[FFmpeg-devel] [PATCH v7] avcodec/jpeg2000: Fix FF_DWT97_INT to pass the conformance testing defined in ISO/IEC 15444-4

WATANABE Osamu owatanab at es.takushoku-u.ac.jp
Fri Nov 8 07:21:39 EET 2024


Hi Pierre,

This message is not an error but a warning that stays there before this patch.

I think the warning is a separate issue. 

Best,
Osamu

> 2024/11/08 9:02、Pierre-Anthony Lemieux <pal at sandflow.com>のメール:
> 
> ?I get with the following error with
> jpeg2000/itu-iso/codestreams_profile1/p1_03.j2k:
> 
> `[jpeg2000 @ 0x5c20680] Mid mismatch 1 in pass 18 of 22`
> 
>> On Sat, Nov 2, 2024 at 9:49?PM Osamu Watanabe
>> <owatanab at es.takushoku-u.ac.jp> wrote:
>> 
>> Fix for the integer version of the inverse 9-7 DWT processing
>> (FF_DWT97_INT, https://trac.ffmpeg.org/ticket/10123), which is activated with
>> `-flags +bitexact`.
>> 
>> I went through the code path for the DWT 9-7 transform (integer) and improved
>> precision to match conformance codestream.
>> 
>> As a result, the encoded codestream size is slightly larger for a given Q value.
>> For example, `-flags +bitexact -i lena.pnm -q: 20 -format j2k -y tmp.j2c`
>> gives 13K (HEAD) and 19K (with this patch).
>> 
>> This commit also updates the source and reference files for affected FATE tests.
>> 
>> Signed-off-by: Osamu Watanabe <owatanab at es.takushoku-u.ac.jp>
>> ---
>> libavcodec/jpeg2000.c                    |  11 +-
>> libavcodec/jpeg2000dec.c                 | 150 +++++++++++++----------
>> libavcodec/jpeg2000dwt.c                 |  47 +++----
>> libavcodec/jpeg2000dwt.h                 |   1 +
>> libavcodec/jpeg2000htdec.c               |   9 +-
>> libavcodec/tests/jpeg2000dwt.c           |   5 +
>> tests/ref/fate/j2k-dwt                   |  40 +++---
>> tests/ref/fate/jpeg2000-dcinema          |   4 +-
>> tests/ref/fate/jpeg2000dec-p0_04         |   2 +-
>> tests/ref/fate/jpeg2000dec-p0_05         |   2 +-
>> tests/ref/fate/jpeg2000dec-p0_09         |   2 +-
>> tests/ref/vsynth/vsynth1-jpeg2000-97     |   8 +-
>> tests/ref/vsynth/vsynth2-jpeg2000-97     |   8 +-
>> tests/ref/vsynth/vsynth3-jpeg2000-97     |   8 +-
>> tests/ref/vsynth/vsynth_lena-jpeg2000-97 |   8 +-
>> 15 files changed, 163 insertions(+), 142 deletions(-)
>> 
>> diff --git a/libavcodec/jpeg2000.c b/libavcodec/jpeg2000.c
>> index d6ffb02319..7911500901 100644
>> --- a/libavcodec/jpeg2000.c
>> +++ b/libavcodec/jpeg2000.c
>> @@ -260,9 +260,7 @@ static void init_band_stepsize(AVCodecContext *avctx,
>>                 band->f_stepsize *= F_LFTG_X * F_LFTG_X * 4;
>>                 break;
>>         }
>> -        if (codsty->transform == FF_DWT97) {
>> -            band->f_stepsize *= pow(F_LFTG_K, 2*(codsty->nreslevels2decode - reslevelno) + lband - 2);
>> -        }
>> +        band->f_stepsize *= pow(F_LFTG_K, 2*(codsty->nreslevels2decode - reslevelno) + lband - 2);
>>     }
>> 
>>     if (band->f_stepsize > (INT_MAX >> 15)) {
>> @@ -270,12 +268,7 @@ static void init_band_stepsize(AVCodecContext *avctx,
>>         av_log(avctx, AV_LOG_ERROR, "stepsize out of range\n");
>>     }
>> 
>> -    band->i_stepsize = band->f_stepsize * (1 << 15);
>> -
>> -    /* FIXME: In OpenJPEG code stepsize = stepsize * 0.5. Why?
>> -     * If not set output of entropic decoder is not correct. */
>> -    if (!av_codec_is_encoder(avctx->codec))
>> -        band->f_stepsize *= 0.5;
>> +    band->i_stepsize = (int)floorf(band->f_stepsize * (1 << 15));
>> }
>> 
>> static int init_prec(AVCodecContext *avctx,
>> diff --git a/libavcodec/jpeg2000dec.c b/libavcodec/jpeg2000dec.c
>> index 5b05ff2455..c9d8b025b1 100644
>> --- a/libavcodec/jpeg2000dec.c
>> +++ b/libavcodec/jpeg2000dec.c
>> @@ -1885,14 +1885,15 @@ static void decode_sigpass(Jpeg2000T1Context *t1, int width, int height,
>>                 && !(t1->flags[(y+1) * t1->stride + x+1] & (JPEG2000_T1_SIG | JPEG2000_T1_VIS))) {
>>                     if (ff_mqc_decode(&t1->mqc, t1->mqc.cx_states + ff_jpeg2000_getsigctxno(t1->flags[(y+1) * t1->stride + x+1] & flags_mask, bandno))) {
>>                         int xorbit, ctxno = ff_jpeg2000_getsgnctxno(t1->flags[(y+1) * t1->stride + x+1] & flags_mask, &xorbit);
>> -                        if (t1->mqc.raw)
>> -                             t1->data[(y) * t1->stride + x] = ff_mqc_decode(&t1->mqc, t1->mqc.cx_states + ctxno) ? -mask : mask;
>> -                        else
>> -                             t1->data[(y) * t1->stride + x] = (ff_mqc_decode(&t1->mqc, t1->mqc.cx_states + ctxno) ^ xorbit) ?
