only use alloca on linux. Use malloc on other platforms. Needed for mingw

compatibility.
This commit is contained in:
Benoit Jacob 2008-08-18 16:36:47 +00:00
parent baf0cffedd
commit e778ae2559
3 changed files with 30 additions and 11 deletions

View File

@ -90,11 +90,10 @@ static void ei_cache_friendly_product(
const bool needRhsCopy = (PacketSize>1) && ((rhsStride%PacketSize!=0) || (size_t(rhs)%16!=0)); const bool needRhsCopy = (PacketSize>1) && ((rhsStride%PacketSize!=0) || (size_t(rhs)%16!=0));
Scalar* __restrict__ block = 0; Scalar* __restrict__ block = 0;
const int allocBlockSize = sizeof(Scalar)*l2BlockRows*size; const int allocBlockSize = sizeof(Scalar)*l2BlockRows*size;
if (allocBlockSize>16000000) const bool allocBlockUsingAlloca = EIGEN_USE_ALLOCA && allocBlockSize<=16000000;
block = (Scalar*)malloc(allocBlockSize); block = (Scalar*)ei_alloca_or_malloc(allocBlockUsingAlloca, allocBlockSize);
else Scalar* __restrict__ rhsCopy
block = (Scalar*)alloca(allocBlockSize); = (Scalar*)ei_alloca_or_malloc(true, sizeof(Scalar)*l2BlockSizeAligned*l2BlockSizeAligned);
Scalar* __restrict__ rhsCopy = (Scalar*)alloca(sizeof(Scalar)*l2BlockSizeAligned*l2BlockSizeAligned);
// loops on each L2 cache friendly blocks of the result // loops on each L2 cache friendly blocks of the result
for(int l2i=0; l2i<rows; l2i+=l2BlockRows) for(int l2i=0; l2i<rows; l2i+=l2BlockRows)
@ -341,8 +340,10 @@ static void ei_cache_friendly_product(
} }
} }
if (allocBlockSize>16000000) if (!allocBlockUsingAlloca)
free(block); free(block);
if (!EIGEN_USE_ALLOCA)
free(rhsCopy);
} }
#endif // EIGEN_EXTERN_INSTANTIATIONS #endif // EIGEN_EXTERN_INSTANTIATIONS

