TheRock/rocm-systems/projects/clr/opencl/tests/ocltst/module/perf/OCLPerfDoubleDMA.cpp

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

#include "OCLPerfDoubleDMA.h"

#include <Timer.h>
#include <assert.h>
#include <stdio.h>

#include <cmath>
#include <sstream>
#include <string>

#include "CL/cl.h"
#include "CL/cl_ext.h"

const size_t blockX = 256;
const size_t blockY = 256;
const size_t blockZ = 512;

const size_t chunk = 16;
const size_t size_S = blockX * blockY * blockZ * sizeof(cl_float4);
const size_t size_s = blockX * blockY * chunk * sizeof(cl_float4);
static const int WindowWidth = 80;

const size_t MaxQueues = 3;
bool profEnable = false;

static const char* strKernel =
    "__kernel void dummy(__global float4* out)  \n"
    "{                                          \n"
    "   uint id = get_global_id(0);             \n"
    "   float4 value = (float4)(1.0f, 2.0f, 3.0f, 4.0f);  \n"
    "   uint factorial = 1;                     \n"
    "   for (uint i = 1; i < (id / 0x400); ++i)\n"
    "   {                                       \n"
    "       factorial *= i;                     \n"
    "   }                                       \n"
    "   out[id] = value * factorial;            \n"
    "}                                          \n";

class ProfileQueue {
 public:
  enum Operation { Write = 0, Execute, Read, Total };

  static const char* OperationName[Total];
  static const char StartCommand[Total];
  static const char ExecCommand[Total];

  ProfileQueue() {}
  ~ProfileQueue() {
    for (size_t op = 0; op < Total; ++op) {
      for (size_t idx = 0; idx < events_[op].size(); ++idx) {
        clReleaseEvent(events_[op][idx]);
      }
    }
  }

  void addEvent(Operation op, cl_event event) { events_[op].push_back(event); }

  void findMinMax(cl_long* min, cl_long* max) {
    // Find time min/max ranges for the frame scaling
    for (size_t op = 0; (op < ProfileQueue::Total); ++op) {
      cl_long time;
      if (events_[op].size() == 0) continue;
      clGetEventProfilingInfo(events_[op][0], CL_PROFILING_COMMAND_START, sizeof(cl_long), &time,
                              NULL);
      if (0 == *min) {
        *min = time;
      } else {
        *min = std::min(*min, time);
      }
      clGetEventProfilingInfo(events_[op][events_[op].size() - 1], CL_PROFILING_COMMAND_END,
                              sizeof(cl_long), &time, NULL);
      if (0 == *max) {
        *max = time;
      } else {
        *max = std::max(*max, time);
      }
    }
  }

  void display(cl_long start, cl_long finish) {
    std::string graph;
    graph.resize(WindowWidth + 1);
    graph[WindowWidth] = '\x0';
    cl_long timeFrame = finish - start;
    cl_long interval = timeFrame / WindowWidth;

    // Find time min/max ranges for the frame scaling
    for (size_t op = 0; (op < Total); ++op) {
      if (events_[op].size() == 0) continue;
      cl_long timeStart, timeEnd;
      int begin = 0, end = 0;
      for (size_t idx = 0; idx < events_[op].size(); ++idx) {
        bool cutStart = false;
        clGetEventProfilingInfo(events_[op][idx], CL_PROFILING_COMMAND_START, sizeof(cl_long),
                                &timeStart, NULL);
        clGetEventProfilingInfo(events_[op][idx], CL_PROFILING_COMMAND_END, sizeof(cl_long),
                                &timeEnd, NULL);

