[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.

17 files changed, 644 insertions, 11 deletions

diff --git a/libc/CMakeLists.txt b/libc/CMakeLists.txt
index f35471a..4ffcd55 100644
--- a/libc/CMakeLists.txt
+++ b/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)
diff --git a/libc/benchmarks/CMakeLists.txt b/libc/benchmarks/CMakeLists.txt
index 4978da6..0cff6eb 100644
--- a/libc/benchmarks/CMakeLists.txt
+++ b/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
diff --git a/libc/benchmarks/gpu/BenchmarkLogger.cpp b/libc/benchmarks/gpu/BenchmarkLogger.cpp
new file mode 100644
index 0000000..2e7e8e7
--- /dev/null
+++ b/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
diff --git a/libc/benchmarks/gpu/BenchmarkLogger.h b/libc/benchmarks/gpu/BenchmarkLogger.h
new file mode 100644
index 0000000..332ff14
--- /dev/null
+++ b/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 */
diff --git a/libc/benchmarks/gpu/CMakeLists.txt b/libc/benchmarks/gpu/CMakeLists.txt
new file mode 100644
index 0000000..d167abc
--- /dev/null
+++ b/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)
diff --git a/libc/benchmarks/gpu/LibcGpuBenchmark.cpp b/libc/benchmarks/gpu/LibcGpuBenchmark.cpp
new file mode 100644
index 0000000..69adb0c
--- /dev/null
+++ b/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
diff --git a/libc/benchmarks/gpu/LibcGpuBenchmark.h b/libc/benchmarks/gpu/LibcGpuBenchmark.h
new file mode 100644
index 0000000..59dd589
--- /dev/null
+++ b/libc/benchmarks/gpu/LibcGpuBenchmark.h
@@ -0,0 +1,108 @@
+#ifndef LLVM_LIBC_BENCHMARKS_LIBC_GPU_BENCHMARK_H
+#define LLVM_LIBC_BENCHMARKS_LIBC_GPU_BENCHMARK_H
+
+#include "benchmarks/gpu/BenchmarkLogger.h"
+#include "benchmarks/gpu/timing/timing.h"
+#include "src/__support/CPP/functional.h"
+#include "src/__support/CPP/limits.h"
+#include "src/__support/CPP/string_view.h"
+#include "src/time/clock.h"
+
+#include <stdint.h>
+
+namespace LIBC_NAMESPACE {
+
+namespace benchmarks {
+
+struct BenchmarkOptions {
+  uint32_t initial_iterations = 1;
+  uint32_t max_iterations = 10000000;
+  uint32_t min_samples = 4;
+  uint32_t max_samples = 1000;
+  uint64_t min_duration = 0;                  // in nanoseconds (ns)
+  uint64_t max_duration = 1000 * 1000 * 1000; // 1e9 nanoseconds = 1 second
+  double epsilon = 0.01;
+  double scaling_factor = 1.4;
+};
+
+struct Measurement {
+  uint32_t iterations = 0;
+  uint64_t elapsed_cycles = 0;
+};
+
+class RefinableRuntimeEstimation {
+  uint64_t total_cycles = 0;
+  uint32_t total_iterations = 0;
+
+public:
+  uint64_t update(const Measurement &M) {
+    total_cycles += M.elapsed_cycles;
+    total_iterations += M.iterations;
+    return total_cycles / total_iterations;
+  }
+};
+
+// Tracks the progression of the runtime estimation
+class RuntimeEstimationProgression {
+  RefinableRuntimeEstimation rre;
+
+public:
+  uint64_t current_estimation = 0;
+
+  double compute_improvement(const Measurement &M) {
+    const uint64_t new_estimation = rre.update(M);
+    double ratio =
+        (static_cast<double>(current_estimation) / new_estimation) - 1.0;
+
+    // Get absolute value
+    if (ratio < 0)
+      ratio *= -1;
+
+    current_estimation = new_estimation;
+    return ratio;
+  }
+};
+
+struct BenchmarkResult {
+  uint64_t cycles = 0;
+  double standard_deviation = 0;
+  uint64_t min = UINT64_MAX;
+  uint64_t max = 0;
+  uint32_t samples = 0;
+  uint32_t total_iterations = 0;
+  clock_t total_time = 0;
+};
+
+BenchmarkResult benchmark(const BenchmarkOptions &options,
+                          cpp::function<uint64_t(void)> wrapper_func);
+
+class Benchmark {
+  const cpp::function<uint64_t(void)> func;
+  const cpp::string_view name;
+
+public:
+  Benchmark(cpp::function<uint64_t(void)> func, char const *name)
+      : func(func), name(name) {
+    add_benchmark(this);
+  }
+
+  static void run_benchmarks();
+
+protected:
+  static void add_benchmark(Benchmark *benchmark);
+
+private:
+  BenchmarkResult run() {
+    BenchmarkOptions options;
+    return benchmark(options, func);
+  }
+  const cpp::string_view get_name() const { return name; }
+};
+} // namespace benchmarks
