1 /****************************************************************************** 2 * 3 * Copyright (C) 2014 The Android Open Source Project 4 * Copyright 2003 - 2004 Open Interface North America, Inc. All rights reserved. 5 * 6 * Licensed under the Apache License, Version 2.0 (the "License"); 7 * you may not use this file except in compliance with the License. 8 * You may obtain a copy of the License at: 9 * 10 * http://www.apache.org/licenses/LICENSE-2.0 11 * 12 * Unless required by applicable law or agreed to in writing, software 13 * distributed under the License is distributed on an "AS IS" BASIS, 14 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 15 * See the License for the specific language governing permissions and 16 * limitations under the License. 17 * 18 ******************************************************************************/ 19 20 /********************************************************************************** 21 $Revision: #1 $ 22 ***********************************************************************************/ 23 24 /** @file 25 26 This file, along with synthesis-generated.c, contains the synthesis 27 filterbank routines. The operations performed correspond to the 28 operations described in A2DP Appendix B, Figure 12.3. Several 29 mathematical optimizations are performed, particularly for the 30 8-subband case. 31 32 One important optimization is to note that the "matrixing" operation 33 can be decomposed into the product of a type II discrete cosine kernel 34 and another, sparse matrix. 35 36 According to Fig 12.3, in the 8-subband case, 37 @code 38 N[k][i] = cos((i+0.5)*(k+4)*pi/8), k = 0..15 and i = 0..7 39 @endcode 40 41 N can be factored as R * C2, where C2 is an 8-point type II discrete 42 cosine kernel given by 43 @code 44 C2[k][i] = cos((i+0.5)*k*pi/8)), k = 0..7 and i = 0..7 45 @endcode 46 47 R turns out to be a sparse 16x8 matrix with the following non-zero 48 entries: 49 @code 50 R[k][k+4] = 1, k = 0..3 51 R[k][abs(12-k)] = -1, k = 5..15 52 @endcode 53 54 The spec describes computing V[0..15] as N * R. 55 @code 56 V[0..15] = N * R = (R * C2) * R = R * (C2 * R) 57 @endcode 58 59 C2 * R corresponds to computing the discrete cosine transform of R, so 60 V[0..15] can be computed by taking the DCT of R followed by assignment 61 and selective negation of the DCT result into V. 62 63 Although this was derived empirically using GNU Octave, it is 64 formally demonstrated in, e.g., Liu, Chi-Min and Lee, 65 Wen-Chieh. "A Unified Fast Algorithm for Cosine Modulated 66 Filter Banks in Current Audio Coding Standards." Journal of 67 the AES 47 (December 1999): 1061. 