>> -                                               -mask : mask;
>> -
>> +                        if (t1->mqc.raw) {
>> +                            t1->data[(y) * t1->stride + x] |= ff_mqc_decode(&t1->mqc, t1->mqc.cx_states + ctxno) << 31;
>> +                            t1->data[(y) * t1->stride + x] |= mask;
>> +                        } else {
>> +                            t1->data[(y) * t1->stride + x] |= (ff_mqc_decode(&t1->mqc, t1->mqc.cx_states + ctxno) ^ xorbit) << 31;
>> +                            t1->data[(y) * t1->stride + x] |= mask;
>> +                        }
>>                         ff_jpeg2000_set_significance(t1, x, y,
>> -                                                     t1->data[(y) * t1->stride + x] < 0);
>> +                                                     t1->data[(y) * t1->stride + x] & INT32_MIN);
>>                     }
>>                     t1->flags[(y + 1) * t1->stride + x + 1] |= JPEG2000_T1_VIS;
>>                 }
>> @@ -1902,11 +1903,10 @@ static void decode_sigpass(Jpeg2000T1Context *t1, int width, int height,
>> static void decode_refpass(Jpeg2000T1Context *t1, int width, int height,
>>                            int bpno, int vert_causal_ctx_csty_symbol)
>> {
>> -    int phalf, nhalf;
>> +    int phalf;
>>     int y0, x, y;
>> 
>>     phalf = 1 << (bpno - 1);
>> -    nhalf = -phalf;
>> 
>>     for (y0 = 0; y0 < height; y0 += 4)
>>         for (x = 0; x < width; x++)
>> @@ -1915,10 +1915,13 @@ static void decode_refpass(Jpeg2000T1Context *t1, int width, int height,
>>                     int flags_mask = (vert_causal_ctx_csty_symbol && y == y0 + 3) ?
>>                         ~(JPEG2000_T1_SIG_S | JPEG2000_T1_SIG_SW | JPEG2000_T1_SIG_SE | JPEG2000_T1_SGN_S) : -1;
>>                     int ctxno = ff_jpeg2000_getrefctxno(t1->flags[(y + 1) * t1->stride + x + 1] & flags_mask);
>> -                    int r     = ff_mqc_decode(&t1->mqc,
>> -                                              t1->mqc.cx_states + ctxno)
>> -                                ? phalf : nhalf;
>> -                    t1->data[(y) * t1->stride + x]          += t1->data[(y) * t1->stride + x] < 0 ? -r : r;
>> +                    t1->data[(y) * t1->stride + x] |= phalf;
>> +                    if (ff_mqc_decode(&t1->mqc, t1->mqc.cx_states + ctxno))
>> +                        t1->data[(y) * t1->stride + x] |= phalf << 1;
>> +                    else {
>> +                        t1->data[(y) * t1->stride + x] &= ~(phalf << 1);
>> +
>> +                    }
>>                     t1->flags[(y + 1) * t1->stride + x + 1] |= JPEG2000_T1_REF;
>>                 }
>> }
>> @@ -1966,11 +1969,9 @@ static void decode_clnpass(const Jpeg2000DecoderContext *s, Jpeg2000T1Context *t
>>                     int xorbit;
>>                     int ctxno = ff_jpeg2000_getsgnctxno(t1->flags[(y + 1) * t1->stride + x + 1] & flags_mask,
>>                                                         &xorbit);
>> -                    t1->data[(y) * t1->stride + x] = (ff_mqc_decode(&t1->mqc,
>> -                                                    t1->mqc.cx_states + ctxno) ^
>> -                                      xorbit)
>> -                                     ? -mask : mask;
>> -                    ff_jpeg2000_set_significance(t1, x, y, t1->data[(y) * t1->stride + x] < 0);
>> +                    t1->data[(y) * t1->stride + x] |= (ff_mqc_decode(&t1->mqc, t1->mqc.cx_states + ctxno) ^ xorbit) << 31;
>> +                    t1->data[(y) * t1->stride + x] |= mask;
>> +                    ff_jpeg2000_set_significance(t1, x, y, t1->data[(y) * t1->stride + x] & INT32_MIN);
>>                 }
>>                 dec = 0;
>>                 t1->flags[(y + 1) * t1->stride + x + 1] &= ~JPEG2000_T1_VIS;
>> @@ -1991,9 +1992,9 @@ static void decode_clnpass(const Jpeg2000DecoderContext *s, Jpeg2000T1Context *t
>> 
>> static int decode_cblk(const Jpeg2000DecoderContext *s, Jpeg2000CodingStyle *codsty,
>>                        Jpeg2000T1Context *t1, Jpeg2000Cblk *cblk,
>> -                       int width, int height, int bandpos, uint8_t roi_shift)
>> +                       int width, int height, int bandpos, uint8_t roi_shift, const int M_b)
>> {
>> -    int passno = cblk->npasses, pass_t = 2, bpno = cblk->nonzerobits - 1;
>> +    int passno = cblk->npasses, pass_t = 2, bpno = cblk->nonzerobits - 1 + 31 - M_b - 1 - roi_shift;
>>     int pass_cnt = 0;
>>     int vert_causal_ctx_csty_symbol = codsty->cblk_style & JPEG2000_CBLK_VSC;
