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

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

#ifdef _WIN32
#define NOMINMAX
#endif /* _WIN32 */

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

#ifdef __linux__
#include <sys/sysinfo.h>
#else
#include <windows.h>
#include <sysinfoapi.h>
#endif

static constexpr size_t ONE_MB = 1024 * 1024;
static constexpr size_t DEV_MEM_ALIGNMENT = 256;

HIP_TEST_CASE(Unit_hipMalloc_Positive_Basic) {
  constexpr size_t page_size = 4096;
  void* ptr = nullptr;
  const auto alloc_size =
      GENERATE_COPY(10, page_size / 2, page_size, page_size * 3 / 2, page_size * 2);
  HIP_CHECK(hipMalloc(&ptr, alloc_size));
  CHECK(ptr != nullptr);
  CHECK(reinterpret_cast<intptr_t>(ptr) % DEV_MEM_ALIGNMENT == 0);
  HIP_CHECK(hipFree(ptr));
}

HIP_TEST_CASE(Unit_hipMalloc_Positive_Zero_Size) {
  void* ptr = reinterpret_cast<void*>(0x1);
  HIP_CHECK(hipMalloc(&ptr, 0));
  REQUIRE(ptr == nullptr);
}

HIP_TEST_CASE(Unit_hipMalloc_Positive_Alignment) {
  void *ptr1 = nullptr, *ptr2 = nullptr;
  HIP_CHECK(hipMalloc(&ptr1, 1));
  HIP_CHECK(hipMalloc(&ptr2, 10));
  CHECK(reinterpret_cast<intptr_t>(ptr1) % DEV_MEM_ALIGNMENT == 0);
  CHECK(reinterpret_cast<intptr_t>(ptr2) % DEV_MEM_ALIGNMENT == 0);
  HIP_CHECK(hipFree(ptr1));
  HIP_CHECK(hipFree(ptr2));
}

HIP_TEST_CASE(Unit_hipMalloc_Negative_Parameters) {
  SECTION("ptr == nullptr") {
    HIP_CHECK_ERROR(hipMalloc(nullptr, 4096), hipErrorInvalidValue);
  }
  SECTION("size == max size_t") {
    void* ptr;
    HIP_CHECK_ERROR(hipMalloc(&ptr, std::numeric_limits<size_t>::max()), hipErrorOutOfMemory);
  }
  (void)hipGetLastError();
}

// Commenting this due to defect SWDEV-501675, used in below commented tests
#if 0
/**
/**
 * Local Function to Get total RAM size in bytes
 */
static inline size_t getTotalRAM() {
#ifdef __linux__
  struct sysinfo info {};
  sysinfo(&info);
  return info.totalram;
#else
  MEMORYSTATUSEX statex;
  statex.dwLength = sizeof(statex);
  GlobalMemoryStatusEx(&statex);
  return (statex.ullTotalPhys);
#endif
}

/**
 * Local Function to Get Available RAM size in bytes
 */
static inline size_t getAvailableRAM() {
#ifdef __linux__
  struct sysinfo info {};
  sysinfo(&info);
  return info.freeram;
#else
  MEMORYSTATUSEX statex;
  statex.dwLength = sizeof(statex);
  GlobalMemoryStatusEx(&statex);
  return (statex.ullAvailPhys);
#endif
}
#endif

/**
 * In addKernel function, all elements of the array a increased by 1
 */
static __global__ void addKernel(char* a, size_t size) {
  size_t idx = blockIdx.x * blockDim.x + threadIdx.x;
  size_t stride = blockDim.x * gridDim.x;
  for (size_t i = idx; i < size; i += stride) {
    a[i] += 1;
  }
}

/**
 * Local Function to perform some operations on device memory and validate it
 */
static void performOperations(char* devMem, size_t size) {
  /**
   * Validating 1MB or less than that, since the most of the scenarios uses
   * this function allocates memory in GB's and execution time will be more
   */
  const size_t sizeToCheck = (size < ONE_MB) ? size : ONE_MB;
  constexpr char value = 'A';

  HIP_CHECK(hipMemset(devMem, value, sizeToCheck));
  addKernel<<<1, 1>>>(devMem, sizeToCheck);
  HIP_CHECK(hipGetLastError());
  std::vector<char> arrToCheck(sizeToCheck, 0);

