[FFmpeg-devel] [RFC v2 3/3] daaladec: Implement a native Daala decoder
Rostislav Pehlivanov
atomnuker at gmail.com
Tue Dec 29 22:34:41 CET 2015
Hi,
Thanks for the feedback.
>This is an infinite loop that can hang the decoder.
Will try to get the upstream to accept a fix.
>overflows in the entropy decoding system
The assertions should be able to catch them, so they should be fine.
They'll trigger if there's a desync with the bitstream. I might add more in
case they're not enough.
>What is the point of having DaalaSharedContext in addition to
DaalaBitstreamHeader?
I was looking at VP9 when starting to write this and it had this structure.
I'll remove the shared context one.
>Third, the decoder seems to return only the first frame, repeatedly,
>instead of the following frames.
You did read that it can only decode I-frames (keyframes). So you need to
encode your files with -k 1 to make all frames keyframes.
>First, this fails building with --enable-shared, because
daala_find_p_format
>is not static.
Fixed yesterday, but sent the same patch twice by mistake. It's been fixed,
but I'd like to post a v3 of the patch to address some of the issues found.
Thanks,
Rostislav
On 29 December 2015 at 15:55, Andreas Cadhalpun <
andreas.cadhalpun at googlemail.com> wrote:
> On 29.12.2015 03:12, Rostislav Pehlivanov wrote:
> > Hmm, kinda odd.
> > I've attached a v2 of the RFC Patch which does this:
> > 1. Fixed some typos.
> > 2. Fixed some missing newlines at the end of daalatab*
> > 3. Moves the pix_fmt finding function inside the main decoder file.
> >
> > Try building now.
>
> First, this fails building with --enable-shared, because
> daala_find_p_format
> is not static.
>
> Second, this decoder crashes, hangs and triggers lots of undefined
> behavior. :-/
>
> Third, the decoder seems to return only the first frame, repeatedly,
> instead of the following frames.
>
> More details are interleaved in the diff below.
>
> > diff --git a/libavcodec/daala.h b/libavcodec/daala.h
> > new file mode 100644
> > index 0000000..535e78f
> > --- /dev/null
> > +++ b/libavcodec/daala.h
> > @@ -0,0 +1,78 @@
> [...]
> > +typedef struct DaalaBitstreamHeader {
> > + uint8_t key_frame;
> > + uint8_t bipred;
> > + uint8_t ref_num;
> > + uint8_t act_mask;
> > + uint8_t qm;
> > + uint8_t haar;
> > + uint8_t golden;
> > + uint8_t pvq_qm[DAALA_MAX_PLANES][DAALA_QM_SIZE];
> > +} DaalaBitstreamHeader;
> > +
> > +typedef struct DaalaSharedContext {
> > + DaalaBitstreamHeader h;
> > +} DaalaSharedContext;
>
> What is the point of having DaalaSharedContext in addition to
> DaalaBitstreamHeader?
>
> > diff --git a/libavcodec/daala_entropy.h b/libavcodec/daala_entropy.h
> > new file mode 100644
> > index 0000000..3fdcaef
> > --- /dev/null
> > +++ b/libavcodec/daala_entropy.h
> > @@ -0,0 +1,554 @@
> [...]
> > +/* Updates the generic exponential probability model */
> > +static av_always_inline void daalaent_exp_model_update(DaalaCDF *c, int
> *ex, int x,
> > + int xs, int id,
> int integrate)
> > +{
> > + int i, xenc;
> > + ent_rng *cdf = &c->cdf[id*c->y];
> > + if (cdf[15] + c->inc > 32767) {
> > + for (i = 0; i < 16; i++)
> > + cdf[i] = (cdf[i] >> 1) + i + 1;
> > + }
> > + xenc = FFMIN(15, xs);
> > + for (i = xenc; i < 16; i++)
> > + cdf[i] += c->inc;
> > + x = FFMIN(x, 32767);
> > + *ex += ((x << 16) - *ex) >> integrate;
>
> This subtraction can overflow.
>
> [...]
> > +/* "+derf | It was a hack for the screen coding wavelet tools." */
>
> A rather bad hack...
