feat: support CCL Reduce

examples/ccl/reduce.cc

@@ -0,0 +1,307 @@
+/**
+ * InfiniCCL Example: Thread-per-GPU Single-Node Reduce
+ *
+ * This example creates one native CCL rank per GPU and validates a rooted
+ * reduction without an MPI launcher.
+ */
+
+#include <unistd.h>
+
+#include <array>
+#include <atomic>
+#include <charconv>
+#include <cmath>
+#include <cstdlib>
+#include <cstring>
+#include <iomanip>
+#include <iostream>
+#include <limits>
+#include <string>
+#include <system_error>
+#include <thread>
+#include <vector>
+
+// Public API
+#include "infiniccl.h"
+
+// Example-Specific Utilities
+#include "utils.h"
+
+// Internal Headers (Accessible via example-specific include paths, technically
+// not public APIs)
+#include "backend_manifest.h"
+
+using namespace infini::ccl;
+
+namespace {
+
+constexpr int kRoot = 0;
+
+struct RunState {
+  std::atomic<bool> correct{true};
+  std::atomic<int> completed{0};
+};
+
+struct ThreadArgs {
+  int rank;
+  int size;
+  infinicclUniqueId id;
+  size_t count;
+  int warmup_iterations;
+  int profile_iterations;
+  RunState *state;
+};
+
+template <typename T>
+bool ParsePositiveNumber(const char *text, T *value) {
+  if (!text || !value) {
+    return false;
+  }
+
+  T parsed{};
+  const char *end = text + std::strlen(text);
+  const auto result = std::from_chars(text, end, parsed);
+  if (result.ec != std::errc{} || result.ptr != end || parsed <= 0) {
+    return false;
+  }
+
+  *value = parsed;
+  return true;
+}
+
+float ExpectedValue(int world_size) {
+  return static_cast<float>(world_size) *
+         (static_cast<float>(world_size) + 1.0f) / 2.0f;
+}
+
+void PrintReduceMetrics(size_t count, double elapsed_ms) {
+  constexpr double kBytesPerMiB = 1024.0 * 1024.0;
+  constexpr double kBytesPerGB = 1.0e9;
+  const double payload_bytes = static_cast<double>(count) * sizeof(float);
+  const auto original_flags = std::cout.flags();
+  const auto original_precision = std::cout.precision();
+
+  std::cout << "Data size:      " << count << " floats (" << std::fixed
+            << std::setprecision(2) << payload_bytes / kBytesPerMiB << " MiB)"
+            << std::endl;
+  std::cout << "Time:           " << std::setprecision(3) << elapsed_ms << " ms"
+            << std::endl;
+  if (elapsed_ms > 0.0 && std::isfinite(elapsed_ms)) {
+    const double algorithm_bandwidth =
+        payload_bytes / kBytesPerGB / (elapsed_ms / 1000.0);
+    // `nccl-tests` uses a bus-bandwidth correction factor of 1 for Reduce.
+    const double bus_bandwidth = algorithm_bandwidth;
+    std::cout << "Throughput:     " << std::setprecision(2) << bus_bandwidth
+              << " GB/s (Bus BW)" << std::endl;
+    std::cout << "Alg Bandwidth:  " << std::setprecision(2)
+              << algorithm_bandwidth << " GB/s" << std::endl;
+  } else {
+    std::cout << "Throughput:     N/A (Bus BW)" << std::endl;
+    std::cout << "Alg Bandwidth:  N/A" << std::endl;
+  }
+
+  std::cout.flags(original_flags);
+  std::cout.precision(original_precision);
+}
+
+void PrintResult(bool correct, float expected, float actual, size_t count,
+                 double elapsed_ms) {
+  constexpr const char *kGreen = "\033[32m";
+  constexpr const char *kRed = "\033[31m";
+  constexpr const char *kReset = "\033[0m";
+
+  std::cout << "\n=== CCL Reduce Results ===" << std::endl;
+  std::cout << "Correct: "
+            << (correct ? (kGreen + std::string("YES") + kReset)
+                        : (kRed + std::string("NO") + kReset))
+            << std::endl;
+  std::cout << "Root rank: " << kRoot << std::endl;
+  std::cout << "Expect:  " << expected << std::endl;
+  std::cout << "Actual:  " << actual << std::endl;
+  PrintReduceMetrics(count, elapsed_ms);
+}
+
+void WaitForAll(RunState *state, int world_size) {
+  state->completed.fetch_add(1, std::memory_order_acq_rel);
+  while (state->completed.load(std::memory_order_acquire) < world_size) {
+    std::this_thread::yield();
+  }
+}
+
+void WorkerThread(ThreadArgs args) {
+  constexpr Device::Type kDevType =
+      ListGetBest<DevicePriority>(EnabledDevices{});
+  using Rt = Runtime<kDevType>;
+
+  CHECK_RT(Rt, Rt::SetDevice(args.rank));
+
+  std::array<char, 256> hostname{};
+  if (gethostname(hostname.data(), hostname.size()) != 0) {
+    std::cerr << "Failed to query the hostname for the Reduce worker."
