RetroArch/audio/filters/eq.c

/*  RetroArch - A frontend for libretro.
 *  Copyright (C) 2010-2014 - Hans-Kristian Arntzen
 *
 *  RetroArch 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 Found-
 *  ation, either version 3 of the License, or (at your option) any later version.
 *
 *  RetroArch is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY;
 *  without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR
 *  PURPOSE.  See the GNU General Public License for more details.
 *
 *  You should have received a copy of the GNU General Public License along with RetroArch.
 *  If not, see <http://www.gnu.org/licenses/>.
 */

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

#include "fft/fft.c"

#ifndef M_PI
#define M_PI 3.1415926535897932384626433832795
#endif

#ifndef min
#define min(a, b) ((a) < (b) ? (a) : (b))
#endif

struct eq_data
{
   fft_t *fft;
   float buffer[8 * 1024];

   float *save;
   float *block;
   fft_complex_t *filter;
   fft_complex_t *fftblock;
   unsigned block_size;
   unsigned block_ptr;
};

struct eq_gain
{
   float freq;
   float gain; // Linear.
};

static void eq_free(void *data)
{
   struct eq_data *eq = (struct eq_data*)data;
   if (!eq)
      return;

   fft_free(eq->fft);
   free(eq->save);
   free(eq->block);
   free(eq->fftblock);
   free(eq->filter);
   free(eq);
}

static void eq_process(void *data, struct dspfilter_output *output,
      const struct dspfilter_input *input)
{
   struct eq_data *eq = (struct eq_data*)data;

   output->samples = eq->buffer;
   output->frames  = 0;

   float *out = eq->buffer;
   const float *in = input->samples;
   unsigned input_frames = input->frames;

   while (input_frames)
   {
      unsigned write_avail = eq->block_size - eq->block_ptr;
      if (input_frames < write_avail)
         write_avail = input_frames;

      memcpy(eq->block + eq->block_ptr * 2, in, write_avail * 2 * sizeof(float));

      in += write_avail * 2;
      input_frames -= write_avail;
      eq->block_ptr += write_avail;

      // Convolve a new block.
      if (eq->block_ptr == eq->block_size)
      {
         unsigned i, c;

         for (c = 0; c < 2; c++)
         {
            fft_process_forward(eq->fft, eq->fftblock, eq->block + c, 2);
            for (i = 0; i < 2 * eq->block_size; i++)
               eq->fftblock[i] = fft_complex_mul(eq->fftblock[i], eq->filter[i]);
            fft_process_inverse(eq->fft, out + c, eq->fftblock, 2);
         }

         // Overlap add method, so add in saved block now.
         for (i = 0; i < 2 * eq->block_size; i++)
            out[i] += eq->save[i];

         // Save block for later.
         memcpy(eq->save, out + 2 * eq->block_size, 2 * eq->block_size * sizeof(float));

         out += eq->block_size * 2;
         output->frames += eq->block_size;
         eq->block_ptr = 0;
      }
   }
}

static int gains_cmp(const void *a_, const void *b_)
{
   const struct eq_gain *a = (const struct eq_gain*)a_;
   const struct eq_gain *b = (const struct eq_gain*)b_;
   if (a->freq < b->freq)
      return -1;
   else if (a->freq > b->freq)
      return 1;
   else
      return 0;
}

static void generate_response(fft_complex_t *response,
      const struct eq_gain *gains, unsigned num_gains, unsigned samples)
{
   unsigned i;

   float start_freq = 0.0f;
   float start_gain = 1.0f;

   float end_freq = 1.0f;
   float end_gain = 1.0f;

   if (num_gains)
   {
      end_freq = gains->freq;
      end_gain = gains->gain;
      num_gains--;
      gains++;
   }

   // Create a response by linear interpolation between known frequency sample points.
   for (i = 0; i <= samples; i++)
   {
      float freq = (float)i / samples;

      while (freq >= end_freq)
      {
         if (num_gains)
         {
            start_freq = end_freq;
            start_gain = end_gain;
            end_freq = gains->freq;
            end_gain = gains->gain;

            gains++;
            num_gains--;
         }
         else
         {
            start_freq = end_freq;
            start_gain = end_gain;
            end_freq = 1.0f;
            end_gain = 1.0f;
            break;
         }
      }

      float lerp = 0.5f;
      // Edge case where i == samples.
      if (end_freq > start_freq)
         lerp = (freq - start_freq) / (end_freq - start_freq);
      float gain = (1.0f - lerp) * start_gain + lerp * end_gain;

      response[i].real = gain;
      response[i].imag = 0.0f;
      response[2 * samples - i].real = gain;
      response[2 * samples - i].imag = 0.0f;
   }
}

// Modified Bessel function of first order.
// Check Wiki for mathematical definition ...
static inline double kaiser_besseli0(double x)
{
   unsigned i;
   double sum = 0.0;

   double factorial = 1.0;
   double factorial_mult = 0.0;
   double x_pow = 1.0;
   double two_div_pow = 1.0;
   double x_sqr = x * x;

