GPGPU emulator for those who does not have GPU.

[woensug-choi]

The GPGPU-SIM repo https://github.com/gpgpu-sim/gpgpu-sim_distribution

This could benefit those who do not have a CUDA-capable NVIDIA card or any GPU card.

Testing on having it on our dockwater docker image

[woensug-choi]

Note,
-cuda-dev rocker

-c dockwater

• docker run wihtout -gpus all

docker run --rm -it -v /home/woensug/.gitconfig:/home/None/.gitconfig:ro -v /home/woensug:/home/woensug --name noetic_runtime -e DISPLAY -e TERM -e QT_X11_NO_MITSHM=1 -e XAUTHORITY=/tmp/.docker7ua80rmz.xauth -v /tmp/.docker7ua80rmz.xauth:/tmp/.docker7ua80rmz.xauth -v /tmp/.X11-unix:/tmp/.X11-unix -v /etc/localtime:/etc/localtime:ro 5f75307ef91f

install gpgpu dependencies

install gcc 5

https://askubuntu.com/questions/1235819/ubuntu-20-04-gcc-version-lower-than-gcc-7

add gcc 5 and 9

https://www.fosslinux.com/39386/how-to-install-multiple-versions-of-gcc-and-g-on-ubuntu-20-04.htm

select gcc 5

add PATHs

source gpgpu

compile gpgpu

link cudart.so at cmakelist
target_link_libraries(nps_multibeam_sonar_ros_plugin "libcudart.so")

select gcc 9

catkin_make

cp config files to ~/.ros

Segmentation fault when running it. Probably caused by catkin_make compiled with gcc-9 and gpgpu-sim compiled with gcc-5.
Can't compile with same gcc version..

[woensug-choi]

 * Copyright 2020 Naval Postgraduate School
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 *
*/
#include <cuda.h>
#include "cuda_runtime.h"
#include "cuda_runtime_api.h"
#include "device_launch_parameters.h"
#include <thrust/complex.h>

#include <stdio.h>
#include <iostream>
#include <complex>
#include <valarray>

#include <opencv2/core.hpp>
#include <opencv2/core/core.hpp>

// #include <math.h>
#include <assert.h>

// For complex numbers
#include <thrust/complex.h>
#include <cuComplex.h>

// For rand() function
#include <unistd.h>
#include <curand.h>
#include <curand_kernel.h>

// For FFT
#include <cufft.h>
#include <cufftw.h>
#include <thrust/device_vector.h>
#include <list>

#include <chrono>

#define BLOCK_SIZE 32

typedef std::complex<float> Complex;
typedef std::valarray<Complex> CArray;
typedef std::valarray<CArray> CArray2D;

__global__ void cuda_hello(){
  printf("Hello World from GPU!\n");
}

static inline void _safe_cuda_call(cudaError err, const char *msg,
                                   const char *file_name, const int line_number)
{
  if (err != cudaSuccess)
  {
    fprintf(stderr, "%s\n\nFile: %s\n\nLine Number: %d\n\nReason: %s\n",
            msg, file_name, line_number, cudaGetErrorString(err));
    std::cin.get();
    exit(EXIT_FAILURE);
  }
}

#define SAFE_CALL(call, msg) _safe_cuda_call((call), (msg), __FILE__, __LINE__)

///////////////////////////////////////////////////////////////////////////
// Incident Angle Calculation Function
// incidence angle is target's normal angle accounting for the ray's azimuth
// and elevation
__device__ float compute_incidence(float azimuth, float elevation, float *normal)
{
  // ray normal from camera azimuth and elevation
  float camera_x = cosf(-azimuth) * cosf(elevation);
  float camera_y = sinf(-azimuth) * cosf(elevation);
  float camera_z = sinf(elevation);
  float ray_normal[3] = {camera_x, camera_y, camera_z};

  // target normal with axes compensated to camera axes
  float target_normal[3] = {normal[2], -normal[0], -normal[1]};

  // dot product
  float dot_product = ray_normal[0] * target_normal[0]
                      + ray_normal[1] * target_normal[1]
                      + ray_normal[2] * target_normal[2];

  return M_PI - acosf(dot_product);
}

///////////////////////////////////////////////////////////////////////////
__device__ __host__ float unnormalized_sinc(float t)
{
  if (abs(t) < 1E-8)
    return 1.0;
  else
    return sin(t) / t;
}

///////////////////////////////////////////////////////////////////////////
template <typename T>
__global__ void column_sums_reduce(const T *__restrict__ in, T *__restrict__ out, size_t width, size_t height)
{

  __shared__ T sdata[BLOCK_SIZE][BLOCK_SIZE + 1];
  size_t idx = threadIdx.x + blockDim.x * blockIdx.x;
  size_t width_stride = gridDim.x * blockDim.x;
  size_t full_width = (width & (~((unsigned long long)(BLOCK_SIZE - 1)))) + ((width & (BLOCK_SIZE - 1)) ? BLOCK_SIZE : 0); // round up to next block
  for (size_t w = idx; w < full_width; w += width_stride)
  { // grid-stride loop across matrix width
    sdata[threadIdx.y][threadIdx.x] = 0;
    size_t in_ptr = w + threadIdx.y * width;
    for (size_t h = threadIdx.y; h < height; h += BLOCK_SIZE)
    { // block-stride loop across matrix height
      sdata[threadIdx.y][threadIdx.x] += (w < width) ? in[in_ptr] : 0;
      in_ptr += width * BLOCK_SIZE;
    }
    __syncthreads();
    T my_val = sdata[threadIdx.x][threadIdx.y];
    for (int i = warpSize >> 1; i > 0; i >>= 1) // warp-wise parallel sum reduction
      my_val += __shfl_xor_sync(0xFFFFFFFFU, my_val, i);
    __syncthreads();
    if (threadIdx.x == 0)
      sdata[0][threadIdx.y] = my_val;
    __syncthreads();
    if ((threadIdx.y == 0) && ((w) < width))
      out[w] = sdata[0][threadIdx.x];
  }
}

