Add perf benchmark utility. (#2977)

* Add perf benchmark utility.

* Update documentation.

* Fix.

* Fix doc.

---------

Co-authored-by: Yong He <yhe@nvidia.com>

15 files changed, 145 insertions, 4 deletions

diff --git a/source/core/slang-performance-profiler.cpp b/source/core/slang-performance-profiler.cpp
new file mode 100644
index 000000000..f5f74f99d
--- /dev/null
+++ b/source/core/slang-performance-profiler.cpp
@@ -0,0 +1,52 @@
+#include "slang-performance-profiler.h"
+#include "slang-dictionary.h"
+
+namespace Slang
+{
+    struct FuncProfileInfo
+    {
+        int invocationCount = 0;
+        std::chrono::nanoseconds duration = std::chrono::nanoseconds::zero();
+    };
+    class PerformanceProfilerImpl : public PerformanceProfiler
+    {
+    public:
+        OrderedDictionary<const char*, FuncProfileInfo> data;
+
+        virtual FuncProfileContext enterFunction(const char* funcName) override
+        {
+            auto entry = data.tryGetValue(funcName);
+            if (!entry)
+            {
+                data.add(funcName, FuncProfileInfo());
+                entry = data.tryGetValue(funcName);
+            }
+            entry->invocationCount++;
+            FuncProfileContext ctx;
+            ctx.funcName = funcName;
+            ctx.startTime = std::chrono::high_resolution_clock::now();
+            return ctx;
+        }
+        virtual void exitFunction(FuncProfileContext ctx) override
+        {
+            auto endTime = std::chrono::high_resolution_clock::now();
+            auto duration = endTime - ctx.startTime;
+            auto entry = data.tryGetValue(ctx.funcName);
+            entry->duration += duration;
+        }
+        virtual void getResult(StringBuilder& out) override
+        {
+            for (auto func : data)
+            {
+                out << func.key << ": \t";
+                out << func.value.invocationCount << "\t" << func.value.duration.count()/1000000 << "\n";
+            }
+        }
+    };
+
+    PerformanceProfiler* Slang::PerformanceProfiler::getProfiler()
+    {
+        static PerformanceProfilerImpl profiler = PerformanceProfilerImpl();
+        return &profiler;
+    }
+}
diff --git a/source/core/slang-performance-profiler.h b/source/core/slang-performance-profiler.h
new file mode 100644
index 000000000..323940ade
--- /dev/null
+++ b/source/core/slang-performance-profiler.h
@@ -0,0 +1,42 @@
+#ifndef SLANG_CORE_PERFORMANCE_PROFILER_H
+#define SLANG_CORE_PERFORMANCE_PROFILER_H
+
+#include "slang-string.h"
+#include <chrono>
+
+namespace Slang
+{
+
+struct FuncProfileContext
+{
+    const char* funcName = nullptr;
+    std::chrono::time_point<std::chrono::high_resolution_clock> startTime;
+};
+
+class PerformanceProfiler
+{
+public:
+    virtual FuncProfileContext enterFunction(const char* funcName) = 0;
+    virtual void exitFunction(FuncProfileContext context) = 0;
+    virtual void getResult(StringBuilder& out) = 0;
+public:
+    static PerformanceProfiler* getProfiler();
+};
+
+struct PerformanceProfilerFuncRAIIContext
+{
+    FuncProfileContext context;
+    PerformanceProfilerFuncRAIIContext(const char* funcName)
+    {
+        context = PerformanceProfiler::getProfiler()->enterFunction(funcName);
+    }
+    ~PerformanceProfilerFuncRAIIContext()
+    {
+        PerformanceProfiler::getProfiler()->exitFunction(context);
+    }
+};
+
+#define SLANG_PROFILE PerformanceProfilerFuncRAIIContext _profileContext(__func__)
+}
+
+#endif
diff --git a/source/slang/slang-compiler.h b/source/slang/slang-compiler.h
index 7082eccd7..e6385c76d 100755
--- a/source/slang/slang-compiler.h
+++ b/source/slang/slang-compiler.h
@@ -2646,6 +2646,7 @@ namespace Slang
         virtual SLANG_NO_THROW void SLANG_MCALL setDiagnosticFlags(SlangDiagnosticFlags flags) SLANG_OVERRIDE;
         virtual SLANG_NO_THROW void SLANG_MCALL setDebugInfoFormat(SlangDebugInfoFormat format) SLANG_OVERRIDE;
         virtual SLANG_NO_THROW void SLANG_MCALL setReportDownstreamTime(bool value) SLANG_OVERRIDE;
+        virtual SLANG_NO_THROW void SLANG_MCALL setReportPerfBenchmark(bool value) SLANG_OVERRIDE;
         void setHLSLToVulkanLayoutOptions(int targetIndex, HLSLToVulkanLayoutOptions* vulkanLayoutOptions);
 
