1 ;// 2 ;// Copyright (C) 2007-2008 ARM Limited 3 ;// 4 ;// Licensed under the Apache License, Version 2.0 (the "License"); 5 ;// you may not use this file except in compliance with the License. 6 ;// You may obtain a copy of the License at 7 ;// 8 ;// http://www.apache.org/licenses/LICENSE-2.0 9 ;// 10 ;// Unless required by applicable law or agreed to in writing, software 11 ;// distributed under the License is distributed on an "AS IS" BASIS, 12 ;// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 ;// See the License for the specific language governing permissions and 14 ;// limitations under the License. 15 ;// 16 ;// 17 ;// 18 ;// File Name: omxVCM4P2_MCReconBlock_s.s 19 ;// OpenMAX DL: v1.0.2 20 ;// Revision: 9641 21 ;// Date: Thursday, February 7, 2008 22 ;// 23 ;// 24 ;// 25 ;// 26 ;// Description: 27 ;// 28 ;// 29 30 ;// Include standard headers 31 INCLUDE omxtypes_s.h 32 INCLUDE armCOMM_s.h 33 34 ;// Import symbols required from other files 35 36 M_VARIANTS ARM1136JS 37 38 ;// *************************************************************************** 39 ;// ARM1136JS implementation 40 ;// *************************************************************************** 41 IF ARM1136JS 42 43 ;// *************************************************************************** 44 ;// MACRO DEFINITIONS 45 ;// *************************************************************************** 46 ;// Description: 47 ;// 48 ;// dest[j] = (x[j] + y[j] + round) >> 1, j=0..3 49 ;// 50 ;// Similar to UHADD8 instruction, but with a rounding value of 1 added to 51 ;// each sum before dividing by two, if round is 1 52 ;// 53 ;// Syntax: 54 ;// M_UHADD8R $dest, $x, $y, $round, $mask 55 ;// 56 ;// Inputs: 57 ;// $x four packed bytes, x[3] : x[2] : x[1] : x[0] 58 ;// $y four packed bytes, y[3] : y[2] : y[1] : y[0] 59 ;// $round 0 if no rounding to be added, 1 if rounding to be done 60 ;// $mask some register set to 0x80808080 61 ;// 62 ;// Outputs: 63 ;// $dest four packed bytes, z[3] : z[2] : z[1] : z[0] 64 65 MACRO 66 M_UHADD8R $dest, $x, $y, $round, $mask 67 IF $round = 1 68 IF $dest /= $y 69 MVN $dest, $x 70 UHSUB8 $dest, $y, $dest 71 EOR $dest, $dest, $mask 72 ELSE 73 MVN $dest, $y 74 UHSUB8 $dest, $x, $dest 75 EOR $dest, $dest, $mask 76 ENDIF 77 ELSE 78 UHADD8 $dest, $x, $y 79 ENDIF 80 MEND 81 ;// *************************************************************************** 82 ;// Description: 83 ;// Load 8 bytes from $pSrc (aligned or unaligned locations) 84 ;// 85 ;// Syntax: 86 ;// M_LOAD_X $pSrc, $srcStep, $out0, $out1, $scratch, $offset 87 ;// 88 ;// Inputs: 89 ;// $pSrc 4 byte aligned source pointer to an address just less than 90 ;// or equal to the data location 91 ;// $srcStep The stride on source 92 ;// $scratch A scratch register, used internally for temp calculations 93 ;// $offset Difference of source data location to the source pointer 94 ;// Use when $offset != 0 (unaligned load) 95 ;// 96 ;// Outputs: 97 ;// $pSrc In case the macro accepts stride, it increments the pSrc by 98 ;// that value, else unchanged 99 ;// $out0 four packed bytes, z[3] : z[2] : z[1] : z[0] 100 ;// $out1 four packed bytes, z[7] : z[6] : z[5] : z[4] 101 ;// 102 ;// Note: {$out0, $out1, $scratch} should be registers with ascending 103 ;// register numbering. In case offset is 0, $scratch is not modified. 