__global__ void gpu_matrix_mult(float *a, float *b, float *c, int m, int n, int k)
{
  int row = blockIdx.y * blockDim.y + threadIdx.y;
  int col = blockIdx.x * blockDim.x + threadIdx.x;
  float sum = 0;
  if (col < k && row < m)
  {
    for (int i = 0; i < n; i++)
    {
      sum += a[row * n + i] * b[i * k + col];
    }
    c[row * k + col] = sum;
  }
}

__global__ void gpu_diag_matrix_mult(float *Val, int *RowPtr, float *diagVals, int total_rows)
{
  const int row = threadIdx.x + blockIdx.x * blockDim.x;
  if (row < total_rows)
  {
    for (int i = RowPtr[row]; i < RowPtr[row + 1]; i++)
    {
      Val[i] = diagVals[row] * Val[i];
    }
  }
}

///////////////////////////////////////////////////////////////////////////
// Sonar Claculation Function
__global__ void sonar_calculation(thrust::complex<float> *P_Beams,
                                  float *depth_image,
                                  float *normal_image,
                                  int width,
                                  int height,
                                  int depth_image_step,
                                  int normal_image_step,
                                  float *rand_image,
                                  int rand_image_step,
                                  float *reflectivity_image,
                                  int reflectivity_image_step,
                                  float hPixelSize,
                                  float vPixelSize,
                                  float hFOV,
                                  float vFOV,
                                  float beam_azimuthAngleWidth,
                                  float beam_elevationAngleWidth,
                                  float ray_azimuthAngleWidth,
                                  float ray_elevationAngleWidth,
                                  float soundSpeed,
                                  float sourceTerm,
                                  int nBeams, int nRays,
                                  int raySkips,
                                  float sonarFreq, float delta_f,
                                  int nFreq, float bandwidth,
                                  float maxDistance,
                                  float attenuation,
                                  float area_scaler)
{
  // 2D Index of current thread
  const int beam = blockIdx.x * blockDim.x + threadIdx.x;
  const int ray = blockIdx.y * blockDim.y + threadIdx.y;

  //Only valid threads perform memory I/O
  if ((beam < width) && (ray < height) && (ray % raySkips == 0))
  {
    // Location of the image pixel
    const int depth_index = ray * depth_image_step / sizeof(float) + beam;
    const int normal_index = ray * normal_image_step / sizeof(float) + (3 * beam);
    const int rand_index = ray * rand_image_step / sizeof(float) + (2 * beam);
    const int reflectivity_index = ray * reflectivity_image_step / sizeof(float) + beam;

    // Input parameters for ray processing
    float distance = depth_image[depth_index] * 1.0f;
    float normal[3] = {normal_image[normal_index],
                      normal_image[normal_index + 1],
                      normal_image[normal_index + 2]};

    // Calculate ray angles
    double fl = static_cast<double>(width) / (2.0 * tan(hFOV/2.0));
    float ray_azimuthAngle = atan2(static_cast<double>(beam) -
                        0.5 * static_cast<double>(width-1), fl);
    float ray_elevationAngle = atan2(static_cast<double>(ray) -
                        0.5 * static_cast<double>(height-1), fl);

    // Beam pattern
    // float azimuthBeamPattern = abs(unnormalized_sinc(M_PI * 0.884
    // 				/ ray_azimuthAngleWidth * sin(ray_azimuthAngle)));
    // only one column of rays for each beam at beam center, interference calculated later
    float azimuthBeamPattern = 1.0;
    float elevationBeamPattern = abs(unnormalized_sinc(M_PI * 0.884
      				                    / (beam_elevationAngleWidth) * sin(ray_elevationAngle)));

    // incidence angle
    float incidence = acos(normal[2]); // compute_incidence(ray_azimuthAngle, ray_elevationAngle, normal);

    // ----- Point scattering model ------ //
    // Gaussian noise generated using opencv RNG
    float xi_z = rand_image[rand_index];
    float xi_y = rand_image[rand_index + 1];

    // Calculate amplitude
    thrust::complex<float> randomAmps = thrust::complex<float>(xi_z / sqrt(2.0), xi_y / sqrt(2.0));
    thrust::complex<float> lambert_sqrt =
        thrust::complex<float>(sqrt(reflectivity_image[reflectivity_index]) * cos(incidence), 0.0);
    thrust::complex<float> beamPattern =
        thrust::complex<float>(azimuthBeamPattern * elevationBeamPattern, 0.0);
    thrust::complex<float> targetArea_sqrt = thrust::complex<float>(sqrt(distance * area_scaler), 0.0);
    thrust::complex<float> propagationTerm =
        thrust::complex<float>(1.0 / (distance*distance) * exp(-2.0 * attenuation * distance), 0.0);
    thrust::complex<float> amplitude = randomAmps * thrust::complex<float>(sourceTerm, 0.0)
                                     * propagationTerm * beamPattern * lambert_sqrt * targetArea_sqrt;

    // Max distance cut-off
    if (distance > maxDistance)
      amplitude = thrust::complex<float>(0.0, 0.0);