>
> > +static inline int daalaent_decode_unary(DaalaEntropy *e)
> > +{
> > + int rval = 0;
> > + while (!daalaent_decode_bits(e, 1))
>
> This is an infinite loop that can hang the decoder.
>
> > + rval++;
> > + return rval;
> > +}
> [...]
> > +/* Decodes quantized coefficients from the bitsteam */
> > +static inline void daalaent_decode_laplace_vector(DaalaEntropy *e,
> dctcoef *y,
> > + int n, int k, dctcoef
> *curr,
> > + const dctcoef *means)
> > +{
> > + int i, exp_q8, mean_k_q8, mean_sum_ex_q8, sum_ex = 0, kn = k,
> ran_delta = 0;
> > + if (k <= 1) {
> > + daalaent_decode_laplace_delta(e, y, n, k, curr, means);
> > + return;
> > + }
> > + if (!k) {
> > + curr[DAALAENT_PVQ_COUNT] = DAALAENT_PVQ_NOVAL;
> > + curr[DAALAENT_PVQ_COUNT_EX] = DAALAENT_PVQ_NOVAL;
> > + curr[DAALAENT_PVQ_K] = 0;
> > + curr[DAALAENT_PVQ_SUM_EX] = 0;
> > + memset(y, 0, n*sizeof(dctcoef));
> > + return;
> > + }
> > + mean_k_q8 = means[DAALAENT_PVQ_K];
> > + mean_sum_ex_q8 = means[DAALAENT_PVQ_SUM_EX];
> > + if (mean_k_q8 < 1 << 23)
> > + exp_q8 = 256*mean_k_q8/(1 + mean_sum_ex_q8);
>
> This multiplication can overflow.
>
> > + else
> > + exp_q8 = mean_k_q8/(1 + (mean_sum_ex_q8 >> 8));
> > + for (i = 0; i < n; i++) {
> > + int x, ex;
> > + if (!kn)
> > + break;
> > + if (kn <= 1 && i != n - 1) {
> > + daalaent_decode_laplace_delta(e, y + i, n - i, kn, curr,
> means);
> > + ran_delta = 1;
> > + i = n;
> > + break;
> > + }
> > + ex = (2*exp_q8*kn + (n - i))/(2*(n - i));
> > + if (ex > kn*256)
> > + ex = kn*256;
> > + sum_ex += (2*256*kn + (n - i))/(2*(n - i));
>
> All these multiplications involving kn can overflow.
>
> > + if (i != n - 1)
> > + x = daalaent_decode_laplace_pvq(e, ex, kn);
> > + else
> > + x = kn;
> > + y[i] = x*daalaent_cphase(e, x);
> > + kn -= abs(x);
> > + }
> > + memset(&y[i], 0, (n - i)*sizeof(dctcoef)); /* Zero the rest */
> > + if (!ran_delta) {
> > + curr[DAALAENT_PVQ_COUNT] = DAALAENT_PVQ_NOVAL;
> > + curr[DAALAENT_PVQ_COUNT_EX] = DAALAENT_PVQ_NOVAL;
> > + }
> > + curr[DAALAENT_PVQ_K] = k - kn;
> > + curr[DAALAENT_PVQ_SUM_EX] = sum_ex;
> > +}
> > +
> > +/* Expectation value is in Q16 */
> > +static inline int daalaent_decode_generic(DaalaEntropy *e, DaalaCDF *c,
> int *ex,
> > + int max, int integrate)
> > +{
> > + int rval, lsb = 0, log_ex = daalaent_log_ex(*ex);
> > + const int shift = FFMAX(0, (log_ex - 5) >> 1);
> > + const int id = FFMIN(DAALAENT_MODEL_TAB - 1, log_ex);
> > + const int ms = (max + (1 << shift >> 1)) >> shift;
> > + int xs = (max == -1) ? 16 : FFMIN(ms + 1, 16);
> > + ent_rng *cdf = &c->cdf[id*c->y];
> > + if (!max)
> > + return 0;
> > + if ((xs = daalaent_decode_cdf(e, cdf, xs, 0, CDF_UNSCALED)) == 15) {
> > + int g = ((2*(*ex) >> 8) + (1 << shift >> 1)) >> shift;
>
> This multiplication can overflow, causing g to become negative, e.g. -256.
