[libc] NVPTX Profiling (#92009)

PR for adding microbenchmarking infrastructure for NVPTX. `nvlink`
cannot perform LTO, so we cannot inline `libc` functions and this
function call overhead is not adjusted for during microbenchmarking.

Author: jameshu15869

libc/CMakeLists.txt

@@ -401,9 +401,7 @@ if(LLVM_INCLUDE_TESTS)
   add_subdirectory(fuzzing)
 endif()
 
-if(LIBC_INCLUDE_BENCHMARKS)
-  add_subdirectory(benchmarks)
-endif()
+add_subdirectory(benchmarks)
 
 if (LIBC_INCLUDE_DOCS)
   add_subdirectory(docs)

libc/benchmarks/CMakeLists.txt

@@ -1,3 +1,13 @@
+if(LIBC_TARGET_OS_IS_GPU)
+  add_subdirectory(gpu)
+  return()
+endif()
+
+# The CPU build depends on Google benchmark.
+if(NOT LIBC_INCLUDE_BENCHMARKS)
+  return()
+endif()
+
 find_package(Threads)
 
 set(LLVM_LINK_COMPONENTS

libc/benchmarks/gpu/BenchmarkLogger.cpp

@@ -0,0 +1,97 @@
+#include "benchmarks/gpu/BenchmarkLogger.h"
+#include "src/__support/CPP/string.h"
+#include "src/__support/CPP/string_view.h"
+#include "src/__support/OSUtil/io.h"               // write_to_stderr
+#include "src/__support/big_int.h"                 // is_big_int
+#include "src/__support/macros/properties/types.h" // LIBC_TYPES_HAS_INT128
+#include "src/__support/uint128.h"
+
+#include <stdint.h>
+
+namespace LIBC_NAMESPACE {
+namespace benchmarks {
+
+// cpp::string_view specialization
+template <>
+BenchmarkLogger &
+    BenchmarkLogger::operator<< <cpp::string_view>(cpp::string_view str) {
+  LIBC_NAMESPACE::write_to_stderr(str);
+  return *this;
+}
+
+// cpp::string specialization
+template <>
+BenchmarkLogger &BenchmarkLogger::operator<< <cpp::string>(cpp::string str) {
+  return *this << static_cast<cpp::string_view>(str);
+}
+
+// const char* specialization
+template <>
+BenchmarkLogger &BenchmarkLogger::operator<< <const char *>(const char *str) {
+  return *this << cpp::string_view(str);
+}
+
+// char* specialization
+template <> BenchmarkLogger &BenchmarkLogger::operator<< <char *>(char *str) {
+  return *this << cpp::string_view(str);
+}
+
+// char specialization
+template <> BenchmarkLogger &BenchmarkLogger::operator<<(char ch) {
+  return *this << cpp::string_view(&ch, 1);
+}
+
+// bool specialization
+template <> BenchmarkLogger &BenchmarkLogger::operator<<(bool cond) {
+  return *this << (cond ? "true" : "false");
+}
+
+// void * specialization
+template <> BenchmarkLogger &BenchmarkLogger::operator<<(void *addr) {
+  return *this << "0x" << cpp::to_string(reinterpret_cast<uintptr_t>(addr));
+}
+
+template <typename T> BenchmarkLogger &BenchmarkLogger::operator<<(T t) {
+  if constexpr (is_big_int_v<T> ||
+                (cpp::is_integral_v<T> && cpp::is_unsigned_v<T> &&
+                 (sizeof(T) > sizeof(uint64_t)))) {
+    static_assert(sizeof(T) % 8 == 0, "Unsupported size of UInt");
+    const IntegerToString<T, radix::Hex::WithPrefix> buffer(t);
+    return *this << buffer.view();
+  } else {
+    return *this << cpp::to_string(t);
+  }
+}
+
+// is_integral specializations
+// char is already specialized to handle character
+template BenchmarkLogger &BenchmarkLogger::operator<< <short>(short);
+template BenchmarkLogger &BenchmarkLogger::operator<< <int>(int);
+template BenchmarkLogger &BenchmarkLogger::operator<< <long>(long);
+template BenchmarkLogger &BenchmarkLogger::operator<< <long long>(long long);
+template BenchmarkLogger &
+    BenchmarkLogger::operator<< <unsigned char>(unsigned char);
+template BenchmarkLogger &
+    BenchmarkLogger::operator<< <unsigned short>(unsigned short);
+template BenchmarkLogger &
+    BenchmarkLogger::operator<< <unsigned int>(unsigned int);
+template BenchmarkLogger &
+    BenchmarkLogger::operator<< <unsigned long>(unsigned long);
+template BenchmarkLogger &
+    BenchmarkLogger::operator<< <unsigned long long>(unsigned long long);
+
+#ifdef LIBC_TYPES_HAS_INT128
+template BenchmarkLogger &
+    BenchmarkLogger::operator<< <__uint128_t>(__uint128_t);
+#endif // LIBC_TYPES_HAS_INT128
+template BenchmarkLogger &BenchmarkLogger::operator<< <UInt<128>>(UInt<128>);
+template BenchmarkLogger &BenchmarkLogger::operator<< <UInt<192>>(UInt<192>);
+template BenchmarkLogger &BenchmarkLogger::operator<< <UInt<256>>(UInt<256>);
+template BenchmarkLogger &BenchmarkLogger::operator<< <UInt<320>>(UInt<320>);
+
+// TODO: Add floating point formatting once it's supported by StringStream.
+
+BenchmarkLogger log;
+
+} // namespace benchmarks
+} // namespace LIBC_NAMESPACE