+              << std::endl;
+    std::exit(EXIT_FAILURE);
+  }
+  hostname.back() = '\0';
+  std::cout << "[Rank " << args.rank << "] Host: " << hostname.data()
+            << " | GPU: " << Device::StringFromType(kDevType) << " | Device "
+            << args.rank << std::endl;
+
+  infinicclComm_t comm = nullptr;
+  CHECK_INFINI(infinicclCommInitRank(&comm, args.size, args.id, args.rank));
+
+  const bool is_root = args.rank == kRoot;
+  const size_t total_bytes = args.count * sizeof(float);
+  std::vector<float> h_send(args.count, static_cast<float>(args.rank + 1));
+  std::vector<float> h_recv(is_root ? args.count : 0, 0.0f);
+
+  float *d_send = nullptr;
+  float *d_recv = nullptr;
+  CHECK_RT(Rt, Rt::Malloc(reinterpret_cast<void **>(&d_send), total_bytes));
+  if (is_root) {
+    CHECK_RT(Rt, Rt::Malloc(reinterpret_cast<void **>(&d_recv), total_bytes));
+  }
+  CHECK_RT(Rt, Rt::Memcpy(d_send, h_send.data(), total_bytes,
+                          Rt::MemcpyHostToDevice));
+  CHECK_RT(Rt, Rt::StreamSynchronize(nullptr));
+
+  for (int i = 0; i < args.warmup_iterations; ++i) {
+    CHECK_INFINI(infinicclReduce(d_send, d_recv, args.count, infinicclFloat32,
+                                 infinicclSum, kRoot, comm, nullptr));
+  }
+  CHECK_RT(Rt, Rt::StreamSynchronize(nullptr));
+
+  Timer timer;
+  for (int i = 0; i < args.profile_iterations; ++i) {
+    CHECK_INFINI(infinicclReduce(d_send, d_recv, args.count, infinicclFloat32,
+                                 infinicclSum, kRoot, comm, nullptr));
+  }
+  CHECK_RT(Rt, Rt::StreamSynchronize(nullptr));
+  const double elapsed_ms =
+      timer.ElapsedMs() / static_cast<double>(args.profile_iterations);
+
+  bool local_correct = true;
+  const float expected = ExpectedValue(args.size);
+  if (is_root) {
+    CHECK_RT(Rt, Rt::Memcpy(h_recv.data(), d_recv, total_bytes,
+                            Rt::MemcpyDeviceToHost));
+    CHECK_RT(Rt, Rt::StreamSynchronize(nullptr));
+    local_correct = Validator::ValidateResult(
+        h_recv.data(), args.count, expected, args.rank, false, "Reduce");
+    if (!local_correct) {
+      args.state->correct.store(false, std::memory_order_relaxed);
+    }
+  }
+
+  WaitForAll(args.state, args.size);
+  if (is_root) {
+    PrintResult(args.state->correct.load(std::memory_order_acquire), expected,
+                h_recv.front(), args.count, elapsed_ms);
+  }
+
+  CHECK_RT(Rt, Rt::Free(d_send));
+  if (is_root) {
+    CHECK_RT(Rt, Rt::Free(d_recv));
+  }
+  CHECK_INFINI(infinicclCommDestroy(comm));
+}
+
+void PrintUsage(const char *program) {
+  std::cout << "Usage: " << program << " [options]\n"
+            << "Options:\n"
+            << "  -g <num_gpus>      Number of GPUs (default: 8)\n"
+            << "  -w <warmup_iters>  Warm-up iterations (default: 2)\n"
+            << "  -p <profile_iters> Profile iterations (default: 20)\n"