>
> > + ent_win decay = FFMAX(2, FFMIN(254, 256*g/(g + 256)));
>
> In that case this gives a division by zero crash.
>
> > +static inline void daalaent_decode_init(DaalaEntropy *e, const uint8_t
> *buf,
> > + int buf_size)
> > +{
> > + e->rbuf = buf;
> > + e->erbuf = buf + buf_size;
>
> This doesn't check if buf_size is 0. In that case e->end_window is always
> 0,
> causing daalaent_decode_bits to always return 0, leading to the infinite
> loop
> in daalaent_decode_unary.
>
> > + e->buf = buf;
> > + e->ebuf = buf + buf_size;
> > + e->err = 0;
> > + e->diff = 0;
> > + e->range = 32768;
> > + e->count = -15;
> > + e->eos_offset = 18 - DAALAENT_WSIZE;
> > + e->end_window = 0;
> > + e->end_window_size = 0;
> > + daalaent_fillup(e);
> > +}
> [...]
> > diff --git a/libavcodec/daala_pvq.h b/libavcodec/daala_pvq.h
> > new file mode 100644
> > index 0000000..ea8a86b
> > --- /dev/null
> > +++ b/libavcodec/daala_pvq.h
> > @@ -0,0 +1,369 @@
> [...]
> > +static inline void daalapvq_synth(dctcoef *xcoeff, dctcoef *ypulse,
> dctcoef *ref,
> > + int n, double gr, uint8_t ref_p,
> double gain,
> > + double theta, const int16_t *qmatrix,
> > + const int16_t *qmatrix_inv)
> > +{
> > + int i, m, nn = n - ref_p, s = 0, yy = 0;
> > + double scale, r[DAALAPVQ_MAX_PART_SIZE], x[DAALAPVQ_MAX_PART_SIZE];
> > + if (ref_p) {
> > + for (i = 0; i < n; i++)
> > + r[i] = ref[i]*qmatrix[i]*DAALA_QM_SCALE_UNIT;
>
> The 'ref[i]*qmatrix[i]' multiplication can overflow.
>
> > + }
> > + m = !ref_p ? 0 : daalapvq_householder_c(r, n, gr, &s);
> > + for (i = 0; i < nn; i++)
> > + yy += ypulse[i]*ypulse[i];
> > + scale = !yy ? 0 : gain/sqrt(yy);
> > + if (!ref_p) {
> > + for (i = 0; i < n; i++)
> > + xcoeff[i] =
> lrint((ypulse[i]*scale)*(qmatrix_inv[i]*DAALA_QM_INV_SCALE_UNIT));
> > + } else {
> > + scale *= sin(theta);
> > + for (i = 0; i < m; i++)
> > + x[i] = ypulse[i]*scale;
> > + x[m] = -s*gain*cos(theta);
> > + for (i = m; i < nn; i++)
> > + x[i+1] = ypulse[i]*scale;
> > + daalapvq_householder_a(x, r, n);
> > + for (i = 0; i < n; i++)
> > + xcoeff[i] =
> lrint(x[i]*qmatrix_inv[i]*DAALA_QM_INV_SCALE_UNIT);
> > + }
> > +}
> > +
> > +static av_always_inline void daalapvq_adapt_shuffle(int *dst, int *src,
> int spd,
> > + int idx, int mul)
> > +{
> > + if (src[idx] < 1)
> > + return;
> > + dst[idx+0] += (mul*src[idx+0] - dst[idx+0]) >> spd;
>
> This multiplication can overflow.
>
> [...]
