annotate ffmpeg/libavcodec/bfin/hpel_pixels_bfin.S @ 13:844d341cf643 tip

Back up before ISMIR
author Yading Song <yading.song@eecs.qmul.ac.uk>
date Thu, 31 Oct 2013 13:17:06 +0000
parents 6840f77b83aa
children
rev   line source
yading@10 1 /*
yading@10 2 * Blackfin Pixel Operations
yading@10 3 * Copyright (C) 2007 Marc Hoffman <marc.hoffman@analog.com>
yading@10 4 *
yading@10 5 * This file is part of FFmpeg.
yading@10 6 *
yading@10 7 * FFmpeg is free software; you can redistribute it and/or
yading@10 8 * modify it under the terms of the GNU Lesser General Public
yading@10 9 * License as published by the Free Software Foundation; either
yading@10 10 * version 2.1 of the License, or (at your option) any later version.
yading@10 11 *
yading@10 12 * FFmpeg is distributed in the hope that it will be useful,
yading@10 13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
yading@10 14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
yading@10 15 * Lesser General Public License for more details.
yading@10 16 *
yading@10 17 * You should have received a copy of the GNU Lesser General Public
yading@10 18 * License along with FFmpeg; if not, write to the Free Software
yading@10 19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
yading@10 20 */
yading@10 21 #include "config_bfin.h"
yading@10 22
yading@10 23 /*
yading@10 24 motion compensation
yading@10 25 primitives
yading@10 26
yading@10 27 * Halfpel motion compensation with rounding (a+b+1)>>1.
yading@10 28 * This is an array[4][4] of motion compensation funcions for 4
yading@10 29 * horizontal blocksizes (8,16) and the 4 halfpel positions<br>
yading@10 30 * *pixels_tab[ 0->16xH 1->8xH ][ xhalfpel + 2*yhalfpel ]
yading@10 31 * @param block destination where the result is stored
yading@10 32 * @param pixels source
yading@10 33 * @param line_size number of bytes in a horizontal line of block
yading@10 34 * @param h height
yading@10 35
yading@10 36 */
yading@10 37
yading@10 38 DEFUN(put_pixels8uc,mL1,
yading@10 39 (uint8_t *block, const uint8_t *s0, const uint8_t *s1,
yading@10 40 int dest_size, int line_size, int h)):
yading@10 41 i3=r0; // dest
yading@10 42 i0=r1; // src0
yading@10 43 i1=r2; // src1
yading@10 44 r0=[sp+12]; // dest_size
yading@10 45 r2=[sp+16]; // line_size
yading@10 46 p0=[sp+20]; // h
yading@10 47 [--sp] = (r7:6);
yading@10 48 r0+=-4;
yading@10 49 m3=r0;
yading@10 50 r2+=-8;
yading@10 51 m0=r2;
yading@10 52 LSETUP(pp8$0,pp8$1) LC0=P0;
yading@10 53 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
yading@10 54
yading@10 55 pp8$0: DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
yading@10 56 R6 = BYTEOP1P(R1:0,R3:2) || R0 = [I0++M0]|| R2 =[I1++M0];
yading@10 57 R7 = BYTEOP1P(R1:0,R3:2)(R) || R0 = [I0++] || [I3++] = R6 ;
yading@10 58 pp8$1: DISALGNEXCPT || R2 = [I1++] || [I3++M3] = R7;
yading@10 59
yading@10 60 (r7:6) = [sp++];
yading@10 61 RTS;
yading@10 62 DEFUN_END(put_pixels8uc)
yading@10 63
yading@10 64 DEFUN(put_pixels16uc,mL1,
yading@10 65 (uint8_t *block, const uint8_t *s0, const uint8_t *s1,