+} // namespace LIBC_NAMESPACE
+
+#define BENCHMARK(SuiteName, TestName, Func)                                   \
+  LIBC_NAMESPACE::benchmarks::Benchmark SuiteName##_##TestName##_Instance(     \
+      Func, #SuiteName "." #TestName);
+
+#endif
diff --git a/libc/benchmarks/gpu/LibcGpuBenchmarkMain.cpp b/libc/benchmarks/gpu/LibcGpuBenchmarkMain.cpp
new file mode 100644
index 0000000..97366e5
--- /dev/null
+++ b/libc/benchmarks/gpu/LibcGpuBenchmarkMain.cpp
@@ -0,0 +1,6 @@
+#include "LibcGpuBenchmark.h"
+
+extern "C" int main(int argc, char **argv, char **envp) {
+  LIBC_NAMESPACE::benchmarks::Benchmark::run_benchmarks();
+  return 0;
+}
diff --git a/libc/benchmarks/gpu/src/CMakeLists.txt b/libc/benchmarks/gpu/src/CMakeLists.txt
new file mode 100644
index 0000000..42eb4f7
--- /dev/null
+++ b/libc/benchmarks/gpu/src/CMakeLists.txt
@@ -0,0 +1 @@
+add_subdirectory(ctype)
diff --git a/libc/benchmarks/gpu/src/ctype/CMakeLists.txt b/libc/benchmarks/gpu/src/ctype/CMakeLists.txt
new file mode 100644
index 0000000..79f0142
--- /dev/null
+++ b/libc/benchmarks/gpu/src/ctype/CMakeLists.txt
@@ -0,0 +1,21 @@
+add_custom_target(libc-gpu-ctype-benchmarks)
+
+add_benchmark(
+  isalnum_benchmark
+  SUITE
+    libc-gpu-ctype-benchmarks
+  SRCS
+    isalnum_benchmark.cpp
+  DEPENDS
+    libc.src.ctype.isalnum
+)
+
+add_benchmark(
+  isalpha_benchmark
+  SUITE
+    libc-gpu-ctype-benchmarks
+  SRCS
+    isalpha_benchmark.cpp
+  DEPENDS
+    libc.src.ctype.isalpha
+)
diff --git a/libc/benchmarks/gpu/src/ctype/isalnum_benchmark.cpp b/libc/benchmarks/gpu/src/ctype/isalnum_benchmark.cpp
new file mode 100644
index 0000000..4050bc0
--- /dev/null
+++ b/libc/benchmarks/gpu/src/ctype/isalnum_benchmark.cpp
@@ -0,0 +1,9 @@
+#include "benchmarks/gpu/LibcGpuBenchmark.h"
+
+#include "src/ctype/isalnum.h"
+
+uint64_t BM_IsAlnum() {
+  char x = 'c';
+  return LIBC_NAMESPACE::latency(LIBC_NAMESPACE::isalnum, x);
+}
+BENCHMARK(LlvmLibcIsAlNumGpuBenchmark, IsAlnumWrapper, BM_IsAlnum);
diff --git a/libc/benchmarks/gpu/src/ctype/isalpha_benchmark.cpp b/libc/benchmarks/gpu/src/ctype/isalpha_benchmark.cpp
new file mode 100644
index 0000000..2038eb8
--- /dev/null
+++ b/libc/benchmarks/gpu/src/ctype/isalpha_benchmark.cpp
@@ -0,0 +1,9 @@
+#include "benchmarks/gpu/LibcGpuBenchmark.h"
+
+#include "src/ctype/isalpha.h"
+
+uint64_t BM_IsAlpha() {
+  char x = 'c';
+  return LIBC_NAMESPACE::latency(LIBC_NAMESPACE::isalpha, x);
+}
+BENCHMARK(LlvmLibcIsAlphaGpuBenchmark, IsAlpha, BM_IsAlpha);
diff --git a/libc/benchmarks/gpu/timing/CMakeLists.txt b/libc/benchmarks/gpu/timing/CMakeLists.txt
new file mode 100644
index 0000000..8bbc7e3
--- /dev/null
+++ b/libc/benchmarks/gpu/timing/CMakeLists.txt
@@ -0,0 +1,12 @@
+foreach(target nvptx)
+  add_subdirectory(${target})
+  list(APPEND target_gpu_timing libc.benchmarks.gpu.timing.${target}.${target}_timing)
+endforeach()
+
+add_header_library(
+  timing
+  HDRS
+    timing.h
+  DEPENDS
+    ${target_gpu_timing}
+)
diff --git a/libc/benchmarks/gpu/timing/nvptx/CMakeLists.txt b/libc/benchmarks/gpu/timing/nvptx/CMakeLists.txt
new file mode 100644
index 0000000..9958e16
--- /dev/null
+++ b/libc/benchmarks/gpu/timing/nvptx/CMakeLists.txt
@@ -0,0 +1,7 @@
+add_header_library(
+  nvptx_timing
+  HDRS
+    timing.h
+  DEPENDS
+    libc.src.__support.common
+)
diff --git a/libc/benchmarks/gpu/timing/nvptx/timing.h b/libc/benchmarks/gpu/timing/nvptx/timing.h
new file mode 100644
index 0000000..d3851a7
--- /dev/null
+++ b/libc/benchmarks/gpu/timing/nvptx/timing.h
@@ -0,0 +1,99 @@
+//===------------- NVPTX implementation of timing utils ---------*- 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_UTILS_GPU_TIMING_NVPTX
+#define LLVM_LIBC_UTILS_GPU_TIMING_NVPTX
+
+#include "src/__support/GPU/utils.h"
+#include "src/__support/common.h"
+#include "src/__support/macros/attributes.h"
+#include "src/__support/macros/config.h"
+
+#include <stdint.h>
+
+namespace LIBC_NAMESPACE {
+
+// Returns the overhead associated with calling the profiling region. This
+// allows us to substract the constant-time overhead from the latency to
+// obtain a true result. This can vary with system load.