> > +static inline void daalapvq_decode_vector(DaalaEntropy *e, DaalaPVQ
> *pvq,
> > + dctcoef *out, dctcoef *ref,
> > + const double beta,
> > + uint8_t key_frame, int p,
> > + uint8_t *skip_rest,
> > + uint8_t has_err, int band_idx,
> > + int qm_off, enum DaalaBsize
> bsize)
> > +{
> > + int i, k;
> > + int qg = 0, skip = 0, itheta = (!!key_frame), has_ref = !key_frame;
> > + double qcg, gain, theta = 0.0f, gr = 0.0f, gain_off = 0.0f;
> > + dctcoef tmp[DAALAPVQ_MAX_PART_SIZE] = {0};
> > +
> > + const int robust = has_err || key_frame;
> > + const int band_len = pvq->size[band_idx];
> > + const int16_t *qmatrix = &pvq->qmatrix[qm_off];
> > + const int16_t *qmatrix_inv = &pvq->qmatrix_inv[qm_off];
> > +
> > + if (!skip_rest[(band_idx + 2) % 3]) {
> > + int iloc = (!!p)*DAALA_NBSIZES*DAALAPVQ_PARTITIONS_MAX +
> bsize*DAALAPVQ_PARTITIONS_MAX + band_idx;
> > + i = daalaent_decode_cdf_adapt(e, &pvq->pvqtheta_gain_cdf, iloc,
> 8 + 7*pvq->skip[band_idx]);
> > + if (!key_frame && i >= 10)
> > + i++;
> > + if (key_frame && i >= 8)
> > + i++;
> > + if (i >= 8) {
> > + i -= 8;
> > + skip_rest[0] = skip_rest[1] = skip_rest[2] = 1;
> > + }
> > + qg = i & 1;
> > + itheta = (i >> 1) - 1;
> > + has_ref = !(itheta == -1);
> > + }
> > + if (qg) {
> > + int *ex = pvq->pvqgain_ex[p][bsize] + band_idx, ex_tmp = *ex;
> > + DaalaCDF *mcdf = has_ref ? &pvq->pvqgain_ref_mcdf :
> &pvq->pvqgain_noref_mcdf;
> > + qg = 1 + daalaent_decode_generic(e, mcdf, &ex_tmp, -1, 2);
> > + *ex += ((qg << 16) - *ex) >> 2;
>
> The left shift can overflow, also the subtraction, and the final addition,
> too.
>
> [...]
> > diff --git a/libavcodec/daaladec.c b/libavcodec/daaladec.c
> > new file mode 100644
> > index 0000000..3501a6b
> > --- /dev/null
> > +++ b/libavcodec/daaladec.c
> > @@ -0,0 +1,804 @@
> [...]
> > +/* Get DC level */
> > +static void get_haar_dc_sb(DaalaContext *s, HaarGradient *g, dctcoef
> *d, int x, int y,
> > + uint8_t p, uint8_t lim_pass)
> > +{
> > + int q, q_dc;
> > + int xdec = s->fmt->dec[p][0];
> > + const int aw = s->width >> xdec;
> > + const int ln = DAALA_LOG_BSIZE_MAX - xdec;
> > + dctcoef dc_pred = 0, **dc_buf = s->haar_dc_buf[p];
> > + if (!s->quantizer[p])
> > + q_dc = 1;
> > + else
> > + q_dc = FFMAX(1,
> s->quantizer[p]*s->s.h.pvq_qm[p][daala_get_qm_idx(DAALA_NBSIZES - 1, 0)] >>
> 4);
> > + if (x > 0 && y > 0) {
> > + if (lim_pass) { /* ALERT: coeffs could change */
> > + dc_pred = 22*dc_buf[x-1][y-0] - 9*dc_buf[x-1][y-1] +
> > + 15*dc_buf[x+0][y-1] + 4*dc_buf[x+1][y-1];
> > + } else {
> > + dc_pred = 23*dc_buf[x-1][y-0] - 10*dc_buf[x-1][y-1] +
> 19*dc_buf[x-0][y-1];
> > + }
>
> These two dc_pred calculations can overflow in various ways.