         EndToEndCompileRequest(
@@ -2691,6 +2692,9 @@ namespace Slang
         bool m_isCommandLineCompile = false;
 
         bool m_reportDownstreamCompileTime = false;
+
+        // If set, will print out compiler performance benchmark results.
+        bool m_reportPerfBenchmark = false;
         
         String m_diagnosticOutput;
 
diff --git a/source/slang/slang-diagnostic-defs.h b/source/slang/slang-diagnostic-defs.h
index 92f79eac5..1832a3b46 100644
--- a/source/slang/slang-diagnostic-defs.h
+++ b/source/slang/slang-diagnostic-defs.h
@@ -143,7 +143,7 @@ DIAGNOSTIC(    99, Error, unknownDebugOption, "unknown debug option, known optio
 DIAGNOSTIC(  100, Error, failedToLoadDownstreamCompiler, "failed to load downstream compiler '$0'")
 DIAGNOSTIC(  101, Error, downstreamCompilerDoesntSupportWholeProgramCompilation, "downstream compiler '$0' doesn't support whole program compilation")
 DIAGNOSTIC(  102, Note,  downstreamCompileTime, "downstream compile time: $0s")
-
+DIAGNOSTIC(  103, Note,  performanceBenchmarkResult, "compiler performance benchmark:\n$0")
 DIAGNOSTIC(99999, Note, noteFailedToLoadDynamicLibrary, "failed to load dynamic library '$0'")
 
 //
diff --git a/source/slang/slang-emit.cpp b/source/slang/slang-emit.cpp
index 41230d169..d71cc5507 100644
--- a/source/slang/slang-emit.cpp
+++ b/source/slang/slang-emit.cpp
@@ -89,6 +89,7 @@
 #include "../compiler-core/slang-artifact-associated-impl.h"
 
 #include "../core/slang-castable.h"
+#include "../core/slang-performance-profiler.h"
 
 #include <assert.h>
 
@@ -199,6 +200,7 @@ Result linkAndOptimizeIR(
     LinkingAndOptimizationOptions const&    options,
     LinkedIR&                               outLinkedIR)
 {
+    SLANG_PROFILE;
     auto session = codeGenContext->getSession();
     auto sink = codeGenContext->getSink();
     auto target = codeGenContext->getTargetFormat();
diff --git a/source/slang/slang-ir-autodiff.cpp b/source/slang/slang-ir-autodiff.cpp
index 6d56736ad..2231e0ce7 100644
--- a/source/slang/slang-ir-autodiff.cpp
+++ b/source/slang/slang-ir-autodiff.cpp
@@ -6,6 +6,7 @@
 #include "slang-ir-single-return.h"
 #include "slang-ir-ssa-simplification.h"
 #include "slang-ir-validate.h"
+#include "../core/slang-performance-profiler.h"
 
 namespace Slang
 {
@@ -1726,6 +1727,7 @@ bool processAutodiffCalls(
     DiagnosticSink*                     sink,
     IRAutodiffPassOptions const&)
 {
+    SLANG_PROFILE;
     bool modified = false;
 
     // Create shared context for all auto-diff related passes
diff --git a/source/slang/slang-ir-inline.cpp b/source/slang/slang-ir-inline.cpp
index f825a9461..db4bb50a8 100644
--- a/source/slang/slang-ir-inline.cpp
+++ b/source/slang/slang-ir-inline.cpp
@@ -3,6 +3,7 @@
 
 #include "slang-ir-ssa-simplification.h"
 
+#include "../core/slang-performance-profiler.h"
 // This file provides general facilities for inlining function calls.
 
 //
@@ -671,6 +672,8 @@ struct MandatoryEarlyInliningPass : InliningPassBase
 
 void performMandatoryEarlyInlining(IRModule* module)
 {
+    SLANG_PROFILE;
+
     MandatoryEarlyInliningPass pass(module);
     pass.considerAllCallSites();
 }
@@ -785,6 +788,8 @@ struct ForceInliningPass : InliningPassBase
 
 void performForceInlining(IRModule* module)
 {
+    SLANG_PROFILE;
+
     ForceInliningPass pass(module);
     pass.considerAllCallSites();
 }
diff --git a/source/slang/slang-ir-legalize-types.cpp b/source/slang/slang-ir-legalize-types.cpp
index 162498bf0..a88d43db3 100644
--- a/source/slang/slang-ir-legalize-types.cpp
+++ b/source/slang/slang-ir-legalize-types.cpp
@@ -11,6 +11,7 @@
 // that the concrete type of everything is known.
 
