Submit FS#10974: Speed up mpc-codec through changing the dct32-algorithm. +5% on PP5020, +12% on M5. · tsiry-sandratraina.com/rockbox-zig@cc57237

+702 -768

2 changed files

expand all

apps

codecs

libmusepack

synth_filter.c

synth_filter_arm.S

+407 -405

apps/codecs/libmusepack/synth_filter.c

··· 43 43 #undef _ 44 44 45 45 #if defined(MPC_FIXED_POINT) 46 - #if defined(OPTIMIZE_FOR_SPEED) 47 - // round at compile time to +/- 2^14 as a pre-shift before 32=32x32-multiply 48 - #define D(value) (MPC_SHR_RND(value, 3)) 46 + #if defined(CPU_ARM) 47 + // do not up-scale D-values to achieve higher speed in smull/mlal 48 + // operations. saves ~14/8 = 1.75 cycles per multiplication 49 + #define D(value) (value) 49 50 50 - // round at runtime to +/- 2^17 as a pre-shift before 32=32x32-multiply 51 - // samples are 18.14 fixed point. 30.2 after this shift, whereas the 52 - // 15.2 bits are significant (not including sign) 53 - #define MPC_V_PRESHIFT(X) MPC_SHR_RND(X, 12) 54 - 55 - // in this configuration a post-shift by >>1 is needed after synthesis 56 - #else 57 - #if defined(CPU_ARM) 58 - // do not up-scale D-values to achieve higher speed in smull/mlal 59 - // operations. saves ~14/8 = 1.75 cycles per multiplication 60 - #define D(value) (value) 61 - 62 - // in this configuration a post-shift by >>16 is needed after synthesis 63 - #else 64 - // saturate to +/- 2^31 (= value << (31-17)), D-values are +/- 2^17 65 - #define D(value) (value << (14)) 66 - #endif 67 - // do not perform pre-shift 68 - #define MPC_V_PRESHIFT(X) (X) 69 - #endif 51 + // in this configuration a post-shift by >>16 is needed after synthesis 52 + #else 53 + // saturate to +/- 2^31 (= value << (31-17)), D-values are +/- 2^17 54 + #define D(value) (value << (14)) 55 + #endif 70 56 #else 71 57 // IMPORTANT: internal scaling is somehow strange for floating point, therefore we scale the coefficients Di_opt 72 58 // by the correct amount to have proper scaled output 73 59 #define D(value) MAKE_MPC_SAMPLE((double)value*(double)(0x1000)) 74 - 75 - // do not perform pre-shift 76 - #define MPC_V_PRESHIFT(X) (X) 77 60 #endif 78 61 79 62 // Di_opt coefficients are +/- 2^17 (pre-shifted by <<16) ··· 115 98 116 99 #undef D 117 100 118 - // needed to prevent from internal overflow in calculate_V (see below) 119 - #define OVERFLOW_FIX 2 120 - 121 - // V-coefficients were expanded (<<) by V_COEFFICIENT_EXPAND 122 - #define V_COEFFICIENT_EXPAND 27 101 + // DCT32-coefficients were expanded (<<) by DCT32_COEFFICIENT_EXPAND 102 + #define DCT32_COEFFICIENT_EXPAND 31 123 103 124 104 #if defined(MPC_FIXED_POINT) 125 - #if defined(OPTIMIZE_FOR_SPEED) 126 - // define 32=32x32-multiplication for DCT-coefficients with samples, vcoef will be pre-shifted on creation 127 - // samples are rounded to +/- 2^19 as pre-shift before 32=32x32-multiply 128 - #define MPC_MULTIPLY_V(sample, vcoef) ( MPC_SHR_RND(sample, 12) * vcoef ) 129 - 130 - // pre- and postscale are used to avoid internal overflow in synthesis calculation 131 - // samples are s15.0, v-coefs are 4.12 -> internal format is s19.12 132 - #define MPC_MULTIPLY_V_PRESCALE(sample, vcoef) ( MPC_SHR_RND(sample, (12+OVERFLOW_FIX)) * vcoef ) 133 - #define MPC_MULTIPLY_V_POSTSCALE(sample, vcoef) ( MPC_SHR_RND(sample, (12-OVERFLOW_FIX)) * vcoef ) 134 - #define MPC_V_POSTSCALE(sample) (sample<<OVERFLOW_FIX) 135 - 136 - // round to +/- 2^16 as pre-shift before 32=32x32-multiply 137 - #define MPC_MAKE_INVCOS(value) (MPC_SHR_RND(value, 15)) 138 - #else 139 - // define 64=32x32-multiplication for DCT-coefficients with samples. Via usage of MPC_FRACT highly optimized assembler might be used 140 - // MULTIPLY_FRACT will do >>32 after multiplication, as V-coef were expanded by V_COEFFICIENT_EXPAND we'll correct this on the result. 141 - // Will loose 5bit accuracy on result in fract part without effect on final audio result 142 - #define MPC_MULTIPLY_V(sample, vcoef) ( (MPC_MULTIPLY_FRACT(sample, vcoef)) << (32-V_COEFFICIENT_EXPAND) ) 143 - 144 - // pre- and postscale are used to avoid internal overflow in synthesis calculation 145 - // samples are s15.14, v-coefs are 4.27 -> internal format is s19.12 146 - #define MPC_MULTIPLY_V_PRESCALE(sample, vcoef) ( (MPC_MULTIPLY_FRACT(sample, vcoef)) << (32-V_COEFFICIENT_EXPAND-OVERFLOW_FIX) ) 147 - #define MPC_MULTIPLY_V_POSTSCALE(sample, vcoef) ( (MPC_MULTIPLY_FRACT(sample, vcoef)) << (32-V_COEFFICIENT_EXPAND+OVERFLOW_FIX) ) 148 - #define MPC_V_POSTSCALE(sample) (sample<<OVERFLOW_FIX) 149 - 150 - // directly use accurate 32bit-coefficients 151 - #define MPC_MAKE_INVCOS(value) (value) 152 - #endif 105 + // define 64=32x32-multiplication for DCT-coefficients with samples. Via usage of MPC_FRACT highly optimized assembler might be used 106 + // MULTIPLY_FRACT will perform >>32 after multiplication, as coef were expanded by DCT32_COEFFICIENT_EXPAND we'll correct this on the result. 107 + // Will loose 4 bit accuracy on result in fract part without effect on final audio result 108 + #define MPC_DCT32_MUL(sample, coef) (MPC_MULTIPLY_FRACT(sample,coef) << (32-DCT32_COEFFICIENT_EXPAND)) 109 + #define MPC_DCT32_SHIFT(sample) (sample) 153 110 #else 154 111 // for floating point use the standard multiplication macro 155 - #define MPC_MULTIPLY_V (sample, vcoef) ( MPC_MULTIPLY(sample, vcoef) ) 156 - #define MPC_MULTIPLY_V_PRESCALE (sample, vcoef) ( MPC_MULTIPLY(sample, vcoef) ) 157 - #define MPC_MULTIPLY_V_POSTSCALE(sample, vcoef) ( MPC_MULTIPLY(sample, vcoef) ) 158 - #define MPC_V_POSTSCALE(sample) (sample) 159 - 160 - // downscale the accurate 32bit-coefficients and convert to float 161 - #define MPC_MAKE_INVCOS(value) MAKE_MPC_SAMPLE((double)value/(double)(1<<V_COEFFICIENT_EXPAND)) 112 + #define MPC_DCT32_MUL(sample, coef) (MPC_MULTIPLY(sample, coef) ) 113 + #define MPC_DCT32_SHIFT(sample) (sample) 162 114 #endif 163 115 164 - // define constants for DCT-synthesis 165 - // INVCOSxx = (0.5 / cos(xx*PI/64)) << 27, <<27 to saturate to +/- 2^31 166 - #define INVCOS01 MPC_MAKE_INVCOS( 67189797) 167 - #define INVCOS02 MPC_MAKE_INVCOS( 67433575) 168 - #define INVCOS03 MPC_MAKE_INVCOS( 67843164) 169 - #define INVCOS04 MPC_MAKE_INVCOS( 68423604) 170 - #define INVCOS05 MPC_MAKE_INVCOS( 69182167) 171 - #define INVCOS06 MPC_MAKE_INVCOS( 70128577) 172 - #define INVCOS07 MPC_MAKE_INVCOS( 71275330) 173 - #define INVCOS08 MPC_MAKE_INVCOS( 72638111) 174 - #define INVCOS09 MPC_MAKE_INVCOS( 74236348) 175 - #define INVCOS10 MPC_MAKE_INVCOS( 76093940) 176 - #define INVCOS11 MPC_MAKE_INVCOS( 78240207) 177 - #define INVCOS12 MPC_MAKE_INVCOS( 80711144) 178 - #define INVCOS13 MPC_MAKE_INVCOS( 83551089) 179 - #define INVCOS14 MPC_MAKE_INVCOS( 86814950) 180 - #define INVCOS15 MPC_MAKE_INVCOS( 90571242) 181 - #define INVCOS16 MPC_MAKE_INVCOS( 94906266) 182 - #define INVCOS17 MPC_MAKE_INVCOS( 99929967) 183 - #define INVCOS18 MPC_MAKE_INVCOS( 105784321) 184 - #define INVCOS19 MPC_MAKE_INVCOS( 112655602) 185 - #define INVCOS20 MPC_MAKE_INVCOS( 120792764) 186 - #define INVCOS21 MPC_MAKE_INVCOS( 130535899) 187 - #define INVCOS22 MPC_MAKE_INVCOS( 142361749) 188 - #define INVCOS23 MPC_MAKE_INVCOS( 156959571) 189 - #define INVCOS24 MPC_MAKE_INVCOS( 175363913) 190 - #define INVCOS25 MPC_MAKE_INVCOS( 199201203) 191 - #define INVCOS26 MPC_MAKE_INVCOS( 231182936) 192 - #define INVCOS27 MPC_MAKE_INVCOS( 276190692) 193 - #define INVCOS28 MPC_MAKE_INVCOS( 343988688) 194 - #define INVCOS29 MPC_MAKE_INVCOS( 457361460) 195 - #define INVCOS30 MPC_MAKE_INVCOS( 684664578) 196 - #define INVCOS31 MPC_MAKE_INVCOS(1367679739) 197 - 116 + /****************************************************************************** 117 + * mpc_dct32(const int *in, int *out) 118 + * 119 + * mpc_dct32 is a dct32 with in[32]->dct[32] that contains the mirroring from 120 + * dct[32] to the expected out[64]. The symmetry is 121 + * out[16] = 0, 122 + * out[ 0..15] = dct[ 0..15], 123 + * out[32..17] = -dct[ 0..15], 124 + * out[33..48] = -dct[16..31], 125 + * out[63..48] = -dct[16..31]. 