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

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

/**
Testcase Scenarios :
 (TestCase 1)::
 1) Execute atomicAdd in multi threaded scenario by diverging the data across
 multiple threads and validate the output at the end of all operations.
 2) Execute atomicAddNoRet in multi threaded scenario by diverging the data
 across multiple threads and validate the output at the end of all operations.
 (TestCase 2)::
 3) Execute atomicAdd API and validate the result.
 4) Execute atomicAddNoRet API and validate the result.
 (TestCase 3)::
 5) atomicadd/NoRet negative scenarios (TBD).
*/

#include <hip_test_kernels.hh>
#include <hip_test_common.hh>
#include <hip_test_checkers.hh>
/*
 * Defines initial and increment values
 */
#define INCREMENT_VALUE 10
#define INT_INITIAL_VALUE 10
#define FLOAT_INITIAL_VALUE 10.50
#define DOUBLE_INITIAL_VALUE 200.12
#define LONG_INITIAL_VALUE 10000
#define UNSIGNED_INITIAL_VALUE 20

#if HT_NVIDIA
// atomicAddNoRet is unavailable in cuda
template <typename T> __device__ void atomicAddNoRet(T* x, int y) {
  atomicAdd(x, static_cast<T>(y));
}
#endif

bool p_atomicNoRet = false;

template <typename T> __global__ void atomicnoret_manywaves(T* C_d) {
  atomicAddNoRet(C_d, INCREMENT_VALUE);
}

template <typename T> __global__ void atomic_manywaves(T* C_d) { atomicAdd(C_d, INCREMENT_VALUE); }

template <typename T> __global__ void atomicnoret_simple(T* C_d) {
  atomicAddNoRet(C_d, INCREMENT_VALUE);
}

template <typename T> __global__ void atomic_simple(T* C_d) { atomicAdd(C_d, INCREMENT_VALUE); }

template <typename T> bool atomictest_manywaves(const T& initial_val) {
  unsigned int ThreadsperBlock = 10;
  unsigned int numBlocks = 1;
  T memSize = sizeof(T);
  T* hOData = reinterpret_cast<T*>(malloc(memSize));
  *hOData = initial_val;
  T* dOData;
  HIP_CHECK(hipMalloc(&dOData, memSize));
  // copy host memory to device to initialize to zero
  HIP_CHECK(hipMemcpy(dOData, hOData, memSize, hipMemcpyHostToDevice));

  // execute the kernel
  hipLaunchKernelGGL(atomic_manywaves, dim3(numBlocks), dim3(ThreadsperBlock), 0, 0, dOData);

  // Copy result from device to host
  HIP_CHECK(hipMemcpy(hOData, dOData, memSize, hipMemcpyDeviceToHost));
  REQUIRE(hOData[0] ==
          initial_val + static_cast<T>(INCREMENT_VALUE * (ThreadsperBlock * numBlocks)));

  // Cleanup memory
  free(hOData);
  HIP_CHECK(hipFree(dOData));

  return true;
}

template <typename T> bool atomictestnoret_manywaves(const T& initial_val) {
  unsigned int ThreadsperBlock = 10;
  unsigned int numBlocks = 1;
  T memSize = sizeof(T);
  T* hOData = reinterpret_cast<T*>(malloc(memSize));
  *hOData = initial_val;
  T* dOData;
  HIP_CHECK(hipMalloc(&dOData, memSize));
  // copy host memory to device to initialize to zero
  HIP_CHECK(hipMemcpy(dOData, hOData, memSize, hipMemcpyHostToDevice));

  // execute the kernel
  hipLaunchKernelGGL(atomicnoret_manywaves, dim3(numBlocks), dim3(ThreadsperBlock), 0, 0, dOData);

  // Copy result from device to host
  HIP_CHECK(hipMemcpy(hOData, dOData, memSize, hipMemcpyDeviceToHost));
  REQUIRE(hOData[0] == initial_val + (INCREMENT_VALUE * (ThreadsperBlock * numBlocks)));