  HIP_CHECK(hipMemcpy(arrToCheck.data(), devMem, sizeToCheck, hipMemcpyDeviceToHost));

  for (size_t i = 0; i < sizeToCheck; i++) {
    INFO("At index : " << i << " Got value : " << arrToCheck[i] << " Expected value : B ");
    REQUIRE(arrToCheck[i] == 'B');
  }
}

/**
 * Test Description
 * ------------------------
 * - This test case tests the following scenario:-
 * - 1) Get the Available VRAM size
 * - 2) Allocate 90% of that using hipMalloc
 * - 3) Perform operations on device memory
 * Test source
 * ------------------------
 * - unit/memory/hipMalloc.cc
 */
HIP_TEST_CASE(Unit_hipMalloc_Allocate90PercentOfDeviceMemory) {
  char* devMem = nullptr;
  size_t freeVRAM = 0, totalVRAM = 0;
  HIP_CHECK(hipMemGetInfo(&freeVRAM, &totalVRAM));
  INFO("Available device memory : " << freeVRAM);

  /**
   * Allocating 90% of available VRAM.
   * Avoided allocating total available VRAM just for stability
   * and to keep some buffer memory.
   */
  size_t size = (freeVRAM * 9) / 10;
  INFO("Size going to allocate : " << size);

  HIP_CHECK(hipMalloc(&devMem, size));
  REQUIRE(devMem != nullptr);

  performOperations(devMem, size);
  HIP_CHECK(hipFree(devMem));
}

/**
 * Test Description
 * ------------------------
 * - APU-only. Allocates a single device buffer expected to exceed the
 *   dedicated-VRAM carveout, then exercises it with memory operations
 *   (memset, copy back, verify head and tail). Validates that hipMalloc
 *   honours the spill path for large allocations on unified memory and
 *   that the resulting buffer remains accessible for those operations.
 * Test source
 * ------------------------
 * - unit/memory/hipMalloc.cc
 */
HIP_TEST_CASE(Unit_hipMalloc_Positive_APU_LargeAllocSpill) {
  hipDeviceProp_t prop{};
  HIP_CHECK(hipGetDeviceProperties(&prop, 0));
  if (!prop.integrated) {
    HIP_SKIP_TEST("dGPU --- APU spill regression test does not apply");
    return;
  }
  // Assumes the dedicated-VRAM carveout is smaller than 5 GiB.
  constexpr size_t size = static_cast<size_t>(5) << 30;
  constexpr size_t headroom = static_cast<size_t>(1) << 30;
  if (prop.totalGlobalMem < size + headroom) {
    HIP_SKIP_TEST("APU totalGlobalMem too small for this allocation plus headroom");
    return;
  }

  char* devMem = nullptr;
  HIP_CHECK(hipMalloc(&devMem, size));
  REQUIRE(devMem != nullptr);

  constexpr int fill = 0xCD;
  HIP_CHECK(hipMemset(devMem, fill, size));

  // Sample-verify head and tail; full read-back is bandwidth-bound at GiB
  // scale but touching both ends catches page-table or aperture mismatches.
  constexpr size_t sample = static_cast<size_t>(64) * 1024;
  std::vector<char> head(sample), tail(sample);
  HIP_CHECK(hipMemcpy(head.data(), devMem, sample, hipMemcpyDeviceToHost));
  HIP_CHECK(hipMemcpy(tail.data(), devMem + (size - sample), sample,
                      hipMemcpyDeviceToHost));
  for (size_t i = 0; i < sample; ++i) {
    REQUIRE(head[i] == static_cast<char>(fill));
    REQUIRE(tail[i] == static_cast<char>(fill));
  }

  HIP_CHECK(hipFree(devMem));
}

// Commenting this due to defect SWDEV-501675
#if 0
/**
 * Test Description
 * ------------------------
 * - This test case tests the following scenario:-
 * - 1) Get the Available VRAM size
 * - 2) Allocate 110% of that using hipMalloc
 * - 3) Check the behaviour in linux and windows
 * - 4) In linux it should give hipErrorOutOfMemory
 * - 5) In windows it should success by using the Shared GPU memory
 * - 6) Perform operations on device memory
 * Test source
 * ------------------------
 * - unit/memory/hipMalloc.cc
 * Test requirements
 * ------------------------
 * - HIP_VERSION >= 6.4
 */
HIP_TEST_CASE(Unit_hipMalloc_Allocate110PercentOfDeviceMemory) {
  char *devMem = nullptr;
  size_t freeVRAM = 0, totalVRAM = 0;
  HIP_CHECK(hipMemGetInfo(&freeVRAM, &totalVRAM));
  INFO("Available Free device memory : " << freeVRAM);