+            << "  -n <count>         Elements reduced per rank (default: "
+               "1048576)\n";
+}
+
+}  // namespace
+
+int main(int argc, char **argv) {
+  int num_gpus = 8;
+  int warmup_iterations = 2;
+  int profile_iterations = 20;
+  size_t count = 1 << 20;
+
+  int opt = 0;
+  while ((opt = getopt(argc, argv, "g:w:p:n:h")) != -1) {
+    bool parsed = false;
+    switch (opt) {
+      case 'g':
+        parsed = ParsePositiveNumber(optarg, &num_gpus);
+        break;
+      case 'w':
+        parsed = ParsePositiveNumber(optarg, &warmup_iterations);
+        break;
+      case 'p':
+        parsed = ParsePositiveNumber(optarg, &profile_iterations);
+        break;
+      case 'n':
+        parsed = ParsePositiveNumber(optarg, &count);
+        break;
+      case 'h':
+        PrintUsage(argv[0]);
+        return EXIT_SUCCESS;
+      default:
+        PrintUsage(argv[0]);
+        return EXIT_FAILURE;
+    }
+
+    if (!parsed) {
+      std::cerr << "Invalid positive numeric option for Reduce." << std::endl;
+      return EXIT_FAILURE;
+    }
+  }
+
+  if (optind != argc) {
+    std::cerr << "Unexpected positional argument for Reduce." << std::endl;
+    return EXIT_FAILURE;
+  }
+  if (count > std::numeric_limits<size_t>::max() / sizeof(float)) {
+    std::cerr << "Reduce buffer size overflows `size_t`." << std::endl;
+    return EXIT_FAILURE;
+  }
+
+  std::array<char, 256> hostname{};
+  if (gethostname(hostname.data(), hostname.size()) != 0) {
+    std::cerr << "Failed to query the hostname for Reduce." << std::endl;
+    return EXIT_FAILURE;
+  }
+  hostname.back() = '\0';
+  std::cout << "[Main Process] Host: " << hostname.data()
+            << " | Target GPUs: " << num_gpus << std::endl;
+  std::cout << "[Main Process] Count: " << count
+            << " floats | Warmup: " << warmup_iterations
+            << " | Profile: " << profile_iterations << " | Root: " << kRoot
+            << std::endl;
+
+  infinicclUniqueId shared_id{};
+  CHECK_INFINI(infinicclGetUniqueId(&shared_id));
+
+  RunState state;
+  std::vector<std::thread> threads;
+  threads.reserve(num_gpus);
+  for (int rank = 0; rank < num_gpus; ++rank) {
+    ThreadArgs args{rank,  num_gpus,          shared_id,
+                    count, warmup_iterations, profile_iterations,
+                    &state};
+    threads.emplace_back(WorkerThread, args);
+  }
+
+  for (auto &thread : threads) {
+    if (thread.joinable()) {
+      thread.join();
+    }
+  }
+
+  const bool correct = state.correct.load(std::memory_order_acquire);
+  if (correct) {
+    std::cout << "[Main Process] CCL Reduce validation passed." << std::endl;
+  } else {
+    std::cerr << "[Main Process] CCL Reduce validation failed." << std::endl;
+  }
+  std::cout
+      << "[Main Process] All worker threads joined. InfiniCCL finalized safely."
+      << std::endl;
+  return correct ? EXIT_SUCCESS : EXIT_FAILURE;
+}