    // Summation of Echo returned from a signal (frequency domain)
    for (size_t f = 0; f < nFreq; f++)
    {
      float freq;
      if (nFreq % 2 == 0)
        freq = delta_f * (-nFreq / 2.0 + f*1.0f + 1.0);
      else
        freq = delta_f * (-(nFreq - 1) / 2.0 + f*1.0f + 1.0);
      float kw = 2.0 * M_PI * freq / soundSpeed; // wave vector

      // Transmit spectrum, frequency domain
      thrust::complex<float> kernel = exp(thrust::complex<float>(0.0f, 2.0f * distance * kw)) * amplitude;
      P_Beams[beam * nFreq * (int)(nRays / raySkips) + (int)(ray / raySkips) * nFreq + f] =
          thrust::complex<float>(kernel.real() , kernel.imag());
    }
  }
}

///////////////////////////////////////////////////////////////////////////
// Sonar Claculation main function
int main()
{
  int height = 50;
  int width = 10;
  cv::Mat depth_image = cv::Mat::ones(height, width, CV_32FC1);
  cv::Mat normal_image = cv::Mat::ones(height, width, CV_32FC3);
  cv::Mat rand_image = cv::Mat::ones(height, width, CV_32FC2);
  double _hFOV = 0.2;
  double _vFOV = 0.2;
  double _soundSpeed = 1500;
  double _maxDistance = 2.0;
  double _sourceLevel = 150.0;
  int _nBeams = height; int _nRays = width;
  int _raySkips = 1;
  double _vPixelSize = _vFOV / height;
  double _hPixelSize =_hFOV / width;
  double _beam_azimuthAngleWidth = _hPixelSize;
  double _beam_elevationAngleWidth = _vFOV/180*3.141592;
  double _ray_azimuthAngleWidth = _hPixelSize;
  double _ray_elevationAngleWidth = _vPixelSize*(_raySkips+1);
  double _sonarFreq = 900e3;
  double _bandwidth = 29.5e4;
  float max_T = _maxDistance*2.0/_soundSpeed;
  float _delta_f = 1.0/max_T;
  int _nFreq = ceil(_bandwidth/_delta_f);
  cv::Mat reflectivity_image = cv::Mat::ones(height, width, CV_32FC1);
  double _attenuation = 0.1;
  float *window;
  window = new float[_nFreq];
  float **beamCorrector;
  beamCorrector = new float*[_nBeams];
  for (int i = 0; i < _nBeams; i++)
      beamCorrector[i] = new float[_nBeams];
  float beamCorrectorSum = 10;
  bool debugFlag = false;


  auto start = std::chrono::high_resolution_clock::now();
  auto stop = std::chrono::high_resolution_clock::now();
  auto duration = std::chrono::duration_cast<std::chrono::microseconds>(stop - start);
  if (debugFlag)
    start = std::chrono::high_resolution_clock::now();

  // ----  Allocation of properties parameters  ---- //
  const float hPixelSize = (float)_hPixelSize;
  const float vPixelSize = (float)_vPixelSize;
  const float hFOV = (float)_hFOV;
  const float vFOV = (float)_vFOV;
  const float beam_elevationAngleWidth = (float)_beam_elevationAngleWidth;
  const float beam_azimuthAngleWidth = (float)_beam_azimuthAngleWidth;
  const float ray_elevationAngleWidth = (float)_ray_elevationAngleWidth;
  const float ray_azimuthAngleWidth = (float)_ray_azimuthAngleWidth;
  const float soundSpeed = (float)_soundSpeed;
  const float maxDistance = (float)_maxDistance;
  const float sonarFreq = (float)_sonarFreq;
  const float bandwidth = (float)_bandwidth;
  const float attenuation = (float)_attenuation;
  const int nBeams = _nBeams;
  const int nRays = _nRays;
  const int nFreq = _nFreq;
  const int raySkips = _raySkips;

  //#######################################################//
  //###############    Sonar Calculation   ################//
  //#######################################################//
  // ---------   Calculation parameters   --------- //
  const float max_distance = maxDistance;
  // Signal
  const float delta_f = bandwidth/nFreq;
  // Precalculation
  const float area_scaler = ray_azimuthAngleWidth * ray_elevationAngleWidth;
  const float sourceLevel = (float)_sourceLevel;                     // db re 1 muPa;
  const float pref = 1e-6;                                           // 1 micro pascal (muPa);
  const float sourceTerm = sqrt(pow(10, (sourceLevel / 10))) * pref; // source term

  // ---------   Allocate GPU memory for image   --------- //
  //Calculate total number of bytes of input and output image
  const int depth_image_Bytes = depth_image.step * depth_image.rows;
  const int normal_image_Bytes = normal_image.step * normal_image.rows;
  const int rand_image_Bytes = rand_image.step * rand_image.rows;
  const int reflectivity_image_Bytes = reflectivity_image.step * reflectivity_image.rows;

  //Allocate device memory
  float *d_depth_image, *d_normal_image, *d_rand_image, *d_reflectivity_image;
  SAFE_CALL(cudaMalloc((void **)&d_depth_image, depth_image_Bytes), "CUDA Malloc Failed");
  SAFE_CALL(cudaMalloc((void **)&d_normal_image, normal_image_Bytes), "CUDA Malloc Failed");
  SAFE_CALL(cudaMalloc((void **)&d_rand_image, rand_image_Bytes), "CUDA Malloc Failed");
  SAFE_CALL(cudaMalloc((void **)&d_reflectivity_image, reflectivity_image_Bytes), "CUDA Malloc Failed");

