fftw_convolver.c

/*
 * (c) Copyright 2001 - 2004, 2006, 2009, 2013 -- Anders Torger
 *
 * This program is open source. For license terms, see the LICENSE file.
 *
 */
#include "defs.h"

#include <stdlib.h>
#include <stdio.h>
#include <errno.h>
#include <string.h>
#include <inttypes.h>
#include <math.h>
#ifdef __OS_SUNOS__
#include <ieeefp.h>
#endif

#include <fftw3.h>

#include "convolver.h"
#include "log2.h"
#include "emalloc.h"
#include "bfrun.h"
#include "dai.h"
#include "bit.h"
#include "swap.h"
#include "dither.h"
#include "asmprot.h"
#include "pinfo.h"
#include "inout.h"
#include "timestamp.h"
#include "numunion.h"

#define ifftplans fftplan_table[1][0]
#define ifftplans_inplace fftplan_table[1][1]
#define fftplans fftplan_table[0][0]
#define fftplans_inplace fftplan_table[0][1]
static void *fftplan_table[2][2][32];
static uint32_t fftplan_generated[2][2];
static int realsize = 0;

static int n_fft, n_fft2, fft_order;

#define OPT_CODE_GCC   0
#define OPT_CODE_SSE   1
#define OPT_CODE_SSE2  2
#define OPT_CODE_NEON  3
static int opt_code;

#if defined(__ARCH_IA32__) || defined(__ARCH_X86_64__)
static inline void
cpuid(uint32_t op,
      uint32_t *eax,
      uint32_t *ebx,
      uint32_t *ecx,
      uint32_t *edx)
{
    asm volatile ("cpuid" : "=a" (*eax), "=b" (*ebx), "=c" (*ecx),
		  "=d" (*edx) : "a" (op));
}
static void
decide_opt_code(void)
{
    uint32_t level, junk, cap;
    numunion_t vendor[2];

    opt_code = OPT_CODE_GCC;

    vendor[1].u8[4] = '\0';
    cpuid(0x00000000, &level, &vendor->u32[0], &vendor->u32[2],
          &vendor->u32[1]);
    if (strcmp((char *)vendor->u8, "GenuineIntel") == 0 &&
        level >= 0x00000001)
    {
	cpuid(0x00000001, &junk, &junk, &junk, &cap);
        if (realsize == 8) {
            if ((cap & (1 << 26)) != 0) {
                opt_code = OPT_CODE_SSE2;
                pinfo("SSE2 capability detected -- optimisation enabled.\n");
            }
	} else {
            if ((cap & (1 << 25)) != 0) {
                opt_code = OPT_CODE_SSE;
                pinfo("SSE capability detected -- optimisation enabled.\n");
            }
	}
    }
}
#else
static void
decide_opt_code(void)
{
#if defined(__ARCH_ARM__)
    opt_code = OPT_CODE_NEON;
#else
    opt_code = OPT_CODE_GCC;
#endif
}
#endif

static void *
create_fft_plan(int length,
                bool_t inplace,
                bool_t invert)
{
    void *plan, *buf[2];

    buf[0] = emallocaligned(length * realsize);
    memset(buf[0], 0, length * realsize);
    buf[1] = buf[0];
    if (!inplace) {
        buf[1] = emallocaligned(length * realsize);
        memset(buf[1], 0, length * realsize);
    }
    if (realsize == 4) {
        plan = fftwf_plan_r2r_1d(length, buf[0], buf[1],
                                 invert ? FFTW_HC2R : FFTW_R2HC,
                                 FFTW_MEASURE);
    } else {
        plan = fftw_plan_r2r_1d(length, buf[0], buf[1],
                                invert ? FFTW_HC2R : FFTW_R2HC,
                                FFTW_MEASURE);
    }
    efree(buf[0]);
    if (!inplace) {
        efree(buf[1]);
    }
    return plan;
}

#define real_t float
#define REALSIZE 4
#define RAW2REAL_NAME raw2realf
#define MIXNSCALE_NAME mixnscalef
#define CONVOLVE_INPLACE_NAME convolve_inplacef
#define CONVOLVE_NAME convolvef
#define CONVOLVE_ADD_NAME convolve_addf
#define DIRAC_CONVOLVE_INPLACE_NAME dirac_convolve_inplacef
#define DIRAC_CONVOLVE_NAME dirac_convolvef
#include "raw2real.h"
#include "fftw_convfuns.h"
#undef real_t
#undef REALSIZE
#undef RAW2REAL_NAME
#undef MIXNSCALE_NAME
#undef CONVOLVE_INPLACE_NAME
#undef CONVOLVE_NAME
#undef CONVOLVE_ADD_NAME
#undef DIRAC_CONVOLVE_INPLACE_NAME
#undef DIRAC_CONVOLVE_NAME

