[libc-commits] [libc] [libc] NVPTX Profiling Draft (PR #92009)

[Joseph Huber via libc-commits]

================
@@ -0,0 +1,122 @@
+#ifndef LLVM_LIBC_BENCHMARKS_LIBC_GPU_BENCHMARK_H
+#define LLVM_LIBC_BENCHMARKS_LIBC_GPU_BENCHMARK_H
+
+#include "benchmarks/gpu/timing/timing.h"
+
+#include "benchmarks/gpu/BenchmarkLogger.h"
+
+#include <stddef.h>
+#include <stdint.h>
+
+namespace LIBC_NAMESPACE {
+
+namespace libc_gpu_benchmarks {
+
+struct BenchmarkOptions {
+  uint32_t InitialIterations = 1;
+  uint32_t MaxIterations = 10000000;
+  uint32_t MinSamples = 4;
+  uint32_t MaxSamples = 1000;
+  double Epsilon = 0.01;
+  double ScalingFactor = 1.4;
+};
+
+struct Measurement {
+  size_t Iterations = 0;
+  uint64_t ElapsedCycles = 0;
+};
+
+class RefinableRuntimeEstimation {
+  uint64_t TotalCycles = 0;
+  size_t TotalIterations = 0;
+
+public:
+  uint64_t Update(const Measurement &M) {
+    TotalCycles += M.ElapsedCycles;
+    TotalIterations += M.Iterations;
+    return TotalCycles / TotalIterations;
+  }
+};
+
+// Tracks the progression of the runtime estimation
+class RuntimeEstimationProgression {
+  RefinableRuntimeEstimation RRE;
+
+public:
+  uint64_t CurrentEstimation = 0;
+
+  double ComputeImprovement(const Measurement &M) {
+    const uint64_t NewEstimation = RRE.Update(M);
+    double Ratio = ((double)CurrentEstimation / NewEstimation) - 1.0;
+
+    // Get absolute value
+    if (Ratio < 0) {
+      Ratio *= -1;
+    }
+
+    CurrentEstimation = NewEstimation;
+    return Ratio;
+  }
+};
+
+struct BenchmarkResult {
+  uint64_t Cycles = 0;
+  size_t Samples = 0;
+  size_t TotalIterations = 0;
+};
+
+BenchmarkResult benchmark(const BenchmarkOptions &Options,
+                          uint64_t (*WrapperFunc)());
+
+class Benchmark {
+  Benchmark *Next = nullptr;
+
+public:
+  virtual ~Benchmark() {}
+  virtual void SetUp() {}
+  virtual void TearDown() {}
+
+  static int runBenchmarks();
+
+protected:
+  static void addBenchmark(Benchmark *);
+
+private:
+  virtual void Run() = 0;
+  virtual const char *getName() const = 0;
+
+  static Benchmark *Start;
+  static Benchmark *End;
+};
+
+class WrapperBenchmark : public Benchmark {
+  using BenchmarkWrapperFunction = uint64_t (*)();
+  BenchmarkWrapperFunction Func;
+  const char *Name;
+
+public:
+  WrapperBenchmark(BenchmarkWrapperFunction Func, char const *Name)
+      : Func(Func), Name(Name) {
+    addBenchmark(this);
+  }
+
+private:
+  void Run() override {
+    BenchmarkOptions Options;
+    auto result = benchmark(Options, Func);
+    constexpr auto GREEN = "\033[32m";
----------------
jhuber6 wrote:

Keep in mind that this will need to handle multiple thread sin the future, but I think we can safely restrict this to a single block just to make it easier to manage.

https://github.com/llvm/llvm-project/pull/92009