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

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

#include "OCLPerfSVMAlloc.h"

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

#include <sstream>
#include <string>

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

// Quiet pesky warnings
#ifdef WIN_OS
#define SNPRINTF sprintf_s
#else
#define SNPRINTF snprintf
#endif

#define NUM_SIZES 5
#define NUM_CG_FLAGS 3
#define NUM_FG_FLAGS 3

static size_t sizeList[NUM_SIZES] = {
    0x040000, 0x080000, 0x100000, 0x200000, 0x400000,
};

#if defined(CL_VERSION_2_0)
static const cl_svm_mem_flags CGFlags[NUM_CG_FLAGS] = {
    CL_MEM_READ_WRITE,
    CL_MEM_WRITE_ONLY,
    CL_MEM_READ_ONLY,
};
static const cl_svm_mem_flags FGFlags[NUM_FG_FLAGS] = {
    0,
    CL_MEM_SVM_FINE_GRAIN_BUFFER,
    CL_MEM_SVM_FINE_GRAIN_BUFFER | CL_MEM_SVM_ATOMICS,
};
#endif

static const char* strKernel =
    "__kernel void dummy(__global uint* out)    \n"
    "{                                          \n"
    "   uint id = get_global_id(0);             \n"
    "   uint value = 1;                         \n"
    "   if ((int)get_local_id(0) < 0)           \n"
    "       out[id] = value;                    \n"
    "}                                          \n";

OCLPerfSVMAlloc::OCLPerfSVMAlloc() {
  _numSubTests = NUM_CG_FLAGS * NUM_FG_FLAGS * NUM_SIZES + NUM_SIZES;
  failed_ = false;
  skip_ = false;
}

OCLPerfSVMAlloc::~OCLPerfSVMAlloc() {}

void OCLPerfSVMAlloc::open(unsigned int test, char* units, double& conversion,
                           unsigned int deviceId) {
  OCLTestImp::open(test, units, conversion, deviceId);
  CHECK_RESULT((error_ != CL_SUCCESS), "Error opening test");

#if defined(CL_VERSION_2_0)
  FGSystem_ = (test >= (NUM_CG_FLAGS * NUM_FG_FLAGS * NUM_SIZES));
  testFGFlag_ = (test / (NUM_SIZES * NUM_CG_FLAGS)) % NUM_FG_FLAGS;
  testCGFlag_ = (test / NUM_SIZES) % NUM_CG_FLAGS;
  testSize_ = test % NUM_SIZES;

  cl_device_svm_capabilities caps;
  error_ = clGetDeviceInfo(devices_[deviceId], CL_DEVICE_SVM_CAPABILITIES,
                           sizeof(cl_device_svm_capabilities), &caps, NULL);
  CHECK_RESULT(error_ != CL_SUCCESS, "clGetDeviceInfo failed");
  if ((caps & CL_DEVICE_SVM_COARSE_GRAIN_BUFFER) == 0) {
    skip_ = true;  // Should never happen as OCL 2.0 devices are required to
                   // support coarse grain SVM
    testDescString = "Coarse Grain Buffer  NOT supported. Test Skipped.";
    return;
  } else if (testFGFlag_ > 0 && (caps & CL_DEVICE_SVM_FINE_GRAIN_BUFFER) == 0) {
    skip_ = true;  // No support for fine grain buffer SVM
    testDescString = "Fine Grain Buffer NOT supported. Test Skipped.";
    return;
  } else if (FGSystem_ && (caps & CL_DEVICE_SVM_FINE_GRAIN_SYSTEM) == 0) {
    skip_ = true;  // No support for fine grain system SVM
    testDescString = "Fine Grain System NOT supported. Test Skipped.";
    return;
  } else if (testFGFlag_ == 2 && (caps & CL_DEVICE_SVM_ATOMICS) == 0) {
    skip_ = true;  // No support for fine grain system SVM
    testDescString = "SVM Atomic        NOT supported. Test Skipped.";
    return;
  }

  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], "-cl-std=CL2.0", 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");
  return;
#else
  skip_ = true;
  testDescString = "SVM NOT supported for < 2.0 builds. Test Skipped.";
  return;
#endif
}