libc/benchmarks/gpu/BenchmarkLogger.h

@@ -0,0 +1,27 @@
+//===-- Utilities to log to standard output during tests --------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_LIBC_BENCHMARKS_GPU_BENCHMARKLOGGER_H
+#define LLVM_LIBC_BENCHMARKS_GPU_BENCHMARKLOGGER_H
+
+namespace LIBC_NAMESPACE {
+namespace benchmarks {
+
+// A class to log to standard output in the context of hermetic tests.
+struct BenchmarkLogger {
+  constexpr BenchmarkLogger() = default;
+  template <typename T> BenchmarkLogger &operator<<(T);
+};
+
+// A global TestLogger instance to be used in tests.
+extern BenchmarkLogger log;
+
+} // namespace benchmarks
+} // namespace LIBC_NAMESPACE
+
+#endif /* LLVM_LIBC_BENCHMARKS_GPU_BENCHMARKLOGGER_H */

libc/benchmarks/gpu/CMakeLists.txt

@@ -0,0 +1,56 @@
+add_subdirectory(timing)
+
+add_custom_target(gpu-benchmark)
+
+function(add_benchmark benchmark_name)
+  cmake_parse_arguments(
+    "BENCHMARK"
+    "" # Optional arguments
+    "" # Single value arguments
+    "LINK_LIBRARIES" # Multi-value arguments
+    ${ARGN}
+  )
+  if(NOT libc.src.time.clock IN_LIST TARGET_LLVMLIBC_ENTRYPOINTS)
+    message(FATAL_ERROR "target does not support clock")
+  endif()
+  add_libc_hermetic(
+    ${benchmark_name}
+    IS_BENCHMARK
+    LINK_LIBRARIES
+      LibcGpuBenchmark.hermetic
+      ${BENCHMARK_LINK_LIBRARIES}
+    ${BENCHMARK_UNPARSED_ARGUMENTS}
+  )
+  get_fq_target_name(${benchmark_name} fq_target_name)
+  add_dependencies(gpu-benchmark ${fq_target_name})
+endfunction(add_benchmark)
+
+add_unittest_framework_library(
+  LibcGpuBenchmark
+  SRCS
+    LibcGpuBenchmark.cpp
+    LibcGpuBenchmarkMain.cpp
+    BenchmarkLogger.cpp
+  HDRS
+    LibcGpuBenchmark.h
+    BenchmarkLogger.h
+  DEPENDS
+    libc.src.__support.big_int
+    libc.src.__support.c_string
+    libc.src.__support.CPP.string
+    libc.src.__support.CPP.string_view
+    libc.src.__support.CPP.type_traits
+    libc.src.__support.CPP.functional
+    libc.src.__support.CPP.limits
+    libc.src.__support.CPP.algorithm
+    libc.src.__support.fixed_point.fx_rep
+    libc.src.__support.macros.properties.types
+    libc.src.__support.OSUtil.osutil
+    libc.src.__support.uint128
+    libc.src.__support.FPUtil.sqrt
+    libc.src.__support.fixedvector
+    libc.src.time.clock
+    libc.benchmarks.gpu.timing.timing
+)
+
+add_subdirectory(src)