  //Copy data from OpenCV input image to device memory
  SAFE_CALL(cudaMemcpy(d_depth_image, depth_image.ptr(), depth_image_Bytes,
                cudaMemcpyHostToDevice), "CUDA Memcpy Failed");
  SAFE_CALL(cudaMemcpy(d_normal_image, normal_image.ptr(), normal_image_Bytes,
                cudaMemcpyHostToDevice),"CUDA Memcpy Failed");
  SAFE_CALL(cudaMemcpy(d_rand_image, rand_image.ptr(), rand_image_Bytes,
                cudaMemcpyHostToDevice),"CUDA Memcpy Failed");
  SAFE_CALL(cudaMemcpy(d_reflectivity_image, reflectivity_image.ptr(), reflectivity_image_Bytes,
                cudaMemcpyHostToDevice), "CUDA Memcpy Failed");

  //Specify a reasonable block size
  const dim3 block(BLOCK_SIZE, BLOCK_SIZE);

  //Calculate grid size to cover the whole image
  const dim3 grid((depth_image.cols + block.x - 1) / block.x,
                  (depth_image.rows + block.y - 1) / block.y);

  // Beam data array
  thrust::complex<float> *P_Beams;
  thrust::complex<float> *d_P_Beams;
  const int P_Beams_N = nBeams * (int)(nRays / raySkips) * (nFreq + 1);
  const int P_Beams_Bytes = sizeof(thrust::complex<float>) * P_Beams_N;
  SAFE_CALL(cudaMallocHost((void **)&P_Beams, P_Beams_Bytes), "CUDA Malloc Failed");
  SAFE_CALL(cudaMalloc((void **)&d_P_Beams, P_Beams_Bytes), "CUDA Malloc Failed");

  //Launch the beamor conversion kernel
  sonar_calculation<<<grid, block>>>(d_P_Beams,
                                      d_depth_image,
                                      d_normal_image,
                                      normal_image.cols,
                                      normal_image.rows,
                                      depth_image.step,
                                      normal_image.step,
                                      d_rand_image,
                                      rand_image.step,
                                      d_reflectivity_image,
                                      reflectivity_image.step,
                                      hPixelSize,
                                      vPixelSize,
                                      hFOV,
                                      vFOV,
                                      beam_azimuthAngleWidth,
                                      beam_elevationAngleWidth,
                                      ray_azimuthAngleWidth,
                                      ray_elevationAngleWidth,
                                      soundSpeed,
                                      sourceTerm,
                                      nBeams, nRays,
                                      raySkips,
                                      sonarFreq, delta_f,
                                      nFreq, bandwidth,
                                      max_distance,
                                      attenuation,
                                      area_scaler);

  // Synchronize to check for any kernel launch errors
  SAFE_CALL(cudaDeviceSynchronize(), "Kernel Launch Failed");

  //Copy back data from destination device meory to OpenCV output image
  SAFE_CALL(cudaMemcpy(P_Beams, d_P_Beams, P_Beams_Bytes,
                        cudaMemcpyDeviceToHost), "CUDA Memcpy Failed");

  // Free GPU memory
  cudaFree(d_depth_image);
  cudaFree(d_normal_image);
  cudaFree(d_rand_image);
  cudaFree(d_reflectivity_image);
  cudaFree(d_P_Beams);

  // For calc time measure
  if (debugFlag)
  {
    stop = std::chrono::high_resolution_clock::now();
    duration = std::chrono::duration_cast<std::chrono::microseconds>(stop - start);
    printf("GPU Sonar Computation Time %lld/100 [s]\n",
            static_cast<long long int>(duration.count() / 10000));
    start = std::chrono::high_resolution_clock::now();
  }

  //########################################################//
  //#########   Summation, Culling and windowing   #########//
  //########################################################//
  // Preallocate an array for return
  CArray2D P_Beams_F(CArray(nFreq), nBeams);
  // GPU grids and rows
  unsigned int grid_rows, grid_cols;
  dim3 dimBlock(BLOCK_SIZE, BLOCK_SIZE);

  // GPU Ray summation using column sum
  float *P_Ray_real, *P_Ray_imag;
  float *d_P_Ray_real, *d_P_Ray_imag;
  const int P_Ray_N = (int)(nRays / raySkips) * (nFreq);
  const int P_Ray_Bytes = sizeof(float) * P_Ray_N;
  float *P_Ray_F_real, *P_Ray_F_imag;
  float *d_P_Ray_F_real, *d_P_Ray_F_imag;
  const int P_Ray_F_N = (nFreq)*1;
  const int P_Ray_F_Bytes = sizeof(float) * P_Ray_F_N;
  cudaMallocHost((void **)&P_Ray_real, P_Ray_Bytes);
  cudaMallocHost((void **)&P_Ray_imag, P_Ray_Bytes);
  cudaMallocHost((void **)&P_Ray_F_real, P_Ray_F_Bytes);
  cudaMallocHost((void **)&P_Ray_F_imag, P_Ray_F_Bytes);
  SAFE_CALL(cudaMalloc((void **)&d_P_Ray_real, P_Ray_Bytes), "CUDA Malloc Failed");
  SAFE_CALL(cudaMalloc((void **)&d_P_Ray_imag, P_Ray_Bytes), "CUDA Malloc Failed");
  SAFE_CALL(cudaMalloc((void **)&d_P_Ray_F_real, P_Ray_F_Bytes), "CUDA Malloc Failed");
  SAFE_CALL(cudaMalloc((void **)&d_P_Ray_F_imag, P_Ray_F_Bytes), "CUDA Malloc Failed");

  dim3 dimGrid_Ray((nFreq + BLOCK_SIZE - 1) / BLOCK_SIZE);

  for (size_t beam = 0; beam < nBeams; beam ++)
  {
    for (size_t ray = 0; ray < (int)(nRays / raySkips); ray++)
    {
      for (size_t f = 0; f < nFreq; f++)
      {
        P_Ray_real[ray * nFreq + f] =
            P_Beams[beam * nFreq * (int)(nRays / raySkips) + ray * nFreq + f].real();
        P_Ray_imag[ray * nFreq + f] =
            P_Beams[beam * nFreq * (int)(nRays / raySkips) + ray * nFreq + f].imag();
      }
    }

