| 1 | // Licensed to the .NET Foundation under one or more agreements. |
| 2 | // The .NET Foundation licenses this file to you under the MIT license. |
| 3 | // See the LICENSE file in the project root for more information. |
| 4 | |
| 5 | #include "stdafx.h" |
| 6 | |
| 7 | #include "cycletimer.h" |
| 8 | #include "winbase.h" |
| 9 | #include "winwrap.h" |
| 10 | #include "assert.h" |
| 11 | #include "utilcode.h" |
| 12 | |
| 13 | bool CycleTimer::GetThreadCyclesS(unsigned __int64* cycles) |
| 14 | { |
| 15 | BOOL res = FALSE; |
| 16 | res = QueryThreadCycleTime(GetCurrentThread(), cycles); |
| 17 | return res != FALSE; |
| 18 | } |
| 19 | |
| 20 | static const int SampleLoopSize = 1000000; |
| 21 | |
| 22 | // static |
| 23 | double CycleTimer::CyclesPerSecond() |
| 24 | { |
| 25 | // Windows does not provide a way of converting cycles to time -- reasonably enough, |
| 26 | // since the frequency of a machine may vary, due, e.g., to power management. |
| 27 | // Windows *does* allow you to translate QueryPerformanceCounter counts into time, |
| 28 | // however. So we'll assume that the clock speed stayed constant, and measure both the |
| 29 | // QPC counts and cycles of a short loop, to get a conversion factor. |
| 30 | LARGE_INTEGER lpFrequency; |
| 31 | if (!QueryPerformanceFrequency(&lpFrequency)) return 0.0; |
| 32 | // Otherwise... |
| 33 | LARGE_INTEGER qpcStart; |
| 34 | unsigned __int64 cycleStart; |
| 35 | if (!QueryPerformanceCounter(&qpcStart)) return 0.0; |
| 36 | if (!GetThreadCyclesS(&cycleStart)) return 0.0; |
| 37 | volatile int sum = 0; |
| 38 | for (int k = 0; k < SampleLoopSize; k++) |
| 39 | { |
| 40 | sum += k; |
| 41 | } |
| 42 | LARGE_INTEGER qpcEnd; |
| 43 | if (!QueryPerformanceCounter(&qpcEnd)) return 0.0; |
| 44 | unsigned __int64 cycleEnd; |
| 45 | if (!GetThreadCyclesS(&cycleEnd)) return 0.0; |
| 46 | |
| 47 | double qpcTicks = ((double)qpcEnd.QuadPart) - ((double)qpcStart.QuadPart); |
| 48 | double secs = (qpcTicks / ((double)lpFrequency.QuadPart)); |
| 49 | double cycles = ((double)cycleEnd) - ((double)cycleStart); |
| 50 | return cycles / secs; |
| 51 | } |
| 52 | |
| 53 | // static |
| 54 | unsigned __int64 CycleTimer::QueryOverhead() |
| 55 | { |
| 56 | unsigned __int64 tot = 0; |
| 57 | unsigned __int64 startCycles; |
| 58 | unsigned __int64 endCycles; |
| 59 | const int N = 1000; |
| 60 | bool b = GetThreadCyclesS(&startCycles); assert(b); |
| 61 | for (int i = 0; i < N; i++) |
| 62 | { |
| 63 | b = GetThreadCyclesS(&endCycles); assert(b); |
| 64 | tot += (endCycles-startCycles); |
| 65 | startCycles = endCycles; |
| 66 | } |
| 67 | return tot/N; |
| 68 | } |
| 69 | |
| 70 | // static |
| 71 | void CycleTimer::InterlockedAddU64(unsigned __int64* loc, unsigned __int64 amount) |
| 72 | { |
| 73 | volatile __int64* vloc = (volatile __int64*)loc; |
| 74 | unsigned __int64 prev = *vloc; |
| 75 | for (;;) |
| 76 | { |
| 77 | unsigned __int64 next = prev + amount; |
| 78 | __int64 snext = (__int64)next; |
| 79 | __int64 sprev = (__int64)prev; |
| 80 | __int64 res = InterlockedCompareExchange64(vloc, snext, sprev); |
| 81 | if (res == sprev) return; |
| 82 | else prev = (unsigned __int64)res; |
| 83 | } |
| 84 | } |
| 85 | |
| 86 | |