126 + * The cos-tab has the format s0.31. 127 + *****************************************************************************/ 198 128 void 199 - mpc_calculate_new_V ( const MPC_SAMPLE_FORMAT * Sample, MPC_SAMPLE_FORMAT * V ) 200 - ICODE_ATTR_MPC_LARGE_IRAM; 129 + mpc_dct32(const MPC_SAMPLE_FORMAT *in, MPC_SAMPLE_FORMAT *v) 130 + ICODE_ATTR_MPC_LARGE_IRAM; 201 131 202 - void 203 - mpc_calculate_new_V ( const MPC_SAMPLE_FORMAT * Sample, MPC_SAMPLE_FORMAT * V ) 132 + void 133 + mpc_dct32(const MPC_SAMPLE_FORMAT *in, MPC_SAMPLE_FORMAT *v) 204 134 { 205 - // Calculating new V-buffer values for left channel 206 - // calculate new V-values (ISO-11172-3, p. 39) 207 - // based upon fast-MDCT algorithm by Byeong Gi Lee 208 - MPC_SAMPLE_FORMAT A[16]; 209 - MPC_SAMPLE_FORMAT B[16]; 210 - MPC_SAMPLE_FORMAT tmp; 135 + MPC_SAMPLE_FORMAT t0, t1, t2, t3, t4, t5, t6, t7; 136 + MPC_SAMPLE_FORMAT t8, t9, t10, t11, t12, t13, t14, t15; 137 + MPC_SAMPLE_FORMAT t16, t17, t18, t19, t20, t21, t22, t23; 138 + MPC_SAMPLE_FORMAT t24, t25, t26, t27, t28, t29, t30, t31; 139 + MPC_SAMPLE_FORMAT t32, t33, t34, t35, t36, t37, t38, t39; 140 + MPC_SAMPLE_FORMAT t40, t41, t42, t43, t44, t45, t46, t47; 141 + MPC_SAMPLE_FORMAT t48, t49, t50, t51, t52, t53, t54, t55; 142 + MPC_SAMPLE_FORMAT t56, t57, t58, t59, t60, t61, t62, t63; 143 + MPC_SAMPLE_FORMAT t64, t65, t66, t67, t68, t69, t70, t71; 144 + MPC_SAMPLE_FORMAT t72, t73, t74, t75, t76, t77, t78, t79; 145 + MPC_SAMPLE_FORMAT t80, t81, t82, t83, t84, t85, t86, t87; 146 + MPC_SAMPLE_FORMAT t88, t89, t90, t91, t92, t93, t94, t95; 147 + MPC_SAMPLE_FORMAT t96, t97, t98, t99, t100, t101, t102, t103; 148 + MPC_SAMPLE_FORMAT t104, t105, t106, t107, t108, t109, t110, t111; 149 + MPC_SAMPLE_FORMAT t112, t113, t114, t115, t116, t117, t118, t119; 150 + MPC_SAMPLE_FORMAT t120, t121, t122, t123, t124, t125, t126, t127; 151 + MPC_SAMPLE_FORMAT t128, t129, t130, t131, t132, t133, t134, t135; 152 + MPC_SAMPLE_FORMAT t136, t137, t138, t139, t140, t141, t142, t143; 153 + MPC_SAMPLE_FORMAT t144, t145, t146, t147, t148, t149, t150, t151; 154 + MPC_SAMPLE_FORMAT t152, t153, t154, t155, t156, t157, t158, t159; 155 + MPC_SAMPLE_FORMAT t160, t161, t162, t163, t164, t165, t166, t167; 156 + MPC_SAMPLE_FORMAT t168, t169, t170, t171, t172, t173, t174, t175; 157 + MPC_SAMPLE_FORMAT t176; 211 158 212 - A[ 0] = Sample[ 0] + Sample[31]; 213 - A[ 1] = Sample[ 1] + Sample[30]; 214 - A[ 2] = Sample[ 2] + Sample[29]; 215 - A[ 3] = Sample[ 3] + Sample[28]; 216 - A[ 4] = Sample[ 4] + Sample[27]; 217 - A[ 5] = Sample[ 5] + Sample[26]; 218 - A[ 6] = Sample[ 6] + Sample[25]; 219 - A[ 7] = Sample[ 7] + Sample[24]; 220 - A[ 8] = Sample[ 8] + Sample[23]; 221 - A[ 9] = Sample[ 9] + Sample[22]; 222 - A[10] = Sample[10] + Sample[21]; 223 - A[11] = Sample[11] + Sample[20]; 224 - A[12] = Sample[12] + Sample[19]; 225 - A[13] = Sample[13] + Sample[18]; 226 - A[14] = Sample[14] + Sample[17]; 227 - A[15] = Sample[15] + Sample[16]; 228 - // 16 adds 159 + /* costab[i] = cos(PI / (2 * 32) * i) */ 160 + #define costab01 (0x7fd8878e) /* 0.998795456 */ 161 + #define costab02 (0x7f62368f) /* 0.995184727 */ 162 + #define costab03 (0x7e9d55fc) /* 0.989176510 */ 163 + #define costab04 (0x7d8a5f40) /* 0.980785280 */ 164 + #define costab05 (0x7c29fbee) /* 0.970031253 */ 165 + #define costab06 (0x7a7d055b) /* 0.956940336 */ 166 + #define costab07 (0x78848414) /* 0.941544065 */ 167 + #define costab08 (0x7641af3d) /* 0.923879533 */ 168 + #define costab09 (0x73b5ebd1) /* 0.903989293 */ 169 + #define costab10 (0x70e2cbc6) /* 0.881921264 */ 170 + #define costab11 (0x6dca0d14) /* 0.857728610 */ 171 + #define costab12 (0x6a6d98a4) /* 0.831469612 */ 172 + #define costab13 (0x66cf8120) /* 0.803207531 */ 173 + #define costab14 (0x62f201ac) /* 0.773010453 */ 174 + #define costab15 (0x5ed77c8a) /* 0.740951125 */ 175 + #define costab16 (0x5a82799a) /* 0.707106781 */ 176 + #define costab17 (0x55f5a4d2) /* 0.671558955 */ 177 + #define costab18 (0x5133cc94) /* 0.634393284 */ 178 + #define costab19 (0x4c3fdff4) /* 0.595699304 */ 179 + #define costab20 (0x471cece7) /* 0.555570233 */ 180 + #define costab21 (0x41ce1e65) /* 0.514102744 */ 181 + #define costab22 (0x3c56ba70) /* 0.471396737 */ 182 + #define costab23 (0x36ba2014) /* 0.427555093 */ 183 + #define costab24 (0x30fbc54d) /* 0.382683432 */ 184 + #define costab25 (0x2b1f34eb) /* 0.336889853 */ 185 + #define costab26 (0x25280c5e) /* 0.290284677 */ 186 + #define costab27 (0x1f19f97b) /* 0.242980180 */ 187 + #define costab28 (0x18f8b83c) /* 0.195090322 */ 188 + #define costab29 (0x12c8106f) /* 0.146730474 */ 189 + #define costab30 (0x0c8bd35e) /* 0.098017140 */ 190 + #define costab31 (0x0647d97c) /* 0.049067674 */ 229 191 230 - B[ 0] = A[ 0] + A[15]; 231 - B[ 1] = A[ 1] + A[14]; 232 - B[ 2] = A[ 2] + A[13]; 233 - B[ 3] = A[ 3] + A[12]; 234 - B[ 4] = A[ 4] + A[11]; 235 - B[ 5] = A[ 5] + A[10]; 236 - B[ 6] = A[ 6] + A[ 9]; 237 - B[ 7] = A[ 7] + A[ 8]; 238 - B[ 8] = MPC_MULTIPLY_V((A[ 0] - A[15]), INVCOS02); 239 - B[ 9] = MPC_MULTIPLY_V((A[ 1] - A[14]), INVCOS06); 240 - B[10] = MPC_MULTIPLY_V((A[ 2] - A[13]), INVCOS10); 241 - B[11] = MPC_MULTIPLY_V((A[ 3] - A[12]), INVCOS14); 242 - B[12] = MPC_MULTIPLY_V((A[ 4] - A[11]), INVCOS18); 243 - B[13] = MPC_MULTIPLY_V((A[ 5] - A[10]), INVCOS22); 244 - B[14] = MPC_MULTIPLY_V((A[ 6] - A[ 9]), INVCOS26); 245 - B[15] = MPC_MULTIPLY_V((A[ 7] - A[ 8]), INVCOS30); 246 - // 8 adds, 8 subs, 8 muls, 8 shifts 192 + t0 = in[ 0] + in[31]; t16 = MPC_DCT32_MUL(in[ 0] - in[31], costab01); 193 + t1 = in[15] + in[16]; t17 = MPC_DCT32_MUL(in[15] - in[16], costab31); 247 194 248 - A[ 0] = B[ 0] + B[ 7]; 249 - A[ 1] = B[ 1] + B[ 6]; 250 - A[ 2] = B[ 2] + B[ 5]; 251 - A[ 3] = B[ 3] + B[ 4]; 252 - A[ 4] = MPC_MULTIPLY_V((B[ 0] - B[ 7]), INVCOS04); 253 - A[ 5] = MPC_MULTIPLY_V((B[ 1] - B[ 6]), INVCOS12); 254 - A[ 6] = MPC_MULTIPLY_V((B[ 2] - B[ 5]), INVCOS20); 255 - A[ 7] = MPC_MULTIPLY_V((B[ 3] - B[ 4]), INVCOS28); 256 - A[ 8] = B[ 8] + B[15]; 257 - A[ 9] = B[ 9] + B[14]; 258 - A[10] = B[10] + B[13]; 259 - A[11] = B[11] + B[12]; 260 - A[12] = MPC_MULTIPLY_V((B[ 8] - B[15]), INVCOS04); 261 - A[13] = MPC_MULTIPLY_V((B[ 9] - B[14]), INVCOS12); 262 - A[14] = MPC_MULTIPLY_V((B[10] - B[13]), INVCOS20); 263 - A[15] = MPC_MULTIPLY_V((B[11] - B[12]), INVCOS28); 264 - // 8 adds, 8 subs, 8 muls, 8 shifts 195 + t41 = t16 + t17; 196 + t59 = MPC_DCT32_MUL(t16 - t17, costab02); 197 + t33 = t0 + t1; 198 + t50 = MPC_DCT32_MUL(t0 - t1, costab02); 265 199 266 - B[ 0] = A[ 0] + A[ 3]; 267 - B[ 1] = A[ 1] + A[ 2]; 268 - B[ 2] = MPC_MULTIPLY_V((A[ 0] - A[ 3]), INVCOS08); 269 - B[ 3] = MPC_MULTIPLY_V((A[ 1] - A[ 2]), INVCOS24); 270 - B[ 4] = A[ 4] + A[ 7]; 271 - B[ 5] = A[ 5] + A[ 6]; 272 - B[ 6] = MPC_MULTIPLY_V((A[ 4] - A[ 7]), INVCOS08); 273 - B[ 7] = MPC_MULTIPLY_V((A[ 5] - A[ 6]), INVCOS24); 274 - B[ 8] = A[ 8] + A[11]; 275 - B[ 9] = A[ 9] + A[10]; 276 - B[10] = MPC_MULTIPLY_V((A[ 8] - A[11]), INVCOS08); 277 - B[11] = MPC_MULTIPLY_V((A[ 9] - A[10]), INVCOS24); 278 - B[12] = A[12] + A[15]; 279 - B[13] = A[13] + A[14]; 280 - B[14] = MPC_MULTIPLY_V((A[12] - A[15]), INVCOS08); 281 - B[15] = MPC_MULTIPLY_V((A[13] - A[14]), INVCOS24); 