    SAFE_CALL(cudaMemcpy(d_P_Ray_real, P_Ray_real, P_Ray_Bytes, cudaMemcpyHostToDevice),
              "CUDA Memcpy Failed");
    SAFE_CALL(cudaMemcpy(d_P_Ray_imag, P_Ray_imag, P_Ray_Bytes, cudaMemcpyHostToDevice),
              "CUDA Memcpy Failed");

    column_sums_reduce<<<dimGrid_Ray, dimBlock>>>(d_P_Ray_real, d_P_Ray_F_real,
                                                  nFreq, (int)(nRays / raySkips));
    column_sums_reduce<<<dimGrid_Ray, dimBlock>>>(d_P_Ray_imag, d_P_Ray_F_imag,
                                                  nFreq, (int)(nRays / raySkips));

    SAFE_CALL(cudaMemcpy(P_Ray_F_real, d_P_Ray_F_real, P_Ray_F_Bytes,
                          cudaMemcpyDeviceToHost), "CUDA Memcpy Failed");
    SAFE_CALL(cudaMemcpy(P_Ray_F_imag, d_P_Ray_F_imag, P_Ray_F_Bytes,
                          cudaMemcpyDeviceToHost), "CUDA Memcpy Failed");
    // SAFE_CALL(cudaDeviceSynchronize(), "Kernel Launch Failed");

    for (size_t f = 0; f < nFreq; f++)
      P_Beams_F[beam][f] = Complex(P_Ray_F_real[f], P_Ray_F_imag[f]);
  }

  // free memory
  cudaFreeHost(P_Beams);
  cudaFreeHost(P_Ray_real);
  cudaFreeHost(P_Ray_imag);
  cudaFreeHost(P_Ray_F_real);
  cudaFreeHost(P_Ray_F_imag);
  cudaFree(d_P_Ray_real);
  cudaFree(d_P_Ray_imag);
  cudaFree(d_P_Ray_F_real);
  cudaFree(d_P_Ray_F_imag);

  if (debugFlag)
  {
    stop = std::chrono::high_resolution_clock::now();
    duration = std::chrono::duration_cast<std::chrono::microseconds>(stop - start);
    printf("Sonar Ray Summation %lld/100 [s]\n",
          static_cast<long long int>(duration.count() / 10000));
    start = std::chrono::high_resolution_clock::now();
  }

  // -------------- Beam culling correction -----------------//
  // beamCorrector and beamCorrectorSum is precalculated at parent cpp
  float *P_Beams_Cor_real, *P_Beams_Cor_imag;
  // float *P_Beams_Cor_F_real, *P_Beams_Cor_F_imag;
  float *P_Beams_Cor_real_tmp, *P_Beams_Cor_imag_tmp;
  float *d_P_Beams_Cor_real, *d_P_Beams_Cor_imag;
  float *d_P_Beams_Cor_F_real, *d_P_Beams_Cor_F_imag;
  const int P_Beams_Cor_N = nBeams * nFreq;
  const int P_Beams_Cor_Bytes = sizeof(float) * P_Beams_Cor_N;
  cudaMallocHost((void **)&P_Beams_Cor_real, P_Beams_Cor_Bytes);
  cudaMallocHost((void **)&P_Beams_Cor_imag, P_Beams_Cor_Bytes);
  cudaMallocHost((void **)&P_Beams_Cor_real_tmp, P_Beams_Cor_Bytes);
  cudaMallocHost((void **)&P_Beams_Cor_imag_tmp, P_Beams_Cor_Bytes);
  // cudaMallocHost((void **)&P_Beams_Cor_F_real, P_Beams_Cor_Bytes);
  // cudaMallocHost((void **)&P_Beams_Cor_F_imag, P_Beams_Cor_Bytes);
  SAFE_CALL(cudaMalloc((void **)&d_P_Beams_Cor_real, P_Beams_Cor_Bytes), "CUDA Malloc Failed");
  SAFE_CALL(cudaMalloc((void **)&d_P_Beams_Cor_imag, P_Beams_Cor_Bytes), "CUDA Malloc Failed");
  SAFE_CALL(cudaMalloc((void **)&d_P_Beams_Cor_F_real, P_Beams_Cor_Bytes), "CUDA Malloc Failed");
  SAFE_CALL(cudaMalloc((void **)&d_P_Beams_Cor_F_imag, P_Beams_Cor_Bytes), "CUDA Malloc Failed");

  float *beamCorrector_lin, *d_beamCorrector_lin;
  const int beamCorrector_lin_N = nBeams * nBeams;
  const int beamCorrector_lin_Bytes = sizeof(float) * beamCorrector_lin_N;
  cudaMallocHost((void **)&beamCorrector_lin, beamCorrector_lin_Bytes);
  SAFE_CALL(cudaMalloc((void **)&d_beamCorrector_lin, beamCorrector_lin_Bytes), "CUDA Malloc Failed");