104 105 MACRO 106 M_LOAD_X $pSrc, $srcStep, $out0, $out1, $scratch, $offset 107 IF $offset = 0 108 LDM $pSrc, {$out0, $out1} 109 ADD $pSrc, $pSrc, $srcStep 110 ELSE 111 LDM $pSrc, {$out0, $out1, $scratch} 112 ADD $pSrc, $pSrc, $srcStep 113 114 MOV $out0, $out0, LSR #8 * $offset 115 ORR $out0, $out0, $out1, LSL #(32 - 8 * ($offset)) 116 MOV $out1, $out1, LSR #8 * $offset 117 ORR $out1, $out1, $scratch, LSL #(32 - 8 * ($offset)) 118 ENDIF 119 MEND 120 121 ;// *************************************************************************** 122 ;// Description: 123 ;// Loads three words for X interpolation, update pointer to next row. For 124 ;// X interpolation, given a truncated-4byteAligned source pointer, 125 ;// invariably three continous words are required from there to get the 126 ;// nine bytes from the source pointer for filtering. 127 ;// 128 ;// Syntax: 129 ;// M_LOAD_XINT $pSrc, $srcStep, $offset, $word0, $word1, $word2, $word3 130 ;// 131 ;// Inputs: 132 ;// $pSrc 4 byte aligned source pointer to an address just less than 133 ;// or equal to the data location 134 ;// 135 ;// $srcStep The stride on source 136 ;// 137 ;// $offset Difference of source data location to the source pointer 138 ;// Use when $offset != 0 (unaligned load) 139 ;// 140 ;// Outputs: 141 ;// $pSrc Incremented by $srcStep 142 ;// 143 ;// $word0, $word1, $word2, $word3 144 ;// Three of these are outputs based on the $offset parameter. 145 ;// The outputs are specifically generated to be processed by 146 ;// the M_EXT_XINT macro. Following is the illustration to show 147 ;// how the nine bytes are spanned for different offsets from 148 ;// notTruncatedForAlignmentSourcePointer. 149 ;// 150 ;// ------------------------------------------------------ 151 ;// | Offset | Aligned Ptr | word0 | word1 | word2 | word3 | 152 ;// |------------------------------------------------------| 153 ;// | 0 | 0 | 0123 | 4567 | 8xxx | | 154 ;// | 1 | -1 | x012 | 3456 | 78xx | | 155 ;// | 2 | -2 | xx01 | 2345 | 678x | | 156 ;// | 3 | -3 | xxx0 | | 1234 | 5678 | 157 ;// ------------------------------------------------------ 158 ;// 159 ;// where the numbering (0-8) is to designate the 9 bytes from 160 ;// start of a particular row. The illustration doesn't take in 161 ;// account the positioning of bytes with in the word and the 162 ;// macro combination with M_EXT_XINT will work only in little 163 ;// endian environs 164 ;// 165 ;// Note: {$word0, $word1, $word2, $word3} should be registers with ascending 166 ;// register numbering 167 168 MACRO 169 M_LOAD_XINT $pSrc, $srcStep, $offset, $word0, $word1, $word2, $word3 170 IF $offset /= 3 171 LDM $pSrc, {$word0, $word1, $word2} 172 ELSE 173 LDM $pSrc, {$word0, $word2, $word3} 174 ENDIF 175 ADD $pSrc, $pSrc, $srcStep 176 MEND 177 178 ;// *************************************************************************** 179 ;// Description: 180 ;// Extract four registers of four pixels for X interpolation 181 ;// 182 ;// Syntax: 183 ;// M_EXT_XINT $offset, $word0, $word1, $word2, $word3 184 ;// 185 ;// Inputs: 186 ;// $offset Difference of source data location to the source pointer 187 ;// Use when $offset != 0 (unaligned load) 188 ;// 189 ;// $word0, $word1, $word2, $word3 190 ;// Three of these are inputs based on the $offset parameter. 