#define real_t double
#define REALSIZE 8
#define RAW2REAL_NAME raw2reald
#define MIXNSCALE_NAME mixnscaled
#define CONVOLVE_INPLACE_NAME convolve_inplaced
#define CONVOLVE_NAME convolved
#define CONVOLVE_ADD_NAME convolve_addd
#define DIRAC_CONVOLVE_INPLACE_NAME dirac_convolve_inplaced
#define DIRAC_CONVOLVE_NAME dirac_convolved
#include "raw2real.h"
#include "fftw_convfuns.h"
#undef real_t
#undef REALSIZE
#undef RAW2REAL_NAME
#undef MIXNSCALE_NAME
#undef CONVOLVE_INPLACE_NAME
#undef CONVOLVE_NAME
#undef CONVOLVE_ADD_NAME
#undef DIRAC_CONVOLVE_INPLACE_NAME
#undef DIRAC_CONVOLVE_NAME

void
convolver_raw2cbuf(void *rawbuf,
		   void *cbuf,
		   void *next_cbuf,
		   struct buffer_format *bf,
                   void (*postprocess)(void *realbuf,
                                       int n_samples,
                                       void *arg),
                   void *pp_arg)
{
    if (realsize == 4) {
        raw2realf(next_cbuf, (void *)&((uint8_t *)rawbuf)[bf->byte_offset],
                  bf->sf.bytes,
                  bf->sf.isfloat, bf->sample_spacing, bf->sf.swap, n_fft2);
    } else {
        raw2reald(next_cbuf, (void *)&((uint8_t *)rawbuf)[bf->byte_offset],
                  bf->sf.bytes,
                  bf->sf.isfloat, bf->sample_spacing, bf->sf.swap, n_fft2);
    }
    if (postprocess != NULL) {
        postprocess(next_cbuf, n_fft2, pp_arg);
    }
        
    memcpy(&((uint8_t *)cbuf)[n_fft2 * realsize], next_cbuf, n_fft2 * realsize);
}

void
convolver_time2freq(void *input_cbuf,
		    void *output_cbuf)
{
    void *fftplan;

    if (input_cbuf == output_cbuf) {
        fftplan = fftplans_inplace[fft_order];
    } else {
        fftplan = fftplans[fft_order];
    }
    if (realsize == 4) {
        fftwf_execute_r2r((const fftwf_plan)fftplan,
                          (float *)input_cbuf, (float *)output_cbuf);
    } else {
        fftw_execute_r2r((const fftw_plan)fftplan,
                         (double *)input_cbuf, (double *)output_cbuf);
    }
}

void
convolver_mixnscale(void *input_cbufs[],
                    void *output_cbuf,
                    double scales[],
                    int n_bufs,
                    int mixmode)
{
    if (realsize == 4) {
        mixnscalef(input_cbufs, output_cbuf, scales, n_bufs, mixmode);
    } else {
        mixnscaled(input_cbufs, output_cbuf, scales, n_bufs, mixmode);
    }
}

void
convolver_convolve_inplace(void *cbuf,
                           void *coeffs)
{
    if (realsize == 4) {
        convolve_inplacef(cbuf, coeffs);
    } else {
        convolve_inplaced(cbuf, coeffs);
    }
}

void
convolver_convolve(void *input_cbuf,
                   void *coeffs,
                   void *output_cbuf)
{
    if (realsize == 4) {
        convolvef(input_cbuf, coeffs, output_cbuf);
    } else {
        convolved(input_cbuf, coeffs, output_cbuf);
    }
}

void
convolver_convolve_add(void *input_cbuf,
		       void *coeffs,
		       void *output_cbuf)
{
    /*
#define real_t double
    real_t *_b = (real_t *)emallocaligned(n_fft * sizeof(real_t));
    real_t *_c = (real_t *)emallocaligned(n_fft * sizeof(real_t));
    real_t *_d = (real_t *)emallocaligned(n_fft * sizeof(real_t));

