TheRock/rocm-systems/projects/hip-tests/catch/performance/scenarios/compute/hipPerfMandelbrot.cc

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

/**
 * @addtogroup hipPerfMandelbrot hipPerfMandelbrot
 * @{
 * @ingroup PerformanceTestCompute
 */

#include <hip_test_common.hh>
#include <hip/hip_vector_types.h>
#include <hip/math_functions.h>
#include <vector>
#include <string>
#include <map>

typedef struct {
  double x;
  double y;
  double width;
} coordRec;

coordRec coords[] = {
    {0.0, 0.0, 4.0},                                     // Whole set
    {0.0, 0.0, 0.00001},                                 // All black
    {-0.0180789661868, 0.6424294066162, 0.00003824140},  // Hit detail
};

static unsigned int numCoords = sizeof(coords) / sizeof(coordRec);

template <typename T> __global__ void float_mad_kernel(uint* out, uint width, T xPos, T yPos,
                                                       T xStep, T yStep, uint maxIter) {
  int tid = (blockIdx.x * blockDim.x + threadIdx.x);
  int i = tid % width;
  int j = tid / width;
  float x0 = static_cast<float>(xPos + xStep * i);
  float y0 = static_cast<float>(yPos + yStep * j);

  float x = x0;
  float y = y0;

  uint iter = 0;
  float tmp;
  for (iter = 0; (x * x + y * y <= 4.0f) && (iter < maxIter); iter++) {
    tmp = x;
    x = fma(-y, y, fma(x, x, x0));
    y = fma(2.0f * tmp, y, y0);
  }
  out[tid] = iter;
}

template <typename T> __global__ void float_mandel_unroll_kernel(uint* out, uint width, T xPos,
                                                                 T yPos, T xStep, T yStep,
                                                                 uint maxIter) {
  int tid = (blockIdx.x * blockDim.x + threadIdx.x);
  int i = tid % width;
  int j = tid / width;
  float x0 = static_cast<float>(xPos + xStep * static_cast<float>(i));
  float y0 = static_cast<float>(yPos + yStep * static_cast<float>(j));

  float x = x0;
  float y = y0;

#define FAST
  uint iter = 0;
  float tmp;
  int stay;
  int ccount = 0;
  stay = (x * x + y * y) <= 4.0;
  float savx = x;
  float savy = y;
#ifdef FAST
  for (iter = 0; (iter < maxIter); iter += 16) {
#else
  for (iter = 0; stay && (iter < maxIter); iter += 16) {
#endif
    x = savx;
    y = savy;

    // Two iterations
    tmp = fma(-y, y, fma(x, x, x0));
    y = fma(2.0f * x, y, y0);
    x = fma(-y, y, fma(tmp, tmp, x0));
    y = fma(2.0f * tmp, y, y0);

    // Two iterations
    tmp = fma(-y, y, fma(x, x, x0));
    y = fma(2.0f * x, y, y0);
    x = fma(-y, y, fma(tmp, tmp, x0));
    y = fma(2.0f * tmp, y, y0);

    // Two iterations
    tmp = fma(-y, y, fma(x, x, x0));
    y = fma(2.0f * x, y, y0);
    x = fma(-y, y, fma(tmp, tmp, x0));
    y = fma(2.0f * tmp, y, y0);

    // Two iterations
    tmp = fma(-y, y, fma(x, x, x0));
    y = fma(2.0f * x, y, y0);
    x = fma(-y, y, fma(tmp, tmp, x0));
    y = fma(2.0f * tmp, y, y0);

    // Two iterations
    tmp = fma(-y, y, fma(x, x, x0));
    y = fma(2.0f * x, y, y0);
    x = fma(-y, y, fma(tmp, tmp, x0));
    y = fma(2.0f * tmp, y, y0);

    // Two iterations
    tmp = fma(-y, y, fma(x, x, x0));
    y = fma(2.0f * x, y, y0);
    x = fma(-y, y, fma(tmp, tmp, x0));
    y = fma(2.0f * tmp, y, y0);