View File

@ -548,7 +548,7 @@ struct ei_cache_friendly_product_selector<ProductType,LhsRows,ColMajor,HasDirect
_res = &res.coeffRef(0); _res = &res.coeffRef(0);
else else
{ {
_res = (Scalar*)alloca(sizeof(Scalar)*res.size()); _res = (Scalar*)ei_alloca_or_malloc(true, sizeof(Scalar)*res.size());
Map<Matrix<Scalar,DestDerived::RowsAtCompileTime,1> >(_res, res.size()) = res; Map<Matrix<Scalar,DestDerived::RowsAtCompileTime,1> >(_res, res.size()) = res;
} }
ei_cache_friendly_product_colmajor_times_vector(res.size(), ei_cache_friendly_product_colmajor_times_vector(res.size(),
@ -557,6 +557,8 @@ struct ei_cache_friendly_product_selector<ProductType,LhsRows,ColMajor,HasDirect
if (!EvalToRes) if (!EvalToRes)
res = Map<Matrix<Scalar,DestDerived::SizeAtCompileTime,1> >(_res, res.size()); res = Map<Matrix<Scalar,DestDerived::SizeAtCompileTime,1> >(_res, res.size());
if(!EIGEN_USE_ALLOCA) free(_res);
} }
}; };
@ -591,7 +593,7 @@ struct ei_cache_friendly_product_selector<ProductType,1,LhsOrder,LhsAccess,RhsCo
_res = &res.coeffRef(0); _res = &res.coeffRef(0);
else else
{ {
_res = (Scalar*)alloca(sizeof(Scalar)*res.size()); _res = (Scalar*)ei_alloca_or_malloc(true, sizeof(Scalar)*res.size());
Map<Matrix<Scalar,DestDerived::SizeAtCompileTime,1> >(_res, res.size()) = res; Map<Matrix<Scalar,DestDerived::SizeAtCompileTime,1> >(_res, res.size()) = res;
} }
ei_cache_friendly_product_colmajor_times_vector(res.size(), ei_cache_friendly_product_colmajor_times_vector(res.size(),
@ -600,6 +602,8 @@ struct ei_cache_friendly_product_selector<ProductType,1,LhsOrder,LhsAccess,RhsCo
if (!EvalToRes) if (!EvalToRes)
res = Map<Matrix<Scalar,DestDerived::SizeAtCompileTime,1> >(_res, res.size()); res = Map<Matrix<Scalar,DestDerived::SizeAtCompileTime,1> >(_res, res.size());
if(!EIGEN_USE_ALLOCA) free(_res);
} }
}; };
@ -621,11 +625,13 @@ struct ei_cache_friendly_product_selector<ProductType,LhsRows,RowMajor,HasDirect
_rhs = &product.rhs().const_cast_derived().coeffRef(0); _rhs = &product.rhs().const_cast_derived().coeffRef(0);
else else
{ {
_rhs = (Scalar*)alloca(sizeof(Scalar)*product.rhs().size()); _rhs = (Scalar*)ei_alloca_or_malloc(true, sizeof(Scalar)*product.rhs().size());
Map<Matrix<Scalar,Rhs::SizeAtCompileTime,1> >(_rhs, product.rhs().size()) = product.rhs(); Map<Matrix<Scalar,Rhs::SizeAtCompileTime,1> >(_rhs, product.rhs().size()) = product.rhs();
} }
ei_cache_friendly_product_rowmajor_times_vector(&product.lhs().const_cast_derived().coeffRef(0,0), product.lhs().stride(), ei_cache_friendly_product_rowmajor_times_vector(&product.lhs().const_cast_derived().coeffRef(0,0), product.lhs().stride(),
_rhs, product.rhs().size(), res); _rhs, product.rhs().size(), res);
if(!EIGEN_USE_ALLOCA) free(_rhs);
} }
}; };
@ -647,11 +653,13 @@ struct ei_cache_friendly_product_selector<ProductType,1,LhsOrder,LhsAccess,RhsCo
_lhs = &product.lhs().const_cast_derived().coeffRef(0); _lhs = &product.lhs().const_cast_derived().coeffRef(0);
else else
{ {
_lhs = (Scalar*)alloca(sizeof(Scalar)*product.lhs().size()); _lhs = (Scalar*)ei_alloca_or_malloc(true, sizeof(Scalar)*product.lhs().size());
Map<Matrix<Scalar,Lhs::SizeAtCompileTime,1> >(_lhs, product.lhs().size()) = product.lhs(); Map<Matrix<Scalar,Lhs::SizeAtCompileTime,1> >(_lhs, product.lhs().size()) = product.lhs();
} }
ei_cache_friendly_product_rowmajor_times_vector(&product.rhs().const_cast_derived().coeffRef(0,0), product.rhs().stride(), ei_cache_friendly_product_rowmajor_times_vector(&product.rhs().const_cast_derived().coeffRef(0,0), product.rhs().stride(),
_lhs, product.lhs().size(), res); _lhs, product.lhs().size(), res);
if(!EIGEN_USE_ALLOCA) free(_lhs);
} }
}; };

View File

@ -52,7 +52,9 @@ using Eigen::Matrix; \
using Eigen::MatrixBase; using Eigen::MatrixBase;
#ifdef NDEBUG #ifdef NDEBUG
#define EIGEN_NO_DEBUG # ifndef EIGEN_NO_DEBUG
# define EIGEN_NO_DEBUG
# endif
#endif #endif
#ifndef ei_assert #ifndef ei_assert
@ -146,4 +148,12 @@ friend class Eigen::MatrixBase<Derived>;
#define EIGEN_ENUM_MIN(a,b) (((int)a <= (int)b) ? (int)a : (int)b) #define EIGEN_ENUM_MIN(a,b) (((int)a <= (int)b) ? (int)a : (int)b)
#define EIGEN_ENUM_MAX(a,b) (((int)a >= (int)b) ? (int)a : (int)b) #define EIGEN_ENUM_MAX(a,b) (((int)a >= (int)b) ? (int)a : (int)b)
#ifdef __linux__
# define EIGEN_USE_ALLOCA 1
# define ei_alloca_or_malloc(condition, size) (condition?alloca(size):malloc(size))
#else
# define EIGEN_USE_ALLOCA 0
# define ei_alloca_or_malloc(condition, size) malloc(size)
#endif
#endif // EIGEN_MACROS_H #endif // EIGEN_MACROS_H