>>     int term_cnt = 0;
>> @@ -2068,22 +2069,25 @@ static int decode_cblk(const Jpeg2000DecoderContext *s, Jpeg2000CodingStyle *cod
>>         av_log(s->avctx, AV_LOG_WARNING, "Synthetic End of Stream Marker Read.\n");
>>     }
>> 
>> +    /* Reconstruct the sample values */
>> +    for (int y = 0; y < height; y++) {
>> +        for (int x = 0; x < width; x++) {
>> +            int32_t sign, n, val;
>> +            const uint32_t mask  = UINT32_MAX >> (M_b + 1); // bit mask for ROI detection
>> +
>> +            n = x + (y * t1->stride);
>> +            val = t1->data[n];
>> +            sign = val & INT32_MIN;
>> +            val &= INT32_MAX;
>> +            /* ROI shift, if necessary */
>> +            if (roi_shift && (((uint32_t)val & ~mask) == 0))
>> +                val <<= roi_shift;
>> +            t1->data[n] = val | sign; /* NOTE: Binary point for reconstruction value is located in 31 - M_b */
>> +        }
>> +    }
>>     return 1;
>> }
>> 
>> -static inline int roi_shift_param(Jpeg2000Component *comp,
>> -                                   int quan_parameter)
>> -{
>> -    uint8_t roi_shift;
>> -    int val;
>> -    roi_shift = comp->roi_shift;
>> -    val = (quan_parameter < 0)?-quan_parameter:quan_parameter;
>> -
>> -    if (val > (1 << roi_shift))
>> -        return (quan_parameter < 0)?-(val >> roi_shift):(val >> roi_shift);
>> -    return quan_parameter;
>> -}
>> -
>> /* TODO: Verify dequantization for lossless case
>>  * comp->data can be float or int
>>  * band->stepsize can be float or int
>> @@ -2093,50 +2097,86 @@ static inline int roi_shift_param(Jpeg2000Component *comp,
>> /* Float dequantization of a codeblock.*/
>> static void dequantization_float(int x, int y, Jpeg2000Cblk *cblk,
>>                                  Jpeg2000Component *comp,
>> -                                 Jpeg2000T1Context *t1, Jpeg2000Band *band)
>> +                                 Jpeg2000T1Context *t1, Jpeg2000Band *band, const int M_b)
>> {
>>     int i, j;
>>     int w = cblk->coord[0][1] - cblk->coord[0][0];
>> +    const int downshift = 31 - M_b;
>> +    float fscale = band->f_stepsize;
>> +    fscale /= (float)(1 << downshift);
>>     for (j = 0; j < (cblk->coord[1][1] - cblk->coord[1][0]); ++j) {
>>         float *datap = &comp->f_data[(comp->coord[0][1] - comp->coord[0][0]) * (y + j) + x];
>>         int *src = t1->data + j*t1->stride;
>> -        for (i = 0; i < w; ++i)
>> -            datap[i] = src[i] * band->f_stepsize;
>> +        for (i = 0; i < w; ++i) {
>> +            int val = src[i];
>> +            if (val < 0) // Convert sign-magnitude to two's complement
>> +                val = -(val & INT32_MAX);
>> +            datap[i] = (float)val * fscale;
>> +        }
>>     }
>> }
>> 
>> /* Integer dequantization of a codeblock.*/
>> static void dequantization_int(int x, int y, Jpeg2000Cblk *cblk,
>>                                Jpeg2000Component *comp,
>> -                               Jpeg2000T1Context *t1, Jpeg2000Band *band)
>> +                               Jpeg2000T1Context *t1, Jpeg2000Band *band, const int M_b)
>> {
>>     int i, j;
>> +    const int downshift = 31 - M_b;
>>     int w = cblk->coord[0][1] - cblk->coord[0][0];
>>     for (j = 0; j < (cblk->coord[1][1] - cblk->coord[1][0]); ++j) {
>>         int32_t *datap = &comp->i_data[(comp->coord[0][1] - comp->coord[0][0]) * (y + j) + x];
>>         int *src = t1->data + j*t1->stride;
>>         if (band->i_stepsize == 32768) {
>> -            for (i = 0; i < w; ++i)
>> -                datap[i] = src[i] / 2;
>> +            for (i = 0; i < w; ++i) {
>> +                int val = src[i];
>> +                if (val < 0)  // Convert sign-magnitude to two's complement
>> +                    val = -((val & INT32_MAX) >> downshift);
>> +                else
>> +                    val >>= downshift;
>> +                datap[i] = val;
>> +            }
>>         } else {
>>             // This should be VERY uncommon
>> -            for (i = 0; i < w; ++i)
>> -                datap[i] = (src[i] * (int64_t)band->i_stepsize) / 65536;
>> +            for (i = 0; i < w; ++i) {
>> +                int val = src[i];
>> +                if (val < 0)  // Convert sign-magnitude to two's complement
>> +                    val = -((val & INT32_MAX) >> downshift);
>> +                else
>> +                    val >>= downshift;
>> +                datap[i] = (val * (int64_t)band->i_stepsize) / 65536;
>> +            }
>>         }
>>     }