282 - // 8 adds, 8 subs, 8 muls, 8 shifts 200 + t2 = in[ 7] + in[24]; t18 = MPC_DCT32_MUL(in[ 7] - in[24], costab15); 201 + t3 = in[ 8] + in[23]; t19 = MPC_DCT32_MUL(in[ 8] - in[23], costab17); 283 202 284 - A[ 0] = B[ 0] + B[ 1]; 285 - A[ 1] = MPC_MULTIPLY_V((B[ 0] - B[ 1]), INVCOS16); 286 - A[ 2] = B[ 2] + B[ 3]; 287 - A[ 3] = MPC_MULTIPLY_V((B[ 2] - B[ 3]), INVCOS16); 288 - A[ 4] = B[ 4] + B[ 5]; 289 - A[ 5] = MPC_MULTIPLY_V((B[ 4] - B[ 5]), INVCOS16); 290 - A[ 6] = B[ 6] + B[ 7]; 291 - A[ 7] = MPC_MULTIPLY_V((B[ 6] - B[ 7]), INVCOS16); 292 - A[ 8] = B[ 8] + B[ 9]; 293 - A[ 9] = MPC_MULTIPLY_V((B[ 8] - B[ 9]), INVCOS16); 294 - A[10] = B[10] + B[11]; 295 - A[11] = MPC_MULTIPLY_V((B[10] - B[11]), INVCOS16); 296 - A[12] = B[12] + B[13]; 297 - A[13] = MPC_MULTIPLY_V((B[12] - B[13]), INVCOS16); 298 - A[14] = B[14] + B[15]; 299 - A[15] = MPC_MULTIPLY_V((B[14] - B[15]), INVCOS16); 300 - // 8 adds, 8 subs, 8 muls, 8 shifts 203 + t42 = t18 + t19; 204 + t60 = MPC_DCT32_MUL(t18 - t19, costab30); 205 + t34 = t2 + t3; 206 + t51 = MPC_DCT32_MUL(t2 - t3, costab30); 301 207 302 - // multiple used expressions: -(A[12] + A[14] + A[15]) 303 - V[48] = -A[ 0]; 304 - V[ 0] = A[ 1]; 305 - V[40] = -A[ 2] - (V[ 8] = A[ 3]); 306 - V[36] = -((V[ 4] = A[ 5] + (V[12] = A[ 7])) + A[ 6]); 307 - V[44] = - A[ 4] - A[ 6] - A[ 7]; 308 - V[ 6] = (V[10] = A[11] + (V[14] = A[15])) + A[13]; 309 - V[38] = (V[34] = -(V[ 2] = A[ 9] + A[13] + A[15]) - A[14]) + A[ 9] - A[10] - A[11]; 310 - V[46] = (tmp = -(A[12] + A[14] + A[15])) - A[ 8]; 311 - V[42] = tmp - A[10] - A[11]; 312 - // 9 adds, 9 subs 208 + t4 = in[ 3] + in[28]; t20 = MPC_DCT32_MUL(in[ 3] - in[28], costab07); 209 + t5 = in[12] + in[19]; t21 = MPC_DCT32_MUL(in[12] - in[19], costab25); 313 210 314 - A[ 0] = MPC_MULTIPLY_V_PRESCALE((Sample[ 0] - Sample[31]), INVCOS01); 315 - A[ 1] = MPC_MULTIPLY_V_PRESCALE((Sample[ 1] - Sample[30]), INVCOS03); 316 - A[ 2] = MPC_MULTIPLY_V_PRESCALE((Sample[ 2] - Sample[29]), INVCOS05); 317 - A[ 3] = MPC_MULTIPLY_V_PRESCALE((Sample[ 3] - Sample[28]), INVCOS07); 318 - A[ 4] = MPC_MULTIPLY_V_PRESCALE((Sample[ 4] - Sample[27]), INVCOS09); 319 - A[ 5] = MPC_MULTIPLY_V_PRESCALE((Sample[ 5] - Sample[26]), INVCOS11); 320 - A[ 6] = MPC_MULTIPLY_V_PRESCALE((Sample[ 6] - Sample[25]), INVCOS13); 321 - A[ 7] = MPC_MULTIPLY_V_PRESCALE((Sample[ 7] - Sample[24]), INVCOS15); 322 - A[ 8] = MPC_MULTIPLY_V_PRESCALE((Sample[ 8] - Sample[23]), INVCOS17); 323 - A[ 9] = MPC_MULTIPLY_V_PRESCALE((Sample[ 9] - Sample[22]), INVCOS19); 324 - A[10] = MPC_MULTIPLY_V_PRESCALE((Sample[10] - Sample[21]), INVCOS21); 325 - A[11] = MPC_MULTIPLY_V_PRESCALE((Sample[11] - Sample[20]), INVCOS23); 326 - A[12] = MPC_MULTIPLY_V_PRESCALE((Sample[12] - Sample[19]), INVCOS25); 327 - A[13] = MPC_MULTIPLY_V_PRESCALE((Sample[13] - Sample[18]), INVCOS27); 328 - A[14] = MPC_MULTIPLY_V_PRESCALE((Sample[14] - Sample[17]), INVCOS29); 329 - A[15] = MPC_MULTIPLY_V_PRESCALE((Sample[15] - Sample[16]), INVCOS31); 330 - // 16 subs, 16 muls, 16 shifts 211 + t43 = t20 + t21; 212 + t61 = MPC_DCT32_MUL(t20 - t21, costab14); 213 + t35 = t4 + t5; 214 + t52 = MPC_DCT32_MUL(t4 - t5, costab14); 331 215 332 - B[ 0] = A[ 0] + A[15]; 333 - B[ 1] = A[ 1] + A[14]; 334 - B[ 2] = A[ 2] + A[13]; 335 - B[ 3] = A[ 3] + A[12]; 336 - B[ 4] = A[ 4] + A[11]; 337 - B[ 5] = A[ 5] + A[10]; 338 - B[ 6] = A[ 6] + A[ 9]; 339 - B[ 7] = A[ 7] + A[ 8]; 340 - B[ 8] = MPC_MULTIPLY_V((A[ 0] - A[15]), INVCOS02); 341 - B[ 9] = MPC_MULTIPLY_V((A[ 1] - A[14]), INVCOS06); 342 - B[10] = MPC_MULTIPLY_V((A[ 2] - A[13]), INVCOS10); 343 - B[11] = MPC_MULTIPLY_V((A[ 3] - A[12]), INVCOS14); 344 - B[12] = MPC_MULTIPLY_V((A[ 4] - A[11]), INVCOS18); 345 - B[13] = MPC_MULTIPLY_V((A[ 5] - A[10]), INVCOS22); 346 - B[14] = MPC_MULTIPLY_V((A[ 6] - A[ 9]), INVCOS26); 347 - B[15] = MPC_MULTIPLY_V((A[ 7] - A[ 8]), INVCOS30); 348 - // 8 adds, 8 subs, 8 muls, 8 shift 216 + t6 = in[ 4] + in[27]; t22 = MPC_DCT32_MUL(in[ 4] - in[27], costab09); 217 + t7 = in[11] + in[20]; t23 = MPC_DCT32_MUL(in[11] - in[20], costab23); 349 218 350 - A[ 0] = B[ 0] + B[ 7]; 351 - A[ 1] = B[ 1] + B[ 6]; 352 - A[ 2] = B[ 2] + B[ 5]; 353 - A[ 3] = B[ 3] + B[ 4]; 354 - A[ 4] = MPC_MULTIPLY_V((B[ 0] - B[ 7]), INVCOS04); 355 - A[ 5] = MPC_MULTIPLY_V((B[ 1] - B[ 6]), INVCOS12); 356 - A[ 6] = MPC_MULTIPLY_V((B[ 2] - B[ 5]), INVCOS20); 357 - A[ 7] = MPC_MULTIPLY_V((B[ 3] - B[ 4]), INVCOS28); 358 - A[ 8] = B[ 8] + B[15]; 359 - A[ 9] = B[ 9] + B[14]; 360 - A[10] = B[10] + B[13]; 361 - A[11] = B[11] + B[12]; 362 - A[12] = MPC_MULTIPLY_V((B[ 8] - B[15]), INVCOS04); 363 - A[13] = MPC_MULTIPLY_V((B[ 9] - B[14]), INVCOS12); 364 - A[14] = MPC_MULTIPLY_V((B[10] - B[13]), INVCOS20); 365 - A[15] = MPC_MULTIPLY_V((B[11] - B[12]), INVCOS28); 366 - // 8 adds, 8 subs, 8 muls, 8 shift 219 + t44 = t22 + t23; 220 + t62 = MPC_DCT32_MUL(t22 - t23, costab18); 221 + t36 = t6 + t7; 222 + t53 = MPC_DCT32_MUL(t6 - t7, costab18); 367 223 368 - B[ 0] = A[ 0] + A[ 3]; 369 - B[ 1] = A[ 1] + A[ 2]; 370 - B[ 2] = MPC_MULTIPLY_V((A[ 0] - A[ 3]), INVCOS08); 371 - B[ 3] = MPC_MULTIPLY_V((A[ 1] - A[ 2]), INVCOS24); 372 - B[ 4] = A[ 4] + A[ 7]; 373 - B[ 5] = A[ 5] + A[ 6]; 374 - B[ 6] = MPC_MULTIPLY_V((A[ 4] - A[ 7]), INVCOS08); 375 - B[ 7] = MPC_MULTIPLY_V((A[ 5] - A[ 6]), INVCOS24); 376 - B[ 8] = A[ 8] + A[11]; 377 - B[ 9] = A[ 9] + A[10]; 378 - B[10] = MPC_MULTIPLY_V((A[ 8] - A[11]), INVCOS08); 379 - B[11] = MPC_MULTIPLY_V((A[ 9] - A[10]), INVCOS24); 380 - B[12] = A[12] + A[15]; 381 - B[13] = A[13] + A[14]; 382 - B[14] = MPC_MULTIPLY_V((A[12] - A[15]), INVCOS08); 383 - B[15] = MPC_MULTIPLY_V((A[13] - A[14]), INVCOS24); 384 - // 8 adds, 8 subs, 8 muls, 8 shift 224 + t8 = in[ 1] + in[30]; t24 = MPC_DCT32_MUL(in[ 1] - in[30], costab03); 225 + t9 = in[14] + in[17]; t25 = MPC_DCT32_MUL(in[14] - in[17], costab29); 385 226 386 - A[ 0] = MPC_V_POSTSCALE((B[ 0] + B[ 1])); 387 - A[ 1] = MPC_MULTIPLY_V_POSTSCALE((B[ 0] - B[ 1]), INVCOS16); 388 - A[ 2] = MPC_V_POSTSCALE((B[ 2] + B[ 3])); 389 - A[ 3] = MPC_MULTIPLY_V_POSTSCALE((B[ 2] - B[ 3]), INVCOS16); 390 - A[ 4] = MPC_V_POSTSCALE((B[ 4] + B[ 5])); 391 - A[ 5] = MPC_MULTIPLY_V_POSTSCALE((B[ 4] - B[ 5]), INVCOS16); 392 - A[ 6] = MPC_V_POSTSCALE((B[ 6] + B[ 7])); 393 - A[ 7] = MPC_MULTIPLY_V_POSTSCALE((B[ 6] - B[ 7]), INVCOS16); 394 - A[ 8] = MPC_V_POSTSCALE((B[ 8] + B[ 9])); 395 - A[ 9] = MPC_MULTIPLY_V_POSTSCALE((B[ 8] - B[ 9]), INVCOS16); 396 - A[10] = MPC_V_POSTSCALE((B[10] + B[11])); 397 - A[11] = MPC_MULTIPLY_V_POSTSCALE((B[10] - B[11]), INVCOS16); 398 - A[12] = MPC_V_POSTSCALE((B[12] + B[13])); 399 - A[13] = MPC_MULTIPLY_V_POSTSCALE((B[12] - B[13]), INVCOS16); 400 - A[14] = MPC_V_POSTSCALE((B[14] + B[15])); 401 - A[15] = MPC_MULTIPLY_V_POSTSCALE((B[14] - B[15]), INVCOS16); 402 - // 8 adds, 8 subs, 8 muls, 8 shift 227 + t45 = t24 + t25; 228 + t63 = MPC_DCT32_MUL(t24 - t25, costab06); 229 + t37 = t8 + t9; 230 + t54 = MPC_DCT32_MUL(t8 - t9, costab06); 403 231 404 - // multiple used expressions: A[ 4]+A[ 6]+A[ 7], A[ 9]+A[13]+A[15] 405 - V[ 5] = (V[11] = (V[13] = A[ 7] + (V[15] = A[15])) + A[11]) + A[ 5] + A[13]; 406 - V[ 7] = (V[ 9] = A[ 3] + A[11] + A[15]) + A[13]; 407 - V[33] = -(V[ 1] = A[ 1] + A[ 9] + A[13] + A[15]) - A[14]; 408 - V[35] = -(V[ 3] = A[ 5] + A[ 7] + A[ 9] + A[13] + A[15]) - A[ 6] - A[14]; 409 - V[37] = (tmp = -(A[10] + A[11] + A[13] + A[14] + A[15])) - A[ 5] - A[ 6] - A[ 7]; 410 - V[39] = tmp - A[ 2] - A[ 3]; 411 - V[41] = (tmp += A[13] - A[12]) - A[ 2] - A[ 3]; 412 - V[43] = tmp - A[ 4] - A[ 6] - A[ 7]; 413 - V[47] = (tmp = -(A[ 8] + A[12] + A[14] + A[15])) - A[ 0]; 414 - V[45] = tmp - A[ 4] - A[ 6] - A[ 7]; 415 - // 22 adds, 18 subs 232 + t10 = in[ 6] + in[25]; t26 = MPC_DCT32_MUL(in[ 6] - in[25], costab13); 233 + t11 = in[ 9] + in[22]; t27 = MPC_DCT32_MUL(in[ 9] - in[22], costab19); 234 + 235 + t46 = t26 + t27; 236 + t64 = MPC_DCT32_MUL(t26 - t27, costab26); 237 + t38 = t10 + t11; 238 + t55 = MPC_DCT32_MUL(t10 - t11, costab26); 239 + 240 + t12 = in[ 2] + in[29]; t28 = MPC_DCT32_MUL(in[ 2] - in[29], costab05); 241 + t13 = in[13] + in[18]; t29 = MPC_DCT32_MUL(in[13] - in[18], costab27); 242 + 243 + t47 = t28 + t29; 244 + t65 = MPC_DCT32_MUL(t28 - t29, costab10); 245 + t39 = t12 + t13; 246 + t56 = MPC_DCT32_MUL(t12 - t13, costab10); 247 + 248 + t14 = in[ 5] + in[26]; t30 = MPC_DCT32_MUL(in[ 5] - in[26], costab11); 249 + t15 = in[10] + in[21]; t31 = MPC_DCT32_MUL(in[10] - in[21], costab21); 250 + 251 + t48 = t30 + t31; 252 + t66 = MPC_DCT32_MUL(t30 - t31, costab22); 253 + t40 = t14 + t15; 254 + t57 = MPC_DCT32_MUL(t14 - t15, costab22); 255 + 256 + t69 = t33 + t34; t89 = MPC_DCT32_MUL(t33 - t34, costab04); 257 + t70 = t35 + t36; t90 = MPC_DCT32_MUL(t35 - t36, costab28); 258 + t71 = t37 + t38; t91 = MPC_DCT32_MUL(t37 - t38, costab12); 259 + t72 = t39 + t40; t92 = MPC_DCT32_MUL(t39 - t40, costab20); 260 + t73 = t41 + t42; t94 = MPC_DCT32_MUL(t41 - t42, costab04); 261 + t74 = t43 + t44; t95 = MPC_DCT32_MUL(t43 - t44, costab28); 262 + t75 = t45 + t46; t96 = MPC_DCT32_MUL(t45 - t46, costab12); 263 + t76 = t47 + t48; t97 = MPC_DCT32_MUL(t47 - t48, costab20); 264 + 265 + t78 = t50 + t51; t100 = MPC_DCT32_MUL(t50 - t51, costab04); 266 + t79 = t52 + t53; t101 = MPC_DCT32_MUL(t52 - t53, costab28); 267 + t80 = t54 + t55; t102 = MPC_DCT32_MUL(t54 - t55, costab12); 268 + t81 = t56 + t57; t103 = MPC_DCT32_MUL(t56 - t57, costab20); 269 + 270 + t83 = t59 + t60; t106 = MPC_DCT32_MUL(t59 - t60, costab04); 271 + t84 = t61 + t62; t107 = MPC_DCT32_MUL(t61 - t62, costab28); 272 + t85 = t63 + t64; t108 = MPC_DCT32_MUL(t63 - t64, costab12); 273 + t86 = t65 + t66; t109 = MPC_DCT32_MUL(t65 - t66, costab20); 274 + 275 + t113 = t69 + t70; 276 + t114 = t71 + t72; 277 + 278 + /* 0 */ v[48] = -MPC_DCT32_SHIFT(t113 + t114); 279 + /* 16 */ v[32] = -(v[ 0] = MPC_DCT32_SHIFT(MPC_DCT32_MUL(t113 - t114, costab16))); 280 + 281 + t115 = t73 + t74; 282 + t116 = t75 + t76; 283 + 284 + t32 = t115 + t116; 285 + 286 + /* 1 */ v[49] = v[47] = -MPC_DCT32_SHIFT(t32); 287 + 288 + t118 = t78 + t79; 289 + t119 = t80 + t81; 290 + 291 + t58 = t118 + t119; 292 + 293 + /* 2 */ v[50] = v[46] = -MPC_DCT32_SHIFT(t58); 294 + 295 + t121 = t83 + t84; 296 + t122 = t85 + t86; 297 + 298 + t67 = t121 + t122; 299 + 300 + t49 = (t67 * 2) - t32; 301 + 302 + /* 3 */ v[51] = v[45] = -MPC_DCT32_SHIFT(t49); 303 + 304 + t125 = t89 + t90; 305 + t126 = t91 + t92; 306 + 307 + t93 = t125 + t126; 308 + 309 + /* 4 */ v[52] = v[44] = -MPC_DCT32_SHIFT(t93); 310 + 311 + t128 = t94 + t95; 312 + t129 = t96 + t97; 313 + 314 + t98 = t128 + t129; 315 + 316 + t68 = (t98 * 2) - t49; 317 + 318 + /* 5 */ v[53] = v[43] = -MPC_DCT32_SHIFT(t68); 319 + 320 + t132 = t100 + t101; 321 + t133 = t102 + t103; 322 + 323 + t104 = t132 + t133; 324 + 325 + t82 = (t104 * 2) - t58; 326 + 327 + /* 6 */ v[54] = v[42] = -MPC_DCT32_SHIFT(t82); 328 + 329 + t136 = t106 + t107; 330 + t137 = t108 + t109; 331 + 332 + t110 = t136 + t137; 333 + 334 + t87 = (t110 * 2) - t67; 335 + 336 + t77 = (t87 * 2) - t68; 337 + 338 + /* 7 */ v[55] = v[41] = -MPC_DCT32_SHIFT(t77); 339 + 340 + t141 = MPC_DCT32_MUL(t69 - t70, costab08); 341 + t142 = MPC_DCT32_MUL(t71 - t72, costab24); 342 + t143 = t141 + t142; 343 + 344 + /* 8 */ v[56] = v[40] = -MPC_DCT32_SHIFT(t143); 345 + /* 24 */ v[24] = -(v[ 8] = MPC_DCT32_SHIFT((MPC_DCT32_MUL(t141 - t142, costab16) * 2) - t143)); 346 + 347 + t144 = MPC_DCT32_MUL(t73 - t74, costab08); 348 + t145 = MPC_DCT32_MUL(t75 - t76, costab24); 349 + t146 = t144 + t145; 350 + 351 + t88 = (t146 * 2) - t77; 352 + 353 + /* 9 */ v[57] = v[39] = -MPC_DCT32_SHIFT(t88); 354 + 355 + t148 = MPC_DCT32_MUL(t78 - t79, costab08); 356 + t149 = MPC_DCT32_MUL(t80 - t81, costab24); 357 + t150 = t148 + t149; 358 + 359 + t105 = (t150 * 2) - t82; 360 + 361 + /* 10 */ v[58] = v[38] = -MPC_DCT32_SHIFT(t105); 362 + 363 + t152 = MPC_DCT32_MUL(t83 - t84, costab08); 364 + t153 = MPC_DCT32_MUL(t85 - t86, costab24); 365 + t154 = t152 + t153; 366 + 367 + t111 = (t154 * 2) - t87; 368 + 369 + t99 = (t111 * 2) - t88; 370 + 371 + /* 11 */ v[59] = v[37] = -MPC_DCT32_SHIFT(t99); 372 + 373 + t157 = MPC_DCT32_MUL(t89 - t90, costab08); 374 + t158 = MPC_DCT32_MUL(t91 - t92, costab24); 375 + t159 = t157 + t158; 376 + 377 + t127 = (t159 * 2) - t93; 378 + 379 + /* 12 */ v[60] = v[36] = -MPC_DCT32_SHIFT(t127); 416 380 417 - V[32] = -(V[ 0] = MPC_V_PRESHIFT(V[ 0])); 418 - V[31] = -(V[ 1] = MPC_V_PRESHIFT(V[ 1])); 419 - V[30] = -(V[ 2] = MPC_V_PRESHIFT(V[ 2])); 420 - V[29] = -(V[ 3] = MPC_V_PRESHIFT(V[ 3])); 421 - V[28] = -(V[ 4] = MPC_V_PRESHIFT(V[ 4])); 422 - V[27] = -(V[ 5] = MPC_V_PRESHIFT(V[ 5])); 423 - V[26] = -(V[ 6] = MPC_V_PRESHIFT(V[ 6])); 424 - V[25] = -(V[ 7] = MPC_V_PRESHIFT(V[ 7])); 425 - V[24] = -(V[ 8] = MPC_V_PRESHIFT(V[ 8])); 426 - V[23] = -(V[ 9] = MPC_V_PRESHIFT(V[ 9])); 427 - V[22] = -(V[10] = MPC_V_PRESHIFT(V[10])); 428 - V[21] = -(V[11] = MPC_V_PRESHIFT(V[11])); 429 - V[20] = -(V[12] = MPC_V_PRESHIFT(V[12])); 430 - V[19] = -(V[13] = MPC_V_PRESHIFT(V[13])); 431 - V[18] = -(V[14] = MPC_V_PRESHIFT(V[14])); 432 - V[17] = -(V[15] = MPC_V_PRESHIFT(V[15])); 433 - // 16 adds, 16 shifts (OPTIMIZE_FOR_SPEED only) 381 + t160 = (MPC_DCT32_MUL(t125 - t126, costab16) * 2) - t127; 434 382 435 - V[63] = (V[33] = MPC_V_PRESHIFT(V[33])); 436 - V[62] = (V[34] = MPC_V_PRESHIFT(V[34])); 437 - V[61] = (V[35] = MPC_V_PRESHIFT(V[35])); 438 - V[60] = (V[36] = MPC_V_PRESHIFT(V[36])); 439 - V[59] = (V[37] = MPC_V_PRESHIFT(V[37])); 440 - V[58] = (V[38] = MPC_V_PRESHIFT(V[38])); 441 - V[57] = (V[39] = MPC_V_PRESHIFT(V[39])); 442 - V[56] = (V[40] = MPC_V_PRESHIFT(V[40])); 443 - V[55] = (V[41] = MPC_V_PRESHIFT(V[41])); 444 - V[54] = (V[42] = MPC_V_PRESHIFT(V[42])); 445 - V[53] = (V[43] = MPC_V_PRESHIFT(V[43])); 446 - V[52] = (V[44] = MPC_V_PRESHIFT(V[44])); 447 - V[51] = (V[45] = MPC_V_PRESHIFT(V[45])); 448 - V[50] = (V[46] = MPC_V_PRESHIFT(V[46])); 449 - V[49] = (V[47] = MPC_V_PRESHIFT(V[47])); 450 - V[48] = (V[48] = MPC_V_PRESHIFT(V[48])); 451 - // 16 adds, 16 shifts (OPTIMIZE_FOR_SPEED only) 452 - 453 - // OPTIMIZE_FOR_SPEED total: 143 adds, 107 subs, 80 muls, 112 shifts 454 - // total: 111 adds, 107 subs, 80 muls, 80 shifts 383 + /* 20 */ v[28] = -(v[ 4] = MPC_DCT32_SHIFT(t160)); 384 + /* 28 */ v[20] = -(v[12] = MPC_DCT32_SHIFT((((MPC_DCT32_MUL(t157 - t158, costab16) * 2) - t159) * 2) - t160)); 385 + 386 + t161 = MPC_DCT32_MUL(t94 - t95, costab08); 387 + t162 = MPC_DCT32_MUL(t96 - t97, costab24); 388 + t163 = t161 + t162; 389 + 390 + t130 = (t163 * 2) - t98; 391 + 392 + t112 = (t130 * 2) - t99; 393 + 394 + /* 13 */ v[61] = v[35] = -MPC_DCT32_SHIFT(t112); 395 + 396 + t164 = (MPC_DCT32_MUL(t128 - t129, costab16) * 2) - t130; 397 + 398 + t166 = MPC_DCT32_MUL(t100 - t101, costab08); 399 + t167 = MPC_DCT32_MUL(t102 - t103, costab24); 400 + t168 = t166 + t167; 401 + 402 + t134 = (t168 * 2) - t104; 403 + 404 + t120 = (t134 * 2) - t105; 405 + 406 + /* 14 */ v[62] = v[34] = -MPC_DCT32_SHIFT(t120); 407 + 408 + t135 = (MPC_DCT32_MUL(t118 - t119, costab16) * 2) - t120; 409 + 410 + /* 18 */ v[30] = -(v[ 2] = MPC_DCT32_SHIFT(t135)); 411 + 412 + t169 = (MPC_DCT32_MUL(t132 - t133, costab16) * 2) - t134; 413 + 414 + t151 = (t169 * 2) - t135; 415 + 416 + /* 22 */ v[26] = -(v[ 6] = MPC_DCT32_SHIFT(t151)); 417 + 418 + t170 = (((MPC_DCT32_MUL(t148 - t149, costab16) * 2) - t150) * 2) - t151; 419 + 420 + /* 26 */ v[22] = -(v[10] = MPC_DCT32_SHIFT(t170)); 421 + /* 30 */ v[18] = -(v[14] = MPC_DCT32_SHIFT((((((MPC_DCT32_MUL(t166 - t167, costab16) * 2) - t168) * 2) - t169) * 2) - t170)); 422 + 423 + t171 = MPC_DCT32_MUL(t106 - t107, costab08); 424 + t172 = MPC_DCT32_MUL(t108 - t109, costab24); 425 + t173 = t171 + t172; 426 + 427 + t138 = (t173 * 2) - t110; 428 + 429 + t123 = (t138 * 2) - t111; 430 + 431 + t139 = (MPC_DCT32_MUL(t121 - t122, costab16) * 2) - t123; 432 + 433 + t117 = (t123 * 2) - t112; 434 + 435 + /* 15 */ v[63] = v[33] =-MPC_DCT32_SHIFT(t117); 436 + 437 + t124 = (MPC_DCT32_MUL(t115 - t116, costab16) * 2) - t117; 438 + 439 + /* 17 */ v[31] = -(v[ 1] = MPC_DCT32_SHIFT(t124)); 440 + 441 + t131 = (t139 * 2) - t124; 442 + 443 + /* 19 */ v[29] = -(v[ 3] = MPC_DCT32_SHIFT(t131)); 444 + 445 + t140 = (t164 * 2) - t131; 446 + 447 + /* 21 */ v[27] = -(v[ 5] = MPC_DCT32_SHIFT(t140)); 448 + 449 + t174 = (MPC_DCT32_MUL(t136 - t137, costab16) * 2) - t138; 450 + 451 + t155 = (t174 * 2) - t139; 452 + 453 + t147 = (t155 * 2) - t140; 454 + 455 + /* 23 */ v[25] = -(v[ 7] = MPC_DCT32_SHIFT(t147)); 456 + 457 + t156 = (((MPC_DCT32_MUL(t144 - t145, costab16) * 2) - t146) * 2) - t147; 458 + 459 + /* 25 */ v[23] = -(v[ 9] = MPC_DCT32_SHIFT(t156)); 460 + 461 + t175 = (((MPC_DCT32_MUL(t152 - t153, costab16) * 2) - t154) * 2) - t155; 462 + 463 + t165 = (t175 * 2) - t156; 464 + 465 + /* 27 */ v[21] = -(v[11] = MPC_DCT32_SHIFT(t165)); 466 + 467 + t176 = (((((MPC_DCT32_MUL(t161 - t162, costab16) * 2) - t163) * 2) - t164) * 2) - t165; 468 + 469 + /* 29 */ v[19] = -(v[13] = MPC_DCT32_SHIFT(t176)); 470 + /* 31 */ v[17] = -(v[15] = MPC_DCT32_SHIFT((((((((MPC_DCT32_MUL(t171 - t172, costab16) * 2) - t173) * 2) - t174) * 2) - t175) * 2) - t176)); 455 471 } 456 472 457 473 #if defined(CPU_ARM) ··· 465 481 const MPC_SAMPLE_FORMAT * V, 466 482 const MPC_SAMPLE_FORMAT * D) 467 483 { 468 - mpc_int32_t k; 484 + mpc_int32_t k; 469 485 470 - #if defined(OPTIMIZE_FOR_SPEED) 471 - // 32=32x32-multiply (FIXED_POINT) 472 - for ( k = 0; k < 32; k++, D += 16, V++ ) 473 - { 474 - *Data = V[ 0]*D[ 0] + V[ 96]*D[ 1] + V[128]*D[ 2] + V[224]*D[ 3] 475 - + V[256]*D[ 4] + V[352]*D[ 5] + V[384]*D[ 6] + V[480]*D[ 7] 476 - + V[512]*D[ 8] + V[608]*D[ 9] + V[640]*D[10] + V[736]*D[11] 477 - + V[768]*D[12] + V[864]*D[13] + V[896]*D[14] + V[992]*D[15]; 478 - *Data >>= 1; // post shift to compensate for pre-shifting 479 - Data += 1; 480 - // total: 32 * (16 muls, 15 adds) 481 - } 482 - #else 483 - #if defined(CPU_COLDFIRE) 484 - // 64=32x32-multiply assembler for Coldfire 485 - for ( k = 0; k < 32; k++, D += 16, V++ ) 486 - { 487 - asm volatile ( 488 - "movem.l (%[D]), %%d0-%%d3 \n\t" 489 - "move.l (%[V]), %%a5 \n\t" 490 - "mac.l %%d0, %%a5, (96*4, %[V]), %%a5, %%acc0 \n\t" 491 - "mac.l %%d1, %%a5, (128*4, %[V]), %%a5, %%acc0\n\t" 492 - "mac.l %%d2, %%a5, (224*4, %[V]), %%a5, %%acc0\n\t" 493 - "mac.l %%d3, %%a5, (256*4, %[V]), %%a5, %%acc0\n\t" 494 - "movem.l (4*4, %[D]), %%d0-%%d3 \n\t" 495 - "mac.l %%d0, %%a5, (352*4, %[V]), %%a5, %%acc0\n\t" 496 - "mac.l %%d1, %%a5, (384*4, %[V]), %%a5, %%acc0\n\t" 497 - "mac.l %%d2, %%a5, (480*4, %[V]), %%a5, %%acc0\n\t" 498 - "mac.l %%d3, %%a5, (512*4, %[V]), %%a5, %%acc0\n\t" 499 - "movem.l (8*4, %[D]), %%d0-%%d3 \n\t" 500 - "mac.l %%d0, %%a5, (608*4, %[V]), %%a5, %%acc0\n\t" 501 - "mac.l %%d1, %%a5, (640*4, %[V]), %%a5, %%acc0\n\t" 502 - "mac.l %%d2, %%a5, (736*4, %[V]), %%a5, %%acc0\n\t" 503 - "mac.l %%d3, %%a5, (768*4, %[V]), %%a5, %%acc0\n\t" 504 - "movem.l (12*4, %[D]), %%d0-%%d3 \n\t" 505 - "mac.l %%d0, %%a5, (864*4, %[V]), %%a5, %%acc0\n\t" 506 - "mac.l %%d1, %%a5, (896*4, %[V]), %%a5, %%acc0\n\t" 507 - "mac.l %%d2, %%a5, (992*4, %[V]), %%a5, %%acc0\n\t" 508 - "mac.l %%d3, %%a5, %%acc0 \n\t" 509 - "movclr.l %%acc0, %%d0 \n\t" 510 - "lsl.l #1, %%d0 \n\t" 511 - "move.l %%d0, (%[Data])+ \n" 512 - : [Data] "+a" (Data) 513 - : [V] "a" (V), [D] "a" (D) 514 - : "d0", "d1", "d2", "d3", "a5"); 515 - } 516 - #else 517 - // 64=64x64-multiply (FIXED_POINT) or float=float*float (!FIXED_POINT) in C 518 - for ( k = 0; k < 32; k++, D += 16, V++ ) 519 - { 520 - *Data = MPC_MULTIPLY_EX(V[ 0],D[ 0],30) + MPC_MULTIPLY_EX(V[ 96],D[ 1],30) 521 - + MPC_MULTIPLY_EX(V[128],D[ 2],30) + MPC_MULTIPLY_EX(V[224],D[ 3],30) 522 - + MPC_MULTIPLY_EX(V[256],D[ 4],30) + MPC_MULTIPLY_EX(V[352],D[ 5],30) 523 - + MPC_MULTIPLY_EX(V[384],D[ 6],30) + MPC_MULTIPLY_EX(V[480],D[ 7],30) 524 - + MPC_MULTIPLY_EX(V[512],D[ 8],30) + MPC_MULTIPLY_EX(V[608],D[ 9],30) 525 - + MPC_MULTIPLY_EX(V[640],D[10],30) + MPC_MULTIPLY_EX(V[736],D[11],30) 526 - + MPC_MULTIPLY_EX(V[768],D[12],30) + MPC_MULTIPLY_EX(V[864],D[13],30) 527 - + MPC_MULTIPLY_EX(V[896],D[14],30) + MPC_MULTIPLY_EX(V[992],D[15],30); 528 - Data += 1; 529 - // total: 16 muls, 15 adds, 16 shifts 530 - } 531 - #endif 532 - #endif 486 + #if defined(CPU_COLDFIRE) 487 + // 64=32x32-multiply assembler for Coldfire 488 + for ( k = 0; k < 32; k++, D += 16, V++ ) 489 + { 490 + asm volatile ( 491 + "movem.l (%[D]), %%d0-%%d3 \n\t" 492 + "move.l (%[V]), %%a5 \n\t" 493 + "mac.l %%d0, %%a5, (96*4, %[V]), %%a5, %%acc0 \n\t" 494 + "mac.l %%d1, %%a5, (128*4, %[V]), %%a5, %%acc0\n\t" 495 + "mac.l %%d2, %%a5, (224*4, %[V]), %%a5, %%acc0\n\t" 496 + "mac.l %%d3, %%a5, (256*4, %[V]), %%a5, %%acc0\n\t" 497 + "movem.l (4*4, %[D]), %%d0-%%d3 \n\t" 498 + "mac.l %%d0, %%a5, (352*4, %[V]), %%a5, %%acc0\n\t" 499 + "mac.l %%d1, %%a5, (384*4, %[V]), %%a5, %%acc0\n\t" 500 + "mac.l %%d2, %%a5, (480*4, %[V]), %%a5, %%acc0\n\t" 501 + "mac.l %%d3, %%a5, (512*4, %[V]), %%a5, %%acc0\n\t" 502 + "movem.l (8*4, %[D]), %%d0-%%d3 \n\t" 503 + "mac.l %%d0, %%a5, (608*4, %[V]), %%a5, %%acc0\n\t" 504 + "mac.l %%d1, %%a5, (640*4, %[V]), %%a5, %%acc0\n\t" 505 + "mac.l %%d2, %%a5, (736*4, %[V]), %%a5, %%acc0\n\t" 506 + "mac.l %%d3, %%a5, (768*4, %[V]), %%a5, %%acc0\n\t" 507 + "movem.l (12*4, %[D]), %%d0-%%d3 \n\t" 508 + "mac.l %%d0, %%a5, (864*4, %[V]), %%a5, %%acc0\n\t" 509 + "mac.l %%d1, %%a5, (896*4, %[V]), %%a5, %%acc0\n\t" 510 + "mac.l %%d2, %%a5, (992*4, %[V]), %%a5, %%acc0\n\t" 511 + "mac.l %%d3, %%a5, %%acc0 \n\t" 512 + "movclr.l %%acc0, %%d0 \n\t" 513 + "lsl.l #1, %%d0 \n\t" 514 + "move.l %%d0, (%[Data])+ \n" 515 + : [Data] "+a" (Data) 516 + : [V] "a" (V), [D] "a" (D) 517 + : "d0", "d1", "d2", "d3", "a5"); 518 + } 519 + #else 520 + // 64=64x64-multiply (FIXED_POINT) or float=float*float (!FIXED_POINT) in C 521 + for ( k = 0; k < 32; k++, D += 16, V++ ) 522 + { 523 + *Data = MPC_MULTIPLY_EX(V[ 0],D[ 0],30) + MPC_MULTIPLY_EX(V[ 96],D[ 1],30) 524 + + MPC_MULTIPLY_EX(V[128],D[ 2],30) + MPC_MULTIPLY_EX(V[224],D[ 3],30) 525 + + MPC_MULTIPLY_EX(V[256],D[ 4],30) + MPC_MULTIPLY_EX(V[352],D[ 5],30) 526 + + MPC_MULTIPLY_EX(V[384],D[ 6],30) + MPC_MULTIPLY_EX(V[480],D[ 7],30) 527 + + MPC_MULTIPLY_EX(V[512],D[ 8],30) + MPC_MULTIPLY_EX(V[608],D[ 9],30) 528 + + MPC_MULTIPLY_EX(V[640],D[10],30) + MPC_MULTIPLY_EX(V[736],D[11],30) 529 + + MPC_MULTIPLY_EX(V[768],D[12],30) + MPC_MULTIPLY_EX(V[864],D[13],30) 530 + + MPC_MULTIPLY_EX(V[896],D[14],30) + MPC_MULTIPLY_EX(V[992],D[15],30); 531 + Data += 1; 532 + // total: 16 muls, 15 adds, 16 shifts 533 + } 534 + #endif /* COLDFIRE */ 533 535 } 534 536 #endif /* CPU_ARM */ 535 537 ··· 543 545 for ( n = 0; n < 36; n++, Y += 32, OutData += 32 ) 544 546 { 545 547 V -= 64; 546 - mpc_calculate_new_V ( Y, V ); 548 + mpc_dct32(Y, V); 547 549 mpc_decoder_windowing_D( OutData, V, Di_opt ); 548 550 } 549 551 }

