TheRock/rocm-systems/projects/hip-tests/catch/performance/scenarios/vmm/hipPerfVMMAlloc.cc

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

/**
 * @addtogroup hipMemCreate hipMemCreate
 * @{
 * @ingroup PerformanceTestVMM
 * `hipMemCreate(hipMemGenericAllocationHandle_t* handle, size_t size,
 *               const hipMemAllocationProp* prop, unsigned long long flags)` -
 * Creates a memory allocation described by the properties and size.
 */

#include <hip_test_common.hh>

/**
 * Test Description
 * ------------------------
 *  - Verify hipPerfBufferCopySpeed status.
 * Test source
 * ------------------------
 * - performance/scenarios/vmm/hipPerfVMMAlloc.cc
 * Test requirements
 * ------------------------
 *  - HIP_VERSION >= 5.6
 */

// Control Variables
constexpr bool single_map = false;
constexpr bool debug_failure = false;
constexpr size_t kMB = (1024 * 1024);
constexpr size_t kGB = (1024 * 1024 * 1024);
constexpr size_t chunk_size = (64 * kMB);

bool CheckVMMSupportedOnDevice(int deviceId) {
  int value = 0;
  hipDeviceAttribute_t attr = hipDeviceAttributeVirtualMemoryManagementSupported;
  HIP_CHECK(hipDeviceGetAttribute(&value, attr, deviceId));
  return static_cast<bool>(value);
}

bool GetVMMGranularityOnDevice(int deviceId, size_t& granularity) {
  hipMemAllocationProp prop{};
  prop.type = hipMemAllocationTypePinned;
  prop.location.type = hipMemLocationTypeDevice;
  prop.location.id = deviceId;  // Current Devices
  HIP_CHECK(
      hipMemGetAllocationGranularity(&granularity, &prop, hipMemAllocationGranularityMinimum));
  return true;
}

template <typename T> size_t GetSizeN(size_t total_size) {
  if (total_size % sizeof(T) != 0) {
    INFO("Size " << total_size << " is not a multiple of size of type T");
    assert(false);
  }
  return (total_size / sizeof(T));
}

bool ValidateUsingCopy(int deviceId, void* dev_ptr, size_t data_size,
                       std::chrono::microseconds& h2d_elapsed,
                       std::chrono::microseconds& d2h_elapsed) {
  // Get Host Data
  std::vector<int> A_h(data_size), B_h(data_size);
  size_t size_n = GetSizeN<int>(data_size);

  for (size_t idx = 0; idx < size_n; ++idx) {
    A_h[idx] = idx;
    B_h[idx] = 0;
  }

  HIP_CHECK(hipSetDevice(deviceId));
  auto start = std::chrono::high_resolution_clock::now();
  HIP_CHECK(hipMemcpy(dev_ptr, A_h.data(), data_size, hipMemcpyHostToDevice));
  auto end = std::chrono::high_resolution_clock::now();
  h2d_elapsed = std::chrono::duration_cast<std::chrono::microseconds>(end - start);

  start = std::chrono::high_resolution_clock::now();
  HIP_CHECK(hipMemcpy(B_h.data(), dev_ptr, data_size, hipMemcpyDeviceToHost));
  end = std::chrono::high_resolution_clock::now();
  d2h_elapsed = std::chrono::duration_cast<std::chrono::microseconds>(end - start);

  if (debug_failure) {
    REQUIRE(true == std::equal(B_h.begin(), B_h.end(), A_h.data()));
  } else {
    assert(A_h.size() == B_h.size());
    for (size_t idx = 0; idx < A_h.size(); ++idx) {
      if (A_h[idx] != B_h[idx]) {
        std::cout << "Failed at first index: " << idx << " Expected: " << A_h[idx]
                  << " Value: " << B_h[idx] << std::endl;
        break;
      }
    }
  }
  return true;
}

bool TestOnDevice(int deviceId) {
  HIP_CHECK(hipSetDevice(deviceId));
  // Check if VMM is supported
  if (!CheckVMMSupportedOnDevice(deviceId)) {
    INFO("VMM is not suppored on device: " << deviceId);
    return false;
  }
  // Get VMM granularity
  size_t granularity = 0;
  if (!GetVMMGranularityOnDevice(deviceId, granularity)) {
    INFO("Cannot get granularity on device: " << deviceId);
    return false;
  }