    memcpy(_b, input_cbuf, n_fft * sizeof(real_t));
    memcpy(_c, coeffs, n_fft * sizeof(real_t));
    memcpy(_d, output_cbuf, n_fft * sizeof(real_t));
    */
    switch (opt_code) {
#if defined(__ARCH_IA32__) || defined(__ARCH_X86_64__) || defined(__ARCH_ARM__)
    case OPT_CODE_NEON:
    case OPT_CODE_SSE:
	convolver_sse_convolve_add(input_cbuf, coeffs, output_cbuf,
                                   n_fft >> 3);
	break;
#ifdef __SSE2__
    case OPT_CODE_SSE2:
	convolver_sse2_convolve_add(input_cbuf, coeffs, output_cbuf,
                                    n_fft >> 3);
	break;
#endif
#endif	
    default:
    case OPT_CODE_GCC:
        if (realsize == 4) {
            convolve_addf(input_cbuf, coeffs, output_cbuf);
        } else {
            convolve_addd(input_cbuf, coeffs, output_cbuf);
        }
    }
    /*
    {
	real_t d1s, d2s, err, e;
	int n;
        
        d1s = _d[0] + _b[0] * _c[0];
        d2s = _d[4] + _b[4] * _c[4];
        for (n = 0; n < n_fft; n += 8) {
            _d[n+0] += _b[n+0] * _c[n+0] - _b[n+4] * _c[n+4];
            _d[n+1] += _b[n+1] * _c[n+1] - _b[n+5] * _c[n+5];
            _d[n+2] += _b[n+2] * _c[n+2] - _b[n+6] * _c[n+6];
            _d[n+3] += _b[n+3] * _c[n+3] - _b[n+7] * _c[n+7];
            
            _d[n+4] += _b[n+0] * _c[n+4] + _b[n+4] * _c[n+0];
            _d[n+5] += _b[n+1] * _c[n+5] + _b[n+5] * _c[n+1];
            _d[n+6] += _b[n+2] * _c[n+6] + _b[n+6] * _c[n+2];
            _d[n+7] += _b[n+3] * _c[n+7] + _b[n+7] * _c[n+3];
        }
        _d[0] = d1s;
        _d[4] = d2s;

        err = 0;
        for (n = 0; n < n_fft; n++) {
            e = (((real_t *)output_cbuf)[n] - _d[n]);
            if (e < 0.0) {
                err -= e;
            } else {
                err += e;
            }
        }
        if (err > 0.0) {
            fprintf(stderr, "err: %.10e\n", err);
        }
        efree(_b);
        efree(_c);
        efree(_d);
#undef real_t
    }
    */
}

void
convolver_crossfade_inplace(void *input_cbuf,
                            void *crossfade_cbuf,
                            void *buffer_cbuf)
{
    void *buf1, *buf2;
    double scale, d;
    float f;
    int n;

    buf1 = buffer_cbuf;
    buf2 = &((uint8_t *)buffer_cbuf)[n_fft * realsize];
    scale = 1.0;
    convolver_mixnscale(&crossfade_cbuf, buffer_cbuf, &scale, 1,
                        CONVOLVER_MIXMODE_OUTPUT);
    convolver_freq2time(buffer_cbuf, crossfade_cbuf);
    convolver_mixnscale(&input_cbuf, buffer_cbuf, &scale, 1,
                        CONVOLVER_MIXMODE_OUTPUT);
    convolver_freq2time(buffer_cbuf, buffer_cbuf);
    if (realsize == 4) {
        f = 1.0 / (float)(n_fft2 - 1);
        for (n = 0; n < n_fft2; n++) {
            ((float *)buffer_cbuf)[n] = 
                ((float *)crossfade_cbuf)[n] * (1.0 - f * (float)n) +
                ((float *)buffer_cbuf)[n] * f * (float)n;
        }
    } else {
        d = 1.0 / (double)(n_fft2 - 1);
        for (n = 0; n < n_fft2; n++) {
            ((double *)buf1)[n] = 
                ((double *)buf1)[n] * (1.0 - d * (double)n) +
                ((double *)buf2)[n] * d * (double)n;
        }
    }
    convolver_time2freq(buffer_cbuf, buffer_cbuf);
    scale = 1.0 / (double)n_fft;
    convolver_mixnscale(&buffer_cbuf, input_cbuf, &scale, 1,
                        CONVOLVER_MIXMODE_INPUT);
}
    