+295 -363

apps/codecs/libmusepack/synth_filter_arm.S

··· 1 - /*************************************************************************** 2 - * __________ __ ___. 3 - * Open \______ \ ____ ____ | | _\_ |__ _______ ___ 4 - * Source | _// _ \_/ ___\| |/ /| __ \ / _ \ \/ / 5 - * Jukebox | | ( <_> ) \___| < | \_\ ( <_> > < < 6 - * Firmware |____|_ /\____/ \___ >__|_ \|___ /\____/__/\_ \ 7 - * \/ \/ \/ \/ \/ 8 - * $Id$ 9 - * 10 - * Copyright (C) 2008 by Andree Buschmann 11 - * 12 - * This program is free software; you can redistribute it and/or 13 - * modify it under the terms of the GNU General Public License 14 - * as published by the Free Software Foundation; either version 2 15 - * of the License, or (at your option) any later version. 16 - * 17 - * This software is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY 18 - * KIND, either express or implied. 19 - * 20 - ****************************************************************************/ 21 - 22 - #include "mpc_config.h" 23 - 24 - .section .text, "ax", %progbits 25 - 26 - #if defined(OPTIMIZE_FOR_SPEED) 27 - /**************************************************************************** 28 - * void mpc_decoder_windowing_D(...) 29 - * 30 - * 2nd step within synthesis filter. Does the dewindowing. 31 - * 32=32x32 multiplies (OPTIMIZE_FOR_SPEED) 32 - * Uses pre-shifted V[] and D[] values. D[] will always be the second operand 33 - * of mul/mla to achieve higher speed as D[] has lower amplitude than V[]. 34 - ****************************************************************************/ 35 - .align 2 36 - .global mpc_decoder_windowing_D 37 - .type mpc_decoder_windowing_D, %function 38 - mpc_decoder_windowing_D: 39 - /* r0 = Data[] */ 40 - /* r1 = V[] */ 41 - /* r2 = D[] */ 42 - /* lr = counter */ 43 - 44 - stmfd sp!, {r4-r11, lr} 45 - 46 - mov lr, #32 47 - .loop32: 48 - ldmia r2!, { r3-r10 } /* load D[00..07] */ 49 - ldr r11, [r1] /* 0 */ 50 - mul r12, r11, r3 51 - ldr r11, [r1, #96*4] /* 1 */ 52 - mla r12, r11, r4, r12 53 - ldr r11, [r1, #128*4] /* 2 */ 54 - mla r12, r11, r5, r12 55 - ldr r11, [r1, #224*4] /* 3 */ 56 - mla r12, r11, r6, r12 57 - ldr r11, [r1, #256*4] /* 4 */ 58 - mla r12, r11, r7, r12 59 - ldr r11, [r1, #352*4] /* 5 */ 60 - mla r12, r11, r8, r12 61 - ldr r11, [r1, #384*4] /* 6 */ 62 - mla r12, r11, r9, r12 63 - ldr r11, [r1, #480*4] /* 7 */ 64 - mla r12, r11, r10, r12 65 - ldmia r2!, { r3-r10 } /* load D[08..15] */ 66 - ldr r11, [r1, #512*4] /* 8 */ 67 - mla r12, r11, r3, r12 68 - ldr r11, [r1, #608*4] /* 9 */ 69 - mla r12, r11, r4, r12 70 - ldr r11, [r1, #640*4] /* 10 */ 71 - mla r12, r11, r5, r12 72 - ldr r11, [r1, #736*4] /* 11 */ 73 - mla r12, r11, r6, r12 74 - ldr r11, [r1, #768*4] /* 12 */ 75 - mla r12, r11, r7, r12 76 - ldr r11, [r1, #864*4] /* 13 */ 77 - mla r12, r11, r8, r12 78 - ldr r11, [r1, #896*4] /* 14 */ 79 - mla r12, r11, r9, r12 80 - ldr r11, [r1, #992*4] /* 15 */ 81 - mla r12, r11, r10, r12 82 - mov r12, r12, asr #1 /* post shift to compensate for pre-shifting */ 83 - str r12, [r0], #4 /* store Data */ 84 - add r1, r1, #4 /* V++ */ 85 - 86 - subs lr, lr, #1 87 - bgt .loop32 88 - 89 - ldmfd sp!, {r4-r11, pc} 90 - .mpc_dewindowing_end: 91 - .size mpc_decoder_windowing_D,.mpc_dewindowing_end-mpc_decoder_windowing_D 92 - #else 93 - /**************************************************************************** 94 - * void mpc_decoder_windowing_D(...) 95 - * 96 - * 2nd step within synthesis filter. Does the dewindowing. 97 - * 64=32x32 multiplies 98 - * Uses un-shifted D[]-values. D[] will always be the second operand of 99 - * smull/smlal to achieve higher speed as D[] has lower amplitude than V[]. 100 - ****************************************************************************/ 101 - .align 2 102 - .global mpc_decoder_windowing_D 103 - .type mpc_decoder_windowing_D, %function 104 - #if 0 105 - mpc_decoder_windowing_D: 106 - /* r0 = Data[] */ 107 - /* r1 = V[] */ 108 - /* r2 = D[] */ 109 - /* lr = counter */ 110 - /************************************************************************ 111 - * Reference implementation. 112 - ***********************************************************************/ 113 - stmfd sp!, {r4-r8, lr} 114 - 115 - mov lr, #32 116 - .loop32: 117 - ldmia r2!, { r3-r6 } /* load D[00..03] */ 118 - ldr r7, [r1] /* 0 */ 119 - smull r8, r12, r7, r3 120 - ldr r7, [r1, #96*4] /* 1 */ 121 - smlal r8, r12, r7, r4 122 - ldr r7, [r1, #128*4] /* 2 */ 123 - smlal r8, r12, r7, r5 124 - ldr r7, [r1, #224*4] /* 3 */ 125 - smlal r8, r12, r7, r6 126 - ldmia r2!, { r3-r6 } /* load D[04..07] */ 127 - ldr r7, [r1, #256*4] /* 4 */ 128 - smlal r8, r12, r7, r3 129 - ldr r7, [r1, #352*4] /* 5 */ 130 - smlal r8, r12, r7, r4 131 - ldr r7, [r1, #384*4] /* 6 */ 132 - smlal r8, r12, r7, r5 133 - ldr r7, [r1, #480*4] /* 7 */ 134 - smlal r8, r12, r7, r6 135 - ldmia r2!, { r3-r6 } /* load D[08..11] */ 136 - ldr r7, [r1, #512*4] /* 8 */ 137 - smlal r8, r12, r7, r3 138 - ldr r7, [r1, #608*4] /* 9 */ 139 - smlal r8, r12, r7, r4 140 - ldr r7, [r1, #640*4] /* 10 */ 141 - smlal r8, r12, r7, r5 142 - ldr r7, [r1, #736*4] /* 11 */ 143 - smlal r8, r12, r7, r6 144 - ldmia r2!, { r3-r6 } /* load D[12..15] */ 145 - ldr r7, [r1, #768*4] /* 12 */ 146 - smlal r8, r12, r7, r3 147 - ldr r7, [r1, #864*4] /* 13 */ 148 - smlal r8, r12, r7, r4 149 - ldr r7, [r1, #896*4] /* 14 */ 150 - smlal r8, r12, r7, r5 151 - ldr r7, [r1, #992*4] /* 15 */ 152 - smlal r8, r12, r7, r6 153 - mov r8, r8, lsr #16 154 - orr r8, r8, r12, lsl #16 /* (lo>>16) || (hi<<16) */ 155 - str r8, [r0], #4 /* store Data */ 156 - add r1, r1, #4 /* V++ */ 157 - 158 - subs lr, lr, #1 159 - bgt .loop32 160 - 161 - ldmfd sp!, {r4-r8, pc} 162 - #else 163 - mpc_decoder_windowing_D: 164 - /* r0 = Data[] */ 165 - /* r1 = V[] */ 166 - /* r2 = D[] */ 167 - /* lr = counter */ 168 - /************************************************************************ 169 - * Further speed up through making use of symmetries within D[]-window. 170 - * The row V[00] can be extracted as it has symmetries within this single 171 - * row. 8 smull/mlal and 8 ldr's can be saved at the cost of 2 add's. 172 - * The rows V[01..15] are symmetric to V[31..17]. 15 x 16 ldr's can be 173 - * saved at the cost of 15 x 4 + 1 add's. 174 - * The row V[16] can be extracted as it has symmetries within this single 175 - * row. 8 smull/mlal and 8 ldr's can be saved. 176 - ***********************************************************************/ 177 - stmfd sp!, {r4-r11, lr} 178 - 179 - /****************************************** 180 - * row 0 with internal symmetry 181 - *****************************************/ 182 - add r2, r2, #4 /* D+=1, r2 = D[01] as D[00] = 0 */ 183 - ldmia r2!, { r3-r6 } /* load D[01..04] */ 184 - ldr r7 , [r1, #96*4] /* 1 */ 185 - ldr r10, [r1, #992*4] /* 15 */ 186 - rsb r10, r10, r7 /* V[01] - V[15] */ 187 - smull r8, r9, r10, r3 188 - ldr r7 , [r1, #128*4] /* 2 */ 189 - ldr r10, [r1, #896*4] /* 14 */ 190 - add r10, r10, r7 /* V[02] + V[14] */ 191 - smlal r8, r9, r10, r4 192 - ldr r7 , [r1, #224*4] /* 3 */ 193 - ldr r10, [r1, #864*4] /* 13 */ 194 - rsb r10, r10, r7 /* V[03] - V[13] */ 195 - smlal r8, r9, r10, r5 196 - ldr r7 , [r1, #256*4] /* 4 */ 197 - ldr r10, [r1, #768*4] /* 12 */ 198 - add r10, r10, r7 /* V[04] + V[12] */ 199 - smlal r8, r9, r10, r6 200 - ldmia r2!, { r3-r6 } /* load D[05..08] */ 201 - ldr r7 , [r1, #352*4] /* 5 */ 202 - ldr r10, [r1, #736*4] /* 11 */ 203 - rsb r10, r10, r7 /* V[05] - V[11] */ 204 - smlal r8, r9, r10, r3 205 - ldr r7 , [r1, #384*4] /* 6 */ 206 - ldr r10, [r1, #640*4] /* 10 */ 207 - add r10, r10, r7 /* V[06] + V[10] */ 208 - smlal r8, r9, r10, r4 209 - ldr r7 , [r1, #480*4] /* 7 */ 210 - ldr r10, [r1, #608*4] /* 9 */ 211 - rsb r10, r10, r7 /* V[07] - V[09] */ 212 - smlal r8, r9, r10, r5 213 - ldr r10, [r1, #512*4] /* 8 */ 214 - smlal r8, r9, r10, r6 215 - mov r8, r8, lsr #16 216 - orr r8, r8, r9, lsl #16 /* (lo>>16) || (hi<<16) */ 217 - str r8, [r0], #4 /* store Data */ 218 - add r1, r1, #4 /* V+=1, r1 = V[01] */ 219 - add r2, r2, #7*4 /* D+=7, r2 = D[16] */ 220 - 221 - /****************************************** 222 - * rows 01..15 are symmetrc to rows 31..17 223 - * r8 = lo, r9 = hi of 01..15 224 - * r1 = V[01..15] 225 - * r10 = lo, r11 = hi of 31..17 226 - * r12 = V[31..16] 227 - *****************************************/ 228 - mov lr, #15 229 - add r12, r1, #30*4 /* r12 = V[31] */ 230 - .loop15: 231 - ldmia r2!, { r3-r6 } /* load D[00..03] */ 232 - ldr r7, [r12, #768*4] /* 12 */ 233 - smull r10, r11, r7, r6 234 - ldr r7, [r12, #864*4] /* 13 */ 235 - smlal r10, r11, r7, r5 236 - ldr r7, [r12, #896*4] /* 14 */ 237 - smlal r10, r11, r7, r4 238 - ldr r7, [r12, #992*4] /* 15 */ 239 - smlal r10, r11, r7, r3 240 - ldr r7, [r1] /* 0 */ 241 - smull r8, r9, r7, r3 242 - ldr r7, [r1, #96*4] /* 1 */ 243 - smlal r8, r9, r7, r4 244 - ldr r7, [r1, #128*4] /* 2 */ 245 - smlal r8, r9, r7, r5 246 - ldr r7, [r1, #224*4] /* 3 */ 247 - smlal r8, r9, r7, r6 248 - ldmia r2!, { r3-r6 } /* load D[04..07] */ 249 - ldr r7, [r1, #256*4] /* 4 */ 250 - smlal r8, r9, r7, r3 251 - ldr r7, [r1, #352*4] /* 5 */ 252 - smlal r8, r9, r7, r4 253 - ldr r7, [r1, #384*4] /* 6 */ 254 - smlal r8, r9, r7, r5 255 - ldr r7, [r1, #480*4] /* 7 */ 256 - smlal r8, r9, r7, r6 257 - ldr r7, [r12, #512*4] /* 8 */ 258 - smlal r10, r11, r7, r6 259 - ldr r7, [r12, #608*4] /* 9 */ 260 - smlal r10, r11, r7, r5 261 - ldr r7, [r12, #640*4] /* 10 */ 262 - smlal r10, r11, r7, r4 263 - ldr r7, [r12, #736*4] /* 11 */ 264 - smlal r10, r11, r7, r3 265 - ldmia r2!, { r3-r6 } /* load D[08..11] */ 266 - ldr r7, [r12, #256*4] /* 4 */ 267 - smlal r10, r11, r7, r6 268 - ldr r7, [r12, #352*4] /* 5 */ 269 - smlal r10, r11, r7, r5 270 - ldr r7, [r12, #384*4] /* 6 */ 271 - smlal r10, r11, r7, r4 272 - ldr r7, [r12, #480*4] /* 7 */ 273 - smlal r10, r11, r7, r3 274 - ldr r7, [r1, #512*4] /* 8 */ 275 - smlal r8, r9, r7, r3 276 - ldr r7, [r1, #608*4] /* 9 */ 277 - smlal r8, r9, r7, r4 278 - ldr r7, [r1, #640*4] /* 10 */ 279 - smlal r8, r9, r7, r5 280 - ldr r7, [r1, #736*4] /* 11 */ 281 - smlal r8, r9, r7, r6 282 - ldmia r2!, { r3-r6 } /* load D[12..15] */ 283 - ldr r7, [r1, #768*4] /* 12 */ 284 - smlal r8, r9, r7, r3 285 - ldr r7, [r1, #864*4] /* 13 */ 286 - smlal r8, r9, r7, r4 287 - ldr r7, [r1, #896*4] /* 14 */ 288 - smlal r8, r9, r7, r5 289 - ldr r7, [r1, #992*4] /* 15 */ 290 - smlal r8, r9, r7, r6 291 - ldr r7, [r12] /* 0 */ 292 - smlal r10, r11, r7, r6 293 - ldr r7, [r12, #96*4] /* 1 */ 294 - smlal r10, r11, r7, r5 295 - ldr r7, [r12, #128*4] /* 2 */ 296 - smlal r10, r11, r7, r4 297 - ldr r7, [r12, #224*4] /* 3 */ 298 - smlal r10, r11, r7, r3 299 - /* store Data[01..15] */ 300 - mov r8, r8, lsr #16 301 - orr r8, r8, r9, lsl #16 /* (lo>>16) || (hi<<16) */ 302 - str r8, [r0] /* store Data */ 303 - /* store Data[31..17] */ 304 - add r0, r0, lr, asl #3 /* r0 = r0 + 2*lr [words] */ 305 - mov r10, r10, lsr #16 306 - orr r10, r10, r11, lsl #16 /* (lo>>16) || (hi<<16) */ 307 - rsb r10, r10, #0 /* r10 = -r10 */ 308 - str r10, [r0], #4 /* store Data */ 309 - sub r0, r0, lr, asl #3 /* r0 = r0 - 2*lr [words] */ 310 - /* correct adresses for next loop */ 311 - sub r12, r12, #4 /* r12 = V-- */ 312 - add r1, r1, #4 /* r1 = V++ */ 313 - /* next loop */ 314 - subs lr, lr, #1 315 - bgt .loop15 316 - 317 - /****************************************** 318 - * V[16] with internal symmetry 319 - *****************************************/ 320 - ldmia r2!, { r3-r6 } /* load D[00..03] */ 321 - ldr r7 , [r1] /* 0 */ 322 - ldr r10, [r1, #992*4] /* 15 */ 323 - rsb r10, r10, r7 /* V[00] - V[15] */ 324 - smull r8, r9, r10, r3 325 - ldr r7 , [r1, #96*4] /* 1 */ 326 - ldr r10, [r1, #896*4] /* 14 */ 327 - rsb r10, r10, r7 /* V[01] - V[14] */ 328 - smlal r8, r9, r10, r4 329 - ldr r7 , [r1, #128*4] /* 2 */ 330 - ldr r10, [r1, #864*4] /* 13 */ 331 - rsb r10, r10, r7 /* V[02] - V[13] */ 332 - smlal r8, r9, r10, r5 333 - ldr r7 , [r1, #224*4] /* 3 */ 334 - ldr r10, [r1, #768*4] /* 12 */ 335 - rsb r10, r10, r7 /* V[03] - V[12] */ 336 - smlal r8, r9, r10, r6 337 - ldmia r2!, { r3-r6 } /* load D[04..07] */ 338 - ldr r7 , [r1, #256*4] /* 4 */ 339 - ldr r10, [r1, #736*4] /* 11 */ 340 - rsb r10, r10, r7 /* V[04] - V[11] */ 341 - smlal r8, r9, r10, r3 342 - ldr r7 , [r1, #352*4] /* 5 */ 343 - ldr r10, [r1, #640*4] /* 10 */ 344 - rsb r10, r10, r7 /* V[05] - V[10] */ 345 - smlal r8, r9, r10, r4 346 - ldr r7 , [r1, #384*4] /* 6 */ 347 - ldr r10, [r1, #608*4] /* 9 */ 348 - rsb r10, r10, r7 /* V[06] - V[09] */ 349 - smlal r8, r9, r10, r5 350 - ldr r7 , [r1, #480*4] /* 7 */ 351 - ldr r10, [r1, #512*4] /* 8 */ 352 - rsb r10, r10, r7 /* V[07] - V[08] */ 353 - smlal r8, r9, r10, r6 354 - mov r8, r8, lsr #16 355 - orr r8, r8, r9, lsl #16 /* (lo>>16) || (hi<<16) */ 356 - str r8, [r0], #4 /* store Data */ 357 - add r1, r1, #4 /* V++ */ 358 - 359 - ldmfd sp!, {r4-r11, pc} 360 - #endif 361 - .mpc_dewindowing_end: 362 - .size mpc_decoder_windowing_D,.mpc_dewindowing_end-mpc_decoder_windowing_D 363 - #endif 1 + /*************************************************************************** 2 + * __________ __ ___. 3 + * Open \______ \ ____ ____ | | _\_ |__ _______ ___ 4 + * Source | _// _ \_/ ___\| |/ /| __ \ / _ \ \/ / 5 + * Jukebox | | ( <_> ) \___| < | \_\ ( <_> > < < 6 + * Firmware |____|_ /\____/ \___ >__|_ \|___ /\____/__/\_ \ 7 + * \/ \/ \/ \/ \/ 8 + * $Id$ 9 + * 10 + * Copyright (C) 2008 by Andree Buschmann 11 + * 12 + * This program is free software; you can redistribute it and/or 13 + * modify it under the terms of the GNU General Public License 14 + * as published by the Free Software Foundation; either version 2 15 + * of the License, or (at your option) any later version. 16 + * 17 + * This software is distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY 18 + * KIND, either express or implied. 19 + * 20 + ****************************************************************************/ 21 + 22 + #include "mpc_config.h" 23 + 24 + .section .text, "ax", %progbits 25 + 26 + /**************************************************************************** 27 + * void mpc_decoder_windowing_D(...) 28 + * 29 + * 2nd step within synthesis filter. Does the dewindowing. 30 + * 64=32x32 multiplies 31 + * Uses un-shifted D[]-values. D[] will always be the second operand of 32 + * smull/smlal to achieve higher speed as D[] has lower amplitude than V[]. 33 + ****************************************************************************/ 34 + .align 2 35 + .global mpc_decoder_windowing_D 36 + .type mpc_decoder_windowing_D, %function 37 + #if 0 38 + mpc_decoder_windowing_D: 39 + /* r0 = Data[] */ 40 + /* r1 = V[] */ 41 + /* r2 = D[] */ 42 + /* lr = counter */ 43 + /************************************************************************ 44 + * Reference implementation. 45 + ***********************************************************************/ 46 + stmfd sp!, {r4-r8, lr} 47 + 48 + mov lr, #32 49 + .loop32: 50 + ldmia r2!, { r3-r6 } /* load D[00..03] */ 51 + ldr r7, [r1] /* 0 */ 52 + smull r8, r12, r7, r3 53 + ldr r7, [r1, #96*4] /* 1 */ 54 + smlal r8, r12, r7, r4 55 + ldr r7, [r1, #128*4] /* 2 */ 56 + smlal r8, r12, r7, r5 57 + ldr r7, [r1, #224*4] /* 3 */ 58 + smlal r8, r12, r7, r6 59 + ldmia r2!, { r3-r6 } /* load D[04..07] */ 60 + ldr r7, [r1, #256*4] /* 4 */ 61 + smlal r8, r12, r7, r3 62 + ldr r7, [r1, #352*4] /* 5 */ 63 + smlal r8, r12, r7, r4 64 + ldr r7, [r1, #384*4] /* 6 */ 65 + smlal r8, r12, r7, r5 66 + ldr r7, [r1, #480*4] /* 7 */ 67 + smlal r8, r12, r7, r6 68 + ldmia r2!, { r3-r6 } /* load D[08..11] */ 69 + ldr r7, [r1, #512*4] /* 8 */ 70 + smlal r8, r12, r7, r3 71 + ldr r7, [r1, #608*4] /* 9 */ 72 + smlal r8, r12, r7, r4 73 + ldr r7, [r1, #640*4] /* 10 */ 74 + smlal r8, r12, r7, r5 75 + ldr r7, [r1, #736*4] /* 11 */ 76 + smlal r8, r12, r7, r6 77 + ldmia r2!, { r3-r6 } /* load D[12..15] */ 78 + ldr r7, [r1, #768*4] /* 12 */ 79 + smlal r8, r12, r7, r3 80 + ldr r7, [r1, #864*4] /* 13 */ 81 + smlal r8, r12, r7, r4 82 + ldr r7, [r1, #896*4] /* 14 */ 83 + smlal r8, r12, r7, r5 84 + ldr r7, [r1, #992*4] /* 15 */ 85 + smlal r8, r12, r7, r6 86 + mov r8, r8, lsr #16 87 + orr r8, r8, r12, lsl #16 /* (lo>>16) || (hi<<16) */ 88 + str r8, [r0], #4 /* store Data */ 89 + add r1, r1, #4 /* V++ */ 90 + 91 + subs lr, lr, #1 92 + bgt .loop32 93 + 94 + ldmfd sp!, {r4-r8, pc} 95 + #else 96 + mpc_decoder_windowing_D: 97 + /* r0 = Data[] */ 98 + /* r1 = V[] */ 99 + /* r2 = D[] */ 100 + /* lr = counter */ 101 + /************************************************************************ 102 + * Further speed up through making use of symmetries within D[]-window. 