  // Allocating 110% of available VRAM
  size_t size = freeVRAM * 1.1;
  INFO("Size Planning to allocate : " << size);

  // Determine amount of memory which is needed from RAM
  size_t requiredMemFromRAM = size - freeVRAM;
  INFO("Required Memory From RAM : " << requiredMemFromRAM);

#ifdef __linux__
  INFO("LINUX");
  HIP_CHECK_ERROR(hipMalloc(&devMem, size), hipErrorOutOfMemory);
#else
  INFO("WINDOWS");
  size_t totalRAM = getTotalRAM();
  size_t availableRAM = getAvailableRAM();
  INFO("Total RAM : " << totalRAM);
  INFO("Available RAM : " << availableRAM);

  /**
   * Validating size of Required Memory From RAM,
   *  1) Since we can use half of RAM as shared GPU memory,
   *     Required Memory From RAM should be less than that.
   *  2) And Required Memory From RAM should be less than Available RAM.
   * If Required Memory From RAM is crossing the threshold,
   * allocating 90% available VRAM
   */
  if ((requiredMemFromRAM < (totalRAM / 2)) &&
      (requiredMemFromRAM < availableRAM)) {
    INFO("Required Memory From RAM is Available");
  } else {
    INFO("Required Memory From RAM is Not available, "
         "allocating 90% of available VRAM");
    // Avoided using total available VRAM for stability, to keep buffer memory
    size = freeVRAM * 0.9;
  }
  INFO("Size going to allocate : " << size);

  HIP_CHECK(hipMalloc(&devMem, size));
  REQUIRE(devMem != nullptr);
  performOperations(devMem, size);
  HIP_CHECK(hipFree(devMem));
#endif
}
#endif

// Commenting this due to defect SWDEV-501675
#if 0
/**
 * Test Description
 * ------------------------
 * - This test case tests the following scenario:-
 * - 1) Get the Available VRAM size, Total RAM size and Available RAM size
 * - 2) In case if Available RAM size is greater than half of Total RAM,
 * -    allocate 90% of half of Total RAM
 * - 3) In case if Available RAM size is less than half of Total RAM,
 * -    allocate 50% of half of Available RAM
 * - 4) Check the behaviour in linux and windows,
 * -    In linux it should give hipErrorOutOfMemory and
 * -    in windows it should success by using the Shared GPU memory
 * - 5) Perform operations on device memory
 * Test source
 * ------------------------
 * - unit/memory/hipMalloc.cc
 * Test requirements
 * ------------------------
 * - HIP_VERSION >= 6.4
 */
HIP_TEST_CASE(Unit_hipMalloc_AllocateAvailableVRAMAndPossibleRAM) {
  char *devMem = nullptr;
  size_t freeVRAM = 0, totalVRAM = 0;
  HIP_CHECK(hipMemGetInfo(&freeVRAM, &totalVRAM));
  INFO("Available Free device memory : " << freeVRAM);

  size_t totalRAM = getTotalRAM();
  size_t availableRAM = getAvailableRAM();
  INFO("Total RAM : " << totalRAM);
  INFO("Available RAM : " << availableRAM);

  // Determine usable Shared GPU Memory by GPU From RAM which is half of RAM
  size_t usableSharedGPUMemFromRam = totalRAM / 2;

  if (availableRAM > usableSharedGPUMemFromRam) {
    /*
     * If Available RAM is greater than the usable Shared GPU Memory
     * Allocating 90% of Usable Shared GPU Memory
     * Not allocating 100% of Usable Shared GPU Memory just to keep
     * buffer memory and for stability
     */
    usableSharedGPUMemFromRam = usableSharedGPUMemFromRam * 0.9;
  } else {
    /*
     * If Available RAM is less than the usable Shared GPU Memory
     * Allocating 50% of RAM only just to keep buffer memory and for stability
     */
    usableSharedGPUMemFromRam = availableRAM * 0.5;
  }
  INFO("Usable Shared GPU Memory from RAM : " << usableSharedGPUMemFromRam);

  size_t size = freeVRAM + usableSharedGPUMemFromRam;
  INFO("Size going to allocate : " << size);