>
> > + dc_pred = (dc_pred + 16) >> 5;
> > + } else {
> > + dc_pred += x > 0 ? dc_buf[x-1][y-0] : 0;
> > + dc_pred += y > 0 ? dc_buf[x-0][y-1] : 0;
> > + }
> > + q = daalaent_decode_generic(&s->e, &s->haar_dc_mcdf[p],
> &s->haar_sb_ex[p], -1, 2);
> > + q *= daalaent_cphase(&s->e, q);
> > + q = q*q_dc + dc_pred;
>
> This multiplication can overflow.
>
> [...]
> > +/* Haar block decoding and transform */
> > +static void decode_block_haar(DaalaContext *s, int x, int y, int p,
> enum DaalaBsize bsize)
> > +{
> > + int i, j, k, l, n = 1 << (bsize + 2);
> > + const int dx = x << bsize, dy = y << bsize;
> > + const int aw = s->width >> s->fmt->dec[p][0];
> > + const int boffset = (dy << 2)*aw + (dx << 2);
> > +
> > + dctcoef tree[4][4];
> > + dctcoef pred[DAALA_BSIZE_MAX*DAALA_BSIZE_MAX];
> > + dctcoef tpred[DAALA_BSIZE_MAX*DAALA_BSIZE_MAX];
> > +
> > + daala_calc_prediction(s, pred, s->dcoef[p], dx, dy, p, bsize);
> > + memcpy(tpred, pred, n*n*sizeof(dctcoef));
> > +
> > + tree[0][0] = daalaent_decode_cdf_adapt(&s->e, &s->haar_bit_cdf, p,
> 16);
> > + if (tree[0][0] == 15)
> > + tree[0][0] += daalaent_decode_unary(&s->e);
> > +
> > + if (tree[0][0] > 24) {
> > + s->e.err = 1;
> > + return;
> > + } else if (tree[0][0] > 1) {
> > + int tmp = daalaent_decode_bits(&s->e, tree[0][0] - 1);
> > + tree[0][0] = (1 << (tree[0][0] - 1)) | tmp;
> > + }
> > +
> > + tree[1][1] = decode_haar_coeff_tree_split(s, tree[0][0],
> 3, 0);
> > + tree[0][1] = decode_haar_coeff_tree_split(s, tree[0][0] -
> tree[1][1], 4, 0);
> > + tree[1][0] = tree[0][0] - tree[1][1] - tree[0][1];
> > +
> > + decode_tree_sum(s, pred, 1, 0, tree[0][1], bsize + 2, 0);
> > + decode_tree_sum(s, pred, 0, 1, tree[1][0], bsize + 2, 1);
> > + decode_tree_sum(s, pred, 1, 1, tree[1][1], bsize + 2, 2);
> > +
> > + for (i = 0; i < n; i++) {
> > + for (j = (i == 0); j < n; j++)
> > + pred[i*n + j] *= daalaent_cphase(&s->e, pred[i*n + j]);
> > + }
> > + for (i = 0; i < 3; i++) { /* Direction */
> > + for (j = 0; j < bsize+2; j++) { /* Level */
> > + int bo = (((i + 1) >> 1) << j)*n + (((i + 1) & 1) << j);
> > + int q = !s->quantizer[p] ? 1 :
> s->quantizer[p]*daala_haar_qm[i == 2][j] >> 4;
> > + for (k = 0; k < 1 << j; k++)
> > + for (l = 0; l < 1 << j; l++)
> > + pred[bo + k*n + l] = q*pred[bo + k*n + l] +
> tpred[bo + k*n + l];
>
> The q*pred[] multiplication can overflow.
>
> [...]
> > diff --git a/libavcodec/daaladsp.c b/libavcodec/daaladsp.c
> > new file mode 100644
> > index 0000000..ba03520
> > --- /dev/null
> > +++ b/libavcodec/daaladsp.c
> > @@ -0,0 +1,2123 @@
> [...]
> > +static av_always_inline void idct_1D_4(pixel *x, int xstride, const
> pixel y[4])
> > +{
> > + int t2h, t0 = y[0], t1 = y[1], t2 = y[2], t3 = y[3];
> > + t3 += (t1*18293 + 8192) >> 14;
> > + t1 -= (t3*21407 + 16384) >> 15;
> > + t3 += (t1*23013 + 16384) >> 15;
> > + t2 = t0 - t2;
>
> These four lines can overflow.
