[libc-commits] [libc] [libc] Use Atomics in GPU Benchmarks (PR #98842)

[Joseph Huber via libc-commits]

================
@@ -12,60 +13,113 @@ namespace LIBC_NAMESPACE_DECL {
 namespace benchmarks {
 
 FixedVector<Benchmark *, 64> benchmarks;
-cpp::array<BenchmarkResult, 1024> results;
 
 void Benchmark::add_benchmark(Benchmark *benchmark) {
   benchmarks.push_back(benchmark);
 }
 
-BenchmarkResult
-reduce_results(const 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;
+struct AtomicBenchmarkSums {
+  cpp::Atomic<uint64_t> cycles_sum = 0;
+  cpp::Atomic<uint64_t> standard_deviation_sum = 0;
+  cpp::Atomic<uint64_t> min = UINT64_MAX;
+  cpp::Atomic<uint64_t> max = 0;
+  cpp::Atomic<uint32_t> samples_sum = 0;
+  cpp::Atomic<uint32_t> iterations_sum = 0;
+  cpp::Atomic<clock_t> time_sum = 0;
+  cpp::Atomic<uint64_t> active_threads = 0;
+
+  void reset() {
+    active_threads.store(0, cpp::MemoryOrder::RELAXED);
+    cycles_sum.store(0, cpp::MemoryOrder::RELAXED);
+    standard_deviation_sum.store(0, cpp::MemoryOrder::RELAXED);
+    min.store(UINT64_MAX, cpp::MemoryOrder::RELAXED);
+    max.store(0, cpp::MemoryOrder::RELAXED);
+    samples_sum.store(0, cpp::MemoryOrder::RELAXED);
+    iterations_sum.store(0, cpp::MemoryOrder::RELAXED);
+    time_sum.store(0, cpp::MemoryOrder::RELAXED);
   }
-  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 update(const BenchmarkResult &result) {
+    gpu::memory_fence();
+    active_threads.fetch_add(1, cpp::MemoryOrder::RELAXED);
+
+    cycles_sum.fetch_add(result.cycles, cpp::MemoryOrder::RELAXED);
+    standard_deviation_sum.fetch_add(
+        static_cast<uint64_t>(result.standard_deviation),
+        cpp::MemoryOrder::RELAXED);
+
+    // Perform a CAS loop to atomically update the min
+    uint64_t orig_min = min.load(cpp::MemoryOrder::RELAXED);
+    while (!min.compare_exchange_strong(
+        orig_min, cpp::min(orig_min, result.min), cpp::MemoryOrder::ACQUIRE,
+        cpp::MemoryOrder::RELAXED)) {
+    }
+
+    // Perform a CAS loop to atomically update the max
+    uint64_t orig_max = max.load(cpp::MemoryOrder::RELAXED);
+    while (!max.compare_exchange_strong(
+        orig_max, cpp::max(orig_max, result.max), cpp::MemoryOrder::ACQUIRE,
+        cpp::MemoryOrder::RELAXED)) {
+    }
+
+    samples_sum.fetch_add(result.samples, cpp::MemoryOrder::RELAXED);
+    iterations_sum.fetch_add(result.total_iterations,
+                             cpp::MemoryOrder::RELAXED);
+    time_sum.fetch_add(result.total_time, cpp::MemoryOrder::RELAXED);
+    gpu::memory_fence();
----------------
jhuber6 wrote:

```suggestion
    cpp::atomic_thread_fence(cpp::MemoryOrder::RELEASE);
```

https://github.com/llvm/llvm-project/pull/98842