  // Measure CPU time of allocation taken
  size_t start_size = 1 * kGB;
  size_t max_size = 64 * kGB;
  for (size_t size_idx = start_size; size_idx <= max_size; (size_idx <<= 1)) {
    void* dev_ptr = nullptr;
    // This seems to be a completely blocking call, measuring CPU time in this test for now.
    // Create Memory Reservation
    auto start = std::chrono::high_resolution_clock::now();
    HIP_CHECK(hipMemAddressReserve(&dev_ptr, size_idx, granularity, nullptr, 0));
    auto end = std::chrono::high_resolution_clock::now();
    std::chrono::microseconds reserve_elapsed =
        std::chrono::duration_cast<std::chrono::microseconds>(end - start);
    std::vector<hipMemGenericAllocationHandle_t> physmem_handles;
    std::chrono::microseconds alloc_elapsed;
    std::chrono::microseconds map_elapsed;
    assert(size_idx % chunk_size == 0);
    size_t chunk_max = size_idx / chunk_size;
    size_t chunk_idx = 0;

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

    if (freeVRAM < size_idx) {
      WARN("Further free device memory unavailable, hence exiting!");
      break;
    }

    if (single_map) {
      // Create Physical memory
      hipMemAllocationProp prop{};
      prop.type = hipMemAllocationTypePinned;
      prop.location.type = hipMemLocationTypeDevice;
      prop.location.id = deviceId;
      hipMemGenericAllocationHandle_t handle;

      start = std::chrono::high_resolution_clock::now();
      HIP_CHECK(hipMemCreate(&handle, size_idx, &prop, 0));
      end = std::chrono::high_resolution_clock::now();
      alloc_elapsed = std::chrono::duration_cast<std::chrono::microseconds>(end - start);
      physmem_handles.push_back(handle);

      // Map the memory
      start = std::chrono::high_resolution_clock::now();
      HIP_CHECK(hipMemMap(dev_ptr, size_idx, 0, handle, 0));
      end = std::chrono::high_resolution_clock::now();
      map_elapsed = std::chrono::duration_cast<std::chrono::microseconds>(end - start);
    } else {
      start = std::chrono::high_resolution_clock::now();
      while (chunk_idx < chunk_max) {
        hipMemAllocationProp prop{};
        prop.type = hipMemAllocationTypePinned;
        prop.location.type = hipMemLocationTypeDevice;
        prop.location.id = deviceId;
        hipMemGenericAllocationHandle_t handle;
        HIP_CHECK(hipMemCreate(&handle, chunk_size, &prop, 0));
        physmem_handles.push_back(handle);
        ++chunk_idx;
      }
      end = std::chrono::high_resolution_clock::now();
      alloc_elapsed = std::chrono::duration_cast<std::chrono::microseconds>(end - start);

      chunk_idx = 0;
      start = std::chrono::high_resolution_clock::now();
      while (chunk_idx < chunk_max) {
        // Use chunk size to map multiple maps
        uint64_t uiptr = reinterpret_cast<uint64_t>(dev_ptr);
        uiptr = uiptr + chunk_idx * chunk_size;
        HIP_CHECK(hipMemMap(reinterpret_cast<void*>(uiptr), chunk_size, 0,
                            physmem_handles[chunk_idx], 0));
        ++chunk_idx;
      }
      end = std::chrono::high_resolution_clock::now();
      map_elapsed = std::chrono::duration_cast<std::chrono::microseconds>(end - start);

      chunk_idx = 0;
      while (chunk_idx < chunk_max) {
        // Set access
        hipMemAccessDesc accessDesc = {};
        accessDesc.location.type = hipMemLocationTypeDevice;
        accessDesc.location.id = deviceId;
        accessDesc.flags = hipMemAccessFlagsProtReadWrite;
        uint64_t uiptr = reinterpret_cast<uint64_t>(dev_ptr);
        uiptr = uiptr + chunk_idx * chunk_size;
        // Make the address accessible to GPU 0
        HIP_CHECK(hipMemSetAccess(reinterpret_cast<void*>(uiptr), chunk_size, &accessDesc, 1));
        ++chunk_idx;
      }
    }