>
> > + t2h = DAALA_DCT_RSHIFT(t2, 1);
> > + t0 -= t2h - DAALA_DCT_RSHIFT(t3, 1);
> > + t1 = t2h - t1;
>
> These two, as well.
>
> > + *(x + 0*xstride) = (pixel)t0;
> > + *(x + 1*xstride) = (pixel)(t2 - t1);
>
> This subtraction can overflow.
>
> > + *(x + 2*xstride) = (pixel)t1;
> > + *(x + 3*xstride) = (pixel)(t0 - t3);
>
> This one, too.
>
> > +}
> > +
> > +static av_always_inline void idct_1D_8(pixel *x, int xstride, const
> pixel y[16])
> > +{
> > + int t1h, t4h, t6h, t0 = y[0], t1 = y[1], t2 = y[2], t3 = y[3], t4 =
> y[4];
> > + int t5 = y[5], t6 = y[6], t7 = y[7];
> > + t5 -= (t3*2485 + 4096) >> 13;
> > + t3 += (t5*18205 + 16384) >> 15;
> > + t5 -= (t3*2485 + 4096) >> 13;
> > + t7 -= (t1*3227 + 16384) >> 15;
> > + t1 += (t7*6393 + 16384) >> 15;
> > + t7 -= (t1*3227 + 16384) >> 15;
> > + t1 += t3;
>
> These seven lines can overflow.
>
> > + t1h = DAALA_DCT_RSHIFT(t1, 1);
> > + t3 = t1h - t3;
>
> This line, too.
>
> > + t5 += t7;
> > + t7 = DAALA_DCT_RSHIFT(t5, 1) - t7;
> > + t3 += (t5*7489 + 4096) >> 13;
> > + t5 -= (t3*11585 + 8192) >> 14;
> > + t3 -= (t5*19195 + 16384) >> 15;
> > + t6 += (t2*21895 + 16384) >> 15;
> > + t2 -= (t6*15137 + 8192) >> 14;
> > + t6 += (t2*21895 + 16384) >> 15;
> > + t0 += (t4*13573 + 16384) >> 15;
> > + t4 -= (t0*11585 + 8192) >> 14;
> > + t0 += (t4*13573 + 16384) >> 15;
>
> These nine lines can also overflow.
>
> > + t4 = t2 - t4;
> > + t4h = DAALA_DCT_RSHIFT(t4, 1);
> > + t2 = t4h - t2;
> > + t6 = t0 - t6;
> > + t6h = DAALA_DCT_RSHIFT(t6, 1);
> > + t0 -= t6h;
> > + t7 = t6h - t7;
> > + t6 -= t7;
> > + t2 += DAALA_DCT_RSHIFT(t3, 1);
> > + t3 = t2 - t3;
> > + t5 += t4h;
> > + t4 -= t5;
>
> This can overflow.
>
> > + t0 += t1h;
> > + t1 = t0 - t1;
> > + *(x + 0*xstride) = (pixel)t0;
> > + *(x + 1*xstride) = (pixel)t4;
> > + *(x + 2*xstride) = (pixel)t2;
> > + *(x + 3*xstride) = (pixel)t6;
> > + *(x + 4*xstride) = (pixel)t7;
> > + *(x + 5*xstride) = (pixel)t3;
> > + *(x + 6*xstride) = (pixel)t5;
> > + *(x + 7*xstride) = (pixel)t1;
> > +}
> > +
> > +static av_always_inline void idct_1D_16(pixel *x, int xstride, const
> pixel y[16])
> > +{
> > + int tfh, tbh, t1h, tdh, t8h, tch, tah, t2h;
> > + int t0 = y[0], t1 = y[1], t2 = y[2], t3 = y[3], t4 = y[4], t5 =
> y[5];
> > + int t6 = y[6], t7 = y[7], t8 = y[8], t9 = y[9], ta = y[10], tb =
> y[11];
> > + int tc = y[12], td = y[13], te = y[14], tf = y[15];
> > + t1 += (tf*13573 + 16384) >> 15;
> > + tf -= (t1*11585 + 8192) >> 14;
> > + t1 += ((tf*13573 + 16384) >> 15)+t7;
> > + td -= (t3*10947 + 8192) >> 14;
> > + t3 += (td*15137 + 8192) >> 14;
> > + t5 += (tb*10947 + 8192) >> 14;
> > + tb -= (t5*15137 + 8192) >> 14;
> > + t5 += (tb*10947 + 8192) >> 14;
> > + td += t5 - ((t3*21895 + 16384) >> 15);
>
> These nine lines can overflow.