   // Approximate. This is an infinite sum.
   // Luckily, it converges rather fast.
   for (i = 0; i < 18; i++)
   {
      sum += x_pow * two_div_pow / (factorial * factorial);

      factorial_mult += 1.0;
      x_pow *= x_sqr;
      two_div_pow *= 0.25;
      factorial *= factorial_mult;
   }

   return sum;
}

static inline double kaiser_window(double index, double beta)
{
   return kaiser_besseli0(beta * sqrt(1 - index * index));
}

static void create_filter(struct eq_data *eq, unsigned size_log2,
      struct eq_gain *gains, unsigned num_gains, double beta, const char *filter_path)
{
   int i;
   int half_block_size = eq->block_size >> 1;
   double window_mod = 1.0 / kaiser_window(0.0, beta);

   fft_t *fft = fft_new(size_log2);
   float *time_filter = (float*)calloc(eq->block_size * 2 + 1, sizeof(*time_filter));
   if (!fft || !time_filter)
      goto end;

   // Make sure bands are in correct order.
   qsort(gains, num_gains, sizeof(*gains), gains_cmp);

   // Compute desired filter response.
   generate_response(eq->filter, gains, num_gains, half_block_size);

   // Get equivalent time-domain filter.
   fft_process_inverse(fft, time_filter, eq->filter, 1);

   // ifftshift() to create the correct linear phase filter.
   // The filter response was designed with zero phase, which won't work unless we compensate
   // for the repeating property of the FFT here by flipping left and right blocks.
   for (i = 0; i < half_block_size; i++)
   {
      float tmp = time_filter[i + half_block_size];
      time_filter[i + half_block_size] = time_filter[i];
      time_filter[i] = tmp;
   }

   // Apply a window to smooth out the frequency repsonse.
   for (i = 0; i < (int)eq->block_size; i++)
   {
      // Kaiser window.
      double phase = (double)i / eq->block_size;
      phase = 2.0 * (phase - 0.5);
      time_filter[i] *= window_mod * kaiser_window(phase, beta);
   }

   // Debugging.
   if (filter_path)
   {
      FILE *file = fopen(filter_path, "w");
      if (file)
      {
         for (i = 0; i < (int)eq->block_size - 1; i++)
            fprintf(file, "%.8f\n", time_filter[i + 1]);
         fclose(file);
      }
   }

   // Padded FFT to create our FFT filter.
   // Make our even-length filter odd by discarding the first coefficient.
   // For some interesting reason, this allows us to design an odd-length linear phase filter.
   fft_process_forward(eq->fft, eq->filter, time_filter + 1, 1);

end:
   fft_free(fft);
   free(time_filter);
}

static void *eq_init(const struct dspfilter_info *info,
      const struct dspfilter_config *config, void *userdata)
{
   unsigned i;
   struct eq_data *eq = (struct eq_data*)calloc(1, sizeof(*eq));
   if (!eq)
      return NULL;

   const float default_freq[] = { 0.0f, info->input_rate };
   const float default_gain[] = { 0.0f, 0.0f };

   float beta;
   config->get_float(userdata, "window_beta", &beta, 4.0f);

   int size_log2;
   config->get_int(userdata, "block_size_log2", &size_log2, 8);
   unsigned size = 1 << size_log2;

   struct eq_gain *gains = NULL;
   float *frequencies, *gain;
   unsigned num_freq, num_gain;
   config->get_float_array(userdata, "frequencies", &frequencies, &num_freq, default_freq, 2);
   config->get_float_array(userdata, "gains", &gain, &num_gain, default_gain, 2);

   char *filter_path = NULL;
   if (!config->get_string(userdata, "impulse_response_output", &filter_path, ""))
   {
      config->free(filter_path);
      filter_path = NULL;
   }

   num_gain = num_freq = min(num_gain, num_freq);

   gains = (struct eq_gain*)calloc(num_gain, sizeof(*gains));
   if (!gains)
      goto error;

   for (i = 0; i < num_gain; i++)
   {
      gains[i].freq = frequencies[i] / (0.5f * info->input_rate);
      gains[i].gain = pow(10.0, gain[i] / 20.0);
   }
   config->free(frequencies);
   config->free(gain);

   eq->block_size = size;

   eq->save     = (float*)calloc(    size, 2 * sizeof(*eq->save));
   eq->block    = (float*)calloc(2 * size, 2 * sizeof(*eq->block));
   eq->fftblock = (fft_complex_t*)calloc(2 * size, sizeof(*eq->fftblock));
   eq->filter   = (fft_complex_t*)calloc(2 * size, sizeof(*eq->filter));

   // Use an FFT which is twice the block size with zero-padding
   // to make circular convolution => proper convolution.
   eq->fft = fft_new(size_log2 + 1);

   if (!eq->fft || !eq->fftblock || !eq->save || !eq->block || !eq->filter)
      goto error;

   create_filter(eq, size_log2, gains, num_gain, beta, filter_path);
   config->free(filter_path);
   filter_path = NULL;

   free(gains);
   return eq;

error:
   free(gains);
   eq_free(eq);
   return NULL;
}

static const struct dspfilter_implementation eq_plug = {
   eq_init,
   eq_process,
   eq_free,

   DSPFILTER_API_VERSION,
   "Linear-Phase FFT Equalizer",
   "eq",
};

#ifdef HAVE_FILTERS_BUILTIN
#define dspfilter_get_implementation eq_dspfilter_get_implementation
#endif

const struct dspfilter_implementation *dspfilter_get_implementation(dspfilter_simd_mask_t mask)
{
   (void)mask;
   return &eq_plug;
}

#undef dspfilter_get_implementation