1<?xml version="1.0" encoding="US-ASCII"?> 2<!DOCTYPE rfc SYSTEM "rfc2629.dtd"> 3<?rfc toc="yes"?> 4<?rfc tocompact="yes"?> 5<?rfc tocdepth="3"?> 6<?rfc tocindent="yes"?> 7<?rfc symrefs="yes"?> 8<?rfc sortrefs="yes"?> 9<?rfc comments="yes"?> 10<?rfc inline="yes"?> 11<?rfc compact="yes"?> 12<?rfc subcompact="no"?> 13<rfc category="std" docName="draft-ietf-codec-opus-update-06" 14 ipr="trust200902"> 15 <front> 16 <title abbrev="Opus Update">Updates to the Opus Audio Codec</title> 17 18<author initials="JM" surname="Valin" fullname="Jean-Marc Valin"> 19<organization>Mozilla Corporation</organization> 20<address> 21<postal> 22<street>331 E. Evelyn Avenue</street> 23<city>Mountain View</city> 24<region>CA</region> 25<code>94041</code> 26<country>USA</country> 27</postal> 28<phone>+1 650 903-0800</phone> 29<email>jmvalin@jmvalin.ca</email> 30</address> 31</author> 32 33<author initials="K." surname="Vos" fullname="Koen Vos"> 34<organization>vocTone</organization> 35<address> 36<postal> 37<street></street> 38<city></city> 39<region></region> 40<code></code> 41<country></country> 42</postal> 43<phone></phone> 44<email>koenvos74@gmail.com</email> 45</address> 46</author> 47 48 49 50 <date day="19" month="June" year="2017" /> 51 52 <abstract> 53 <t>This document addresses minor issues that were found in the specification 54 of the Opus audio codec in <xref target="RFC6716">RFC 6716</xref>.</t> 55 </abstract> 56 </front> 57 58 <middle> 59 <section title="Introduction"> 60 <t>This document addresses minor issues that were discovered in the reference 61 implementation of the Opus codec that serves as the specification in 62 <xref target="RFC6716">RFC 6716</xref>. Only issues affecting the decoder are 63 listed here. An up-to-date implementation of the Opus encoder can be found at 64 https://opus-codec.org/.</t> 65 <t> 66 Some of the changes in this document update normative behaviour in a way that requires 67 new test vectors. The English text of the specification is unaffected, only 68 the C implementation is. The updated specification remains fully compatible with 69 the original specification. 70 </t> 71 72 <t> 73 Note: due to RFC formatting conventions, lines exceeding the column width 74 in the patch are split using a backslash character. The backslashes 75 at the end of a line and the white space at the beginning 76 of the following line are not part of the patch. A properly formatted patch 77 including all changes is available at 78 <eref target="https://jmvalin.ca/misc_stuff/opus_update.patch"/>. (EDITOR: 79 change to an ietf.org link when ready) 80 </t> 81 82 </section> 83 84 <section title="Terminology"> 85 <t>The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", 86 "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this 87 document are to be interpreted as described in <xref 88 target="RFC2119">RFC 2119</xref>.</t> 89 </section> 90 91 <section title="Stereo State Reset in SILK"> 92 <t>The reference implementation does not reinitialize the stereo state 93 during a mode switch. The old stereo memory can produce a brief impulse 94 (i.e. single sample) in the decoded audio. This can be fixed by changing 95 silk/dec_API.c at line 72: 96 </t> 97<figure> 98<artwork><![CDATA[ 99 for( n = 0; n < DECODER_NUM_CHANNELS; n++ ) { 100 ret = silk_init_decoder( &channel_state[ n ] ); 101 } 102+ silk_memset(&((silk_decoder *)decState)->sStereo, 0, 103+ sizeof(((silk_decoder *)decState)->sStereo)); 104+ /* Not strictly needed, but it's cleaner that way */ 105+ ((silk_decoder *)decState)->prev_decode_only_middle = 0; 106 107 return ret; 108 } 109]]></artwork> 110</figure> 111 <t> 112 This change affects the normative part of the decoder, although the 113 amount of change is too small to make a significant impact on testvectors. 114 </t> 115 </section> 116 117 <section anchor="padding" title="Parsing of the Opus Packet Padding"> 118 <t>It was discovered that some invalid packets of very large size could trigger 119 an out-of-bounds read in the Opus packet parsing code responsible for padding. 