68 69 Given the shift operation performed prior to computing V[0..15], it is 70 clear that V[0..159] represents a rolling history of the 10 most 71 recent groups of blocks input to the synthesis operation. Interpreting 72 the matrix N in light of its factorization into C2 and R, R's 73 sparseness has implications for interpreting the values in V. In 74 particular, there is considerable redundancy in the values stored in 75 V. Furthermore, since R[4][0..7] are all zeros, one out of every 16 76 values in V will be zero regardless of the input data. Within each 77 block of 16 values in V, fully half of them are redundant or 78 irrelevant: 79 80 @code 81 V[ 0] = DCT[4] 82 V[ 1] = DCT[5] 83 V[ 2] = DCT[6] 84 V[ 3] = DCT[7] 85 V[ 4] = 0 86 V[ 5] = -DCT[7] = -V[3] (redundant) 87 V[ 6] = -DCT[6] = -V[2] (redundant) 88 V[ 7] = -DCT[5] = -V[1] (redundant) 89 V[ 8] = -DCT[4] = -V[0] (redundant) 90 V[ 9] = -DCT[3] 91 V[10] = -DCT[2] 92 V[11] = -DCT[1] 93 V[12] = -DCT[0] 94 V[13] = -DCT[1] = V[11] (redundant) 95 V[14] = -DCT[2] = V[10] (redundant) 96 V[15] = -DCT[3] = V[ 9] (redundant) 97 @endcode 98 99 Since the elements of V beyond 15 were originally computed the same 100 way during a previous run, what holds true for V[x] also holds true 101 for V[x+16]. Thus, so long as care is taken to maintain the mapping, 102 we need only actually store the unique values, which correspond to the 103 output of the DCT, in some cases inverted. In fact, instead of storing 104 V[0..159], we could store DCT[0..79] which would contain a history of 105 DCT results. More on this in a bit. 106 107 Going back to figure 12.3 in the spec, it should be clear that the 108 vector U need not actually be explicitly constructed, but that with 109 suitable indexing into V during the window operation, the same end can 110 be accomplished. In the same spirit of the pseudocode shown in the 111 figure, the following is the construction of W without using U: 112 113 @code 114 for i=0 to 79 do 115 W[i] = D[i]*VSIGN(i)*V[remap_V(i)] where remap_V(i) = 32*(int(i/16)) + (i % 16) + (i % 16 >= 8 ? 16 : 0) 116 and VSIGN(i) maps i%16 into {1, 1, 1, 1, 0, -1, -1, -1, -1, 1, 1, 1, 1, 1, 1 } 117 These values correspond to the 118 signs of the redundant values as 119 shown in the explanation three 120 paragraphs above. 121 @endcode 122 123 We saw above how V[4..8,13..15] (and by extension 124 V[(4..8,13..15)+16*n]) can be defined in terms of other elements 125 within the subblock of V. V[0..3,9..12] correspond to DCT elements. 126 127 @code 128 for i=0 to 79 do 129 W[i] = D[i]*DSIGN(i)*DCT[remap_DCT(i)] 130 @endcode 131 132 The DCT is calculated using the Arai-Agui-Nakajima factorization, 133 which saves some computation by producing output that needs to be 134 multiplied by scaling factors before being used. 135 136 @code 137 for i=0 to 79 do 138 W[i] = D[i]*SCALE[i%8]*AAN_DCT[remap_DCT(i)] 139 @endcode 140 141 D can be premultiplied with the DCT scaling factors to yield 142 143 @code 144 for i=0 to 79 do 145 W[i] = DSCALED[i]*AAN_DCT[remap_DCT(i)] where DSCALED[i] = D[i]*SCALE[i%8] 146 @endcode 147 148 The output samples X[0..7] are defined as sums of W: 149 150 @code 151 X[j] = sum{i=0..9}(W[j+8*i]) 152 @endcode 153 154 @ingroup codec_internal 155 */ 156 157 /** 158 @addtogroup codec_internal 159 @{ 160 */ 161 162 #include "oi_codec_sbc_private.h" 163 164 const OI_INT32 dec_window_4[21] = { 165 0, /* +0.00000000E+00 */ 166 97, /* +5.36548976E-04 */ 167 