From bacf73cc68af8b7654cc920b07635f926b785b84 Mon Sep 17 00:00:00 2001 From: kaizhangNV <149626564+kaizhangNV@users.noreply.github.com> Date: Tue, 4 Jun 2024 18:24:12 -0700 Subject: Add APIs to get profile of compile time (#4242) * Add APIs to get profile of compile time Add serial time measurement Add profiler to measure lots of stages in slang compilation, and it can accumulate the time spent in each thread in multi-threads case and finally report a serial timing info. * Add invocation times to the profiler * Simplify the profiler and provide a 'clear' option Change the profiler design to only return the thread_local profiler to user. We create a ISlangProfiler interface to carry the thread_local variable PerformanceProfilerImpl profiler to user. In addition, we provide a new option in the input parameter to control whether or not user want to clear the previous profile data. So spGetCompileProfile() can always returns a fresh new profiling data. * Change to use slang container List Stop using std::vector, instead use slang's container List. Generate a UUID for ISlangProfiler --- source/core/slang-performance-profiler.cpp | 75 +++++++++++++++++++++++++++--- source/core/slang-performance-profiler.h | 31 ++++++++++++ 2 files changed, 100 insertions(+), 6 deletions(-) (limited to 'source/core') diff --git a/source/core/slang-performance-profiler.cpp b/source/core/slang-performance-profiler.cpp index 7ca33f615..f8e934491 100644 --- a/source/core/slang-performance-profiler.cpp +++ b/source/core/slang-performance-profiler.cpp @@ -3,11 +3,6 @@ namespace Slang { - struct FuncProfileInfo - { - int invocationCount = 0; - std::chrono::nanoseconds duration = std::chrono::nanoseconds::zero(); - }; class PerformanceProfilerImpl : public PerformanceProfiler { public: @@ -39,9 +34,14 @@ namespace Slang for (auto func : data) { out << func.key << ": \t"; - out << func.value.invocationCount << "\t" << func.value.duration.count() / 1000000.0f << "\n"; + auto milliseconds = std::chrono::duration_cast< std::chrono::milliseconds >(func.value.duration); + out << func.value.invocationCount << "\t" << milliseconds.count() << "ms\n"; } } + virtual void clear() override + { + data.clear(); + } }; PerformanceProfiler* Slang::PerformanceProfiler::getProfiler() @@ -49,4 +49,67 @@ namespace Slang thread_local static PerformanceProfilerImpl profiler = PerformanceProfilerImpl(); return &profiler; } + + SlangProfiler::SlangProfiler(PerformanceProfiler* profiler) + { + PerformanceProfilerImpl* profilerImpl = static_cast(profiler); + size_t entryCount = profilerImpl->data.getCount(); + + m_profilEntries.reserve(entryCount); + + int index = 0; + for (auto func : profilerImpl->data) + { + ProfileInfo profileEntry {}; + size_t strSize = std::min(sizeof(profileEntry.funcName) - 1, strlen(func.key)); + + if (strSize > 0) + { + strncpy_s(profileEntry.funcName, strSize, func.key, strSize); + } + profileEntry.invocationCount = func.value.invocationCount; + profileEntry.duration = func.value.duration; + + m_profilEntries.insert(index, profileEntry); + index++; + } + } + + ISlangUnknown* SlangProfiler::getInterface(const Guid& guid) + { + if(guid == SlangProfiler::getTypeGuid()) + return static_cast(this); + else + return nullptr; + } + + size_t SlangProfiler::getEntryCount() + { + return m_profilEntries.getCount(); + } + + const char* SlangProfiler::getEntryName(uint32_t index) + { + if (index >= (uint32_t)m_profilEntries.getCount()) + return nullptr; + + return m_profilEntries[index].funcName; + } + + long SlangProfiler::getEntryTimeMS(uint32_t index) + { + if (index >= (uint32_t)m_profilEntries.getCount()) + return 0; + + auto milliseconds = std::chrono::duration_cast< std::chrono::milliseconds >(m_profilEntries[index].duration); + return (long)milliseconds.count(); + } + + uint32_t SlangProfiler::getEntryInvocationTimes(uint32_t index) + { + if (index >= (uint32_t)m_profilEntries.getCount()) + return 0; + + return m_profilEntries[index].invocationCount; + } } diff --git a/source/core/slang-performance-profiler.h b/source/core/slang-performance-profiler.h index 323940ade..f7c62bb45 100644 --- a/source/core/slang-performance-profiler.h +++ b/source/core/slang-performance-profiler.h @@ -3,10 +3,19 @@ #include "slang-string.h" #include +#include +#include "../../slang-com-helper.h" +#include "../core/slang-list.h" namespace Slang { +struct FuncProfileInfo +{ + int invocationCount = 0; + std::chrono::nanoseconds duration = std::chrono::nanoseconds::zero(); +}; + struct FuncProfileContext { const char* funcName = nullptr; @@ -19,6 +28,7 @@ public: virtual FuncProfileContext enterFunction(const char* funcName) = 0; virtual void exitFunction(FuncProfileContext context) = 0; virtual void getResult(StringBuilder& out) = 0; + virtual void clear() = 0; public: static PerformanceProfiler* getProfiler(); }; @@ -36,6 +46,27 @@ struct PerformanceProfilerFuncRAIIContext } }; +struct SlangProfiler: public ISlangProfiler, public RefObject +{ +public: + SLANG_REF_OBJECT_IUNKNOWN_ALL + struct ProfileInfo + { + char funcName[256] = {0}; + int invocationCount = 0; + std::chrono::nanoseconds duration = std::chrono::nanoseconds::zero(); + }; + SlangProfiler(PerformanceProfiler * profiler); + ISlangUnknown* getInterface(const Guid& guid); + + virtual SLANG_NO_THROW size_t SLANG_MCALL getEntryCount() override; + virtual SLANG_NO_THROW const char* SLANG_MCALL getEntryName(uint32_t index) override; + virtual SLANG_NO_THROW long SLANG_MCALL getEntryTimeMS(uint32_t index) override; + virtual SLANG_NO_THROW uint32_t SLANG_MCALL getEntryInvocationTimes(uint32_t index) override; +private: + List m_profilEntries; +}; + #define SLANG_PROFILE PerformanceProfilerFuncRAIIContext _profileContext(__func__) } -- cgit v1.2.3