TheRock/rocm-systems/projects/hip-tests/catch/unit/memory/hipDrvMemcpy3DAsync_old.cc

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

/**
 * @addtogroup hipDrvMemcpy3DAsync hipDrvMemcpy3DAsync
 * @{
 * @ingroup MemoryTest
 * `hipMemcpy3DAsync(const hipMemcpy3DParms* p, hipStream_t stream)` -
 * Copies data between 3D objects.
 */

/*
 * Test Scenarios
 * 1. Verifying hipDrvMemcpy3DAsync API for H2A,A2A,A2H scenarios
 * 2. Verifying hipDrvMemcpy3DAsync API for H2D,D2D,D2H scenarios
 * 3. Verifying Negative Scenarios
 * 4. Verifying Extent validation scenarios by passing 0
 * 5. Verifying hipDrvMemcpy3DAsync API by allocating Memory in
 *    one GPU and trigger hipDrvMemcpy3DAsync from peer GPU for
 *    H2D,D2D,D2H scenarios
 * 6. Verifying hipDrvMemcpy3DAsync API by allocating Memory in
 *    one GPU and trigger hipDrvMemcpy3DAsync from peer GPU for
 *    H2A,A2A,A2H scenarios
 *
 *    Scenarios 3 is temporarily excluded in AMD platform
 *    Scenario 5&6 are excluded in CUDA platform
 */

#include <hip_test_common.hh>
#include <hip_test_checkers.hh>

template <typename T> class DrvMemcpy3DAsync {
  int width, height, depth;
  unsigned int size;
  hipArray_Format formatKind;
  hipArray_t arr, arr1;
  hipStream_t stream;
  size_t pitch_D, pitch_E;
  HIP_MEMCPY3D myparms;
  hipDeviceptr_t D_m, E_m;
  T* hData{nullptr};

 public:
  DrvMemcpy3DAsync(int l_width, int l_height, int l_depth, hipArray_Format l_format);
  DrvMemcpy3DAsync() = delete;
  void AllocateMemory();
  void SetDefaultData();
  void HostArray_DrvMemcpy3DAsync(bool device_context_change = false);
  void HostDevice_DrvMemcpy3DAsync(bool device_context_change = false);
  void Extent_Validation();
  void NegativeTests();
  void DeAllocateMemory();
};

/* Intializes class variables */
template <typename T> DrvMemcpy3DAsync<T>::DrvMemcpy3DAsync(int l_width, int l_height, int l_depth,
                                                            hipArray_Format l_format) {
  width = l_width;
  height = l_height;
  depth = l_depth;
  formatKind = l_format;
}

/* Allocating Memory */
template <typename T> void DrvMemcpy3DAsync<T>::AllocateMemory() {
  size = width * height * depth * sizeof(T);
  hData = reinterpret_cast<T*>(malloc(size));
  memset(hData, 0, size);
  for (int i = 0; i < depth; i++) {
    for (int j = 0; j < height; j++) {
      for (int k = 0; k < width; k++) {
        hData[i * width * height + j * width + k] = i * width * height + j * width + k;
      }
    }
  }
  HIP_CHECK(hipStreamCreate(&stream));
  HIP_CHECK(hipMallocPitch(reinterpret_cast<void**>(&D_m), &pitch_D, width * sizeof(T), height));
  HIP_CHECK(hipMallocPitch(reinterpret_cast<void**>(&E_m), &pitch_E, width * sizeof(T), height));
  HIP_ARRAY3D_DESCRIPTOR* desc;
  desc = reinterpret_cast<HIP_ARRAY3D_DESCRIPTOR*>(malloc(sizeof(HIP_ARRAY3D_DESCRIPTOR)));
  desc->Format = formatKind;
  desc->NumChannels = 1;
  desc->Width = width;
  desc->Height = height;
  desc->Depth = depth;
  desc->Flags = hipArrayDefault;
  HIP_CHECK(hipArray3DCreate(&arr, desc));
  HIP_CHECK(hipArray3DCreate(&arr1, desc));
}

/* Setting the default data */
template <typename T> void DrvMemcpy3DAsync<T>::SetDefaultData() {
  memset(&myparms, 0x0, sizeof(HIP_MEMCPY3D));
  myparms.srcXInBytes = 0;
  myparms.srcY = 0;
  myparms.srcZ = 0;
  myparms.srcLOD = 0;
  myparms.dstXInBytes = 0;
  myparms.dstY = 0;
  myparms.dstZ = 0;
  myparms.dstLOD = 0;
  myparms.WidthInBytes = width * sizeof(T);
  myparms.Height = height;
  myparms.Depth = depth;
}

