GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-04-20 16:49:38 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fix a regression when using OpenMP, and fix bug #714: the number of threads might be lower than the number of requested ones

Browse Source
This commit is contained in:
Gael Guennebaud 2015-02-18 15:19:23 +01:00
parent 548b781380
commit c7bb1e8ea8
3 changed files with 33 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

25
Eigen/src/Core/products/GeneralMatrixMatrix.h
View File

@ -217,8 +217,9 @@ struct gemm_functor
: m_lhs(lhs), m_rhs(rhs), m_dest(dest), m_actualAlpha(actualAlpha), m_blocking(blocking) : m_lhs(lhs), m_rhs(rhs), m_dest(dest), m_actualAlpha(actualAlpha), m_blocking(blocking)
{} {}
void initParallelSession() const void initParallelSession(Index num_threads) const
{ {
m_blocking.initParallel(m_lhs.rows(), m_rhs.cols(), m_lhs.cols(), num_threads);
m_blocking.allocateA(); m_blocking.allocateA();
} }
@ -276,7 +277,7 @@ class level3_blocking
}; };
template<int StorageOrder, typename _LhsScalar, typename _RhsScalar, int MaxRows, int MaxCols, int MaxDepth, int KcFactor> template<int StorageOrder, typename _LhsScalar, typename _RhsScalar, int MaxRows, int MaxCols, int MaxDepth, int KcFactor>
class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, MaxDepth, KcFactor, true> class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, MaxDepth, KcFactor, true /* == FiniteAtCompileTime */>
: public level3_blocking< : public level3_blocking<
typename conditional<StorageOrder==RowMajor,_RhsScalar,_LhsScalar>::type, typename conditional<StorageOrder==RowMajor,_RhsScalar,_LhsScalar>::type,
typename conditional<StorageOrder==RowMajor,_LhsScalar,_RhsScalar>::type> typename conditional<StorageOrder==RowMajor,_LhsScalar,_RhsScalar>::type>
@ -299,7 +300,7 @@ class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, M
public: public:
gemm_blocking_space(Index /*rows*/, Index /*cols*/, Index /*depth*/, int /*num_threads*/, bool /*full_rows = false*/) gemm_blocking_space(Index /*rows*/, Index /*cols*/, Index /*depth*/, Index /*num_threads*/, bool /*full_rows = false*/)
{ {
this->m_mc = ActualRows; this->m_mc = ActualRows;
this->m_nc = ActualCols; this->m_nc = ActualCols;
@ -307,6 +308,9 @@ class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, M
this->m_blockA = m_staticA; this->m_blockA = m_staticA;
this->m_blockB = m_staticB; this->m_blockB = m_staticB;
} }
void initParallel(Index, Index, Index, Index)
{}
inline void allocateA() {} inline void allocateA() {}
inline void allocateB() {} inline void allocateB() {}
@ -331,7 +335,7 @@ class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, M
public: public:
gemm_blocking_space(Index rows, Index cols, Index depth, int num_threads, bool l3_blocking) gemm_blocking_space(Index rows, Index cols, Index depth, Index num_threads, bool l3_blocking)
{ {
this->m_mc = Transpose ? cols : rows; this->m_mc = Transpose ? cols : rows;
this->m_nc = Transpose ? rows : cols; this->m_nc = Transpose ? rows : cols;
@ -351,6 +355,19 @@ class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, M
m_sizeA = this->m_mc * this->m_kc; m_sizeA = this->m_mc * this->m_kc;
m_sizeB = this->m_kc * this->m_nc; m_sizeB = this->m_kc * this->m_nc;
} }
void initParallel(Index rows, Index cols, Index depth, Index num_threads)
{
this->m_mc = Transpose ? cols : rows;
this->m_nc = Transpose ? rows : cols;
this->m_kc = depth;
eigen_internal_assert(this->m_blockA==0 && this->m_blockB==0);
Index m = this->m_mc;
computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, m, this->m_nc, num_threads);
m_sizeA = this->m_mc * this->m_kc;
m_sizeB = this->m_kc * this->m_nc;
}
void allocateA() void allocateA()
{ {

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

@ -120,25 +120,28 @@ void parallelize_gemm(const Functor& func, Index rows, Index cols, bool transpos
return func(0,rows, 0,cols); return func(0,rows, 0,cols);
Eigen::initParallel(); Eigen::initParallel();
func.initParallelSession(); func.initParallelSession(threads);
if(transpose) if(transpose)
std::swap(rows,cols); std::swap(rows,cols);
Index blockCols = (cols / threads) & ~Index(0x3);
Index blockRows = (rows / threads);
blockRows = (blockRows/Functor::Traits::mr)*Functor::Traits::mr;
ei_declare_aligned_stack_constructed_variable(GemmParallelInfo<Index>,info,threads,0); ei_declare_aligned_stack_constructed_variable(GemmParallelInfo<Index>,info,threads,0);
#pragma omp parallel num_threads(threads) #pragma omp parallel num_threads(threads)
{ {
Index i = omp_get_thread_num(); Index i = omp_get_thread_num();
// Note that the actual number of threads might be lower than the number of request ones.
Index actual_threads = omp_get_num_threads();
Index blockCols = (cols / actual_threads) & ~Index(0x3);
Index blockRows = (rows / actual_threads);
blockRows = (blockRows/Functor::Traits::mr)*Functor::Traits::mr;
Index r0 = i*blockRows; Index r0 = i*blockRows;
Index actualBlockRows = (i+1==threads) ? rows-r0 : blockRows; Index actualBlockRows = (i+1==actual_threads) ? rows-r0 : blockRows;
Index c0 = i*blockCols; Index c0 = i*blockCols;
Index actualBlockCols = (i+1==threads) ? cols-c0 : blockCols; Index actualBlockCols = (i+1==actual_threads) ? cols-c0 : blockCols;
info[i].lhs_start = r0; info[i].lhs_start = r0;
info[i].lhs_length = actualBlockRows; info[i].lhs_length = actualBlockRows;

3
test/product_large.cpp
View File

@ -64,8 +64,7 @@ void test_product_large()
#endif #endif
// Regression test for bug 714: // Regression test for bug 714:
#ifdef EIGEN_HAS_OPENMP #if defined EIGEN_HAS_OPENMP
std::cout << "Testing omp_set_dynamic(1)\n";
omp_set_dynamic(1); omp_set_dynamic(1);
for(int i = 0; i < g_repeat; i++) { for(int i = 0; i < g_repeat; i++) {
CALL_SUBTEST_6( product(Matrix<float,Dynamic,Dynamic>(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) ); CALL_SUBTEST_6( product(Matrix<float,Dynamic,Dynamic>(internal::random<int>(1,EIGEN_TEST_MAX_SIZE), internal::random<int>(1,EIGEN_TEST_MAX_SIZE))) );