103 + * The row V[00] can be extracted as it has symmetries within this single 104 + * row. 8 smull/mlal and 8 ldr's can be saved at the cost of 2 add's. 105 + * The rows V[01..15] are symmetric to V[31..17]. 15 x 16 ldr's can be 106 + * saved at the cost of 15 x 4 + 1 add's. 107 + * The row V[16] can be extracted as it has symmetries within this single 108 + * row. 8 smull/mlal and 8 ldr's can be saved. 109 + ***********************************************************************/ 110 + stmfd sp!, {r4-r11, lr} 111 + 112 + /****************************************** 113 + * row 0 with internal symmetry 114 + *****************************************/ 115 + add r2, r2, #4 /* D+=1, r2 = D[01] as D[00] = 0 */ 116 + ldmia r2!, { r3-r6 } /* load D[01..04] */ 117 + ldr r7 , [r1, #96*4] /* 1 */ 118 + ldr r10, [r1, #992*4] /* 15 */ 119 + rsb r10, r10, r7 /* V[01] - V[15] */ 120 + smull r8, r9, r10, r3 121 + ldr r7 , [r1, #128*4] /* 2 */ 122 + ldr r10, [r1, #896*4] /* 14 */ 123 + add r10, r10, r7 /* V[02] + V[14] */ 124 + smlal r8, r9, r10, r4 125 + ldr r7 , [r1, #224*4] /* 3 */ 126 + ldr r10, [r1, #864*4] /* 13 */ 127 + rsb r10, r10, r7 /* V[03] - V[13] */ 128 + smlal r8, r9, r10, r5 129 + ldr r7 , [r1, #256*4] /* 4 */ 130 + ldr r10, [r1, #768*4] /* 12 */ 131 + add r10, r10, r7 /* V[04] + V[12] */ 132 + smlal r8, r9, r10, r6 133 + ldmia r2!, { r3-r6 } /* load D[05..08] */ 134 + ldr r7 , [r1, #352*4] /* 5 */ 135 + ldr r10, [r1, #736*4] /* 11 */ 136 + rsb r10, r10, r7 /* V[05] - V[11] */ 137 + smlal r8, r9, r10, r3 138 + ldr r7 , [r1, #384*4] /* 6 */ 139 + ldr r10, [r1, #640*4] /* 10 */ 140 + add r10, r10, r7 /* V[06] + V[10] */ 141 + smlal r8, r9, r10, r4 142 + ldr r7 , [r1, #480*4] /* 7 */ 143 + ldr r10, [r1, #608*4] /* 9 */ 144 + rsb r10, r10, r7 /* V[07] - V[09] */ 145 + smlal r8, r9, r10, r5 146 + ldr r10, [r1, #512*4] /* 8 */ 147 + smlal r8, r9, r10, r6 148 + mov r8, r8, lsr #16 149 + orr r8, r8, r9, lsl #16 /* (lo>>16) || (hi<<16) */ 150 + str r8, [r0], #4 /* store Data */ 151 + add r1, r1, #4 /* V+=1, r1 = V[01] */ 152 + add r2, r2, #7*4 /* D+=7, r2 = D[16] */ 153 + 154 + /****************************************** 155 + * rows 01..15 are symmetrc to rows 31..17 156 + * r8 = lo, r9 = hi of 01..15 157 + * r1 = V[01..15] 158 + * r10 = lo, r11 = hi of 31..17 159 + * r12 = V[31..16] 160 + *****************************************/ 161 + mov lr, #15 162 + add r12, r1, #30*4 /* r12 = V[31] */ 163 + .loop15: 164 + ldmia r2!, { r3-r6 } /* load D[00..03] */ 165 + ldr r7, [r12, #768*4] /* 12 */ 166 + smull r10, r11, r7, r6 167 + ldr r7, [r12, #864*4] /* 13 */ 168 + smlal r10, r11, r7, r5 169 + ldr r7, [r12, #896*4] /* 14 */ 170 + smlal r10, r11, r7, r4 171 + ldr r7, [r12, #992*4] /* 15 */ 172 + smlal r10, r11, r7, r3 173 + ldr r7, [r1] /* 0 */ 174 + smull r8, r9, r7, r3 175 + ldr r7, [r1, #96*4] /* 1 */ 176 + smlal r8, r9, r7, r4 177 + ldr r7, [r1, #128*4] /* 2 */ 178 + smlal r8, r9, r7, r5 179 + ldr r7, [r1, #224*4] /* 3 */ 180 + smlal r8, r9, r7, r6 181 + ldmia r2!, { r3-r6 } /* load D[04..07] */ 182 + ldr r7, [r1, #256*4] /* 4 */ 183 + smlal r8, r9, r7, r3 184 + ldr r7, [r1, #352*4] /* 5 */ 185 + smlal r8, r9, r7, r4 186 + ldr r7, [r1, #384*4] /* 6 */ 187 + smlal r8, r9, r7, r5 188 + ldr r7, [r1, #480*4] /* 7 */ 189 + smlal r8, r9, r7, r6 190 + ldr r7, [r12, #512*4] /* 8 */ 191 + smlal r10, r11, r7, r6 192 + ldr r7, [r12, #608*4] /* 9 */ 193 + smlal r10, r11, r7, r5 194 + ldr r7, [r12, #640*4] /* 10 */ 195 + smlal r10, r11, r7, r4 196 + ldr r7, [r12, #736*4] /* 11 */ 197 + smlal r10, r11, r7, r3 198 + ldmia r2!, { r3-r6 } /* load D[08..11] */ 199 + ldr r7, [r12, #256*4] /* 4 */ 200 + smlal r10, r11, r7, r6 201 + ldr r7, [r12, #352*4] /* 5 */ 202 + smlal r10, r11, r7, r5 203 + ldr r7, [r12, #384*4] /* 6 */ 204 + smlal r10, r11, r7, r4 205 + ldr r7, [r12, #480*4] /* 7 */ 206 + smlal r10, r11, r7, r3 207 + ldr r7, [r1, #512*4] /* 8 */ 208 + smlal r8, r9, r7, r3 209 + ldr r7, [r1, #608*4] /* 9 */ 210 + smlal r8, r9, r7, r4 211 + ldr r7, [r1, #640*4] /* 10 */ 212 + smlal r8, r9, r7, r5 213 + ldr r7, [r1, #736*4] /* 11 */ 214 + smlal r8, r9, r7, r6 215 + ldmia r2!, { r3-r6 } /* load D[12..15] */ 216 + ldr r7, [r1, #768*4] /* 12 */ 217 + smlal r8, r9, r7, r3 218 + ldr r7, [r1, #864*4] /* 13 */ 219 + smlal r8, r9, r7, r4 220 + ldr r7, [r1, #896*4] /* 14 */ 221 + smlal r8, r9, r7, r5 222 + ldr r7, [r1, #992*4] /* 15 */ 223 + smlal r8, r9, r7, r6 224 + ldr r7, [r12] /* 0 */ 225 + smlal r10, r11, r7, r6 226 + ldr r7, [r12, #96*4] /* 1 */ 227 + smlal r10, r11, r7, r5 228 + ldr r7, [r12, #128*4] /* 2 */ 229 + smlal r10, r11, r7, r4 230 + ldr r7, [r12, #224*4] /* 3 */ 231 + smlal r10, r11, r7, r3 232 + /* store Data[01..15] */ 233 + mov r8, r8, lsr #16 234 + orr r8, r8, r9, lsl #16 /* (lo>>16) || (hi<<16) */ 235 + str r8, [r0] /* store Data */ 236 + /* store Data[31..17] */ 237 + add r0, r0, lr, asl #3 /* r0 = r0 + 2*lr [words] */ 238 + mov r10, r10, lsr #16 239 + orr r10, r10, r11, lsl #16 /* (lo>>16) || (hi<<16) */ 240 + rsb r10, r10, #0 /* r10 = -r10 */ 241 + str r10, [r0], #4 /* store Data */ 242 + sub r0, r0, lr, asl #3 /* r0 = r0 - 2*lr [words] */ 243 + /* correct adresses for next loop */ 244 + sub r12, r12, #4 /* r12 = V-- */ 245 + add r1, r1, #4 /* r1 = V++ */ 246 + /* next loop */ 247 + subs lr, lr, #1 248 + bgt .loop15 249 + 250 + /****************************************** 251 + * V[16] with internal symmetry 252 + *****************************************/ 253 + ldmia r2!, { r3-r6 } /* load D[00..03] */ 254 + ldr r7 , [r1] /* 0 */ 255 + ldr r10, [r1, #992*4] /* 15 */ 256 + rsb r10, r10, r7 /* V[00] - V[15] */ 257 + smull r8, r9, r10, r3 258 + ldr r7 , [r1, #96*4] /* 1 */ 259 + ldr r10, [r1, #896*4] /* 14 */ 260 + rsb r10, r10, r7 /* V[01] - V[14] */ 261 + smlal r8, r9, r10, r4 262 + ldr r7 , [r1, #128*4] /* 2 */ 263 + ldr r10, [r1, #864*4] /* 13 */ 264 + rsb r10, r10, r7 /* V[02] - V[13] */ 265 + smlal r8, r9, r10, r5 266 + ldr r7 , [r1, #224*4] /* 3 */ 267 + ldr r10, [r1, #768*4] /* 12 */ 268 + rsb r10, r10, r7 /* V[03] - V[12] */ 269 + smlal r8, r9, r10, r6 270 + ldmia r2!, { r3-r6 } /* load D[04..07] */ 271 + ldr r7 , [r1, #256*4] /* 4 */ 272 + ldr r10, [r1, #736*4] /* 11 */ 273 + rsb r10, r10, r7 /* V[04] - V[11] */ 274 + smlal r8, r9, r10, r3 275 + ldr r7 , [r1, #352*4] /* 5 */ 276 + ldr r10, [r1, #640*4] /* 10 */ 277 + rsb r10, r10, r7 /* V[05] - V[10] */ 278 + smlal r8, r9, r10, r4 279 + ldr r7 , [r1, #384*4] /* 6 */ 280 + ldr r10, [r1, #608*4] /* 9 */ 281 + rsb r10, r10, r7 /* V[06] - V[09] */ 282 + smlal r8, r9, r10, r5 283 + ldr r7 , [r1, #480*4] /* 7 */ 284 + ldr r10, [r1, #512*4] /* 8 */ 285 + rsb r10, r10, r7 /* V[07] - V[08] */ 286 + smlal r8, r9, r10, r6 287 + mov r8, r8, lsr #16 288 + orr r8, r8, r9, lsl #16 /* (lo>>16) || (hi<<16) */ 289 + str r8, [r0], #4 /* store Data */ 290 + add r1, r1, #4 /* V++ */ 291 + 292 + ldmfd sp!, {r4-r11, pc} 293 + #endif 294 + .mpc_dewindowing_end: 295 + .size mpc_decoder_windowing_D,.mpc_dewindowing_end-mpc_decoder_windowing_D

Configure Feed

Configure Feed