TheRock/rocm-systems/projects/hip-tests/catch/unit/stream/hipStreamWithCUMask.cc

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

/**
Testcase Scenarios :

1)Validates functionality of hipStreamAddCallback with created stream.

2)Validates functionality of stream with cu mask.

3)Create a stream with all CU masks disabled (0x00000000).
Verify that default CU mask is set for the stream.

4)Size is greater than physical CU number. In this case the extra elements
are ignored and hipExtStreamCreateWithCUMask must return hipSuccess.

5)Negative Testing of hipExtStreamCreateWithCUMask.
*/


#include <hip_test_common.hh>
#include <hip_test_kernels.hh>
#include <chrono>
#include <thread>
#include <iostream>
#include <vector>

#define NUM_CU_PARTITIONS 4
#define CONSTANT 1.618f
#define SIZE_INBYTES_OF_MB (1024 * 1024)
#define GRIDSIZE 512
#define BLOCKSIZE 256
#define ZERO_MASK 0x00000000


namespace hipExtStreamCreateWithCUMaskTest {

float *A_h, *C_h;
bool cbDone = false;
bool isPassed = true;
size_t N = 4 * SIZE_INBYTES_OF_MB;


// Make a default CU mask bit-array where all CUs are active
// this default mask is expected to be returned when there is no
// custom or global CU mask defined.
void createDefaultCUMask(std::vector<uint32_t>* pdefaultCUMask, int numOfCUs) {
  uint32_t temp = 0;
  uint32_t bit_index = 0;
  for (int i = 0; i < numOfCUs; i++) {
    temp |= 1UL << bit_index;
    if (bit_index >= 32) {
      (*pdefaultCUMask).push_back(temp);
      temp = 0;
      bit_index = 0;
      temp |= 1UL << bit_index;
    }
    bit_index += 1;
  }
  if (bit_index != 0) {
    (*pdefaultCUMask).push_back(temp);
  }
}
// Create masks of disabled CU masks.
void createDisabledCUMask(std::vector<uint32_t>* pdisabledCUMask, int numOfCUs) {
  uint32_t temp = ZERO_MASK;
  uint32_t bit_index = 0;
  for (int i = 0; i < numOfCUs; i++) {
    if (bit_index >= 32) {
      (*pdisabledCUMask).push_back(temp);
      temp = ZERO_MASK;
      bit_index = 0;
    }
    bit_index += 1;
  }
  if (bit_index != 0) {
    (*pdisabledCUMask).push_back(temp);
  }
}

void Callback(hipStream_t, hipError_t status, void* userData) {
  isPassed = true;
  HIP_CHECK(status);
  REQUIRE(userData == nullptr);
  for (size_t i = 0; i < N; i++) {
    if (C_h[i] != A_h[i] * A_h[i]) {
      UNSCOPED_INFO("Data mismatch at index: " << i);
      isPassed = false;
      break;
    }
  }
  cbDone = true;
}
}  // namespace hipExtStreamCreateWithCUMaskTest

using hipExtStreamCreateWithCUMaskTest::A_h;
using hipExtStreamCreateWithCUMaskTest::C_h;
using hipExtStreamCreateWithCUMaskTest::Callback;
using hipExtStreamCreateWithCUMaskTest::cbDone;
using hipExtStreamCreateWithCUMaskTest::createDefaultCUMask;
using hipExtStreamCreateWithCUMaskTest::createDisabledCUMask;
using hipExtStreamCreateWithCUMaskTest::isPassed;
using hipExtStreamCreateWithCUMaskTest::N;


/**
 * Scenario: Validates functionality of hipStreamAddCallback with created stream.
 */
HIP_TEST_CASE(Unit_hipExtStreamCreateWithCUMask_ValidateCallbackFunc) {
  float *A_d, *C_d;
  size_t Nbytes = N * sizeof(float);
  cbDone = false;
  A_h = reinterpret_cast<float*>(malloc(Nbytes));
  REQUIRE(A_h != nullptr);
  C_h = reinterpret_cast<float*>(malloc(Nbytes));
  REQUIRE(C_h != nullptr);

  // Fill with Phi + i
  for (size_t i = 0; i < N; i++) {
    A_h[i] = CONSTANT + i;
  }

  HIP_CHECK(hipMalloc(&A_d, Nbytes));
  HIP_CHECK(hipMalloc(&C_d, Nbytes));

  hipStream_t mystream;
  std::vector<uint32_t> defaultCUMask;
  HIP_CHECK(hipSetDevice(0));
  hipDeviceProp_t props;
  HIP_CHECK(hipGetDeviceProperties(&props, 0));
  createDefaultCUMask(&defaultCUMask, props.multiProcessorCount);