        // Continue if out of the frame scope
        if (timeStart >= finish) continue;
        if (timeEnd <= start) continue;

        if (timeStart <= start) {
          timeStart = start;
          cutStart = true;
        }

        if (timeEnd >= finish) {
          timeEnd = finish;
        }

        // Readjust time to the frame
        timeStart -= start;
        timeEnd -= start;
        timeStart = static_cast<cl_long>(floor(static_cast<float>(timeStart) / interval + 0.5f));
        timeEnd = static_cast<cl_long>(floor(static_cast<float>(timeEnd) / interval + 0.5f));
        begin = static_cast<int>(timeStart);
        // Idle from end to begin
        for (int c = end; c < begin; ++c) {
          graph[c] = '-';
        }
        end = static_cast<int>(timeEnd);
        for (int c = begin; c < end; ++c) {
          if ((c == begin) && !cutStart) {
            graph[c] = StartCommand[op];
          } else {
            graph[c] = ExecCommand[op];
          }
        }
        if ((begin == end) && (end < WindowWidth)) {
          graph[begin] = '+';
        }
      }
      if (end < WindowWidth) {
        for (int c = end; c < WindowWidth; ++c) {
          graph[c] = '-';
        }
      }
      printf("%s\n", graph.c_str());
    }
  }

 private:
  // Profiling events
  std::vector<cl_event> events_[Total];
};

const char* ProfileQueue::OperationName[Total] = {"BufferWrite", "KernelExecution", "BufferRead"};
const char ProfileQueue::StartCommand[Total] = {'W', 'X', 'R'};
const char ProfileQueue::ExecCommand[Total] = {'>', '#', '<'};

class Profile {
 public:
  Profile(bool profEna, int numQueues)
      : profileEna_(profEna), numQueues_(numQueues), min_(0), max_(0), execTime_(0) {}

  ~Profile() {}

  void addEvent(int queue, ProfileQueue::Operation op, cl_event event) {
    if (profileEna_) {
      profQueue[queue].addEvent(op, event);
    }
  }

  cl_long findExecTime() {
    if (execTime_ != 0) return execTime_;
    for (int q = 0; q < numQueues_; ++q) {
      profQueue[q].findMinMax(&min_, &max_);
    }
    execTime_ = max_ - min_;
    return execTime_;
  }

  void display(cl_long start, cl_long finish) {
    if (!profileEna_) return;
    printf("\n ----------- Time frame %.3f (us), scale 1:%.0f\n", (float)(finish - start) / 1000,
           (float)(finish - start) / (1000 * WindowWidth));
    for (size_t op = 0; (op < ProfileQueue::Total); ++op) {
      printf("%s - %c%c; ", ProfileQueue::OperationName[op], ProfileQueue::StartCommand[op],
             ProfileQueue::ExecCommand[op]);
    }
    printf("\n");
    for (int q = 0; q < numQueues_; ++q) {
      printf("CommandQueue #%d\n", q);
      profQueue[q].display(min_ + start, min_ + finish);
    }
  }

 private:
  bool profileEna_;
  int numQueues_;     //!< Total number of queues
  cl_long min_;       //!< Min HW timestamp
  cl_long max_;       //!< Max HW timestamp
  cl_long execTime_;  //!< Profile time
  ProfileQueue profQueue[MaxQueues];
};

OCLPerfDoubleDMA::OCLPerfDoubleDMA() {
  _numSubTests = 2 * MaxQueues * 2;
  failed_ = false;
}

OCLPerfDoubleDMA::~OCLPerfDoubleDMA() {}

void OCLPerfDoubleDMA::open(unsigned int test, char* units, double& conversion,
                            unsigned int deviceId) {
  _deviceId = deviceId;
  OCLTestImp::open(test, units, conversion, deviceId);
  CHECK_RESULT((error_ != CL_SUCCESS), "Error opening test");
  test_ = test;
  cl_device_type deviceType;
  error_ = _wrapper->clGetDeviceInfo(devices_[deviceId], CL_DEVICE_TYPE, sizeof(deviceType),
                                     &deviceType, NULL);
  CHECK_RESULT((error_ != CL_SUCCESS), "CL_DEVICE_TYPE failed");

  if (!(deviceType & CL_DEVICE_TYPE_GPU)) {
    printf("GPU device is required for this test!\n");
    failed_ = true;
    return;
  }
  program_ = _wrapper->clCreateProgramWithSource(context_, 1, &strKernel, NULL, &error_);
  CHECK_RESULT((error_ != CL_SUCCESS), "clCreateProgramWithSource()  failed");
  error_ = _wrapper->clBuildProgram(program_, 1, &devices_[deviceId], NULL, NULL, NULL);
  if (error_ != CL_SUCCESS) {
    char programLog[1024];
    _wrapper->clGetProgramBuildInfo(program_, devices_[deviceId], CL_PROGRAM_BUILD_LOG, 1024,
                                    programLog, 0);
    printf("\n%s\n", programLog);
    fflush(stdout);
  }
  CHECK_RESULT((error_ != CL_SUCCESS), "clBuildProgram() failed");
  kernel_ = _wrapper->clCreateKernel(program_, "dummy", &error_);
  CHECK_RESULT((error_ != CL_SUCCESS), "clCreateKernel() failed");