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

void OCLPerfSVMAlloc::run(void) {
  if (skip_) {
    return;
  }

  if (failed_) {
    return;
  }
#if defined(CL_VERSION_2_0)
  cl_uint* buffer = NULL;
  CPerfCounter timer;
  void* hostPtr = NULL;

  size_t bufSize = sizeList[testSize_] * sizeof(cl_int4);
  size_t iter = 100;

  cl_mem_flags flags = CGFlags[testCGFlag_] | FGFlags[testFGFlag_];

  timer.Reset();
  timer.Start();

  size_t gws[1] = {bufSize / sizeof(cl_int4)};
  size_t lws[1] = {64};

  for (size_t i = 0; i < iter; ++i) {
    if (!FGSystem_) {
      buffer = (cl_uint*)clSVMAlloc(context_, flags, bufSize, 0);
    } else {
      buffer = (cl_uint*)malloc(bufSize);
    }
    CHECK_RESULT(buffer == 0, "Allocation failed");

    error_ = _wrapper->clSetKernelArgSVMPointer(kernel_, 0, buffer);
    CHECK_RESULT((error_ != CL_SUCCESS), "clSetKernelArg() failed");

    error_ = _wrapper->clEnqueueNDRangeKernel(cmdQueues_[_deviceId], kernel_, 1, NULL, gws, lws, 0,
                                              NULL, NULL);
    CHECK_RESULT((error_ != CL_SUCCESS), "clEnqueueNDRangeKernel() failed");

    _wrapper->clFinish(cmdQueues_[_deviceId]);

    if (!FGSystem_) {
      clSVMFree(context_, (void*)buffer);
    } else {
      free(buffer);
    }
  }

  timer.Stop();

  CPerfCounter timer2;
  timer2.Reset();
  size_t numN = 100;

  if (!FGSystem_) {
    buffer = (cl_uint*)clSVMAlloc(context_, flags, bufSize, 0);
  } else {
    buffer = (cl_uint*)malloc(bufSize);
  }
  CHECK_RESULT(buffer == 0, "Allocation failed");

  timer2.Start();
  for (size_t i = 0; i < numN; ++i) {
    error_ = _wrapper->clSetKernelArgSVMPointer(kernel_, 0, buffer);
    CHECK_RESULT((error_ != CL_SUCCESS), "clSetKernelArg() failed");

    error_ = _wrapper->clEnqueueNDRangeKernel(cmdQueues_[_deviceId], kernel_, 1, NULL, gws, lws, 0,
                                              NULL, NULL);
    CHECK_RESULT((error_ != CL_SUCCESS), "clEnqueueNDRangeKernel() failed");
  }
  _wrapper->clFinish(cmdQueues_[_deviceId]);
  timer2.Stop();

  if (!FGSystem_) {
    clSVMFree(context_, (void*)buffer);
  } else {
    free(buffer);
  }

  char pFlags[5];
  pFlags[0] = (testCGFlag_ == 0 || testCGFlag_ == 2) ? 'R' : '_';  // CL_MEM_READ_ONLY
  pFlags[1] = (testCGFlag_ == 0 || testCGFlag_ == 1) ? 'W' : '_';  // CL_MEM_WRITE_ONLY
  pFlags[2] = (testFGFlag_ == 1 || testFGFlag_ == 2) ? 'F' : '_';  // CL_MEM_SVM_FINE_GRAIN_BUFFER
  pFlags[3] = (testFGFlag_ == 2) ? 'A' : '_';                      // CL_MEM_SVM_ATOMICS

  char buf[256];

  if (!FGSystem_ && (testFGFlag_ == 0)) {
    SNPRINTF(buf, sizeof(buf), "Coarse Grain Buffer Alloc + Free (GB/s) for %6d KB, flags=%4s",
             (int)bufSize / 1024, pFlags);
  } else if (!FGSystem_ && (testFGFlag_ > 0)) {
    SNPRINTF(buf, sizeof(buf), "Fine Grain Buffer   Alloc + Free (GB/s) for %6d KB, flags=%4s",
             (int)bufSize / 1024, pFlags);
  } else if (FGSystem_) {
    SNPRINTF(buf, sizeof(buf), "Fine Grain System   Alloc + Free (GB/s) for %6d KB, flags=N/A ",
             (int)bufSize / 1024);
  }

  testDescString = buf;
  double sec1 = timer.GetElapsedTime();
  double sec2 = timer2.GetElapsedTime();
  _perfInfo = static_cast<float>((bufSize * (double)(1e-09)) / (sec1 / iter - sec2 / numN));
#endif
}

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