RetroArch/audio/sinc.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/>.
 */

// Bog-standard windowed SINC implementation.
// Only suitable as an upsampler, as cutoff frequency isn't dynamically configurable (yet).

#include "resampler.h"
#include "../libretro.h"
#include "../performance.h"
#include <math.h>
#include <stdint.h>
#include <stdlib.h>
#include <string.h>
#include <stdio.h>
#include "../msvc/msvc_compat.h"

#ifndef RESAMPLER_TEST
#include "../general.h"
#else
#define RARCH_LOG(...) fprintf(stderr, __VA_ARGS__)
#endif

#ifdef __SSE__
#include <xmmintrin.h>
#endif

// Rough SNR values for upsampling:
// LOWEST: 40 dB
// LOWER: 55 dB
// NORMAL: 70 dB
// HIGHER: 110 dB
// HIGHEST: 140 dB

// TODO, make all this more configurable.
#if defined(SINC_LOWEST_QUALITY)
#define SINC_WINDOW_LANCZOS
#define CUTOFF 0.98
#define PHASE_BITS 12
#define SINC_COEFF_LERP 0
#define SUBPHASE_BITS 10
#define SIDELOBES 2
#define ENABLE_AVX 0
#elif defined(SINC_LOWER_QUALITY)
#define SINC_WINDOW_LANCZOS
#define CUTOFF 0.98
#define PHASE_BITS 12
#define SUBPHASE_BITS 10
#define SINC_COEFF_LERP 0
#define SIDELOBES 4
#define ENABLE_AVX 0
#elif defined(SINC_HIGHER_QUALITY)
#define SINC_WINDOW_KAISER
#define SINC_WINDOW_KAISER_BETA 10.5
#define CUTOFF 0.90
#define PHASE_BITS 10
#define SUBPHASE_BITS 14
#define SINC_COEFF_LERP 1
#define SIDELOBES 32
#define ENABLE_AVX 1
#elif defined(SINC_HIGHEST_QUALITY)
#define SINC_WINDOW_KAISER
#define SINC_WINDOW_KAISER_BETA 14.5
#define CUTOFF 0.962
#define PHASE_BITS 10
#define SUBPHASE_BITS 14
#define SINC_COEFF_LERP 1
#define SIDELOBES 128
#define ENABLE_AVX 1
#else
#define SINC_WINDOW_KAISER
#define SINC_WINDOW_KAISER_BETA 5.5
#define CUTOFF 0.825
#define PHASE_BITS 8
#define SUBPHASE_BITS 16
#define SINC_COEFF_LERP 1
#define SIDELOBES 8
#define ENABLE_AVX 0
#endif

// For the little amount of taps we're using,
// SSE1 is faster than AVX for some reason.
// AVX code is kept here though as by increasing number
// of sinc taps, the AVX code is clearly faster than SSE1.

#if defined(__AVX__) && ENABLE_AVX
#include <immintrin.h>
#endif

#define PHASES (1 << (PHASE_BITS + SUBPHASE_BITS))

#define TAPS (SIDELOBES * 2)
#define SUBPHASE_MASK ((1 << SUBPHASE_BITS) - 1)
#define SUBPHASE_MOD (1.0f / (1 << SUBPHASE_BITS))

typedef struct rarch_sinc_resampler
{
   float *phase_table;
   float *buffer_l;
   float *buffer_r;

   unsigned taps;

   unsigned ptr;
   uint32_t time;

   // A buffer for phase_table, buffer_l and buffer_r are created in a single calloc().
   // Ensure that we get as good cache locality as we can hope for.
   float *main_buffer;
} rarch_sinc_resampler_t;

static inline double sinc(double val)
{
   if (fabs(val) < 0.00001)
      return 1.0;
   else
      return sin(val) / val;
}

#if defined(SINC_WINDOW_LANCZOS)
static inline double window_function(double index)
{
   return sinc(M_PI * index);
}
#elif defined(SINC_WINDOW_KAISER)
// Modified Bessel function of first order.
// Check Wiki for mathematical definition ...
static inline double 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 window_function(double index)
{
   return besseli0(SINC_WINDOW_KAISER_BETA * sqrt(1 - index * index));
}
#else
#error "No SINC window function defined."
#endif

static void init_sinc_table(rarch_sinc_resampler_t *resamp, double cutoff,
      float *phase_table, int phases, int taps, bool calculate_delta)
{
   int i, j, p;
   double window_mod = window_function(0.0); // Need to normalize w(0) to 1.0.
   int stride = calculate_delta ? 2 : 1;