#ifdef __linux__
  INFO("LINUX");
  HIP_CHECK_ERROR(hipMalloc(&devMem, size), hipErrorOutOfMemory);
#else
  INFO("WINDOWS");
  HIP_CHECK(hipMalloc(&devMem, size));
  REQUIRE(devMem != nullptr);
  performOperations(devMem, size);
  HIP_CHECK(hipFree(devMem));
#endif
}

/**
 * Test Description
 * ------------------------
 * - This test case tests the following scenario:-
 * - 1) Get the Total RAM size
 * - 2) Allocate 1MB more than that using hipMalloc
 * - 3) Check the behaviour in linux and windows.
 * -    In linux, windows it should give hipErrorOutOfMemory
 * Test source
 * ------------------------
 * - unit/memory/hipMalloc.cc
 * Test requirements
 * ------------------------
 * - HIP_VERSION >= 6.4
 */
HIP_TEST_CASE(Unit_hipMalloc_AllocateMoreThanTotalRAM) {
  char *devMem = nullptr;

  size_t totalRAM = getTotalRAM();
  INFO("Total RAM : " << totalRAM);

  size_t size = totalRAM + ONE_MB;
  INFO("Size going to allocate : " << size);

  HIP_CHECK_ERROR(hipMalloc(&devMem, size), hipErrorOutOfMemory);
}

/**
 * Test Description
 * ------------------------
 * - This test case tests the following scenario:-
 * - 1) Get the Total VRAM size
 * - 2) Allocate 1MB more than total VRAM size using hipMalloc
 * - 3) Check the behaviour in linux and windows
 * - 4) In linux it should give hipErrorOutOfMemory
 * - 5) In windows it should success using shared GPU memory
 * - 6) Perform operations on device memory
 * Test source
 * ------------------------
 * - unit/memory/hipMalloc.cc
 * Test requirements
 * ------------------------
 * - HIP_VERSION >= 6.4
 */
HIP_TEST_CASE(Unit_hipMalloc_AllocateMoreThanTotalVRAM) {
  char *devMem = nullptr;

  size_t freeVRAM = 0, totalVRAM = 0;
  HIP_CHECK(hipMemGetInfo(&freeVRAM, &totalVRAM));
  INFO("Available total device memory : " << totalVRAM);

  size_t size = totalVRAM + ONE_MB;
  INFO("Size Planning to allocate : " << size);

#ifdef __linux__
  INFO("LINUX");
  HIP_CHECK_ERROR(hipMalloc(&devMem, size), hipErrorOutOfMemory);
#else
  INFO("WINDOWS");

  size_t requiredMemFromRAM = size - freeVRAM;
  INFO("Required Memory From RAM : " << requiredMemFromRAM);

  size_t totalRAM = getTotalRAM();
  size_t availableRAM = getAvailableRAM();
  INFO("Total RAM : " << totalRAM);
  INFO("Available RAM : " << availableRAM);

  /**
   * Validating size of Required Memory From RAM,
   * 1) Since we can use half of RAM as shared GPU memory,
   *    Required Memory From RAM should be less than that.
   * 2) And Required Memory From RAM should be less than Available RAM.
   * If Required Memory From RAM is crossing the threshold,
   * allocating 90% available VRAM
   */
  if ((requiredMemFromRAM < (totalRAM / 2)) &&
      (requiredMemFromRAM < availableRAM)) {
    INFO("Required Memory From RAM is Available");
  } else {
    INFO("Required Memory From RAM is Not available, "
         "allocating 90% of available VRAM");
    // Avoided using total available VRAM for stability, to keep buffer memory
    size = freeVRAM * 0.9;
  }
  INFO("Size going to allocate : " << size);

  HIP_CHECK(hipMalloc(&devMem, size));
  REQUIRE(devMem != nullptr);
  performOperations(devMem, size);
  HIP_CHECK(hipFree(devMem));
#endif
}
#endif