static void cook_imlt(COOKContext *q, float* inbuffer, float* outbuffer,
float* mlt_tmp){
int i;
/* prerotation */
for(i=0 ; i<q->mlt_size ; i+=2){
outbuffer[i] = (q->mlt_presin[i/2] * inbuffer[q->mlt_size-1-i]) +
(q->mlt_precos[i/2] * inbuffer[i]);
outbuffer[i+1] = (q->mlt_precos[i/2] * inbuffer[q->mlt_size-1-i]) -
(q->mlt_presin[i/2] * inbuffer[i]);
}
/* FFT */
ff_fft_permute(&q->fft_ctx, (FFTComplex *) outbuffer);
ff_fft_calc (&q->fft_ctx, (FFTComplex *) outbuffer);
/* postrotation */
for(i=0 ; i<q->mlt_size ; i+=2){
mlt_tmp[i] = (q->mlt_postcos[(q->mlt_size-1-i)/2] * outbuffer[i+1]) +
(q->mlt_postcos[i/2] * outbuffer[i]);
mlt_tmp[q->mlt_size-1-i] = (q->mlt_postcos[(q->mlt_size-1-i)/2] * outbuffer[i]) -
(q->mlt_postcos[i/2] * outbuffer[i+1]);
}
/* window and reorder */
for(i=0 ; i<q->mlt_size/2 ; i++){
outbuffer[i] = mlt_tmp[q->mlt_size/2-1-i] * q->mlt_window[i];
outbuffer[q->mlt_size-1-i]= mlt_tmp[q->mlt_size/2-1-i] *
q->mlt_window[q->mlt_size-1-i];
outbuffer[q->mlt_size+i]= mlt_tmp[q->mlt_size/2+i] *
q->mlt_window[q->mlt_size-1-i];
outbuffer[2*q->mlt_size-1-i]= -(mlt_tmp[q->mlt_size/2+i] *
q->mlt_window[i]);
}
}
static void imdct_c(MDCTContext *s, const FFTSample *input, FFTSample *tmp)
{
int k, n4, n2, n, j;
const uint16_t *revtab = s->fft.revtab;
const FFTSample *tcos = s->tcos;
const FFTSample *tsin = s->tsin;
const FFTSample *in1, *in2;
FFTComplex *z = (FFTComplex *)tmp;
n = 1 << s->nbits;
n2 = n >> 1;
n4 = n >> 2;
/* pre rotation */
in1 = input;
in2 = input + n2 - 1;
for(k = 0; k < n4; k++) {
j=revtab[k];
CMUL(z[j].re, z[j].im, *in2, *in1, tcos[k], tsin[k]);
in1 += 2;
in2 -= 2;
}
ff_fft_calc(&s->fft, z);
/* post rotation + reordering */
/* XXX: optimize */
for(k = 0; k < n4; k++) {
CMUL(z[k].re, z[k].im, z[k].re, z[k].im, tcos[k], tsin[k]);
}
}
static void imc_imdct256(IMCContext *q) {
int i;
float re, im;
/* prerotation */
for(i=0; i < COEFFS/2; i++) {
q->samples[i].re = -(q->pre_coef1[i] * q->CWdecoded[COEFFS-1-i*2]) -
(q->pre_coef2[i] * q->CWdecoded[i*2]);
q->samples[i].im = (q->pre_coef2[i] * q->CWdecoded[COEFFS-1-i*2]) -
(q->pre_coef1[i] * q->CWdecoded[i*2]);
}
/* FFT */
ff_fft_permute(&q->fft, q->samples);
ff_fft_calc (&q->fft, q->samples);
/* postrotation, window and reorder */
for(i = 0; i < COEFFS/2; i++) {
re = (q->samples[i].re * q->post_cos[i]) + (-q->samples[i].im * q->post_sin[i]);
im = (-q->samples[i].im * q->post_cos[i]) - (q->samples[i].re * q->post_sin[i]);
q->out_samples[i*2] = (q->mdct_sine_window[COEFFS-1-i*2] * q->last_fft_im[i]) + (q->mdct_sine_window[i*2] * re);