  // (nfreq x nBeams) * (nBeams x nBeams) = (nfreq x nBeams)
  for (size_t beam = 0; beam < nBeams; beam ++)
  {
    for (size_t f = 0; f < nFreq; f++)
    {
      P_Beams_Cor_real[f * nBeams + beam] = P_Beams_F[beam][f].real() * 1.0f;
      P_Beams_Cor_imag[f * nBeams + beam] = P_Beams_F[beam][f].imag() * 1.0f;
    }
    for (size_t beam_other = 0; beam_other < nBeams; beam_other ++)
      beamCorrector_lin[beam_other * nBeams + beam] = beamCorrector[beam][beam_other];
  }

  SAFE_CALL(cudaMemcpy(d_P_Beams_Cor_real, P_Beams_Cor_real, P_Beams_Cor_Bytes,
                        cudaMemcpyHostToDevice),
            "CUDA Memcpy Failed");
  SAFE_CALL(cudaMemcpy(d_P_Beams_Cor_imag, P_Beams_Cor_imag, P_Beams_Cor_Bytes,
                        cudaMemcpyHostToDevice),
            "CUDA Memcpy Failed");
  SAFE_CALL(cudaMemcpy(d_beamCorrector_lin, beamCorrector_lin, beamCorrector_lin_Bytes,
                        cudaMemcpyHostToDevice),
            "CUDA Memcpy Failed");

  grid_rows = (nFreq + BLOCK_SIZE - 1) / BLOCK_SIZE;
  grid_cols = (nBeams + BLOCK_SIZE - 1) / BLOCK_SIZE;
  dim3 dimGrid_Beam(grid_cols, grid_rows);

  gpu_matrix_mult<<<dimGrid_Beam, dimBlock>>>(d_P_Beams_Cor_real, d_beamCorrector_lin,
                                              d_P_Beams_Cor_F_real, nFreq, nBeams, nBeams);
  // SAFE_CALL(cudaDeviceSynchronize(), "Kernel Launch Failed");

  gpu_matrix_mult<<<dimGrid_Beam, dimBlock>>>(d_P_Beams_Cor_imag, d_beamCorrector_lin,
                                              d_P_Beams_Cor_F_imag, nFreq, nBeams, nBeams);
  // SAFE_CALL(cudaDeviceSynchronize(), "Kernel Launch Failed");

  //Copy back data from destination device meory
  SAFE_CALL(cudaMemcpy(P_Beams_Cor_real_tmp, d_P_Beams_Cor_F_real, P_Beams_Cor_Bytes,
                        cudaMemcpyDeviceToHost),
            "CUDA Memcpy Failed");
  SAFE_CALL(cudaMemcpy(P_Beams_Cor_imag_tmp, d_P_Beams_Cor_F_imag, P_Beams_Cor_Bytes,
                        cudaMemcpyDeviceToHost),
            "CUDA Memcpy Failed");
  // SAFE_CALL(cudaDeviceSynchronize(), "Kernel Launch Failed");

  // ---------------    Windowing   ----------------- //
  // float *window_diag, *d_window;
  // const int window_N = nFreq * 1;
  // const int window_Bytes = sizeof(float) * window_N;
  // window_diag = (float *)malloc(window_Bytes);
  // SAFE_CALL(cudaMalloc((void **)&d_window, window_Bytes), "CUDA Malloc Failed");

  // int *diag_ptr, *d_diag_ptr;
  // const int diag_ptr_N = nBeams * 1;
  // const int diag_ptr_Bytes = sizeof(int) * diag_ptr_N;
  // diag_ptr = (int *)malloc(diag_ptr_Bytes);
  // SAFE_CALL(cudaMalloc((void **)&d_diag_ptr, diag_ptr_Bytes), "CUDA Malloc Failed");

  // // (nBeams x nfreq) * (1 x nFreq) = (nBeams x nFreq)
  // for (size_t beam = 0; beam < nBeams; beam ++)
  // {
  //   for (size_t f = 0; f < nFreq; f++)
  //   { // Transpose
  //     P_Beams_Cor_real[beam * nFreq + f] = P_Beams_Cor_real_tmp[f * nBeams + beam];
  //     P_Beams_Cor_imag[beam * nFreq + f] = P_Beams_Cor_imag_tmp[f * nBeams + beam];
  //     window_diag[f] = window[f];
  //   }
  //   diag_ptr[beam] = (int)beam;
  // }
  // SAFE_CALL(cudaMemcpy(d_P_Beams_Cor_real, P_Beams_Cor_real, P_Beams_Cor_Bytes,
  //                      cudaMemcpyHostToDevice),
  //           "CUDA Memcpy Failed");
  // SAFE_CALL(cudaMemcpy(d_P_Beams_Cor_imag, P_Beams_Cor_imag, P_Beams_Cor_Bytes,
  //                      cudaMemcpyHostToDevice),
  //           "CUDA Memcpy Failed");
  // SAFE_CALL(cudaMemcpy(d_window, window_diag, window_Bytes,
  //                      cudaMemcpyHostToDevice),
  //           "CUDA Memcpy Failed");
  // SAFE_CALL(cudaMemcpy(d_diag_ptr, diag_ptr, diag_ptr_Bytes,
  //                      cudaMemcpyHostToDevice),
  //           "CUDA Memcpy Failed");

  // grid_rows = (nFreq + BLOCK_SIZE - 1) / BLOCK_SIZE;
  // grid_cols = (nFreq + BLOCK_SIZE - 1) / BLOCK_SIZE;
  // dim3 dimGrid_window(grid_cols, grid_rows);
  // gpu_diag_matrix_mult<<<dimGrid_window, dimBlock>>>(d_P_Beams_Cor_real, d_diag_ptr, d_window, nFreq);
  // SAFE_CALL(cudaDeviceSynchronize(), "Kernel Launch Failed");

