TheRock/rocm-systems/projects/clr/opencl/tests/ocltst/module/perf/OCLPerfKernelThroughput.h

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

/*******************************************************************************
 * Kernel Throughput
 *
 *
 *
 *
 *
 *
 ******************************************************************************/

#ifndef _OCL_KernelThroughput_H_
#define _OCL_KernelThroughput_H_

#ifdef WIN32
#include "xmmintrin.h"
#endif

#include "OCLTestImp.h"
// #include <sstream>
// #define WIN32_LEAN_AND_MEAN //Restricts windows.h to include only the core
// API. #include "windows.h" #undef Yield #include <process.h> #include
//<xmmintrin.h> #include <emmintrin.h> #include <pmmintrin.h>

#define LARGE_INT long long
#define UNSIGNED_LARGE_INT unsigned long long
#define MAX_LOOP_ITER 10
typedef cl_float4 float4;
typedef void (*CPUKernel)(__m128*, __m128*, unsigned int);

class OCLPerfKernelThroughput : public OCLTestImp {
 public:
  OCLPerfKernelThroughput();
  virtual ~OCLPerfKernelThroughput();

 public:
  virtual void open(unsigned int test, char* units, double& conversion, unsigned int deviceID);
  virtual void run(void);
  virtual unsigned int close(void);

  std::string shader_;
  void genShaderMadds();
  void genShaderMatrixMultiply();
  void checkData();
  // void allocateBuffers();
  void launchKernel();

  // test parameters
  int kernelTypeIdx_;
  int memPathIdx_;
  int numElementsIdx_;
  int workSizeIdx_;
  float gold_;
  double _reqDataSize;
  bool _dataSizeTooBig;

  // device attributes
  cl_uint maxComputeUnits_;
  cl_uint maxClockFrequency_;

  LARGE_INT numComputeUnits_;
  LARGE_INT numWorkGroupsPerComputeUnit_;
  LARGE_INT numThreads_;
  cl_uint work_dim_;
  size_t* global_work_size_;
  size_t* local_work_size_;

  // opencl objects
  cl_context context_;
  cl_command_queue cmd_queue_;
  cl_program program_;
  cl_kernel kernel_;
  cl_int error_;

  // buffer sizes

  // kernel-specific values
  int flopsPerByte_;
  int matrixDim1_, matrixDim2_;

  // buffers
  size_t input1BufferSize_;
  size_t input2BufferSize_;
  size_t output1BufferSize_;
  cl_mem input1Buffer_;
  cl_mem input2Buffer_;
  cl_mem output1Buffer_;
  float* input1Ptr_;
  float* input2Ptr_;
  float* output1Ptr_;

  // performance results
  float bandwidth_;      // GB/s
  float gflops_;         // GFlop/s
  float avgKernelTime_;  // microseconds
};

#endif  // _OCL_KernelThroughput_H_