    // Two iterations
    tmp = fma(-y, y, fma(x, x, x0));
    y = fma(2.0f * x, y, y0);
    x = fma(-y, y, fma(tmp, tmp, x0));
    y = fma(2.0f * tmp, y, y0);

    // Two iterations
    tmp = fma(-y, y, fma(x, x, x0));
    y = fma(2.0f * x, y, y0);
    x = fma(-y, y, fma(tmp, tmp, x0));
    y = fma(2.0f * tmp, y, y0);

    stay = (x * x + y * y) <= 4.0;
    savx = (stay ? x : savx);
    savy = (stay ? y : savy);
    ccount += stay * 16;
#ifdef FAST
    if (!stay) break;
#endif
  }
  // Handle remainder
  if (!stay) {
    iter = 16;
    do {
      x = savx;
      y = savy;
      stay = ((x * x + y * y) <= 4.0) && (ccount < maxIter);
      tmp = x;
      x = fma(-y, y, fma(x, x, x0));
      y = fma(2.0f * tmp, y, y0);
      ccount += stay;
      iter--;
      savx = (stay ? x : savx);
      savy = (stay ? y : savy);
    } while (stay && iter);
  }
  out[tid] = (uint)ccount;
}

template <typename T> __global__ void double_mad_kernel(uint* out, uint width, T xPos, T yPos,
                                                        T xStep, T yStep, uint maxIter) {
  int tid = (blockIdx.x * blockDim.x + threadIdx.x);
  int i = tid % width;
  int j = tid / width;
  double x0 = static_cast<double>(xPos + xStep * i);
  double y0 = static_cast<double>(yPos + yStep * j);

  double x = x0;
  double y = y0;

  uint iter = 0;
  double tmp;
  for (iter = 0; (x * x + y * y <= 4.0f) && (iter < maxIter); iter++) {
    tmp = x;
    x = fma(-y, y, fma(x, x, x0));
    y = fma(2.0f * tmp, y, y0);
  }
  out[tid] = iter;
};

template <typename T> __global__ void double_mandel_unroll_kernel(uint* out, uint width, T xPos,
                                                                  T yPos, T xStep, T yStep,
                                                                  uint maxIter) {
  int tid = (blockIdx.x * blockDim.x + threadIdx.x);

  int i = tid % width;
  int j = tid / width;
  double x0 = static_cast<double>(xPos + xStep * static_cast<double>(i));
  double y0 = static_cast<double>(yPos + yStep * static_cast<double>(j));

  double x = x0;
  double y = y0;

#define FAST
  uint iter = 0;
  double tmp;
  int stay;
  int ccount = 0;
  stay = (x * x + y * y) <= 4.0;
  double savx = x;
  double savy = y;
#ifdef FAST
  for (iter = 0; (iter < maxIter); iter += 16)
#else
  for (iter = 0; stay && (iter < maxIter); iter += 16)
#endif
  {
    x = savx;
    y = savy;

    // Two iterations
    tmp = fma(-y, y, fma(x, x, x0));
    y = fma(2.0f * x, y, y0);
    x = fma(-y, y, fma(tmp, tmp, x0));
    y = fma(2.0f * tmp, y, y0);

    // Two iterations
    tmp = fma(-y, y, fma(x, x, x0));
    y = fma(2.0f * x, y, y0);
    x = fma(-y, y, fma(tmp, tmp, x0));
    y = fma(2.0f * tmp, y, y0);

    // Two iterations
    tmp = fma(-y, y, fma(x, x, x0));
    y = fma(2.0f * x, y, y0);
    x = fma(-y, y, fma(tmp, tmp, x0));
    y = fma(2.0f * tmp, y, y0);

    // Two iterations
    tmp = fma(-y, y, fma(x, x, x0));
    y = fma(2.0f * x, y, y0);
    x = fma(-y, y, fma(tmp, tmp, x0));
    y = fma(2.0f * tmp, y, y0);

