| 1 |
// **************************************************************************** |
| 2 |
// * |
| 3 |
// * XVID MPEG-4 VIDEO CODEC |
| 4 |
// * - IA64 forward discrete cosine transform - |
| 5 |
// * |
| 6 |
// * Copyright(C) 2002 Stephan Krause, Ingo-Marc Weber, Daniel Kallfass |
| 7 |
// * |
| 8 |
// * This program is free software; you can redistribute it and/or modify it |
| 9 |
// * under the terms of the GNU General Public License as published by |
| 10 |
// * the Free Software Foundation; either version 2 of the License, or |
| 11 |
// * (at your option) any later version. |
| 12 |
// * |
| 13 |
// * This program is distributed in the hope that it will be useful, |
| 14 |
// * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 15 |
// * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 16 |
// * GNU General Public License for more details. |
| 17 |
// * |
| 18 |
// * You should have received a copy of the GNU General Public License |
| 19 |
// * along with this program; if not, write to the Free Software |
| 20 |
// * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
| 21 |
// * |
| 22 |
// * $Id: fdct_ia64.s,v 1.6 2009-02-19 17:07:29 Isibaar Exp $ |
| 23 |
// * |
| 24 |
// ***************************************************************************/ |
| 25 |
// |
| 26 |
// **************************************************************************** |
| 27 |
// * |
| 28 |
// * fdct_ia64.s, IA-64 optimized forward DCT |
| 29 |
// * |
| 30 |
// * Completed version provided by Intel at AppNote AP-922 |
| 31 |
// * http://developer.intel.com/software/products/college/ia32/strmsimd/ |
| 32 |
// * Copyright (C) 1999 Intel Corporation, |
| 33 |
// * |
| 34 |
// * This version was implemented during an IA-64 practical training at |
| 35 |
// * the University of Karlsruhe (http://i44w3.info.uni-karlsruhe.de/) |
| 36 |
// * |
| 37 |
// ***************************************************************************** |
| 38 |
// |
| 39 |
// ***************************************************************************** |
| 40 |
// * |
| 41 |
// * Revision history: |
| 42 |
// * |
| 43 |
// * 24.07.2002 Initial Version |
| 44 |
// * |
| 45 |
// ***************************************************************************** |
| 46 |
|
| 47 |
|
| 48 |
// This is a fast precise implementation of 8x8 Discrete Cosine Transform |
| 49 |
// published in Intel Application Note 922 from 1999 and optimized for IA-64. |
| 50 |
// |
| 51 |
// An unoptimized "straight forward" version can be found at the end of this file. |
| 52 |
|
| 53 |
|
| 54 |
.pred.safe_across_calls p1-p5,p16-p63 |
| 55 |
.text |
| 56 |
.align 16 |
| 57 |
.global fdct_ia64# |
| 58 |
.proc fdct_ia64# |
| 59 |
fdct_ia64: |
| 60 |
.prologue |
| 61 |
alloc r14 = ar.pfs, 1, 56, 0, 0 |
| 62 |
// Save constants |
| 63 |
mov r31 = 0x32ec // c0 = tan(1pi/16) |
| 64 |
mov r30 = 0x6a0a // c1 = tan(2pi/16) |
| 65 |
mov r29 = 0xab0e // c2 = tan(3pi/16) |
| 66 |
mov r28 = 0xb505 // g4 = cos(4pi/16) |
| 67 |
mov r27 = 0xd4db // g3 = cos(3pi/16) |
| 68 |
mov r26 = 0xec83 // g2 = cos(2pi/16) |
| 69 |
mov r25 = 0xfb15 // g1 = cos(1pi/16) |
| 70 |
mov r24 = 0x0002 // correction bit for descaling |
| 71 |
mov r23 = 0x0004 // correction bit for descaling |
| 72 |
|
| 73 |
// Load Matrix into registers |
| 74 |
|
| 75 |
add loc0 = r0, r32 |
| 76 |
add loc2 = 16, r32 |
| 77 |
add loc4 = 32, r32 |
| 78 |
add loc6 = 48, r32 |
| 79 |
add loc8 = 64, r32 |
| 80 |
add loc10 = 80, r32 |
| 81 |
add loc12 = 96, r32 |
| 82 |
add loc14 = 112, r32 |
| 83 |
add loc1 = 8, r32 |
| 84 |
add loc3 = 24, r32 |
| 85 |
add loc5 = 40, r32 |
| 86 |
add loc7 = 56, r32 |
| 87 |
add loc9 = 72, r32 |
| 88 |
add loc11 = 88, r32 |
| 89 |
add loc13 = 104, r32 |
| 90 |
add loc15 = 120, r32 |
| 91 |
;; |
| 92 |
ld8 loc16 = [loc0] |
| 93 |
ld8 loc17 = [loc2] |
| 94 |
ld8 loc18 = [loc4] |
| 95 |
ld8 loc19 = [loc6] |
| 96 |
ld8 loc20 = [loc8] |
| 97 |
ld8 loc21 = [loc10] |
| 98 |
ld8 loc22 = [loc12] |
| 99 |
ld8 loc23 = [loc14] |
| 100 |
ld8 loc24 = [loc1] |
| 101 |
ld8 loc25 = [loc3] |
| 102 |
ld8 loc26 = [loc5] |
| 103 |
ld8 loc27 = [loc7] |
| 104 |
mux2 r26 = r26, 0x00 |
| 105 |
ld8 loc28 = [loc9] |
| 106 |
mux2 r31 = r31, 0x00 |
| 107 |
mux2 r25 = r25, 0x00 |
| 108 |
ld8 loc29 = [loc11] |
| 109 |
mux2 r30 = r30, 0x00 |
| 110 |
mux2 r29 = r29, 0x00 |
| 111 |
ld8 loc30 = [loc13] |
| 112 |
mux2 r28 = r28, 0x00 |
| 113 |
mux2 r27 = r27, 0x00 |
| 114 |
ld8 loc31 = [loc15] |
| 115 |
mux2 r24 = r24, 0x00 |
| 116 |
mux2 r23 = r23, 0x00 |
| 117 |
;; |
| 118 |
pshl2 loc16 = loc16, 3 |
| 119 |
pshl2 loc17 = loc17, 3 |
| 120 |
pshl2 loc18 = loc18, 3 |
| 121 |
pshl2 loc19 = loc19, 3 |
| 122 |
pshl2 loc20 = loc20, 3 |
| 123 |
pshl2 loc21 = loc21, 3 |
| 124 |
pshl2 loc22 = loc22, 3 |
| 125 |
pshl2 loc23 = loc23, 3 |
| 126 |
;; |
| 127 |
pshl2 loc24 = loc24, 3 |
| 128 |
|
| 129 |
// ******************* |
| 130 |
// column-DTC 1st half |
| 131 |
// ******************* |
| 132 |
|
| 133 |
psub2 loc37 = loc17, loc22 // t5 = x1 - x6 |
| 134 |
pshl2 loc25 = loc25, 3 |
| 135 |
pshl2 loc26 = loc26, 3 |
| 136 |
psub2 loc38 = loc18, loc21 // t6 = x2 - x5 |
| 137 |
pshl2 loc27 = loc27, 3 |
| 138 |
pshl2 loc28 = loc28, 3 |
| 139 |
;; |
| 140 |
padd2 loc32 = loc16, loc23 // t0 = x0 + x7 |
| 141 |
pshl2 loc29 = loc29, 3 |
| 142 |
pshl2 loc30 = loc30, 3 |
| 143 |
padd2 loc33 = loc17, loc22 // t1 = x1 + x6 |
| 144 |
padd2 loc40 = loc37, loc38 // buf0 = t5 + t6 |
| 145 |
psub2 loc41 = loc37, loc38 // buf1 = t5 - t6 |
| 146 |
;; |
| 147 |
padd2 loc34 = loc18, loc21 // t2 = x2 + x5 |
| 148 |
pshl2 loc31 = loc31, 3 |
| 149 |
padd2 loc35 = loc19, loc20 // t3 = x3 + x4 |
| 150 |
psub2 loc36 = loc16, loc23 // t4 = x0 - x7 |
| 151 |
pmpyshr2 loc37 = loc40, r28, 16 // t5 = buf0 * g4 |
| 152 |
pmpyshr2 loc38 = loc41, r28, 16 // t6 = buf1 * g4 |
| 153 |
;; |
| 154 |
psub2 loc39 = loc19, loc20 // t7 = x3 - x4 |
| 155 |
padd2 loc37 = loc37, loc40 // t5 = t5 + buf1 |
| 156 |
padd2 loc38 = loc38, loc41 // t6 = t6 + buf2 |
| 157 |
|
| 158 |
padd2 loc16 = loc32, loc35 // x0 = t0 + t3 |
| 159 |
padd2 loc17 = loc33, loc34 // x1 = t1 + t2 |
| 160 |
psub2 loc18 = loc32, loc35 // x2 = t0 - t3 |
| 161 |
;; |
| 162 |
psub2 loc19 = loc33, loc34 // x3 = t1 - t2 |
| 163 |
padd2 loc20 = loc36, loc37 // x4 = t4 + t5 |
| 164 |
padd2 loc21 = loc38, loc39 // x5 = t6 + t7 |
| 165 |
psub2 loc22 = loc36, loc37 // x6 = t4 - t5 |
| 166 |
psub2 loc23 = loc38, loc39 // x7 = t6 - t7 |