  // gpu_diag_matrix_mult<<<dimGrid_window, dimBlock>>>(d_P_Beams_Cor_imag, d_diag_ptr, d_window, nFreq);
  // SAFE_CALL(cudaDeviceSynchronize(), "Kernel Launch Failed");

  // //Copy back data from destination device meory
  // SAFE_CALL(cudaMemcpy(P_Beams_Cor_F_real, d_P_Beams_Cor_real, P_Beams_Cor_Bytes,
  //                      cudaMemcpyDeviceToHost),
  //           "CUDA Memcpy Failed");
  // SAFE_CALL(cudaMemcpy(P_Beams_Cor_F_imag, d_P_Beams_Cor_imag, P_Beams_Cor_Bytes,
  //                      cudaMemcpyDeviceToHost),
  //           "CUDA Memcpy Failed");
  // SAFE_CALL(cudaDeviceSynchronize(), "Kernel Launch Failed");

  // Return
  for (size_t beam = 0; beam < nBeams; beam ++)
    for (size_t f = 0; f < nFreq; f++)
      P_Beams_F[beam][f] =
          Complex(P_Beams_Cor_real_tmp[f * nBeams + beam] / beamCorrectorSum,
                  P_Beams_Cor_imag_tmp[f * nBeams + beam] / beamCorrectorSum);

  // Free memory
  cudaFree(d_P_Beams_Cor_imag);
  cudaFree(d_P_Beams_Cor_real);
  cudaFree(d_P_Beams_Cor_F_imag);
  cudaFree(d_P_Beams_Cor_F_real);
  cudaFree(d_beamCorrector_lin);
  // cudaFree(d_window);
  // cudaFree(d_diag_ptr);
  cudaFreeHost(P_Beams_Cor_real);
  cudaFreeHost(P_Beams_Cor_imag);
  // cudaFreeHost(P_Beams_Cor_F_real);
  // cudaFreeHost(P_Beams_Cor_F_imag);
  cudaFreeHost(P_Beams_Cor_real_tmp);
  cudaFreeHost(P_Beams_Cor_imag_tmp);
  cudaFreeHost(beamCorrector_lin);
  // cudaFreeHost(window_diag);
  // cudaFreeHost(diag_ptr);

  // For calc time measure
  if (debugFlag)
  {
    stop = std::chrono::high_resolution_clock::now();
    duration = std::chrono::duration_cast<std::chrono::microseconds>(stop - start);
    printf("GPU Window & Correction %lld/100 [s]\n",
          static_cast<long long int>(duration.count() / 10000));
    start = std::chrono::high_resolution_clock::now();
  }

  //#################################################//
  //###################   FFT   #####################//
  //#################################################//
  // SAFE_CALL(cudaDeviceSynchronize(), "Kernel Launch Failed");
  const int DATASIZE = nFreq;
  const int BATCH = nBeams;
  // --- Host side input data allocation and initialization
  cufftComplex *hostInputData = (cufftComplex *)malloc(
      DATASIZE * BATCH * sizeof(cufftComplex));
  for (int beam = 0; beam < BATCH; beam++)
  {
    for (int f = 0; f < DATASIZE; f++)
    {
      if (f < nFreq)
        hostInputData[beam * DATASIZE + f] =
            make_cuComplex(P_Beams_F[beam][f].real() * 1.0f,
                            P_Beams_F[beam][f].imag() * 1.0f);
      else
        hostInputData[beam * DATASIZE + f] =
            (make_cuComplex(0.f, 0.f)); // zero padding
    }
  }

  // --- Device side input data allocation and initialization
  cufftComplex *deviceInputData;
  SAFE_CALL(cudaMalloc((void **)&deviceInputData,
                        DATASIZE * BATCH * sizeof(cufftComplex)),
                        "FFT CUDA Malloc Failed");
  SAFE_CALL(cudaMemcpy(deviceInputData, hostInputData,
                        DATASIZE * BATCH * sizeof(cufftComplex),
                        cudaMemcpyHostToDevice),
                        "FFT CUDA Memcopy Failed");

  // --- Host side output data allocation
  cufftComplex *hostOutputData =
      (cufftComplex *)malloc(DATASIZE * BATCH * sizeof(cufftComplex));

  // --- Device side output data allocation
  cufftComplex *deviceOutputData;
  cudaMalloc((void **)&deviceOutputData,
              DATASIZE * BATCH * sizeof(cufftComplex));

  // --- Batched 1D FFTs
  cufftHandle handle;
  int rank = 1;         // --- 1D FFTs
  int n[] = {DATASIZE}; // --- Size of the Fourier transform
  // --- Distance between two successive input/output elements
  int istride = 1, ostride = 1;
  int idist = DATASIZE, odist = DATASIZE; // --- Distance between batches
  // --- Input/Output size with pitch (ignored for 1D transforms)
  int inembed[] = {0};
  int onembed[] = {0};
  int batch = BATCH; // --- Number of batched executions
  cufftPlanMany(&handle, rank, n,
                inembed, istride, idist,
                onembed, ostride, odist, CUFFT_C2C, batch);

  cufftExecC2C(handle, deviceInputData, deviceOutputData, CUFFT_FORWARD);