 #include "../compiler-core/slang-name.h"
+#include "../core/slang-performance-profiler.h"
 
 #include "slang-ir.h"
 #include "slang-ir-clone.h"
@@ -3843,6 +3844,8 @@ void legalizeResourceTypes(
     IRModule*       module,
     DiagnosticSink* sink)
 {
+    SLANG_PROFILE;
+
     SLANG_UNUSED(sink);
 
     IRResourceTypeLegalizationContext context(module);
@@ -3853,6 +3856,8 @@ void legalizeExistentialTypeLayout(
     IRModule*       module,
     DiagnosticSink* sink)
 {
+    SLANG_PROFILE;
+
     SLANG_UNUSED(module);
     SLANG_UNUSED(sink);
 
diff --git a/source/slang/slang-ir-link.cpp b/source/slang/slang-ir-link.cpp
index f57919e44..081fd7486 100644
--- a/source/slang/slang-ir-link.cpp
+++ b/source/slang/slang-ir-link.cpp
@@ -9,6 +9,7 @@
 #include "slang-ir-autodiff.h"
 #include "slang-module-library.h"
 
+#include "../core/slang-performance-profiler.h"
 #include "../compiler-core/slang-artifact.h"
 
 namespace Slang
@@ -1428,6 +1429,8 @@ static bool _isPublicOrHLSLExported(IRInst* inst)
 LinkedIR linkIR(
     CodeGenContext* codeGenContext)
 {
+    SLANG_PROFILE;
+
     auto linkage = codeGenContext->getLinkage();
     auto program = codeGenContext->getProgram();
     auto session = codeGenContext->getSession();
diff --git a/source/slang/slang-ir-loop-unroll.cpp b/source/slang/slang-ir-loop-unroll.cpp
index 05e829827..36a2abbfc 100644
--- a/source/slang/slang-ir-loop-unroll.cpp
+++ b/source/slang/slang-ir-loop-unroll.cpp
@@ -7,6 +7,7 @@
 #include "slang-ir-util.h"
 #include "slang-ir-simplify-cfg.h"
 #include "slang-ir-dce.h"
+#include "../core/slang-performance-profiler.h"
 
 namespace Slang
 {
@@ -508,6 +509,8 @@ bool unrollLoopsInFunc(
 
 bool unrollLoopsInModule(IRModule* module, DiagnosticSink* sink)
 {
+    SLANG_PROFILE;
+
     for (auto inst : module->getGlobalInsts())
     {
         if (auto genFunc = as<IRGeneric>(inst))
diff --git a/source/slang/slang-ir-lower-generics.cpp b/source/slang/slang-ir-lower-generics.cpp
index fc810e54e..6a5abbf8c 100644
--- a/source/slang/slang-ir-lower-generics.cpp
+++ b/source/slang/slang-ir-lower-generics.cpp
@@ -15,7 +15,7 @@
 #include "slang-ir-witness-table-wrapper.h"
 #include "slang-ir-ssa-simplification.h"
 #include "slang-ir-util.h"
-
+#include "../core/slang-performance-profiler.h"
 
 namespace Slang
 {
@@ -181,6 +181,8 @@ namespace Slang
         IRModule*               module,
         DiagnosticSink*         sink)
     {
+        SLANG_PROFILE;
+
         SharedGenericsLoweringContext sharedContext;
         sharedContext.targetReq = targetReq;
         sharedContext.module = module;
diff --git a/source/slang/slang-ir-specialize.cpp b/source/slang/slang-ir-specialize.cpp
index 80dcec82b..48b083b7d 100644
--- a/source/slang/slang-ir-specialize.cpp
+++ b/source/slang/slang-ir-specialize.cpp
@@ -7,6 +7,7 @@
 #include "slang-ir-ssa-simplification.h"
 #include "slang-ir-lower-witness-lookup.h"
 #include "slang-ir-dce.h"
+#include "../core/slang-performance-profiler.h"
 