| 167 |
;; |
| 168 |
pmpyshr2 loc42 = loc18, r30, 16 // buf2 = x2 * c1 |
| 169 |
padd2 loc32 = loc16, loc17 // t0 = x0 + x1 |
| 170 |
pmpyshr2 loc43 = loc19, r30, 16 // buf3 = x3 * c1 |
| 171 |
pmpyshr2 loc44 = loc20, r31, 16 // buf4 = x4 * c0 |
| 172 |
psub2 loc33 = loc16, loc17 // t1 = x0 - x1 |
| 173 |
pmpyshr2 loc46 = loc22, r29, 16 // buf6 = x6 * c2 |
| 174 |
pmpyshr2 loc47 = loc23, r29, 16 // buf7 = x7 * c2 |
| 175 |
;; |
| 176 |
padd2 loc46 = loc46, loc22 // buf6 = buf6 + x6 |
| 177 |
pmpyshr2 loc45 = loc21, r31, 16 // buf5 = x5 * c0 |
| 178 |
padd2 loc47 = loc47, loc23 // buf7 = buf7 + x7 |
| 179 |
padd2 loc34 = loc18, loc43 // t2 = x2 + (x3 * c1) |
| 180 |
psub2 loc35 = loc42, loc19 // t3 = (c1 * x2) - x3 |
| 181 |
psub2 loc37 = loc44, loc21 // t5 = (c1 * x4) - x5 |
| 182 |
;; |
| 183 |
padd2 loc36 = loc20, loc45 // t4 = x4 + (x5 * c1) |
| 184 |
padd2 loc38 = loc22, loc47 // t6 = x6 + (x7 * c1) |
| 185 |
pmpyshr2 loc16 = loc32, r28, 16 // x0 = t0 * g4 |
| 186 |
pmpyshr2 loc17 = loc33, r28, 16 // x1 = t1 * g4 |
| 187 |
psub2 loc39 = loc46, loc23 // t7 = (c1 * x6) - x7 |
| 188 |
;; |
| 189 |
padd2 loc48 = loc16, loc32 // y0 = x0 + t0 |
| 190 |
pmpyshr2 loc18 = loc34, r26, 16 // x2 = t2 * g2 |
| 191 |
pmpyshr2 loc19 = loc35, r26, 16 // x3 = t3 * g2 |
| 192 |
padd2 loc52 = loc17, loc33 // y4 = x1 + t1 |
| 193 |
pmpyshr2 loc20 = loc36, r25, 16 // x4 = t4 * g1 |
| 194 |
pmpyshr2 loc21 = loc37, r25, 16 // x5 = t5 * g1 |
| 195 |
;; |
| 196 |
padd2 loc50 = loc18, loc34 // y2 = x2 + t2 |
| 197 |
pmpyshr2 loc22 = loc38, r27, 16 // x6 = t6 * g3 |
| 198 |
pmpyshr2 loc23 = loc39, r27, 16 // x7 = t7 * g3 |
| 199 |
padd2 loc55 = loc21, loc37 // y7 = x5 + t5 |
| 200 |
padd2 loc49 = loc20, loc36 // y1 = x4 + t4 |
| 201 |
padd2 loc54 = loc19, loc35 // y6 = x3 + t3 |
| 202 |
;; |
| 203 |
padd2 loc51 = loc22, loc38 // y3 = x6 + t6 |
| 204 |
padd2 loc53 = loc23, loc39 // y5 = x7 + t7 |
| 205 |
|
| 206 |
//divide by 4 |
| 207 |
|
| 208 |
padd2 loc48 = loc48, r24 |
| 209 |
padd2 loc49 = loc49, r24 |
| 210 |
padd2 loc50 = loc50, r24 |
| 211 |
padd2 loc52 = loc52, r24 |
| 212 |
;; |
| 213 |
padd2 loc51 = loc51, r24 |
| 214 |
pshr2 loc48 = loc48, 2 |
| 215 |
padd2 loc53 = loc53, r24 |
| 216 |
pshr2 loc49 = loc49, 2 |
| 217 |
padd2 loc54 = loc54, r24 |
| 218 |
pshr2 loc50 = loc50, 2 |
| 219 |
padd2 loc55 = loc55, r24 |
| 220 |
pshr2 loc52 = loc52, 2 |
| 221 |
;; |
| 222 |
pshr2 loc51 = loc51, 2 |
| 223 |
pshr2 loc53 = loc53, 2 |
| 224 |
pshr2 loc54 = loc54, 2 |
| 225 |
pshr2 loc55 = loc55, 2 |
| 226 |
|
| 227 |
|
| 228 |
// ******************* |
| 229 |
// column-DTC 2nd half |
| 230 |
// ******************* |
| 231 |
|
| 232 |
psub2 loc37 = loc25, loc30 // t5 = x1.2 - x6.2 |
| 233 |
psub2 loc38 = loc26, loc29 // t6 = x2.2 - x5.2 |
| 234 |
padd2 loc32 = loc24, loc31 // t0 = x0.2 + x7.2 |
| 235 |
padd2 loc33 = loc25, loc30 // t1 = x1.2 + x6.2 |
| 236 |
;; |
| 237 |
padd2 loc34 = loc26, loc29 // t2 = x2.2 + x5.2 |
| 238 |
psub2 loc41 = loc37, loc38 // buf1 = t5 - t6 |
| 239 |
padd2 loc40 = loc37, loc38 // buf0 = t5 + t6 |
| 240 |
padd2 loc35 = loc27, loc28 // t3 = x3.2 + x4.2 |
| 241 |
;; |
| 242 |
psub2 loc36 = loc24, loc31 // t4 = x0.2 - x7.2 |
| 243 |
pmpyshr2 loc37 = loc40, r28, 16 // t5 = buf0 * g4 |
| 244 |
pmpyshr2 loc38 = loc41, r28, 16 // t6 = buf1 * g4 |
| 245 |
;; |
| 246 |
psub2 loc39 = loc27, loc28 // t7 = x3.2 - x4.2 |
| 247 |
padd2 loc37 = loc37, loc40 // t5 = t5 + buf1 |
| 248 |
padd2 loc38 = loc38, loc41 // t6 = t6 + buf2 |
| 249 |
|
| 250 |
padd2 loc16 = loc32, loc35 // x0 = t0 + t3 |
| 251 |
padd2 loc17 = loc33, loc34 // x1 = t1 + t2 |
| 252 |
psub2 loc18 = loc32, loc35 // x2 = t0 - t3 |
| 253 |
;; |
| 254 |
psub2 loc19 = loc33, loc34 // x3 = t1 - t2 |
| 255 |
padd2 loc20 = loc36, loc37 // x4 = t4 + t5 |
| 256 |
padd2 loc21 = loc38, loc39 // x5 = t6 + t7 |
| 257 |
psub2 loc22 = loc36, loc37 // x6 = t4 - t5 |
| 258 |
psub2 loc23 = loc38, loc39 // x7 = t6 - t7 |
| 259 |
;; |
| 260 |
pmpyshr2 loc42 = loc18, r30, 16 // buf2 = x2 * c1 |
| 261 |
padd2 loc32 = loc16, loc17 // t0 = x0 + x1 |
| 262 |
pmpyshr2 loc43 = loc19, r30, 16 // buf3 = x3 * c1 |
| 263 |
pmpyshr2 loc44 = loc20, r31, 16 // buf4 = x4 * c0 |
| 264 |
psub2 loc33 = loc16, loc17 // t1 = x0 - x1 |
| 265 |
pmpyshr2 loc45 = loc21, r31, 16 // buf5 = x5 * c0 |
| 266 |
pmpyshr2 loc46 = loc22, r29, 16 // buf6 = x6 * c2 |
| 267 |
pmpyshr2 loc47 = loc23, r29, 16 // buf7 = x7 * c2 |
| 268 |
;; |
| 269 |
padd2 loc34 = loc18, loc43 // t2 = x2 + buf3 |
| 270 |
padd2 loc46 = loc46, loc22 // buf6 = buf6 + x6 |
| 271 |
padd2 loc47 = loc47, loc23 // buf7 = buf7 + x7 |
| 272 |
psub2 loc35 = loc42, loc19 // t3 = buf2 - x3 |
| 273 |
padd2 loc36 = loc20, loc45 // t4 = x4 + buf5 |
| 274 |
pmpyshr2 loc16 = loc32, r28, 16 // x0 = t0 * g4 |
| 275 |
;; |
| 276 |
psub2 loc37 = loc44, loc21 // t5 = buf4 - x5 |
| 277 |
padd2 loc38 = loc22, loc47 // t6 = x6 + buf7 |
| 278 |
psub2 loc39 = loc46, loc23 // t7 = buf6 - x7 |
| 279 |
pmpyshr2 loc17 = loc33, r28, 16 // x1 = t1 * g4 |
| 280 |
;; |
| 281 |
padd2 loc40 = loc16, loc32 // y0.2 = x0 + t0 |
| 282 |
pmpyshr2 loc18 = loc34, r26, 16 // x2 = t2 * g2 |
| 283 |
pmpyshr2 loc19 = loc35, r26, 16 // x3 = t3 * g2 |
| 284 |
padd2 loc44 = loc17, loc33 // y4.2 = x1 + t1 |
| 285 |
pmpyshr2 loc20 = loc36, r25, 16 // x4 = t4 * g1 |
| 286 |
pmpyshr2 loc21 = loc37, r25, 16 // x5 = t5 * g1 |
| 287 |
;; |
| 288 |
padd2 loc42 = loc18, loc34 // y2.2 = x2 + t2 |
| 289 |
pmpyshr2 loc22 = loc38, r27, 16 // x6 = t6 * g3 |
| 290 |
pmpyshr2 loc23 = loc39, r27, 16 // x7 = t7 * g3 |
| 291 |
padd2 loc47 = loc21, loc37 // y7.2 = x5 + t5 |
| 292 |
padd2 loc41 = loc20, loc36 // y1.2 = x4 + t4 |
| 293 |
padd2 loc46 = loc19, loc35 // y6.2 = x3 + t3 |
| 294 |
;; |
| 295 |
padd2 loc43 = loc22, loc38 // y3.2 = x6 + t6 |
| 296 |
|
| 297 |
// ******************* |
| 298 |
// transpose matrix |
| 299 |
// ******************* |
| 300 |
|
| 301 |
mix2.r loc32 = loc48, loc49 // tmp0 = mixr y0, y1 |
| 302 |
mix2.l loc33 = loc48, loc49 // tmp1 = mixl y0, y1 |
| 303 |
padd2 loc45 = loc23, loc39 // y5.2 = x7 + t7 |
| 304 |
mix2.r loc34 = loc50, loc51 // tmp2 = mixr y2, y3 |
| 305 |
mix2.l loc35 = loc50, loc51 // tmp3 = mixl y2, y3 |
| 306 |
;; |
| 307 |
|
| 308 |
//divide by 4 |
| 309 |
|
| 310 |
padd2 loc40 = loc40, r24 |
| 311 |
padd2 loc41 = loc41, r24 |
| 312 |
mix4.r loc16 = loc32, loc34 // x0 = mixr tmp0, tmp2 |
| 313 |
padd2 loc42 = loc42, r24 |
| 314 |
padd2 loc43 = loc43, r24 |
| 315 |
