timing.h

Go to the documentation of this file.

//
// Copyright 2016 Pixar
//
// Licensed under the terms set forth in the LICENSE.txt file available at
// https://openusd.org/license.
//
#ifndef PXR_BASE_ARCH_TIMING_H
#define PXR_BASE_ARCH_TIMING_H
 
 
#include "pxr/pxr.h"
#include "pxr/base/arch/api.h"
#include "pxr/base/arch/defines.h"
#include "pxr/base/arch/inttypes.h"
 
 
#if defined(ARCH_OS_LINUX) && defined(ARCH_CPU_INTEL)
#include <x86intrin.h>
#elif defined(ARCH_OS_DARWIN)
#include <mach/mach_time.h>
#elif defined(ARCH_OS_WINDOWS)
#include <intrin.h>
#endif
 
#include <algorithm>
#include <atomic>
#include <iterator>
#include <numeric>
 
PXR_NAMESPACE_OPEN_SCOPE
 
inline uint64_t
ArchGetTickTime()
{
#if defined(ARCH_OS_DARWIN)
    // On Darwin we'll use mach_absolute_time().
    return mach_absolute_time();
#elif defined(ARCH_CPU_INTEL)
    // On Intel we'll use the rdtsc instruction.
    return __rdtsc();
#elif defined (ARCH_CPU_ARM)
    uint64_t result;
    #if defined(ARCH_COMPILER_MSVC)
    // MSVC does not support inline assembly on ARM64 platforms
    // 0x5F02 == ARM64_CNTVCT - manually calculated value avoids <windows.h>
    result = _ReadStatusReg(0x5F02);
    #else
    __asm __volatile("mrs       %0, CNTVCT_EL0" : "=&r" (result));
    #endif
    return result;
#else
#error Unknown architecture.
#endif
}
 
 
inline uint64_t
ArchGetStartTickTime()
{
    uint64_t t;
#if defined (ARCH_OS_DARWIN) || \
    (defined (ARCH_CPU_ARM) && defined (ARCH_COMPILER_MSVC))
    return ArchGetTickTime();
#elif defined (ARCH_CPU_ARM)
    std::atomic_signal_fence(std::memory_order_seq_cst);
    asm volatile("mrs %0, cntvct_el0" : "=r"(t));
    std::atomic_signal_fence(std::memory_order_seq_cst);
#elif defined (ARCH_COMPILER_MSVC)
    _mm_lfence();
    std::atomic_signal_fence(std::memory_order_seq_cst);
    t = __rdtsc();
    _mm_lfence();
    std::atomic_signal_fence(std::memory_order_seq_cst);
#elif defined(ARCH_CPU_INTEL) && \
    (defined(ARCH_COMPILER_CLANG) || defined(ARCH_COMPILER_GCC))
    // Prevent reorders by the compiler.
    std::atomic_signal_fence(std::memory_order_seq_cst);
    asm volatile(
        "lfence\n\t"
        "rdtsc\n\t"
        "shl $32, %%rdx\n\t"
        "or %%rdx, %0\n\t"
        "lfence"
        : "=a"(t)
        :
        // rdtsc writes rdx
        // shl modifies cc flags
        : "rdx", "cc");
#else
#error "Unsupported architecture."
#endif
  return t;
}
 
inline uint64_t
ArchGetStopTickTime()
{
    uint64_t t;
#if defined (ARCH_OS_DARWIN) || \
    (defined (ARCH_CPU_ARM) && defined (ARCH_COMPILER_MSVC))
    return ArchGetTickTime();
#elif defined (ARCH_CPU_ARM)
    std::atomic_signal_fence(std::memory_order_seq_cst);
    asm volatile("mrs %0, cntvct_el0" : "=r"(t));
    std::atomic_signal_fence(std::memory_order_seq_cst);
#elif defined (ARCH_COMPILER_MSVC)
    std::atomic_signal_fence(std::memory_order_seq_cst);
    unsigned aux;
    t = __rdtscp(&aux);
    _mm_lfence();
    std::atomic_signal_fence(std::memory_order_seq_cst);
#elif defined(ARCH_CPU_INTEL) && \
    (defined(ARCH_COMPILER_CLANG) || defined(ARCH_COMPILER_GCC))
    std::atomic_signal_fence(std::memory_order_seq_cst);
    asm volatile(
        "rdtscp\n\t"
        "shl $32, %%rdx\n\t"
        "or %%rdx, %0\n\t"
        "lfence"
        : "=a"(t)
        :
        // rdtscp writes rcx & rdx
        // shl modifies cc flags
        : "rcx", "rdx", "cc");
#else
#error "Unsupported architecture."
#endif
  return t;
}
 