   double sidelobes = taps / 2.0;
   for (i = 0; i < phases; i++)
   {
      for (j = 0; j < taps; j++)
      {
         int n = j * phases + i;
         double window_phase = (double)n / (phases * taps); // [0, 1).
         window_phase = 2.0 * window_phase - 1.0; // [-1, 1)
         double sinc_phase = sidelobes * window_phase;

         float val = cutoff * sinc(M_PI * sinc_phase * cutoff) * window_function(window_phase) / window_mod;
         phase_table[i * stride * taps + j] = val;
      }
   }

   if (calculate_delta)
   {
      for (p = 0; p < phases - 1; p++)
      {
         for (j = 0; j < taps; j++)
         {
            float delta = phase_table[(p + 1) * stride * taps + j] - phase_table[p * stride * taps + j];
            phase_table[(p * stride + 1) * taps + j] = delta;
         }
      }

      int phase = phases - 1;
      for (j = 0; j < taps; j++)
      {
         int n = j * phases + (phase + 1);
         double window_phase = (double)n / (phases * taps); // (0, 1].
         window_phase = 2.0 * window_phase - 1.0; // (-1, 1]
         double sinc_phase = sidelobes * window_phase;

         float val = cutoff * sinc(M_PI * sinc_phase * cutoff) * window_function(window_phase) / window_mod;
         float delta = (val - phase_table[phase * stride * taps + j]);
         phase_table[(phase * stride + 1) * taps + j] = delta;
      }
   }
}

// No memalign() for us on Win32 ...
static void *aligned_alloc__(size_t boundary, size_t size)
{
   void *ptr = malloc(boundary + size + sizeof(uintptr_t));
   if (!ptr)
      return NULL;

   uintptr_t addr = ((uintptr_t)ptr + sizeof(uintptr_t) + boundary) & ~(boundary - 1);
   void **place   = (void**)addr;
   place[-1]      = ptr;

   return (void*)addr;
}

static void aligned_free__(void *ptr)
{
   void **p = (void**)ptr;
   free(p[-1]);
}

#if !(defined(__AVX__) && ENABLE_AVX) && !defined(__SSE__)
static inline void process_sinc_C(rarch_sinc_resampler_t *resamp, float *out_buffer)
{
   unsigned i;
   float sum_l = 0.0f;
   float sum_r = 0.0f;
   const float *buffer_l = resamp->buffer_l + resamp->ptr;
   const float *buffer_r = resamp->buffer_r + resamp->ptr;

   unsigned taps  = resamp->taps;
   unsigned phase = resamp->time >> SUBPHASE_BITS;
#if SINC_COEFF_LERP
   const float *phase_table = resamp->phase_table + phase * taps * 2;
   const float *delta_table = phase_table + taps;
   float delta = (float)(resamp->time & SUBPHASE_MASK) * SUBPHASE_MOD;
#else
   const float *phase_table = resamp->phase_table + phase * taps;
#endif

   for (i = 0; i < taps; i++)
   {
#if SINC_COEFF_LERP
      float sinc_val = phase_table[i] + delta_table[i] * delta;
#else
      float sinc_val = phase_table[i];
#endif
      sum_l         += buffer_l[i] * sinc_val;
      sum_r         += buffer_r[i] * sinc_val;
   }

   out_buffer[0] = sum_l;
   out_buffer[1] = sum_r;
}
#endif

#if defined(__AVX__) && ENABLE_AVX
#define process_sinc_func process_sinc
static void process_sinc(rarch_sinc_resampler_t *resamp, float *out_buffer)
{
   unsigned i;
   __m256 sum_l = _mm256_setzero_ps();
   __m256 sum_r = _mm256_setzero_ps();

   const float *buffer_l = resamp->buffer_l + resamp->ptr;
   const float *buffer_r = resamp->buffer_r + resamp->ptr;

   unsigned taps = resamp->taps;
   unsigned phase = resamp->time >> SUBPHASE_BITS;
#if SINC_COEFF_LERP
   const float *phase_table = resamp->phase_table + phase * taps * 2;
   const float *delta_table = phase_table + taps;
   __m256 delta = _mm256_set1_ps((float)(resamp->time & SUBPHASE_MASK) * SUBPHASE_MOD);
#else
   const float *phase_table = resamp->phase_table + phase * taps;
#endif

   for (i = 0; i < taps; i += 8)
   {
      __m256 buf_l = _mm256_loadu_ps(buffer_l + i);
      __m256 buf_r = _mm256_loadu_ps(buffer_r + i);