191 ;// The inputs are specifically selected to be processed by 192 ;// the M_EXT_XINT macro. 193 ;// 194 ;// ------------------------------------------------------ 195 ;// | Offset | Aligned Ptr | word0 | word1 | word2 | word3 | 196 ;// |------------------------------------------------------| 197 ;// | 0 | 0 | 0123 | 4567 | 8xxx | yyyy | 198 ;// | 1 | -1 | x012 | 3456 | 78xx | yyyy | 199 ;// | 2 | -2 | xx01 | 2345 | 678x | yyyy | 200 ;// | 3 | -3 | xxx0 | yyyy | 1234 | 5678 | 201 ;// ------------------------------------------------------ 202 ;// 203 ;// Outputs: 204 ;// $word0, $word1, $word2, $word3 205 ;// Bytes from the original source pointer (not truncated for 206 ;// 4 byte alignment) as shown in the table. 207 ;// ------------------------------- 208 ;// | word0 | word1 | word2 | word3 | 209 ;// |-------------------------------| 210 ;// | 0123 | 4567 | 1234 | 5678 | 211 ;// ------------------------------- 212 ;// 213 ;// Note: {$word0, $word1, $word2, $word3} should be registers with ascending 214 ;// register numbering 215 216 MACRO 217 M_EXT_XINT $offset, $word0, $word1, $word2, $word3 218 IF $offset = 0 219 ; $word0 and $word1 are ok 220 ; $word2, $word3 are just 8 shifted versions 221 MOV $word3, $word1, LSR #8 222 ORR $word3, $word3, $word2, LSL #24 223 MOV $word2, $word0, LSR #8 224 ORR $word2, $word2, $word1, LSL #24 225 ELIF $offset = 3 226 ; $word2 and $word3 are ok (taken care while loading itself) 227 ; set $word0 & $word1 228 MOV $word0, $word0, LSR #24 229 ORR $word0, $word0, $word2, LSL #8 230 MOV $word1, $word2, LSR #24 231 ORR $word1, $word1, $word3, LSL #8 232 ELSE 233 MOV $word0, $word0, LSR #8 * $offset 234 ORR $word0, $word0, $word1, LSL #(32 - 8 * ($offset)) 235 MOV $word1, $word1, LSR #8 * $offset 236 ORR $word1, $word1, $word2, LSL #(32 - 8 * ($offset)) 237 238 MOV $word3, $word1, LSR #8 239 ORR $word3, $word3, $word2, LSL #(32 - 8 * (($offset)+1)) 240 MOV $word2, $word0, LSR #8 241 ORR $word2, $word2, $word1, LSL #24 242 ENDIF 243 MEND 244 245 ;// *************************************************************************** 246 ;// Description: 247 ;// Computes half-sum and xor of two inputs and puts them in the input 248 ;// registers in that order 249 ;// 250 ;// Syntax: 251 ;// M_HSUM_XOR $v0, $v1, $tmp 252 ;// 253 ;// Inputs: 254 ;// $v0 a, first input 255 ;// $v1 b, second input 256 ;// $tmp scratch register 257 ;// 258 ;// Outputs: 259 ;// $v0 (a + b)/2 260 ;// $v1 a ^ b 261 262 MACRO 263 M_HSUM_XOR $v0, $v1, $tmp 264 UHADD8 $tmp, $v0, $v1 ;// s0 = a + b 265 EOR $v1, $v0, $v1 ;// l0 = a ^ b 266 MOV $v0, $tmp ;// s0 267 MEND 268 ;// *************************************************************************** 269 ;// Description: 270 ;// Calculates average of 4 values (a,b,c,d) for HalfPixelXY predict type in 271 ;// mcReconBlock module. Very specific to the implementation of 272 ;// M_MCRECONBLOCK_HalfPixelXY done here. Uses "tmp" as scratch register and 273 ;// "yMask" for mask variable "0x1010101x" set in it. In yMask 4 lsbs are 274 ;// not significant and are used by the callee for row counter (y) 275 ;// 276 ;// Some points to note are: 277 ;// 1. Input is pair of pair-averages and Xors 278 ;// 2. $sum1 and $lsb1 are not modified and hence can be reused in another 279 ;// running average 280 ;// 3. Output is in the first argument 281 ;// 282 ;// Syntax: 283 ;// M_AVG4 $sum0, $lsb0, $sum1, $lsb1, $rndVal 284 ;// 285 ;// Inputs: 286 ;// $sum0 (a + b) >> 1, where a and b are 1st and 2nd inputs to be averaged 287 ;// $lsb0 (a ^ b) 288 ;// $sum1 (c + d) >> 1. Not modified 289 ;// $lsb1 (c ^ d) Not modified 290 ;// $rndVal Assembler Variable. 