TheRock/rocm-systems/projects/hip-tests/catch/unit/deviceLib/threadfence_system.cc

/*
 * Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
 *
 * SPDX-License-Identifier: MIT
 */

#include <hip_test_common.hh>
#include <hip/hip_runtime.h>

#include <atomic>
#include <thread>

__host__ __device__ void fence_system() {
#ifdef __HIP_DEVICE_COMPILE__
  __threadfence_system();
#else
  std::atomic_thread_fence(std::memory_order_seq_cst);
#endif
}

__host__ __device__ void round_robin(const int id, const int num_dev, const int num_iter,
                                     volatile int* data, volatile int* flag) {
  for (int i = 0; i < num_iter; i++) {
    while (*flag % num_dev != id) fence_system();  // invalid the cache for read

    (*data)++;
    fence_system();  // make sure the store to data is sequenced before the store to flag
    (*flag)++;
    fence_system();  // invalid the cache to flush out flag
  }
}

__global__ void gpu_round_robin(const int id, const int num_dev, const int num_iter,
                                volatile int* data, volatile int* flag) {
  round_robin(id, num_dev, num_iter, data, flag);
}

HIP_TEST_CASE(Unit_threadfence_system) {
  int num_gpus = 0;
  HIP_CHECK(hipGetDeviceCount(&num_gpus));
  REQUIRE(num_gpus > 0);

  volatile int* data = nullptr;
  if (hipHostMalloc(&data, sizeof(int), hipHostMallocCoherent) != hipSuccess) {
    HIP_SKIP_TEST(HipTest::SkipReason::kCoherentHostAllocFailed);
  }

  constexpr int init_data = 1000;
  *data = init_data;

  volatile int* flag = nullptr;
  if (hipHostMalloc(&flag, sizeof(int), hipHostMallocCoherent) != hipSuccess) {
    HIP_CHECK(hipHostFree((void*)data));
    HIP_SKIP_TEST(HipTest::SkipReason::kCoherentHostAllocFailed);
  }
  *flag = 0;

  // number of rounds per device
  constexpr int num_iter = 1000;

  // one CPU thread + 1 kernel/GPU
  const int num_dev = num_gpus + 1;

  int next_id = 0;
  std::vector<std::thread> threads;

  // create a CPU thread for the round_robin
  threads.push_back(std::thread(round_robin, next_id++, num_dev, num_iter, data, flag));

  // run one thread per GPU
  dim3 dim_block(1, 1, 1);
  dim3 dim_grid(1, 1, 1);

  // launch one kernel per device for the round robin
  for (; next_id < num_dev; ++next_id) {
    threads.push_back(std::thread([=]() {
      HIP_CHECK(hipSetDevice(next_id - 1));
      hipLaunchKernelGGL(gpu_round_robin, dim_grid, dim_block, 0, 0x0, next_id, num_dev, num_iter,
                         data, flag);
      HIP_CHECK(hipGetLastError());
      HIP_CHECK(hipDeviceSynchronize());
    }));
  }

  for (auto& t : threads) {
    t.join();
  }

  int expected_data = init_data + num_dev * num_iter;
  int expected_flag = num_dev * num_iter;

  bool passed = *data == expected_data && *flag == expected_flag;

  HIP_CHECK(hipHostFree((void*)data));
  HIP_CHECK(hipHostFree((void*)flag));

  REQUIRE(passed == true);
}