Chris@49: // Copyright (C) 2012-2013 NICTA (www.nicta.com.au)
Chris@49: // Copyright (C) 2012-2013 Conrad Sanderson
Chris@49: // 
Chris@49: // This Source Code Form is subject to the terms of the Mozilla Public
Chris@49: // License, v. 2.0. If a copy of the MPL was not distributed with this
Chris@49: // file, You can obtain one at http://mozilla.org/MPL/2.0/.
Chris@49: 
Chris@49: 
Chris@49: //! \addtogroup memory
Chris@49: //! @{
Chris@49: 
Chris@49: 
Chris@49: class memory
Chris@49:   {
Chris@49:   public:
Chris@49:   
Chris@49:                         arma_inline             static uword enlarge_to_mult_of_chunksize(const uword n_elem);
Chris@49:   
Chris@49:   template<typename eT> arma_inline arma_malloc static eT*   acquire(const uword n_elem);
Chris@49:   
Chris@49:   template<typename eT> arma_inline arma_malloc static eT*   acquire_chunked(const uword n_elem);
Chris@49:   
Chris@49:   template<typename eT> arma_inline             static void  release(eT* mem);
Chris@49:   
Chris@49:   
Chris@49:   template<typename eT> arma_inline static bool      is_aligned(const eT*  mem);
Chris@49:   template<typename eT> arma_inline static void mark_as_aligned(      eT*& mem);
Chris@49:   template<typename eT> arma_inline static void mark_as_aligned(const eT*& mem);
Chris@49:   };
Chris@49: 
Chris@49: 
Chris@49: 
Chris@49: arma_inline
Chris@49: uword
Chris@49: memory::enlarge_to_mult_of_chunksize(const uword n_elem)
Chris@49:   {
Chris@49:   const uword chunksize = arma_config::spmat_chunksize;
Chris@49:   
Chris@49:   // this relies on integer division
Chris@49:   const uword n_elem_mod = ((n_elem % chunksize) != 0) ? ((n_elem / chunksize) + 1) * chunksize : n_elem;
Chris@49:   
Chris@49:   return n_elem_mod;
Chris@49:   }
Chris@49: 
Chris@49: 
Chris@49: 
Chris@49: template<typename eT>
Chris@49: arma_inline
Chris@49: arma_malloc
Chris@49: eT*
Chris@49: memory::acquire(const uword n_elem)
Chris@49:   {
Chris@49:   #if   defined(ARMA_USE_TBB_ALLOC)
Chris@49:     {
Chris@49:     return ( (eT *) scalable_malloc(sizeof(eT)*n_elem) );
Chris@49:     }
Chris@49:   #elif defined(ARMA_USE_MKL_ALLOC)
Chris@49:     {
Chris@49:     return ( (eT *) mkl_malloc( sizeof(eT)*n_elem, 128 ) );
Chris@49:     }
Chris@49:   #elif defined(ARMA_HAVE_POSIX_MEMALIGN)
Chris@49:     {
Chris@49:     eT* memptr;
Chris@49:     
Chris@49:     const size_t alignment = 16;  // change the 16 to 64 if you wish to align to the cache line
Chris@49:     
Chris@49:     int status = posix_memalign((void **)&memptr, ( (alignment >= sizeof(void*)) ? alignment : sizeof(void*) ), sizeof(eT)*n_elem);
Chris@49:     
Chris@49:     return (status == 0) ? memptr : NULL;
Chris@49:     }
Chris@49:   #elif defined(_MSC_VER)
Chris@49:     {
Chris@49:     return ( (eT *) _aligned_malloc( sizeof(eT)*n_elem, 16 ) );  // lives in malloc.h
Chris@49:     }
Chris@49:   #else
Chris@49:     {
Chris@49:     //return ( new(std::nothrow) eT[n_elem] );
Chris@49:     return ( (eT *) malloc(sizeof(eT)*n_elem) );
Chris@49:     }
Chris@49:   #endif
Chris@49:   
Chris@49:   // TODO: for mingw, use __mingw_aligned_malloc
Chris@49:   }
Chris@49: 
Chris@49: 
Chris@49: 
Chris@49: //! get memory in multiples of chunks, holding at least n_elem
Chris@49: template<typename eT>
Chris@49: arma_inline
Chris@49: arma_malloc
Chris@49: eT*
Chris@49: memory::acquire_chunked(const uword n_elem)
Chris@49:   {
Chris@49:   const uword n_elem_mod = memory::enlarge_to_mult_of_chunksize(n_elem);