 namespace Slang
 {
@@ -2340,6 +2341,7 @@ bool specializeModule(
     IRModule*   module,
     DiagnosticSink* sink)
 {
+    SLANG_PROFILE;
     SpecializationContext context;
     context.module = module;
     context.sink = sink;
diff --git a/source/slang/slang-ir-ssa-simplification.cpp b/source/slang/slang-ir-ssa-simplification.cpp
index 66b4ceccb..43c9f8d2e 100644
--- a/source/slang/slang-ir-ssa-simplification.cpp
+++ b/source/slang/slang-ir-ssa-simplification.cpp
@@ -11,6 +11,7 @@
 #include "slang-ir-remove-unused-generic-param.h"
 #include "slang-ir-redundancy-removal.h"
 #include "slang-ir-propagate-func-properties.h"
+#include "../core/slang-performance-profiler.h"
 
 namespace Slang
 {
@@ -18,6 +19,7 @@ namespace Slang
     // until no more changes are possible.
     void simplifyIR(IRModule* module, DiagnosticSink* sink)
     {
+        SLANG_PROFILE;
         bool changed = true;
         const int kMaxIterations = 8;
         const int kMaxFuncIterations = 16;
diff --git a/source/slang/slang-options.cpp b/source/slang/slang-options.cpp
index 09c46b7ef..ad061a4e5 100644
--- a/source/slang/slang-options.cpp
+++ b/source/slang/slang-options.cpp
@@ -72,6 +72,7 @@ enum class OptionKind
     InputFilesRemain,
     EmitIr,
     ReportDownstreamTime,
+    ReportPerfBenchmark,
 
     SourceEmbedStyle,
     SourceEmbedName,
@@ -427,6 +428,7 @@ void initCommandOptions(CommandOptions& options)
         { OptionKind::DumpWarningDiagnostics, "-dump-warning-diagnostics", nullptr, "Dump to output list of warning diagnostic numeric and name ids." },
         { OptionKind::InputFilesRemain, "--", nullptr, "Treat the rest of the command line as input files."},
         { OptionKind::ReportDownstreamTime, "-report-downstream-time", nullptr, "Reports the time spent in the downstream compiler." },
+        { OptionKind::ReportPerfBenchmark, "-report-perf-benchmark", nullptr, "Reports compiler performance benchmark results." },
         { OptionKind::SourceEmbedStyle, "-source-embed-style", "-source-embed-style <source-embed-style>",
         "If source embedding is enabled, defines the style used. When enabled (with any style other than `none`), "
         "will write compile results into embeddable source for the target language. "
@@ -1867,6 +1869,11 @@ SlangResult OptionsParser::_parse(
                 m_compileRequest->setReportDownstreamTime(true);
                 break;
             }
+            case OptionKind::ReportPerfBenchmark:
+            {
+                m_compileRequest->setReportPerfBenchmark(true);
+                break;
+            }
             case OptionKind::ModuleName:
             {
                 CommandLineArg moduleName;
diff --git a/source/slang/slang.cpp b/source/slang/slang.cpp
index daa9cda3b..ffeabb0bd 100644
--- a/source/slang/slang.cpp
+++ b/source/slang/slang.cpp
@@ -7,7 +7,7 @@
 #include "../core/slang-type-text-util.h"
 #include "../core/slang-type-convert-util.h"
 #include "../core/slang-castable.h"
-
+#include "../core/slang-performance-profiler.h"
 // Artifact
 #include "../compiler-core/slang-artifact-impl.h"
 #include "../compiler-core/slang-artifact-desc-util.h"
@@ -4813,6 +4813,11 @@ void EndToEndCompileRequest::setReportDownstreamTime(bool value)
     m_reportDownstreamCompileTime = value;
 }
 
+void EndToEndCompileRequest::setReportPerfBenchmark(bool value)
+{
+    m_reportPerfBenchmark = value;
+}
+
 void EndToEndCompileRequest::setDiagnosticCallback(SlangDiagnosticCallback callback, void const* userData)
 {
     ComPtr<ISlangWriter> writer(new CallbackWriter(callback, userData, WriterFlag::IsConsole));
@@ -5181,7 +5186,12 @@ SlangResult EndToEndCompileRequest::EndToEndCompileRequest::compile()
         double downstreamTime = downstreamEndTime - downstreamStartTime;
         String downstreamTimeStr = String(downstreamTime, "%.2f");
         getSink()->diagnose(SourceLoc(), Diagnostics::downstreamCompileTime, downstreamTimeStr);
-
+    }
+    if (m_reportPerfBenchmark)
+    {
+        StringBuilder perfResult;
+        PerformanceProfiler::getProfiler()->getResult(perfResult);
+        getSink()->diagnose(SourceLoc(), Diagnostics::performanceBenchmarkResult, perfResult.produceString());
     }
 
     // Repro dump handling