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

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

#include <hip_test_common.hh>
#include <algorithm>
#include <initializer_list>
#include <memory>
#include "hip/driver_types.h"
#include <cstring>
#include <vector>
#include <limits>
#include <random>
#include <hip_test_checkers.hh>
#include <hip_test_kernels.hh>
#ifdef __HIP_PLATFORM_NVIDIA__
#include "DriverContext.hh"
#endif

/**
 * @brief Test hipMalloc3D, hipMallocPitch and hipMemAllocPitch with multiple input values.
 *        Checks that the memory has been allocated with the specified pitch and extent sizes.
 */

static void validateMemory(void* devPtr, hipExtent extent, size_t pitch) {
  INFO("Width: " << extent.width << " Height: " << extent.height << " Depth: " << extent.depth);

  const size_t theoreticalAllocatedMemory{pitch * extent.height * extent.depth};
  if (theoreticalAllocatedMemory == 0) {
    return; /* If there was no memory allocated then we don't need to do further checks. */
  }

  std::unique_ptr<char[]> hostPtr{new char[theoreticalAllocatedMemory]};
  std::memset(hostPtr.get(), 2, theoreticalAllocatedMemory);

  hipPitchedPtr devPitchedPtr{devPtr, pitch, extent.width, extent.height};
  hipPitchedPtr hostPitchedPtr{hostPtr.get(), pitch, extent.width, extent.height};

  HIP_CHECK(hipMemset3D(devPitchedPtr, 1, extent));

  hipMemcpy3DParms params{};
  params.srcPtr = devPitchedPtr;
  params.kind = hipMemcpyKind::hipMemcpyDeviceToHost;
  params.dstPtr = hostPitchedPtr;
  params.extent = extent;
  HIP_CHECK(hipMemcpy3D(&params))

  bool mismatch = false;
  for (size_t width = 0; width < extent.width; ++width) {
    for (size_t height = 0; height < extent.height; ++height) {
      for (size_t depth = 0; depth < extent.depth; ++depth) {
        char* reinterpretedPtr = reinterpret_cast<char*>(hostPtr.get());
        size_t index = (pitch * extent.height * depth) + (pitch * height) + width;
        if (*(reinterpretedPtr + index) != 1) {
          mismatch = true;
        }
      }
    }
  }
  REQUIRE(!mismatch);
}

class ExtentGenerator {
 public:
  static constexpr size_t totalRandomValues{20};
  static constexpr size_t seed{1337};

  std::uniform_int_distribution<size_t> width_distribution{1, 1024};
  std::uniform_int_distribution<size_t> height_distribution{1, 100};
  std::uniform_int_distribution<size_t> depth_distribution{1, 100};

  std::vector<hipExtent> extents2D{};
  std::vector<hipExtent> extents3D{};

  static ExtentGenerator& getInstance() {
    static ExtentGenerator instance;
    return instance;
  }

 private:
  ExtentGenerator() {
    std::mt19937 randomGenerator{seed};
    extents3D = std::vector<hipExtent>{hipExtent{0, 0, 0}, hipExtent{1, 0, 0}, hipExtent{0, 1, 0},
                                       hipExtent{0, 0, 1}};

    for (size_t i = 0; i < totalRandomValues; ++i) {
      extents3D.push_back(hipExtent{width_distribution(randomGenerator),
                                    height_distribution(randomGenerator),
                                    depth_distribution(randomGenerator)});
    }

    extents2D = std::vector<hipExtent>{hipExtent{0, 0, 1}, hipExtent{1, 0, 1}, hipExtent{0, 1, 1}};

    for (size_t i = 0; i < totalRandomValues; ++i) {
      extents2D.push_back(
          hipExtent{width_distribution(randomGenerator), height_distribution(randomGenerator), 1});
    }
  }
};

enum class AllocationApi { hipMalloc3D, hipMallocPitch, hipMemAllocPitch };

hipExtent generateExtent(AllocationApi api) {
  hipExtent extent;
  if (api == AllocationApi::hipMalloc3D) {
    auto& extents3D = ExtentGenerator::getInstance().extents3D;
    // At level_0: 4 edge cases + 4 random = 8 extents (vs 24 normally)
    auto end3D = isQuickLevel() ? extents3D.begin() + std::min(extents3D.size(), size_t(8))
                                : extents3D.end();
    extent = GENERATE_REF(from_range(extents3D.begin(), end3D));
  } else {
    auto& extents2D = ExtentGenerator::getInstance().extents2D;
    extent = GENERATE_REF(from_range(extents2D.begin(), extents2D.end()));
  }

  return extent;
}


HIP_TEST_CASE(Unit_hipMalloc3D_ValidatePitch) {

  hipPitchedPtr hipPitchedPtr;
  hipExtent validExtent{generateExtent(AllocationApi::hipMalloc3D)};