libc/benchmarks/gpu/LibcGpuBenchmark.cpp

@@ -0,0 +1,140 @@
+#include "LibcGpuBenchmark.h"
+#include "src/__support/CPP/algorithm.h"
+#include "src/__support/CPP/array.h"
+#include "src/__support/CPP/string.h"
+#include "src/__support/FPUtil/sqrt.h"
+#include "src/__support/GPU/utils.h"
+#include "src/__support/fixedvector.h"
+#include "src/time/gpu/time_utils.h"
+
+namespace LIBC_NAMESPACE {
+namespace benchmarks {
+
+FixedVector<Benchmark *, 64> benchmarks;
+cpp::array<BenchmarkResult, 1024> results;
+
+void Benchmark::add_benchmark(Benchmark *benchmark) {
+  benchmarks.push_back(benchmark);
+}
+
+BenchmarkResult reduce_results(cpp::array<BenchmarkResult, 1024> &results) {
+  BenchmarkResult result;
+  uint64_t cycles_sum = 0;
+  double standard_deviation_sum = 0;
+  uint64_t min = UINT64_MAX;
+  uint64_t max = 0;
+  uint32_t samples_sum = 0;
+  uint32_t iterations_sum = 0;
+  clock_t time_sum = 0;
+  uint64_t num_threads = gpu::get_num_threads();
+  for (uint64_t i = 0; i < num_threads; i++) {
+    BenchmarkResult current_result = results[i];
+    cycles_sum += current_result.cycles;
+    standard_deviation_sum += current_result.standard_deviation;
+    min = cpp::min(min, current_result.min);
+    max = cpp::max(max, current_result.max);
+    samples_sum += current_result.samples;
+    iterations_sum += current_result.total_iterations;
+    time_sum += current_result.total_time;
+  }
+  result.cycles = cycles_sum / num_threads;
+  result.standard_deviation = standard_deviation_sum / num_threads;
+  result.min = min;
+  result.max = max;
+  result.samples = samples_sum / num_threads;
+  result.total_iterations = iterations_sum / num_threads;
+  result.total_time = time_sum / num_threads;
+  return result;
+}
+
+void Benchmark::run_benchmarks() {
+  uint64_t id = gpu::get_thread_id();
+  gpu::sync_threads();
+
+  for (Benchmark *benchmark : benchmarks)
+    results[id] = benchmark->run();
+  gpu::sync_threads();
+  if (id == 0) {
+    for (Benchmark *benchmark : benchmarks) {
+      BenchmarkResult all_results = reduce_results(results);
+      constexpr auto GREEN = "\033[32m";
+      constexpr auto RESET = "\033[0m";
+      log << GREEN << "[ RUN      ] " << RESET << benchmark->get_name() << '\n';
+      log << GREEN << "[       OK ] " << RESET << benchmark->get_name() << ": "
+          << all_results.cycles << " cycles, " << all_results.min << " min, "
+          << all_results.max << " max, " << all_results.total_iterations
+          << " iterations, " << all_results.total_time << " ns, "
+          << static_cast<long>(all_results.standard_deviation) << " stddev\n";
+    }
+  }
+  gpu::sync_threads();
+}
+
+BenchmarkResult benchmark(const BenchmarkOptions &options,
+                          cpp::function<uint64_t(void)> wrapper_func) {
+  BenchmarkResult result;
+  RuntimeEstimationProgression rep;
+  uint32_t total_iterations = 0;
+  uint32_t iterations = options.initial_iterations;
+  if (iterations < 1u)
+    iterations = 1;
+
+  uint32_t samples = 0;
+  uint64_t total_time = 0;
+  uint64_t best_guess = 0;
+  uint64_t total_cycles = 0;
+  uint64_t cycles_squared = 0;
+  uint64_t min = UINT64_MAX;
+  uint64_t max = 0;
+
+  uint64_t overhead = UINT64_MAX;
+  int overhead_iterations = 10;
+  for (int i = 0; i < overhead_iterations; i++)
+    overhead = cpp::min(overhead, LIBC_NAMESPACE::overhead());
+
+  for (uint64_t time_budget = options.max_duration; time_budget >= 0;) {
+    uint64_t sample_cycles = 0;
+    const clock_t start = static_cast<double>(clock());
+    for (uint32_t i = 0; i < iterations; i++) {
+      auto wrapper_intermediate = wrapper_func();
+      uint64_t result = wrapper_intermediate - overhead;
+      max = cpp::max(max, result);
+      min = cpp::min(min, result);
+      sample_cycles += result;
+    }
+    const clock_t end = clock();
+    const clock_t duration_ns =
+        ((end - start) * 1000 * 1000 * 1000) / CLOCKS_PER_SEC;
+    total_time += duration_ns;
+    time_budget -= duration_ns;
+    samples++;
+    total_cycles += sample_cycles;
+    cycles_squared += sample_cycles * sample_cycles;
+
+    total_iterations += iterations;
+    const double change_ratio =
+        rep.compute_improvement({iterations, sample_cycles});
+    best_guess = rep.current_estimation;
+
+    if (samples >= options.max_samples || iterations >= options.max_iterations)
+      break;
+    if (total_time >= options.min_duration && samples >= options.min_samples &&
+        change_ratio < options.epsilon)
+      break;
+
+    iterations *= options.scaling_factor;
+  }
+  result.cycles = best_guess;
+  result.standard_deviation = fputil::sqrt<double>(
+      static_cast<double>(cycles_squared) / total_iterations -
+      static_cast<double>(best_guess * best_guess));
+  result.min = min;
+  result.max = max;
+  result.samples = samples;
+  result.total_iterations = total_iterations;
+  result.total_time = total_time;
+  return result;
+};
+
+} // namespace benchmarks
+} // namespace LIBC_NAMESPACE