>> }
>> 
>> static void dequantization_int_97(int x, int y, Jpeg2000Cblk *cblk,
>>                                Jpeg2000Component *comp,
>> -                               Jpeg2000T1Context *t1, Jpeg2000Band *band)
>> +                               Jpeg2000T1Context *t1, Jpeg2000Band *band, const int M_b)
>> {
>>     int i, j;
>>     int w = cblk->coord[0][1] - cblk->coord[0][0];
>> +    float fscale = band->f_stepsize;
>> +    const int downshift = 31 - M_b;
>> +    const int PRESCALE = 6; // At least 6 is required to pass the conformance tests in ISO/IEC 15444-4
>> +    int scale;
>> +
>> +    fscale /= (float)(1 << downshift);
>> +    fscale *= (float)(1 << PRESCALE);
>> +    fscale *= (float)(1 << (16 + I_PRESHIFT));
>> +    scale = (int)(fscale + 0.5);
>> +    band->i_stepsize = scale;
>>     for (j = 0; j < (cblk->coord[1][1] - cblk->coord[1][0]); ++j) {
>>         int32_t *datap = &comp->i_data[(comp->coord[0][1] - comp->coord[0][0]) * (y + j) + x];
>>         int *src = t1->data + j*t1->stride;
>> -        for (i = 0; i < w; ++i)
>> -            datap[i] = (src[i] * (int64_t)band->i_stepsize + (1<<15)) >> 16;
>> +        for (i = 0; i < w; ++i) {
>> +            int val = src[i];
>> +            if (val < 0) // Convert sign-magnitude to two's complement
>> +                val = -(val & INT32_MAX);
>> +            // Shifting down to prevent overflow in dequantization
>> +            val = (val + (1 << (PRESCALE - 1))) >> PRESCALE;
>> +            datap[i] = RSHIFT(val * (int64_t)band->i_stepsize, 16);
>> +        }
>>     }
>> }
>> 
>> @@ -2168,18 +2208,6 @@ static inline void mct_decode(const Jpeg2000DecoderContext *s, Jpeg2000Tile *til
>>     s->dsp.mct_decode[tile->codsty[0].transform](src[0], src[1], src[2], csize);
>> }
>> 
>> -static inline void roi_scale_cblk(Jpeg2000Cblk *cblk,
>> -                                  Jpeg2000Component *comp,
>> -                                  Jpeg2000T1Context *t1)
>> -{
>> -    int i, j;
>> -    int w = cblk->coord[0][1] - cblk->coord[0][0];
>> -    for (j = 0; j < (cblk->coord[1][1] - cblk->coord[1][0]); ++j) {
>> -        int *src = t1->data + j*t1->stride;
>> -        for (i = 0; i < w; ++i)
>> -            src[i] = roi_shift_param(comp, src[i]);
>> -    }
>> -}
>> 
>> static inline int tile_codeblocks(const Jpeg2000DecoderContext *s, Jpeg2000Tile *tile)
>> {
>> @@ -2242,7 +2270,7 @@ static inline int tile_codeblocks(const Jpeg2000DecoderContext *s, Jpeg2000Tile
>>                             ret = decode_cblk(s, codsty, &t1, cblk,
>>                                               cblk->coord[0][1] - cblk->coord[0][0],
>>                                               cblk->coord[1][1] - cblk->coord[1][0],
>> -                                              bandpos, comp->roi_shift);
>> +                                              bandpos, comp->roi_shift, M_b);
>> 
>>                         if (ret)
>>                             coded = 1;
>> @@ -2251,14 +2279,12 @@ static inline int tile_codeblocks(const Jpeg2000DecoderContext *s, Jpeg2000Tile
>>                         x = cblk->coord[0][0] - band->coord[0][0];
>>                         y = cblk->coord[1][0] - band->coord[1][0];
>> 
>> -                        if (comp->roi_shift)
>> -                            roi_scale_cblk(cblk, comp, &t1);
>>                         if (codsty->transform == FF_DWT97)
>> -                            dequantization_float(x, y, cblk, comp, &t1, band);
>> +                            dequantization_float(x, y, cblk, comp, &t1, band, M_b);
>>                         else if (codsty->transform == FF_DWT97_INT)
>> -                            dequantization_int_97(x, y, cblk, comp, &t1, band);
>> +                            dequantization_int_97(x, y, cblk, comp, &t1, band, M_b);
>>                         else
>> -                            dequantization_int(x, y, cblk, comp, &t1, band);
>> +                            dequantization_int(x, y, cblk, comp, &t1, band, M_b);
>>                    } /* end cblk */
>>                 } /*end prec */
>>             } /* end band */
>> diff --git a/libavcodec/jpeg2000dwt.c b/libavcodec/jpeg2000dwt.c
>> index 34e33553f7..9ee8122658 100644
>> --- a/libavcodec/jpeg2000dwt.c
>> +++ b/libavcodec/jpeg2000dwt.c
>> @@ -39,13 +39,12 @@
>> 
>> /* Lifting parameters in integer format.
>>  * Computed as param = (float param) * (1 << 16) */
>> -#define I_LFTG_ALPHA  103949ll
>> -#define I_LFTG_BETA     3472ll
>> -#define I_LFTG_GAMMA   57862ll