yading@10 66 int dest_size, int line_size, int h)):
yading@10 67 link 0;
yading@10 68 [--sp] = (r7:6);
yading@10 69 i3=r0; // dest
yading@10 70 i0=r1; // src0
yading@10 71 i1=r2; // src1
yading@10 72 r0=[fp+20]; // dest_size
yading@10 73 r2=[fp+24]; // line_size
yading@10 74 p0=[fp+28]; // h
yading@10 75
yading@10 76
yading@10 77 r0+=-12;
yading@10 78 m3=r0; // line_size
yading@10 79 r2+=-16;
yading@10 80 m0=r2;
yading@10 81
yading@10 82 LSETUP(pp16$0,pp16$1) LC0=P0;
yading@10 83 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
yading@10 84
yading@10 85 pp16$0: DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
yading@10 86 R6 = BYTEOP1P(R1:0,R3:2) || R0 = [I0++] || R2 =[I1++];
yading@10 87 R7 = BYTEOP1P(R1:0,R3:2)(R) || R1 = [I0++] || R3 =[I1++];
yading@10 88 [I3++] = R6;
yading@10 89 R6 = BYTEOP1P(R1:0,R3:2) || R0 = [I0++M0] || R2 =[I1++M0];
yading@10 90 R7 = BYTEOP1P(R1:0,R3:2)(R) || R0 = [I0++] || [I3++] = R7 ;
yading@10 91 [I3++] = R6;
yading@10 92 pp16$1: DISALGNEXCPT || R2 = [I1++] || [I3++M3] = R7;
yading@10 93
yading@10 94 (r7:6) = [sp++];
yading@10 95 unlink;
yading@10 96 RTS;
yading@10 97 DEFUN_END(put_pixels16uc)
yading@10 98
yading@10 99
yading@10 100
yading@10 101
yading@10 102
yading@10 103
yading@10 104 DEFUN(put_pixels8uc_nornd,mL1,
yading@10 105 (uint8_t *block, const uint8_t *s0, const uint8_t *s1,
yading@10 106 int line_size, int h)):
yading@10 107 i3=r0; // dest
yading@10 108 i0=r1; // src0
yading@10 109 i1=r2; // src1
yading@10 110 r2=[sp+12]; // line_size
yading@10 111 p0=[sp+16]; // h
yading@10 112 [--sp] = (r7:6);
yading@10 113 r2+=-4;
yading@10 114 m3=r2;
yading@10 115 r2+=-4;
yading@10 116 m0=r2;
yading@10 117 LSETUP(pp8$2,pp8$3) LC0=P0;
yading@10 118 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
yading@10 119
yading@10 120 pp8$2: DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
yading@10 121 R6 = BYTEOP1P(R1:0,R3:2)(T) || R0 = [I0++M0]|| R2 =[I1++M0];
yading@10 122 R7 = BYTEOP1P(R1:0,R3:2)(T,R) || R0 = [I0++] || [I3++] = R6 ;
yading@10 123 pp8$3: DISALGNEXCPT || R2 = [I1++] || [I3++M3] = R7;
yading@10 124
yading@10 125 (r7:6) = [sp++];
yading@10 126 RTS;
yading@10 127 DEFUN_END(put_pixels8uc_nornd)
yading@10 128
yading@10 129 DEFUN(put_pixels16uc_nornd,mL1,
yading@10 130 (uint8_t *block, const uint8_t *s0, const uint8_t *s1,
yading@10 131 int line_size, int h)):
yading@10 132 i3=r0; // dest
yading@10 133 i0=r1; // src0
yading@10 134 i1=r2; // src1
yading@10 135 r2=[sp+12]; // line_size
yading@10 136 p0=[sp+16]; // h
yading@10 137
yading@10 138 [--sp] = (r7:6);
yading@10 139 r2+=-12;
yading@10 140 m3=r2; // line_size
yading@10 141 r2+=-4;
yading@10 142 m0=r2;
yading@10 143
yading@10 144 LSETUP(pp16$2,pp16$3) LC0=P0;
yading@10 145 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
yading@10 146
yading@10 147 pp16$2:
yading@10 148 DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
yading@10 149 R6 = BYTEOP1P(R1:0,R3:2)(T) || R0 = [I0++] || R2 =[I1++];
