1/* 2 * idct BlackFin 3 * 4 * Copyright (C) 2007 Marc Hoffman <marc.hoffman@analog.com> 5 * 6 * This file is part of FFmpeg. 7 * 8 * FFmpeg is free software; you can redistribute it and/or 9 * modify it under the terms of the GNU Lesser General Public 10 * License as published by the Free Software Foundation; either 11 * version 2.1 of the License, or (at your option) any later version. 12 * 13 * FFmpeg 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 GNU 16 * Lesser General Public License for more details. 17 * 18 * You should have received a copy of the GNU Lesser General Public 19 * License along with FFmpeg; if not, write to the Free Software 20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA 21 */ 22/* 23 This blackfin DSP code implements an 8x8 inverse type II DCT. 24 25Prototype : void ff_bfin_idct(DCTELEM *in) 26 27Registers Used : A0, A1, R0-R7, I0-I3, B0, B2, B3, M0-M2, L0-L3, P0-P5, LC0. 28 29Performance : 30 Code Size : 498 Bytes. 31 Cycle Count : 417 Cycles 32 33 34----------------------------------------------------------- 35FFMPEG conformance testing results 36----------------------------------------------------------- 37 38dct-test: modified with the following 39 dct_error("BFINidct", 1, ff_bfin_idct, idct, test); 40produces the following output 41 42root:/u/ffmpeg/bhead/libavcodec> ./dct-test -i 43ffmpeg DCT/IDCT test 44 45 8 15 -2 21 24 17 0 10 46 2 -10 -5 -5 -3 7 -14 -3 47 2 -13 -10 -19 18 -6 6 -2 48 9 4 16 -3 9 12 10 15 49 15 -9 -2 10 1 16 0 -15 50 -15 5 7 3 13 0 13 20 51 -6 -15 24 9 -18 1 9 -22 52 -8 25 23 2 -7 0 30 13 53IDCT BFINidct: err_inf=1 err2=0.01002344 syserr=0.00150000 maxout=266 blockSumErr=64 54IDCT BFINidct: 88.3 kdct/s 55 56*/ 57 58#include "config_bfin.h" 59 60#ifdef __FDPIC__ 61.section .l1.data.B,"aw",@progbits 62#else 63.data 64#endif 65 66.align 4; 67coefs: 68.short 0x5a82; // C4 69.short 0x5a82; // C4 70.short 0x30FC; //cos(3pi/8) C6 71.short 0x7642; //cos(pi/8) C2 72.short 0x18F9; //cos(7pi/16) 73.short 0x7D8A; //cos(pi/16) 74.short 0x471D; //cos(5pi/16) 75.short 0x6A6E; //cos(3pi/16) 76.short 0x18F9; //cos(7pi/16) 77.short 0x7D8A; //cos(pi/16) 78 79#ifdef __FDPIC__ 80.section .l1.data.A,"aw",@progbits 81#endif 82 83vtmp: .space 256 84 85#define TMP0 FP-8 86#define TMP1 FP-12 87#define TMP2 FP-16 88 89 90.text 91DEFUN(idct,mL1, 92 (DCTELEM *block)): 93 94/********************** Function Prologue *********************************/ 95 link 16; 96 [--SP] = (R7:4, P5:3); // Push the registers onto the stack. 