  size_t bufSize = size_s;
  cl_mem buffer;
  if (test_ >= (2 * MaxQueues)) {
    profEnable = true;
  }
  test_ %= 2 * MaxQueues;
  size_t numBufs = (test_ % MaxQueues) + 1;
  for (size_t b = 0; b < numBufs; ++b) {
    buffer = _wrapper->clCreateBuffer(context_, CL_MEM_READ_WRITE, bufSize, NULL, &error_);
    CHECK_RESULT((error_ != CL_SUCCESS), "clCreateBuffer() failed");
    buffers_.push_back(buffer);
  }

  buffer = _wrapper->clCreateBuffer(context_, CL_MEM_READ_WRITE | CL_MEM_ALLOC_HOST_PTR, size_S,
                                    NULL, &error_);
  CHECK_RESULT((error_ != CL_SUCCESS), "clCreateBuffer() failed");
  buffers_.push_back(buffer);
}

static void CL_CALLBACK notify_callback(const char* errinfo, const void* private_info, size_t cb,
                                        void* user_data) {}

void OCLPerfDoubleDMA::run(void) {
  if (failed_) {
    return;
  }
  CPerfCounter timer;
  const int numQueues = (test_ % MaxQueues) + 1;
  const bool useKernel = ((test_ / MaxQueues) > 0);
  const int numBufs = numQueues;
  Profile profile(profEnable, numQueues);

  std::vector<cl_command_queue> cmdQueues(numQueues);
  int q;
  cl_command_queue_properties qProp = (profEnable) ? CL_QUEUE_PROFILING_ENABLE : 0;
  for (q = 0; q < numQueues; ++q) {
    cl_command_queue cmdQueue =
        _wrapper->clCreateCommandQueue(context_, devices_[_deviceId], qProp, &error_);
    CHECK_RESULT((error_ != CL_SUCCESS), "clCreateCommandQueue() failed");
    cmdQueues[q] = cmdQueue;
  }

  float* Data_s = (float*)_wrapper->clEnqueueMapBuffer(cmdQueues[0], buffers_[numBufs], CL_TRUE,
                                                       CL_MAP_READ | CL_MAP_WRITE, 0, size_S, 0,
                                                       NULL, NULL, &error_);

  size_t gws[1] = {size_s / (4 * sizeof(float))};
  size_t lws[1] = {256};

  // Warm-up
  for (q = 0; q < numQueues; ++q) {
    error_ |= _wrapper->clEnqueueWriteBuffer(cmdQueues[q], buffers_[q], CL_FALSE, 0, size_s,
                                             (char*)Data_s, 0, NULL, NULL);
    error_ |= _wrapper->clSetKernelArg(kernel_, 0, sizeof(cl_mem), (void*)&buffers_[q]);
    error_ |=
        _wrapper->clEnqueueNDRangeKernel(cmdQueues[q], kernel_, 1, NULL, gws, lws, 0, NULL, NULL);
    error_ |= _wrapper->clEnqueueReadBuffer(cmdQueues[q], buffers_[q], CL_FALSE, 0, size_s,
                                            (char*)Data_s, 0, NULL, NULL);
    error_ |= _wrapper->clFinish(cmdQueues[q]);
  }

  CHECK_RESULT_NO_RETURN((error_ != CL_SUCCESS), "Execution failed");

  size_t s_done = 0;
  cl_event r[MaxQueues] = {0}, w[MaxQueues] = {0}, x[MaxQueues] = {0};