mix4.r loc17 = loc33, loc35 // x1 = mixr tmp1, tmp3 |
| 316 |
padd2 loc44 = loc44, r24 |
| 317 |
padd2 loc45 = loc45, r24 |
| 318 |
mix4.l loc18 = loc32, loc34 // x2 = mixl tmp0, tmp2 |
| 319 |
padd2 loc46 = loc46, r24 |
| 320 |
padd2 loc47 = loc47, r24 |
| 321 |
mix4.l loc19 = loc33, loc35 // x3 = mixl tmp1, tmp3 |
| 322 |
;; |
| 323 |
pshr2 loc40 = loc40, 2 |
| 324 |
pshr2 loc41 = loc41, 2 |
| 325 |
pshr2 loc42 = loc42, 2 |
| 326 |
pshr2 loc43 = loc43, 2 |
| 327 |
mix2.r loc32 = loc52, loc53 // tmp0 = mixr y4, y5 |
| 328 |
mix2.l loc33 = loc52, loc53 // tmp1 = mixl y4, y5 |
| 329 |
mix2.r loc34 = loc54, loc55 // tmp2 = mixr y6, y7 |
| 330 |
mix2.l loc35 = loc54, loc55 // tmp3 = mixl y6, y7 |
| 331 |
;; |
| 332 |
pshr2 loc44 = loc44, 2 |
| 333 |
pshr2 loc45 = loc45, 2 |
| 334 |
pshr2 loc46 = loc46, 2 |
| 335 |
pshr2 loc47 = loc47, 2 |
| 336 |
mix4.r loc24 = loc32, loc34 // x0.2 = mixr tmp0, tmp2 |
| 337 |
mix4.r loc25 = loc33, loc35 // x1.2 = mixr tmp1, tmp3 |
| 338 |
mix4.l loc26 = loc32, loc34 // x2.2 = mixl tmp0, tmp2 |
| 339 |
mix4.l loc27 = loc33, loc35 // x3.2 = mixl tmp1, tmp3 |
| 340 |
;; |
| 341 |
mix2.r loc32 = loc40, loc41 // tmp0 = mixr y0.2, y1.2 |
| 342 |
mix2.l loc33 = loc40, loc41 // tmp1 = mixl y0.2, y1.2 |
| 343 |
mix2.r loc34 = loc42, loc43 // tmp2 = mixr y2.2, y3.2 |
| 344 |
mix2.l loc35 = loc42, loc43 // tmp3 = mixl y2.2, y3.2 |
| 345 |
;; |
| 346 |
mix4.r loc20 = loc32, loc34 // x4 = mixr tmp0, tmp2 |
| 347 |
mix4.r loc21 = loc33, loc35 // x5 = mixr tmp1, tmp3 |
| 348 |
mix4.l loc22 = loc32, loc34 // x6 = mixl tmp0, tmp2 |
| 349 |
mix4.l loc23 = loc33, loc35 // x7 = mixl tmp1, tmp3 |
| 350 |
;; |
| 351 |
mix2.r loc32 = loc44, loc45 // tmp0 = mixr y4.2, y5.2 |
| 352 |
mix2.l loc33 = loc44, loc45 // tmp1 = mixl y4.2, y5.2 |
| 353 |
mix2.r loc34 = loc46, loc47 // tmp2 = mixr y6.2, y6.2 |
| 354 |
mix2.l loc35 = loc46, loc47 // tmp3 = mixl y6.2, y6.2 |
| 355 |
;; |
| 356 |
mix4.r loc28 = loc32, loc34 // x4.2 = mixr tmp0, tmp2 |
| 357 |
mix4.r loc29 = loc33, loc35 // x5.2 = mixr tmp1, tmp3 |
| 358 |
mix4.l loc30 = loc32, loc34 // x6.2 = mixl tmp0, tmp2 |
| 359 |
mix4.l loc31 = loc33, loc35 // x7.2 = mixl tmp1, tmp3 |
| 360 |
|
| 361 |
// ******************* |
| 362 |
// row-DTC 1st half |
| 363 |
// ******************* |
| 364 |
|
| 365 |
psub2 loc37 = loc17, loc22 // t5 = x1 - x6 |
| 366 |
psub2 loc38 = loc18, loc21 // t6 = x2 - x5 |
| 367 |
;; |
| 368 |
padd2 loc32 = loc16, loc23 // t0 = x0 + x7 |
| 369 |
padd2 loc33 = loc17, loc22 // t1 = x1 + x6 |
| 370 |
padd2 loc34 = loc18, loc21 // t2 = x2 + x5 |
| 371 |
psub2 loc41 = loc37, loc38 // buf1 = t5 - t6 |
| 372 |
padd2 loc40 = loc37, loc38 // buf0 = t5 + t6 |
| 373 |
padd2 loc35 = loc19, loc20 // t3 = x3 + x4 |
| 374 |
;; |
| 375 |
psub2 loc36 = loc16, loc23 // t4 = x0 - x7 |
| 376 |
pmpyshr2 loc37 = loc40, r28, 16 // t5 = buf0 * g4 |
| 377 |
pmpyshr2 loc38 = loc41, r28, 16 // t6 = buf1 * g4 |
| 378 |
;; |
| 379 |
psub2 loc39 = loc19, loc20 // t7 = x3 - x4 |
| 380 |
padd2 loc37 = loc37, loc40 // t5 = t5 + buf1 |
| 381 |
padd2 loc38 = loc38, loc41 // t6 = t6 + buf2 |
| 382 |
|
| 383 |
padd2 loc16 = loc32, loc35 // x0 = t0 + t3 |
| 384 |
padd2 loc17 = loc33, loc34 // x1 = t1 + t2 |
| 385 |
psub2 loc18 = loc32, loc35 // x2 = t0 - t3 |
| 386 |
;; |
| 387 |
psub2 loc19 = loc33, loc34 // x3 = t1 - t2 |
| 388 |
padd2 loc20 = loc36, loc37 // x4 = t4 + t5 |
| 389 |
padd2 loc21 = loc38, loc39 // x5 = t6 + t7 |
| 390 |
psub2 loc22 = loc36, loc37 // x6 = t4 - t5 |
| 391 |
psub2 loc23 = loc38, loc39 // x7 = t6 - t7 |
| 392 |
;; |
| 393 |
pmpyshr2 loc42 = loc18, r30, 16 // buf2 = x2 * c1 |
| 394 |
padd2 loc32 = loc16, loc17 // t0 = x0 + x1 |
| 395 |
pmpyshr2 loc43 = loc19, r30, 16 // buf3 = x3 * c1 |
| 396 |
pmpyshr2 loc44 = loc20, r31, 16 // buf4 = x4 * c0 |
| 397 |
psub2 loc33 = loc16, loc17 // t1 = x0 - x1 |
| 398 |
pmpyshr2 loc45 = loc21, r31, 16 // buf5 = x5 * c0 |
| 399 |
pmpyshr2 loc46 = loc22, r29, 16 // buf6 = x6 * c2 |
| 400 |
pmpyshr2 loc47 = loc23, r29, 16 // buf7 = x7 * c2 |
| 401 |
;; |
| 402 |
padd2 loc46 = loc46, loc22 // buf6 = buf6 + x6 |
| 403 |
padd2 loc47 = loc47, loc23 // buf7 = buf7 + x7 |
| 404 |
padd2 loc34 = loc18, loc43 // t2 = x2 + (x3 * c1) |
| 405 |
;; |
| 406 |
psub2 loc35 = loc42, loc19 // t3 = (c1 * x2) - x3 |
| 407 |
padd2 loc36 = loc20, loc45 // t4 = x4 + (x5 * c1) |
| 408 |
psub2 loc37 = loc44, loc21 // t5 = (c1 * x4) - x5 |
| 409 |
padd2 loc38 = loc22, loc47 // t6 = x6 + (buf7 * c1) |
| 410 |
psub2 loc39 = loc46, loc23 // t7 = (c1 * buf6) - x7 |
| 411 |
pmpyshr2 loc16 = loc32, r28, 16 // x0 = t0 * g4 |
| 412 |
pmpyshr2 loc17 = loc33, r28, 16 // x1 = t1 * g4 |
| 413 |
;; |
| 414 |
padd2 loc48 = loc16, loc32 // y0 = x0 + t0 |
| 415 |
pmpyshr2 loc18 = loc34, r26, 16 // x2 = t2 * g2 |
| 416 |
pmpyshr2 loc19 = loc35, r26, 16 // x3 = t3 * g2 |
| 417 |
padd2 loc52 = loc17, loc33 // y4 = x1 + t1 |
| 418 |
pmpyshr2 loc20 = loc36, r25, 16 // x4 = t4 * g1 |
| 419 |
pmpyshr2 loc21 = loc37, r25, 16 // x5 = t5 * g1 |
| 420 |
;; |
| 421 |
padd2 loc50 = loc18, loc34 // y2 = x2 + t2 |
| 422 |
pmpyshr2 loc22 = loc38, r27, 16 // x6 = t6 * g3 |
| 423 |
pmpyshr2 loc23 = loc39, r27, 16 // x7 = t7 * g3 |
| 424 |
padd2 loc55 = loc21, loc37 // y7 = x5 + t5 |
| 425 |
padd2 loc49 = loc20, loc36 // y1 = x4 + t4 |
| 426 |
padd2 loc54 = loc19, loc35 // y6 = x3 + t3 |
| 427 |
;; |
| 428 |
padd2 loc51 = loc22, loc38 // y3 = x6 + t6 |
| 429 |
padd2 loc53 = loc23, loc39 // y5 = x7 + t7 |
| 430 |
|
| 431 |
// ******************* |
| 432 |
// row-DTC 2nd half |
| 433 |
// ******************* |
| 434 |
|
| 435 |
psub2 loc37 = loc25, loc30 // t5 = x1.2 - x6.2 |
| 436 |
psub2 loc38 = loc26, loc29 // t6 = x2.2 - x5.2 |
| 437 |
padd2 loc32 = loc24, loc31 // t0 = x0.2 + x7.2 |
| 438 |
padd2 loc33 = loc25, loc30 // t1 = x1.2 + x6.2 |
| 439 |
;; |
| 440 |
padd2 loc34 = loc26, loc29 // t2 = x2.2 + x5.2 |
| 441 |
psub2 loc41 = loc37, loc38 // buf1 = t5 - t6 |
| 442 |
padd2 loc40 = loc37, loc38 // buf0 = t5 + t6 |
| 443 |
padd2 loc35 = loc27, loc28 // t3 = x3.2 + x4.2 |
| 444 |
;; |
| 445 |
psub2 loc36 = loc24, loc31 // t4 = x0.2 - x7.2 |
| 446 |
pmpyshr2 loc37 = loc40, r28, 16 // t5 = buf0 * g4 |
| 447 |
pmpyshr2 loc38 = loc41, r28, 16 // t6 = buf1 * g4 |
| 448 |
;; |
| 449 |
psub2 loc39 = loc27, loc28 // t7 = x3.2 - x4.2 |
| 450 |
padd2 loc37 = loc37, loc40 // t5 = t5 + buf1 |
| 451 |
padd2 loc38 = loc38, loc41 // t6 = t6 + buf2 |
| 452 |
|
| 453 |
padd2 loc16 = loc32, loc35 // x0 = t0 + t3 |