>> -#define I_LFTG_DELTA   29066ll
>> -#define I_LFTG_K       80621ll
>> -#define I_LFTG_X       53274ll
>> -#define I_PRESHIFT 8
>> +#define I_LFTG_ALPHA_PRIME   38413ll // = 103949 - 65536, (= alpha - 1.0)
>> +#define I_LFTG_BETA           3472ll
>> +#define I_LFTG_GAMMA         57862ll
>> +#define I_LFTG_DELTA         29066ll
>> +#define I_LFTG_K             80621ll
>> +#define I_LFTG_X             53274ll
>> 
>> static inline void extend53(int *p, int i0, int i1)
>> {
>> @@ -234,8 +233,11 @@ static void sd_1d97_int(int *p, int i0, int i1)
>>     extend97_int(p, i0, i1);
>>     i0++; i1++;
>> 
>> -    for (i = (i0>>1) - 2; i < (i1>>1) + 1; i++)
>> -        p[2 * i + 1] -= (I_LFTG_ALPHA * (p[2 * i]     + p[2 * i + 2]) + (1 << 15)) >> 16;
>> +    for (i = (i0>>1) - 2; i < (i1>>1) + 1; i++) {
>> +        const int64_t sum = p[2 * i] + p[2 * i + 2];
>> +        p[2 * i + 1] -= sum;
>> +        p[2 * i + 1] -= (I_LFTG_ALPHA_PRIME * sum + (1 << 15)) >> 16;
>> +    }
>>     for (i = (i0>>1) - 1; i < (i1>>1) + 1; i++)
>>         p[2 * i]     -= (I_LFTG_BETA  * (p[2 * i - 1] + p[2 * i + 1]) + (1 << 15)) >> 16;
>>     for (i = (i0>>1) - 1; i < (i1>>1); i++)
>> @@ -276,7 +278,7 @@ static void dwt_encode97_int(DWTContext *s, int *t)
>> 
>>             // copy back and deinterleave
>>             for (i =   mv; i < lv; i+=2, j++)
>> -                t[w*j + lp] = ((l[i] * I_LFTG_X) + (1 << 15)) >> 16;
>> +                t[w*j + lp] = l[i];
>>             for (i = 1-mv; i < lv; i+=2, j++)
>>                 t[w*j + lp] = l[i];
>>         }
>> @@ -293,7 +295,7 @@ static void dwt_encode97_int(DWTContext *s, int *t)
>> 
>>             // copy back and deinterleave
>>             for (i =   mh; i < lh; i+=2, j++)
>> -                t[w*lp + j] = ((l[i] * I_LFTG_X) + (1 << 15)) >> 16;
>> +                t[w*lp + j] = l[i];
>>             for (i = 1-mh; i < lh; i+=2, j++)
>>                 t[w*lp + j] = l[i];
>>         }
>> @@ -301,7 +303,7 @@ static void dwt_encode97_int(DWTContext *s, int *t)
>>     }
>> 
>>     for (i = 0; i < w * h; i++)
>> -        t[i] = (t[i] + ((1<<I_PRESHIFT)>>1)) >> I_PRESHIFT;
>> +        t[i] = (t[i] + ((1<<(I_PRESHIFT))>>1)) >> (I_PRESHIFT);
>> }
>> 
>> static void sr_1d53(unsigned *p, int i0, int i1)
>> @@ -471,8 +473,11 @@ static void sr_1d97_int(int32_t *p, int i0, int i1)
>>     for (i = (i0 >> 1); i < (i1 >> 1) + 1; i++)
>>         p[2 * i]     += (I_LFTG_BETA  * (p[2 * i - 1] + (int64_t)p[2 * i + 1]) + (1 << 15)) >> 16;
>>     /* step 6 */
>> -    for (i = (i0 >> 1); i < (i1 >> 1); i++)
>> -        p[2 * i + 1] += (I_LFTG_ALPHA * (p[2 * i]     + (int64_t)p[2 * i + 2]) + (1 << 15)) >> 16;
>> +    for (i = (i0 >> 1); i < (i1 >> 1); i++) {
>> +        const int64_t sum = p[2 * i] + (int64_t) p[2 * i + 2];
>> +        p[2 * i + 1] += sum;
>> +        p[2 * i + 1] += (I_LFTG_ALPHA_PRIME * sum + (1 << 15)) >> 16;
>> +    }
>> }
>> 
>> static void dwt_decode97_int(DWTContext *s, int32_t *t)
>> @@ -486,9 +491,6 @@ static void dwt_decode97_int(DWTContext *s, int32_t *t)
>>     /* position at index O of line range [0-5,w+5] cf. extend function */
>>     line += 5;
>> 
>> -    for (i = 0; i < w * h; i++)
>> -        data[i] *= 1LL << I_PRESHIFT;
>> -
>>     for (lev = 0; lev < s->ndeclevels; lev++) {
>>         int lh = s->linelen[lev][0],
>>             lv = s->linelen[lev][1],
>> @@ -500,9 +502,9 @@ static void dwt_decode97_int(DWTContext *s, int32_t *t)
>>         l = line + mh;
>>         for (lp = 0; lp < lv; lp++) {
>>             int i, j = 0;
>> -            // rescale with interleaving
>> +            // interleaving
>>             for (i = mh; i < lh; i += 2, j++)
>> -                l[i] = ((data[w * lp + j] * I_LFTG_K) + (1 << 15)) >> 16;
>> +                l[i] = data[w * lp + j];
>>             for (i = 1 - mh; i < lh; i += 2, j++)
>>                 l[i] = data[w * lp + j];
>> 
>> @@ -516,9 +518,9 @@ static void dwt_decode97_int(DWTContext *s, int32_t *t)
>>         l = line + mv;
>>         for (lp = 0; lp < lh; lp++) {
>>             int i, j = 0;
>> -            // rescale with interleaving
>> +            // interleaving
>>             for (i = mv; i < lv; i += 2, j++)
>> -                l[i] = ((data[w * j + lp] * I_LFTG_K) + (1 << 15)) >> 16;
>> +                l[i] = data[w * j + lp];
>>             for (i = 1 - mv; i < lv; i += 2, j++)
>>                 l[i] = data[w * j + lp];
>> 
>> @@ -530,7 +532,8 @@ static void dwt_decode97_int(DWTContext *s, int32_t *t)