97 B0 = R0; // Pointer to Input matrix 98 RELOC(R1, P3, coefs); // Pointer to Coefficients 99 RELOC(R2, P3, vtmp); // Pointer to Temporary matrix 100 B3 = R1; 101 B2 = R2; 102 L3 = 20; // L3 is used for making the coefficient array 103 // circular. 104 // MUST BE RESTORED TO ZERO at function exit. 105 M1 = 16 (X); // All these registers are initialized for 106 M3 = 8(X); // modifying address offsets. 107 108 I0 = B0; // I0 points to Input Element (0, 0). 109 I2 = B0; // I2 points to Input Element (0, 0). 110 I2 += M3 || R0.H = W[I0]; 111 // Element 0 is read into R0.H 112 I1 = I2; // I1 points to input Element (0, 6). 113 I1 += 4 || R0.L = W[I2++]; 114 // I2 points to input Element (0, 4). 115 // Element 4 is read into R0.L. 116 P2 = 8 (X); 117 P3 = 32 (X); 118 P4 = -32 (X); 119 P5 = 98 (X); 120 R7 = 0x8000(Z); 121 I3 = B3; // I3 points to Coefficients 122 P0 = B2; // P0 points to array Element (0, 0) of temp 123 P1 = B2; 124 R7 = [I3++] || [TMP2]=R7; // Coefficient C4 is read into R7.H and R7.L. 125 MNOP; 126 NOP; 127 128 /* 129 * A1 = Y0 * cos(pi/4) 130 * A0 = Y0 * cos(pi/4) 131 * A1 = A1 + Y4 * cos(pi/4) 132 * A0 = A0 - Y4 * cos(pi/4) 133 * load: 134 * R1=(Y2,Y6) 135 * R7=(C2,C6) 136 * res: 137 * R3=Y0, R2=Y4 138 */ 139 A1=R7.H*R0.H, A0=R7.H*R0.H (IS) || I0+= 4 || R1.L=W[I1++]; 140 R3=(A1+=R7.H*R0.L), R2=(A0-=R7.H*R0.L) (IS) || R1.H=W[I0--] || R7=[I3++]; 141 142 LSETUP (.0, .1) LC0 = P2; // perform 8 1d idcts 143 144 P2 = 112 (X); 145 P1 = P1 + P2; // P1 points to element (7, 0) of temp buffer. 146 P2 = -94(X); 147 148.0: 149 /* 150 * A1 = Y2 * cos(3pi/8) 151 * A0 = Y2 * cos(pi/8) 152 * A1 = A1 - Y6 * cos(pi/8) 153 * A0 = A0 + Y6 * cos(3pi/8) 154 * R5 = (Y1,Y7) 155 * R7 = (C1,C7) 156 * res: 157 * R1=Y2, R0=Y6 158 */ 159 A1=R7.L*R1.H, A0=R7.H*R1.H (IS) || I0+=4 || R5.H=W[I0]; 160 R1=(A1-=R7.H*R1.L), R0=(A0+=R7.L*R1.L) (IS) || R5.L=W[I1--] || R7=[I3++]; 161 /* 162 * Y0 = Y0 + Y6. 163 * Y4 = Y4 + Y2. 164 * Y2 = Y4 - Y2. 165 * Y6 = Y0 - Y6. 166 * R3 is saved 167 * R6.l=Y3 168 * note: R3: Y0, R2: Y4, R1: Y2, R0: Y6 169 */ 170 R3=R3+R0, R0=R3-R0; 171 R2=R2+R1, R1=R2-R1 || [TMP0]=R3 || R6.L=W[I0--]; 172 /* 173 * Compute the odd portion (1,3,5,7) even is done. 174 * 175 * Y1 = C7 * Y1 - C1 * Y7 + C3 * Y5 - C5 * Y3. 