| 454 |
padd2 loc17 = loc33, loc34 // x1 = t1 + t2 |
| 455 |
psub2 loc18 = loc32, loc35 // x2 = t0 - t3 |
| 456 |
;; |
| 457 |
psub2 loc19 = loc33, loc34 // x3 = t1 - t2 |
| 458 |
padd2 loc20 = loc36, loc37 // x4 = t4 + t5 |
| 459 |
padd2 loc21 = loc38, loc39 // x5 = t6 + t7 |
| 460 |
psub2 loc22 = loc36, loc37 // x6 = t4 - t5 |
| 461 |
psub2 loc23 = loc38, loc39 // x7 = t6 - t7 |
| 462 |
;; |
| 463 |
pmpyshr2 loc42 = loc18, r30, 16 // buf2 = x2 * c1 |
| 464 |
padd2 loc32 = loc16, loc17 // t0 = x0 + x1 |
| 465 |
pmpyshr2 loc43 = loc19, r30, 16 // buf3 = x3 * c1 |
| 466 |
pmpyshr2 loc44 = loc20, r31, 16 // buf4 = x4 * c0 |
| 467 |
psub2 loc33 = loc16, loc17 // t1 = x0 - x1 |
| 468 |
pmpyshr2 loc45 = loc21, r31, 16 // buf5 = x5 * c0 |
| 469 |
pmpyshr2 loc46 = loc22, r29, 16 // buf6 = x6 * c2 |
| 470 |
pmpyshr2 loc47 = loc23, r29, 16 // buf7 = x7 * c2 |
| 471 |
;; |
| 472 |
padd2 loc46 = loc46, loc22 // buf6 = buf6 + x6 |
| 473 |
padd2 loc47 = loc47, loc23 // buf7 = buf7 + x7 |
| 474 |
padd2 loc34 = loc18, loc43 // t2 = x2 + (x3 * c1) |
| 475 |
;; |
| 476 |
psub2 loc35 = loc42, loc19 // t3 = (c1 * x2) - x3 |
| 477 |
padd2 loc36 = loc20, loc45 // t4 = x4 + (x5 * c1) |
| 478 |
psub2 loc37 = loc44, loc21 // t5 = (c1 * x4) - x5 |
| 479 |
padd2 loc38 = loc22, loc47 // t6 = x6 + (buf7 * c1) |
| 480 |
psub2 loc39 = loc46, loc23 // t7 = (c1 * buf6) - x7 |
| 481 |
pmpyshr2 loc16 = loc32, r28, 16 // x0 = t0 * g4 |
| 482 |
pmpyshr2 loc17 = loc33, r28, 16 // x1 = t1 * g4 |
| 483 |
;; |
| 484 |
padd2 loc40 = loc16, loc32 // y0.2 = x0 + t0 |
| 485 |
pmpyshr2 loc18 = loc34, r26, 16 // x2 = t2 * g2 |
| 486 |
pmpyshr2 loc19 = loc35, r26, 16 // x3 = t3 * g2 |
| 487 |
padd2 loc44 = loc17, loc33 // y4.2 = x1 + t1 |
| 488 |
pmpyshr2 loc20 = loc36, r25, 16 // x4 = t4 * g1 |
| 489 |
pmpyshr2 loc21 = loc37, r25, 16 // x5 = t5 * g1 |
| 490 |
;; |
| 491 |
padd2 loc42 = loc18, loc34 // y2.2 = x2 + t2 |
| 492 |
pmpyshr2 loc22 = loc38, r27, 16 // x6 = t6 * g3 |
| 493 |
pmpyshr2 loc23 = loc39, r27, 16 // x7 = t7 * g3 |
| 494 |
padd2 loc46 = loc19, loc35 // y6.2 = x3 + t3 |
| 495 |
nop.i 0x0 |
| 496 |
nop.i 0x0 |
| 497 |
;; |
| 498 |
|
| 499 |
// ******************* |
| 500 |
// Transpose matrix |
| 501 |
// ******************* |
| 502 |
padd2 loc41 = loc20, loc36 // y1.2 = x4 + t4 |
| 503 |
mix2.l loc32 = loc49, loc48 // tmp0 = mixr y1, y0 |
| 504 |
mix2.r loc33 = loc49, loc48 // tmp1 = mixl y1, y0 |
| 505 |
padd2 loc47 = loc21, loc37 // y7.2 = x5 + t5 |
| 506 |
mix2.l loc34 = loc51, loc50 // tmp2 = mixr y3, y2 |
| 507 |
mix2.r loc35 = loc51, loc50 // tmp3 = mixl y3, y2 |
| 508 |
;; |
| 509 |
padd2 loc43 = loc22, loc38 // y3.2 = x6 + t6 |
| 510 |
mix4.l loc16 = loc34, loc32 // x0 = mixr tmp2, tmp0 |
| 511 |
mix4.l loc17 = loc35, loc33 // x1 = mixr tmp3, tmp1 |
| 512 |
padd2 loc45 = loc23, loc39 // y5.2 = x7 + t7 |
| 513 |
mix4.r loc18 = loc34, loc32 // x2 = mixl tmp2, tmp0 |
| 514 |
mix4.r loc19 = loc35, loc33 // x3 = mixl tmp3, tmp1 |
| 515 |
;; |
| 516 |
padd2 loc16 = loc16, r23 |
| 517 |
mix2.l loc32 = loc41, loc40 // tmp0 = mixr y0.2, y1.2 |
| 518 |
mix2.r loc33 = loc41, loc40 // tmp1 = mixl y0.2, y1.2 |
| 519 |
padd2 loc17 = loc17, r23 |
| 520 |
mix2.l loc34 = loc43, loc42 // tmp2 = mixr y2.2, y3.2 |
| 521 |
mix2.r loc35 = loc43, loc42 // tmp3 = mixl y2.2, y3.2 |
| 522 |
;; |
| 523 |
padd2 loc18 = loc18, r23 |
| 524 |
mix4.l loc20 = loc34, loc32 // x4 = mixr tmp2, tmp0 |
| 525 |
mix4.l loc21 = loc35, loc33 // x5 = mixr tmp3, tmp1 |
| 526 |
padd2 loc19 = loc19, r23 |
| 527 |
mix4.r loc22 = loc34, loc32 // x6 = mixl tmp2, tmp0 |
| 528 |
mix4.r loc23 = loc35, loc33 // x7 = mixl tmp3, tmp1 |
| 529 |
;; |
| 530 |
padd2 loc20 = loc20, r23 |
| 531 |
mix2.l loc32 = loc53, loc52 // tmp0 = mixr y5, y4 |
| 532 |
mix2.r loc33 = loc53, loc52 // tmp1 = mixl y5, y4 |
| 533 |
padd2 loc21 = loc21, r23 |
| 534 |
mix2.l loc34 = loc55, loc54 // tmp2 = mixr y7, y6 |
| 535 |
mix2.r loc35 = loc55, loc54 // tmp3 = mixl y7, y6 |
| 536 |
;; |
| 537 |
padd2 loc22 = loc22, r23 |
| 538 |
mix4.l loc24 = loc34, loc32 // x0.2 = mixr tmp2, tmp0 |
| 539 |
mix4.l loc25 = loc35, loc33 // x1.2 = mixr tmp3, tmp1 |
| 540 |
padd2 loc23 = loc23, r23 |
| 541 |
mix4.r loc26 = loc34, loc32 // x2.2 = mixl tmp2, tmp0 |
| 542 |
mix4.r loc27 = loc35, loc33 // x3.2 = mixl tmp3, tmp1 |
| 543 |
;; |
| 544 |
padd2 loc24 = loc24, r23 |
| 545 |
mix2.l loc32 = loc45, loc44 // tmp0 = mixr y4.2, y5.2 |
| 546 |
mix2.r loc33 = loc45, loc44 // tmp1 = mixl y4.2, y5.2 |
| 547 |
padd2 loc25 = loc25, r23 |
| 548 |
mix2.l loc34 = loc47, loc46 // tmp2 = mixr y6.2, y6.2 |
| 549 |
mix2.r loc35 = loc47, loc46 // tmp3 = mixl y6.2, y6.2 |
| 550 |
;; |
| 551 |
padd2 loc26 = loc26, r23 |
| 552 |
mix4.l loc28 = loc34, loc32 // x4.2 = mixr tmp2, tmp0 |
| 553 |
mix4.l loc29 = loc35, loc33 // x5.2 = mixr tmp3, tmp1 |
| 554 |
padd2 loc27 = loc27, r23 |
| 555 |
mix4.r loc30 = loc34, loc32 // x6.2 = mixl tmp2, tmp0 |
| 556 |
mix4.r loc31 = loc35, loc33 // x7.2 = mixl tmp3, tmp1 |
| 557 |
;; |
| 558 |
// ******************* |
| 559 |
// Descale |
| 560 |
// ******************* |
| 561 |
padd2 loc28 = loc28, r23 |
| 562 |
pshr2 loc16 = loc16, 3 |
| 563 |
pshr2 loc17 = loc17, 3 |
| 564 |
padd2 loc29 = loc29, r23 |
| 565 |
pshr2 loc18 = loc18, 3 |
| 566 |
pshr2 loc19 = loc19, 3 |
| 567 |
padd2 loc30 = loc30, r23 |
| 568 |
pshr2 loc20 = loc20, 3 |
| 569 |
pshr2 loc21 = loc21, 3 |
| 570 |
padd2 loc31 = loc31, r23 |
| 571 |
pshr2 loc22 = loc22, 3 |
| 572 |
pshr2 loc23 = loc23, 3 |
| 573 |
;; |
| 574 |
pshr2 loc24 = loc24, 3 |
| 575 |
pshr2 loc25 = loc25, 3 |
| 576 |
pshr2 loc26 = loc26, 3 |
| 577 |
pshr2 loc27 = loc27, 3 |
| 578 |
pshr2 loc28 = loc28, 3 |
| 579 |
pshr2 loc29 = loc29, 3 |
| 580 |
pshr2 loc30 = loc30, 3 |
| 581 |
pshr2 loc31 = loc31, 3 |
| 582 |
;; |
| 583 |
// ******************* |
| 584 |
// Store matrix |
| 585 |
// ******************* |
| 586 |
st8 [loc0] = loc16 |
| 587 |
st8 [loc1] = loc24 |
| 588 |
st8 [loc2] = loc17 |
| 589 |
st8 [loc3] = loc25 |
| 590 |
st8 [loc4] = loc18 |
| 591 |
st8 [loc5] = loc26 |
| 592 |
st8 [loc6] = loc19 |
| 593 |
st8 [loc7] = loc27 |
| 594 |
st8 [loc8] = loc20 |
| 595 |
st8 [loc9] = loc28 |
| 596 |
st8 [loc10] = loc21 |
| 597 |
st8 [loc11] = loc29 |
| 598 |
st8 [loc12] = loc22 |
| 599 |
st8 [loc13] = loc30 |
| 600 |
st8 [loc14] = loc23 |
| 601 |
st8 [loc15] = loc31 |
| 602 |
|
| 603 |
mov ar.pfs = r14 |
| 604 |
br.ret.sptk.many b0 |
| 605 |
.endp fdct_ia64# |