  HIP_CHECK(hipMalloc3D(&hipPitchedPtr, validExtent));
  validateMemory(hipPitchedPtr.ptr, validExtent, hipPitchedPtr.pitch);
  HIP_CHECK(hipFree(hipPitchedPtr.ptr));
}

HIP_TEST_CASE(Unit_hipMemAllocPitch_ValidatePitch) {

  size_t pitch = 0;
  hipDeviceptr_t ptr;
  hipExtent validExtent{generateExtent(AllocationApi::hipMemAllocPitch)};
  unsigned int elementSizeBytes = GENERATE(4, 8, 16);

  if (validExtent.width == 0 || validExtent.height == 0) {
    return;
  }
// hipMemAllocPitch is driver API hence explicit init is required on NVidia plaform.
#ifdef __HIP_PLATFORM_NVIDIA__
  DriverContext ctx;
#endif
  HIP_CHECK(
      hipMemAllocPitch(&ptr, &pitch, validExtent.width, validExtent.height, elementSizeBytes));
  validateMemory(reinterpret_cast<void*>(ptr), validExtent, pitch);
  HIP_CHECK(hipFree(reinterpret_cast<void*>(ptr)));
}

HIP_TEST_CASE(Unit_hipMallocPitch_ValidatePitch) {

  size_t pitch = 0;
  void* ptr;
  hipExtent validExtent{generateExtent(AllocationApi::hipMemAllocPitch)};
  HIP_CHECK(hipMallocPitch(&ptr, &pitch, validExtent.width, validExtent.height));
  validateMemory(ptr, validExtent, pitch);
  HIP_CHECK(hipFree(ptr));
}

HIP_TEST_CASE(Unit_hipMalloc3D_Negative) {

  SECTION("Invalid ptr") {
    hipExtent validExtent{1, 1, 1};
    HIP_CHECK_ERROR(hipMalloc3D(nullptr, validExtent), hipErrorInvalidValue);
  }

  hipPitchedPtr ptr;
  constexpr size_t maxSizeT = std::numeric_limits<size_t>::max();

#if HT_NVIDIA
  // TODO-MATCH-ERRORCODE
  SECTION("Max size_t width") {
    hipExtent validExtent{maxSizeT, 1, 1};
    HIP_CHECK_ERROR(hipMalloc3D(&ptr, validExtent), hipErrorInvalidValue);
  }
#endif

  SECTION("Max size_t height") {
    hipExtent validExtent{1, maxSizeT, 1};
    HIP_CHECK_ERROR(hipMalloc3D(&ptr, validExtent), hipErrorOutOfMemory);
  }

  SECTION("Max size_t depth") {
    hipExtent validExtent{1, 1, maxSizeT};
    HIP_CHECK_ERROR(hipMalloc3D(&ptr, validExtent), hipErrorOutOfMemory);
  }

#if HT_NVIDIA
  // TODO-MATCH-ERRORCODE
  SECTION("Max size_t all dimensions") {
    hipExtent validExtent{maxSizeT, maxSizeT, maxSizeT};
    HIP_CHECK_ERROR(hipMalloc3D(&ptr, validExtent), hipErrorInvalidValue);
  }
#endif
}

HIP_TEST_CASE(Unit_hipMallocPitch_Negative) {

  size_t pitch = 0;
  void* ptr;
  constexpr size_t maxSizeT = std::numeric_limits<size_t>::max();

  SECTION("Invalid ptr") {
    HIP_CHECK_ERROR(hipMallocPitch(nullptr, &pitch, 1, 1), hipErrorInvalidValue);
  }