/*
This function verifies the negative scenarios of
hipDrvMemcpy3DAsync API
*/
template <typename T> void DrvMemcpy3DAsync<T>::NegativeTests() {
  HIP_CHECK(hipSetDevice(0));
  AllocateMemory();
  SetDefaultData();
  int deviceId;
  HIP_CHECK(hipGetDevice(&deviceId));
  unsigned int MaxPitch;
  HIP_CHECK(hipDeviceGetAttribute(reinterpret_cast<int*>(&MaxPitch), hipDeviceAttributeMaxPitch,
                                  deviceId));
  myparms.srcHost = hData;
  myparms.dstArray = arr;
  myparms.srcPitch = width * sizeof(T);
  myparms.srcHeight = height;
  myparms.srcMemoryType = hipMemoryTypeHost;
  myparms.dstMemoryType = hipMemoryTypeArray;

  SECTION("Passing nullptr to Source Host") {
    myparms.srcHost = nullptr;
    REQUIRE(hipDrvMemcpy3DAsync(&myparms, stream) != hipSuccess);
  }

  SECTION("Passing both dst host and device") {
    myparms.dstHost = hData;
    myparms.dstArray = nullptr;
    myparms.dstDevice = D_m;
    myparms.WidthInBytes = pitch_D;
    myparms.dstMemoryType = hipMemoryTypeDevice;
    REQUIRE(hipDrvMemcpy3DAsync(&myparms, stream) != hipSuccess);
  }

  SECTION("Passing max value to WidthInBytes") {
    myparms.WidthInBytes = std::numeric_limits<int>::max();
    myparms.Height = std::numeric_limits<int>::max();
    myparms.Depth = std::numeric_limits<int>::max();
    REQUIRE(hipDrvMemcpy3DAsync(&myparms, stream) != hipSuccess);
  }

  SECTION("Passing width > max width size") {
    myparms.WidthInBytes = width * sizeof(T) + 1;
    REQUIRE(hipDrvMemcpy3DAsync(&myparms, stream) != hipSuccess);
  }

  SECTION("Passing height > max height size") {
    myparms.Height = height + 1;
    REQUIRE(hipDrvMemcpy3DAsync(&myparms, stream) != hipSuccess);
  }

  SECTION("Passing depth > max depth size") {
    myparms.Depth = depth + 1;
    REQUIRE(hipDrvMemcpy3DAsync(&myparms, stream) != hipSuccess);
  }

  SECTION("widthinbytes + srcXinBytes is out of bound") {
    myparms.srcXInBytes = 1;
    myparms.dstArray = nullptr;
    myparms.dstDevice = hipDeviceptr_t(D_m);
    myparms.dstPitch = pitch_D;
    myparms.dstHeight = height;
    myparms.dstMemoryType = hipMemoryTypeDevice;
    REQUIRE(hipDrvMemcpy3DAsync(&myparms, stream) != hipSuccess);
  }

  SECTION("widthinbytes + dstXinBytes is out of bound") {
    myparms.dstXInBytes = pitch_D;
    myparms.dstArray = nullptr;
    myparms.dstDevice = hipDeviceptr_t(D_m);
    myparms.dstPitch = pitch_D;
    myparms.dstHeight = height;
    myparms.dstMemoryType = hipMemoryTypeDevice;
    REQUIRE(hipDrvMemcpy3DAsync(&myparms, stream) != hipSuccess);
  }

  SECTION("srcY + height is out of bound") {
    myparms.srcY = 1;
    myparms.dstArray = nullptr;
    myparms.dstDevice = hipDeviceptr_t(D_m);
    myparms.dstPitch = pitch_D;
    myparms.dstHeight = height;
    myparms.dstMemoryType = hipMemoryTypeDevice;
    REQUIRE(hipDrvMemcpy3DAsync(&myparms, stream) != hipSuccess);
  }

  SECTION("dstY + height out of bounds") {
    myparms.dstY = 1;
    myparms.dstArray = nullptr;
    myparms.dstDevice = hipDeviceptr_t(D_m);
    myparms.dstPitch = pitch_D;
    myparms.dstHeight = height;
    myparms.dstMemoryType = hipMemoryTypeDevice;
    REQUIRE(hipDrvMemcpy3DAsync(&myparms, stream) != hipSuccess);
  }