q->out_samples[COEFFS-1-i*2] = (q->mdct_sine_window[i*2] * q->last_fft_im[i]) - (q->mdct_sine_window[COEFFS-1-i*2] * re);
q->last_fft_im[i] = im;
}
}
/**
* Compute inverse MDCT of size N = 2^nbits
* @param output N samples
* @param input N/2 samples
* @param tmp N/2 samples
*/
void ff_imdct_calc(MDCTContext *s, FFTSample *output,
const FFTSample *input, FFTSample *tmp)
{
int k, n8, n4, n2, n, j;
const uint16_t *revtab = s->fft.revtab;
const FFTSample *tcos = s->tcos;
const FFTSample *tsin = s->tsin;
const FFTSample *in1, *in2;
FFTComplex *z = (FFTComplex *)tmp;
n = 1 << s->nbits;
n2 = n >> 1;
n4 = n >> 2;
n8 = n >> 3;
/* pre rotation */
in1 = input;
in2 = input + n2 - 1;
for(k = 0; k < n4; k++) {
j=revtab[k];
CMUL(z[j].re, z[j].im, *in2, *in1, tcos[k], tsin[k]);
in1 += 2;
in2 -= 2;
}
ff_fft_calc(&s->fft, z);
/* post rotation + reordering */
/* XXX: optimize */
for(k = 0; k < n4; k++) {
CMUL(z[k].re, z[k].im, z[k].re, z[k].im, tcos[k], tsin[k]);
}
for(k = 0; k < n8; k++) {
output[2*k] = -z[n8 + k].im;
output[n2-1-2*k] = z[n8 + k].im;
output[2*k+1] = z[n8-1-k].re;
output[n2-1-2*k-1] = -z[n8-1-k].re;
output[n2 + 2*k]=-z[k+n8].re;
output[n-1- 2*k]=-z[k+n8].re;
output[n2 + 2*k+1]=z[n8-k-1].im;
output[n-2 - 2 * k] = z[n8-k-1].im;
}
}
/**
* Compute MDCT of size N = 2^nbits
* @param input N samples
* @param out N/2 samples
*/
void ff_mdct_calc_c(FFTContext *s, FFTSample *out, const FFTSample *input)
{
int i, j, n, n8, n4, n2, n3;
FFTSample re, im;
const uint16_t *revtab = s->revtab;
const FFTSample *tcos = s->tcos;
const FFTSample *tsin = s->tsin;
FFTComplex *x = (FFTComplex *)out;
n = 1 << s->mdct_bits;
n2 = n >> 1;
n4 = n >> 2;
n8 = n >> 3;
n3 = 3 * n4;
/* pre rotation */
for(i=0;i<n8;i++) {
re = -input[2*i+3*n4] - input[n3-1-2*i];
im = -input[n4+2*i] + input[n4-1-2*i];
j = revtab[i];
CMUL(x[j].re, x[j].im, re, im, -tcos[i], tsin[i]);
re = input[2*i] - input[n2-1-2*i];
im = -(input[n2+2*i] + input[n-1-2*i]);
j = revtab[n8 + i];
CMUL(x[j].re, x[j].im, re, im, -tcos[n8 + i], tsin[n8 + i]);
}
ff_fft_calc(s, x);
/* post rotation */
for(i=0;i<n8;i++) {
FFTSample r0, i0, r1, i1;
CMUL(i1, r0, x[n8-i-1].re, x[n8-i-1].im, -tsin[n8-i-1], -tcos[n8-i-1]);
CMUL(i0, r1, x[n8+i ].re, x[n8+i ].im, -tsin[n8+i ], -tcos[n8+i ]);
x[n8-i-1].re = r0;
x[n8-i-1].im = i0;
x[n8+i ].re = r1;
x[n8+i ].im = i1;
}
}
/**
* Compute MDCT of size N = 2^nbits
* @param input N samples
* @param out N/2 samples
* @param tmp temporary storage of N/2 samples
*/
void ff_mdct_calc(MDCTContext *s, FFTSample *out,
const FFTSample *input, FFTSample *tmp)
{
int i, j, n, n8, n4, n2, n3;