void
convolver_dirac_convolve_inplace(void *cbuf)
{
    if (realsize == 4) {
        dirac_convolve_inplacef(cbuf);
    } else {
        dirac_convolve_inplaced(cbuf);
    }
}

void
convolver_dirac_convolve(void *input_cbuf,
                         void *output_cbuf)
{
    if (realsize == 4) {
        dirac_convolvef(input_cbuf, output_cbuf);
    } else {
        dirac_convolved(input_cbuf, output_cbuf);
    }
}

void
convolver_freq2time(void *input_cbuf,
		    void *output_cbuf)
{
    void *ifftplan;

    if (input_cbuf == output_cbuf) {
        ifftplan = ifftplans_inplace[fft_order];
    } else {
        ifftplan = ifftplans[fft_order];
    }
    if (realsize == 4) {
        fftwf_execute_r2r((const fftwf_plan)ifftplan,
                          (float *)input_cbuf, (float *)output_cbuf);
    } else {
        fftw_execute_r2r((const fftw_plan)ifftplan,
                         (double *)input_cbuf, (double *)output_cbuf);
    }
}

void
convolver_convolve_eval(void *input_cbuf,
			void *buffer_cbuf, /* 1.5 x size */
			void *output_cbuf)
{
    if (realsize == 4) {
        fftwf_execute_r2r((const fftwf_plan)ifftplans[fft_order],
                          (float *)input_cbuf,
                          &((float *)buffer_cbuf)[n_fft2]);
        fftwf_execute_r2r((const fftwf_plan)fftplans[fft_order],
                          (float *)buffer_cbuf,
                          (float *)output_cbuf);
    } else {
        fftw_execute_r2r((const fftw_plan)ifftplans[fft_order],
                         (double *)input_cbuf,
                         &((double *)buffer_cbuf)[n_fft2]);
        fftw_execute_r2r((const fftw_plan)fftplans[fft_order],
                         (double *)buffer_cbuf,
                         (double *)output_cbuf);
    }
    memcpy(buffer_cbuf, &((uint8_t *)buffer_cbuf)[n_fft2 * realsize],
	   n_fft2 * realsize);
}

#define real_t float
#define REALSIZE 4
#define REAL2RAW_NAME real2rawf_hp_tpdf
#define REAL2INT_CALL ditherf_real2int_hp_tpdf(((float *)realbuf)[n], rmin,    \
                                               rmax, imin, imax, overflow,     \
                                               dither_state, n)
#define REAL2RAW_EXTRA_PARAMS , struct dither_state *dither_state
#include "real2raw.h"
#undef REAL2RAW_NAME
#undef REAL2INT_CALL
#undef REAL2RAW_EXTRA_PARAMS

#define REAL2RAW_NAME real2rawf_no_dither
#define REAL2INT_CALL ditherd_real2int_no_dither(((float *)realbuf)[n], rmin,  \
                                                 rmax, imin, imax, overflow)
#define REAL2RAW_EXTRA_PARAMS 
#include "real2raw.h"
#undef REAL2RAW_NAME
#undef REAL2INT_CALL
#undef REAL2RAW_EXTRA_PARAMS
#undef REALSIZE
#undef real_t

#define real_t double
#define REALSIZE 8
#define REAL2RAW_NAME real2rawd_hp_tpdf
#define REAL2INT_CALL ditherd_real2int_hp_tpdf(((double *)realbuf)[n], rmin,   \
                                               rmax, imin, imax, overflow,     \
                                               dither_state, n)
#define REAL2RAW_EXTRA_PARAMS , struct dither_state *dither_state
#include "real2raw.h"
#undef REAL2RAW_NAME
#undef REAL2INT_CALL
#undef REAL2RAW_EXTRA_PARAMS