| 606 |
.common fdct#,8,8 |
| 607 |
|
| 608 |
|
| 609 |
|
| 610 |
|
| 611 |
|
| 612 |
|
| 613 |
|
| 614 |
|
| 615 |
//*********************************************** |
| 616 |
//* Here is a version of the DCT implementation * |
| 617 |
//* unoptimized in terms of command ordering. * |
| 618 |
//* This version is about 30% slower but * |
| 619 |
//* easier understand. * |
| 620 |
//*********************************************** |
| 621 |
// |
| 622 |
// .pred.safe_across_calls p1-p5,p16-p63 |
| 623 |
//.text |
| 624 |
// .align 16 |
| 625 |
// .global fdct_ia64# |
| 626 |
// .proc fdct_ia64# |
| 627 |
//fdct_ia64: |
| 628 |
// .prologue |
| 629 |
// alloc r14 = ar.pfs, 1, 56, 0, 0 |
| 630 |
// |
| 631 |
// // ******************* |
| 632 |
// // Save constants |
| 633 |
// // ******************* |
| 634 |
// mov r31 = 0x32ec // c0 = tan(1pi/16) |
| 635 |
// mov r30 = 0x6a0a // c1 = tan(2pi/16) |
| 636 |
// mov r29 = 0xab0e // c2 = tan(3pi/16) |
| 637 |
// mov r28 = 0xb505 // g4 = cos(4pi/16) |
| 638 |
// mov r27 = 0xd4db // g3 = cos(3pi/16) |
| 639 |
// mov r26 = 0xec83 // g2 = cos(2pi/16) |
| 640 |
// mov r25 = 0xfb15 // g1 = cos(1pi/16) |
| 641 |
// mov r24 = 0x0002 // correction bit for descaling |
| 642 |
// mov r23 = 0x0004 // correction bit for descaling |
| 643 |
// |
| 644 |
// // ************************** |
| 645 |
// // Load Matrix into registers |
| 646 |
// // ************************** |
| 647 |
// |
| 648 |
// add loc0 = r0, r32 |
| 649 |
// ;; |
| 650 |
// mux2 r31 = r31, 0x00 |
| 651 |
// mux2 r30 = r30, 0x00 |
| 652 |
// mux2 r29 = r29, 0x00 |
| 653 |
// mux2 r28 = r28, 0x00 |
| 654 |
// mux2 r27 = r27, 0x00 |
| 655 |
// mux2 r26 = r26, 0x00 |
| 656 |
// mux2 r25 = r25, 0x00 |
| 657 |
// mux2 r24 = r24, 0x00 |
| 658 |
// mux2 r23 = r23, 0x00 |
| 659 |
// ld8 loc16 = [loc0] |
| 660 |
// add loc2 = 16, r32 |
| 661 |
// add loc4 = 32, r32 |
| 662 |
// add loc6 = 48, r32 |
| 663 |
// add loc8 = 64, r32 |
| 664 |
// add loc10 = 80, r32 |
| 665 |
// ;; |
| 666 |
// ld8 loc17 = [loc2] |
| 667 |
// ld8 loc18 = [loc4] |
| 668 |
// add loc12 = 96, r32 |
| 669 |
// ld8 loc19 = [loc6] |
| 670 |
// ld8 loc20 = [loc8] |
| 671 |
// add loc14 = 112, r32 |
| 672 |
// ;; |
| 673 |
// ld8 loc21 = [loc10] |
| 674 |
// ld8 loc22 = [loc12] |
| 675 |
// add loc1 = 8, r32 |
| 676 |
// ld8 loc23 = [loc14] |
| 677 |
// add loc3 = 24, r32 |
| 678 |
// add loc5 = 40, r32 |
| 679 |
// ;; |
| 680 |
// ld8 loc24 = [loc1] |
| 681 |
// ld8 loc25 = [loc3] |
| 682 |
// add loc7 = 56, r32 |
| 683 |
// ld8 loc26 = [loc5] |
| 684 |
// add loc9 = 72, r32 |
| 685 |
// add loc11 = 88, r32 |
| 686 |
// ;; |
| 687 |
// ld8 loc27 = [loc7] |
| 688 |
// ld8 loc28 = [loc9] |
| 689 |
// add loc13 = 104, r32 |
| 690 |
// ld8 loc29 = [loc11] |
| 691 |
// add loc15 = 120, r32 |
| 692 |
// ;; |
| 693 |
// ld8 loc30 = [loc13] |
| 694 |
// ld8 loc31 = [loc15] |
| 695 |
// ;; |
| 696 |
// // ****** |
| 697 |
// // Scale |
| 698 |
// // ****** |
| 699 |
// pshl2 loc16 = loc16, 3 |
| 700 |
// pshl2 loc17 = loc17, 3 |
| 701 |
// pshl2 loc18 = loc18, 3 |
| 702 |
// pshl2 loc19 = loc19, 3 |
| 703 |
// pshl2 loc20 = loc20, 3 |
| 704 |
// pshl2 loc21 = loc21, 3 |
| 705 |
// pshl2 loc22 = loc22, 3 |
| 706 |
// pshl2 loc23 = loc23, 3 |
| 707 |
// pshl2 loc24 = loc24, 3 |
| 708 |
// pshl2 loc25 = loc25, 3 |
| 709 |
// pshl2 loc26 = loc26, 3 |
| 710 |
// pshl2 loc27 = loc27, 3 |
| 711 |
// pshl2 loc28 = loc28, 3 |
| 712 |
// pshl2 loc29 = loc29, 3 |
| 713 |
// pshl2 loc30 = loc30, 3 |
| 714 |
// pshl2 loc31 = loc31, 3 |
| 715 |
// ;; |
| 716 |
// |
| 717 |
// // ******************* |
| 718 |
// // column-DTC 1st half |
| 719 |
// // ******************* |
| 720 |
// |
| 721 |
// padd2 loc32 = loc16, loc23 // t0 = x0 + x7 |
| 722 |
// padd2 loc33 = loc17, loc22 // t1 = x1 + x6 |
| 723 |
// padd2 loc34 = loc18, loc21 // t2 = x2 + x5 |
| 724 |
// padd2 loc35 = loc19, loc20 // t3 = x3 + x4 |
| 725 |
// psub2 loc36 = loc16, loc23 // t4 = x0 - x7 |
| 726 |
// psub2 loc37 = loc17, loc22 // t5 = x1 - x6 |
| 727 |
// psub2 loc38 = loc18, loc21 // t6 = x2 - x5 |
| 728 |
// psub2 loc39 = loc19, loc20 // t7 = x3 - x4 |
| 729 |
// ;; |
| 730 |
// padd2 loc40 = loc37, loc38 // buf0 = t5 + t6 |
| 731 |
// psub2 loc41 = loc37, loc38 // buf1 = t5 - t6 |
| 732 |
// ;; |
| 733 |
// pmpyshr2 loc37 = loc40, r28, 16 // t5 = buf0 * g4 |
| 734 |
// pmpyshr2 loc38 = loc41, r28, 16 // t6 = buf1 * g4 |
| 735 |
// ;; |
| 736 |
// padd2 loc37 = loc37, loc40 // t5 = t5 + buf1 |
| 737 |
// padd2 loc38 = loc38, loc41 // t6 = t6 + buf2 |
| 738 |
// ;; |
| 739 |
// padd2 loc16 = loc32, loc35 // x0 = t0 + t3 |
| 740 |
// padd2 loc17 = loc33, loc34 // x1 = t1 + t2 |
| 741 |
// psub2 loc18 = loc32, loc35 // x2 = t0 - t3 |
| 742 |
// psub2 loc19 = loc33, loc34 // x3 = t1 - t2 |
| 743 |
// padd2 loc20 = loc36, loc37 // x4 = t4 + t5 |
| 744 |
// padd2 loc21 = loc38, loc39 // x5 = t6 + t7 |
| 745 |
// psub2 loc22 = loc36, loc37 // x6 = t4 - t5 |
| 746 |
// psub2 loc23 = loc38, loc39 // x7 = t6 - t7 |
| 747 |
// ;; |
| 748 |
// |
| 749 |
// padd2 loc32 = loc16, loc17 // t0 = x0 + x1 |
| 750 |
// psub2 loc33 = loc16, loc17 // t1 = x0 - x1 |
| 751 |
// pmpyshr2 loc42 = loc18, r30, 16 // buf2 = x2 * c1 |
| 752 |
// pmpyshr2 loc43 = loc19, r30, 16 // buf3 = x3 * c1 |
| 753 |
// pmpyshr2 loc44 = loc20, r31, 16 // buf4 = x4 * c0 |
| 754 |
// pmpyshr2 loc45 = loc21, r31, 16 // buf5 = x5 * c0 |
| 755 |
// pmpyshr2 loc46 = loc22, r29, 16 // buf6 = x6 * c2 |
| 756 |
// pmpyshr2 loc47 = loc23, r29, 16 // buf7 = x7 * c2 |
| 757 |
// ;; |
| 758 |
// padd2 loc46 = loc46, loc22 // buf6 = buf6 + x6 |
| 759 |
// padd2 loc47 = loc47, loc23 // buf7 = buf7 + x7 |
| 760 |
// ;; |
| 761 |
// padd2 loc34 = loc18, loc43 // t2 = x2 + (x3 * c1) |
| 762 |
// psub2 loc35 = loc42, loc19 // t3 = (c1 * x2) - x3 |
| 763 |
// padd2 loc36 = loc20, loc45 // t4 = x4 + (x5 * c1) |
| 764 |
// psub2 loc37 = loc44, loc21 // t5 = (c1 * x4) - x5 |
| 765 |
// padd2 loc38 = loc22, loc47 // t6 = x6 + (x7 * c1) |
| 766 |
// psub2 loc39 = loc46, loc23 // t7 = (c1 * x6) - x7 |
| 767 |
// ;; |
| 768 |
// pmpyshr2 loc16 = loc32, r28, 16 // x0 = t0 * g4 |
| 769 |
// pmpyshr2 loc17 = loc33, r28, 16 // x1 = t1 * g4 |
| 770 |
// pmpyshr2 loc18 = loc34, r26, 16 // x2 = t2 * g2 |
| 771 |
// pmpyshr2 loc19 = loc35, r26, 16 // x3 = t3 * g2 |
| 772 |
// pmpyshr2 loc20 = loc36, r25, 16 // x4 = t4 * g1 |
| 773 |
// pmpyshr2 loc21 = loc37, r25, 16 // x5 = t5 * g1 |
| 774 |