120 This is due to an integer overflow if the signaled padding exceeds 2^31-1 bytes 121 (the actual packet may be smaller). The code can be fixed by applying the following 122 changes at line 596 of src/opus_decoder.c: 123 </t> 124<figure> 125<artwork><![CDATA[ 126 /* Padding flag is bit 6 */ 127 if (ch&0x40) 128 { 129- int padding=0; 130 int p; 131 do { 132 if (len<=0) 133 return OPUS_INVALID_PACKET; 134 p = *data++; 135 len--; 136- padding += p==255 ? 254: p; 137+ len -= p==255 ? 254: p; 138 } while (p==255); 139- len -= padding; 140 } 141]]></artwork> 142</figure> 143 <t>This packet parsing issue is limited to reading memory up 144 to about 60 kB beyond the compressed buffer. This can only be triggered 145 by a compressed packet more than about 16 MB long, so it's not a problem 146 for RTP. In theory, it <spanx style="emph">could</spanx> crash a file 147 decoder (e.g. Opus in Ogg) if the memory just after the incoming packet 148 is out-of-range, but our attempts to trigger such a crash in a production 149 application built using an affected version of the Opus decoder failed.</t> 150 </section> 151 152 <section anchor="resampler" title="Resampler buffer"> 153 <t>The SILK resampler had the following issues: 154 <list style="numbers"> 155 <t>The calls to memcpy() were using sizeof(opus_int32), but the type of the 156 local buffer was opus_int16.</t> 157 <t>Because the size was wrong, this potentially allowed the source 158 and destination regions of the memcpy() to overlap. 159 We <spanx style="emph">believe</spanx> that nSamplesIn is at least fs_in_khZ, 160 which is at least 8. 161 Since RESAMPLER_ORDER_FIR_12 is only 8, that should not be a problem once 162 the type size is fixed.</t> 163 <t>The size of the buffer used RESAMPLER_MAX_BATCH_SIZE_IN, but the 164 data stored in it was actually _twice_ the input batch size 165 (nSamplesIn<<1).</t> 166 </list></t> 167 <t> 168 The fact that the code never produced any error in testing (including when run under the 169 Valgrind memory debugger), suggests that in practice 170 the batch sizes are reasonable enough that none of the issues above 171 was ever a problem. However, proving that is non-obvious. 172 </t> 173 <t>The code can be fixed by applying the following changes to line 78 of silk/resampler_private_IIR_FIR.c: 174 </t> 175<figure> 176<artwork><![CDATA[ 177 ) 178 { 179 silk_resampler_state_struct *S = \ 180(silk_resampler_state_struct *)SS; 181 opus_int32 nSamplesIn; 182 opus_int32 max_index_Q16, index_increment_Q16; 183- opus_int16 buf[ RESAMPLER_MAX_BATCH_SIZE_IN + \ 184RESAMPLER_ORDER_FIR_12 ]; 185+ opus_int16 buf[ 2*RESAMPLER_MAX_BATCH_SIZE_IN + \ 186RESAMPLER_ORDER_FIR_12 ]; 187 188 /* Copy buffered samples to start of buffer */ 189- silk_memcpy( buf, S->sFIR, RESAMPLER_ORDER_FIR_12 \ 190* sizeof( opus_int32 ) ); 191+ silk_memcpy( buf, S->sFIR, RESAMPLER_ORDER_FIR_12 \ 192* sizeof( opus_int16 ) ); 193 194 /* Iterate over blocks of frameSizeIn input samples */ 195 index_increment_Q16 = S->invRatio_Q16; 196 while( 1 ) { 197 nSamplesIn = silk_min( inLen, S->batchSize ); 198 199 /* Upsample 2x */ 200 silk_resampler_private_up2_HQ( S->sIIR, &buf[ \ 201RESAMPLER_ORDER_FIR_12 ], in, nSamplesIn ); 202 203 max_index_Q16 = silk_LSHIFT32( nSamplesIn, 16 + 1 \ 204); /* + 1 because 2x upsampling */ 205 out = silk_resampler_private_IIR_FIR_INTERPOL( out, \ 206buf, max_index_Q16, index_increment_Q16 ); 207 in += nSamplesIn; 208 inLen -= nSamplesIn; 209 210 if( inLen > 0 ) { 211 /* More iterations to do; copy last part of \ 212filtered signal to beginning of buffer */ 213- silk_memcpy( buf, &buf[ nSamplesIn << 1 ], \ 214RESAMPLER_ORDER_FIR_12 * sizeof( opus_int32 ) ); 215+ silk_memmove( buf, &buf[ nSamplesIn << 1 ], \ 216RESAMPLER_ORDER_FIR_12 * sizeof( opus_int16 ) ); 217 } else { 218 break; 219 } 220 } 221 222 /* Copy last part of filtered signal to the state for \ 223the next call */ 224- silk_memcpy( S->sFIR, &buf[ nSamplesIn << 1 ], \ 225RESAMPLER_ORDER_FIR_12 * sizeof( opus_int32 ) ); 226+ silk_memcpy( S->sFIR, &buf[ nSamplesIn << 1 ], \ 