diff options
Diffstat (limited to 'thirdparty/embree/kernels/common/alloc.h')
| -rw-r--r-- | thirdparty/embree/kernels/common/alloc.h | 72 |
1 files changed, 58 insertions, 14 deletions
diff --git a/thirdparty/embree/kernels/common/alloc.h b/thirdparty/embree/kernels/common/alloc.h index 4458e35c24..12769df2c8 100644 --- a/thirdparty/embree/kernels/common/alloc.h +++ b/thirdparty/embree/kernels/common/alloc.h @@ -8,6 +8,10 @@ #include "scene.h" #include "primref.h" +#if defined(APPLE) && defined(__aarch64__) +#include <mutex> +#endif + namespace embree { class FastAllocator @@ -26,7 +30,7 @@ namespace embree public: struct ThreadLocal2; - enum AllocationType { ALIGNED_MALLOC, OS_MALLOC, SHARED, ANY_TYPE }; + enum AllocationType { ALIGNED_MALLOC, EMBREE_OS_MALLOC, SHARED, ANY_TYPE }; /*! Per thread structure holding the current memory block. */ struct __aligned(64) ThreadLocal @@ -132,7 +136,11 @@ namespace embree { assert(alloc_i); if (alloc.load() == alloc_i) return; +#if defined(APPLE) && defined(__aarch64__) + std::scoped_lock lock(mutex); +#else Lock<SpinLock> lock(mutex); +#endif //if (alloc.load() == alloc_i) return; // not required as only one thread calls bind if (alloc.load()) { alloc.load()->bytesUsed += alloc0.getUsedBytes() + alloc1.getUsedBytes(); @@ -150,7 +158,11 @@ namespace embree { assert(alloc_i); if (alloc.load() != alloc_i) return; +#if defined(APPLE) && defined(__aarch64__) + std::scoped_lock lock(mutex); +#else Lock<SpinLock> lock(mutex); +#endif if (alloc.load() != alloc_i) return; // required as a different thread calls unbind alloc.load()->bytesUsed += alloc0.getUsedBytes() + alloc1.getUsedBytes(); alloc.load()->bytesFree += alloc0.getFreeBytes() + alloc1.getFreeBytes(); @@ -161,7 +173,11 @@ namespace embree } public: +#if defined(APPLE) && defined(__aarch64__) + std::mutex mutex; +#else SpinLock mutex; //!< required as unbind is called from other threads +#endif std::atomic<FastAllocator*> alloc; //!< parent allocator ThreadLocal alloc0; ThreadLocal alloc1; @@ -169,7 +185,7 @@ namespace embree FastAllocator (Device* device, bool osAllocation) : device(device), slotMask(0), usedBlocks(nullptr), freeBlocks(nullptr), use_single_mode(false), defaultBlockSize(PAGE_SIZE), estimatedSize(0), - growSize(PAGE_SIZE), maxGrowSize(maxAllocationSize), log2_grow_size_scale(0), bytesUsed(0), bytesFree(0), bytesWasted(0), atype(osAllocation ? OS_MALLOC : ALIGNED_MALLOC), + growSize(PAGE_SIZE), maxGrowSize(maxAllocationSize), log2_grow_size_scale(0), bytesUsed(0), bytesFree(0), bytesWasted(0), atype(osAllocation ? EMBREE_OS_MALLOC : ALIGNED_MALLOC), primrefarray(device,0) { for (size_t i=0; i<MAX_THREAD_USED_BLOCK_SLOTS; i++) @@ -206,7 +222,7 @@ namespace embree void setOSallocation(bool flag) { - atype = flag ? OS_MALLOC : ALIGNED_MALLOC; + atype = flag ? EMBREE_OS_MALLOC : ALIGNED_MALLOC; } private: @@ -217,7 +233,11 @@ namespace embree ThreadLocal2* alloc = thread_local_allocator2; if (alloc == nullptr) { thread_local_allocator2 = alloc = new ThreadLocal2; +#if defined(APPLE) && defined(__aarch64__) + std::scoped_lock lock(s_thread_local_allocators_lock); +#else Lock<SpinLock> lock(s_thread_local_allocators_lock); +#endif s_thread_local_allocators.push_back(make_unique(alloc)); } return alloc; @@ -227,7 +247,11 @@ namespace embree __forceinline void join(ThreadLocal2* alloc) { +#if defined(APPLE) && defined(__aarch64__) + std::scoped_lock lock(s_thread_local_allocators_lock); +#else Lock<SpinLock> lock(thread_local_allocators_lock); +#endif thread_local_allocators.push_back(alloc); } @@ -492,7 +516,11 @@ namespace embree /* parallel block creation in case of no freeBlocks, avoids single global mutex */ if (likely(freeBlocks.load() == nullptr)) { +#if defined(APPLE) && defined(__aarch64__) + std::scoped_lock lock(slotMutex[slot]); +#else Lock<SpinLock> lock(slotMutex[slot]); +#endif if (myUsedBlocks == threadUsedBlocks[slot]) { const size_t alignedBytes = (bytes+(align-1)) & ~(align-1); const size_t allocSize = max(min(growSize,maxGrowSize),alignedBytes); @@ -505,7 +533,11 @@ namespace embree /* if this fails allocate new block */ { - Lock<SpinLock> lock(mutex); +#if