// pmpyshr2 loc22 = loc38, r27, 16 // x6 = t6 * g3 |
| 775 |
// pmpyshr2 loc23 = loc39, r27, 16 // x7 = t7 * g3 |
| 776 |
// ;; |
| 777 |
// padd2 loc48 = loc16, loc32 // y0 = x0 + t0 |
| 778 |
// padd2 loc49 = loc20, loc36 // y1 = x4 + t4 |
| 779 |
// padd2 loc50 = loc18, loc34 // y2 = x2 + t2 |
| 780 |
// padd2 loc51 = loc22, loc38 // y3 = x6 + t6 |
| 781 |
// padd2 loc52 = loc17, loc33 // y4 = x1 + t1 |
| 782 |
// padd2 loc53 = loc23, loc39 // y5 = x7 + t7 |
| 783 |
// padd2 loc54 = loc19, loc35 // y6 = x3 + t3 |
| 784 |
// padd2 loc55 = loc21, loc37 // y7 = x5 + t5 |
| 785 |
// ;; |
| 786 |
// |
| 787 |
// // ******************* |
| 788 |
// // column-DTC 2nd half |
| 789 |
// // ******************* |
| 790 |
// |
| 791 |
// padd2 loc32 = loc24, loc31 // t0 = x0.2 + x7.2 |
| 792 |
// padd2 loc33 = loc25, loc30 // t1 = x1.2 + x6.2 |
| 793 |
// padd2 loc34 = loc26, loc29 // t2 = x2.2 + x5.2 |
| 794 |
// padd2 loc35 = loc27, loc28 // t3 = x3.2 + x4.2 |
| 795 |
// psub2 loc36 = loc24, loc31 // t4 = x0.2 - x7.2 |
| 796 |
// psub2 loc37 = loc25, loc30 // t5 = x1.2 - x6.2 |
| 797 |
// psub2 loc38 = loc26, loc29 // t6 = x2.2 - x5.2 |
| 798 |
// psub2 loc39 = loc27, loc28 // t7 = x3.2 - x4.2 |
| 799 |
// ;; |
| 800 |
// padd2 loc40 = loc37, loc38 // buf0 = t5 + t6 |
| 801 |
// psub2 loc41 = loc37, loc38 // buf1 = t5 - t6 |
| 802 |
// ;; |
| 803 |
// pmpyshr2 loc37 = loc40, r28, 16 // t5 = buf0 * g4 |
| 804 |
// pmpyshr2 loc38 = loc41, r28, 16 // t6 = buf1 * g4 |
| 805 |
// ;; |
| 806 |
// padd2 loc37 = loc37, loc40 // t5 = t5 + buf1 |
| 807 |
// padd2 loc38 = loc38, loc41 // t6 = t6 + buf2 |
| 808 |
// ;; |
| 809 |
// padd2 loc16 = loc32, loc35 // x0 = t0 + t3 |
| 810 |
// padd2 loc17 = loc33, loc34 // x1 = t1 + t2 |
| 811 |
// psub2 loc18 = loc32, loc35 // x2 = t0 - t3 |
| 812 |
// psub2 loc19 = loc33, loc34 // x3 = t1 - t2 |
| 813 |
// padd2 loc20 = loc36, loc37 // x4 = t4 + t5 |
| 814 |
// padd2 loc21 = loc38, loc39 // x5 = t6 + t7 |
| 815 |
// psub2 loc22 = loc36, loc37 // x6 = t4 - t5 |
| 816 |
// psub2 loc23 = loc38, loc39 // x7 = t6 - t7 |
| 817 |
// ;; |
| 818 |
// padd2 loc32 = loc16, loc17 // t0 = x0 + x1 |
| 819 |
// psub2 loc33 = loc16, loc17 // t1 = x0 - x1 |
| 820 |
// pmpyshr2 loc42 = loc18, r30, 16 // buf2 = x2 * c1 |
| 821 |
// pmpyshr2 loc43 = loc19, r30, 16 // buf3 = x3 * c1 |
| 822 |
// pmpyshr2 loc44 = loc20, r31, 16 // buf4 = x4 * c0 |
| 823 |
// pmpyshr2 loc45 = loc21, r31, 16 // buf5 = x5 * c0 |
| 824 |
// pmpyshr2 loc46 = loc22, r29, 16 // buf6 = x6 * c2 |
| 825 |
// pmpyshr2 loc47 = loc23, r29, 16 // buf7 = x7 * c2 |
| 826 |
// ;; |
| 827 |
// padd2 loc46 = loc46, loc22 // buf6 = buf6 + x6 |
| 828 |
// padd2 loc47 = loc47, loc23 // buf7 = buf7 + x7 |
| 829 |
// ;; |
| 830 |
// padd2 loc34 = loc18, loc43 // t2 = x2 + (x3 * c1) |
| 831 |
// psub2 loc35 = loc42, loc19 // t3 = (c1 * x2) - x3 |
| 832 |
// padd2 loc36 = loc20, loc45 // t4 = x4 + (x5 * c1) |
| 833 |
// psub2 loc37 = loc44, loc21 // t5 = (c1 * x4) - x5 |
| 834 |
// padd2 loc38 = loc22, loc47 // t6 = x6 + (x7 * c1) |
| 835 |
// psub2 loc39 = loc46, loc23 // t7 = (c1 * x6) - x7 |
| 836 |
// ;; |
| 837 |
// pmpyshr2 loc16 = loc32, r28, 16 // x0 = t0 * g4 |
| 838 |
// pmpyshr2 loc17 = loc33, r28, 16 // x1 = t1 * g4 |
| 839 |
// pmpyshr2 loc18 = loc34, r26, 16 // x2 = t2 * g2 |
| 840 |
// pmpyshr2 loc19 = loc35, r26, 16 // x3 = t3 * g2 |
| 841 |
// pmpyshr2 loc20 = loc36, r25, 16 // x4 = t4 * g1 |
| 842 |
// pmpyshr2 loc21 = loc37, r25, 16 // x5 = t5 * g1 |
| 843 |
// pmpyshr2 loc22 = loc38, r27, 16 // x6 = t6 * g3 |
| 844 |
// pmpyshr2 loc23 = loc39, r27, 16 // x7 = t7 * g3 |
| 845 |
// ;; |
| 846 |
// padd2 loc40 = loc16, loc32 // y0.2 = x0 + t0 |
| 847 |
// padd2 loc41 = loc20, loc36 // y1.2 = x4 + t4 |
| 848 |
// padd2 loc42 = loc18, loc34 // y2.2 = x2 + t2 |
| 849 |
// padd2 loc43 = loc22, loc38 // y3.2 = x6 + t6 |
| 850 |
// padd2 loc44 = loc17, loc33 // y4.2 = x1 + t1 |
| 851 |
// padd2 loc45 = loc23, loc39 // y5.2 = x7 + t7 |
| 852 |
// padd2 loc46 = loc19, loc35 // y6.2 = x3 + t3 |
| 853 |
// padd2 loc47 = loc21, loc37 // y7.2 = x5 + t5 |
| 854 |
// ;; |
| 855 |
// padd2 loc40 = loc40, r24 // add r24 to correct rounding |
| 856 |
// padd2 loc41 = loc41, r24 |
| 857 |
// padd2 loc42 = loc42, r24 |
| 858 |
// padd2 loc43 = loc43, r24 |
| 859 |
// padd2 loc44 = loc44, r24 |
| 860 |
// padd2 loc45 = loc45, r24 |
| 861 |
// padd2 loc46 = loc46, r24 |
| 862 |
// padd2 loc47 = loc47, r24 |
| 863 |
// padd2 loc48 = loc48, r24 |
| 864 |
// padd2 loc49 = loc49, r24 |
| 865 |
// padd2 loc50 = loc50, r24 |
| 866 |
// padd2 loc51 = loc51, r24 |
| 867 |
// padd2 loc52 = loc52, r24 |
| 868 |
// padd2 loc53 = loc53, r24 |
| 869 |
// padd2 loc54 = loc54, r24 |
| 870 |
// padd2 loc55 = loc55, r24 |
| 871 |
// ;; |
| 872 |
// pshr2 loc40 = loc40, 2 // Divide all matrix elements through 4 |
| 873 |
// pshr2 loc41 = loc41, 2 |
| 874 |
// pshr2 loc42 = loc42, 2 |
| 875 |
// pshr2 loc43 = loc43, 2 |
| 876 |
// pshr2 loc44 = loc44, 2 |
| 877 |
// pshr2 loc45 = loc45, 2 |
| 878 |
// pshr2 loc46 = loc46, 2 |
| 879 |
// pshr2 loc47 = loc47, 2 |
| 880 |
// pshr2 loc48 = loc48, 2 |
| 881 |
// pshr2 loc49 = loc49, 2 |
| 882 |
// pshr2 loc50 = loc50, 2 |
| 883 |
// pshr2 loc51 = loc51, 2 |
| 884 |
// pshr2 loc52 = loc52, 2 |
| 885 |
// pshr2 loc53 = loc53, 2 |
| 886 |
// pshr2 loc54 = loc54, 2 |
| 887 |
// pshr2 loc55 = loc55, 2 |
| 888 |
// ;; |
| 889 |
// |
| 890 |
// // ***************** |
| 891 |
// // Transpose matrix |
| 892 |
// // ***************** |
| 893 |
// |
| 894 |
// mix2.r loc32 = loc48, loc49 // tmp0 = mixr y0, y1 |
| 895 |
// mix2.l loc33 = loc48, loc49 // tmp1 = mixl y0, y1 |
| 896 |
// mix2.r loc34 = loc50, loc51 // tmp2 = mixr y2, y3 |
| 897 |
// mix2.l loc35 = loc50, loc51 // tmp3 = mixl y2, y3 |
| 898 |
// ;; |
| 899 |
// mix4.r loc16 = loc32, loc34 // x0 = mixr tmp0, tmp2 |
| 900 |
// mix4.r loc17 = loc33, loc35 // x1 = mixr tmp1, tmp3 |
| 901 |
// mix4.l loc18 = loc32, loc34 // x2 = mixl tmp0, tmp2 |
| 902 |
// mix4.l loc19 = loc33, loc35 // x3 = mixl tmp1, tmp3 |
| 903 |
// ;; |
| 904 |
// mix2.r loc32 = loc40, loc41 // tmp0 = mixr y0.2, y1.2 |
| 905 |
// mix2.l loc33 = loc40, loc41 // tmp1 = mixl y0.2, y1.2 |
| 906 |
// mix2.r loc34 = loc42, loc43 // tmp2 = mixr y2.2, y3.2 |
| 907 |
// mix2.l loc35 = loc42, loc43 // tmp3 = mixl y2.2, y3.2 |
| 908 |
// ;; |
| 909 |
// mix4.r loc20 = loc32, loc34 // x4 = mixr tmp0, tmp2 |
| 910 |