FFTSample re, im, re1, im1;
const uint16_t *revtab = s->fft.revtab;
const FFTSample *tcos = s->tcos;
const FFTSample *tsin = s->tsin;
FFTComplex *x = (FFTComplex *)tmp;
n = 1 << s->nbits;
n2 = n >> 1;
n4 = n >> 2;
n8 = n >> 3;
n3 = 3 * n4;
/* pre rotation */
for(i=0;i<n8;i++) {
re = -input[2*i+3*n4] - input[n3-1-2*i];
im = -input[n4+2*i] + input[n4-1-2*i];
j = revtab[i];
CMUL(x[j].re, x[j].im, re, im, -tcos[i], tsin[i]);
re = input[2*i] - input[n2-1-2*i];
im = -(input[n2+2*i] + input[n-1-2*i]);
j = revtab[n8 + i];
CMUL(x[j].re, x[j].im, re, im, -tcos[n8 + i], tsin[n8 + i]);
}
ff_fft_calc(&s->fft, x);
/* post rotation */
for(i=0;i<n4;i++) {
re = x[i].re;
im = x[i].im;
CMUL(re1, im1, re, im, -tsin[i], -tcos[i]);
out[2*i] = im1;
out[n2-1-2*i] = re1;
}
}
static void ff_dct_calc_c(DCTContext *s, FFTSample *data)
{
int n = 1<<s->nbits;
int i;
#define ROTATE(i,n) (-M_PI*((n)-0.5f)*(i)/(n))
if (s->inverse) {
for(i=0; i < n; i++) {
s->data[i].re = (float) (2 * data[i] * cos(ROTATE(i,n)));
s->data[i].im = (float) (2 * data[i] * sin(ROTATE(i,n)));
}
s->data[n].re = 0;
s->data[n].im = 0;
for(i=0; i<n-1; i++) {
s->data[n+i+1].re = (float) (-2 * data[n - (i+1)] * cos(ROTATE(n+i+1,n)));
s->data[n+i+1].im = (float) (-2 * data[n - (i+1)] * sin(ROTATE(n+i+1,n)));
}
}else{
for(i=0; i < n; i++) {
s->data[i].re = data[n - (i+1)];
s->data[i].im = 0;
s->data[n+i].re = data[i];
s->data[n+i].im = 0;
}
}
ff_fft_permute(&s->fft, s->data);
ff_fft_calc(&s->fft, s->data);
if (s->inverse) {
for(i=0; i < n; i++)
data[i] = s->data[n-(i+1)].re / (2 * n);
}else {
for(i=0; i < n; i++)
data[i] = (float) (s->data[i].re / (2 * cos(ROTATE(i,n))));
}
#undef ROTATE
}
/**
* Compute the middle half of the inverse MDCT of size N = 2^nbits,
* thus excluding the parts that can be derived by symmetry
* @param output N/2 samples
* @param input N/2 samples
*/
void ff_imdct_half_c(MDCTContext *s, FFTSample *output, const FFTSample *input)
{
int k, n8, n4, n2, n, j;
const uint16_t *revtab = s->fft.revtab;
const FFTSample *tcos = s->tcos;
const FFTSample *tsin = s->tsin;
const FFTSample *in1, *in2;
FFTComplex *z = (FFTComplex *)output;
n = 1 << s->nbits;
n2 = n >> 1;
n4 = n >> 2;
n8 = n >> 3;
/* pre rotation */
in1 = input;
in2 = input + n2 - 1;
for(k = 0; k < n4; k++) {
j=revtab[k];
CMUL(z[j].re, z[j].im, *in2, *in1, tcos[k], tsin[k]);
in1 += 2;
in2 -= 2;
}
ff_fft_calc(&s->fft, z);
/* post rotation + reordering */
output += n4;
for(k = 0; k < n8; k++) {
FFTSample r0, i0, r1, i1;
CMUL(r0, i1, z[n8-k-1].im, z[n8-k-1].re, tsin[n8-k-1], tcos[n8-k-1]);