yading@10 150 R7 = BYTEOP1P(R1:0,R3:2)(T,R) || R1 = [I0++] || R3 =[I1++];
yading@10 151 [I3++] = R6;
yading@10 152
yading@10 153 R6 = BYTEOP1P(R1:0,R3:2)(T) || R0 = [I0++M0] || R2 =[I1++M0];
yading@10 154 R7 = BYTEOP1P(R1:0,R3:2)(T,R) || R0 = [I0++] || [I3++] = R7 ;
yading@10 155 [I3++] = R6;
yading@10 156 pp16$3: DISALGNEXCPT || R2 = [I1++] || [I3++M3] = R7;
yading@10 157
yading@10 158 (r7:6) = [sp++];
yading@10 159
yading@10 160 RTS;
yading@10 161 DEFUN_END(put_pixels16uc_nornd)
yading@10 162
yading@10 163 DEFUN(z_put_pixels16_xy2,mL1,
yading@10 164 (uint8_t *block, const uint8_t *s0,
yading@10 165 int dest_size, int line_size, int h)):
yading@10 166 link 0;
yading@10 167 [--sp] = (r7:4);
yading@10 168 i3=r0; // dest
yading@10 169 i0=r1; // src0--> pixels
yading@10 170 i1=r1; // src1--> pixels + line_size
yading@10 171 r2+=-12;
yading@10 172 m2=r2; // m2=dest_width-4
yading@10 173 r2=[fp+20];
yading@10 174 m3=r2; // line_size
yading@10 175 p0=[fp+24]; // h
yading@10 176 r2+=-16;
yading@10 177 i1+=m3; /* src1 + line_size */
yading@10 178 m0=r2; /* line-size - 20 */
yading@10 179
yading@10 180 B0 = I0;
yading@10 181 B1 = I1;
yading@10 182 B3 = I3;
yading@10 183
yading@10 184 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
yading@10 185
yading@10 186 LSETUP(LS$16E,LE$16E) LC0=P0;
yading@10 187 LS$16E: DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
yading@10 188 R4 = BYTEOP2P (R3:2,R1:0) (RNDL) || R0 = [I0++] || R2 =[I1++];
yading@10 189 R5 = BYTEOP2P (R3:2,R1:0) (RNDL,R) || R1 = [I0++] || [I3++] = R4 ;
yading@10 190 DISALGNEXCPT || R3 = [I1++] || [I3++] = R5;
yading@10 191 R4 = BYTEOP2P (R3:2,R1:0) (RNDL) || R0 = [I0++M0]|| R2 = [I1++M0];
yading@10 192 R5 = BYTEOP2P (R3:2,R1:0) (RNDL,R) || R0 = [I0++] || [I3++] = R4 ;
yading@10 193 LE$16E: DISALGNEXCPT || R2 = [I1++] || [I3++M2] = R5;
yading@10 194
yading@10 195 M1 = 1;
yading@10 196 I3 = B3;
yading@10 197 I1 = B1;
yading@10 198 I0 = B0;
yading@10 199
yading@10 200 I0 += M1;
yading@10 201 I1 += M1;
yading@10 202
yading@10 203 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
yading@10 204 LSETUP(LS$16O,LE$16O) LC0=P0;
yading@10 205 LS$16O: DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
yading@10 206 R4 = BYTEOP2P (R3:2,R1:0) (RNDH) || R0 = [I0++] || R2 =[I1++];
yading@10 207 R5 = BYTEOP2P (R3:2,R1:0) (RNDH,R) || R1 = [I0++] || R6 =[I3++];
yading@10 208 R4 = R4 +|+ R6 || R7 = [I3--];
yading@10 209 R5 = R5 +|+ R7 || [I3++] = R4;
yading@10 210 DISALGNEXCPT || R3 =[I1++] || [I3++] = R5;
yading@10 211 R4 = BYTEOP2P (R3:2,R1:0) (RNDH) || R0 = [I0++M0]|| R2 = [I1++M0];
yading@10 212 R5 = BYTEOP2P (R3:2,R1:0) (RNDH,R) || R0 = [I0++] || R6 = [I3++];
yading@10 213 R4 = R4 +|+ R6 || R7 = [I3--];
yading@10 214 R5 = R5 +|+ R7 || [I3++] = R4;
yading@10 215 LE$16O: DISALGNEXCPT || R2 = [I1++] || [I3++M2] = R5;
yading@10 216
yading@10 217 (r7:4) = [sp++];
yading@10 218 unlink;
yading@10 219 rts;
yading@10 220 DEFUN_END(z_put_pixels16_xy2)
yading@10 221
yading@10 222 DEFUN(put_pixels16_xy2_nornd,mL1,
yading@10 223 (uint8_t *block, const uint8_t *s0,