#define REAL2RAW_NAME real2rawd_no_dither
#define REAL2INT_CALL ditherd_real2int_no_dither(((double *)realbuf)[n], rmin, \
                                                 rmax, imin, imax, overflow)
#define REAL2RAW_EXTRA_PARAMS 
#include "real2raw.h"
#undef REAL2RAW_NAME
#undef REAL2INT_CALL
#undef REAL2RAW_EXTRA_PARAMS
#undef REALSIZE
#undef real_t


void
convolver_cbuf2raw(void *cbuf,
		   void *outbuf,
		   struct buffer_format *bf,
		   bool_t apply_dither,
		   void *dither_state,
		   struct bfoverflow *overflow)
{
    if (realsize == 4) {
        if (apply_dither && !bf->sf.isfloat) {
            dither_preloop_real2int_hp_tpdf(dither_state, n_fft2);
            real2rawf_hp_tpdf((void *)&((uint8_t *)outbuf)[bf->byte_offset],
                              cbuf, bf->sf.sbytes << 3, bf->sf.bytes,
                              bf->sf.isfloat, bf->sample_spacing,
                              bf->sf.swap, n_fft2, overflow, dither_state);
        } else {
            real2rawf_no_dither((void *)&((uint8_t *)outbuf)[bf->byte_offset],
                                cbuf, bf->sf.sbytes << 3, bf->sf.bytes,
                                bf->sf.isfloat, bf->sample_spacing,
                                bf->sf.swap, n_fft2, overflow);
        }
    } else {
        if (apply_dither && !bf->sf.isfloat) {
            dither_preloop_real2int_hp_tpdf(dither_state, n_fft2);
            real2rawd_hp_tpdf((void *)&((uint8_t *)outbuf)[bf->byte_offset],
                              cbuf, bf->sf.sbytes << 3, bf->sf.bytes,
                              bf->sf.isfloat, bf->sample_spacing,
                              bf->sf.swap, n_fft2, overflow, dither_state);
        } else {
            real2rawd_no_dither((void *)&((uint8_t *)outbuf)[bf->byte_offset],
                                cbuf, bf->sf.sbytes << 3,
                                bf->sf.bytes,
                                bf->sf.isfloat, bf->sample_spacing,
                                bf->sf.swap, n_fft2, overflow);
        }
    }
}

int
convolver_cbufsize(void)
{
    return n_fft * realsize;
}

void *
convolver_coeffs2cbuf(void *coeffs,
		      int n_coeffs,
		      double scale,
                      void *optional_dest)
{
    void *rcoeffs, *coeffs_data;
    int n, len;

    len = (n_coeffs > n_fft2) ? n_fft2 : n_coeffs;
    rcoeffs = emallocaligned(n_fft * realsize);
    memset(rcoeffs, 0, n_fft * realsize);

    if (realsize == 4) {
        for (n = 0; n < len; n++) {
            ((float *)rcoeffs)[n_fft2 + n] = ((float *)coeffs)[n] *
                (float)scale;
            if (!finite((double)((float *)rcoeffs)[n_fft2 + n])) {
                fprintf(stderr, "NaN or Inf value among coefficients.\n");
                return NULL;
            }
        }
        fftwf_execute_r2r((const fftwf_plan)fftplans_inplace[fft_order],
                          (float *)rcoeffs, (float *)rcoeffs);
    } else {
        for (n = 0; n < len; n++) {
            ((double *)rcoeffs)[n_fft2 + n] = ((double *)coeffs)[n] * scale;
            if (!finite(((double *)rcoeffs)[n_fft2 + n])) {
                fprintf(stderr, "NaN or Inf value among coefficients.\n");
                return NULL;
            }
        }
        fftw_execute_r2r((const fftw_plan)fftplans_inplace[fft_order],
                         (double *)rcoeffs, (double *)rcoeffs);
    }

    scale = 1.0 / (double)n_fft;
    if (optional_dest != NULL) {
        coeffs_data = optional_dest;
    } else {
        coeffs_data = emallocaligned(n_fft * realsize);
    }
    convolver_mixnscale(&rcoeffs, coeffs_data, &scale, 1,
			CONVOLVER_MIXMODE_INPUT);
    efree(rcoeffs);
    
    return coeffs_data;
}

void
convolver_runtime_coeffs2cbuf(void *src,  /* nfft / 2 */
                              void *dest) /* nfft */
{
    static void *tmp = NULL;
    double scale;
    
    if (tmp == NULL) {
        tmp = emallocaligned(n_fft * realsize);
    }
    memset(dest, 0, n_fft2 * realsize);
    memcpy(&((uint8_t *)dest)[n_fft2 * realsize], src, n_fft2 * realsize);
    if (realsize == 4) {
        fftwf_execute_r2r((const fftwf_plan)fftplans[fft_order],
                          (float *)dest, (float *)tmp);
    } else {
        fftw_execute_r2r((const fftw_plan)fftplans[fft_order],
                         (double *)dest, (double *)tmp);
    }
    scale = 1.0 / (double)n_fft;
    convolver_mixnscale(&tmp, dest, &scale, 1, CONVOLVER_MIXMODE_INPUT);
}

