Commit a29bd934 authored by vaclav's avatar vaclav
Browse files

Merge branch 'OPTIMIZE_FFT_STACK' into 'main' 

FFT: reduce stack and harmonize functions

See merge request !2884
parents 7543c759 a5d8c358

lib_com/basop_util.c

+60 −0
Original line number Diff line number Diff line
@@ -714,6 +714,66 @@ Word16 getScaleFactor16( /* o: measured headroom in range [ 
    return i;

}



#ifdef OPTIMIZE_FFT_STACK

/* o: measured headroom in range [0..31], 0 if all x[i] == 0 */

Word16 getScaleFactor32_cmplx(

    cmplx *x,          /* i: array containing 32-bit data */

    const Word16 len_x /* i: length of the array to scan  */

)

{

    Word16 i, i_min, i_max, i_re, i_im;

    Word32 x_min_re, x_max_re, x_min_im, x_max_im;



    x_max_re = 0;

    move32();

    x_min_re = 0;

    move32();

    x_max_im = 0;

    move32();

    x_min_im = 0;

    move32();

    FOR( i = 0; i < len_x; i++ )

    {

        if ( x[i].re >= 0 )

            x_max_re = L_max( x_max_re, x[i].re );

        if ( x[i].re < 0 )

            x_min_re = L_min( x_min_re, x[i].re );

        if ( x[i].im >= 0 )

            x_max_im = L_max( x_max_im, x[i].im );

        if ( x[i].im < 0 )

            x_min_im = L_min( x_min_im, x[i].im );

    }



    i_max = 0x20;

    move16();

    i_min = 0x20;

    move16();



    if ( x_max_re != 0 )

        i_max = norm_l( x_max_re );



    if ( x_min_re != 0 )

        i_min = norm_l( x_min_re );



    i_re = s_and( s_min( i_max, i_min ), 0x1F );



    i_max = 0x20;

    move16();

    i_min = 0x20;

    move16();



    if ( x_max_im != 0 )

        i_max = norm_l( x_max_im );



    if ( x_min_im != 0 )

        i_min = norm_l( x_min_im );



    i_im = s_and( s_min( i_max, i_min ), 0x1F );



    return s_min( i_re, i_im );

}

#endif





/********************************************************************/

/*!

lib_com/basop_util.h

+12 −3
Original line number Diff line number Diff line
@@ -216,10 +216,19 @@ void BASOP_Util_Sqrt_InvSqrt_MantExp( Word16 mantissa, /*!< mantissa */ 
    and   -32768 <= x <= -16384 for negative x

*/



Word16 getScaleFactor16(                       /* o: measured headroom in range [0..15], 0 if all x[i] == 0 */

/* o: measured headroom in range [0..15], 0 if all x[i] == 0 */

Word16 getScaleFactor16(

    const Word16 *x,      /* i: array containing 16-bit data */

    const Word16 len_x ); /* i: length of the array to scan  */



#ifdef OPTIMIZE_FFT_STACK

/* o: measured headroom in range [0..31], 0 if all x[i] == 0 */

Word16 getScaleFactor32_cmplx(

    cmplx *x,          /* i: array containing 32-bit data */

    const Word16 len_x /* i: length of the array to scan  */

);

#endif



/********************************************************************/

/*!

  \brief   Calculates the scalefactor needed to normalize input array

lib_com/edct_fx.c

+316 −26

File changed.

Preview size limit exceeded, changes collapsed.

lib_com/fft_fx.c

+130 −2
Original line number Diff line number Diff line
@@ -109,11 +109,13 @@ static void fft5_8( Word16 n1, Word32 *zRe, Word32 *zIm, const Word16 *Idx ); 
static void fft4_5( Word32 *x, Word32 *y, const Word16 *Idx );

static void fft5_4( Word16 n1, Word32 *zRe, Word32 *zIm, const Word16 *Idx );



#ifndef HARMONIZE_DCT

void DoRTFTn_fx_ivas(

    Word32 *x,     /* i/o: real part of input and output data       */

    Word32 *y,     /* i/o: imaginary part of input and output data  */

    const Word16 n /* i  : size of the FFT n=(2^k) up to 1024       */

);

#endif

/*-----------------------------------------------------------------*

 * fft15_shift2()

 * 15-point FFT with 2-point circular shift
@@ -2438,6 +2440,7 @@ static void cftmdl( 
    return;

}



#ifndef HARMONIZE_DCT

static void cftbsub(

    Word16 n,       // Q0

    Word32 *a,      // Qx
@@ -2733,6 +2736,7 @@ void edct2_fx_ivas( 
        }

    }

}

#endif



void DoRTFTn_fx_ivas(

    Word32 *x,     /* i/o: real part of input and output data Qx      */
@@ -2740,7 +2744,6 @@ void DoRTFTn_fx_ivas( 
    const Word16 n /* i  : size of the FFT up to 1024 Q0*/

)