0 for rounding, 1 for no rounding 291 ;// 292 ;// Outputs: 293 ;// $sum0 (a + b + c + d + 1) / 4 : If no rounding 294 ;// (a + b + c + d + 2) / 4 : If rounding 295 296 MACRO 297 M_AVG4 $sum0, $lsb0, $sum1, $lsb1, $rndVal 298 LCLS OP1 299 LCLS OP2 300 IF $rndVal = 0 ;// rounding case 301 OP1 SETS "AND" 302 OP2 SETS "ORR" 303 ELSE ;// Not rounding case 304 OP1 SETS "ORR" 305 OP2 SETS "AND" 306 ENDIF 307 308 LCLS lsb2 309 LCLS sum2 310 LCLS dest 311 312 lsb2 SETS "tmp" 313 sum2 SETS "$lsb0" 314 dest SETS "$sum0" 315 316 $OP1 $lsb0, $lsb0, $lsb1 ;// e0 = e0 & e1 317 EOR $lsb2, $sum0, $sum1 ;// e2 = s0 ^ s1 318 $OP2 $lsb2, $lsb2, $lsb0 ;// e2 = e2 | e0 319 AND $lsb2, $lsb2, yMask, LSR # 4 ;// e2 = e2 & mask 320 UHADD8 $sum2, $sum0, $sum1 ;// s2 = (s0 + s1)/2 321 UADD8 $dest, $sum2, $lsb2 ;// dest = s2 + e2 322 MEND 323 ;// *************************************************************************** 324 ;// Motion compensation handler macros 325 ;// *************************************************************************** 326 ;// Description: 327 ;// Implement motion compensation routines using the named registers in 328 ;// callee function. Each of the following 4 implement the 4 predict type 329 ;// Each handles 8 cases each ie all the combinations of 4 types of source 330 ;// alignment offsets and 2 types of rounding flag 331 ;// 332 ;// Syntax: 333 ;// M_MCRECONBLOCK_IntegerPixel $rndVal, $offset 334 ;// M_MCRECONBLOCK_HalfPixelX $rndVal, $offset 335 ;// M_MCRECONBLOCK_HalfPixelY $rndVal, $offset 336 ;// M_MCRECONBLOCK_HalfPixelXY $rndVal, $offset 337 ;// 338 ;// Inputs: 339 ;// $rndVal Assembler Variable. 0 for rounding, 1 for no rounding 340 ;// $offset $pSrc MOD 4 value. Offset from 4 byte aligned location. 341 ;// 342 ;// Outputs: 343 ;// Outputs come in the named registers of the callee functions 344 ;// The macro loads the data from the source pointer, processes it and 345 ;// stores in the destination pointer. Does the whole prediction cycle 346 ;// of Motion Compensation routine for a particular predictType 347 ;// After this only residue addition to the predicted values remain 348 349 MACRO 350 M_MCRECONBLOCK_IntegerPixel $rndVal, $offset 351 ;// Algorithmic Description: 352 ;// This handles motion compensation for IntegerPixel predictType. Both 353 ;// rounding cases are handled by the same code base. It is just a copy 354 ;// from source to destination. Two lines are done per loop to reduce 355 ;// stalls. Loop has been software pipelined as well for that purpose. 