filtermain.cpp

Go to the documentation of this file.

/*
  GPU multi-rate FIR filter bank example software

  Oxford e-Research Centre, Oxford University

  Centre for Digital Music, Queen Mary, University of London.

  This program is free software: you can redistribute it and/or modify
  it under the terms of the GNU General Public License as published by
  the Free Software Foundation, either version 3 of the License, 
  or (at your option) any later version.
  See the file COPYING included with this distribution for more information.
*/

#include <cstdio>
#include <cstdlib>
#include <cstring>

#include <sys/time.h>
#include "filters.h"

// Program main

using namespace std;


int main( int argc, char** argv) 
{
  unsigned int N = 100; // total number of input blocks before resampling

  int bufb_size = B_SIZE + OFFSET;

// the total input signal to loop through is
// essentially used to check results and provide multiple
// block iterations.

  float *h_in[MAX_RATES], *h_out[MAX_RATES], *h_reference[MAX_RATES]; // input and output arrays
  int out_sz[MAX_RATES];  //total output size for all filters

  //timing
  struct timeval t_start, t_end;
  double t_gpu = 0.0, t_mcpu = 0.0;

  // parameters, filter and GPU arrays structures
  // object pointers can be passed to these for C++ code
  filter_arrays farr;
  params gparams;
  gpu_arrays gpuarrays;
  int ratenumf[MAX_RATES];
 
  // command line arguments
  cmd_args args;

  //deaults
  args.nrates = 3;
  args.nf = 60;
  args.insize = 1024;
  args.rconst = 0;
  args.tim=1;

  read_command_line(argc, argv, &args);

  //initialise parameters
  int nf = args.nf;
  int numrates = args.nrates;
  int rem= nf%numrates;

  printf("\nGPU FIR filter parameters\n------\n");
  printf("\nTotal number of input blocks = %d\n", N);
  gparams.nfilters = nf; // not sure if this is needed
  printf("\nTotal number of filters = %d\n", gparams.nfilters);
  gparams.fsize = B_SIZE;
  printf("\nFilter size = %d\n", gparams.fsize);
  gparams.nrates = numrates;
  printf("\nNumber of sampling rates = %d\n", gparams.nrates);
  gparams.streams = 1;
  printf("\nCUDA streams flag = %d\n", gparams.streams);

  //dividing sampling rates equally...
  for (int i=0; i< numrates; ++i){
   ratenumf[i]  = nf/numrates;
  }
  if (rem > 0){
    for (int i=0; i<rem; ++i)
      ratenumf[i]++;
  }

  int in_sz = args.insize; 
  if (numrates == 1 && args.rconst==1){
      gpuarrays.osize[0] = in_sz;
  }
  else{
    for (int i=0; i< numrates; ++i){
      gpuarrays.osize[i] = in_sz/rdiv[i];
      if (gpuarrays.osize[i] < B_SIZE ){
    printf("\n\nInput size for rate %d is  shorter than filter size.\nChose a longer input block or a shorter filter.\n", i);
    printf("\nFilter size = %d\n", B_SIZE);
    printf("\nDecimated  size = %d\n", gpuarrays.osize[i]);
    printf("\nInitial input block size = %d\n", in_sz);
    printf("\nDecimation factor = %d\n\n", rdiv[i]);
    exit(EXIT_FAILURE);
      }
    }
  }

  int out_blk_sz = 0;

  for (int i=0; i<numrates; i++){
    gparams.rnumf[i] = ratenumf[i];
    printf("\nNumber of filters for rate %d = %d\n", i, gparams.rnumf[i]);
    gpuarrays.isize[i] = gpuarrays.osize[i] + gparams.fsize;
    printf("\nGPU Input size for rate %d = %d\n", i, gpuarrays.isize[i]);
    out_sz[i] = N * gpuarrays.osize[i];
    out_blk_sz +=ratenumf[i] *gpuarrays.osize[i];
    printf("\nGPU output size for 1 block for this rate: %d\n", ratenumf[i] *gpuarrays.osize[i]);
  }
    printf("\ntotal output size for 1 block for all rates: %d\n", out_blk_sz);
    printf("----------------\n");

  int oindex = 0;
  int pos[MAX_RATES];
  int sum = 0;
  pos[0] = 0;
  for (int r=1; r<numrates; r++){
    sum+=ratenumf[r-1];
    pos[r] = sum;
  }