>
> > + tf = t9 - tf;
> > + tb += t3;
> > + tfh = DAALA_DCT_RSHIFT(tf, 1);
> > + t9 -= tfh;
> > + tbh = DAALA_DCT_RSHIFT(tb, 1);
> > + t3 += tfh - tbh;
> > + t1h = DAALA_DCT_RSHIFT(t1, 1);
> > + t7 = t1h - t7 + tbh;
> > + tdh = DAALA_DCT_RSHIFT(td, 1);
> > + t5 += t1h - tdh;
> > + t9 = tdh - t9;
> > + td -= t9;
> > + tf = t3 - tf;
> > + t1 -= t5 + ((tf*20055 + 16384) >> 15);
> > + tf += (t1*23059 + 8192) >> 14;
> > + t1 -= (tf*5417 + 4096) >> 13;
>
> These three lines, too.
>
> > + tb = t7 - tb;
> > + t9 += (t7*14101 + 8192) >> 14;
>
> This line, too.
>
> > + t7 += (t9*3363 + 4096) >> 13;
> > + t9 -= (t7*12905 + 8192) >> 14;
> > + tb -= (td*4379 + 8192) >> 14;
> > + td += (tb*20435 + 8192) >> 14;
> > + tb -= (td*17515 + 16384) >> 15;
> > + t3 += (t5*851 + 4096) >> 13;
> > + t5 += (t3*14699 + 8192) >> 14;
> > + t3 -= (t5*1035 + 1024) >> 11;
> > + t6 -= (ta*7335 + 16384) >> 15;
> > + ta -= (t6*12873 + 8192) >> 14;
> > + te += (t2*2873 + 1024) >> 11;
> > + t2 += (te*9041 + 16384) >> 15;
> > + t6 = DAALA_DCT_RSHIFT(t2, 1) - t6 - ((ta*8593 + 8192) >> 14);
> > + te = DAALA_DCT_RSHIFT(ta, 1) - te + ((t2*2275 + 1024) >> 11);
>
> These thirteen lines, too.
>
> > + t2 -= t6;
> > + ta -= te;
> > + t6 -= (ta*13573 + 16384) >> 15;
> > + ta += (t6*11585 + 8192) >> 14;
> > + t6 -= (ta*13573 + 16384) >> 15;
> > + tc += (t4*9147 + 4096) >> 13;
> > + t4 -= (tc*10703 + 8192) >> 14;
> > + tc += (t4*23013 + 16384) >> 15;
>
> These six lines, too.
>
> [...]
> > +static av_always_inline void idct_1D_32(pixel *x, int xstride, const
> pixel y[32])
> > +{
> > + int t0 = y[0], tg = y[1], t8 = y[2], to = y[3], t4 = y[4], tk
> = y[5];
> > + int tc = y[6], ts = y[7], t2 = y[8], ti = y[9], ta = y[10], tq
> = y[11];
> > + int t6 = y[12], tm = y[13], te = y[14], tu = y[15], t1 = y[16], th
> = y[17];
> > + int t9 = y[18], tp = y[19], t5 = y[20], tl = y[21], td = y[22], tt
> = y[23];
> > + int t3 = y[24], tj = y[25], tb = y[26], tr = y[27], t7 = y[28], tn
> = y[29];
> > + int tf = y[30], tv = y[31];
> > + OD_IDCT_32(t0, tg, t8, to, t4, tk, tc, ts, t2, ti, ta, tq, t6, tm,
> te, tu,
> > + t1, th, t9, tp, t5, tl, td, tt, t3, tj, tb, tr, t7, tn,
> tf, tv);
>
> This can overflow.
