Simplify compiling GPU code for tests

[gromacs.git] / src / gromacs / utility / alignedallocator.cpp

/*
 * This file is part of the GROMACS molecular simulation package.
 *
 * Copyright (c) 2015,2017,2018,2019, by the GROMACS development team, led by
 * Mark Abraham, David van der Spoel, Berk Hess, and Erik Lindahl,
 * and including many others, as listed in the AUTHORS file in the
 * top-level source directory and at http://www.gromacs.org.
 *
 * GROMACS is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public License
 * as published by the Free Software Foundation; either version 2.1
 * of the License, or (at your option) any later version.
 *
 * GROMACS is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with GROMACS; if not, see
 * http://www.gnu.org/licenses, or write to the Free Software Foundation,
 * Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA.
 *
 * If you want to redistribute modifications to GROMACS, please
 * consider that scientific software is very special. Version
 * control is crucial - bugs must be traceable. We will be happy to
 * consider code for inclusion in the official distribution, but
 * derived work must not be called official GROMACS. Details are found
 * in the README & COPYING files - if they are missing, get the
 * official version at http://www.gromacs.org.
 *
 * To help us fund GROMACS development, we humbly ask that you cite
 * the research papers on the package. Check out http://www.gromacs.org.
 */
/*! \internal \file
 * \brief
 * Implements AlignedAllocator.
 *
 * \author Erik Lindahl <erik.lindahl@gmail.com>
 * \author Mark Abraham <mark.j.abraham@gmail.com>
 * \ingroup module_utility
 */
#include "gmxpre.h"

#include "alignedallocator.h"

#include "config.h"

#include <cstdlib>

#include <memory>

#if HAVE_MM_MALLOC_H
#    include <mm_malloc.h>
#elif HAVE_MALLOC_H
#    include <malloc.h>
#elif HAVE_XMMINTRIN_H
#    include <xmmintrin.h>
#endif

#ifdef HAVE_UNISTD_H
#    include <unistd.h>
#endif

#if GMX_NATIVE_WINDOWS
#    include <windows.h> // only for the page size query purposes
#endif

#include "gromacs/utility/gmxassert.h"

namespace gmx
{

namespace
{

/*! \brief Allocate aligned memory in a fully portable way
 *
 *  \param bytes  Amount of memory (bytes) to allocate. The routine will return
 *                nullptr if the allocation fails. However, note that asking for
 *                zero bytes will return a pointer that is non-null and properly
 *                aligned (but obviously you cannot use it, since you promised
 *                not to access data beyond the 0 bytes you asked for).
 *
 *  \param alignment  Alignment specification in bytes, must be a power of 2.
 *
 * \return Nonzero pointer if the allocation worked, otherwise nullptr.
 *  This routine should only be called from alignedMalloc(), which also does
 *  the checking for valid values. This particular function is used for platforms
 *  where we have no control of the alignment of memory returned by the system.
 *  Instead, we increase the amount of memory requested internally such that we
 *  both can create a pointer inside this memory that fulfills the memory
 *  alignment requested, and that we have room to store the original pointer
 *  just before this area.
 *
 *  \note This is an internal routine that should only be called from
 *        gmx::alignedMalloc(). Just like system-provided routines, it provides
 *        memory that is aligned - but not padded.
 */
gmx_unused void* alignedMallocGeneric(std::size_t bytes, std::size_t alignment)
{
    // The amount of extra memory (beyound what the user asked for) we need is:
    // - sizeof(void *), to store the original pointer
    // - alignment, to make sure we have an aligned pointer in the area
    void* pMalloc = malloc(bytes + sizeof(void*) + alignment);

    if (pMalloc == nullptr)
    {
        return nullptr;
    }

    // Convert pMalloc to size_t (so we work with raw bytes), add the space we
    // need to save the original pointer, and (alignment-1) bytes, and then mask
    // out the lowest bits.
    std::size_t mask     = ~static_cast<std::size_t>(alignment - 1);
    void*       pAligned = reinterpret_cast<void*>(
            (reinterpret_cast<std::size_t>(pMalloc) + sizeof(void*) + alignment - 1) & mask);