356 ;// 357 ;// M_LOAD_X loads a whole row in two registers and then they are stored 358 359 CaseIntegerPixelRnd0Offset$offset 360 CaseIntegerPixelRnd1Offset$offset 361 M_LOAD_X pSrc, srcStep, tmp1, tmp2, tmp3, $offset 362 M_LOAD_X pSrc, srcStep, tmp3, tmp4, tmp5, $offset 363 YloopIntegerPixelOffset$offset 364 SUBS y, y, #2 365 STRD tmp1, tmp2, [pDst], dstStep 366 STRD tmp3, tmp4, [pDst], dstStep 367 M_LOAD_X pSrc, srcStep, tmp1, tmp2, tmp3, $offset 368 M_LOAD_X pSrc, srcStep, tmp3, tmp4, tmp5, $offset 369 BGT YloopIntegerPixelOffset$offset 370 371 B SwitchPredictTypeEnd 372 MEND 373 ;// *************************************************************************** 374 MACRO 375 M_MCRECONBLOCK_HalfPixelX $rndVal, $offset 376 ;// Algorithmic Description: 377 ;// This handles motion compensation for HalfPixelX predictType. The two 378 ;// rounding cases are handled by the different code base and spanned by 379 ;// different macro calls. Loop has been software pipelined to reduce 380 ;// stalls. 381 ;// 382 ;// Filtering involves averaging a pixel with the next horizontal pixel. 383 ;// M_LOAD_XINT and M_EXT_XINT combination generate 4 registers, 2 with 384 ;// all pixels in a row with 4 pixel in each register and another 2 385 ;// registers with pixels corresponding to one horizontally shifted pixel 386 ;// corresponding to the initial row pixels. These are set of packed 387 ;// registers appropriate to do 4 lane SIMD. 388 ;// After that M_UHADD8R macro does the averaging taking care of the 389 ;// rounding as required 390 391 CaseHalfPixelXRnd$rndVal.Offset$offset 392 IF $rndVal = 0 393 LDR mask, =0x80808080 394 ENDIF 395 396 M_LOAD_XINT pSrc, srcStep, $offset, tmp1, tmp2, tmp3, tmp4 397 YloopHalfPixelXRnd$rndVal.Offset$offset 398 SUBS y, y, #1 399 M_EXT_XINT $offset, tmp1, tmp2, tmp3, tmp4 400 M_UHADD8R tmp5, tmp1, tmp3, (1-$rndVal), mask 401 M_UHADD8R tmp6, tmp2, tmp4, (1-$rndVal), mask 402 STRD tmp5, tmp6, [pDst], dstStep 403 M_LOAD_XINT pSrc, srcStep, $offset, tmp1, tmp2, tmp3, tmp4 404 BGT YloopHalfPixelXRnd$rndVal.Offset$offset 405 406 B SwitchPredictTypeEnd 407 MEND 408 ;// *************************************************************************** 409 MACRO 410 M_MCRECONBLOCK_HalfPixelY $rndVal, $offset 411 ;// Algorithmic Description: 412 ;// This handles motion compensation for HalfPixelY predictType. The two 413 ;// rounding cases are handled by the different code base and spanned by 414 ;// different macro calls. PreLoading is used to avoid reload of same data. 415 ;// 416 ;// Filtering involves averaging a pixel with the next vertical pixel. 417 ;// M_LOAD_X generates 2 registers with all pixels in a row with 4 pixel in 418 ;// each register. These are set of packed registers appropriate to do 419 ;// 4 lane SIMD. After that M_UHADD8R macro does the averaging taking care 420 ;// of the rounding as required 421 422 CaseHalfPixelYRnd$rndVal.Offset$offset 423 IF $rndVal = 0 424 LDR mask, =0x80808080 425 ENDIF 426 427 M_LOAD_X pSrc, srcStep, tmp1, tmp2, tmp5, $offset ;// Pre-load 428 YloopHalfPixelYRnd$rndVal.Offset$offset 429 SUBS y, y, #2 430 ;// Processing one line 431 M_LOAD_X pSrc, srcStep, tmp3, tmp4, tmp5, $offset 432 M_UHADD8R tmp1, tmp1, tmp3, (1-$rndVal), mask 433 M_UHADD8R tmp2, tmp2, tmp4, (1-$rndVal), mask 434 STRD tmp1, tmp2, [pDst], dstStep 435 ;// Processing another line 436 M_LOAD_X pSrc, srcStep, tmp1, tmp2, tmp5, $offset 437 M_UHADD8R tmp3, tmp3, tmp1, (1-$rndVal), mask 438 M_UHADD8R tmp4, tmp4, tmp2, (1-$rndVal), mask 439 STRD tmp3, tmp4, [pDst], dstStep 440 441 BGT YloopHalfPixelYRnd$rndVal.Offset$offset 442 443 B SwitchPredictTypeEnd 444 MEND 445 ;// *************************************************************************** 446 MACRO 447 