>>     }
>> 
>>     for (i = 0; i < w * h; i++)
>> -        data[i] = (data[i] + ((1LL<<I_PRESHIFT)>>1)) >> I_PRESHIFT;
>> +        // We shift down by `I_PRESHIFT` because the input coefficients `datap[]` were shifted up by `I_PRESHIFT` to secure the precision
>> +        data[i] = (int32_t)(data[i] + ((1LL<<(I_PRESHIFT))>>1)) >> (I_PRESHIFT);
>> }
>> 
>> int ff_jpeg2000_dwt_init(DWTContext *s, int border[2][2],
>> diff --git a/libavcodec/jpeg2000dwt.h b/libavcodec/jpeg2000dwt.h
>> index 718d183ac1..62f0548ac8 100644
>> --- a/libavcodec/jpeg2000dwt.h
>> +++ b/libavcodec/jpeg2000dwt.h
>> @@ -32,6 +32,7 @@
>> #define FF_DWT_MAX_DECLVLS 32 ///< max number of decomposition levels
>> #define F_LFTG_K      1.230174104914001f
>> #define F_LFTG_X      0.812893066115961f
>> +#define I_PRESHIFT 8
>> 
>> enum DWTType {
>>     FF_DWT97,
>> diff --git a/libavcodec/jpeg2000htdec.c b/libavcodec/jpeg2000htdec.c
>> index c47c8d61fe..186a6873ac 100644
>> --- a/libavcodec/jpeg2000htdec.c
>> +++ b/libavcodec/jpeg2000htdec.c
>> @@ -1314,8 +1314,6 @@ ff_jpeg2000_decode_htj2k(const Jpeg2000DecoderContext *s, Jpeg2000CodingStyle *c
>>         jpeg2000_decode_magref_segment(width, height, quad_buf_width, Dref, Lref,
>>                                        pLSB - 1, sample_buf, block_states);
>> 
>> -    pLSB = 31 - M_b;
>> -
>>     /* Reconstruct the sample values */
>>     for (int y = 0; y < height; y++) {
>>         for (int x = 0; x < width; x++) {
>> @@ -1328,12 +1326,7 @@ ff_jpeg2000_decode_htj2k(const Jpeg2000DecoderContext *s, Jpeg2000CodingStyle *c
>>             /* ROI shift, if necessary */
>>             if (roi_shift && (((uint32_t)val & ~mask) == 0))
>>                 val <<= roi_shift;
>> -            /* Convert sign-magnitude to two's complement. */
>> -            if (sign)
>> -                val = -val;
>> -            /* Shift down to 1 bit upper from decimal point for reconstruction value (= 0.5) */
>> -            val >>= (pLSB - 1);
>> -            t1->data[n] = val;
>> +            t1->data[n] = val | sign; /* NOTE: Binary point for reconstruction value is located in 31 - M_b */
>>         }
>>     }
>> free:
>> diff --git a/libavcodec/tests/jpeg2000dwt.c b/libavcodec/tests/jpeg2000dwt.c
>> index 520ecc05a3..9b26440ca8 100644
>> --- a/libavcodec/tests/jpeg2000dwt.c
>> +++ b/libavcodec/tests/jpeg2000dwt.c
>> @@ -46,6 +46,11 @@ static int test_dwt(int *array, int *ref, int border[2][2], int decomp_levels, i
>>         fprintf(stderr, "ff_dwt_encode failed\n");
>>         return 1;
>>     }
>> +    if (type == FF_DWT97_INT) {
>> +        // pre-scaling to simulate dequantization which places the binary point at 1 bit above from LSB
>> +        for (j = 0; j< s->linelen[decomp_levels-1][0] * s->linelen[decomp_levels-1][1]; j++)
>> +            array[j] <<= I_PRESHIFT;
>> +    }
>>     ret = ff_dwt_decode(s, array);
>>     if (ret < 0) {
>>         fprintf(stderr, "ff_dwt_encode failed\n");
>> diff --git a/tests/ref/fate/j2k-dwt b/tests/ref/fate/j2k-dwt
>> index 42415f00f9..9fbaaefd02 100644
>> --- a/tests/ref/fate/j2k-dwt
>> +++ b/tests/ref/fate/j2k-dwt
>> @@ -1,60 +1,60 @@
>> 5/3i, decomp:15 border 151 170 140 183 milli-err2:        0
>> -9/7i, decomp:15 border 151 170 140 183 milli-err2:      544
>> +9/7i, decomp:15 border 151 170 140 183 milli-err2:      112
>> 9/7f, decomp:15 border 151 170 140 183 err2:               0.000
>> 5/3i, decomp:21 border 173 201  81 189 milli-err2:        0
>> -9/7i, decomp:21 border 173 201  81 189 milli-err2:      592
>> +9/7i, decomp:21 border 173 201  81 189 milli-err2:      109
>> 9/7f, decomp:21 border 173 201  81 189 err2:               0.000
>> 5/3i, decomp:22 border 213 227  76 245 milli-err2:        0
>> -9/7i, decomp:22 border 213 227  76 245 milli-err2:      533
>> +9/7i, decomp:22 border 213 227  76 245 milli-err2:      115
>> 9/7f, decomp:22 border 213 227  76 245 err2:               0.000
>> 5/3i, decomp:13 border 134 157 184 203 milli-err2:        0
>> -9/7i, decomp:13 border 134 157 184 203 milli-err2:      535
>> +9/7i, decomp:13 border 134 157 184 203 milli-err2:      109
>> 9/7f, decomp:13 border 134 157 184 203 err2:               0.000
>> 5/3i, decomp: 1 border 204 237   6 106 milli-err2:        0
>> -9/7i, decomp: 1 border 204 237   6 106 milli-err2:      219
>> +9/7i, decomp: 1 border 204 237   6 106 milli-err2:       95
>> 9/7f, decomp: 1 border 204 237   6 106 err2:               0.000
>> 5/3i, decomp:28 border  76 211  13 210 milli-err2:        0