270, /* +1.49188357E-03 */ 168 495, /* +2.73370904E-03 */ 169 694, /* +3.83720193E-03 */ 170 704, /* +3.89205149E-03 */ 171 338, /* +1.86581691E-03 */ 172 -554, /* -3.06012286E-03 */ 173 1974, /* +1.09137620E-02 */ 174 3697, /* +2.04385087E-02 */ 175 5224, /* +2.88757392E-02 */ 176 5824, /* +3.21939290E-02 */ 177 4681, /* +2.58767811E-02 */ 178 1109, /* +6.13245186E-03 */ 179 -5214, /* -2.88217274E-02 */ 180 -14047, /* -7.76463494E-02 */ 181 24529, /* +1.35593274E-01 */ 182 35274, /* +1.94987841E-01 */ 183 44618, /* +2.46636662E-01 */ 184 50984, /* +2.81828203E-01 */ 185 53243, /* +2.94315332E-01 */ 186 }; 187 188 #define DCTII_4_K06_FIX ( 11585)/* S1.14 11585 0.707107*/ 189 190 #define DCTII_4_K08_FIX ( 21407)/* S1.14 21407 1.306563*/ 191 192 #define DCTII_4_K09_FIX (-15137)/* S1.14 -15137 -0.923880*/ 193 194 #define DCTII_4_K10_FIX ( -8867)/* S1.14 -8867 -0.541196*/ 195 196 /** Scales x by y bits to the right, adding a rounding factor. 197 */ 198 #ifndef SCALE 199 #define SCALE(x, y) (((x) + (1 <<((y)-1))) >> (y)) 200 #endif 201 202 #ifndef CLIP_INT16 203 #define CLIP_INT16(x) do { if (x > OI_INT16_MAX) { x = OI_INT16_MAX; } else if (x < OI_INT16_MIN) { x = OI_INT16_MIN; } } while (0) 204 #endif 205 206 /** 207 * Default C language implementation of a 16x32->32 multiply. This function may 208 * be replaced by a platform-specific version for speed. 209 * 210 * @param u A signed 16-bit multiplicand 211 * @param v A signed 32-bit multiplier 212 213 * @return A signed 32-bit value corresponding to the 32 most significant bits 214 * of the 48-bit product of u and v. 215 */ 216 INLINE OI_INT32 default_mul_16s_32s_hi(OI_INT16 u, OI_INT32 v) 217 { 218 OI_UINT16 v0; 219 OI_INT16 v1; 220 221 OI_INT32 w,x; 222 223 v0 = (OI_UINT16)(v & 0xffff); 224 v1 = (OI_INT16) (v >> 16); 225 226 w = v1 * u; 227 x = u * v0; 228 229 return w + (x >> 16); 230 } 231 232 #define MUL_16S_32S_HI(_x, _y) default_mul_16s_32s_hi(_x, _y) 233 234 #define LONG_MULT_DCT(K, sample) (MUL_16S_32S_HI(K, sample)<<2) 235 236 PRIVATE void SynthWindow80_generated(OI_INT16 *pcm, SBC_BUFFER_T const * RESTRICT buffer, OI_UINT strideShift); 237 PRIVATE void SynthWindow112_generated(OI_INT16 *pcm, SBC_BUFFER_T const * RESTRICT buffer, OI_UINT strideShift); 238 PRIVATE void dct2_8(SBC_BUFFER_T * RESTRICT out, OI_INT32 const * RESTRICT x); 239 240 typedef void (*SYNTH_FRAME)(OI_CODEC_SBC_DECODER_CONTEXT *context, OI_INT16 *pcm, OI_UINT blkstart, OI_UINT blkcount); 241 242 #ifndef COPY_BACKWARD_32BIT_ALIGNED_72_HALFWORDS 243 #define COPY_BACKWARD_32BIT_ALIGNED_72_HALFWORDS(dest, src) do { shift_buffer(dest, src, 72); } while (0) 244 #endif 245 246 #ifndef DCT2_8 247 #define DCT2_8(dst, src) dct2_8(dst, src) 248 #endif 249 250 #ifndef SYNTH80 251 #define SYNTH80 SynthWindow80_generated 252 #endif 253 254 #ifndef SYNTH112 255 #define SYNTH112 SynthWindow112_generated 256 #endif 257 258 PRIVATE void OI_SBC_SynthFrame_80(OI_CODEC_SBC_DECODER_CONTEXT *context, OI_INT16 *pcm, OI_UINT blkstart, OI_UINT blkcount) 259 { 260 OI_UINT blk; 261 OI_UINT ch; 262 OI_UINT nrof_channels = context->common.frameInfo.nrof_channels; 263 OI_UINT pcmStrideShift = context->common.pcmStride == 1 ? 