  HIP_CHECK(hipExtStreamCreateWithCUMask(&mystream, defaultCUMask.size(), defaultCUMask.data()));
  HIP_CHECK(hipMemcpyAsync(A_d, A_h, Nbytes, hipMemcpyHostToDevice, mystream));
  const unsigned threadsPerBlock = 1024;
  const int blocks =
      (N % threadsPerBlock == 0) ? (N / threadsPerBlock) : ((N / threadsPerBlock) + 1);
  hipLaunchKernelGGL((HipTest::vector_square), dim3(blocks), dim3(threadsPerBlock), 0, mystream,
                     A_d, C_d, N);
  HIP_CHECK(hipGetLastError());
  HIP_CHECK(hipMemcpyAsync(C_h, C_d, Nbytes, hipMemcpyDeviceToHost, mystream));
  HIP_CHECK(hipStreamAddCallback(mystream, Callback, nullptr, 0));
  while (!cbDone)
    std::this_thread::sleep_for(std::chrono::microseconds(100000));  // Sleep for 100 ms
  HIP_CHECK(hipStreamDestroy(mystream));
  HIP_CHECK(hipFree(reinterpret_cast<void*>(C_d)));
  HIP_CHECK(hipFree(reinterpret_cast<void*>(A_d)));
  free(C_h);
  free(A_h);
  REQUIRE(isPassed == true);
}

/**
 * Scenario: Validates functionality of stream with cu mask.
 */
HIP_TEST_CASE(Unit_hipExtStreamCreateWithCUMask_Functionality) {
  const int KNumPartition = NUM_CU_PARTITIONS;
  float *dA[KNumPartition], *dC[KNumPartition];
  float *hA, *hC;
  size_t N = 25 * SIZE_INBYTES_OF_MB;
  size_t Nbytes = N * sizeof(float);
  std::vector<hipStream_t> streams(KNumPartition);
  std::vector<std::vector<uint32_t>> cuMasks(KNumPartition);
  std::stringstream ss[KNumPartition];

  int nGpu = 0;
  HIP_CHECK(hipGetDeviceCount(&nGpu));
  if (nGpu < 1) {
    HIP_SKIP_TEST(HipTest::SkipReason::kNoGpuDevice);
  }

  static int device = 0;
  HIP_CHECK(hipSetDevice(device));
  hipDeviceProp_t props;
  HIP_CHECK(hipGetDeviceProperties(&props, device));
  INFO("info: running on bus " << "0x" << props.pciBusID << " " << props.name << " with "
                               << props.multiProcessorCount << " CUs");

  hA = new float[Nbytes];
  REQUIRE(hA != nullptr);
  hC = new float[Nbytes];
  REQUIRE(hC != nullptr);
  for (size_t i = 0; i < N; i++) {
    hA[i] = CONSTANT + i;
  }

  int cuCountPerGroup = 1;
  if (props.major >= 10) {
    cuCountPerGroup = 2;  // For gfx >= 10, one work group processor encompasses 2 CUs
  }

  unsigned long mask = pow(2, cuCountPerGroup) - 1;
  for (int np = 0; np < KNumPartition; np++) {
    HIP_CHECK(hipMalloc(&dA[np], Nbytes));
    HIP_CHECK(hipMalloc(&dC[np], Nbytes));
    // make unique CU masks in the multiple of dwords for each stream
    uint32_t temp = 0;
    uint32_t bit_index = cuCountPerGroup * np;
    for (int i = np; i < props.multiProcessorCount; i = i + cuCountPerGroup * 4) {
      temp |= mask << bit_index;
      if (bit_index >= 32) {
        cuMasks[np].push_back(temp);
        temp = 0;
        bit_index = cuCountPerGroup * np;
        temp |= mask << bit_index;
      }
      bit_index += cuCountPerGroup * 4;
    }
    if (bit_index != 0) {
      cuMasks[np].push_back(temp);
    }

    HIP_CHECK(hipExtStreamCreateWithCUMask(&streams[np], cuMasks[np].size(), cuMasks[np].data()));

    HIP_CHECK(hipMemcpy(dA[np], hA, Nbytes, hipMemcpyHostToDevice));

    ss[np] << std::hex << std::setfill('0');
    for (int i = cuMasks[np].size() - 1; i >= 0; i--) {
      ss[np] << std::setw(8) << cuMasks[np][i];
    }
  }

  const unsigned threadsPerBlock = 1024;
  const int blocks =
      (N % threadsPerBlock == 0) ? (N / threadsPerBlock) : ((N / threadsPerBlock) + 1);

  auto single_start = std::chrono::steady_clock::now();
  INFO("info: launch 'vector_square' kernel on one stream " << streams[0] << " with CU mask: 0x"
                                                            << ss[0].str().c_str());

  hipLaunchKernelGGL(HipTest::vector_square, dim3(blocks), dim3(threadsPerBlock), 0, streams[0],
                     dA[0], dC[0], N);
  HIP_CHECK(hipGetLastError());
  HIP_CHECK(hipDeviceSynchronize());

  auto single_end = std::chrono::steady_clock::now();
  std::chrono::duration<double> single_kernel_time = single_end - single_start;