+[[gnu::noinline]] static uint64_t overhead() {
+  volatile uint32_t x = 1;
+  uint32_t y = x;
+  uint64_t start = gpu::processor_clock();
+  asm("" ::"r"(y), "llr"(start));
+  uint32_t result = y;
+  asm("or.b32 %[v_reg], %[v_reg], 0;" ::[v_reg] "r"(result) :);
+  uint64_t stop = gpu::processor_clock();
+  volatile auto storage = result;
+  return stop - start;
+}
+
+// Stimulate a simple function and obtain its latency in clock cycles on the
+// system. This function cannot be inlined or else it will disturb the very
+// delicate balance of hard-coded dependencies.
+template <typename F, typename T>
+[[gnu::noinline]] static LIBC_INLINE uint64_t latency(F f, T t) {
+  // We need to store the input somewhere to guarantee that the compiler will
+  // not constant propagate it and remove the profiling region.
+  volatile T storage = t;
+  T arg = storage;
+  asm("" ::"r"(arg));
+
+  // Get the current timestamp from the clock.
+  gpu::memory_fence();
+  uint64_t start = gpu::processor_clock();
+
+  // This forces the compiler to load the input argument and run the clock cycle
+  // counter before the profiling region.
+  asm("" ::"r"(arg), "llr"(start));
+
+  // Run the function under test and return its value.
+  auto result = f(arg);
+
+  // This inline assembly performs a no-op which forces the result to both be
+  // used and prevents us from exiting this region before it's complete.
+  asm("or.b32 %[v_reg], %[v_reg], 0;" ::[v_reg] "r"(result) :);
+
+  // Obtain the current timestamp after running the calculation and force
+  // ordering.
+  uint64_t stop = gpu::processor_clock();
+  gpu::memory_fence();
+  asm("" ::"r"(stop));
+  volatile T output = result;
+
+  // Return the time elapsed.
+  return stop - start;
+}
+
+template <typename F, typename T1, typename T2>
+static LIBC_INLINE uint64_t latency(F f, T1 t1, T2 t2) {
+  volatile T1 storage = t1;
+  volatile T2 storage2 = t2;
+  T1 arg = storage;
+  T2 arg2 = storage2;
+  asm("" ::"r"(arg), "r"(arg2));
+
+  gpu::memory_fence();
+  uint64_t start = gpu::processor_clock();
+
+  asm("" ::"r"(arg), "r"(arg2), "llr"(start));
+
+  auto result = f(arg, arg2);
+
+  asm("or.b32 %[v_reg], %[v_reg], 0;" ::[v_reg] "r"(result) :);
+
+  uint64_t stop = gpu::processor_clock();
+  gpu::memory_fence();
+  asm("" ::"r"(stop));
+  volatile auto output = result;
+
+  return stop - start;
+}
+} // namespace LIBC_NAMESPACE
+
+#endif // LLVM_LIBC_UTILS_GPU_TIMING_NVPTX
diff --git a/libc/benchmarks/gpu/timing/timing.h b/libc/benchmarks/gpu/timing/timing.h
new file mode 100644
index 0000000..180ea77
--- /dev/null
+++ b/libc/benchmarks/gpu/timing/timing.h
@@ -0,0 +1,22 @@
+//===------------- Implementation of GPU timing utils -----------*- 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_UTILS_GPU_TIMING_H
+#define LLVM_LIBC_UTILS_GPU_TIMING_H
+
+#include "src/__support/macros/properties/architectures.h"
+
+#if defined(LIBC_TARGET_ARCH_IS_AMDGPU)
+#error "amdgpu not yet supported"
+#elif defined(LIBC_TARGET_ARCH_IS_NVPTX)
+#include "nvptx/timing.h"
+#else
+#error "unsupported platform"
+#endif
+
+#endif // LLVM_LIBC_UTILS_GPU_TIMING_H
diff --git a/libc/cmake/modules/LLVMLibCTestRules.cmake b/libc/cmake/modules/LLVMLibCTestRules.cmake
index c8d7c8a..fbeec32 100644
--- a/libc/cmake/modules/LLVMLibCTestRules.cmake
+++ b/libc/cmake/modules/LLVMLibCTestRules.cmake
@@ -526,12 +526,15 @@ function(add_integration_test test_name)
   add_dependencies(${INTEGRATION_TEST_SUITE} ${fq_target_name})
 endfunction(add_integration_test)
 
-# Rule to add a hermetic test. A hermetic test is one whose executable is fully
+# Rule to add a hermetic program. A hermetic program is one whose executable is fully