  SECTION("src pitch greater than Max allowed pitch") {
    myparms.srcMemoryType = hipMemoryTypeDevice;
    myparms.dstMemoryType = hipMemoryTypeHost;
    myparms.srcDevice = D_m;
    myparms.srcHost = nullptr;
    myparms.srcPitch = MaxPitch;
    myparms.srcHeight = height;
    myparms.dstHost = hData;
    myparms.dstArray = nullptr;
    myparms.dstPitch = width * sizeof(T);
    myparms.dstHeight = height;
    REQUIRE(hipDrvMemcpy3DAsync(&myparms, stream) != hipSuccess);
  }

  SECTION("dst pitch greater than Max allowed pitch") {
    myparms.dstDevice = hipDeviceptr_t(D_m);
    myparms.dstArray = nullptr;
    myparms.dstPitch = MaxPitch + 1;
    myparms.dstHeight = height;
    myparms.dstMemoryType = hipMemoryTypeDevice;
    REQUIRE(hipDrvMemcpy3DAsync(&myparms, stream) != hipSuccess);
  }

  SECTION("Nullptr to src/dst device") {
    myparms.dstDevice = hipDeviceptr_t(nullptr);
    myparms.dstArray = nullptr;
    myparms.dstPitch = pitch_D;
    myparms.dstHeight = height;
    myparms.dstMemoryType = hipMemoryTypeDevice;
    REQUIRE(hipDrvMemcpy3DAsync(&myparms, stream) != hipSuccess);
  }

  SECTION("Nullptr to src/dst array") {
    myparms.dstArray = nullptr;
    REQUIRE(hipDrvMemcpy3DAsync(&myparms, stream) != hipSuccess);
  }

  SECTION("Nullptr to hipDrvMemcpy3DAsync") {
    REQUIRE(hipDrvMemcpy3DAsync(nullptr, stream) != hipSuccess);
  }

  DeAllocateMemory();
}
/*
This function verifies the Extent validation scenarios of
hipDrvMemcpy3DAsync API
*/
template <typename T> void DrvMemcpy3DAsync<T>::Extent_Validation() {
  HIP_CHECK(hipSetDevice(0));
  // Allocating the memory
  AllocateMemory();

  // Setting default data
  SetDefaultData();
  myparms.srcMemoryType = hipMemoryTypeHost;
  myparms.dstMemoryType = hipMemoryTypeDevice;
  myparms.srcHost = hData;
  myparms.srcPitch = width * sizeof(T);
  myparms.srcHeight = height;
  myparms.dstDevice = D_m;
  myparms.dstPitch = pitch_D;
  myparms.dstHeight = height;

  SECTION("WidthInBytes is 0") {
    myparms.WidthInBytes = 0;
    HIP_CHECK(hipDrvMemcpy3DAsync(&myparms, stream));
    HIP_CHECK(hipStreamSynchronize(stream));
  }

  SECTION("Height is 0") {
    myparms.Height = 0;
    HIP_CHECK(hipDrvMemcpy3DAsync(&myparms, stream));
    HIP_CHECK(hipStreamSynchronize(stream));
  }

  SECTION("Depth is 0") {
    myparms.Depth = 0;
    HIP_CHECK(hipDrvMemcpy3DAsync(&myparms, stream));
    HIP_CHECK(hipStreamSynchronize(stream));
  }

  DeAllocateMemory();
}
/*
This Function verifies following functionalities of hipDrvMemcpy3DAsync API
1. Host to Device copy
2. Device to Device
3. Device to Host
In the end  validates the results.

This functionality is verified in 2 scenarios
1. Basic scenario on same GPU device
2. Device context change scenario where memory is allocated in 1 GPU
   and hipDrvMemcpy3DAsync API is trigerred from another GPU
*/
template <typename T>
void DrvMemcpy3DAsync<T>::HostDevice_DrvMemcpy3DAsync(bool device_context_change) {
  HIP_CHECK(hipSetDevice(0));
  bool skip_test = false;
  int peerAccess = 0;
  AllocateMemory();
  if (device_context_change) {
    HIP_CHECK(hipDeviceCanAccessPeer(&peerAccess, 0, 1));
    if (!peerAccess) {
      HIP_SKIP_TEST(HipTest::SkipReason::kPeerAccessUnavailable);
      skip_test = true;
    } else {
      HIP_CHECK(hipSetDevice(1));
    }
  }
  if (!skip_test) {
    SetDefaultData();
    myparms.srcMemoryType = hipMemoryTypeHost;
    myparms.dstMemoryType = hipMemoryTypeDevice;
    myparms.srcHost = hData;
    myparms.srcPitch = width * sizeof(T);
    myparms.srcHeight = height;
    myparms.dstDevice = hipDeviceptr_t(D_m);
    myparms.dstPitch = pitch_D;
    myparms.dstHeight = height;
    HIP_CHECK(hipDrvMemcpy3DAsync(&myparms, stream));
    HIP_CHECK(hipStreamSynchronize(stream));