176 * Y7 = C1 * Y1 + C7 * Y7 + C5 * Y5 + C3 * Y3. 177 * Y5 = C5 * Y1 + C3 * Y7 + C7 * Y5 - C1 * Y3. 178 * Y3 = C3 * Y1 - C5 * Y7 - C1 * Y5 - C7 * Y3. 179 */ 180 // R5=(Y1,Y7) R6=(Y5,Y3) // R7=(C1,C7) 181 A1 =R7.L*R5.H, A0 =R7.H*R5.H (IS) || [TMP1]=R2 || R6.H=W[I2--]; 182 A1-=R7.H*R5.L, A0+=R7.L*R5.L (IS) || I0-=4 || R7=[I3++]; 183 A1+=R7.H*R6.H, A0+=R7.L*R6.H (IS) || I0+=M1; // R7=(C3,C5) 184 R3 =(A1-=R7.L*R6.L), R2 =(A0+=R7.H*R6.L) (IS); 185 A1 =R7.L*R5.H, A0 =R7.H*R5.H (IS) || R4=[TMP0]; 186 A1+=R7.H*R5.L, A0-=R7.L*R5.L (IS) || I1+=M1 || R7=[I3++]; // R7=(C1,C7) 187 A1+=R7.L*R6.H, A0-=R7.H*R6.H (IS); 188 R7 =(A1-=R7.H*R6.L), R6 =(A0-=R7.L*R6.L) (IS) || I2+=M1; 189 // R3=Y1, R2=Y7, R7=Y5, R6=Y3 190 191 /* Transpose write column. */ 192 R5.H=R4+R2 (RND12); // Y0=Y0+Y7 193 R5.L=R4-R2 (RND12) || R4 = [TMP1]; // Y7=Y7-Y0 194 R2.H=R1+R7 (RND12) || W[P0++P3]=R5.H; // Y2=Y2+Y5 st Y0 195 R2.L=R1-R7 (RND12) || W[P1++P4]=R5.L || R7=[I3++]; // Y5=Y2-Y5 st Y7 196 R5.H=R0-R3 (RND12) || W[P0++P3]=R2.H || R1.L=W[I1++]; // Y1=Y6-Y1 st Y2 197 R5.L=R0+R3 (RND12) || W[P1++P4]=R2.L || R0.H=W[I0++]; // Y6=Y6+Y1 st Y5 198 R3.H=R4-R6 (RND12) || W[P0++P3]=R5.H || R0.L=W[I2++]; // Y3=Y3-Y4 st Y1 199 R3.L=R4+R6 (RND12) || W[P1++P4]=R5.L || R1.H=W[I0++]; // Y4=Y3+Y4 st Y6 200 201 /* pipeline loop start, + drain Y3, Y4 */ 202 A1=R7.H*R0.H, A0=R7.H*R0.H (IS) || W[P0++P2]= R3.H || R1.H = W[I0--]; 203.1: R3=(A1+=R7.H*R0.L), R2=(A0-=R7.H*R0.L) (IS) || W[P1++P5]= R3.L || R7 = [I3++]; 204 205 206 207 I0 = B2; // I0 points to Input Element (0, 0) 208 I2 = B2; // I2 points to Input Element (0, 0) 209 I2 += M3 || R0.H = W[I0]; 210 // Y0 is read in R0.H 211 I1 = I2; // I1 points to input Element (0, 6) 212 I1 += 4 || R0.L = W[I2++]; 213 // I2 points to input Element (0, 4) 214 // Y4 is read in R0.L 215 P2 = 8 (X); 216 I3 = B3; // I3 points to Coefficients 217 P0 = B0; // P0 points to array Element (0, 0) for writing 218 // output 219 P1 = B0; 220 R7 = [I3++]; // R7.H = C4 and R7.L = C4 221 NOP; 222 223 /* 224 * A1 = Y0 * cos(pi/4) 225 * A0 = Y0 * cos(pi/4) 226 * A1 = A1 + Y4 * cos(pi/4) 227 * A0 = A0 - Y4 * cos(pi/4) 228 * load: 229 * R1=(Y2,Y6) 230 * R7=(C2,C6) 231 * res: 232 * R3=Y0, R2=Y4 233 */ 234 A1=R7.H*R0.H, A0=R7.H*R0.H (IS) || I0+=4 || R1.L=W[I1++]; 235 R3=(A1+=R7.H*R0.L), R2=(A0-=R7.H*R0.L) (IS) || R1.H=W[I0--] || R7=[I3++]; 236 237 LSETUP (.2, .3) LC0 = P2; // peform 8 1d idcts 238 P2 = 112 (X); 239 P1 = P1 + P2; 240 P2 = -94(X); 241 242.2: 243 /* 244 * A1 = Y2 * cos(3pi/8) 245 * A0 = Y2 * cos(pi/8) 246 * A1 = A1 - Y6 * cos(pi/8) 247 * A0 = A0 + Y6 * cos(3pi/8) 248 * R5 = (Y1,Y7) 249 * R7 = (C1,C7) 250 * res: 251 * R1=Y2, R0=Y6 252 */ 253 A1=R7.L*R1.H, A0=R7.H*R1.H (IS) || I0+=4 || R5.H=W[I0]; 254 R1=(A1-=R7.H*R1.L), R0=(A0+=R7.L*R1.L) (IS) || R5.L=W[I1--] || R7=[I3++]; 255 /* 256 * Y0 = Y0 + Y6. 257 * Y4 = Y4 + Y2. 258 * Y2 = Y4 - Y2. 259 * Y6 = Y0 - Y6. 260 * R3 is saved 261 * R6.l=Y3 262 * note: R3: Y0, R2: Y4, R1: Y2, R0: Y6 263 */ 264 R3=R3+R0, R0=R3-R0; 265 R2=R2+R1, R1=R2-R1 || [TMP0]=R3 || R6.L=W[I0--]; 266 /* 267 * Compute the odd portion (1,3,5,7) even is done. 268 * 269 * Y1 = C7 * Y1 - C1 * Y7 + C3 * Y5 - C5 * Y3. 270 * Y7 = C1 * Y1 + C7 * Y7 + C5 * Y5 + C3 * Y3. 271 * Y5 = C5 * Y1 + C3 * Y7 + C7 * Y5 - C1 * Y3. 272 * Y3 = C3 * Y1 - C5 * Y7 - C1 * Y5 - C7 * Y3. 273 */ 274 // R5=(Y1,Y7) R6=(Y5,Y3) // R7=(C1,C7) 275 A1 =R7.L*R5.H, A0 =R7.H*R5.H (IS) || [TMP1]=R2 || R6.H=W[I2--]; 276 A1-=R7.H*R5.L, A0+=R7.L*R5.L (IS) || I0-=4 || R7=[I3++]; 277 A1+=R7.H*R6.H, A0+=R7.L*R6.H (IS) || I0+=M1; // R7=(C3,C5) 278 R3 =(A1-=R7.L*R6.L), R2 =(A0+=R7.H*R6.L) (IS); 279 A1 =R7.L*R5.H, A0 =R7.H*R5.H (IS) || R4=[TMP0]; 280 A1+=R7.H*R5.L, A0-=R7.L*R5.L (IS) || I1+=M1 || R7=[I3++]; // R7=(C1,C7) 281 A1+=R7.L*R6.H, A0-=R7.H*R6.H (IS); 282 R7 =(A1-=R7.H*R6.L), R6 =(A0-=R7.L*R6.L) (IS) || I2+=M1; 283 // R3=Y1, R2=Y7, R7=Y5, R6=Y3 284 285 /* Transpose write column. */ 286 R5.H=R4+R2 (RND20); // Y0=Y0+Y7 287 R5.L=R4-R2 (RND20) || R4 = [TMP1]; // Y7=Y7-Y0 288 R2.H=R1+R7 (RND20) || W[P0++P3]=R5.H; // Y2=Y2+Y5 st Y0 289 R2.L=R1-R7 (RND20) || W[P1++P4]=R5.L || R7=[I3++]; // Y5=Y2-Y5 st Y7 290 R5.H=R0-R3 (RND20) || W[P0++P3]=R2.H || R1.L=W[I1++]; // Y1=Y6-Y1 st Y2 291 R5.L=R0+R3 (RND20) || W[P1++P4]=R2.L || R0.H=W[I0++]; // Y6=Y6+Y1 st Y5 292 R3.H=R4-R6 (RND20) || W[P0++P3]=R5.H || R0.L=W[I2++]; // Y3=Y3-Y4 st Y1 293 R3.L=R4+R6 (RND20) || W[P1++P4]=R5.L || R1.H=W[I0++]; // Y4=Y3+Y4 st Y6 294 295 /* pipeline loop start, + drain Y3, Y4 */ 296 A1=R7.H*R0.H, A0=R7.H*R0.H (IS) || W[P0++P2]= R3.H || R1.H = W[I0--]; 297.3: R3=(A1+=R7.H*R0.L), R2=(A0-=R7.H*R0.L) (IS) || W[P1++P5]= R3.L || R7 = [I3++]; 298 299 L3 = 0; 300 (R7:4,P5:3)=[SP++]; 301 unlink; 302 RTS; 303DEFUN_END(idct) 304 305 306