[llvm-exegesis] Add an analysis mode.
Summary:
The analysis mode gives the user a clustered view of the measurement results.
Next steps are (requires the split ok AsmTemplate.Name into {mnemonic, mode}):
- Show the sched class.
- Highlight any inconsistencies with the checked-in data.
Reviewers: gchatelet
Subscribers: mgorny, llvm-commits, tschuett
Differential Revision: https://reviews.llvm.org/D46865
llvm-svn: 332344
diff --git a/llvm/tools/llvm-exegesis/llvm-exegesis.cpp b/llvm/tools/llvm-exegesis/llvm-exegesis.cpp
index 7768357..fff1877 100644
--- a/llvm/tools/llvm-exegesis/llvm-exegesis.cpp
+++ b/llvm/tools/llvm-exegesis/llvm-exegesis.cpp
@@ -12,8 +12,10 @@
///
//===----------------------------------------------------------------------===//
+#include "lib/Analysis.h"
#include "lib/BenchmarkResult.h"
#include "lib/BenchmarkRunner.h"
+#include "lib/Clustering.h"
#include "lib/Latency.h"
#include "lib/LlvmState.h"
#include "lib/PerfHelper.h"
@@ -23,8 +25,11 @@
#include "llvm/ADT/Twine.h"
#include "llvm/MC/MCInstBuilder.h"
#include "llvm/MC/MCRegisterInfo.h"
+#include "llvm/MC/MCSubtargetInfo.h"
#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Format.h"
#include "llvm/Support/Path.h"
+#include "llvm/Support/TargetRegistry.h"
#include "llvm/Support/TargetSelect.h"
#include <algorithm>
#include <random>
@@ -39,26 +44,41 @@
OpcodeName("opcode-name", llvm::cl::desc("opcode to measure, by name"),
llvm::cl::init(""));
-enum class BenchmarkModeE { Latency, Uops };
-static llvm::cl::opt<BenchmarkModeE>
- BenchmarkMode("benchmark-mode", llvm::cl::desc("the benchmark mode to run"),
- llvm::cl::values(clEnumValN(BenchmarkModeE::Latency,
- "latency", "Instruction Latency"),
- clEnumValN(BenchmarkModeE::Uops, "uops",
- "Uop Decomposition")));
+static llvm::cl::opt<std::string>
+ BenchmarkFile("benchmarks-file", llvm::cl::desc(""), llvm::cl::init("-"));
+
+enum class BenchmarkModeE { Latency, Uops, Analysis };
+static llvm::cl::opt<BenchmarkModeE> BenchmarkMode(
+ "benchmark-mode", llvm::cl::desc("the benchmark mode to run"),
+ llvm::cl::values(
+ clEnumValN(BenchmarkModeE::Latency, "latency", "Instruction Latency"),
+ clEnumValN(BenchmarkModeE::Uops, "uops", "Uop Decomposition"),
+ clEnumValN(BenchmarkModeE::Analysis, "analysis", "Analysis")));
static llvm::cl::opt<unsigned>
NumRepetitions("num-repetitions",
llvm::cl::desc("number of time to repeat the asm snippet"),
llvm::cl::init(10000));
+static llvm::cl::opt<unsigned> AnalysisNumPoints(
+ "analysis-numpoints",
+ llvm::cl::desc("minimum number of points in an analysis cluster"),
+ llvm::cl::init(3));
+
+static llvm::cl::opt<float>
+ AnalysisEpsilon("analysis-epsilon",
+ llvm::cl::desc("dbscan epsilon for analysis clustering"),
+ llvm::cl::init(0.1));
+
namespace exegesis {
-void main() {
- if (OpcodeName.empty() == (OpcodeIndex == 0)) {
+void benchmarkMain() {
+ if (exegesis::pfm::pfmInitialize())
+ llvm::report_fatal_error("cannot initialize libpfm");
+
+ if (OpcodeName.empty() == (OpcodeIndex == 0))
llvm::report_fatal_error(
"please provide one and only one of 'opcode-index' or 'opcode-name'");
- }
llvm::InitializeNativeTarget();
llvm::InitializeNativeTargetAsmPrinter();
@@ -94,10 +114,42 @@
case BenchmarkModeE::Uops:
Runner = llvm::make_unique<UopsBenchmarkRunner>();
break;
+ case BenchmarkModeE::Analysis:
+ llvm_unreachable("not a benchmark");
}
Runner->run(State, Opcode, NumRepetitions > 0 ? NumRepetitions : 1, Filter)
- .writeYamlOrDie("-");
+ .writeYamlOrDie(BenchmarkFile);
+ exegesis::pfm::pfmTerminate();
+}
+
+void analysisMain() {
+ // Read benchmarks.
+ const std::vector<InstructionBenchmark> Points =
+ InstructionBenchmark::readYamlsOrDie(BenchmarkFile);
+ llvm::outs() << "Parsed " << Points.size() << " benchmark points\n";
+ if (Points.empty()) {
+ llvm::errs() << "no benchmarks to analyze\n";
+ return;
+ }
+ // FIXME: Check that all points have the same triple/cpu.
+ // FIXME: Merge points from several runs (latency and uops).
+
+ llvm::InitializeAllTargets();
+ std::string Error;
+ const auto *TheTarget =
+ llvm::TargetRegistry::lookupTarget(Points[0].LLVMTriple, Error);
+ if (!TheTarget) {
+ llvm::errs() << "unknown target '" << Points[0].LLVMTriple << "'\n";
+ return;
+ }
+ std::unique_ptr<llvm::MCSubtargetInfo> STI(TheTarget->createMCSubtargetInfo(
+ Points[0].LLVMTriple, Points[0].CpuName, ""));
+
+ const auto Clustering = llvm::cantFail(InstructionBenchmarkClustering::create(
+ Points, AnalysisNumPoints, AnalysisEpsilon));
+ if (auto Err = printAnalysisClusters(Clustering, *STI, llvm::outs()))
+ llvm::report_fatal_error(std::move(Err));
}
} // namespace exegesis
@@ -105,13 +157,10 @@
int main(int Argc, char **Argv) {
llvm::cl::ParseCommandLineOptions(Argc, Argv, "");
- if (exegesis::pfm::pfmInitialize()) {
- llvm::errs() << "cannot initialize libpfm\n";
- return EXIT_FAILURE;
+ if (BenchmarkMode == BenchmarkModeE::Analysis) {
+ exegesis::analysisMain();
+ } else {
+ exegesis::benchmarkMain();
}
-
- exegesis::main();
-
- exegesis::pfm::pfmTerminate();
return EXIT_SUCCESS;
}