    // Store original pointer. Since we allocated at least sizeof(void *) extra
    // space this is always a valid memory location.
    reinterpret_cast<void**>(pAligned)[-1] = pMalloc;

    return pAligned;
}


/*! \brief Free aligned memory
 *
 *  \param p  Memory pointer previously returned from
 *            gmx::internal::alignedFreePortable().
 *
 *  Since this routine relies on the original pointer being stored just before
 *  the memory area p points to, bad things will happen if you call this routine
 *  with a pointer obtained any other way, or if you call the system free()
 *  with a pointer obtained from std::alignedMalloc().
 *
 * \note  This is an internal routine that should only be called from
 *        gmx::alignedFree().
 */
gmx_unused void alignedFreeGeneric(void* p)
{
    if (p)
    {
        // Pick up the pointer stored just below p, and use that to call free()
        free(reinterpret_cast<void**>(p)[-1]);
    }
}

//! Implement malloc of \c bytes of memory, aligned to \c alignment.
void* mallocImpl(std::size_t bytes, std::size_t alignment)
{
    void* p;

#if HAVE__MM_MALLOC
    p = _mm_malloc(bytes, alignment);
#elif HAVE_POSIX_MEMALIGN
    if (posix_memalign(&p, alignment, bytes) != 0)
    {
        p = nullptr;
    }
#elif HAVE_MEMALIGN
    p = memalign(alignment, bytes);
#elif HAVE__ALIGNED_MALLOC
    p = _aligned_malloc(bytes, alignment);
#else
    p = internal::alignedMallocGeneric(bytes, alignment);
#endif

    return p;
}

//! Free aligned memory allocated with mallocImpl().
void freeImpl(void* p)
{
    if (p)
    {
#if HAVE__MM_MALLOC
        _mm_free(p);
#elif HAVE_POSIX_MEMALIGN || HAVE_MEMALIGN
        free(p);
#elif HAVE__ALIGNED_MALLOC
        _aligned_free(p);
#else
        internal::alignedFreeGeneric(p);
#endif
    }
}

} // namespace

// === AlignedAllocationPolicy

std::size_t AlignedAllocationPolicy::alignment()
{
    // For now we always use 128-byte alignment:
    // 1) IBM Power already has cache lines of 128-bytes, and needs it.
    // 2) x86 has 64 byte cache lines, but since a future AVX-1024 (rumored?)
    //    will need 1024/8=128 byte SIMD alignment, it is safer to use that
    //    already now.
    // 3) The old Pentium4 used 256-byte cache prefetching (but 64-byte lines).
    //    However, it's not worth worrying about performance for P4...
    // 4) ARM & Sparc have 64 byte lines, but will be just fine with
    //    128-byte alignment (nobody knows what the future brings)
    //
    // So, for now we're semi-lazy and just align to 128 bytes!
    //
    // TODO LINCS code is copying this assumption independently (for now)
    return 128;
}

void* AlignedAllocationPolicy::malloc(std::size_t bytes)
{
    // Pad memory at the end with another alignment bytes to avoid false sharing
    auto size = alignment();
    bytes += size;

    return mallocImpl(bytes, size);
}

void AlignedAllocationPolicy::free(void* p)
{
    freeImpl(p);
}

// === PageAlignedAllocationPolicy

//! Return a page size, from a sysconf/WinAPI query if available, or a default guess (4096 bytes).
//! \todo Move this function into sysinfo.cpp where other OS-specific code/includes live
static std::size_t getPageSize()
{
    long pageSize;
#if GMX_NATIVE_WINDOWS
    SYSTEM_INFO si;
    GetNativeSystemInfo(&si);
    pageSize = si.dwPageSize;
#elif defined(_SC_PAGESIZE)
    /* Note that sysconf returns -1 on its error conditions, which we
       don't really need to check, nor can really handle at
       initialization time. */
    pageSize = sysconf(_SC_PAGESIZE);
#elif defined(_SC_PAGE_SIZE)
    pageSize = sysconf(_SC_PAGE_SIZE);
#else
    pageSize = -1;
#endif
    return ((pageSize == -1) ? 4096 // A useful guess
                             : static_cast<std::size_t>(pageSize));
}

/* Implements the "construct on first use" idiom to avoid the static
 * initialization order fiasco where a possible static page-aligned
 * container would be initialized before the alignment variable was.
 *
 * Note that thread-safety of the initialization is guaranteed by the
 * C++11 language standard.
 *
 * The size_t has no destructor, so there is no deinitialization
 * issue.  See https://isocpp.org/wiki/faq/ctors for discussion of
 * alternatives and trade-offs. */
std::size_t PageAlignedAllocationPolicy::alignment()
{
    static size_t thePageSize = getPageSize();
    return thePageSize;
}

void* PageAlignedAllocationPolicy::malloc(std::size_t bytes)
{
    return mallocImpl(bytes, alignment());
}

void PageAlignedAllocationPolicy::free(void* p)
{
    freeImpl(p);
}

} // namespace gmx