defined(APPLE) && defined(__aarch64__) + std::scoped_lock lock(mutex); +#else + Lock<SpinLock> lock(mutex); +#endif if (myUsedBlocks == threadUsedBlocks[slot]) { if (freeBlocks.load() != nullptr) { @@ -527,7 +559,11 @@ namespace embree /*! add new block */ void addBlock(void* ptr, ssize_t bytes) { +#if defined(APPLE) && defined(__aarch64__) + std::scoped_lock lock(mutex); +#else Lock<SpinLock> lock(mutex); +#endif const size_t sizeof_Header = offsetof(Block,data[0]); void* aptr = (void*) ((((size_t)ptr)+maxAlignment-1) & ~(maxAlignment-1)); size_t ofs = (size_t) aptr - (size_t) ptr; @@ -613,8 +649,8 @@ namespace embree bytesWasted(alloc->bytesWasted), stat_all(alloc,ANY_TYPE), stat_malloc(alloc,ALIGNED_MALLOC), - stat_4K(alloc,OS_MALLOC,false), - stat_2M(alloc,OS_MALLOC,true), + stat_4K(alloc,EMBREE_OS_MALLOC,false), + stat_2M(alloc,EMBREE_OS_MALLOC,true), stat_shared(alloc,SHARED) {} AllStatistics (size_t bytesUsed, @@ -707,7 +743,7 @@ namespace embree /* We avoid using os_malloc for small blocks as this could * cause a risk of fragmenting the virtual address space and * reach the limit of vm.max_map_count = 65k under Linux. */ - if (atype == OS_MALLOC && bytesAllocate < maxAllocationSize) + if (atype == EMBREE_OS_MALLOC && bytesAllocate < maxAllocationSize) atype = ALIGNED_MALLOC; /* we need to additionally allocate some header */ @@ -716,7 +752,7 @@ namespace embree bytesReserve = sizeof_Header+bytesReserve; /* consume full 4k pages with using os_malloc */ - if (atype == OS_MALLOC) { + if (atype == EMBREE_OS_MALLOC) { bytesAllocate = ((bytesAllocate+PAGE_SIZE-1) & ~(PAGE_SIZE-1)); bytesReserve = ((bytesReserve +PAGE_SIZE-1) & ~(PAGE_SIZE-1)); } @@ -748,11 +784,11 @@ namespace embree return new (ptr) Block(ALIGNED_MALLOC,bytesAllocate-sizeof_Header,bytesAllocate-sizeof_Header,next,alignment); } } - else if (atype == OS_MALLOC) + else if (atype == EMBREE_OS_MALLOC) { if (device) device->memoryMonitor(bytesAllocate,false); bool huge_pages; ptr = os_malloc(bytesReserve,huge_pages); - return new (ptr) Block(OS_MALLOC,bytesAllocate-sizeof_Header,bytesReserve-sizeof_Header,next,0,huge_pages); + return new (ptr) Block(EMBREE_OS_MALLOC,bytesAllocate-sizeof_Header,bytesReserve-sizeof_Header,next,0,huge_pages); } else assert(false); @@ -796,7 +832,7 @@ namespace embree if (device) device->memoryMonitor(-sizeof_Alloced,true); } - else if (atype == OS_MALLOC) { + else if (atype == EMBREE_OS_MALLOC) { size_t sizeof_This = sizeof_Header+reserveEnd; os_free(this,sizeof_This,huge_pages); if (device) device->memoryMonitor(-sizeof_Alloced,true); @@ -857,7 +893,7 @@ namespace embree bool hasType(AllocationType atype_i, bool huge_pages_i) const { if (atype_i == ANY_TYPE ) return true; - else if (atype == OS_MALLOC) return atype_i == atype && huge_pages_i == huge_pages; + else if (atype == EMBREE_OS_MALLOC) return atype_i == atype && huge_pages_i == huge_pages; else return atype_i == atype; } @@ -906,7 +942,7 @@ namespace embree void print_block() const { if (atype == ALIGNED_MALLOC) std::cout << "A"; - else if (atype == OS_MALLOC) std::cout << "O"; + else if (atype == EMBREE_OS_MALLOC) std::cout << "O"; else if (atype == SHARED) std::cout << "S"; if (huge_pages) std::cout << "H"; size_t bytesUsed = getBlockUsedBytes(); @@ -936,7 +972,11 @@ namespace embree std::atomic<Block*> freeBlocks; std::atomic<Block*> threadBlocks[MAX_THREAD_USED_BLOCK_SLOTS]; - SpinLock slotMutex[MAX_THREAD_USED_BLOCK_SLOTS]; +#if defined(APPLE) && defined(__aarch64__) + std::mutex slotMutex[MAX_THREAD_USED_BLOCK_SLOTS]; +#else + PaddedSpinLock slotMutex[MAX_THREAD_USED_BLOCK_SLOTS]; +#endif bool use_single_mode; size_t defaultBlockSize; @@ -950,7 +990,11 @@ namespace embree static __thread ThreadLocal2* thread_local_allocator2; static SpinLock s_thread_local_allocators_lock; static std::vector<std::unique_ptr<ThreadLocal2>> s_thread_local_allocators; +#if defined(APPLE) && defined(__aarch64__) + std::mutex thread_local_allocators_lock; +#else SpinLock thread_local_allocators_lock; +#endif std::vector<ThreadLocal2*> thread_local_allocators; AllocationType atype; mvector<PrimRef> primrefarray; //!< primrefarray used to allocate nodes |