M_MCRECONBLOCK_HalfPixelXY $rndVal, $offset 448 ;// Algorithmic Description: 449 ;// This handles motion compensation for HalfPixelXY predictType. The two 450 ;// rounding cases are handled by the different code base and spanned by 451 ;// different macro calls. PreLoading is used to avoid reload of same data. 452 ;// 453 ;// Filtering involves averaging a pixel with the next vertical, horizontal 454 ;// and right-down diagonal pixels. Just as in HalfPixelX case, M_LOAD_XINT 455 ;// and M_EXT_XINT combination generates 4 registers with a row and its 456 ;// 1 pixel right shifted version, with 4 pixels in one register. Another 457 ;// call of that macro-combination gets another row. Then M_HSUM_XOR is 458 ;// called to get mutual half-sum and xor combinations of a row with its 459 ;// shifted version as they are inputs to the M_AVG4 macro which computes 460 ;// the 4 element average with rounding. Note that it is the half-sum/xor 461 ;// values that are preserved for next row as they can be re-used in the 462 ;// next call to the M_AVG4 and saves recomputation. 463 ;// Due to lack of register, the row counter and a masking value required 464 ;// in M_AVG4 are packed into a single register yMask where the last nibble 465 ;// holds the row counter values and rest holds the masking variable left 466 ;// shifted by 4 467 468 CaseHalfPixelXYRnd$rndVal.Offset$offset 469 LDR yMask, =((0x01010101 << 4) + 8) 470 471 M_LOAD_XINT pSrc, srcStep, $offset, t00, t01, t10, t11 ;// Load a, a', b, b' 472 M_EXT_XINT $offset, t00, t01, t10, t11 473 M_HSUM_XOR t00, t10, tmp ;// s0, l0 474 M_HSUM_XOR t01, t11, tmp ;// s0', l0' 475 476 YloopHalfPixelXYRnd$rndVal.Offset$offset 477 ;// Processsing one line 478 ;// t00, t01, t10, t11 required from previous loop 479 M_LOAD_XINT pSrc, srcStep, $offset, t20, t21, t30, t31 ;// Load c, c', d, d' 480 SUB yMask, yMask, #2 481 M_EXT_XINT $offset, t20, t21, t30, t31 482 M_HSUM_XOR t20, t30, tmp ;// s1, l1 483 M_HSUM_XOR t21, t31, tmp ;// s1', l1' 484 M_AVG4 t00, t10, t20, t30, $rndVal ;// s0, l0, s1, l1 485 M_AVG4 t01, t11, t21, t31, $rndVal ;// s0', l0', s1', l1' 486 STRD t00, t01, [pDst], dstStep ;// store the average 487 488 ;// Processsing another line 489 ;// t20, t21, t30, t31 required from above 490 M_LOAD_XINT pSrc, srcStep, $offset, t00, t01, t10, t11 ;// Load a, a', b, b' 491 TST yMask, #7 492 M_EXT_XINT $offset, t00, t01, t10, t11 493 M_HSUM_XOR t00, t10, tmp 494 M_HSUM_XOR t01, t11, tmp 495 M_AVG4 t20, t30, t00, t10, $rndVal 496 M_AVG4 t21, t31, t01, t11, $rndVal 497 STRD t20, t21, [pDst], dstStep 498 499 BGT YloopHalfPixelXYRnd$rndVal.Offset$offset 500 501 IF $offset/=3 :LOR: $rndVal/=1 502 B SwitchPredictTypeEnd 503 ENDIF 504 MEND 505 ;// *************************************************************************** 506 ;// Motion compensation handler macros end here 507 ;// *************************************************************************** 508 ;// Description: 509 ;// Populates all 4 kinds of offsets "cases" for each predictType and rndVal 510 ;// combination in the "switch" to prediction processing code segment 511 ;// 512 ;// Syntax: 513 ;// M_CASE_OFFSET $rnd, $predictType 514 ;// 515 ;// Inputs: 516 ;// $rnd 0 for rounding, 1 for no rounding 517 ;// $predictType The prediction mode 518 ;// 519 ;// Outputs: 520 ;// Populated