CMUL(r1, i0, z[n8+k ].im, z[n8+k ].re, tsin[n8+k ], tcos[n8+k ]);
z[n8-k-1].re = r0;
z[n8-k-1].im = i0;
z[n8+k ].re = r1;
z[n8+k ].im = i1;
}
}
int main(int argc, char **argv)
{
FFTComplex *tab, *tab1, *tab_ref;
FFTSample *tab2;
int it, i, c;
int do_speed = 0;
int do_mdct = 0;
int do_inverse = 0;
FFTContext s1, *s = &s1;
MDCTContext m1, *m = &m1;
int fft_nbits, fft_size;
fft_nbits = 9;
for(;;) {
c = getopt(argc, argv, "hsimn:");
if (c == -1)
break;
switch(c) {
case 'h':
help();
break;
case 's':
do_speed = 1;
break;
case 'i':
do_inverse = 1;
break;
case 'm':
do_mdct = 1;
break;
case 'n':
fft_nbits = atoi(optarg);
break;
}
}
fft_size = 1 << fft_nbits;
tab = av_malloc(fft_size * sizeof(FFTComplex));
tab1 = av_malloc(fft_size * sizeof(FFTComplex));
tab_ref = av_malloc(fft_size * sizeof(FFTComplex));
tab2 = av_malloc(fft_size * sizeof(FFTSample));
if (do_mdct) {
if (do_inverse)
av_log(NULL, AV_LOG_INFO,"IMDCT");
else
av_log(NULL, AV_LOG_INFO,"MDCT");
ff_mdct_init(m, fft_nbits, do_inverse);
} else {
if (do_inverse)
av_log(NULL, AV_LOG_INFO,"IFFT");
else
av_log(NULL, AV_LOG_INFO,"FFT");
ff_fft_init(s, fft_nbits, do_inverse);
fft_ref_init(fft_nbits, do_inverse);
}
av_log(NULL, AV_LOG_INFO," %d test\n", fft_size);
/* generate random data */
for(i=0;i<fft_size;i++) {
tab1[i].re = frandom();
tab1[i].im = frandom();
}
/* checking result */
av_log(NULL, AV_LOG_INFO,"Checking...\n");
if (do_mdct) {
if (do_inverse) {
imdct_ref((float *)tab_ref, (float *)tab1, fft_nbits);
ff_imdct_calc(m, tab2, (float *)tab1);
check_diff((float *)tab_ref, tab2, fft_size);
} else {
mdct_ref((float *)tab_ref, (float *)tab1, fft_nbits);
ff_mdct_calc(m, tab2, (float *)tab1);
check_diff((float *)tab_ref, tab2, fft_size / 2);
}
} else {
memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
ff_fft_permute(s, tab);
ff_fft_calc(s, tab);
fft_ref(tab_ref, tab1, fft_nbits);
check_diff((float *)tab_ref, (float *)tab, fft_size * 2);
}
/* do a speed test */
if (do_speed) {
int64_t time_start, duration;
int nb_its;
av_log(NULL, AV_LOG_INFO,"Speed test...\n");
/* we measure during about 1 seconds */
nb_its = 1;
for(;;) {
time_start = gettime();
for(it=0;it<nb_its;it++) {
if (do_mdct) {
if (do_inverse) {
ff_imdct_calc(m, (float *)tab, (float *)tab1);
} else {
ff_mdct_calc(m, (float *)tab, (float *)tab1);
}
} else {
memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
ff_fft_calc(s, tab);
}
}
duration = gettime() - time_start;
if (duration >= 1000000)
break;
nb_its *= 2;
}
av_log(NULL, AV_LOG_INFO,"time: %0.1f us/transform [total time=%0.2f s its=%d]\n",
(double)duration / nb_its,