    // Two iterations
    tmp = fma(-y, y, fma(x, x, x0));
    y = fma(2.0f * x, y, y0);
    x = fma(-y, y, fma(tmp, tmp, x0));
    y = fma(2.0f * tmp, y, y0);

    // Two iterations
    tmp = fma(-y, y, fma(x, x, x0));
    y = fma(2.0f * x, y, y0);
    x = fma(-y, y, fma(tmp, tmp, x0));
    y = fma(2.0f * tmp, y, y0);

    // Two iterations
    tmp = fma(-y, y, fma(x, x, x0));
    y = fma(2.0f * x, y, y0);
    x = fma(-y, y, fma(tmp, tmp, x0));
    y = fma(2.0f * tmp, y, y0);

    // Two iterations
    tmp = fma(-y, y, fma(x, x, x0));
    y = fma(2.0f * x, y, y0);
    x = fma(-y, y, fma(tmp, tmp, x0));
    y = fma(2.0f * tmp, y, y0);

    stay = (x * x + y * y) <= 4.0;
    savx = (stay ? x : savx);
    savy = (stay ? y : savy);
    ccount += stay * 16;
#ifdef FAST
    if (!stay) break;
#endif
  }
  // Handle remainder
  if (!stay) {
    iter = 16;
    do {
      x = savx;
      y = savy;
      stay = ((x * x + y * y) <= 4.0) && (ccount < maxIter);
      tmp = x;
      x = fma(-y, y, fma(x, x, x0));
      y = fma(2.0f * tmp, y, y0);
      ccount += stay;
      iter--;
      savx = (stay ? x : savx);
      savy = (stay ? y : savy);
    } while (stay && iter);
  }
  out[tid] = (uint)ccount;
};

// Expected results for each kernel run at each coord
unsigned long long expectedIters[] = {
    203277748ull, 2147483648ull, 120254651ull, 203277748ull, 2147483648ull, 120254651ull,
    203277748ull, 2147483648ull, 120254651ull, 203315114ull, 2147483648ull, 120042599ull,
    203315114ull, 2147483648ull, 120042599ull, 203280620ull, 2147483648ull, 120485704ull,
    203280620ull, 2147483648ull, 120485704ull, 203280620ull, 2147483648ull, 120485704ull,
    203315114ull, 2147483648ull, 120042599ull, 203315114ull, 2147483648ull, 120042599ull};

class hipPerfMandelBrot {
 public:
  hipPerfMandelBrot();
  ~hipPerfMandelBrot();

  void setNumKernels(unsigned int num) { numKernels = num; }

  unsigned int getNumKernels() { return numKernels; }

  void setNumStreams(unsigned int num) { numStreams = num; }
  unsigned int getNumStreams() { return numStreams; }

  void open(int deviceID);
  bool run(unsigned int testCase);
  void printResults(void);

  // array of funtion pointers
  typedef void (hipPerfMandelBrot::*funPtr)(uint* out, uint width, float xPos, float yPos,
                                            float xStep, float yStep, uint maxIter,
                                            hipStream_t* streams, int blocks, int threads_per_block,
                                            int kernelCnt);

  // Wrappers
  void float_mad(uint* out, uint width, float xPos, float yPos, float xStep, float yStep,
                 uint maxIter, hipStream_t* streams, int blocks, int threads_per_block,
                 int kernelCnt);

  void float_mandel_unroll(uint* out, uint width, float xPos, float yPos, float xStep, float yStep,
                           uint maxIter, hipStream_t* streams, int blocks, int threads_per_block,
                           int kernelCnt);

  void double_mad(uint* out, uint width, float xPos, float yPos, float xStep, float yStep,
                  uint maxIter, hipStream_t* streams, int blocks, int threads_per_block,
                  int kernelCnt);

  void double_mandel_unroll(uint* out, uint width, float xPos, float yPos, float xStep, float yStep,
                            uint maxIter, hipStream_t* streams, int blocks, int threads_per_block,
                            int kernelCnt);

  hipStream_t streams[2];