 # statically linked and consists of pieces drawn only from LLVM's libc. Nothing,
 # including the startup objects, come from the system libc.
 #
+# For the GPU, these can be either tests or benchmarks, depending on the value
+# of the LINK_LIBRARIES arg.
+#
 # Usage:
-#   add_libc_hermetic_test(
+#   add_libc_hermetic(
 #     <target name>
 #     SUITE <the suite to which the test should belong>
 #     SRCS <src1.cpp> [src2.cpp ...]
@@ -543,14 +546,14 @@ endfunction(add_integration_test)
 #     LINK_LIBRARIES <list of linking libraries for this target>
 #     LOADER_ARGS <list of special args to loaders (like the GPU loader)>
 #   )
-function(add_libc_hermetic_test test_name)
+function(add_libc_hermetic test_name)
   if(NOT TARGET libc.startup.${LIBC_TARGET_OS}.crt1)
     message(VERBOSE "Skipping ${fq_target_name} as it is not available on ${LIBC_TARGET_OS}.")
     return()
   endif()
   cmake_parse_arguments(
     "HERMETIC_TEST"
-    "" # No optional arguments
+    "IS_BENCHMARK" # Optional arguments
     "SUITE" # Single value arguments
     "SRCS;HDRS;DEPENDS;ARGS;ENV;COMPILE_OPTIONS;LINK_LIBRARIES;LOADER_ARGS" # Multi-value arguments
     ${ARGN}
@@ -678,7 +681,6 @@ function(add_libc_hermetic_test test_name)
     PRIVATE
       libc.startup.${LIBC_TARGET_OS}.crt1${internal_suffix}
       ${link_libraries}
-      LibcTest.hermetic
       LibcHermeticTestSupport.hermetic
       # The NVIDIA 'nvlink' linker does not currently support static libraries.
       $<$<NOT:$<BOOL:${LIBC_TARGET_ARCHITECTURE_IS_NVPTX}>>:${fq_target_name}.__libc__>)
@@ -714,8 +716,12 @@ function(add_libc_hermetic_test test_name)
   )
 
   add_dependencies(${HERMETIC_TEST_SUITE} ${fq_target_name})
-  add_dependencies(libc-hermetic-tests ${fq_target_name})
-endfunction(add_libc_hermetic_test)
+  if(NOT ${HERMETIC_TEST_IS_BENCHMARK})
+    # If it is a benchmark, it will already have been added to the
+    # gpu-benchmark target
+    add_dependencies(libc-hermetic-tests ${fq_target_name})
+  endif()
+endfunction(add_libc_hermetic)
 
 # A convenience function to add both a unit test as well as a hermetic test.
 function(add_libc_test test_name)
@@ -730,7 +736,12 @@ function(add_libc_test test_name)
     add_libc_unittest(${test_name}.__unit__ ${LIBC_TEST_UNPARSED_ARGUMENTS})
   endif()
   if(LIBC_ENABLE_HERMETIC_TESTS AND NOT LIBC_TEST_UNIT_TEST_ONLY)
-    add_libc_hermetic_test(${test_name}.__hermetic__ ${LIBC_TEST_UNPARSED_ARGUMENTS})
+    add_libc_hermetic(
+      ${test_name}.__hermetic__
+      LINK_LIBRARIES
+        LibcTest.hermetic
+      ${LIBC_TEST_UNPARSED_ARGUMENTS}
+    )
     get_fq_target_name(${test_name} fq_test_name)
     if(TARGET ${fq_test_name}.__hermetic__ AND TARGET ${fq_test_name}.__unit__)
       # Tests like the file tests perform file operations on disk file. If we