  HIP_CHECK(hipMemcpy(hC, dC[0], Nbytes, hipMemcpyDeviceToHost));

  for (size_t i = 0; i < N; i++) {
    REQUIRE(hC[i] == (hA[i] * hA[i]));
  }

  INFO("info: launch 'vector_square' kernel on " << KNumPartition << " streams:");
  auto all_start = std::chrono::steady_clock::now();
  for (int np = 0; np < KNumPartition; np++) {
    INFO("info: launch 'vector_square' kernel on the stream " << streams[np] << " with CU mask: 0x"
                                                              << ss[np].str().c_str());
    hipLaunchKernelGGL(HipTest::vector_square, dim3(blocks), dim3(threadsPerBlock), 0, streams[np],
                       dA[np], dC[np], N);
    HIP_CHECK(hipGetLastError());
  }
  HIP_CHECK(hipDeviceSynchronize());

  auto all_end = std::chrono::steady_clock::now();
  std::chrono::duration<double> all_kernel_time = all_end - all_start;

  for (int np = 0; np < KNumPartition; np++) {
    HIP_CHECK(hipMemcpy(hC, dC[np], Nbytes, hipMemcpyDeviceToHost));
    for (size_t i = 0; i < N; i++) {
      REQUIRE(hC[i] == (hA[i] * hA[i]));
    }
  }

  INFO("info: kernel launched on one stream took: " << single_kernel_time.count() << " seconds");
  INFO("info: kernels launched on " << KNumPartition << " streams took: " << all_kernel_time.count()
                                    << " seconds");
  INFO("info: launching kernels on "
       << KNumPartition << " streams asynchronously is "
       << single_kernel_time.count() / (all_kernel_time.count() / KNumPartition)
       << " times faster per stream than launching on one stream alone");

  delete[] hA;
  delete[] hC;
  for (int np = 0; np < KNumPartition; np++) {
    HIP_CHECK(hipFree(dC[np]));
    HIP_CHECK(hipFree(dA[np]));
    HIP_CHECK(hipStreamDestroy(streams[np]));
  }
}

/**
 * Scenario: Create a stream with all CU masks disabled (0x00000000).
 * Verify that default CU mask is set for the stream.
 */
HIP_TEST_CASE(Unit_hipExtStreamCreateWithCUMask_AllCUsMasked) {
  HIP_CHECK(hipSetDevice(0));
  hipDeviceProp_t props;
  HIP_CHECK(hipGetDeviceProperties(&props, 0));
  // make a CU mask with all CUs disabled.
  std::vector<uint32_t> allCUDisabled;
  createDisabledCUMask(&allCUDisabled, props.multiProcessorCount);
  hipStream_t stream;
  HIP_CHECK(hipExtStreamCreateWithCUMask(&stream, allCUDisabled.size(), allCUDisabled.data()));
  // Verify whether default CU mask is set for the stream.
  uint32_t size = (props.multiProcessorCount / 32) + 1;
  std::vector<uint32_t> cuMask(size);
  std::vector<uint32_t> defaultCUMask;
  createDefaultCUMask(&defaultCUMask, props.multiProcessorCount);
  HIP_CHECK(hipExtStreamGetCUMask(stream, cuMask.size(), &cuMask[0]));
  for (int i = 0; i < static_cast<int>(defaultCUMask.size()); i++) {
    REQUIRE(defaultCUMask[i] == cuMask[i]);
  }
  HIP_CHECK(hipStreamDestroy(stream));
}

/**
 * Scenario: Negative Testing of hipExtStreamCreateWithCUMask.
 */
HIP_TEST_CASE(Unit_hipExtStreamCreateWithCUMask_NegTst) {
  std::vector<uint32_t> defaultCUMask;
  REQUIRE(hipSuccess == hipSetDevice(0));
  hipDeviceProp_t props;
  REQUIRE(hipSuccess == hipGetDeviceProperties(&props, 0));
  createDefaultCUMask(&defaultCUMask, props.multiProcessorCount);
  hipStream_t stream;
  // Negative Scenario 1: stream = nullptr
  SECTION("stream is nullptr") {
    REQUIRE_FALSE(hipExtStreamCreateWithCUMask(nullptr, defaultCUMask.size(),
                                               defaultCUMask.data()) == hipSuccess);
  }
  // Negative Scenario 2: cuMaskSize = 0
  SECTION("cuMaskSize is 0") {
    REQUIRE_FALSE(hipExtStreamCreateWithCUMask(&stream, 0, defaultCUMask.data()) == hipSuccess);
  }
  // Negative Scenario 3: cuMask = nullptr
  SECTION("cuMask is nullptr") {
    REQUIRE_FALSE(hipExtStreamCreateWithCUMask(&stream, defaultCUMask.size(), nullptr) ==
                  hipSuccess);
  }
}