yading@10 224 int line_size, int h)):
yading@10 225 link 0;
yading@10 226 [--sp] = (r7:4);
yading@10 227 i3=r0; // dest
yading@10 228 i0=r1; // src0--> pixels
yading@10 229 i1=r1; // src1--> pixels + line_size
yading@10 230 m3=r2;
yading@10 231 r2+=-12;
yading@10 232 m2=r2;
yading@10 233 r2+=-4;
yading@10 234 i1+=m3; /* src1 + line_size */
yading@10 235 m0=r2; /* line-size - 20 */
yading@10 236 p0=[fp+20]; // h
yading@10 237
yading@10 238 B0=I0;
yading@10 239 B1=I1;
yading@10 240 B3=I3;
yading@10 241
yading@10 242 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
yading@10 243
yading@10 244 LSETUP(LS$16ET,LE$16ET) LC0=P0;
yading@10 245 LS$16ET:DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
yading@10 246 R4 = BYTEOP2P (R3:2,R1:0) (TL) || R0 = [I0++] || R2 =[I1++];
yading@10 247 R5 = BYTEOP2P (R3:2,R1:0) (TL,R) || R1 = [I0++] || [I3++] = R4 ;
yading@10 248 DISALGNEXCPT || R3 = [I1++] || [I3++] = R5;
yading@10 249 R4 = BYTEOP2P (R3:2,R1:0) (TL) || R0 = [I0++M0]|| R2 = [I1++M0];
yading@10 250 R5 = BYTEOP2P (R3:2,R1:0) (TL,R) || R0 = [I0++] || [I3++] = R4 ;
yading@10 251 LE$16ET:DISALGNEXCPT || R2 = [I1++] || [I3++M2] = R5;
yading@10 252
yading@10 253 M1 = 1;
yading@10 254 I3=B3;
yading@10 255 I1=B1;
yading@10 256 I0=B0;
yading@10 257
yading@10 258 I0 += M1;
yading@10 259 I1 += M1;
yading@10 260
yading@10 261 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
yading@10 262 LSETUP(LS$16OT,LE$16OT) LC0=P0;
yading@10 263 LS$16OT:DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
yading@10 264 R4 = BYTEOP2P (R3:2,R1:0) (TH) || R0 = [I0++] || R2 =[I1++];
yading@10 265 R5 = BYTEOP2P (R3:2,R1:0) (TH,R) || R1 = [I0++] || R6 =[I3++];
yading@10 266 R4 = R4 +|+ R6 || R7 = [I3--];
yading@10 267 R5 = R5 +|+ R7 || [I3++] = R4;
yading@10 268 DISALGNEXCPT || R3 =[I1++] || [I3++] = R5;
yading@10 269 R4 = BYTEOP2P (R3:2,R1:0) (TH) || R0 = [I0++M0]|| R2 = [I1++M0];
yading@10 270 R5 = BYTEOP2P (R3:2,R1:0) (TH,R) || R0 = [I0++] || R6 = [I3++];
yading@10 271 R4 = R4 +|+ R6 || R7 = [I3--];
yading@10 272 R5 = R5 +|+ R7 || [I3++] = R4;
yading@10 273 LE$16OT:DISALGNEXCPT || R2 = [I1++] || [I3++M2] = R5;
yading@10 274
yading@10 275 (r7:4) = [sp++];
yading@10 276 unlink;
yading@10 277 rts;
yading@10 278 DEFUN_END(put_pixels16_xy2_nornd)
yading@10 279
yading@10 280 DEFUN(z_put_pixels8_xy2,mL1,
yading@10 281 (uint8_t *block, const uint8_t *s0,
yading@10 282 int dest_size, int line_size, int h)):
yading@10 283 link 0;
yading@10 284 [--sp] = (r7:4);
yading@10 285 i3=r0; // dest
yading@10 286 i0=r1; // src0--> pixels
yading@10 287 i1=r1; // src1--> pixels + line_size
yading@10 288 r2+=-4;
yading@10 289 m2=r2; // m2=dest_width-4
yading@10 290 r2=[fp+20];
yading@10 291 m3=r2; // line_size
yading@10 292 p0=[fp+24]; // h
yading@10 293 r2+=-8;
yading@10 294 i1+=m3; /* src1 + line_size */
yading@10 295 m0=r2; /* line-size - 20 */
yading@10 296
yading@10 297 b0 = I0;
yading@10 298 b1 = I1;
yading@10 299 b3 = I3;
yading@10 300
yading@10 301 LSETUP(LS$8E,LE$8E) LC0=P0;