#if SINC_COEFF_LERP
      __m256 deltas = _mm256_load_ps(delta_table + i);
      __m256 sinc = _mm256_add_ps(_mm256_load_ps(phase_table + i), _mm256_mul_ps(deltas, delta));
#else
      __m256 sinc = _mm256_load_ps(phase_table + i);
#endif
      sum_l       = _mm256_add_ps(sum_l, _mm256_mul_ps(buf_l, sinc));
      sum_r       = _mm256_add_ps(sum_r, _mm256_mul_ps(buf_r, sinc));
   }

   // hadd on AVX is weird, and acts on low-lanes and high-lanes separately.
   __m256 res_l = _mm256_hadd_ps(sum_l, sum_l);
   __m256 res_r = _mm256_hadd_ps(sum_r, sum_r);
   res_l = _mm256_hadd_ps(res_l, res_l);
   res_r = _mm256_hadd_ps(res_r, res_r);
   res_l = _mm256_add_ps(_mm256_permute2f128_ps(res_l, res_l, 1), res_l);
   res_r = _mm256_add_ps(_mm256_permute2f128_ps(res_r, res_r, 1), res_r);

   // This is optimized to mov %xmmN, [mem].
   // There doesn't seem to be any _mm256_store_ss intrinsic.
   _mm_store_ss(out_buffer + 0, _mm256_extractf128_ps(res_l, 0));
   _mm_store_ss(out_buffer + 1, _mm256_extractf128_ps(res_r, 0));
}
#elif defined(__SSE__)
#define process_sinc_func process_sinc
static void process_sinc(rarch_sinc_resampler_t *resamp, float *out_buffer)
{
   unsigned i;
   __m128 sum_l = _mm_setzero_ps();
   __m128 sum_r = _mm_setzero_ps();

   const float *buffer_l = resamp->buffer_l + resamp->ptr;
   const float *buffer_r = resamp->buffer_r + resamp->ptr;

   unsigned taps = resamp->taps;
   unsigned phase = resamp->time >> SUBPHASE_BITS;
#if SINC_COEFF_LERP
   const float *phase_table = resamp->phase_table + phase * taps * 2;
   const float *delta_table = phase_table + taps;
   __m128 delta = _mm_set1_ps((float)(resamp->time & SUBPHASE_MASK) * SUBPHASE_MOD);
#else
   const float *phase_table = resamp->phase_table + phase * taps;
#endif

   for (i = 0; i < taps; i += 4)
   {
      __m128 buf_l = _mm_loadu_ps(buffer_l + i);
      __m128 buf_r = _mm_loadu_ps(buffer_r + i);

#if SINC_COEFF_LERP
      __m128 deltas = _mm_load_ps(delta_table + i);
      __m128 sinc = _mm_add_ps(_mm_load_ps(phase_table + i), _mm_mul_ps(deltas, delta));
#else
      __m128 sinc = _mm_load_ps(phase_table + i);
#endif
      sum_l       = _mm_add_ps(sum_l, _mm_mul_ps(buf_l, sinc));
      sum_r       = _mm_add_ps(sum_r, _mm_mul_ps(buf_r, sinc));
   }

   // Them annoying shuffles :V
   // sum_l = { l3, l2, l1, l0 }
   // sum_r = { r3, r2, r1, r0 }

   __m128 sum = _mm_add_ps(_mm_shuffle_ps(sum_l, sum_r, _MM_SHUFFLE(1, 0, 1, 0)),
         _mm_shuffle_ps(sum_l, sum_r, _MM_SHUFFLE(3, 2, 3, 2)));

   // sum   = { r1, r0, l1, l0 } + { r3, r2, l3, l2 }
   // sum   = { R1, R0, L1, L0 }

   sum = _mm_add_ps(_mm_shuffle_ps(sum, sum, _MM_SHUFFLE(3, 3, 1, 1)), sum);

   // sum   = {R1, R1, L1, L1 } + { R1, R0, L1, L0 }
   // sum   = { X,  R,  X,  L } 

   // Store L
   _mm_store_ss(out_buffer + 0, sum);

   // movehl { X, R, X, L } == { X, R, X, R }
   _mm_store_ss(out_buffer + 1, _mm_movehl_ps(sum, sum));
}
#elif defined(HAVE_NEON)

#if SINC_COEFF_LERP
#error "NEON asm does not support SINC lerp."
#endif

// Need to make this function pointer as Android doesn't have built-in targets
// for NEON and plain ARMv7a.
static void (*process_sinc_func)(rarch_sinc_resampler_t *resamp, float *out_buffer);