// mix4.r loc21 = loc33, loc35 // x5 = mixr tmp1, tmp3 |
| 911 |
// mix4.l loc22 = loc32, loc34 // x6 = mixl tmp0, tmp2 |
| 912 |
// mix4.l loc23 = loc33, loc35 // x7 = mixl tmp1, tmp3 |
| 913 |
// ;; |
| 914 |
// mix2.r loc32 = loc52, loc53 // tmp0 = mixr y4, y5 |
| 915 |
// mix2.l loc33 = loc52, loc53 // tmp1 = mixl y4, y5 |
| 916 |
// mix2.r loc34 = loc54, loc55 // tmp2 = mixr y6, y7 |
| 917 |
// mix2.l loc35 = loc54, loc55 // tmp3 = mixl y6, y7 |
| 918 |
// ;; |
| 919 |
// mix4.r loc24 = loc32, loc34 // x0.2 = mixr tmp0, tmp2 |
| 920 |
// mix4.r loc25 = loc33, loc35 // x1.2 = mixr tmp1, tmp3 |
| 921 |
// mix4.l loc26 = loc32, loc34 // x2.2 = mixl tmp0, tmp2 |
| 922 |
// mix4.l loc27 = loc33, loc35 // x3.2 = mixl tmp1, tmp3 |
| 923 |
// ;; |
| 924 |
// mix2.r loc32 = loc44, loc45 // tmp0 = mixr y4.2, y5.2 |
| 925 |
// mix2.l loc33 = loc44, loc45 // tmp1 = mixl y4.2, y5.2 |
| 926 |
// mix2.r loc34 = loc46, loc47 // tmp2 = mixr y6.2, y6.2 |
| 927 |
// mix2.l loc35 = loc46, loc47 // tmp3 = mixl y6.2, y6.2 |
| 928 |
// ;; |
| 929 |
// mix4.r loc28 = loc32, loc34 // x4.2 = mixr tmp0, tmp2 |
| 930 |
// mix4.r loc29 = loc33, loc35 // x5.2 = mixr tmp1, tmp3 |
| 931 |
// mix4.l loc30 = loc32, loc34 // x6.2 = mixl tmp0, tmp2 |
| 932 |
// mix4.l loc31 = loc33, loc35 // x7.2 = mixl tmp1, tmp3 |
| 933 |
// ;; |
| 934 |
// |
| 935 |
// // ******************* |
| 936 |
// // row-DTC 1st half |
| 937 |
// // ******************* |
| 938 |
// |
| 939 |
// padd2 loc32 = loc16, loc23 // t0 = x0 + x7 |
| 940 |
// padd2 loc33 = loc17, loc22 // t1 = x1 + x6 |
| 941 |
// padd2 loc34 = loc18, loc21 // t2 = x2 + x5 |
| 942 |
// padd2 loc35 = loc19, loc20 // t3 = x3 + x4 |
| 943 |
// psub2 loc36 = loc16, loc23 // t4 = x0 - x7 |
| 944 |
// psub2 loc37 = loc17, loc22 // t5 = x1 - x6 |
| 945 |
// psub2 loc38 = loc18, loc21 // t6 = x2 - x5 |
| 946 |
// psub2 loc39 = loc19, loc20 // t7 = x3 - x4 |
| 947 |
// ;; |
| 948 |
// padd2 loc40 = loc37, loc38 // buf0 = t5 + t6 |
| 949 |
// psub2 loc41 = loc37, loc38 // buf1 = t5 - t6 |
| 950 |
// ;; |
| 951 |
// pmpyshr2 loc37 = loc40, r28, 16 // t5 = buf0 * g4 |
| 952 |
// pmpyshr2 loc38 = loc41, r28, 16 // t6 = buf1 * g4 |
| 953 |
// ;; |
| 954 |
// padd2 loc37 = loc37, loc40 // t5 = t5 + buf1 |
| 955 |
// padd2 loc38 = loc38, loc41 // t6 = t6 + buf2 |
| 956 |
// ;; |
| 957 |
// padd2 loc16 = loc32, loc35 // x0 = t0 + t3 |
| 958 |
// padd2 loc17 = loc33, loc34 // x1 = t1 + t2 |
| 959 |
// psub2 loc18 = loc32, loc35 // x2 = t0 - t3 |
| 960 |
// psub2 loc19 = loc33, loc34 // x3 = t1 - t2 |
| 961 |
// padd2 loc20 = loc36, loc37 // x4 = t4 + t5 |
| 962 |
// padd2 loc21 = loc38, loc39 // x5 = t6 + t7 |
| 963 |
// psub2 loc22 = loc36, loc37 // x6 = t4 - t5 |
| 964 |
// psub2 loc23 = loc38, loc39 // x7 = t6 - t7 |
| 965 |
// ;; |
| 966 |
// padd2 loc32 = loc16, loc17 // t0 = x0 + x1 |
| 967 |
// psub2 loc33 = loc16, loc17 // t1 = x0 - x1 |
| 968 |
// pmpyshr2 loc42 = loc18, r30, 16 // buf2 = x2 * c1 |
| 969 |
// pmpyshr2 loc43 = loc19, r30, 16 // buf3 = x3 * c1 |
| 970 |
// pmpyshr2 loc44 = loc20, r31, 16 // buf4 = x4 * c0 |
| 971 |
// pmpyshr2 loc45 = loc21, r31, 16 // buf5 = x5 * c0 |
| 972 |
// pmpyshr2 loc46 = loc22, r29, 16 // buf6 = x6 * c2 |
| 973 |
// pmpyshr2 loc47 = loc23, r29, 16 // buf7 = x7 * c2 |
| 974 |
// ;; |
| 975 |
// padd2 loc46 = loc46, loc22 // buf6 = buf6 + x6 |
| 976 |
// padd2 loc47 = loc47, loc23 // buf7 = buf7 + x7 |
| 977 |
// ;; |
| 978 |
// padd2 loc34 = loc18, loc43 // t2 = x2 + (x3 * c1) |
| 979 |
// psub2 loc35 = loc42, loc19 // t3 = (c1 * x2) - x3 |
| 980 |
// padd2 loc36 = loc20, loc45 // t4 = x4 + (x5 * c1) |
| 981 |
// psub2 loc37 = loc44, loc21 // t5 = (c1 * x4) - x5 |
| 982 |
// padd2 loc38 = loc22, loc47 // t6 = x6 + (x7 * c1) |
| 983 |
// psub2 loc39 = loc46, loc23 // t7 = (c1 * x6) - x7 |
| 984 |
// ;; |
| 985 |
// pmpyshr2 loc16 = loc32, r28, 16 // x0 = t0 * g4 |
| 986 |
// pmpyshr2 loc17 = loc33, r28, 16 // x1 = t1 * g4 |
| 987 |
// pmpyshr2 loc18 = loc34, r26, 16 // x2 = t2 * g2 |
| 988 |
// pmpyshr2 loc19 = loc35, r26, 16 // x3 = t3 * g2 |
| 989 |
// pmpyshr2 loc20 = loc36, r25, 16 // x4 = t4 * g1 |
| 990 |
// pmpyshr2 loc21 = loc37, r25, 16 // x5 = t5 * g1 |
| 991 |
// pmpyshr2 loc22 = loc38, r27, 16 // x6 = t6 * g3 |
| 992 |
// pmpyshr2 loc23 = loc39, r27, 16 // x7 = t7 * g3 |
| 993 |
// ;; |
| 994 |
// padd2 loc48 = loc16, loc32 // y0 = x0 + t0 |
| 995 |
// padd2 loc49 = loc20, loc36 // y1 = x4 + t4 |
| 996 |
// padd2 loc50 = loc18, loc34 // y2 = x2 + t2 |
| 997 |
// padd2 loc51 = loc22, loc38 // y3 = x6 + t6 |
| 998 |
// padd2 loc52 = loc17, loc33 // y4 = x1 + t1 |
| 999 |
// padd2 loc53 = loc23, loc39 // y5 = x7 + t7 |
| 1000 |
// padd2 loc54 = loc19, loc35 // y6 = x3 + t3 |
| 1001 |
// padd2 loc55 = loc21, loc37 // y7 = x5 + t5 |
| 1002 |
// ;; |
| 1003 |
// |
| 1004 |
// // ******************* |
| 1005 |
// // row-DTC 2nd half |
| 1006 |
// // ******************* |
| 1007 |
// |
| 1008 |
// padd2 loc32 = loc24, loc31 // t0 = x0.2 + x7.2 |
| 1009 |
// padd2 loc33 = loc25, loc30 // t1 = x1.2 + x6.2 |
| 1010 |
// padd2 loc34 = loc26, loc29 // t2 = x2.2 + x5.2 |
| 1011 |
// padd2 loc35 = loc27, loc28 // t3 = x3.2 + x4.2 |
| 1012 |
// psub2 loc36 = loc24, loc31 // t4 = x0.2 - x7.2 |
| 1013 |
// psub2 loc37 = loc25, loc30 // t5 = x1.2 - x6.2 |
| 1014 |
// psub2 loc38 = loc26, loc29 // t6 = x2.2 - x5.2 |
| 1015 |
// psub2 loc39 = loc27, loc28 // t7 = x3.2 - x4.2 |
| 1016 |
// ;; |
| 1017 |
// padd2 loc40 = loc37, loc38 // buf0 = t5 + t6 |
| 1018 |
// psub2 loc41 = loc37, loc38 // buf1 = t5 - t6 |
| 1019 |
// ;; |
| 1020 |
// pmpyshr2 loc37 = loc40, r28, 16 // t5 = buf0 * g4 |
| 1021 |
// pmpyshr2 loc38 = loc41, r28, 16 // t6 = buf1 * g4 |
| 1022 |
// ;; |
| 1023 |
// padd2 loc37 = loc37, loc40 // t5 = t5 + buf1 |
| 1024 |
// padd2 loc38 = loc38, loc41 // t6 = t6 + buf2 |
| 1025 |
// ;; |
| 1026 |
// padd2 loc16 = loc32, loc35 // x0 = t0 + t3 |
| 1027 |
// padd2 loc17 = loc33, loc34 // x1 = t1 + t2 |
| 1028 |
// psub2 loc18 = loc32, loc35 // x2 = t0 - t3 |
| 1029 |
// psub2 loc19 = loc33, loc34 // x3 = t1 - t2 |
| 1030 |
// padd2 loc20 = loc36, loc37 // x4 = t4 + t5 |
| 1031 |
// padd2 loc21 = loc38, loc39 // x5 = t6 + t7 |
| 1032 |
// psub2 loc22 = loc36, loc37 // x6 = t4 - t5 |
| 1033 |
// psub2 loc23 = loc38, loc39 // x7 = t6 - t7 |
| 1034 |
// ;; |
| 1035 |
// padd2 loc32 = loc16, loc17 // t0 = x0 + x1 |
| 1036 |
// psub2 loc33 = loc16, loc17 // t1 = x0 - x1 |
| 1037 |