(double)duration / 1000000.0,
nb_its);
}
if (do_mdct) {
ff_mdct_end(m);
} else {
ff_fft_end(s);
}
return 0;
}
int main(int argc, char **argv)
{
FFTComplex *tab, *tab1, *tab_ref;
FFTSample *tab2;
int it, i, c;
int do_speed = 0;
int err = 1;
enum tf_transform transform = TRANSFORM_FFT;
int do_inverse = 0;
FFTContext s1, *s = &s1;
FFTContext m1, *m = &m1;
RDFTContext r1, *r = &r1;
DCTContext d1, *d = &d1;
int fft_nbits, fft_size, fft_size_2;
double scale = 1.0;
AVLFG prng;
av_lfg_init(&prng, 1);
fft_nbits = 9;
for(;;) {
c = getopt(argc, argv, "hsimrdn:f:");
if (c == -1)
break;
switch(c) {
case 'h':
help();
break;
case 's':
do_speed = 1;
break;
case 'i':
do_inverse = 1;
break;
case 'm':
transform = TRANSFORM_MDCT;
break;
case 'r':
transform = TRANSFORM_RDFT;
break;
case 'd':
transform = TRANSFORM_DCT;
break;
case 'n':
fft_nbits = atoi(optarg);
break;
case 'f':
scale = atof(optarg);
break;
}
}
fft_size = 1 << fft_nbits;
fft_size_2 = fft_size >> 1;
tab = av_malloc(fft_size * sizeof(FFTComplex));
tab1 = av_malloc(fft_size * sizeof(FFTComplex));
tab_ref = av_malloc(fft_size * sizeof(FFTComplex));
tab2 = av_malloc(fft_size * sizeof(FFTSample));
switch (transform) {
case TRANSFORM_MDCT:
av_log(NULL, AV_LOG_INFO,"Scale factor is set to %f\n", scale);
if (do_inverse)
av_log(NULL, AV_LOG_INFO,"IMDCT");
else
av_log(NULL, AV_LOG_INFO,"MDCT");
ff_mdct_init(m, fft_nbits, do_inverse, scale);
break;
case TRANSFORM_FFT:
if (do_inverse)
av_log(NULL, AV_LOG_INFO,"IFFT");
else
av_log(NULL, AV_LOG_INFO,"FFT");
ff_fft_init(s, fft_nbits, do_inverse);
fft_ref_init(fft_nbits, do_inverse);
break;
case TRANSFORM_RDFT:
if (do_inverse)
av_log(NULL, AV_LOG_INFO,"IDFT_C2R");
else
av_log(NULL, AV_LOG_INFO,"DFT_R2C");
ff_rdft_init(r, fft_nbits, do_inverse ? IDFT_C2R : DFT_R2C);
fft_ref_init(fft_nbits, do_inverse);
break;
case TRANSFORM_DCT:
if (do_inverse)
av_log(NULL, AV_LOG_INFO,"DCT_III");
else
av_log(NULL, AV_LOG_INFO,"DCT_II");
ff_dct_init(d, fft_nbits, do_inverse ? DCT_III : DCT_II);
break;
}
av_log(NULL, AV_LOG_INFO," %d test\n", fft_size);
/* generate random data */
for (i = 0; i < fft_size; i++) {
tab1[i].re = frandom(&prng);
tab1[i].im = frandom(&prng);
}
/* checking result */
av_log(NULL, AV_LOG_INFO,"Checking...\n");
switch (transform) {
case TRANSFORM_MDCT:
if (do_inverse) {
imdct_ref((float *)tab_ref, (float *)tab1, fft_nbits);
ff_imdct_calc(m, tab2, (float *)tab1);
err = check_diff((float *)tab_ref, tab2, fft_size, scale);
} else {
mdct_ref((float *)tab_ref, (float *)tab1, fft_nbits);
ff_mdct_calc(m, tab2, (float *)tab1);