yading@10 302 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
yading@10 303 LS$8E: DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
yading@10 304 R4 = BYTEOP2P (R3:2,R1:0) (RNDL) || R0 = [I0++M0] || R2 =[I1++M0];
yading@10 305 R5 = BYTEOP2P (R3:2,R1:0) (RNDL,R) || R0 = [I0++] || [I3++] = R4 ;
yading@10 306 LE$8E: DISALGNEXCPT || R2 = [I1++] || [I3++M2] = R5;
yading@10 307
yading@10 308 M1 = 1;
yading@10 309 I3 = b3;
yading@10 310 I1 = b1;
yading@10 311 I0 = b0;
yading@10 312
yading@10 313 I0 += M1;
yading@10 314 I1 += M1;
yading@10 315
yading@10 316 LSETUP(LS$8O,LE$8O) LC0=P0;
yading@10 317 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
yading@10 318 LS$8O: DISALGNEXCPT || R1 = [I0++] || R3 =[I1++];
yading@10 319 R4 = BYTEOP2P (R3:2,R1:0) (RNDH) || R0 = [I0++M0] || R2 =[I1++M0];
yading@10 320 R5 = BYTEOP2P (R3:2,R1:0) (RNDH,R) || R0 = [I0++] || R6 =[I3++];
yading@10 321 R4 = R4 +|+ R6 || R7 = [I3--];
yading@10 322 R5 = R5 +|+ R7 || [I3++] = R4;
yading@10 323 LE$8O: DISALGNEXCPT || R2 =[I1++] || [I3++M2] = R5;
yading@10 324
yading@10 325 (r7:4) = [sp++];
yading@10 326 unlink;
yading@10 327 rts;
yading@10 328 DEFUN_END(z_put_pixels8_xy2)
yading@10 329
yading@10 330 DEFUN(put_pixels8_xy2_nornd,mL1,
yading@10 331 (uint8_t *block, const uint8_t *s0, int line_size, int h)):
yading@10 332 link 0;
yading@10 333 [--sp] = (r7:4);
yading@10 334 i3=r0; // dest
yading@10 335 i0=r1; // src0--> pixels
yading@10 336 i1=r1; // src1--> pixels + line_size
yading@10 337 m3=r2;
yading@10 338 r2+=-4;
yading@10 339 m2=r2;
yading@10 340 r2+=-4;
yading@10 341 i1+=m3; /* src1 + line_size */
yading@10 342 m0=r2; /* line-size - 20 */
yading@10 343 p0=[fp+20]; // h
yading@10 344
yading@10 345
yading@10 346 b0 = I0;
yading@10 347 b1 = I1;
yading@10 348 b3 = I3;
yading@10 349
yading@10 350 LSETUP(LS$8ET,LE$8ET) LC0=P0;
yading@10 351 DISALGNEXCPT || R0 = [I0++] || R2 =[I1++];
yading@10 352
yading@10 353 LS$8ET: DISALGNEXCPT || R1 = [I0++] || R3 = [I1++];
yading@10 354 R4 = BYTEOP2P (R3:2,R1:0) (TL) || R0 = [I0++M0] || R2 = [I1++M0];
yading@10 355 R5 = BYTEOP2P (R3:2,R1:0) (TL,R) || R0 = [I0++] || [I3++] = R4 ;
yading@10 356 LE$8ET: DISALGNEXCPT || R2 = [I1++] || [I3++M2] = R5;
yading@10 357
yading@10 358 M1 = 1;
yading@10 359 I3 = b3;
yading@10 360 I1 = b1;
yading@10 361 I0 = b0;
yading@10 362
yading@10 363 I0 += M1;
yading@10 364 I1 += M1;
yading@10 365
yading@10 366 LSETUP(LS$8OT,LE$8OT) LC0=P0;
yading@10 367 DISALGNEXCPT || R0 = [I0++] || R2 = [I1++];
yading@10 368
yading@10 369 LS$8OT: DISALGNEXCPT || R1 = [I0++] || R3 = [I1++];
yading@10 370 R4 = BYTEOP2P (R3:2,R1:0) (TH) || R0 = [I0++M0] || R2 = [I1++M0];
yading@10 371 R5 = BYTEOP2P (R3:2,R1:0) (TH,R) || R0 = [I0++] || R6 = [I3++];
yading@10 372 R4 = R4 +|+ R6 || R7 = [I3--];
yading@10 373 R5 = R5 +|+ R7 || [I3++] = R4;
yading@10 374 LE$8OT: DISALGNEXCPT || R2 =[I1++] || [I3++M2] = R5;
yading@10 375
yading@10 376 (r7:4) = [sp++];
yading@10 377 unlink;
yading@10 378 rts;
yading@10 379 DEFUN_END(put_pixels8_xy2_nornd)