Chris@49:   
Chris@49:   return memory::acquire<eT>(n_elem_mod);
Chris@49:   }
Chris@49: 
Chris@49: 
Chris@49: 
Chris@49: template<typename eT>
Chris@49: arma_inline
Chris@49: void
Chris@49: memory::release(eT* mem)
Chris@49:   {
Chris@49:   #if   defined(ARMA_USE_TBB_ALLOC)
Chris@49:     {
Chris@49:     scalable_free( (void *)(mem) );
Chris@49:     }
Chris@49:   #elif defined(ARMA_USE_MKL_ALLOC)
Chris@49:     {
Chris@49:     mkl_free( (void *)(mem) );
Chris@49:     }
Chris@49:   #elif defined(ARMA_HAVE_POSIX_MEMALIGN)
Chris@49:     {
Chris@49:     free( (void *)(mem) );
Chris@49:     }
Chris@49:   #elif defined(_MSC_VER)
Chris@49:     {
Chris@49:     _aligned_free( (void *)(mem) );
Chris@49:     }
Chris@49:   #else
Chris@49:     {
Chris@49:     //delete [] mem;
Chris@49:     free( (void *)(mem) );
Chris@49:     }
Chris@49:   #endif
Chris@49:   
Chris@49:   // TODO: for mingw, use __mingw_aligned_free
Chris@49:   }
Chris@49: 
Chris@49: 
Chris@49: 
Chris@49: template<typename eT>
Chris@49: arma_inline
Chris@49: bool
Chris@49: memory::is_aligned(const eT* mem)
Chris@49:   {
Chris@49:   #if (defined(ARMA_HAVE_ICC_ASSUME_ALIGNED) || defined(ARMA_HAVE_GCC_ASSUME_ALIGNED)) && !defined(ARMA_DONT_CHECK_ALIGNMENT)
Chris@49:     {
Chris@49:     return ((std::ptrdiff_t(mem) & 0x0F) == 0);
Chris@49:     }
Chris@49:   #else
Chris@49:     {
Chris@49:     arma_ignore(mem);
Chris@49:     
Chris@49:     return false;
Chris@49:     }
Chris@49:   #endif
Chris@49:   }
Chris@49: 
Chris@49: 
Chris@49: 
Chris@49: template<typename eT>
Chris@49: arma_inline
Chris@49: void
Chris@49: memory::mark_as_aligned(eT*& mem)
Chris@49:   {
Chris@49:   #if defined(ARMA_HAVE_ICC_ASSUME_ALIGNED)
Chris@49:     {
Chris@49:     __assume_aligned(mem, 16);
Chris@49:     }
Chris@49:   #elif defined(ARMA_HAVE_GCC_ASSUME_ALIGNED)
Chris@49:     {
Chris@49:     mem = (eT*)__builtin_assume_aligned(mem, 16);
Chris@49:     }
Chris@49:   #else
Chris@49:     {
Chris@49:     arma_ignore(mem);
Chris@49:     }
Chris@49:   #endif
Chris@49:   
Chris@49:   // TODO: MSVC?  __assume( (mem & 0x0F) == 0 );
Chris@49:   //
Chris@49:   // http://comments.gmane.org/gmane.comp.gcc.patches/239430
Chris@49:   // GCC __builtin_assume_aligned is similar to ICC's __assume_aligned,
Chris@49:   // so for lvalue first argument ICC's __assume_aligned can be emulated using
Chris@49:   // #define __assume_aligned(lvalueptr, align) lvalueptr = __builtin_assume_aligned (lvalueptr, align) 
Chris@49:   //
Chris@49:   // http://www.inf.ethz.ch/personal/markusp/teaching/263-2300-ETH-spring11/slides/class19.pdf
Chris@49:   // http://software.intel.com/sites/products/documentation/hpc/composerxe/en-us/cpp/lin/index.htm
Chris@49:   // http://d3f8ykwhia686p.cloudfront.net/1live/intel/CompilerAutovectorizationGuide.pdf
Chris@49:   }
Chris@49: 
Chris@49: 
Chris@49: 
Chris@49: template<typename eT>
Chris@49: arma_inline
Chris@49: void
Chris@49: memory::mark_as_aligned(const eT*& mem)
Chris@49:   {
Chris@49:   #if defined(ARMA_HAVE_ICC_ASSUME_ALIGNED)
Chris@49:     {
Chris@49:     __assume_aligned(mem, 16);
Chris@49:     }
Chris@49:   #elif defined(ARMA_HAVE_GCC_ASSUME_ALIGNED)
Chris@49:     {
Chris@49:     mem = (const eT*)__builtin_assume_aligned(mem, 16);
Chris@49:     }
Chris@49:   #else
Chris@49:     {
Chris@49:     arma_ignore(mem);
Chris@49:     }
Chris@49:   #endif
Chris@49:   }
Chris@49: 
Chris@49: 
Chris@49: 
Chris@49: //! @}