 private:
  void setData(void* ptr, unsigned int value);
  void checkData(uint* ptr);

  unsigned int numKernels;
  unsigned int numStreams;

  std::map<std::string, std::vector<double>> results;
  unsigned int width_;
  unsigned int bufSize;
  unsigned int maxIter;
  unsigned int coordIdx;
  volatile unsigned long long totalIters = 0;
  int numCUs;
  static const unsigned int numLoops = 10;
};

hipPerfMandelBrot::hipPerfMandelBrot() {}

hipPerfMandelBrot::~hipPerfMandelBrot() {}

void hipPerfMandelBrot::open(int deviceId) {
  int nGpu = 0;
  HIP_CHECK(hipGetDeviceCount(&nGpu));
  if (nGpu < 1) {
    HIP_SKIP_TEST(HipTest::SkipReason::kNoGpuDevice);
  }
  HIP_CHECK(hipSetDevice(deviceId));
  hipDeviceProp_t props;
  HIP_CHECK(hipGetDeviceProperties(&props, deviceId));

  CONSOLE_PRINT("info: running on bus 0x%x %s with %d CUs and device id: %d\n", props.pciBusID,
                props.name, props.multiProcessorCount, deviceId);

  numCUs = props.multiProcessorCount;
}

void hipPerfMandelBrot::printResults() {
  int numStreams = getNumStreams();

  CONSOLE_PRINT("Measured perf for kernels in GFLOPS on %d streams (s)", numStreams);

  std::map<std::string, std::vector<double>>::iterator itr;
  for (itr = results.begin(); itr != results.end(); itr++) {
    CONSOLE_PRINT("\n%s ", itr->first.c_str());
    for (auto i : results[itr->first]) {
      CONSOLE_PRINT("%10f ", i);
    }
  }
  results.clear();
  CONSOLE_PRINT("\n");
}

// Wrappers for the kernel launches
void hipPerfMandelBrot::float_mad(uint* out, uint width, float xPos, float yPos, float xStep,
                                  float yStep, uint maxIter, hipStream_t* streams, int blocks,
                                  int threads_per_block, int kernelCnt) {
  int streamCnt = getNumStreams();
  hipLaunchKernelGGL(float_mad_kernel<float>, dim3(blocks), dim3(threads_per_block), 0,
                     streams[kernelCnt % streamCnt], out, width, xPos, yPos, xStep, yStep, maxIter);
}

void hipPerfMandelBrot::float_mandel_unroll(uint* out, uint width, float xPos, float yPos,
                                            float xStep, float yStep, uint maxIter,
                                            hipStream_t* streams, int blocks, int threads_per_block,
                                            int kernelCnt) {
  int streamCnt = getNumStreams();
  hipLaunchKernelGGL(float_mandel_unroll_kernel<float>, dim3(blocks), dim3(threads_per_block), 0,
                     streams[kernelCnt % streamCnt], out, width, xPos, yPos, xStep, yStep, maxIter);
}

void hipPerfMandelBrot::double_mad(uint* out, uint width, float xPos, float yPos, float xStep,
                                   float yStep, uint maxIter, hipStream_t* streams, int blocks,
                                   int threads_per_block, int kernelCnt) {
  int streamCnt = getNumStreams();
  hipLaunchKernelGGL(double_mad_kernel<double>, dim3(blocks), dim3(threads_per_block), 0,
                     streams[kernelCnt % streamCnt], out, width, xPos, yPos, xStep, yStep, maxIter);
}

void hipPerfMandelBrot::double_mandel_unroll(uint* out, uint width, float xPos, float yPos,
                                             float xStep, float yStep, uint maxIter,
                                             hipStream_t* streams, int blocks,
                                             int threads_per_block, int kernelCnt) {
  int streamCnt = getNumStreams();
  hipLaunchKernelGGL(float_mandel_unroll_kernel<double>, dim3(blocks), dim3(threads_per_block), 0,
                     streams[kernelCnt % streamCnt], out, width, xPos, yPos, xStep, yStep, maxIter);
}

bool hipPerfMandelBrot::run(unsigned int testCase) {
  unsigned int numStreams = getNumStreams();
  coordIdx = testCase % numCoords;

  funPtr p[] = {&hipPerfMandelBrot::float_mad, &hipPerfMandelBrot::float_mandel_unroll,
                &hipPerfMandelBrot::double_mad, &hipPerfMandelBrot::double_mandel_unroll};