err = check_diff((float *)tab_ref, tab2, fft_size / 2, scale);
}
break;
case TRANSFORM_FFT:
memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
ff_fft_permute(s, tab);
ff_fft_calc(s, tab);
fft_ref(tab_ref, tab1, fft_nbits);
err = check_diff((float *)tab_ref, (float *)tab, fft_size * 2, 1.0);
break;
case TRANSFORM_RDFT:
if (do_inverse) {
tab1[ 0].im = 0;
tab1[fft_size_2].im = 0;
for (i = 1; i < fft_size_2; i++) {
tab1[fft_size_2+i].re = tab1[fft_size_2-i].re;
tab1[fft_size_2+i].im = -tab1[fft_size_2-i].im;
}
memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
tab2[1] = tab1[fft_size_2].re;
ff_rdft_calc(r, tab2);
fft_ref(tab_ref, tab1, fft_nbits);
for (i = 0; i < fft_size; i++) {
tab[i].re = tab2[i];
tab[i].im = 0;
}
err = check_diff((float *)tab_ref, (float *)tab, fft_size * 2, 0.5);
} else {
for (i = 0; i < fft_size; i++) {
tab2[i] = tab1[i].re;
tab1[i].im = 0;
}
ff_rdft_calc(r, tab2);
fft_ref(tab_ref, tab1, fft_nbits);
tab_ref[0].im = tab_ref[fft_size_2].re;
err = check_diff((float *)tab_ref, (float *)tab2, fft_size, 1.0);
}
break;
case TRANSFORM_DCT:
memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
ff_dct_calc(d, tab);
if (do_inverse) {
idct_ref(tab_ref, tab1, fft_nbits);
} else {
dct_ref(tab_ref, tab1, fft_nbits);
}
err = check_diff((float *)tab_ref, (float *)tab, fft_size, 1.0);
break;
}
/* do a speed test */
if (do_speed) {
int64_t time_start, duration;
int nb_its;
av_log(NULL, AV_LOG_INFO,"Speed test...\n");
/* we measure during about 1 seconds */
nb_its = 1;
for(;;) {
time_start = gettime();
for (it = 0; it < nb_its; it++) {
switch (transform) {
case TRANSFORM_MDCT:
if (do_inverse) {
ff_imdct_calc(m, (float *)tab, (float *)tab1);
} else {
ff_mdct_calc(m, (float *)tab, (float *)tab1);
}
break;
case TRANSFORM_FFT:
memcpy(tab, tab1, fft_size * sizeof(FFTComplex));
ff_fft_calc(s, tab);
break;
case TRANSFORM_RDFT:
memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
ff_rdft_calc(r, tab2);
break;
case TRANSFORM_DCT:
memcpy(tab2, tab1, fft_size * sizeof(FFTSample));
ff_dct_calc(d, tab2);
break;
}
}
duration = gettime() - time_start;
if (duration >= 1000000)
break;
nb_its *= 2;
}
av_log(NULL, AV_LOG_INFO,"time: %0.1f us/transform [total time=%0.2f s its=%d]\n",
(double)duration / nb_its,
(double)duration / 1000000.0,
nb_its);
}
switch (transform) {
case TRANSFORM_MDCT:
ff_mdct_end(m);
break;
case TRANSFORM_FFT:
ff_fft_end(s);
break;
case TRANSFORM_RDFT:
ff_rdft_end(r);
break;
case TRANSFORM_DCT:
ff_dct_end(d);
break;
}
av_free(tab);
av_free(tab1);
av_free(tab2);
av_free(tab_ref);
av_free(exptab);
return err;
}