{



    Word16 i;

    Word32 z[2048];
@@ -6427,6 +6430,7 @@ static void fft_lenN( 
 * Complex-value FFT

 *-----------------------------------------------------------------*/



#ifndef HARMONIZE_DCT

void fft_fx(

    Word32 *re,          /* i/o: real part Qx          */

    Word32 *im,          /* i/o: imag part Qx          */
@@ -6434,6 +6438,7 @@ void fft_fx( 
    const Word16 s       /* i  : sign                */

)

{



    cmplx x[960];



    FOR( Word16 j = 0; j < length; j++ )
@@ -6511,6 +6516,73 @@ void fft_fx( 


    return;

}

#else

void fft_fx(

    cmplx *x,           /* i/o: complex data        */

    const Word16 length /* i  : length of fft       */

)

{

    SWITCH( length )

    {

        case 20:

            fft_len20_fx( x );

            BREAK;

        case 40:

            fft_lenN( x, FFT_RotVector_640_fx, 640, 5, 8, 8, 40 );

            BREAK;

        case 64:

            fft_lenN( x, FFT_RotVector_256_fx, 256, 8, 8, 8, 64 );

            BREAK;

        case 80:

            fft_lenN( x, FFT_RotVector_640_fx, 640, 10, 8, 4, 40 );

            BREAK;

        case 100:

            fft_lenN( x, FFT_RotVector_400_fx, 400, 10, 10, 4, 40 );

            BREAK;

        case 120:

            fft_lenN( x, FFT_RotVector_960_fx, 960, 15, 8, 4, 60 );

            BREAK;

        case 128:

            fft_lenN( x, FFT_RotVector_256_fx, 256, 16, 8, 4, 64 );

            BREAK;

        case 160:

            fft_lenN( x, FFT_RotVector_640_fx, 640, 20, 8, 2, 40 );

            BREAK;

        case 200:

            fft_lenN( x, FFT_RotVector_400_fx, 400, 20, 10, 2, 40 );

            BREAK;

        case 240:

            fft_lenN( x, FFT_RotVector_960_fx, 960, 30, 8, 2, 60 );

            BREAK;

        case 256:

            fft_lenN( x, FFT_RotVector_256_fx, 256, 32, 8, 2, 64 );

            BREAK;

        case 320:

            fft_lenN( x, FFT_RotVector_640_fx, 640, 20, 16, 2, 40 );

            BREAK;

        case 400:

            fft_lenN( x, FFT_RotVector_400_fx, 400, 20, 20, 2, 40 );

            BREAK;

        case 480:

            fft_lenN( x, FFT_RotVector_960_fx, 960, 30, 16, 2, 60 );

            BREAK;

        case 600:

            fft_lenN( x, FFT_RotVector_600_fx, 600, 30, 20, 2, 60 );

            BREAK;

        case 640:

            fft_lenN( x, FFT_RotVector_640_fx, 640, 20, 32, 2, 40 );

            BREAK;

        case 960:

            fft_lenN( x, FFT_RotVector_960_fx, 960, 30, 32, 2, 60 );

            BREAK;

        default:

            assert( !"fft length is not supported!" );

    }



    return;

}

#endif





void rfft_fx(

    Word32 *x,           /* i/o: values Qx                   */
@@ -6522,6 +6594,9 @@ void rfft_fx( 
    Word16 i, sizeOfFft2, sizeOfFft4;

    Word32 tmp, t1, t2, t3, t4;

    Word16 s1, s2;

#ifdef HARMONIZE_DCT

    cmplx spec[L_FRAME48k];

#endif



    sizeOfFft2 = shr( length, 1 );

    sizeOfFft4 = shr( length, 2 );
@@ -6592,10 +6667,43 @@ void rfft_fx( 


    SWITCH( isign )

    {



        case -1:



#ifdef HARMONIZE_DCT

            FOR( i = 0; i < sizeOfFft2; i++ )

            {

                spec[i].re = x[2 * i];

                move32();

                spec[i].im = x[2 * i + 1];

                move32();

            }



            fft_fx( spec, sizeOfFft2 );



            FOR( i = 0; i < sizeOfFft4; i++ )

            {

                x[2 * i] = spec[i].re;

                move32();

                x[2 * i + 1] = spec[sizeOfFft2 - i - 1].re;

                move32();



                x[2 * i] = spec[i].im;

                move32();

                x[2 * i + 1] = L_negate( spec[sizeOfFft2 - i - 1].im );

                move32();

            }



            FOR( i = 0; i < sizeOfFft2; i++ )

            {

                x[2 * i] = spec[i].re;

                move32();

                x[2 * i + 1] = spec[i].im;

                move32();

            }

#else

            fft_fx( x, x + 1, sizeOfFft2, 2 );

#endif



            // Qx

            tmp = L_add( x[0], x[1] );

            x[1] = L_sub( x[0], x[1] ); // Qx
@@ -6651,7 +6759,27 @@ void rfft_fx( 
                move32();