// pmpyshr2 loc42 = loc18, r30, 16 // buf2 = x2 * c1 |
| 1038 |
// pmpyshr2 loc43 = loc19, r30, 16 // buf3 = x3 * c1 |
| 1039 |
// pmpyshr2 loc44 = loc20, r31, 16 // buf4 = x4 * c0 |
| 1040 |
// pmpyshr2 loc45 = loc21, r31, 16 // buf5 = x5 * c0 |
| 1041 |
// pmpyshr2 loc46 = loc22, r29, 16 // buf6 = x6 * c2 |
| 1042 |
// pmpyshr2 loc47 = loc23, r29, 16 // buf7 = x7 * c2 |
| 1043 |
// ;; |
| 1044 |
// padd2 loc46 = loc46, loc22 // buf6 = buf6 + x6 |
| 1045 |
// padd2 loc47 = loc47, loc23 // buf7 = buf7 + x7 |
| 1046 |
// ;; |
| 1047 |
// padd2 loc34 = loc18, loc43 // t2 = x2 + (x3 * c1) |
| 1048 |
// psub2 loc35 = loc42, loc19 // t3 = (c1 * x2) - x3 |
| 1049 |
// padd2 loc36 = loc20, loc45 // t4 = x4 + (x5 * c1) |
| 1050 |
// psub2 loc37 = loc44, loc21 // t5 = (c1 * x4) - x5 |
| 1051 |
// padd2 loc38 = loc22, loc47 // t6 = x6 + (x7 * c1) |
| 1052 |
// psub2 loc39 = loc46, loc23 // t7 = (c1 * x6) - x7 |
| 1053 |
// ;; |
| 1054 |
// pmpyshr2 loc16 = loc32, r28, 16 // x0 = t0 * g4 |
| 1055 |
// pmpyshr2 loc17 = loc33, r28, 16 // x1 = t1 * g4 |
| 1056 |
// pmpyshr2 loc18 = loc34, r26, 16 // x2 = t2 * g2 |
| 1057 |
// pmpyshr2 loc19 = loc35, r26, 16 // x3 = t3 * g2 |
| 1058 |
// pmpyshr2 loc20 = loc36, r25, 16 // x4 = t4 * g1 |
| 1059 |
// pmpyshr2 loc21 = loc37, r25, 16 // x5 = t5 * g1 |
| 1060 |
// pmpyshr2 loc22 = loc38, r27, 16 // x6 = t6 * g3 |
| 1061 |
// pmpyshr2 loc23 = loc39, r27, 16 // x7 = t7 * g3 |
| 1062 |
// ;; |
| 1063 |
// padd2 loc40 = loc16, loc32 // y0.2 = x0 + t0 |
| 1064 |
// padd2 loc41 = loc20, loc36 // y1.2 = x4 + t4 |
| 1065 |
// padd2 loc42 = loc18, loc34 // y2.2 = x2 + t2 |
| 1066 |
// padd2 loc43 = loc22, loc38 // y3.2 = x6 + t6 |
| 1067 |
// padd2 loc44 = loc17, loc33 // y4.2 = x1 + t1 |
| 1068 |
// padd2 loc45 = loc23, loc39 // y5.2 = x7 + t7 |
| 1069 |
// padd2 loc46 = loc19, loc35 // y6.2 = x3 + t3 |
| 1070 |
// padd2 loc47 = loc21, loc37 // y7.2 = x5 + t5 |
| 1071 |
// ;; |
| 1072 |
// // ******************* |
| 1073 |
// // Transpose matrix |
| 1074 |
// // ******************* |
| 1075 |
// |
| 1076 |
// mix2.l loc32 = loc49, loc48 // tmp0 = mixr y1, y0 |
| 1077 |
// mix2.r loc33 = loc49, loc48 // tmp1 = mixl y1, y0 |
| 1078 |
// mix2.l loc34 = loc51, loc50 // tmp2 = mixr y3, y2 |
| 1079 |
// mix2.r loc35 = loc51, loc50 // tmp3 = mixl y3, y2 |
| 1080 |
// ;; |
| 1081 |
// mix4.l loc16 = loc34, loc32 // x0 = mixr tmp2, tmp0 |
| 1082 |
// mix4.l loc17 = loc35, loc33 // x1 = mixr tmp3, tmp1 |
| 1083 |
// mix4.r loc18 = loc34, loc32 // x2 = mixl tmp2, tmp0 |
| 1084 |
// mix4.r loc19 = loc35, loc33 // x3 = mixl tmp3, tmp1 |
| 1085 |
// ;; |
| 1086 |
// mix2.l loc32 = loc41, loc40 // tmp0 = mixr y0.2, y1.2 |
| 1087 |
// mix2.r loc33 = loc41, loc40 // tmp1 = mixl y0.2, y1.2 |
| 1088 |
// mix2.l loc34 = loc43, loc42 // tmp2 = mixr y2.2, y3.2 |
| 1089 |
// mix2.r loc35 = loc43, loc42 // tmp3 = mixl y2.2, y3.2 |
| 1090 |
// ;; |
| 1091 |
// mix4.l loc20 = loc34, loc32 // x4 = mixr tmp2, tmp0 |
| 1092 |
// mix4.l loc21 = loc35, loc33 // x5 = mixr tmp3, tmp1 |
| 1093 |
// mix4.r loc22 = loc34, loc32 // x6 = mixl tmp2, tmp0 |
| 1094 |
// mix4.r loc23 = loc35, loc33 // x7 = mixl tmp3, tmp1 |
| 1095 |
// ;; |
| 1096 |
// mix2.l loc32 = loc53, loc52 // tmp0 = mixr y5, y4 |
| 1097 |
// mix2.r loc33 = loc53, loc52 // tmp1 = mixl y5, y4 |
| 1098 |
// mix2.l loc34 = loc55, loc54 // tmp2 = mixr y7, y6 |
| 1099 |
// mix2.r loc35 = loc55, loc54 // tmp3 = mixl y7, y6 |
| 1100 |
// ;; |
| 1101 |
// mix4.l loc24 = loc34, loc32 // x0.2 = mixr tmp2, tmp0 |
| 1102 |
// mix4.l loc25 = loc35, loc33 // x1.2 = mixr tmp3, tmp1 |
| 1103 |
// mix4.r loc26 = loc34, loc32 // x2.2 = mixl tmp2, tmp0 |
| 1104 |
// mix4.r loc27 = loc35, loc33 // x3.2 = mixl tmp3, tmp1 |
| 1105 |
// ;; |
| 1106 |
// mix2.l loc32 = loc45, loc44 // tmp0 = mixr y4.2, y5.2 |
| 1107 |
// mix2.r loc33 = loc45, loc44 // tmp1 = mixl y4.2, y5.2 |
| 1108 |
// mix2.l loc34 = loc47, loc46 // tmp2 = mixr y6.2, y6.2 |
| 1109 |
// mix2.r loc35 = loc47, loc46 // tmp3 = mixl y6.2, y6.2 |
| 1110 |
// ;; |
| 1111 |
// mix4.l loc28 = loc34, loc32 // x4.2 = mixr tmp2, tmp0 |
| 1112 |
// mix4.l loc29 = loc35, loc33 // x5.2 = mixr tmp3, tmp1 |
| 1113 |
// mix4.r loc30 = loc34, loc32 // x6.2 = mixl tmp2, tmp0 |
| 1114 |
// mix4.r loc31 = loc35, loc33 // x7.2 = mixl tmp3, tmp1 |
| 1115 |
// ;; |
| 1116 |
// |
| 1117 |
// // ******** |
| 1118 |
// // descale |
| 1119 |
// // ******** |
| 1120 |
// |
| 1121 |
// padd2 loc16 = loc16, r23 |
| 1122 |
// padd2 loc17 = loc17, r23 |
| 1123 |
// padd2 loc18 = loc18, r23 |
| 1124 |
// padd2 loc19 = loc19, r23 |
| 1125 |
// padd2 loc20 = loc20, r23 |
| 1126 |
// padd2 loc21 = loc21, r23 |
| 1127 |
// padd2 loc22 = loc22, r23 |
| 1128 |
// padd2 loc23 = loc23, r23 |
| 1129 |
// padd2 loc24 = loc24, r23 |
| 1130 |
// padd2 loc25 = loc25, r23 |
| 1131 |
// padd2 loc26 = loc26, r23 |
| 1132 |
// padd2 loc27 = loc27, r23 |
| 1133 |
// padd2 loc28 = loc28, r23 |
| 1134 |
// padd2 loc29 = loc29, r23 |
| 1135 |
// padd2 loc30 = loc30, r23 |
| 1136 |
// padd2 loc31 = loc31, r23 |
| 1137 |
// ;; |
| 1138 |
// pshr2 loc16 = loc16, 3 |
| 1139 |
// pshr2 loc17 = loc17, 3 |
| 1140 |
// pshr2 loc18 = loc18, 3 |
| 1141 |
// pshr2 loc19 = loc19, 3 |
| 1142 |
// pshr2 loc20 = loc20, 3 |
| 1143 |
// pshr2 loc21 = loc21, 3 |
| 1144 |
// pshr2 loc22 = loc22, 3 |
| 1145 |
// pshr2 loc23 = loc23, 3 |
| 1146 |
// pshr2 loc24 = loc24, 3 |
| 1147 |
// pshr2 loc25 = loc25, 3 |
| 1148 |
// pshr2 loc26 = loc26, 3 |
| 1149 |
// pshr2 loc27 = loc27, 3 |
| 1150 |
// pshr2 loc28 = loc28, 3 |
| 1151 |
// pshr2 loc29 = loc29, 3 |
| 1152 |
// pshr2 loc30 = loc30, 3 |
| 1153 |
// pshr2 loc31 = loc31, 3 |
| 1154 |
// ;; |
| 1155 |
// // ************ |
| 1156 |
// // Store Matrix |
| 1157 |
// // ************ |
| 1158 |
// st8 [loc0] = loc16 |
| 1159 |
// st8 [loc1] = loc24 |
| 1160 |
// st8 [loc2] = loc17 |
| 1161 |
// st8 [loc3] = loc25 |
| 1162 |
// st8 [loc4] = loc18 |
| 1163 |
// st8 [loc5] = loc26 |
| 1164 |
// st8 [loc6] = loc19 |
| 1165 |
// st8 [loc7] = loc27 |
| 1166 |
// st8 [loc8] = loc20 |
| 1167 |
// st8 [loc9] = loc28 |
| 1168 |
// st8 [loc10] = loc21 |
| 1169 |
// st8 [loc11] = loc29 |
| 1170 |
// st8 [loc12] = loc22 |
| 1171 |
// st8 [loc13] = loc30 |
| 1172 |
// st8 [loc14] = loc23 |
| 1173 |
// st8 [loc15] = loc31 |
| 1174 |
// |
| 1175 |
// mov ar.pfs = r14 |
| 1176 |
// br.ret.sptk.many b0 |
| 1177 |
// .endp fdct_ia64# |
| 1178 |
// .common fdct#,8,8 |
| 1179 |
// |