  // Maximum iteration count
  maxIter = 32768;

  uint** hPtr = new uint*[numKernels];
  uint** dPtr = new uint*[numKernels];

  // Width is divisible by 4 because the mandelbrot kernel processes 4 pixels at once.
  width_ = 256;

  bufSize = width_ * width_ * sizeof(uint);

  // Create streams for concurrency
  for (uint i = 0; i < numStreams; i++) {
    HIP_CHECK(hipStreamCreate(&streams[i]));
  }

  // Allocate memory on the host and device
  for (uint i = 0; i < numKernels; i++) {
    HIP_CHECK(hipHostMalloc(reinterpret_cast<void**>(&hPtr[i]), bufSize, hipHostMallocDefault));
    setData(hPtr[i], 0xdeadbeef);
    HIP_CHECK(hipMalloc(reinterpret_cast<uint**>(&dPtr[i]), bufSize))
  }

  // Prepare kernel launch parameters
  int threads = (bufSize / sizeof(uint));
  int threads_per_block = 64;
  int blocks = (threads / threads_per_block) + (threads % threads_per_block);

  // Copy memory asynchronously and concurrently from host to device
  for (uint i = 0; i < numKernels; i++) {
    HIP_CHECK(hipMemcpy(dPtr[i], hPtr[i], bufSize, hipMemcpyHostToDevice));
  }

  // Synchronize to make sure all the copies are completed
  HIP_CHECK(hipStreamSynchronize(0));

  int kernelIdx;
  if (testCase == 0 || testCase == 5 || testCase == 10) {
    kernelIdx = 0;
  } else if (testCase == 1 || testCase == 6 || testCase == 11) {
    kernelIdx = 1;
  } else if (testCase == 2 || testCase == 7 || testCase == 12) {
    kernelIdx = 2;
  } else if (testCase == 3 || testCase == 8 || testCase == 13) {
    kernelIdx = 3;
  }
  double totalTime = 0.0;
  for (unsigned int k = 0; k < numLoops; k++) {
    if ((testCase == 0 || testCase == 1 || testCase == 2 || testCase == 5 || testCase == 6 ||
         testCase == 7 || testCase == 10 || testCase == 11 || testCase == 12)) {
      float xStep = static_cast<float>(coords[coordIdx].width / static_cast<double>(width_));
      float yStep = static_cast<float>(-coords[coordIdx].width / static_cast<double>(width_));
      float xPos = static_cast<float>(coords[coordIdx].x - 0.5 * coords[coordIdx].width);
      float yPos = static_cast<float>(coords[coordIdx].y + 0.5 * coords[coordIdx].width);

      // Time the kernel execution
      auto all_start = std::chrono::steady_clock::now();

      for (uint i = 0; i < numKernels; i++) {
        (this->*p[kernelIdx])(dPtr[i], width_, xPos, yPos, xStep, yStep, maxIter, streams, blocks,
                              threads_per_block, i);
      }

      // Synchronize all the concurrent streams to have completed execution
      HIP_CHECK(hipStreamSynchronize(0));

      auto all_end = std::chrono::steady_clock::now();
      std::chrono::duration<double> all_kernel_time = all_end - all_start;
      totalTime += all_kernel_time.count();
    } else {
      double xStep = coords[coordIdx].width / static_cast<double>(width_);
      double yStep = -coords[coordIdx].width / static_cast<double>(width_);
      double xPos = coords[coordIdx].x - 0.5 * coords[coordIdx].width;
      double yPos = coords[coordIdx].y + 0.5 * coords[coordIdx].width;