227RESAMPLER_ORDER_FIR_12 * sizeof( opus_int16 ) ); 228 } 229]]></artwork> 230</figure> 231 </section> 232 233 <section title="Integer wrap-around in inverse gain computation"> 234 <t> 235 It was discovered through decoder fuzzing that some bitstreams could produce 236 integer values exceeding 32-bits in LPC_inverse_pred_gain_QA(), causing 237 a wrap-around. Although the error is harmless in practice, the C standard considers 238 the behavior as undefined, so the following patch to line 87 of silk/LPC_inv_pred_gain.c 239 detects values that do not fit in a 32-bit integer and considers the corresponding filters unstable: 240 </t> 241<figure> 242<artwork><![CDATA[ 243 /* Update AR coefficient */ 244 for( n = 0; n < k; n++ ) { 245- tmp_QA = Aold_QA[ n ] - MUL32_FRAC_Q( \ 246Aold_QA[ k - n - 1 ], rc_Q31, 31 ); 247- Anew_QA[ n ] = MUL32_FRAC_Q( tmp_QA, rc_mult2 , mult2Q ); 248+ opus_int64 tmp64; 249+ tmp_QA = silk_SUB_SAT32( Aold_QA[ n ], MUL32_FRAC_Q( \ 250Aold_QA[ k - n - 1 ], rc_Q31, 31 ) ); 251+ tmp64 = silk_RSHIFT_ROUND64( silk_SMULL( tmp_QA, \ 252rc_mult2 ), mult2Q); 253+ if( tmp64 > silk_int32_MAX || tmp64 < silk_int32_MIN ) { 254+ return 0; 255+ } 256+ Anew_QA[ n ] = ( opus_int32 )tmp64; 257 } 258]]></artwork> 259</figure> 260 </section> 261 262 <section title="Integer wrap-around in LSF decoding"> 263 <t> 264 It was discovered -- also from decoder fuzzing -- that an integer wrap-around could 265 occur when decoding line spectral frequency coefficients from extreme bitstreams. 266 The end result of the wrap-around is an illegal read access on the stack, which 267 the authors do not believe is exploitable but should nonetheless be fixed. The following 268 patch to line 137 of silk/NLSF_stabilize.c prevents the problem: 269 </t> 270<figure> 271<artwork><![CDATA[ 272 /* Keep delta_min distance between the NLSFs */ 273 for( i = 1; i < L; i++ ) 274- NLSF_Q15[i] = silk_max_int( NLSF_Q15[i], \ 275NLSF_Q15[i-1] + NDeltaMin_Q15[i] ); 276+ NLSF_Q15[i] = silk_max_int( NLSF_Q15[i], \ 277silk_ADD_SAT16( NLSF_Q15[i-1], NDeltaMin_Q15[i] ) ); 278 279 /* Last NLSF should be no higher than 1 - NDeltaMin[L] */ 280]]></artwork> 281</figure> 282 283 </section> 284 285 <section title="Cap on Band Energy"> 286 <t>On extreme bit-streams, it is possible for log-domain band energy levels 287 to exceed the maximum single-precision floating point value once converted 288 to a linear scale. This would later cause the decoded values to be NaN, 289 possibly causing problems in the software using the PCM values. This can be 290 avoided with the following patch to line 552 of celt/quant_bands.c: 291 </t> 292<figure> 293<artwork><![CDATA[ 294 { 295 opus_val16 lg = ADD16(oldEBands[i+c*m->nbEBands], 296 SHL16((opus_val16)eMeans[i],6)); 297+ lg = MIN32(QCONST32(32.f, 16), lg); 298 eBands[i+c*m->nbEBands] = PSHR32(celt_exp2(lg),4); 299 } 300 for (;i<m->nbEBands;i++) 301]]></artwork> 302</figure> 303 </section> 304 305 <section title="Hybrid Folding" anchor="folding"> 306 <t>When encoding in hybrid mode at low bitrate, we sometimes only have 307 enough bits to code a single CELT band (8 - 9.6 kHz). When that happens, 308 the second band (CELT band 18, from 9.6 to 12 kHz) cannot use folding 309 because it is wider than the amount already coded, and falls back to 310 LCG noise. Because it can also happen on transients (e.g. stops), it 311 can cause audible pre-echo. 312 </t> 313 <t> 314 To address the issue, we change the folding behavior so that it is 315 never forced to fall back to LCG due to the first band not containing 316 enough coefficients to fold onto the second band. This 317 is achieved by simply repeating part of the first band in the folding 318 of the second band. This changes the code in celt/bands.c around line 1237: 319 </t> 320<figure> 321<artwork><![CDATA[ 322 b = 0; 323 } 324 325- if (resynth && M*eBands[i]-N >= M*eBands[start] && \ 326(update_lowband || lowband_offset==0)) 327+ if (resynth && (M*eBands[i]-N >= M*eBands[start] || \ 328i==start+1) && (update_lowband || lowband_offset==0)) 329 lowband_offset = i; 