +114,12 @@ struct ei_product_triangular_matrix_matrix<Scalar,Index,Mode,true,
IsLower = (Mode&Lower) == Lower IsLower = (Mode&Lower) == Lower
}; };
Index kc = std::min<Index>(Blocking::Max_kc/4,depth); // cache block size along the K direction Index kc = depth; // cache block size along the K direction
Index mc = std::min<Index>(Blocking::Max_mc,rows); // cache block size along the M direction Index mc = rows; // cache block size along the M direction
Index nc = cols; // cache block size along the N direction
computeProductBlockingSizes<Scalar,Scalar>(kc, mc, nc);
// it is better to use smaller blocks along the diagonal
kc /= 4;
Scalar* blockA = ei_aligned_stack_new(Scalar, kc*mc); Scalar* blockA = ei_aligned_stack_new(Scalar, kc*mc);
std::size_t sizeB = kc*Blocking::PacketSize*Blocking::nr + kc*cols; std::size_t sizeB = kc*Blocking::PacketSize*Blocking::nr + kc*cols;
@ -238,8 +242,12 @@ struct ei_product_triangular_matrix_matrix<Scalar,Index,Mode,false,
IsLower = (Mode&Lower) == Lower IsLower = (Mode&Lower) == Lower
}; };
Index kc = std::min<Index>(Blocking::Max_kc/4,depth); // cache block size along the K direction Index kc = depth; // cache block size along the K direction
Index mc = std::min<Index>(Blocking::Max_mc,rows); // cache block size along the M direction Index mc = rows; // cache block size along the M direction
Index nc = cols; // cache block size along the N direction
computeProductBlockingSizes<Scalar,Scalar>(kc, mc, nc);
// it is better to use smaller blocks along the diagonal
kc /= 4;
Scalar* blockA = ei_aligned_stack_new(Scalar, kc*mc); Scalar* blockA = ei_aligned_stack_new(Scalar, kc*mc);
std::size_t sizeB = kc*Blocking::PacketSize*Blocking::nr + kc*cols; std::size_t sizeB = kc*Blocking::PacketSize*Blocking::nr + kc*cols;
@ -273,7 +281,7 @@ struct ei_product_triangular_matrix_matrix<Scalar,Index,Mode,false,
Index rs = IsLower ? std::min(cols,actual_k2) : cols - k2; Index rs = IsLower ? std::min(cols,actual_k2) : cols - k2;
// size of the triangular part // size of the triangular part
Index ts = (IsLower && actual_k2>=cols) ? 0 : actual_kc; Index ts = (IsLower && actual_k2>=cols) ? 0 : actual_kc;
Scalar* geb = blockB+ts*ts; Scalar* geb = blockB+ts*ts;
pack_rhs(geb, &rhs(actual_k2,IsLower ? 0 : k2), rhsStride, alpha, actual_kc, rs); pack_rhs(geb, &rhs(actual_k2,IsLower ? 0 : k2), rhsStride, alpha, actual_kc, rs);

19
Eigen/src/Core/products/TriangularSolverMatrix.h
View File

@ -63,8 +63,12 @@ struct ei_triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conjugate,TriStora
IsLower = (Mode&Lower) == Lower IsLower = (Mode&Lower) == Lower
}; };
Index kc = std::min<Index>(Blocking::Max_kc/4,size); // cache block size along the K direction Index kc = size; // cache block size along the K direction
Index mc = std::min<Index>(Blocking::Max_mc,size); // cache block size along the M direction Index mc = size; // cache block size along the M direction
Index nc = cols; // cache block size along the N direction
computeProductBlockingSizes<Scalar,Scalar>(kc, mc, nc);
// it is better to use smaller blocks along the diagonal
kc /= 4;
Scalar* blockA = ei_aligned_stack_new(Scalar, kc*mc); Scalar* blockA = ei_aligned_stack_new(Scalar, kc*mc);
std::size_t sizeB = kc*Blocking::PacketSize*Blocking::nr + kc*cols; std::size_t sizeB = kc*Blocking::PacketSize*Blocking::nr + kc*cols;
@ -196,8 +200,15 @@ struct ei_triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conjugate,TriStor
IsLower = (Mode&Lower) == Lower IsLower = (Mode&Lower) == Lower
}; };
Index kc = std::min<Index>(Blocking::Max_kc/4,size); // cache block size along the K direction // Index kc = std::min<Index>(Blocking::Max_kc/4,size); // cache block size along the K direction
Index mc = std::min<Index>(Blocking::Max_mc,size); // cache block size along the M direction // Index mc = std::min<Index>(Blocking::Max_mc,size); // cache block size along the M direction
// check that !!!!
Index kc = size; // cache block size along the K direction
Index mc = size; // cache block size along the M direction
Index nc = rows; // cache block size along the N direction
computeProductBlockingSizes<Scalar,Scalar>(kc, mc, nc);
// it is better to use smaller blocks along the diagonal
kc /= 4;
Scalar* blockA = ei_aligned_stack_new(Scalar, kc*mc); Scalar* blockA = ei_aligned_stack_new(Scalar, kc*mc);
std::size_t sizeB = kc*Blocking::PacketSize*Blocking::nr + kc*size; std::size_t sizeB = kc*Blocking::PacketSize*Blocking::nr + kc*size;

10
Eigen/src/Core/util/BlasUtil.h
View File

@ -123,7 +123,7 @@ class ei_const_blas_data_mapper
Index m_stride; Index m_stride;
}; };
// Defines various constant controlling level 3 blocking // Defines various constant controlling register blocking for matrix-matrix algorithms.
template<typename Scalar> template<typename Scalar>
struct ei_product_blocking_traits struct ei_product_blocking_traits
{ {
@ -136,13 +136,7 @@ struct ei_product_blocking_traits
nr = NumberOfRegisters/4, nr = NumberOfRegisters/4,
// register block size along the M direction (currently, this one cannot be modified) // register block size along the M direction (currently, this one cannot be modified)
mr = 2 * PacketSize, mr = 2 * PacketSize
// max cache block size along the K direction
Max_kc = 4 * ei_meta_sqrt<EIGEN_TUNE_FOR_CPU_CACHE_SIZE/(64*sizeof(Scalar))>::ret,
// max cache block size along the M direction
Max_mc = 2*Max_kc
}; };
}; };