>
> [...]
> > +static av_always_inline void idct_1D_64(pixel *x, int xstride, const
> pixel y[64])
> > +{
> > + int t0 = y[0], tw = y[1], tg = y[2], tM = y[3], t8 = y[4], tE
> = y[5];
> > + int to = y[6], tU = y[7], t4 = y[8], tA = y[9], tk = y[10], tQ
> = y[11];
> > + int tc = y[12], tI = y[13], ts = y[14], tY = y[15], t2 = y[16], ty
> = y[17];
> > + int ti = y[18], tO = y[19], ta = y[20], tG = y[21], tq = y[22], tW
> = y[23];
> > + int t6 = y[24], tC = y[25], tm = y[26], tS = y[27], te = y[28], tK
> = y[29];
> > + int tu = y[30], t_ = y[31], t1 = y[32], tx = y[33], th = y[34], tN
> = y[35];
> > + int t9 = y[36], tF = y[37], tp = y[38], tV = y[39], t5 = y[40], tB
> = y[41];
> > + int tl = y[42], tR = y[43], td = y[44], tJ = y[45], tt = y[46], tZ
> = y[47];
> > + int t3 = y[48], tz = y[49], tj = y[50], tP = y[51], tb = y[52], tH
> = y[53];
> > + int tr = y[54], tX = y[55], t7 = y[56], tD = y[57], tn = y[58], tT
> = y[59];
> > + int tf = y[60], tL = y[61], tv = y[62], t = y[63];
> > + OD_IDCT_64(t0, tw, tg, tM, t8, tE, to, tU, t4, tA, tk, tQ, tc, tI,
> ts, tY,
> > + t2, ty, ti, tO, ta, tG, tq, tW, t6, tC, tm, tS, te, tK,
> tu, t_, t1, tx,
> > + th, tN, t9, tF, tp, tV, t5, tB, tl, tR, td, tJ, tt, tZ,
> t3, tz, tj, tP,
> > + tb, tH, tr, tX, t7, tD, tn, tT, tf, tL, tv, t);
>
> This can overflow.
> > +static void postfilter_4x4(pixel *dst, const pixel *src)
> > +{
> > + int t[4];
> > + t[3] = src[0]-src[3];
> > + t[2] = src[1]-src[2];
> > + t[1] = src[1]-(t[2]>>1);
>
> These two lines can overflow.
>
> > + t[0] = src[0]-(t[3]>>1);
> > + t[2] -= (t[3]*FILTER_PARAM_4_3+32)>>6;
> > + t[3] -= (t[2]*FILTER_PARAM_4_2+32)>>6;
> > + #if FILTER_PARAM_4_1 != 64
> > + t[3] = t[3]*(1 << 6)/FILTER_PARAM_4_1;
> > + #endif
> > + #if FILTER_PARAM_4_0 != 64
> > + t[2] = t[2]*(1 << 6)/FILTER_PARAM_4_0;
> > + #endif
>
> These four multiplications can overflow.
> [...]
> > +/* Chroma from luma */
> > +static void daala_cfl_resample(uint8_t *_dst, int dstride, const
> uint8_t *_src,
> > + int istride, int xdec, int ydec, int bs,
> int chroma_bs)
> > +{
> > + int i, j;
> > + const int n = 4 << bs;
> > + pixel *dst = (pixel *)_dst;
> > + pixel *src = (pixel *)_src;
> > + if (!chroma_bs && (xdec || ydec)) {
> > + if (xdec) {
> > + if (ydec) {
> > + daala_sf_ful_up(dst, dstride, src, istride, n, n, n);
> > + for (i = 0; i < 4; i++) {
> > + for (j = 0; j < 4; j++) {
> > + dst[i*dstride + j] =
> (daaladsp_cfl_scale[j][i]*dst[i*dstride + j] + 64) >> 7;
>
> The daaladsp_cfl_scale * dst multiplication can overflow, as well.
>
> Best regards,
> Andreas
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
More information about the ffmpeg-devel
mailing list