      // Time the kernel execution
      auto all_start = std::chrono::steady_clock::now();
      for (uint i = 0; i < numKernels; i++) {
        (this->*p[kernelIdx])(dPtr[i], width_, xPos, yPos, xStep, yStep, maxIter, streams, blocks,
                              threads_per_block, i);
      }
      // Synchronize all the concurrent streams to have completed execution
      HIP_CHECK(hipStreamSynchronize(0));

      auto all_end = std::chrono::steady_clock::now();
      std::chrono::duration<double> all_kernel_time = all_end - all_start;
      totalTime += all_kernel_time.count();
    }
  }

  // Copy data back from device to the host
  for (uint i = 0; i < numKernels; i++) {
    HIP_CHECK(hipMemcpy(hPtr[i], dPtr[i], bufSize, hipMemcpyDeviceToHost));
  }
  for (uint i = 0; i < numKernels; i++) {
    checkData(hPtr[i]);
    int j = 0;
    while ((totalIters != expectedIters[j] && totalIters > expectedIters[j]) && j < 30) {
      j++;
    }

    if (j == 30) {
      CONSOLE_PRINT("Incorrect iteration count detected. ");
    }
  }

  // Compute GFLOPS.  There are 7 FLOPs per iteration
  double perf = (static_cast<double>(totalIters * numKernels) * 7 * static_cast<double>(1e-09)) /
                (totalTime / (double)numLoops);


  std::vector<std::string> kernelName = {"float", "float_unroll", "double", "double_unroll"};

  // Print results except for Warm-up kernel
  if (testCase != 100) {
    results[kernelName[testCase % 4]].push_back(perf);
  }

  for (uint i = 0; i < numStreams; i++) {
    HIP_CHECK(hipStreamDestroy(streams[i]));
  }

  // Free host and device memory
  for (uint i = 0; i < numKernels; i++) {
    HIP_CHECK(hipHostFree(hPtr[i]));
    HIP_CHECK(hipFree(dPtr[i]));
  }
  delete[] hPtr;
  delete[] dPtr;
  return true;
}

void hipPerfMandelBrot::setData(void* ptr, unsigned int value) {
  unsigned int* ptr2 = (unsigned int*)ptr;
  for (unsigned int i = 0; i < width_ * width_; i++) {
    ptr2[i] = value;
  }
}

void hipPerfMandelBrot::checkData(uint* ptr) {
  totalIters = 0;
  for (unsigned int i = 0; i < width_ * width_; i++) {
    totalIters += ptr[i];
  }
}

/**
 * Test Description
 * ------------------------
 *  - Verify the warm-up kernel default stream executes serially.
 *  - verify by running all kernels - sync.
 *  - verify by running all kernels - async.
 * Test source
 * ------------------------
 * - performance/scenarios/compute/hipPerfMandelbrot.cc
 * Test requirements
 * ------------------------
 *  - HIP_VERSION >= 5.6
 */

HIP_TEST_CASE(Performance_hipPerfMandelbrot) {
  hipPerfMandelBrot mandelbrotCompute;
  int deviceId = 0;
  mandelbrotCompute.open(deviceId);
#if HT_AMD
  SECTION("warm-up kernel default stream executes serially") {
    mandelbrotCompute.setNumStreams(1);
    mandelbrotCompute.setNumKernels(1);
    REQUIRE(true == mandelbrotCompute.run(100 /*Random number*/));
  }
#endif
  SECTION("run all - sync") {
    int i = 0;
    do {
      mandelbrotCompute.setNumStreams(1);
      mandelbrotCompute.setNumKernels(1);
      REQUIRE(true == mandelbrotCompute.run(i));
      i++;
    } while (i < 12);
    mandelbrotCompute.printResults();
  }

  SECTION("run all - async") {
    int i = 0;
    do {
      mandelbrotCompute.setNumStreams(2);
      mandelbrotCompute.setNumKernels(2);
      REQUIRE(true == mandelbrotCompute.run(i));
      i++;
    } while (i < 12);
    mandelbrotCompute.printResults();
  }
}

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