    // Device to Device
    SetDefaultData();
    myparms.srcMemoryType = hipMemoryTypeDevice;
    myparms.dstMemoryType = hipMemoryTypeDevice;
    myparms.srcDevice = hipDeviceptr_t(D_m);
    myparms.srcPitch = pitch_D;
    myparms.srcHeight = height;
    myparms.dstDevice = hipDeviceptr_t(E_m);
    myparms.dstPitch = pitch_E;
    myparms.dstHeight = height;
    HIP_CHECK(hipDrvMemcpy3DAsync(&myparms, stream));
    HIP_CHECK(hipStreamSynchronize(stream));
    T* hOutputData = reinterpret_cast<T*>(malloc(size));
    memset(hOutputData, 0, size);

    // Device to host
    SetDefaultData();
    myparms.srcMemoryType = hipMemoryTypeDevice;
    myparms.dstMemoryType = hipMemoryTypeHost;
    myparms.srcDevice = hipDeviceptr_t(E_m);
    myparms.srcPitch = pitch_E;
    myparms.srcHeight = height;
    myparms.dstHost = hOutputData;
    myparms.dstPitch = width * sizeof(T);
    myparms.dstHeight = height;
    HIP_CHECK(hipDrvMemcpy3DAsync(&myparms, stream));
    HIP_CHECK(hipStreamSynchronize(stream));

    HipTest::checkArray(hData, hOutputData, width, height, depth);
    free(hOutputData);
  }
  DeAllocateMemory();
}

/*
This Function verifies following functionalities of hipDrvMemcpy3DAsync API
1. Host to Array copy
2. Array to Array
3. Array to Host
In the end  validates the results.

This functionality is verified in 2 scenarios
1. Basic scenario on same GPU device
2. Device context change scenario where memory is allocated in 1 GPU
   and hipDrvMemcpy3DAsync API is trigerred from another GPU
*/
template <typename T>
void DrvMemcpy3DAsync<T>::HostArray_DrvMemcpy3DAsync(bool device_context_change) {
  HIP_CHECK(hipSetDevice(0));
  bool skip_test = false;
  int peerAccess = 0;
  AllocateMemory();
  if (device_context_change) {
    HIP_CHECK(hipDeviceCanAccessPeer(&peerAccess, 0, 1));
    if (!peerAccess) {
      HIP_SKIP_TEST(HipTest::SkipReason::kPeerAccessUnavailable);
      skip_test = true;
    } else {
      HIP_CHECK(hipSetDevice(1));
    }
  }
  if (!skip_test) {
    SetDefaultData();
    myparms.srcMemoryType = hipMemoryTypeHost;
    myparms.dstMemoryType = hipMemoryTypeArray;
    myparms.srcHost = hData;
    myparms.srcPitch = width * sizeof(T);
    myparms.srcHeight = height;
    myparms.dstArray = arr;
    HIP_CHECK(hipDrvMemcpy3DAsync(&myparms, stream));
    HIP_CHECK(hipStreamSynchronize(stream));
    // Array to Array
    SetDefaultData();
    myparms.srcMemoryType = hipMemoryTypeArray;
    myparms.dstMemoryType = hipMemoryTypeArray;
    myparms.srcArray = arr;
    myparms.dstArray = arr1;
    HIP_CHECK(hipDrvMemcpy3DAsync(&myparms, stream));
    HIP_CHECK(hipStreamSynchronize(stream));
    T* hOutputData = reinterpret_cast<T*>(malloc(size));
    memset(hOutputData, 0, size);
    SetDefaultData();
    // Device to host
    myparms.srcMemoryType = hipMemoryTypeArray;
    myparms.dstMemoryType = hipMemoryTypeHost;
    myparms.srcArray = arr1;
    myparms.dstHost = hOutputData;
    myparms.dstPitch = width * sizeof(T);
    myparms.dstHeight = height;
    HIP_CHECK(hipDrvMemcpy3DAsync(&myparms, stream));
    HIP_CHECK(hipStreamSynchronize(stream));

    HipTest::checkArray(hData, hOutputData, width, height, depth);
    free(hOutputData);
  }
  DeAllocateMemory();
}

