| 1 | // Copyright 2009-2021 Intel Corporation | 
| 2 | // SPDX-License-Identifier: Apache-2.0 | 
| 3 |  | 
| 4 | #pragma once | 
| 5 |  | 
| 6 | #include "intrinsics.h" | 
| 7 | #include "sysinfo.h" | 
| 8 | #include "atomic.h" | 
| 9 |  | 
| 10 | namespace embree | 
| 11 | { | 
| 12 |   /*! system barrier using operating system */ | 
| 13 |   class BarrierSys | 
| 14 |   { | 
| 15 |   public: | 
| 16 |  | 
| 17 |     /*! construction / destruction */ | 
| 18 |     BarrierSys (size_t N = 0); | 
| 19 |     ~BarrierSys (); | 
| 20 |  | 
| 21 |   private: | 
| 22 |     /*! class in non-copyable */ | 
| 23 |     BarrierSys (const BarrierSys& other) DELETED; // do not implement | 
| 24 |     BarrierSys& operator= (const BarrierSys& other) DELETED; // do not implement | 
| 25 |  | 
| 26 |   public: | 
| 27 |     /*! intializes the barrier with some number of threads */ | 
| 28 |     void init(size_t count); | 
| 29 |  | 
| 30 |     /*! lets calling thread wait in barrier */ | 
| 31 |     void wait(); | 
| 32 |  | 
| 33 |   private: | 
| 34 |     void* opaque; | 
| 35 |   }; | 
| 36 |  | 
| 37 |   /*! fast active barrier using atomitc counter */ | 
| 38 |   struct BarrierActive  | 
| 39 |   { | 
| 40 |   public: | 
| 41 |     BarrierActive ()  | 
| 42 |       : cntr(0) {} | 
| 43 |      | 
| 44 |     void reset() { | 
| 45 |       cntr.store(i: 0); | 
| 46 |     } | 
| 47 |  | 
| 48 |     void wait (size_t numThreads)  | 
| 49 |     { | 
| 50 |       cntr++; | 
| 51 |       while (cntr.load() != numThreads)  | 
| 52 |         pause_cpu(); | 
| 53 |     } | 
| 54 |  | 
| 55 |   private: | 
| 56 |     std::atomic<size_t> cntr; | 
| 57 |   }; | 
| 58 |  | 
| 59 |   /*! fast active barrier that does not require initialization to some number of threads */ | 
| 60 |   struct BarrierActiveAutoReset | 
| 61 |   { | 
| 62 |   public: | 
| 63 |     BarrierActiveAutoReset ()  | 
| 64 |       : cntr0(0), cntr1(0) {} | 
| 65 |  | 
| 66 |     void wait (size_t threadCount)  | 
| 67 |     { | 
| 68 |       cntr0.fetch_add(i: 1); | 
| 69 |       while (cntr0 != threadCount) pause_cpu(); | 
| 70 |       cntr1.fetch_add(i: 1); | 
| 71 |       while (cntr1 != threadCount) pause_cpu(); | 
| 72 |       cntr0.fetch_add(i: -1); | 
| 73 |       while (cntr0 != 0) pause_cpu(); | 
| 74 |       cntr1.fetch_add(i: -1); | 
| 75 |       while (cntr1 != 0) pause_cpu(); | 
| 76 |     } | 
| 77 |  | 
| 78 |   private: | 
| 79 |     std::atomic<size_t> cntr0; | 
| 80 |     std::atomic<size_t> cntr1; | 
| 81 |   }; | 
| 82 |  | 
| 83 |   class LinearBarrierActive | 
| 84 |   { | 
| 85 |   public: | 
| 86 |  | 
| 87 |     /*! construction and destruction */ | 
| 88 |     LinearBarrierActive (size_t threadCount = 0); | 
| 89 |     ~LinearBarrierActive(); | 
| 90 |      | 
| 91 |   private: | 
| 92 |     /*! class in non-copyable */ | 
| 93 |     LinearBarrierActive (const LinearBarrierActive& other) DELETED; // do not implement | 
| 94 |     LinearBarrierActive& operator= (const LinearBarrierActive& other) DELETED; // do not implement | 
| 95 |  | 
| 96 |   public: | 
| 97 |     /*! intializes the barrier with some number of threads */ | 
| 98 |     void init(size_t threadCount); | 
| 99 |      | 
| 100 |     /*! thread with threadIndex waits in the barrier */ | 
| 101 |     void wait (const size_t threadIndex); | 
| 102 |      | 
| 103 |   private: | 
| 104 |     volatile unsigned char* count0; | 
| 105 |     volatile unsigned char* count1;  | 
| 106 |     volatile unsigned int mode; | 
| 107 |     volatile unsigned int flag0; | 
| 108 |     volatile unsigned int flag1; | 
| 109 |     volatile size_t threadCount; | 
| 110 |   }; | 
| 111 | } | 
| 112 |  | 
| 113 |  |