recursiveGaussian_cuda.cu

/* Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 *  * Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 *  * Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 *  * Neither the name of NVIDIA CORPORATION nor the names of its
 *    contributors may be used to endorse or promote products derived
 *    from this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
 * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
 * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
 * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */

/*
  Recursive Gaussian filter
  sgreen 8/1/08

  This code sample implements a Gaussian blur using Deriche's recursive method:
  http://citeseer.ist.psu.edu/deriche93recursively.html

  This is similar to the box filter sample in the SDK, but it uses the previous
  outputs of the filter as well as the previous inputs. This is also known as an
  IIR (infinite impulse response) filter, since its response to an input impulse
  can last forever.

  The main advantage of this method is that the execution time is independent of
  the filter width.

  The GPU processes columns of the image in parallel. To avoid uncoalesced reads
  for the row pass we transpose the image and then transpose it back again
  afterwards.

  The implementation is based on code from the CImg library:
  http://cimg.sourceforge.net/
  Thanks to David Tschumperl� and all the CImg contributors!
*/

#include <stdlib.h>
#include <stdio.h>
#include <string.h>

#include <cuda_runtime.h>
#include <helper_cuda.h>
#include <helper_math.h>

#include "recursiveGaussian_kernel.cuh"

#define USE_SIMPLE_FILTER 0

// Round a / b to nearest higher integer value
int iDivUp(int a, int b) { return (a % b != 0) ? (a / b + 1) : (a / b); }

/*
  Transpose a 2D array (see SDK transpose example)
*/
extern "C" void transpose(uint *d_src, uint *d_dest, uint width, int height) {
  dim3 grid(iDivUp(width, BLOCK_DIM), iDivUp(height, BLOCK_DIM), 1);
  dim3 threads(BLOCK_DIM, BLOCK_DIM, 1);
  d_transpose<<<grid, threads>>>(d_dest, d_src, width, height);
  getLastCudaError("Kernel execution failed");
}

/*
  Perform Gaussian filter on a 2D image using CUDA

  Parameters:
  d_src  - pointer to input image in device memory
  d_dest - pointer to destination image in device memory
  d_temp - pointer to temporary storage in device memory
  width  - image width
  height - image height
  sigma  - sigma of Gaussian
  order  - filter order (0, 1 or 2)
*/

// 8-bit RGBA version
extern "C" void gaussianFilterRGBA(uint *d_src, uint *d_dest, uint *d_temp,
                                   int width, int height, float sigma,
                                   int order, int nthreads) {
  // compute filter coefficients
  const float nsigma = sigma < 0.1f ? 0.1f : sigma, alpha = 1.695f / nsigma,
              ema = (float)std::exp(-alpha), ema2 = (float)std::exp(-2 * alpha),
              b1 = -2 * ema, b2 = ema2;

  float a0 = 0, a1 = 0, a2 = 0, a3 = 0, coefp = 0, coefn = 0;

  switch (order) {
    case 0: {
      const float k = (1 - ema) * (1 - ema) / (1 + 2 * alpha * ema - ema2);
      a0 = k;
      a1 = k * (alpha - 1) * ema;
      a2 = k * (alpha + 1) * ema;
      a3 = -k * ema2;
    } break;

    case 1: {
      const float k = (1 - ema) * (1 - ema) / ema;
      a0 = k * ema;
      a1 = a3 = 0;
      a2 = -a0;
    } break;

    case 2: {
      const float ea = (float)std::exp(-alpha),
                  k = -(ema2 - 1) / (2 * alpha * ema),
                  kn = (-2 * (-1 + 3 * ea - 3 * ea * ea + ea * ea * ea) /
                        (3 * ea + 1 + 3 * ea * ea + ea * ea * ea));
      a0 = kn;
      a1 = -kn * (1 + k * alpha) * ema;
      a2 = kn * (1 - k * alpha) * ema;
      a3 = -kn * ema2;
    } break;

    default:
      fprintf(stderr, "gaussianFilter: invalid order parameter!\n");
      return;
  }

  coefp = (a0 + a1) / (1 + b1 + b2);
  coefn = (a2 + a3) / (1 + b1 + b2);

// process columns
#if USE_SIMPLE_FILTER
  d_simpleRecursive_rgba<<<iDivUp(width, nthreads), nthreads>>>(
      d_src, d_temp, width, height, ema);
#else
  d_recursiveGaussian_rgba<<<iDivUp(width, nthreads), nthreads>>>(
      d_src, d_temp, width, height, a0, a1, a2, a3, b1, b2, coefp, coefn);
#endif
  getLastCudaError("Kernel execution failed");

  transpose(d_temp, d_dest, width, height);
  getLastCudaError("transpose: Kernel execution failed");

// process rows
#if USE_SIMPLE_FILTER
  d_simpleRecursive_rgba<<<iDivUp(height, nthreads), nthreads>>>(
      d_dest, d_temp, height, width, ema);
#else
  d_recursiveGaussian_rgba<<<iDivUp(height, nthreads), nthreads>>>(
      d_dest, d_temp, height, width, a0, a1, a2, a3, b1, b2, coefp, coefn);
#endif
  getLastCudaError("Kernel execution failed");

  transpose(d_temp, d_dest, height, width);
}