            }



#ifdef HARMONIZE_DCT

            FOR( i = 0; i < sizeOfFft2; i++ )

            {

                spec[i].re = x[2 * i];

                move32();

                spec[i].im = x[2 * i + 1];

                move32();

            }



            fft_fx( spec, sizeOfFft2 );



            FOR( i = 0; i < sizeOfFft2; i++ )

            {

                x[2 * i] = spec[i].re;

                move32();

                x[2 * i + 1] = spec[i].im;

                move32();

            }

#else

            fft_fx( x, x + 1, sizeOfFft2, 2 );

#endif



            FOR( i = 0; i < length; i += 2 )

            {

lib_com/fft_fx_evs.c

+73 −24
Original line number Diff line number Diff line
@@ -10,7 +10,7 @@ 
#include <assert.h>



/*-----------------------------------------------------------------*

 * Local functions

 * Local constants

 *-----------------------------------------------------------------*/



#define FFT3_ONE_THIRD 21845 /* 1/3 in Q16 */
@@ -19,6 +19,10 @@ 
#define KP951056516_16FX 2042378325 /* EDCT & EMDCT constants Q31*/

#define KP587785252_16FX 1262259213 /* EDCT & EMDCT constants Q31*/



/*-----------------------------------------------------------------*

 * Local function prototypes

 *-----------------------------------------------------------------*/



static void fft5_shift4_16fx( Word16 n1, Word16 *zRe, Word16 *zIm, const Word16 *Idx );

static void fft64_16fx( Word16 *x, Word16 *y, const Word16 *Idx );

static void fft32_5_16fx( Word16 *x, Word16 *y, const Word16 *Idx );
@@ -32,11 +36,6 @@ static void fft64_16fx( Word16 *x, Word16 *y, const Word16 *Idx ); 
static void fft5_32_16fx( Word16 *zRe, Word16 *zIm, const Word16 *Idx );

static void cdftForw_16fx( Word16 n, Word16 *a, const Word16 *ip, const Word32 *w );



#include "math_32.h"



/*-----------------------------------------------------------------*

 * Local functions

 *-----------------------------------------------------------------*/

static void cdftForw_fx( Word16 n, Word32 *a, const Word16 *ip, const Word16 *w );

static void bitrv2_SR_fx( Word16 n, const Word16 *ip, Word32 *a );

static void cftfsub_fx( Word16 n, Word32 *a, const Word16 *w );
@@ -44,16 +43,39 @@ static void cft1st_fx( Word16 n, Word32 *a, const Word16 *w ); 
static void cftmdl_fx( Word16 n, Word16 l, Word32 *a, const Word16 *w );





/*-----------------------------------------------------------------*

 * DoRTFTn_fx()

 *

 *

 *-----------------------------------------------------------------*/



void DoRTFTn_fx(

    Word32 *x, /* i/o : real part of input and output data Q(x)      */

    Word32 *y, /* i/o : imaginary part of input and output data Q(x) */

#ifdef OPTIMIZE_FFT_STACK

    cmplx *spec, /* i/o : complex input and output data                */

#endif

    const Word16 n /* i : size of the FFT up to 1024                     */

)

{



    Word16 i;

    Word32 z[2048], *pt;



#ifdef OPTIMIZE_FFT_STACK

    IF( spec != NULL )

    {

        pt = z;

        FOR( i = 0; i < n; i++ )

        {

            *pt++ = spec[i].re;

            move16();

            *pt++ = spec[i].im;

            move16();

        }

    }

    ELSE

    {

#endif

        pt = z;

        FOR( i = 0; i < n; i++ )

        {
@@ -62,6 +84,9 @@ void DoRTFTn_fx( 
            *pt++ = y[i];

            move16();

        }

#ifdef OPTIMIZE_FFT_STACK

    }

#endif



    IF( EQ_16( n, 16 ) )

    {
@@ -92,6 +117,25 @@ void DoRTFTn_fx( 
        assert( 0 );

    }



#ifdef OPTIMIZE_FFT_STACK

    IF( spec != NULL )

    {

        spec[0].re = z[0];

        move16();

        spec[0].im = z[1];

        move16();

        pt = &z[2];

        FOR( i = n - 1; i >= 1; i-- )

        {

            spec[i].re = *pt++;

            move16();

            spec[i].im = *pt++;

            move16();

        }

    }

    ELSE

    {

#endif

        x[0] = z[0];

        move16();

        y[0] = z[1];
@@ -104,6 +148,9 @@ void DoRTFTn_fx( 
            y[i] = *pt++;

            move16();

        }

#ifdef OPTIMIZE_FFT_STACK

    }

#endif



    return;

}
@@ -124,6 +171,8 @@ static void cdftForw_fx( 


    /* Do FFT */

    cftfsub_fx( n, a, w );



    return;

}



/*-----------------------------------------------------------------*