  // Initialize arrays
  for (int i=0; i<numrates; i++){
    h_in[i] = (float*) malloc((gpuarrays.osize[i]*N + gparams.fsize)*sizeof(float));

    h_out[i] = (float*) malloc(gpuarrays.osize[i]*ratenumf[i]*N*sizeof(float));
    
    h_reference[i] = (float*) malloc(gpuarrays.osize[i]*ratenumf[i]*N*sizeof(float));

    for (int n=0; n<gparams.fsize; ++n)
      h_in[i][n] = 0.0f;

    for (int n=0; n<gpuarrays.osize[i]*N; ++n)
      h_in[i][n + gparams.fsize] = rand() / (float)RAND_MAX; 

    for (int n=0; n<gpuarrays.osize[i]*ratenumf[i]*N; ++n){
      h_out[i][n] = 0.0f;
      h_reference[i][n] = 0.0f;
    }
  }

  // initialize filters with random numbers
  farr.bk = (float*) malloc(nf*gparams.fsize*sizeof(float)); 
  for (int i=0; i < nf; ++i){
    farr.m_offb[i] = OFFSET;
    for (int b=0; b < gparams.fsize;  ++b)
      farr.bk[i*gparams.fsize+b] = rand() / (float)RAND_MAX;
  }
  
  //initialize cpu buffers for each filter 
  for (int f=0; f < nf; ++f){
    for (int i=0; i < bufb_size ; i++){
      farr.buf_in[f][i] = 0.0;
    }
  }   

  printf("\nRunning multirate filter bank test on GPU, CPU, and CPU OpenMP\n");
  //  printf("----------------\n");
  // compute reference solution  
  printf("\ncompute reference solution (1 CPU)\n");

  compute_ref( h_in, h_reference, &gpuarrays, &gparams, &args, &farr, N);

  //reset CPU buffers
  for (int f=0; f < nf; ++f){
    farr.m_offb[f] = OFFSET;
    for (int i=0; i < bufb_size ; ++i)
      farr.buf_in[f][i] = 0.0;
  }

// compute OPENMP solution    
  printf("\nCompute OPENMP solution\n");

  compute_omp( h_in, h_reference, &gpuarrays, &gparams, &args, &farr, N);

 // compute CUDA solution  
 printf("\nCompute cuda solution\n");
 
// init
 cudaMultiFilterFirInit(h_in, h_out, farr.bk, &gparams, &gpuarrays);
 for (int i=0; i<numrates; i++)
   memcpy(gpuarrays.h_in[i], &h_in[i][0], gpuarrays.isize[i]*sizeof(float) );
 //time execution
 oindex = 0;
 if (args.tim)
   gettimeofday(&t_start, NULL);
 
 for (int ii=0; ii<N; ++ii){ // loop through input blocks
   
   //need to copy data to pinned memory for streams...
   for (int i=0; i<numrates; i++){
     if (ii==0)
       memcpy(gpuarrays.h_in[i], &h_in[i][0], gpuarrays.isize[i]*sizeof(float) );
     else
       memcpy(gpuarrays.h_in[i], &h_in[i][ii*gpuarrays.osize[i]], gpuarrays.isize[i]*sizeof(float) );
}

   //call GPU function   
   cudaMultiFilterFirStreams(&gpuarrays, &gparams);
   
   // ... and copy data back from pinned memory...
   for (int i=0; i<numrates; i++){
     oindex = ii*ratenumf[i]*gpuarrays.osize[i];
     cudaDeviceSynchronize(); 
     memcpy(&h_out[i][oindex], gpuarrays.h_out[i], gpuarrays.osize[i]* ratenumf[i] *sizeof(float) );
   }
 }

 if (args.tim){
   cudaDeviceSynchronize();
   gettimeofday(&t_end, NULL); 
   t_gpu = (double) (t_end.tv_sec + (t_end.tv_usec / 1000000.0)  - t_start.tv_sec - (t_start.tv_usec/ 1000000.0)) * 1000.0; 
   printf("Finished GPU FIR\nProcessing gpu took: %f ms\n", t_gpu/(double)N);   
 }

 // check results
 check_results(h_reference, h_out, &gpuarrays, &gparams, N);
  
 //close GPU
 cudaMultiFilterFirClose(&gparams, &gpuarrays);
 
 // cleanup memory
 for (int i = 0; i < numrates; ++i){
   free(h_out[i]);
   free(h_in[i]);
   free(h_reference[i]);
 }
 
 free(farr.bk);
 
}