list of "M_CASE"s for the "M_SWITCH" macro 521 522 MACRO 523 M_CASE_OFFSET $rnd, $predictType 524 M_CASE Case$predictType.Rnd$rnd.Offset0 525 M_CASE Case$predictType.Rnd$rnd.Offset1 526 M_CASE Case$predictType.Rnd$rnd.Offset2 527 M_CASE Case$predictType.Rnd$rnd.Offset3 528 MEND 529 ;// *************************************************************************** 530 ;// Description: 531 ;// Populates all 2 kinds of rounding "cases" for each predictType in the 532 ;// "switch" to prediction processing code segment 533 ;// 534 ;// Syntax: 535 ;// M_CASE_OFFSET $predictType 536 ;// 537 ;// Inputs: 538 ;// $predictType The prediction mode 539 ;// 540 ;// Outputs: 541 ;// Populated list of "M_CASE_OFFSET" macros 542 543 MACRO 544 M_CASE_MCRECONBLOCK $predictType 545 M_CASE_OFFSET 0, $predictType ;// 0 for rounding 546 M_CASE_OFFSET 1, $predictType ;// 1 for no rounding 547 MEND 548 ;// *************************************************************************** 549 ;// Description: 550 ;// Populates all 8 kinds of rounding and offset combinations handling macros 551 ;// for the specified predictType. In case of "IntegerPixel" predictType, 552 ;// rounding is not required so same code segment handles both cases 553 ;// 554 ;// Syntax: 555 ;// M_MCRECONBLOCK $predictType 556 ;// 557 ;// Inputs: 558 ;// $predictType The prediction mode 559 ;// 560 ;// Outputs: 561 ;// Populated list of "M_MCRECONBLOCK_<predictType>" macros for specified 562 ;// predictType. Each 563 ;// M_MCRECONBLOCK_<predictType> $rnd, $offset 564 ;// is an code segment (starting with a label indicating the predictType, 565 ;// rounding and offset combination) 566 ;// Four calls of this macro with the 4 prediction modes populate all the 32 567 ;// handlers 568 569 MACRO 570 M_MCRECONBLOCK $predictType 571 M_MCRECONBLOCK_$predictType 0, 0 572 M_MCRECONBLOCK_$predictType 0, 1 573 M_MCRECONBLOCK_$predictType 0, 2 574 M_MCRECONBLOCK_$predictType 0, 3 575 IF "$predictType" /= "IntegerPixel" ;// If not IntegerPixel then rounding makes a difference 576 M_MCRECONBLOCK_$predictType 1, 0 577 M_MCRECONBLOCK_$predictType 1, 1 578 M_MCRECONBLOCK_$predictType 1, 2 579 M_MCRECONBLOCK_$predictType 1, 3 580 ENDIF 581 MEND 582 ;// *************************************************************************** 583 ;// Input/Output Registers 584 pSrc RN 0 585 srcStep RN 1 586 arg_pSrcResidue RN 2 587 pSrcResidue RN 12 588 pDst RN 3 589 dstStep RN 2 590 predictType RN 10 591 rndVal RN 11 592 mask RN 11 593 594 ;// Local Scratch Registers 595 zero RN 12 596 y RN 14 597 598 tmp1 RN 4 599 tmp2 RN 5 600 tmp3 RN 6 601 tmp4 RN 7 602 tmp5 RN 8 603 tmp6 RN 9 604 tmp7 RN 10 605 tmp8 RN 11 606 tmp9 RN 12 607 608 t00 RN 4 609 t01 RN 5 610 t10 RN 6 611 t11 RN 7 612 t20 RN 8 613 t21 RN 9 614 t30 RN 10 615 t31 RN 11 616 tmp RN 12 617 618 yMask RN 14 619 620 dst RN 1 621 return RN 0 622 623 ;// Allocate memory on stack 624 M_ALLOC4 Stk_pDst, 4 625 M_ALLOC4 Stk_pSrcResidue, 4 626 ;// Function header 627 M_START omxVCM4P2_MCReconBlock, r11 628 ;// Define stack arguments 629 M_ARG Arg_dstStep, 4 630 M_ARG Arg_predictType, 4 631 M_ARG Arg_rndVal, 4 632 ;// Save on stack 633 M_STR pDst, Stk_pDst 634 M_STR arg_pSrcResidue, Stk_pSrcResidue 635 ;// Load argument from the stack 636 M_LDR dstStep, Arg_dstStep 637 M_LDR