bool_t
convolver_verify_cbuf(void *cbufs[],
                      int n_cbufs)
{
    int n, i;
    
    for (n = 0; n < n_cbufs; n++) {
        if (realsize == 4) {
            for (i = 0; i < n_fft; i++) {
                if (!finite((double)((float *)cbufs[n])[i])) {
                    fprintf(stderr, "NaN or Inf value among coefficients.\n");
                    return false;
                }
            }
        } else {
            for (i = 0; i < n_fft; i++) {
                if (!finite(((double *)cbufs[n])[i])) {
                    fprintf(stderr, "NaN or Inf value among coefficients.\n");
                    return false;
                }
            }
        }
    }
    return true;
}

void
convolver_debug_dump_cbuf(const char filename[],
                          void *cbufs[],
                          int n_cbufs)
{
    void *coeffs;
    double scale;
    FILE *stream;
    int n, i;

    if ((stream = fopen(filename, "wt")) == NULL) {
        fprintf(stderr, "Could not open \"%s\" for writing: %s", filename,
                strerror(errno));
        return;
    }
    coeffs = emallocaligned(n_fft * realsize);
    scale = 1.0;    
    for (n = 0; n < n_cbufs; n++) {
        convolver_mixnscale(&cbufs[n], coeffs, &scale, 1,
                            CONVOLVER_MIXMODE_OUTPUT);
        if (realsize == 4) {
            fftwf_execute_r2r((const fftwf_plan)ifftplans_inplace[fft_order],
                              (float *)coeffs, (float *)coeffs);
            for (i = 0; i < n_fft2; i++) {
                fprintf(stream, "%.16e\n", ((float *)coeffs)[n_fft2 + i]);
            }
        } else {
            fftw_execute_r2r((const fftw_plan)ifftplans_inplace[fft_order],
                             (double *)coeffs, (double *)coeffs);
            for (i = 0; i < n_fft2; i++) {
                fprintf(stream, "%.16e\n", ((double *)coeffs)[n_fft2 + i]);
            }
        }
    }
    efree(coeffs);
    fclose(stream);
}

void *
convolver_fftplan(int order,
                  int invert,
                  int inplace)
{
    invert = !!invert;
    inplace = !!inplace;
    if (!bit_isset(&fftplan_generated[invert][inplace], order)) {
        pinfo("Creating %s%sFFTW plan of size %d using wisdom...",
              invert ? "inverse " : "forward ",
              inplace ? "inplace " : "",
              1 << order);
        fftplan_table[invert][inplace][order] =
            create_fft_plan(1 << order, inplace, invert);
        pinfo("finished\n");
        bit_set(&fftplan_generated[invert][inplace], order);
    }
    return fftplan_table[invert][inplace][order];
}

struct _td_conv_t_ {
    void *fftplan;
    void *ifftplan;    
    void *coeffs;
    int blocklen;
};

int
convolver_td_block_length(int n_coeffs)
{
    if (n_coeffs < 1) {
        return -1;
    }
    return 1 << log2_roof(n_coeffs);
}

td_conv_t *
convolver_td_new(void *coeffs,
                 int n_coeffs)
{
    int blocklen, n;
    td_conv_t *tdc;
    double scaled;
    float scalef;
    
    if ((blocklen = convolver_td_block_length(n_coeffs)) == -1) {
        return NULL;
    }
    tdc = emalloc(sizeof(td_conv_t));
    memset(tdc, 0, sizeof(td_conv_t));
    tdc->fftplan = convolver_fftplan(log2_get(blocklen) + 1, false, true);
    tdc->ifftplan = convolver_fftplan(log2_get(blocklen) + 1, true, true);
    tdc->coeffs = emallocaligned(2 * blocklen * realsize);
    tdc->blocklen = blocklen;