/* DeAllocating the memory */
template <typename T> void DrvMemcpy3DAsync<T>::DeAllocateMemory() {
  HIP_CHECK(hipArrayDestroy(arr));
  HIP_CHECK(hipArrayDestroy(arr1));
  HIP_CHECK(hipStreamDestroy(stream));
  free(hData);
}


/**
 * Test Description
 * ------------------------
 *  - This testcase verifies H2D copy in device context
      change scenario for hipDrvMemcpy3DAsync API
 * Test source
 * ------------------------
 *  - unit/memory/hipDrvMemcpy3DAsync_old.cc
 * Test requirements
 * ------------------------
 *  - HIP_VERSION >= 6.0
 */

HIP_TEST_CASE(Unit_hipDrvMemcpy3DAsync_H2DDeviceContextChange) {
  CHECK_IMAGE_SUPPORT
  int numDevices = 0;
  HIP_CHECK(hipGetDeviceCount(&numDevices));
  if (numDevices > 1) {
    DrvMemcpy3DAsync<float> memcpy3d(10, 10, 1, HIP_AD_FORMAT_FLOAT);
    memcpy3d.HostDevice_DrvMemcpy3DAsync(true);
  } else {
    HIP_SKIP_TEST(HipTest::SkipReason::kFewerThanTwoGpus);
  }
}

/**
 * Test Description
 * ------------------------
 *  - This testcase verifies Host to Array copy in device context
      change scenario for hipDrvMemcpy3DAsync API
 * Test source
 * ------------------------
 *  - unit/memory/hipDrvMemcpy3DAsync_old.cc
 * Test requirements
 * ------------------------
 *  - HIP_VERSION >= 6.0
 */

HIP_TEST_CASE(Unit_hipDrvMemcpy3DAsync_Host2ArrayDeviceContextChange) {
  CHECK_IMAGE_SUPPORT
  int numDevices = 0;
  HIP_CHECK(hipGetDeviceCount(&numDevices));
  if (numDevices > 1) {
    DrvMemcpy3DAsync<float> memcpy3d(10, 10, 10, HIP_AD_FORMAT_FLOAT);
    memcpy3d.HostArray_DrvMemcpy3DAsync(true);
  } else {
    HIP_SKIP_TEST(HipTest::SkipReason::kFewerThanTwoGpus);
  }
}

/**
 * Test Description
 * ------------------------
 *  - This testcase performs multidevice size check on DrvMemcpy3DAsync API
      1. Verify with 128 for all height, width & depth value
      2. Verify with 256 for height and 128 for width & depth value
      3. Verify with 256 for width and 128 for height & depth value
      4. Verify with 256 for depth and 128 for height & width value
 * Test source
 * ------------------------
 *  - unit/memory/hipDrvMemcpy3DAsync_old.cc
 * Test requirements
 * ------------------------
 *  - HIP_VERSION >= 6.0
 */

HIP_TEST_CASE(Unit_hipDrvMemcpy3DAsync_multiDevice_Basic_Size_Test) {
  CHECK_IMAGE_SUPPORT
  constexpr int size_128b = 128, size_256b = 256;
  int numDevices = 0;
  HIP_CHECK(hipGetDeviceCount(&numDevices));

  for (int i = 0; i < numDevices; i++) {
    HIP_CHECK(hipSetDevice(i));

    SECTION("Verify with 128 for all height, width & depth value") {
      DrvMemcpy3DAsync<int> memcpy3d(size_128b, size_128b, size_128b, HIP_AD_FORMAT_SIGNED_INT32);
      memcpy3d.HostArray_DrvMemcpy3DAsync();
    }
    SECTION("Verify with 256 for height and 128 for width & depth value") {
      DrvMemcpy3DAsync<int> memcpy3d(size_256b, size_128b, size_128b, HIP_AD_FORMAT_SIGNED_INT32);
      memcpy3d.HostArray_DrvMemcpy3DAsync();
    }
    SECTION("Verify with 256 for width and 128 for height & depth value") {
      DrvMemcpy3DAsync<float> memcpy3d(size_128b, size_256b, size_128b, HIP_AD_FORMAT_FLOAT);
      memcpy3d.HostArray_DrvMemcpy3DAsync();
    }
    SECTION("Verify with 256 for depth and 128 for height & width value") {
      DrvMemcpy3DAsync<int> memcpy3d(size_128b, size_128b, size_256b, HIP_AD_FORMAT_SIGNED_INT32);
      memcpy3d.HostArray_DrvMemcpy3DAsync();
    }
  }
}

/**
 * End doxygen group MemoryTest.
 * @}
 */