330 331+ if (i == start+1) 332+ { 333+ int n1, n2; 334+ int offset; 335+ n1 = M*(eBands[start+1]-eBands[start]); 336+ n2 = M*(eBands[start+2]-eBands[start+1]); 337+ offset = M*eBands[start]; 338+ /* Duplicate enough of the first band folding data to \ 339be able to fold the second band. 340+ Copies no data for CELT-only mode. */ 341+ OPUS_COPY(&norm[offset+n1], &norm[offset+2*n1 - n2], n2-n1); 342+ if (C==2) 343+ OPUS_COPY(&norm2[offset+n1], &norm2[offset+2*n1 - n2], \ 344n2-n1); 345+ } 346+ 347 tf_change = tf_res[i]; 348 if (i>=m->effEBands) 349 { 350]]></artwork> 351</figure> 352 353 <t> 354 as well as line 1260: 355 </t> 356 357<figure> 358<artwork><![CDATA[ 359 fold_start = lowband_offset; 360 while(M*eBands[--fold_start] > effective_lowband); 361 fold_end = lowband_offset-1; 362- while(M*eBands[++fold_end] < effective_lowband+N); 363+ while(++fold_end < i && M*eBands[fold_end] < \ 364effective_lowband+N); 365 x_cm = y_cm = 0; 366 fold_i = fold_start; do { 367 x_cm |= collapse_masks[fold_i*C+0]; 368 369]]></artwork> 370</figure> 371 <t> 372 The fix does not impact compatibility, because the improvement does 373 not depend on the encoder doing anything special. There is also no 374 reasonable way for an encoder to use the original behavior to 375 improve quality over the proposed change. 376 </t> 377 </section> 378 379 <section title="Downmix to Mono" anchor="stereo"> 380 <t>The last issue is not strictly a bug, but it is an issue that has been reported 381 when downmixing an Opus decoded stream to mono, whether this is done inside the decoder 382 or as a post-processing step on the stereo decoder output. Opus intensity stereo allows 383 optionally coding the two channels 180-degrees out of phase on a per-band basis. 384 This provides better stereo quality than forcing the two channels to be in phase, 385 but when the output is downmixed to mono, the energy in the affected bands is cancelled 386 sometimes resulting in audible artefacts. 387 </t> 388 <t>As a work-around for this issue, the decoder MAY choose not to apply the 180-degree 389 phase shift when the output is meant to be downmixed (inside or 390 outside of the decoder). 391 </t> 392 </section> 393 394 395 <section title="New Test Vectors"> 396 <t>Changes in <xref target="folding"/> and <xref target="stereo"/> have 397 sufficient impact on the testvectors to make them fail. For this reason, 398 this document also updates the Opus test vectors. The new test vectors now 399 include two decoded outputs for the same bitstream. The outputs with 400 suffix 'm' do not apply the CELT 180-degree phase shift as allowed in 401 <xref target="stereo"/>, while the outputs without the suffix do. An 402 implementation is compliant as long as it passes either set of vectors. 403 </t> 404 <t> 405 In addition, any Opus implementation 406 that passes the original test vectors from <xref target="RFC6716">RFC 6716</xref> 407 is still compliant with the Opus specification. However, newer implementations 408 SHOULD be based on the new test vectors rather than the old ones. 409 </t> 410 <t>The new test vectors are located at 411 <eref target="https://jmvalin.ca/misc_stuff/opus_newvectors.tar.gz"/>. (EDITOR: 412 change to an ietf.org link when ready) 413 </t> 414 </section> 415 416 <section anchor="IANA" title="IANA Considerations"> 417 <t>This document makes no request of IANA.</t> 418 419 <t>Note to RFC Editor: this section may be removed on publication as an 420 RFC.</t> 421 </section> 422 423 <section anchor="Acknowledgements" title="Acknowledgements"> 424 <t>We would like to thank Juri Aedla for reporting the issue with the parsing of 425 the Opus padding. Also, thanks to Jonathan Lennox and Mark Harris for their 426 feedback on this document.</t> 427 </section> 428 </middle> 429 430 <back> 431 <references title="References"> 432 <?rfc include="http://xml.resource.org/public/rfc/bibxml/reference.RFC.2119.xml"?> 433 <?rfc include="http://xml.resource.org/public/rfc/bibxml/reference.RFC.6716.xml"?> 434 435 436 </references> 437 </back> 438</rfc> 439