  SECTION("Invalid pitch") {
    HIP_CHECK_ERROR(hipMallocPitch(&ptr, nullptr, 1, 1), hipErrorInvalidValue);
  }

#if HT_NVIDIA
  // TODO-MATCH-ERRORCODE
  SECTION("Max size_t width") {
    HIP_CHECK_ERROR(hipMallocPitch(&ptr, &pitch, maxSizeT, 1), hipErrorInvalidValue);
  }
#endif

  SECTION("Max size_t height") {
    HIP_CHECK_ERROR(hipMallocPitch(&ptr, &pitch, 1, maxSizeT), hipErrorOutOfMemory);
  }
}

HIP_TEST_CASE(Unit_hipMallocPitch_Zero_Dims) {

  void* ptr = nullptr;
  size_t pitch = 0;

  SECTION("width == 0") {
    HIP_CHECK(hipMallocPitch(&ptr, &pitch, 0, 1));
    REQUIRE(ptr == nullptr);
  }

  SECTION("height == 0") {
    HIP_CHECK(hipMallocPitch(&ptr, &pitch, 1, 0));
    REQUIRE(ptr == nullptr);
  }
}

HIP_TEST_CASE(Unit_hipMemAllocPitch_Negative) {

  size_t pitch = 0;
  hipDeviceptr_t ptr{};
  unsigned int validElementSizeBytes{4};
  constexpr size_t maxSizeT = std::numeric_limits<size_t>::max();

#if HT_NVIDIA
  /* Device synchronize is used here to initialize the device.
   * Nvidia does not implicitly do it for this Api. And hipInit(0) does not work either.
   */
  HIP_CHECK(hipDeviceSynchronize());

  SECTION("Invalid elementSizeBytes") {
    unsigned int invalidElementSizeBytes = GENERATE(0, 7, 12, 17);
    HIP_CHECK_ERROR(hipMemAllocPitch(&ptr, &pitch, 1, 1, invalidElementSizeBytes),
                    hipErrorInvalidValue);
  }

  SECTION("Zero width") {
    HIP_CHECK_ERROR(hipMemAllocPitch(&ptr, &pitch, 0, 1, validElementSizeBytes),
                    hipErrorInvalidValue);
  }
  SECTION("Zero height") {
    HIP_CHECK_ERROR(hipMemAllocPitch(&ptr, &pitch, 1, 0, validElementSizeBytes),
                    hipErrorInvalidValue);
  }
#endif

  SECTION("Invalid dptr") {
    HIP_CHECK_ERROR(hipMemAllocPitch(nullptr, &pitch, 1, 1, validElementSizeBytes),
                    hipErrorInvalidValue);
  }

  SECTION("Invalid pitch") {
    HIP_CHECK_ERROR(hipMemAllocPitch(&ptr, nullptr, 1, 1, validElementSizeBytes),
                    hipErrorInvalidValue);
  }

#if HT_NVIDIA
  // TODO-MATCH-ERRORCODE
  SECTION("Max size_t width") {
    HIP_CHECK_ERROR(hipMemAllocPitch(&ptr, &pitch, maxSizeT, 1, validElementSizeBytes),
                    hipErrorInvalidValue);
  }
#endif

  SECTION("Max size_t height") {
    HIP_CHECK_ERROR(hipMemAllocPitch(&ptr, &pitch, 1, maxSizeT, validElementSizeBytes),
                    hipErrorOutOfMemory);
  }
}

/*
Test Scenarios of hipMallocPitch API
1. Basic Functionality Scenario
2. Allocate memory using hipMallocPitch API, Launch Kernel validate result.
3. Allocate Memory in small chunks and large chunks and check for possible memory leaks
4. Allocate Memory using hipMallocPitch API, Memcpy2D on the allocated variables.
5. Multithreaded scenario
*/

static constexpr auto SMALLCHUNK_NUMW{4};
static constexpr auto SMALLCHUNK_NUMH{4};
static constexpr auto LARGECHUNK_NUMW{1025};
static constexpr auto LARGECHUNK_NUMH{1000};
static constexpr auto NUM_W{10};
static constexpr auto NUM_H{10};
static constexpr auto COLUMNS{8};
static constexpr auto ROWS{8};
static constexpr auto CHUNK_LOOP{100};