  /*----------  pass2:  copy Data_s to and from GPU Buffers ----------*/
  s_done = 0;
  timer.Reset();
  timer.Start();
  int idx = numBufs - 1;
  // Start from the last so read/write won't go to the same DMA when kernel is
  // executed
  q = numQueues - 1;
  size_t iter = 0;
  while (1) {
    if (0 == r[idx]) {
      error_ |= _wrapper->clEnqueueWriteBuffer(cmdQueues[q], buffers_[idx], CL_FALSE, 0, size_s,
                                               (char*)Data_s + s_done, 0, NULL, &w[idx]);
    } else {
      error_ |= _wrapper->clEnqueueWriteBuffer(cmdQueues[q], buffers_[idx], CL_FALSE, 0, size_s,
                                               (char*)Data_s + s_done, 1, &r[idx], &w[idx]);
      if (!profEnable) {
        error_ |= _wrapper->clReleaseEvent(r[idx]);
      }
    }
    _wrapper->clFlush(cmdQueues[q]);
    profile.addEvent(q, ProfileQueue::Write, w[idx]);

    if (useKernel) {
      // Change the queue
      ++q %= numQueues;
      // Implicit flush of DMA engine on kernel start, because memory dependency
      error_ |= _wrapper->clSetKernelArg(kernel_, 0, sizeof(cl_mem), (void*)&buffers_[idx]);
      error_ |= _wrapper->clEnqueueNDRangeKernel(cmdQueues[q], kernel_, 1, NULL, gws, lws, 1,
                                                 &w[idx], &x[idx]);
      if (!profEnable) {
        error_ |= _wrapper->clReleaseEvent(w[idx]);
      }
      profile.addEvent(q, ProfileQueue::Execute, x[idx]);
    }
    _wrapper->clFlush(cmdQueues[q]);

    // Change the queue
    ++q %= numQueues;
    error_ |= _wrapper->clEnqueueReadBuffer(cmdQueues[q], buffers_[idx], CL_FALSE, 0, size_s,
                                            (char*)Data_s + s_done, 1,
                                            (useKernel) ? &x[idx] : &w[idx], &r[idx]);
    if (!profEnable) {
      error_ |= _wrapper->clReleaseEvent((useKernel) ? x[idx] : w[idx]);
    }
    profile.addEvent(q, ProfileQueue::Read, r[idx]);
    _wrapper->clFlush(cmdQueues[q]);

    if ((s_done += size_s) >= size_S) {
      if (!profEnable) {
        error_ |= _wrapper->clReleaseEvent(r[idx]);
      }
      break;
    }
    ++iter;
    ++idx %= numBufs;
    ++q %= numQueues;
  }

  for (q = 0; q < numQueues; ++q) {
    error_ |= _wrapper->clFinish(cmdQueues[q]);
  }
  timer.Stop();

  error_ =
      _wrapper->clEnqueueUnmapMemObject(cmdQueues[0], buffers_[numBufs], Data_s, 0, NULL, NULL);

  error_ |= _wrapper->clFinish(cmdQueues[0]);
  CHECK_RESULT_NO_RETURN((error_ != CL_SUCCESS), "Execution failed");

  cl_long gpuTimeFrame = profile.findExecTime();
  cl_long oneIter = gpuTimeFrame / iter;

  // Display 4 iterations in the middle
  cl_long startFrame = oneIter * (iter / 2 - 2);
  cl_long finishFrame = oneIter * (iter / 2 + 2);
  profile.display(startFrame, finishFrame);

  for (q = 0; q < numQueues; ++q) {
    error_ = _wrapper->clReleaseCommandQueue(cmdQueues[q]);
    CHECK_RESULT_NO_RETURN((error_ != CL_SUCCESS), "clReleaseCommandQueue() failed");
  }

  double GBytes = (double)(2 * size_S) / (double)(1000 * 1000 * 1000);
  _perfInfo = static_cast<float>(GBytes / timer.GetElapsedTime());

  std::stringstream stream;
  if (useKernel) {
    stream << "Write/Kernel/Read operation ";
  } else {
    stream << "Write/Read operation ";
  }
  stream << numQueues << " queues; profiling " << ((profEnable) ? "enabled" : "disabled")
         << " [GB/s]";

  stream.flags(std::ios::right | std::ios::showbase);
  testDescString = stream.str();
}

unsigned int OCLPerfDoubleDMA::close(void) { return OCLTestImp::close(); }