    // Also measure the memcpy time elapsed
    std::chrono::microseconds h2d_elapsed;
    std::chrono::microseconds d2h_elapsed;

    // Validate using copy
    if (!ValidateUsingCopy(deviceId, dev_ptr, size_idx, h2d_elapsed, d2h_elapsed)) {
      INFO("Validation failed for size: " << size_idx);
    }

    start = std::chrono::high_resolution_clock::now();
    chunk_idx = 0;
    while (chunk_idx < chunk_max) {
      uint64_t uiptr = reinterpret_cast<uint64_t>(dev_ptr);
      uiptr = uiptr + chunk_idx * chunk_size;
      HIP_CHECK(hipMemUnmap(reinterpret_cast<void*>(uiptr), chunk_size));
      ++chunk_idx;
    }
    end = std::chrono::high_resolution_clock::now();
    std::chrono::microseconds unmap_elapsed =
        std::chrono::duration_cast<std::chrono::microseconds>(end - start);

    start = std::chrono::high_resolution_clock::now();
    for (auto& physmem_handle : physmem_handles) {
      HIP_CHECK(hipMemRelease(physmem_handle));
    }
    end = std::chrono::high_resolution_clock::now();
    std::chrono::microseconds release_elapsed =
        std::chrono::duration_cast<std::chrono::microseconds>(end - start);

    start = std::chrono::high_resolution_clock::now();
    HIP_CHECK(hipMemAddressFree(dev_ptr, size_idx));
    end = std::chrono::high_resolution_clock::now();
    std::chrono::microseconds free_elapsed =
        std::chrono::duration_cast<std::chrono::microseconds>(end - start);

    // Print the results
    std::cout << "-------------Size: " << (size_idx / kGB) << " GB----------------" << std::endl;
    std::cout << "Time taken to reserve : " << reserve_elapsed.count()
              << " micro seconds and free: " << free_elapsed.count() << " micro seconds"
              << std::endl;
    std::cout << "Time taken to alloc : " << alloc_elapsed.count()
              << " micro seconds and release: " << release_elapsed.count() << " micro seconds"
              << std::endl;
    std::cout << "Time taken to map : " << map_elapsed.count()
              << " micro seconds and unmap: " << unmap_elapsed.count() << " micro seconds"
              << std::endl;
    std::cout << "Time taken to H2D : " << h2d_elapsed.count()
              << " micro seconds and D2H: " << d2h_elapsed.count() << " micro seconds" << std::endl;
    std::cout << "-------------------------/hipMallocPerf------------------------" << std::endl;

    void* dev_ptr_legacy = nullptr;
    start = std::chrono::high_resolution_clock::now();
    HIP_CHECK(hipMalloc(&dev_ptr_legacy, size_idx));
    end = std::chrono::high_resolution_clock::now();
    std::chrono::microseconds hm_elapsed =
        std::chrono::duration_cast<std::chrono::microseconds>(end - start);
    start = std::chrono::high_resolution_clock::now();
    HIP_CHECK(hipFree(dev_ptr_legacy));
    end = std::chrono::high_resolution_clock::now();
    std::chrono::microseconds hf_elapsed =
        std::chrono::duration_cast<std::chrono::microseconds>(end - start);
    std::cout << "Time taken for hipMalloc : " << hm_elapsed.count()
              << " micro seconds and hipFree: " << hf_elapsed.count() << " micro seconds"
              << std::endl;
    std::cout << "---------------------------------------------------------------" << std::endl;
    std::cout << std::endl;
  }

  return true;
}

HIP_TEST_CASE(Performance_hipPerfVMMAllocSpeed_test) {
  int numDevices = 0;
  HIP_CHECK(hipGetDeviceCount(&numDevices));
  if (numDevices <= 0) {
    HIP_SKIP_TEST(HipTest::SkipReason::kNoGpuDevice);
  } else {
    // Test on Primary Device first
    int deviceId = 0;
    TestOnDevice(deviceId);
  }
}
/**
 * End doxygen group perfVMMTest.
 * @}
 */