streams.cu

#include <cuda_runtime.h>

#include <cstdio>
#include <cstdlib>
#include <unordered_set>
#include <vector>
#include <cstdint>

static void check(cudaError_t e, const char* msg) {
  if (e != cudaSuccess) {
    std::fprintf(stderr, "CUDA error (%s): %s\n", msg, cudaGetErrorString(e));
    std::exit(1);
  }
}

int main() {
  const int N = 10000;

  std::vector<cudaStream_t> streams(N);
  for (int i = 0; i < N; ++i) {
    check(cudaStreamCreateWithFlags(&streams[i], cudaStreamNonBlocking),
          "cudaStreamCreateWithFlags(cudaStreamNonBlocking)");
  }

  std::unordered_set<std::uintptr_t> seen;
  bool any_dup = false;

  for (int i = 0; i < N; ++i) {
    // cudaStream_t is an opaque handle; usually pointer-like.
    std::uintptr_t h = reinterpret_cast<std::uintptr_t>(streams[i]);
    std::printf("Stream %d: handle = 0x%zx\n", i, (size_t)h);

    if (!seen.insert(h).second) {
      any_dup = true;
      std::printf("  >>> DUPLICATE handle detected: 0x%zx (at stream %d)\n",
                  (size_t)h, i);
    }
  }

  std::printf("\nSummary:\n");
  std::printf("Total streams created: %d\n", N);
  std::printf("Unique handles:        %zu\n", seen.size());
  std::printf("%s\n", any_dup ? "⚠️ Duplicates detected" : "✅ No duplicates detected");

  for (int i = 0; i < N; ++i) {
    check(cudaStreamDestroy(streams[i]), "cudaStreamDestroy");
  }

  return 0;
}