#if defined (doxygen) ||                                                       \
    (!defined(ARCH_OS_DARWIN) && defined(ARCH_CPU_INTEL) &&                    \
     (defined(ARCH_COMPILER_CLANG) || defined(ARCH_COMPILER_GCC)))
 
struct ArchIntervalTimer
{
    explicit ArchIntervalTimer(bool start=true)
        : _started(start) {
        if (_started) {
            Start();
        }
    }
 
    void Start() {
        _started = true;
        std::atomic_signal_fence(std::memory_order_seq_cst);
        asm volatile(
            "lfence\n\t"
            "rdtsc\n\t"
            "lfence"
            : "=a"(_startLow), "=d"(_startHigh) :: );
    }
 
    bool IsStarted() const {
        return _started;
    }
 
    uint64_t GetStartTicks() const {
        return (uint64_t(_startHigh) << 32) + _startLow;
    }
 
    uint64_t GetCurrentTicks() {
        return ArchGetStopTickTime();
    }
 
    uint64_t GetElapsedTicks() {
        if (!_started) {
            return 0;
        }
        uint32_t stopLow, stopHigh;
        std::atomic_signal_fence(std::memory_order_seq_cst);
        asm volatile(
            "rdtscp\n\t"
            "lfence"
            : "=a"(stopLow), "=d"(stopHigh)
            :
            // rdtscp writes rcx
            : "rcx");
        return ((uint64_t(stopHigh) << 32) + stopLow) -
            ((uint64_t(_startHigh) << 32) + _startLow);
    }
private:
    bool _started = false;
    uint32_t _startLow = 0, _startHigh = 0;
};
 
#else
 
struct ArchIntervalTimer
{
    explicit ArchIntervalTimer(bool start=true)
        : _started(start) {
        if (_started) {
            _startTicks = ArchGetStartTickTime();
        }
    }
 
    void Start() {
        _started = true;
        _startTicks = ArchGetStartTickTime();
    }
 
    bool IsStarted() const {
        return _started;
    }
 
    uint64_t GetStartTicks() const {
        return _startTicks;
    }
 
    uint64_t GetCurrentTicks() {
        return ArchGetStopTickTime();
    }
 
    uint64_t GetElapsedTicks() {
        if (!_started) {
            return 0;
        }
        return ArchGetStopTickTime() - _startTicks;
    }
private:
    bool _started = false;
    uint64_t _startTicks;
};
 
#endif
 
ARCH_API
uint64_t ArchGetTickQuantum();
 
ARCH_API
uint64_t ArchGetIntervalTimerTickOverhead();
 
 
ARCH_API
int64_t ArchTicksToNanoseconds(uint64_t nTicks);
 
ARCH_API
double ArchTicksToSeconds(uint64_t nTicks);
 
ARCH_API
uint64_t ArchSecondsToTicks(double seconds);
    
ARCH_API
double ArchGetNanosecondsPerTick();
 
ARCH_API
uint64_t
Arch_MeasureExecutionTime(uint64_t maxTicks, bool *reachedConsensus,
                          void const *m, uint64_t (*callM)(void const *, int));
 
template <class Fn>
uint64_t
ArchMeasureExecutionTime(
    Fn const &fn,
    uint64_t maxTicks = 1e7,
    bool *reachedConsensus = nullptr)
{
    auto measureN = [&fn](int nTimes) -> uint64_t {
        ArchIntervalTimer iTimer;
        for (int i = nTimes; i--; ) {
            std::atomic_signal_fence(std::memory_order_seq_cst);
            (void)fn();
            std::atomic_signal_fence(std::memory_order_seq_cst);
        }
        return iTimer.GetElapsedTicks();
    };
 
    using MeasureNType = decltype(measureN);
    
    return Arch_MeasureExecutionTime(
        maxTicks, reachedConsensus,
        static_cast<void const *>(&measureN),
        [](void const *mN, int nTimes) {
            return (*static_cast<MeasureNType const *>(mN))(nTimes);
        });
}
 
 
PXR_NAMESPACE_CLOSE_SCOPE
 
#endif // PXR_BASE_ARCH_TIMING_H