  // Cleanup memory
  free(hOData);
  HIP_CHECK(hipFree(dOData));

  return true;
}

template <typename T> bool atomictest_simple(const T& initial_val) {
  unsigned int ThreadsperBlock = 1;
  unsigned int numBlocks = 1;
  T memSize = sizeof(T);
  T* hOData = reinterpret_cast<T*>(malloc(memSize));
  *hOData = initial_val;
  T* dOData;
  HIP_CHECK(hipMalloc(&dOData, memSize));
  // copy host memory to device to initialize to zero
  HIP_CHECK(hipMemcpy(dOData, hOData, memSize, hipMemcpyHostToDevice));

  // execute the kernel
  hipLaunchKernelGGL(atomic_simple, dim3(numBlocks), dim3(ThreadsperBlock), 0, 0, dOData);

  // Copy result from device to host
  HIP_CHECK(hipMemcpy(hOData, dOData, memSize, hipMemcpyDeviceToHost));
  REQUIRE(hOData[0] == initial_val + INCREMENT_VALUE);

  // Cleanup memory
  free(hOData);
  HIP_CHECK(hipFree(dOData));

  return true;
}

template <typename T> bool atomictestnoret_simple(const T& initial_val) {
  unsigned int ThreadsperBlock = 1;
  unsigned int numBlocks = 1;
  T memSize = sizeof(T);
  T* hOData = reinterpret_cast<T*>(malloc(memSize));
  *hOData = initial_val;
  T* dOData;
  HIP_CHECK(hipMalloc(&dOData, memSize));
  // copy host memory to device to initialize to zero
  HIP_CHECK(hipMemcpy(dOData, hOData, memSize, hipMemcpyHostToDevice));

  // execute the kernel
  hipLaunchKernelGGL(atomicnoret_simple, dim3(numBlocks), dim3(ThreadsperBlock), 0, 0, dOData);

  // Copy result from device to host
  HIP_CHECK(hipMemcpy(hOData, dOData, memSize, hipMemcpyDeviceToHost));
  REQUIRE(hOData[0] == initial_val + INCREMENT_VALUE);

  // Cleanup memory
  free(hOData);
  HIP_CHECK(hipFree(dOData));

  return true;
}

HIP_TEST_CASE(Unit_hipTestAtomicAdd) {
  bool TestPassed = true;

  SECTION("atomic tests with many waves") {
    REQUIRE(TestPassed == atomictest_manywaves<int>(INT_INITIAL_VALUE));
    REQUIRE(TestPassed == atomictest_manywaves<unsigned int>(UNSIGNED_INITIAL_VALUE));
    REQUIRE(TestPassed == atomictest_manywaves<float>(FLOAT_INITIAL_VALUE));
#if HT_AMD
    REQUIRE(TestPassed == atomictest_manywaves<uint64_t>(LONG_INITIAL_VALUE));
    REQUIRE(TestPassed == atomictest_manywaves<double>(DOUBLE_INITIAL_VALUE));
#endif
  }
  SECTION("atomic tests with many waves and no return") {
    REQUIRE(TestPassed == atomictestnoret_manywaves<float>(FLOAT_INITIAL_VALUE));
  }
  SECTION("simple atomic tests") {
    REQUIRE(TestPassed == atomictest_simple<int>(INT_INITIAL_VALUE));
    REQUIRE(TestPassed == atomictest_simple<unsigned int>(UNSIGNED_INITIAL_VALUE));
    REQUIRE(TestPassed == atomictest_simple<float>(FLOAT_INITIAL_VALUE));
#if HT_AMD
    REQUIRE(TestPassed == atomictest_simple<uint64_t>(LONG_INITIAL_VALUE));
    REQUIRE(TestPassed == atomictest_simple<double>(DOUBLE_INITIAL_VALUE));
#endif
  }
  SECTION("Simple atomic test with no return") {
    REQUIRE(TestPassed == atomictestnoret_simple<float>(FLOAT_INITIAL_VALUE));
  }
}