template <typename T> __global__ void copy_var(T* A, T* B, size_t ROWS, size_t pitch_A, size_t pitch_B) {
  const size_t pitcha = pitch_A / sizeof(T), pitchb = pitch_B / sizeof(T);
  const size_t sizea = ROWS * pitcha, sizeb = ROWS * pitchb;

  for (size_t ia = 0, ib = 0; ia < sizea && ib < sizeb; ia += pitcha, ib += sizeb) {
    B[ib] = A[ia];
  }
}
template <typename T> static bool validateResult(T* A, T* B, size_t pitch_A) {
  bool testResult = true;
  for (uint64_t i = 0; i < pitch_A * ROWS; i = i + pitch_A) {
    if (A[i] != B[i]) {
      testResult = false;
      break;
    }
  }
  return testResult;
}
/*
 * This API verifies  memory allocations for small and
 * bigger chunks of data.
 * Two scenarios are verified in this API
 * 1. SmallChunk: Allocates SMALLCHUNK_NUMW in a loop and
 *    releases the memory and verifies the meminfo.
 * 2. LargeChunk: Allocates LARGECHUNK_NUMW in a loop and
 *    releases the memory and verifies the meminfo
 *
 * In both cases, the memory info before allocation and
 * after releasing the memory should be the same
 *
 */
template <typename T> static void MemoryAllocDiffSizes(int gpu) {
  HIP_CHECK(hipSetDevice(gpu));
  std::vector<size_t> array_size;
  array_size.push_back(SMALLCHUNK_NUMH);
  array_size.push_back(LARGECHUNK_NUMH);
  for (auto& sizes : array_size) {
    T* A_d[CHUNK_LOOP];
    size_t pitch_A = 0;
    size_t width;
    if (sizes == SMALLCHUNK_NUMH) {
      width = SMALLCHUNK_NUMW * sizeof(T);
    } else {
      width = LARGECHUNK_NUMW * sizeof(T);
    }
    for (int i = 0; i < CHUNK_LOOP; i++) {
      HIP_CHECK(hipMallocPitch(reinterpret_cast<void**>(&A_d[i]), &pitch_A, width, sizes));
    }
    for (int i = 0; i < CHUNK_LOOP; i++) {
      HIP_CHECK(hipFree(A_d[i]));
    }
  }
}

/*Thread Function */
static void threadFunc(int gpu) { MemoryAllocDiffSizes<float>(gpu); }

/*
 * This testcase verifies the basic scenario of
 * hipMallocPitch API for different datatypes
 *
 */
HIP_TEMPLATE_TEST_CASE(Unit_hipMallocPitch_Basic, int, unsigned int, float) {

  TestType* A_d;
  size_t pitch_A = 0;
  size_t width{NUM_W * sizeof(TestType)};
  REQUIRE(hipMallocPitch(reinterpret_cast<void**>(&A_d), &pitch_A, width, NUM_H) == hipSuccess);
  REQUIRE(width <= pitch_A);
  HIP_CHECK(hipFree(A_d));
}

/*
 * This testcase verifies hipMallocPitch API for small
 * and big chunks of data.
 */
HIP_TEMPLATE_TEST_CASE(Unit_hipMallocPitch_SmallandBigChunks, int, unsigned int,
                   float) {

  MemoryAllocDiffSizes<TestType>(0);
}

/*
 * This testcase verifies the memory allocated by hipMallocPitch API
 * by performing Memcpy2D on the allocated memory.
 */
HIP_TEMPLATE_TEST_CASE(Unit_hipMallocPitch_Memcpy2D, int, float, double) {

  HIP_CHECK(hipSetDevice(0));
  TestType *A_h{nullptr}, *B_h{nullptr}, *C_h{nullptr}, *A_d{nullptr}, *B_d{nullptr};
  size_t pitch_A = 0, pitch_B = 0;
  size_t width{NUM_W * sizeof(TestType)};

  // Allocating memory
  HipTest::initArrays<TestType>(nullptr, nullptr, nullptr, &A_h, &B_h, &C_h, NUM_W * NUM_H, false);
  HIP_CHECK(hipMallocPitch(reinterpret_cast<void**>(&A_d), &pitch_A, width, NUM_H));
  HIP_CHECK(hipMallocPitch(reinterpret_cast<void**>(&B_d), &pitch_B, width, NUM_H));