predictType, Arg_predictType 638 M_LDR rndVal, Arg_rndVal 639 640 MOV y, #8 641 642 AND tmp1, pSrc, #3 643 ORR predictType, tmp1, predictType, LSL #3 644 ORR predictType, predictType, rndVal, LSL #2 645 ;// Truncating source pointer to align to 4 byte location 646 BIC pSrc, pSrc, #3 647 648 ;// Implementation takes care of all combinations of different 649 ;// predictTypes, rounding cases and source pointer offsets to alignment 650 ;// of 4 bytes in different code bases unless one of these parameter wasn't 651 ;// making any difference to the implementation. Below M_CASE_MCRECONBLOCK 652 ;// macros branch into 8 M_CASE macros for all combinations of the 2 653 ;// rounding cases and 4 offsets of the pSrc pointer to the 4 byte 654 ;// alignment. 655 M_SWITCH predictType 656 M_CASE_MCRECONBLOCK IntegerPixel 657 M_CASE_MCRECONBLOCK HalfPixelX 658 M_CASE_MCRECONBLOCK HalfPixelY 659 M_CASE_MCRECONBLOCK HalfPixelXY 660 M_ENDSWITCH 661 662 ;// The M_MCRECONBLOCK macros populate the code bases by calling all 8 663 ;// particular macros (4 in case of IntegerPixel as rounding makes no 664 ;// difference there) to generate the code for all cases of rounding and 665 ;// offsets. LTORG is used to segment the code as code size bloated beyond 666 ;// 4KB. 667 M_MCRECONBLOCK IntegerPixel 668 M_MCRECONBLOCK HalfPixelX 669 LTORG 670 M_MCRECONBLOCK HalfPixelY 671 M_MCRECONBLOCK HalfPixelXY 672 SwitchPredictTypeEnd 673 674 ;// Residue Addition 675 ;// This is done in 2 lane SIMD though loads are further optimized and 676 ;// 4 bytes are loaded in case of destination buffer. Algorithmic 677 ;// details are in inlined comments 678 M_LDR pSrcResidue, Stk_pSrcResidue 679 CMP pSrcResidue, #0 680 BEQ pSrcResidueConditionEnd 681 pSrcResidueNotNull 682 M_LDR pDst, Stk_pDst 683 MOV y, #8 684 SUB dstStep, dstStep, #4 685 Yloop_pSrcResidueNotNull 686 SUBS y, y, #1 687 LDR dst, [pDst] ;// dst = [dcba] 688 LDMIA pSrcResidue!, {tmp1, tmp2} ;// tmp1=[DC] tmp2=[BA] 689 PKHBT tmp3, tmp1, tmp2, LSL #16 ;// Deltaval1 = [C A] 690 PKHTB tmp4, tmp2, tmp1, ASR #16 ;// DeltaVal2 = [D B] 691 UXTB16 tmp1, dst ;// tmp1 = [0c0a] 692 UXTB16 tmp2, dst, ROR #8 ;// tmp2 = [0d0b] 693 QADD16 tmp1, tmp1, tmp3 ;// Add and saturate to 16 bits 694 QADD16 tmp2, tmp2, tmp4 695 USAT16 tmp1, #8, tmp1 696 USAT16 tmp2, #8, tmp2 ;// armClip(0, 255, tmp2) 697 ORR tmp1, tmp1, tmp2, LSL #8 ;// tmp1 = [dcba] 698 STR tmp1, [pDst], #4 699 700 LDR dst, [pDst] 701 LDMIA pSrcResidue!, {tmp1, tmp2} 702 PKHBT tmp3, tmp1, tmp2, LSL #16 703 PKHTB tmp4, tmp2, tmp1, ASR #16 704 UXTB16 tmp1, dst 705 UXTB16 tmp2, dst, ROR #8 706 QADD16 tmp1, tmp1, tmp3 707 QADD16 tmp2, tmp2, tmp4 708 USAT16 tmp1, #8, tmp1 709 USAT16 tmp2, #8, tmp2 710 ORR tmp1, tmp1, tmp2, LSL #8 711 STR tmp1, [pDst], dstStep 712 713 BGT Yloop_pSrcResidueNotNull 714 pSrcResidueConditionEnd 715 716 MOV return, #OMX_Sts_NoErr 717 718 M_END 719 ENDIF ;// ARM1136JS 720 721 ;// *************************************************************************** 722 ;// CortexA8 implementation 723 ;// *************************************************************************** 724 END 725 ;// *************************************************************************** 726 ;// omxVCM4P2_MCReconBlock ends 727 ;// *************************************************************************** 728