// Assumes that taps >= 8, and that taps is a multiple of 8.
void process_sinc_neon_asm(float *out, const float *left, const float *right, const float *coeff, unsigned taps);

static void process_sinc_neon(rarch_sinc_resampler_t *resamp, float *out_buffer)
{
   const float *buffer_l = resamp->buffer_l + resamp->ptr;
   const float *buffer_r = resamp->buffer_r + resamp->ptr;

   unsigned phase = resamp->time >> SUBPHASE_BITS;
   unsigned taps = resamp->taps;
   const float *phase_table = resamp->phase_table + phase * taps;

   process_sinc_neon_asm(out_buffer, buffer_l, buffer_r, phase_table, taps);
}
#else // Plain ol' C99
#define process_sinc_func process_sinc_C
#endif

static void resampler_sinc_process(void *re_, struct resampler_data *data)
{
   rarch_sinc_resampler_t *re = (rarch_sinc_resampler_t*)re_;

   uint32_t ratio = PHASES / data->ratio;

   const float *input = data->data_in;
   float *output      = data->data_out;
   size_t frames         = data->input_frames;
   size_t out_frames     = 0;

   while (frames)
   {
      while (frames && re->time >= PHASES)
      {
         // Push in reverse to make filter more obvious.
         if (!re->ptr)
            re->ptr = re->taps;
         re->ptr--;

         re->buffer_l[re->ptr + re->taps] = re->buffer_l[re->ptr] = *input++;
         re->buffer_r[re->ptr + re->taps] = re->buffer_r[re->ptr] = *input++;

         re->time -= PHASES;
         frames--;
      }

      while (re->time < PHASES)
      {
         process_sinc_func(re, output);
         output += 2;
         out_frames++;
         re->time += ratio;
      }
   }

   data->output_frames = out_frames;
}

static void resampler_sinc_free(void *re)
{
   rarch_sinc_resampler_t *resampler = (rarch_sinc_resampler_t*)re;
   if (resampler)
      aligned_free__(resampler->main_buffer);
   free(resampler);
}

static void *resampler_sinc_new(double bandwidth_mod)
{
   rarch_sinc_resampler_t *re = (rarch_sinc_resampler_t*)calloc(1, sizeof(*re));
   if (!re)
      return NULL;

   memset(re, 0, sizeof(*re));

   re->taps = TAPS;
   double cutoff = CUTOFF;

   // Downsampling, must lower cutoff, and extend number of taps accordingly to keep same stopband attenuation.
   if (bandwidth_mod < 1.0)
   {
      cutoff *= bandwidth_mod;
      re->taps = (unsigned)ceil(re->taps / bandwidth_mod);
   }

   // Be SIMD-friendly.
#if (defined(__AVX__) && ENABLE_AVX) || defined(HAVE_NEON)
   re->taps = (re->taps + 7) & ~7;
#else
   re->taps = (re->taps + 3) & ~3;
#endif

   size_t phase_elems = (1 << PHASE_BITS) * re->taps;
#if SINC_COEFF_LERP
   phase_elems *= 2;
#endif
   size_t elems = phase_elems + 4 * re->taps;

   re->main_buffer = (float*)aligned_alloc__(128, sizeof(float) * elems);
   if (!re->main_buffer)
      goto error;

   re->phase_table = re->main_buffer;
   re->buffer_l = re->main_buffer + phase_elems;
   re->buffer_r = re->buffer_l + 2 * re->taps;

   init_sinc_table(re, cutoff, re->phase_table, 1 << PHASE_BITS, re->taps, SINC_COEFF_LERP);

#if defined(__AVX__) && ENABLE_AVX
   RARCH_LOG("Sinc resampler [AVX]\n");
#elif defined(__SSE__)
   RARCH_LOG("Sinc resampler [SSE]\n");
#elif defined(HAVE_NEON)
   unsigned cpu = rarch_get_cpu_features();
   process_sinc_func = cpu & RETRO_SIMD_NEON ? process_sinc_neon : process_sinc_C;
   RARCH_LOG("Sinc resampler [%s]\n", cpu & RETRO_SIMD_NEON ? "NEON" : "C");
#else
   RARCH_LOG("Sinc resampler [C]\n");
#endif

   RARCH_LOG("SINC params (%u phase bits, %u taps).\n", PHASE_BITS, re->taps);
   return re;

error:
   resampler_sinc_free(re);
   return NULL;
}

const rarch_resampler_t sinc_resampler = {
   resampler_sinc_new,
   resampler_sinc_process,
   resampler_sinc_free,
   "sinc",
};