0 : 1; 264 OI_UINT offset = context->common.filterBufferOffset; 265 OI_INT32 *s = context->common.subdata + 8 * nrof_channels * blkstart; 266 OI_UINT blkstop = blkstart + blkcount; 267 268 for (blk = blkstart; blk < blkstop; blk++) { 269 if (offset == 0) { 270 COPY_BACKWARD_32BIT_ALIGNED_72_HALFWORDS(context->common.filterBuffer[0] + context->common.filterBufferLen - 72, context->common.filterBuffer[0]); 271 if (nrof_channels == 2) { 272 COPY_BACKWARD_32BIT_ALIGNED_72_HALFWORDS(context->common.filterBuffer[1] + context->common.filterBufferLen - 72, context->common.filterBuffer[1]); 273 } 274 offset = context->common.filterBufferLen - 80; 275 } else { 276 offset -= 1*8; 277 } 278 279 for (ch = 0; ch < nrof_channels; ch++) { 280 DCT2_8(context->common.filterBuffer[ch] + offset, s); 281 SYNTH80(pcm + ch, context->common.filterBuffer[ch] + offset, pcmStrideShift); 282 s += 8; 283 } 284 pcm += (8 << pcmStrideShift); 285 } 286 context->common.filterBufferOffset = offset; 287 } 288 289 PRIVATE void OI_SBC_SynthFrame_4SB(OI_CODEC_SBC_DECODER_CONTEXT *context, OI_INT16 *pcm, OI_UINT blkstart, OI_UINT blkcount) 290 { 291 OI_UINT blk; 292 OI_UINT ch; 293 OI_UINT nrof_channels = context->common.frameInfo.nrof_channels; 294 OI_UINT pcmStrideShift = context->common.pcmStride == 1 ? 0 : 1; 295 OI_UINT offset = context->common.filterBufferOffset; 296 OI_INT32 *s = context->common.subdata + 8 * nrof_channels * blkstart; 297 OI_UINT blkstop = blkstart + blkcount; 298 299 for (blk = blkstart; blk < blkstop; blk++) { 300 if (offset == 0) { 301 COPY_BACKWARD_32BIT_ALIGNED_72_HALFWORDS(context->common.filterBuffer[0] + context->common.filterBufferLen - 72,context->common.filterBuffer[0]); 302 if (nrof_channels == 2) { 303 COPY_BACKWARD_32BIT_ALIGNED_72_HALFWORDS(context->common.filterBuffer[1] + context->common.filterBufferLen - 72,context->common.filterBuffer[1]); 304 } 305 offset =context->common.filterBufferLen - 80; 306 } else { 307 offset -= 8; 308 } 309 for (ch = 0; ch < nrof_channels; ch++) { 310 cosineModulateSynth4(context->common.filterBuffer[ch] + offset, s); 311 SynthWindow40_int32_int32_symmetry_with_sum(pcm + ch, 312 context->common.filterBuffer[ch] + offset, 313 pcmStrideShift); 314 s += 4; 315 } 316 pcm += (4 << pcmStrideShift); 317 } 318 context->common.filterBufferOffset = offset; 319 } 320 321 #ifdef SBC_ENHANCED 322 323 PRIVATE void OI_SBC_SynthFrame_Enhanced(OI_CODEC_SBC_DECODER_CONTEXT *context, OI_INT16 *pcm, OI_UINT blkstart, OI_UINT blkcount) 324 { 325 OI_UINT blk; 326 OI_UINT ch; 327 OI_UINT nrof_channels = context->common.frameInfo.nrof_channels; 328 OI_UINT pcmStrideShift = context->common.pcmStride == 1 ? 