>> -9/7i, decomp:28 border  76 211  13 210 milli-err2:      791
>> +9/7i, decomp:28 border  76 211  13 210 milli-err2:      117
>> 9/7f, decomp:28 border  76 211  13 210 err2:               0.000
>> 5/3i, decomp:21 border  76  99  43 123 milli-err2:        0
>> -9/7i, decomp:21 border  76  99  43 123 milli-err2:      686
>> +9/7i, decomp:21 border  76  99  43 123 milli-err2:      104
>> 9/7f, decomp:21 border  76  99  43 123 err2:               0.000
>> 5/3i, decomp:15 border 192 243 174 204 milli-err2:        0
>> -9/7i, decomp:15 border 192 243 174 204 milli-err2:      476
>> +9/7i, decomp:15 border 192 243 174 204 milli-err2:      122
>> 9/7f, decomp:15 border 192 243 174 204 err2:               0.000
>> 5/3i, decomp:21 border  17  68  93 204 milli-err2:        0
>> -9/7i, decomp:21 border  17  68  93 204 milli-err2:      633
>> +9/7i, decomp:21 border  17  68  93 204 milli-err2:      124
>> 9/7f, decomp:21 border  17  68  93 204 err2:               0.000
>> 5/3i, decomp:11 border 142 168  82 174 milli-err2:        0
>> -9/7i, decomp:11 border 142 168  82 174 milli-err2:      696
>> +9/7i, decomp:11 border 142 168  82 174 milli-err2:      115
>> 9/7f, decomp:11 border 142 168  82 174 err2:               0.000
>> 5/3i, decomp:23 border 142 209 171 235 milli-err2:        0
>> -9/7i, decomp:23 border 142 209 171 235 milli-err2:      626
>> +9/7i, decomp:23 border 142 209 171 235 milli-err2:      120
>> 9/7f, decomp:23 border 142 209 171 235 err2:               0.000
>> 5/3i, decomp:30 border  37 185  79 245 milli-err2:        0
>> -9/7i, decomp:30 border  37 185  79 245 milli-err2:      953
>> +9/7i, decomp:30 border  37 185  79 245 milli-err2:      116
>> 9/7f, decomp:30 border  37 185  79 245 err2:               0.000
>> 5/3i, decomp: 5 border 129 236  30 243 milli-err2:        0
>> -9/7i, decomp: 5 border 129 236  30 243 milli-err2:      620
>> +9/7i, decomp: 5 border 129 236  30 243 milli-err2:      117
>> 9/7f, decomp: 5 border 129 236  30 243 err2:               0.000
>> 5/3i, decomp:10 border   5 160 146 247 milli-err2:        0
>> -9/7i, decomp:10 border   5 160 146 247 milli-err2:      797
>> +9/7i, decomp:10 border   5 160 146 247 milli-err2:      117
>> 9/7f, decomp:10 border   5 160 146 247 err2:               0.000
>> 5/3i, decomp: 5 border 104 162   6  47 milli-err2:        0
>> -9/7i, decomp: 5 border 104 162   6  47 milli-err2:      603
>> +9/7i, decomp: 5 border 104 162   6  47 milli-err2:      120
>> 9/7f, decomp: 5 border 104 162   6  47 err2:               0.000
>> 5/3i, decomp:24 border  78 250 102 218 milli-err2:        0
>> -9/7i, decomp:24 border  78 250 102 218 milli-err2:      836
>> +9/7i, decomp:24 border  78 250 102 218 milli-err2:      113
>> 9/7f, decomp:24 border  78 250 102 218 err2:               0.000
>> 5/3i, decomp:28 border  86  98  56  79 milli-err2:        0
>> -9/7i, decomp:28 border  86  98  56  79 milli-err2:      597
>> +9/7i, decomp:28 border  86  98  56  79 milli-err2:      115
>> 9/7f, decomp:28 border  86  98  56  79 err2:               0.000
>> 5/3i, decomp: 6 border  95 238 197 214 milli-err2:        0
>> -9/7i, decomp: 6 border  95 238 197 214 milli-err2:      478
>> +9/7i, decomp: 6 border  95 238 197 214 milli-err2:      114
>> 9/7f, decomp: 6 border  95 238 197 214 err2:               0.000
>> 5/3i, decomp:17 border  77 169  93 165 milli-err2:        0
>> -9/7i, decomp:17 border  77 169  93 165 milli-err2:      616
>> +9/7i, decomp:17 border  77 169  93 165 milli-err2:      124
>> 9/7f, decomp:17 border  77 169  93 165 err2:               0.000
>> 5/3i, decomp:22 border 178 187   7 119 milli-err2:        0
>> -9/7i, decomp:22 border 178 187   7 119 milli-err2:      392
>> +9/7i, decomp:22 border 178 187   7 119 milli-err2:       96
>> 9/7f, decomp:22 border 178 187   7 119 err2:               0.000
>> diff --git a/tests/ref/fate/jpeg2000-dcinema b/tests/ref/fate/jpeg2000-dcinema
>> index cdf8cd4fc6..223d753580 100644
>> --- a/tests/ref/fate/jpeg2000-dcinema
>> +++ b/tests/ref/fate/jpeg2000-dcinema
>> @@ -3,5 +3,5 @@
>> #codec_id 0: rawvideo
>> #dimensions 0: 1920x1080
>> #sar 0: 1/1
>> -0,          0,          0,        1, 12441600, 0xfcf6a127
>> -0,          1,          1,        1, 12441600, 0x577b6a64
>> +0,          0,          0,        1, 12441600, 0x0cf44be1
>> +0,          1,          1,        1, 12441600, 0xefe54482
>> diff --git a/tests/ref/fate/jpeg2000dec-p0_04 b/tests/ref/fate/jpeg2000dec-p0_04
>> index 5de7880c44..4c3ec322e4 100644