  // Initialize the data
  HipTest::setDefaultData<TestType>(NUM_W * NUM_H, A_h, B_h, C_h);

  // Host to Device
  HIP_CHECK(hipMemcpy2D(A_d, pitch_A, A_h, COLUMNS * sizeof(TestType), COLUMNS * sizeof(TestType),
                        ROWS, hipMemcpyHostToDevice));

  // Performs D2D on same GPU device
  HIP_CHECK(hipMemcpy2D(B_d, pitch_B, A_d, pitch_A, COLUMNS * sizeof(TestType), ROWS,
                        hipMemcpyDeviceToDevice));

  // hipMemcpy2D Device to Host
  HIP_CHECK(hipMemcpy2D(B_h, COLUMNS * sizeof(TestType), B_d, pitch_B, COLUMNS * sizeof(TestType),
                        ROWS, hipMemcpyDeviceToHost));

  // Validating the result
  REQUIRE(HipTest::checkArray<TestType>(A_h, B_h, COLUMNS, ROWS) == true);


  // DeAllocating the memory
  HIP_CHECK(hipFree(A_d));
  HIP_CHECK(hipFree(B_d));
  HipTest::freeArrays<TestType>(nullptr, nullptr, nullptr, A_h, B_h, C_h, false);
}


/*
This testcase verifies the hipMallocPitch API in multithreaded
scenario by launching threads in parallel on multiple GPUs
and verifies the hipMallocPitch API with small and big chunks data
*/

HIP_TEST_CASE(Unit_hipMallocPitch_MultiThread) {

  std::vector<std::thread> threadlist;
  int devCnt = 0;

  devCnt = HipTest::getDeviceCount();
  for (int i = 0; i < devCnt; i++) {
    threadlist.push_back(std::thread(threadFunc, i));
  }

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

/*
 * This testcase verifies hipMallocPitch API by
 *  1. Allocating Memory using hipMallocPitch API
 *  2. Launching the kernel and copying the data from the allocated kernel
 *     variable to another kernel variable.
 *  3. Validating the result
 */
HIP_TEMPLATE_TEST_CASE(Unit_hipMallocPitch_KernelLaunch, int, float, double) {

  HIP_CHECK(hipSetDevice(0));
  TestType *A_h{nullptr}, *B_h{nullptr}, *C_h{nullptr}, *A_d{nullptr}, *B_d{nullptr};
  size_t pitch_A = 0, pitch_B = 0;
  size_t width{NUM_W * sizeof(TestType)};

  // Allocating memory
  HipTest::initArrays<TestType>(nullptr, nullptr, nullptr, &A_h, &B_h, &C_h, NUM_W * NUM_H, false);
  HIP_CHECK(hipMallocPitch(reinterpret_cast<void**>(&A_d), &pitch_A, width, NUM_H));
  HIP_CHECK(hipMallocPitch(reinterpret_cast<void**>(&B_d), &pitch_B, width, NUM_H));

  // Host to Device
  HIP_CHECK(hipMemcpy2D(A_d, pitch_A, A_h, COLUMNS * sizeof(TestType), COLUMNS * sizeof(TestType),
                        ROWS, hipMemcpyHostToDevice));

  hipLaunchKernelGGL(copy_var<TestType>, dim3(1), dim3(1), 0, 0, static_cast<TestType*>(A_d),
                     static_cast<TestType*>(B_d), ROWS, pitch_A, pitch_B);
  HIP_CHECK(hipGetLastError());


  // hipMemcpy2D Device to Host
  HIP_CHECK(hipMemcpy2D(B_h, COLUMNS * sizeof(TestType), B_d, pitch_B, COLUMNS * sizeof(TestType),
                        ROWS, hipMemcpyDeviceToHost));

  // Validating the result
  validateResult(A_h, B_h, pitch_A);

  // DeAllocating the memory
  HIP_CHECK(hipFree(A_d));
  HIP_CHECK(hipFree(B_d));
  HipTest::freeArrays<TestType>(nullptr, nullptr, nullptr, A_h, B_h, C_h, false);
}