0 : 1; 329 OI_UINT offset = context->common.filterBufferOffset; 330 OI_INT32 *s = context->common.subdata + 8 * nrof_channels * blkstart; 331 OI_UINT blkstop = blkstart + blkcount; 332 333 for (blk = blkstart; blk < blkstop; blk++) { 334 if (offset == 0) { 335 COPY_BACKWARD_32BIT_ALIGNED_104_HALFWORDS(context->common.filterBuffer[0] +context->common.filterBufferLen - 104, context->common.filterBuffer[0]); 336 if (nrof_channels == 2) { 337 COPY_BACKWARD_32BIT_ALIGNED_104_HALFWORDS(context->common.filterBuffer[1] + context->common.filterBufferLen - 104, context->common.filterBuffer[1]); 338 } 339 offset = context->common.filterBufferLen - 112; 340 } else { 341 offset -= 8; 342 } 343 for (ch = 0; ch < nrof_channels; ++ch) { 344 DCT2_8(context->common.filterBuffer[ch] + offset, s); 345 SYNTH112(pcm + ch, context->common.filterBuffer[ch] + offset, pcmStrideShift); 346 s += 8; 347 } 348 pcm += (8 << pcmStrideShift); 349 } 350 context->common.filterBufferOffset = offset; 351 } 352 353 static const SYNTH_FRAME SynthFrameEnhanced[] = { 354 NULL, /* invalid */ 355 OI_SBC_SynthFrame_Enhanced, /* mono */ 356 OI_SBC_SynthFrame_Enhanced /* stereo */ 357 }; 358 359 #endif 360 361 static const SYNTH_FRAME SynthFrame8SB[] = { 362 NULL, /* invalid */ 363 OI_SBC_SynthFrame_80, /* mono */ 364 OI_SBC_SynthFrame_80 /* stereo */ 365 }; 366 367 368 static const SYNTH_FRAME SynthFrame4SB[] = { 369 NULL, /* invalid */ 370 OI_SBC_SynthFrame_4SB, /* mono */ 371 OI_SBC_SynthFrame_4SB /* stereo */ 372 }; 373 374 PRIVATE void OI_SBC_SynthFrame(OI_CODEC_SBC_DECODER_CONTEXT *context, OI_INT16 *pcm, OI_UINT start_block, OI_UINT nrof_blocks) 375 { 376 OI_UINT nrof_subbands = context->common.frameInfo.nrof_subbands; 377 OI_UINT nrof_channels = context->common.frameInfo.nrof_channels; 378 379 OI_ASSERT(nrof_subbands == 4 || nrof_subbands == 8); 380 if (nrof_subbands == 4) { 381 SynthFrame4SB[nrof_channels](context, pcm, start_block, nrof_blocks); 382 #ifdef SBC_ENHANCED 383 } else if (context->common.frameInfo.enhanced) { 384 SynthFrameEnhanced[nrof_channels](context, pcm, start_block, nrof_blocks); 385 #endif /* SBC_ENHANCED */ 386 } else { 387 SynthFrame8SB[nrof_channels](context, pcm, start_block, nrof_blocks); 388 } 389 } 390 391 392 void SynthWindow40_int32_int32_symmetry_with_sum(OI_INT16 *pcm, SBC_BUFFER_T buffer[80], OI_UINT strideShift) 393 { 394 OI_INT32 pa; 395 OI_INT32 pb; 396 397 /* These values should be zero, since out[2] of the 4-band cosine modulation 398 * is always zero. */ 399 OI_ASSERT(buffer[ 2] == 0); 400 OI_ASSERT(buffer[10] == 0); 401 OI_ASSERT(buffer[18] == 0); 402 OI_ASSERT(buffer[26] == 0); 403 OI_ASSERT(buffer[34] == 0); 404 OI_ASSERT(buffer[42] == 0); 405 OI_ASSERT(buffer[50] == 0); 406 OI_ASSERT(buffer[58] == 0); 407 OI_ASSERT(buffer[66] == 0); 408 OI_ASSERT(buffer[74] == 0); 409 410 411 pa = dec_window_4[ 4] * (buffer[12] + buffer[76]); 412 pa += dec_window_4[ 8] * (buffer[16] - buffer[64]); 413 pa += dec_window_4[12] * (buffer[28] + buffer[60]); 414 pa += dec_window_4[16] * (buffer[32] - buffer[48]); 415 pa += dec_window_4[20] * buffer[44]; 416 pa = SCALE(-pa, 15); 417 CLIP_INT16(pa); 418 pcm[0 << strideShift] = (OI_INT16)pa; 419 420 421 pa = dec_window_4[ 1] * buffer[ 1]; pb = dec_window_4[ 1] * buffer[79]; 422 pb += dec_window_4[ 3] * buffer[ 3]; pa += dec_window_4[ 3] * buffer[77]; 423 pa += dec_window_4[ 5] * buffer[13]; pb += dec_window_4[ 5] * buffer[67]; 424 pb += dec_window_4[ 7] * buffer[15]; pa += dec_window_4[ 7] * buffer[65]; 425 pa += dec_window_4[ 9] * buffer[17]; pb += dec_window_4[ 9] * buffer[63]; 426 pb += dec_window_4[11] * buffer[19]; pa += dec_window_4[11] * buffer[61]; 427 pa += dec_window_4[13] * buffer[29]; pb += dec_window_4[13] * buffer[51]; 428 pb += dec_window_4[15] * buffer[31]; pa += dec_window_4[15] * buffer[49]; 429 pa += dec_window_4[17] * buffer[33]; pb += dec_window_4[17] * buffer[47]; 430 pb += dec_window_4[19] * buffer[35]; pa += dec_window_4[19] * buffer[45]; 431 pa = SCALE(-pa, 15); 432 CLIP_INT16(pa); 433 pcm[1 << strideShift] = (OI_INT16)(pa); 434 pb = SCALE(-pb, 15); 435 CLIP_INT16(pb); 436 pcm[3 << strideShift] = (OI_INT16)(pb); 437 438 439 pa = dec_window_4[2] * (/*buffer[ 2] + */ buffer[78]); /* buffer[ 2] is always zero */ 440 pa += dec_window_4[6] * (buffer[14] /* + buffer[66]*/); /* buffer[66] is always zero */ 441 pa += dec_window_4[10] * (/*buffer[18] + */ buffer[62]); /* buffer[18] is always zero */ 442 pa += dec_window_4[14] * (buffer[30] /* + buffer[50]*/); /* buffer[50] is always zero */ 443 pa += dec_window_4[18] * (/*buffer[34] + */ buffer[46]); /* buffer[34] is always zero */ 444 pa = SCALE(-pa, 15); 445 CLIP_INT16(pa); 446 pcm[2 << strideShift] = (OI_INT16)(pa); 447 } 448 449 450 /** 451 This routine implements the cosine modulation matrix for 4-subband 452 synthesis. This is called "matrixing" in the SBC specification. This 453 matrix, M4, can be factored into an 8-point Type II Discrete Cosine 454 Transform, DCTII_4 and a matrix S4, given here: 455 456 @code 457 __ __ 458 | 0 0 1 0 | 459 | 0 0 0 1 | 460 | 0 0 0 0 | 461 | 0 0 0 -1 | 462 S4 = | 0 0 -1 0 | 463 | 0 -1 0 0 | 464 | -1 0 0 0 | 465 |__ 0 -1 0 0 __| 466 467 M4 * in = S4 * (DCTII_4 * in) 468 @endcode 469 470 (DCTII_4 * in) is computed using a Fast Cosine Transform. The algorithm 471 here is based on an implementation computed by the SPIRAL computer 472 algebra system, manually converted to fixed-point arithmetic. S4 can be 473 implemented using only assignment and negation. 474 */ 475 PRIVATE void cosineModulateSynth4(SBC_BUFFER_T * RESTRICT out, OI_INT32 const * RESTRICT in) 476 { 477 OI_INT32 f0, f1, f2, f3, f4, f7, f8, f9, f10; 478 OI_INT32 y0, y1, y2, y3; 479 480 f0 = (in[0] - in[3]); 481 f1 = (in[0] + in[3]); 482 f2 = (in[1] - in[2]); 483 f3 = (in[1] + in[2]); 484 485 f4 = f1 - f3; 486 487 y0 = -SCALE(f1 + f3, DCT_SHIFT); 488 y2 = -SCALE(LONG_MULT_DCT(DCTII_4_K06_FIX, f4), DCT_SHIFT); 489 f7 = f0 + f2; 490 f8 = LONG_MULT_DCT(DCTII_4_K08_FIX, f0); 491 f9 = LONG_MULT_DCT(DCTII_4_K09_FIX, f7); 492 f10 = LONG_MULT_DCT(DCTII_4_K10_FIX, f2); 493 y3 = -SCALE(f8 + f9, DCT_SHIFT); 494 y1 = -SCALE(f10 - f9, DCT_SHIFT); 495 496 out[0] = (OI_INT16)-y2; 497 out[1] = (OI_INT16)-y3; 498 out[2] = (OI_INT16)0; 499 out[3] = (OI_INT16)y3; 500 out[4] = (OI_INT16)y2; 501 out[5] = (OI_INT16)y1; 502 out[6] = (OI_INT16)y0; 503 out[7] = (OI_INT16)y1; 504 } 505 506 507 508 /** 509 @} 510 */ 511