1 /* Copyright 2016 The TensorFlow Authors. All Rights Reserved. 2 3 Licensed under the Apache License, Version 2.0 (the "License"); 4 you may not use this file except in compliance with the License. 5 You may obtain a copy of the License at 6 7 http://www.apache.org/licenses/LICENSE-2.0 8 9 Unless required by applicable law or agreed to in writing, software 10 distributed under the License is distributed on an "AS IS" BASIS, 11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12 See the License for the specific language governing permissions and 13 limitations under the License. 14 ==============================================================================*/ 15 // This class is designed to get accurate profile for programs. 16 17 #ifndef TENSORFLOW_PLATFORM_PROFILEUTILS_CPU_UTILS_H__ 18 #define TENSORFLOW_PLATFORM_PROFILEUTILS_CPU_UTILS_H__ 19 20 #include <chrono> 21 #include <memory> 22 23 #include "tensorflow/core/platform/macros.h" 24 #include "tensorflow/core/platform/profile_utils/i_cpu_utils_helper.h" 25 #include "tensorflow/core/platform/types.h" 26 27 #if defined(ARMV6) || defined(__ARM_ARCH_7A__) 28 #include <sys/time.h> 29 #endif 30 31 namespace tensorflow { 32 33 namespace profile_utils { 34 35 // CpuUtils is a profiling tool with static functions 36 // designed to be called from multiple classes. 37 // A dedicated class which inherits ICpuUtilsHelper is 38 // stored as a function-local static variable which inherits 39 // GetCpuUtilsHelperSingletonInstance that caches CPU information, 40 // because loading CPU information may take a long time. 41 // Users must call EnableClockCycleProfiling before using CpuUtils. 42 class CpuUtils { 43 public: 44 // Constant for invalid frequency. 45 // This value is returned when the frequency is not obtained somehow. 46 static constexpr int64 INVALID_FREQUENCY = -1; 47 static constexpr uint64 DUMMY_CYCLE_CLOCK = 1; 48 49 // Return current clock cycle. This function is designed to 50 // minimize the overhead to get clock and maximize the accuracy of 51 // time for profile. 52 // This returns unsigned int because there is no guarantee that rdtsc 53 // is less than 2 ^ 61. 54 static inline uint64 GetCurrentClockCycle() { 55 #if defined(__ANDROID__) 56 return GetCpuUtilsHelperSingletonInstance().GetCurrentClockCycle(); 57 // ---------------------------------------------------------------- 58 #elif defined(__x86_64__) || defined(__amd64__) 59 uint64_t high, low; 60 __asm__ volatile("rdtsc" : "=a"(low), "=d"(high)); 61 return (high << 32) | low; 62 // ---------------------------------------------------------------- 63 #elif defined(__aarch64__) 64 // System timer of ARMv8 runs at a different frequency than the CPU's. 65 // The frequency is fixed, typically in the range 1-50MHz. It can because 66 // read at CNTFRQ special register. We assume the OS has set up 67 // the virtual timer properly. 68 uint64_t virtual_timer_value; 69 asm volatile("mrs %0, cntvct_el0" : "=r"(virtual_timer_value)); 70 return virtual_timer_value; 71 // ---------------------------------------------------------------- 72 // V6 is the earliest arm that has a standard cyclecount 73 #elif defined(ARMV6) || defined(__ARM_ARCH_7A__) 74 uint32_t pmccntr; 75 uint32_t pmuseren; 76 uint32_t pmcntenset; 77 // Read the user mode perf monitor counter access permissions. 78 asm volatile("mrc p15, 0, %0, c9, c14, 0" : "=r"(pmuseren)); 79 if (pmuseren & 1) { // Allows reading perfmon counters for user mode code. 80 asm volatile("mrc p15, 0, %0, c9, c12, 1" : "=r"(pmcntenset)); 81 if (pmcntenset & 0x80000000ul) { // Is it counting? 82 asm volatile("mrc p15, 0, %0, c9, c13, 0" : "=r"(pmccntr)); 83 // The counter is set up to count every 64th cyclecount 84 return static_cast<uint64>(pmccntr) * 64; // Should optimize to << 64 85 } 86 } 87 // Returning dummy clock when can't access to the counter 88 return DUMMY_CYCLE_CLOCK; 89 #else 90 // TODO(satok): Support generic way to emulate clock count. 91 // TODO(satok): Support other architectures if wanted. 92 // Returning dummy clock when can't access to the counter 93 return DUMMY_CYCLE_CLOCK; 94 #endif 95 } 96 97 // Return cycle counter frequency. 98 // As this method caches the cpu frequency internally, 99 // the first call will incur overhead, but not subsequent calls. 100 #if (defined(__powerpc__) || \ 101 defined(__ppc__) && (__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)) || \ 102 (defined(__s390x__)) 103 static uint64 GetCycleCounterFrequency(); 104 #else 105 static int64 GetCycleCounterFrequency(); 106 #endif 107 108 // Return micro second per each clock 109 // As this method caches the cpu frequency internally, 110 // the first call will incur overhead, but not subsequent calls. 111 static double GetMicroSecPerClock(); 112 113 // Reset clock cycle 114 // Resetting clock cycle is recommended to prevent 115 // clock cycle counters from overflowing on some platforms. 116 static void ResetClockCycle(); 117 118 // Enable clock cycle profile 119 // You can enable / disable profile if it's supported by the platform 120 static void EnableClockCycleProfiling(bool enable); 121 122 // Return chrono::duration per each clock 123 static std::chrono::duration<double> ConvertClockCycleToTime( 124 const int64 clock_cycle); 125 126 private: 127 class DefaultCpuUtilsHelper : public ICpuUtilsHelper { 128 public: 129 DefaultCpuUtilsHelper() = default; 130 void ResetClockCycle() final {} 131 uint64 GetCurrentClockCycle() final { return DUMMY_CYCLE_CLOCK; } 132 void EnableClockCycleProfiling(bool /* enable */) final {} 133 int64 CalculateCpuFrequency() final { return INVALID_FREQUENCY; } 134 135 private: 136 TF_DISALLOW_COPY_AND_ASSIGN(DefaultCpuUtilsHelper); 137 }; 138 139 // Return cpu frequency. 140 // CAVEAT: as this method calls system call and parse the mssage, 141 // this call may be slow. This is why this class caches the value by 142 // StaticVariableInitializer. 143 static int64 GetCycleCounterFrequencyImpl(); 144 145 // Return a singleton of ICpuUtilsHelper 146 // ICpuUtilsHelper is declared as a function-local static variable 147 // for the following two reasons: 148 // 1. Avoid passing instances to all classes which want 149 // to use profiling tools in CpuUtils 150 // 2. Minimize the overhead of acquiring ICpuUtilsHelper 151 static ICpuUtilsHelper& GetCpuUtilsHelperSingletonInstance(); 152 153 TF_DISALLOW_COPY_AND_ASSIGN(CpuUtils); 154 }; 155 156 } // namespace profile_utils 157 158 } // namespace tensorflow 159 160 #endif // TENSORFLOW_PLATFORM_PROFILEUTILS_CPU_UTILS_H__ 161