  // --- Device->Host copy of the results
  SAFE_CALL(cudaMemcpy(hostOutputData, deviceOutputData,
                        DATASIZE * BATCH * sizeof(cufftComplex),
                        cudaMemcpyDeviceToHost),
                        "FFT CUDA Memcopy Failed");

  cufftDestroy(handle);
  cudaFree(deviceOutputData);
  cudaFree(deviceInputData);
  free(hostInputData);
  free(hostOutputData);


  for (int beam = 0; beam < BATCH; beam++)
  {
    for (int f = 0; f < nFreq; f++)
    {
      P_Beams_F[beam][f] =
          Complex(hostOutputData[beam * DATASIZE + f].x * delta_f,
                  hostOutputData[beam * DATASIZE + f].y * delta_f);
    }
  }

  // For calc time measure
  if (debugFlag)
  {
    stop = std::chrono::high_resolution_clock::now();
    duration = std::chrono::duration_cast<std::chrono::microseconds>(stop - start);
    printf("GPU FFT Calc Time %lld/100 [s]\n",
          static_cast<long long int>(duration.count() / 10000));
  }

  cuda_hello<<<1,1>>>();
  return 0;
}

Works well as a standalone excutable with

FROM osrf/ros:melodic-desktop-full

# Required utilities
RUN apt update \
 && apt install -y --no-install-recommends \
        build-essential \
        cmake \
        cppcheck \
        curl \
        git \
        gnupg \
        libeigen3-dev \
        libgles2-mesa-dev \
        libglu1-mesa-dev \
        libgl1-mesa-glx \
        libgl1-mesa-dri \
        xutils-dev \
        bison \
        zlib1g-dev \
        flex \
        lsb-release \
        pkg-config \
        protobuf-compiler \
        python3-dbg \
        python3-pip \
        python3-venv \
        python3-catkin-tools \
        qtbase5-dev \
        ruby \
        software-properties-common \
        sudo \
        wget \
        gdb \
        psmisc \
        vim \
        nano \
        x11-apps \
 && apt clean

# --- Install GCUDA Libs --- #
# Download Cuda local installer
WORKDIR /tmp
RUN wget https://developer.download.nvidia.com/compute/cuda/11.1.1/local_installers/cuda_11.1.1_455.32.00_linux.run

# Install CUDA Libs (without opengl libraries  to keep GUI libraries)
RUN sh cuda_11.1.1_455.32.00_linux.run --silent --toolkit --no-opengl-libs

# --- Install GPGPU-SIM --- #
# Clone source code
WORKDIR /tmp/gpgpu-sim
ENV VERSION_GPGPU_SIM=4.0.3
RUN curl -fsSL https://github.com/woensug-choi/gpgpu-sim_distribution/archive/v$VERSION_GPGPU_SIM.tar.gz | tar xz
RUN mv gpgpu-sim_distribution-$VERSION_GPGPU_SIM gpgpu-sim_distribution

# Define environment variables
ENV CUDA_INSTALL_PATH="/usr/local/cuda"
ENV PATH="$CUDA_INSTALL_PATH/bin:${PATH}"

# Compile
WORKDIR /tmp/gpgpu-sim/gpgpu-sim_distribution
RUN /bin/bash -c "source setup_environment && make -j$(nproc)"

# Define runtime env. variables
ENV PTX_SIM_MODE_FUNC=1
ENV PTX_SIM_MODE_DETAIL=0

# Install dave dependencies
# Set Ubuntu release
ARG RELEASE=bionic
# Set ROS distribution
ARG DIST=melodic
# Set Gazebo verison
ARG GAZ=gazebo9
RUN apt update \
 && apt install -y --no-install-recommends \
    ${GAZ} \
    lib${GAZ}-dev \
    python-vcstool \
    ros-${DIST}-gazebo-plugins \
    ros-${DIST}-gazebo-ros \
    ros-${DIST}-gazebo-ros-control \
    ros-${DIST}-gazebo-ros-pkgs \
    ros-${DIST}-effort-controllers \
    ros-${DIST}-geographic-info \
    ros-${DIST}-hector-gazebo-plugins \
    ros-${DIST}-joint-state-controller \
    ros-${DIST}-joint-state-publisher \
    ros-${DIST}-joy \
    ros-${DIST}-joy-teleop \
    ros-${DIST}-kdl-parser-py \
    ros-${DIST}-key-teleop \
    ros-${DIST}-move-base \
    ros-${DIST}-robot-localization \
    ros-${DIST}-robot-state-publisher \
    ros-${DIST}-ros-base \
    ros-${DIST}-rqt \
    ros-${DIST}-rqt-common-plugins \
    ros-${DIST}-rqt-robot-plugins \
    ros-${DIST}-rviz \
    ros-${DIST}-teleop-tools \
    ros-${DIST}-teleop-twist-joy \
    ros-${DIST}-teleop-twist-keyboard \
    ros-${DIST}-tf2-geometry-msgs \
    ros-${DIST}-tf2-tools \
    ros-${DIST}-urdfdom-py \
    ros-${DIST}-velodyne-gazebo-plugins \
    ros-${DIST}-velodyne-simulator \
    ros-${DIST}-xacro \
 && apt clean


# Add user
ARG USER_ID=1000
ARG GROUP_ID=1000
RUN groupadd -g ${GROUP_ID} gpgpu-sim \
 && useradd -ms /bin/bash gpgpu-sim -g gpgpu-sim

# RUN COMMAND
# docker run -it --net=host \
#     --env="DISPLAY" \
#     --env="QT_X11_NO_MITSHM=1" \
#     --volume="/tmp/.X11-unix:/tmp/.X11-unix:rw" \
#     -v /home/${USER}:/home/gpgpu-sim \
#     -u $(id -u ${USER}):$(id -g ${USER}) \
#     gpgpu-sim-melodic \
#     bash

# First things first
# source /tmp/gpgpu-sim/gpgpu-sim_distribution/setup_environment