>> --- a/tests/ref/fate/jpeg2000dec-p0_04
>> +++ b/tests/ref/fate/jpeg2000dec-p0_04
>> @@ -3,4 +3,4 @@
>> #codec_id 0: rawvideo
>> #dimensions 0: 640x480
>> #sar 0: 0/1
>> -0,          0,          0,        1,   921600, 0x097d9665
>> +0,          0,          0,        1,   921600, 0x38311bba
>> diff --git a/tests/ref/fate/jpeg2000dec-p0_05 b/tests/ref/fate/jpeg2000dec-p0_05
>> index bb215043a1..410a08a62c 100644
>> --- a/tests/ref/fate/jpeg2000dec-p0_05
>> +++ b/tests/ref/fate/jpeg2000dec-p0_05
>> @@ -3,4 +3,4 @@
>> #codec_id 0: rawvideo
>> #dimensions 0: 1024x1024
>> #sar 0: 0/1
>> -0,          0,          0,        1,  2621440, 0x081f5048
>> +0,          0,          0,        1,  2621440, 0x9608ad8b
>> diff --git a/tests/ref/fate/jpeg2000dec-p0_09 b/tests/ref/fate/jpeg2000dec-p0_09
>> index 1755e7cc7d..ff78bf9dc7 100644
>> --- a/tests/ref/fate/jpeg2000dec-p0_09
>> +++ b/tests/ref/fate/jpeg2000dec-p0_09
>> @@ -3,4 +3,4 @@
>> #codec_id 0: rawvideo
>> #dimensions 0: 17x37
>> #sar 0: 0/1
>> -0,          0,          0,        1,      629, 0x5c9c389d
>> +0,          0,          0,        1,      629, 0xf35d38d6
>> diff --git a/tests/ref/vsynth/vsynth1-jpeg2000-97 b/tests/ref/vsynth/vsynth1-jpeg2000-97
>> index c979ab5c36..820249c6c0 100644
>> --- a/tests/ref/vsynth/vsynth1-jpeg2000-97
>> +++ b/tests/ref/vsynth/vsynth1-jpeg2000-97
>> @@ -1,4 +1,4 @@
>> -5e6d32b7205d31245b0d1f015d08b515 *tests/data/fate/vsynth1-jpeg2000-97.avi
>> -3643886 tests/data/fate/vsynth1-jpeg2000-97.avi
>> -a2262f1da2f49bc196b780a6b47ec4e8 *tests/data/fate/vsynth1-jpeg2000-97.out.rawvideo
>> -stddev:    4.23 PSNR: 35.59 MAXDIFF:   53 bytes:  7603200/  7603200
>> +803c2e8a4d054c5d603eed4c77abe492 *tests/data/fate/vsynth1-jpeg2000-97.avi
>> +4466514 tests/data/fate/vsynth1-jpeg2000-97.avi
>> +c9cf5a4580f10b00056c8d8731d21395 *tests/data/fate/vsynth1-jpeg2000-97.out.rawvideo
>> +stddev:    3.82 PSNR: 36.49 MAXDIFF:   49 bytes:  7603200/  7603200
>> diff --git a/tests/ref/vsynth/vsynth2-jpeg2000-97 b/tests/ref/vsynth/vsynth2-jpeg2000-97
>> index 591f8b6bb3..e40dca5b9a 100644
>> --- a/tests/ref/vsynth/vsynth2-jpeg2000-97
>> +++ b/tests/ref/vsynth/vsynth2-jpeg2000-97
>> @@ -1,4 +1,4 @@
>> -aa5573136c54b1855d8d00efe2a149bd *tests/data/fate/vsynth2-jpeg2000-97.avi
>> -2464134 tests/data/fate/vsynth2-jpeg2000-97.avi
>> -1f63c8b065e847e4c63d57ce23442ea8 *tests/data/fate/vsynth2-jpeg2000-97.out.rawvideo
>> -stddev:    3.21 PSNR: 37.99 MAXDIFF:   26 bytes:  7603200/  7603200
>> +c189c8b89c7aee3ab4f4a5aafdf7568f *tests/data/fate/vsynth2-jpeg2000-97.avi
>> +3225460 tests/data/fate/vsynth2-jpeg2000-97.avi
>> +4c0fbd7af969085d19dfabeb9634cddb *tests/data/fate/vsynth2-jpeg2000-97.out.rawvideo
>> +stddev:    2.55 PSNR: 39.98 MAXDIFF:   22 bytes:  7603200/  7603200
>> diff --git a/tests/ref/vsynth/vsynth3-jpeg2000-97 b/tests/ref/vsynth/vsynth3-jpeg2000-97
>> index 5d9d083791..92376e64cf 100644
>> --- a/tests/ref/vsynth/vsynth3-jpeg2000-97
>> +++ b/tests/ref/vsynth/vsynth3-jpeg2000-97
>> @@ -1,4 +1,4 @@
>> -522e12684aca4262a9d613cb2db7006c *tests/data/fate/vsynth3-jpeg2000-97.avi
>> -85526 tests/data/fate/vsynth3-jpeg2000-97.avi
>> -8def36ad1413ab3a5c2af2e1af4603f9 *tests/data/fate/vsynth3-jpeg2000-97.out.rawvideo
>> -stddev:    4.51 PSNR: 35.04 MAXDIFF:   47 bytes:    86700/    86700
>> +943cbdefa18b4a83175943f4e81e037c *tests/data/fate/vsynth3-jpeg2000-97.avi
>> +95642 tests/data/fate/vsynth3-jpeg2000-97.avi
>> +c4d58f0da2e8be602f54f032b58a581b *tests/data/fate/vsynth3-jpeg2000-97.out.rawvideo
>> +stddev:    4.11 PSNR: 35.84 MAXDIFF:   46 bytes:    86700/    86700
>> diff --git a/tests/ref/vsynth/vsynth_lena-jpeg2000-97 b/tests/ref/vsynth/vsynth_lena-jpeg2000-97
>> index 0539300185..dd2819b8cb 100644
>> --- a/tests/ref/vsynth/vsynth_lena-jpeg2000-97
>> +++ b/tests/ref/vsynth/vsynth_lena-jpeg2000-97
>> @@ -1,4 +1,4 @@
>> -80fe872c8afaad914da6ef037957d93b *tests/data/fate/vsynth_lena-jpeg2000-97.avi
>> -1937216 tests/data/fate/vsynth_lena-jpeg2000-97.avi
>> -1b97333a8dc115a5ba609b0070d89d4d *tests/data/fate/vsynth_lena-jpeg2000-97.out.rawvideo
>> -stddev:    2.82 PSNR: 39.10 MAXDIFF:   24 bytes:  7603200/  7603200
>> +9e2f5705be9d08494530724b625e17a4 *tests/data/fate/vsynth_lena-jpeg2000-97.avi
>> +2599714 tests/data/fate/vsynth_lena-jpeg2000-97.avi
>> +ab207505ec9c8a16bb45621404199e5c *tests/data/fate/vsynth_lena-jpeg2000-97.out.rawvideo
>> +stddev:    2.23 PSNR: 41.16 MAXDIFF:   20 bytes:  7603200/  7603200
>> --
>> 2.43.0
>> 


More information about the ffmpeg-devel mailing list