    memset(tdc->coeffs, 0, blocklen * realsize);
    memset(&((uint8_t *)tdc->coeffs)[(blocklen + n_coeffs) * realsize], 0,
           (blocklen - n_coeffs) * realsize);
    memcpy(&((uint8_t *)tdc->coeffs)[blocklen * realsize], coeffs,
           n_coeffs * realsize);
    if (realsize == 4) {
        fftwf_execute_r2r(tdc->fftplan, tdc->coeffs, tdc->coeffs);
        scalef = 1.0 / (float)(blocklen << 1);
        for (n = 0; n < blocklen << 1; n++) {
            ((float *)tdc->coeffs)[n] *= scalef;
        }
    } else {
        fftw_execute_r2r(tdc->fftplan, tdc->coeffs, tdc->coeffs);
        scaled = 1.0 / (double)(blocklen << 1);
        for (n = 0; n < blocklen << 1; n++) {
            ((double *)tdc->coeffs)[n] *= scaled;
        }
    }
    return tdc;
}

static void
convolve_inplace_ordered(void *cbuf,
                         void *coeffs,
                         int size)
{
    int n, size2 = size >> 1;
    if (realsize == 4) {
        float a, *b = (float *)cbuf, *c = (float *)coeffs;
        
        b[0] *= c[0];
        for (n = 1; n < size2; n++) {
            a = b[n];
            b[n] = a * c[n] - b[size - n] * c[size - n];
            b[size - n] = a * c[size - n] + b[size - n] * c[n];
        }
        b[size2] *= c[size2];
    } else {
        double a, *b = (double *)cbuf, *c = (double *)coeffs;
        
        b[0] *= c[0];
        for (n = 1; n < size2; n++) {
            a = b[n];
            b[n] = a * c[n] - b[size - n] * c[size - n];
            b[size - n] = a * c[size - n] + b[size - n] * c[n];
        }
        b[size2] *= c[size2];
    }
}

void
convolver_td_convolve(td_conv_t *tdc,
                      void *overlap_block)
{
    if (realsize == 4) {
        fftwf_execute_r2r(tdc->fftplan, overlap_block, overlap_block);
        convolve_inplace_ordered(overlap_block, tdc->coeffs,
                                 tdc->blocklen << 1);
        fftwf_execute_r2r(tdc->ifftplan, overlap_block, overlap_block);
    } else {
        fftw_execute_r2r(tdc->fftplan, overlap_block, overlap_block);
        convolve_inplace_ordered(overlap_block, tdc->coeffs,
                                 tdc->blocklen << 1);
        fftw_execute_r2r(tdc->ifftplan, overlap_block, overlap_block);
    }
}

bool_t
convolver_init(const char config_filename[],
	       int length,
               int _realsize)
{
    int order;
    FILE *stream;
    bool_t quiet;

    realsize = _realsize;
    decide_opt_code();

    if (realsize != 4 && realsize != 8) {
	fprintf(stderr, "Invalid real size %d.\n", realsize);
	return false;
    }
    if ((order = log2_get(length)) == -1) {
	fprintf(stderr, "Invalid length %d.\n", length);
	return false;
    }
    order++;
    fft_order = order;
    n_fft = 2 * length;
    n_fft2 = length;

    if ((stream = fopen(config_filename, "rt")) == NULL) {
	if (errno != ENOENT) {
	    fprintf(stderr, "Could not open \"%s\" for reading: %s.\n",
		    config_filename, strerror(errno));
	    return false;
	}
    } else {
        /* We ignore if we can't read wisdom, we just overwrite with new */
        if (realsize == 4) {
            fftwf_import_wisdom_from_file(stream);
        } else {
            fftw_import_wisdom_from_file(stream);
        }
        fclose(stream);
    }

    memset(fftplan_generated, 0, sizeof(fftplan_generated));
    pinfo("Creating 4 FFTW plans of size %d...", 1 << fft_order);
    quiet = bfconf->quiet;
    bfconf->quiet = true;
    convolver_fftplan(fft_order, false, false);
    convolver_fftplan(fft_order, false, true);
    convolver_fftplan(fft_order, true, false);
    convolver_fftplan(fft_order, true, true);
    bfconf->quiet = quiet;
    pinfo("finished.\n");

    /* Wisdom is cumulative, save it each time (and get wiser) */
    if ((stream = fopen(config_filename, "wt")) == NULL) {
        fprintf(stderr, "Warning: could not save wisdom:\n"
                "  could not open \"%s\" for writing: %s.\n",
                config_filename, strerror(errno));
    } else {
        if (realsize == 4) {
            fftwf_export_wisdom_to_file(stream);
        } else {
            fftw_export_wisdom_to_file(stream);
        }
        fclose(stream);
    }
    
    return true;
}