1/* 2 * Copyright (c) 2014 The WebRTC project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11#include "webrtc/modules/audio_coding/codecs/isac/fix/source/codec.h" 12#include "webrtc/modules/audio_coding/codecs/isac/fix/source/fft.h" 13#include "webrtc/modules/audio_coding/codecs/isac/fix/source/settings.h" 14 15// The tables are defined in transform_tables.c file. 16extern const int16_t WebRtcIsacfix_kCosTab1[FRAMESAMPLES/2]; 17extern const int16_t WebRtcIsacfix_kSinTab1[FRAMESAMPLES/2]; 18extern const int16_t WebRtcIsacfix_kCosTab2[FRAMESAMPLES/4]; 19extern const int16_t WebRtcIsacfix_kSinTab2[FRAMESAMPLES/4]; 20 21// MIPS DSPr2 version of the WebRtcIsacfix_Time2Spec function 22// is not bit-exact with the C version. 23// The accuracy of the MIPS DSPr2 version is same or better. 24void WebRtcIsacfix_Time2SpecMIPS(int16_t* inre1Q9, 25 int16_t* inre2Q9, 26 int16_t* outreQ7, 27 int16_t* outimQ7) { 28 int k = FRAMESAMPLES / 2; 29 int32_t tmpreQ16[FRAMESAMPLES / 2], tmpimQ16[FRAMESAMPLES / 2]; 30 int32_t r0, r1, r2, r3, r4, r5, r6, r7, r8, r9; 31 int32_t inre1, inre2, tmpre, tmpim, factor, max, max1; 32 int16_t* cosptr; 33 int16_t* sinptr; 34 35 cosptr = (int16_t*)WebRtcIsacfix_kCosTab1; 36 sinptr = (int16_t*)WebRtcIsacfix_kSinTab1; 37 38 __asm __volatile ( 39 ".set push \n\t" 40 ".set noreorder \n\t" 41 "addiu %[inre1], %[inre1Q9], 0 \n\t" 42 "addiu %[inre2], %[inre2Q9], 0 \n\t" 43 "addiu %[tmpre], %[tmpreQ16], 0 \n\t" 44 "addiu %[tmpim], %[tmpimQ16], 0 \n\t" 45 "addiu %[factor], $zero, 16921 \n\t" 46 "mul %[max], $zero, $zero \n\t" 47 // Multiply with complex exponentials and combine into one complex vector. 48 // Also, calculate the maximal absolute value in the same loop. 49 "1: \n\t" 50#if defined(MIPS_DSP_R2_LE) 51 "lwl %[r0], 0(%[inre1]) \n\t" 52 "lwl %[r2], 0(%[cosptr]) \n\t" 53 "lwl %[r3], 0(%[sinptr]) \n\t" 54 "lwl %[r1], 0(%[inre2]) \n\t" 55 "lwr %[r0], 0(%[inre1]) \n\t" 56 "lwr %[r2], 0(%[cosptr]) \n\t" 57 "lwr %[r3], 0(%[sinptr]) \n\t" 58 "lwr %[r1], 0(%[inre2]) \n\t" 59 "muleq_s.w.phr %[r4], %[r2], %[r0] \n\t" 60 "muleq_s.w.phr %[r5], %[r3], %[r0] \n\t" 61 "muleq_s.w.phr %[r6], %[r3], %[r1] \n\t" 62 "muleq_s.w.phr %[r7], %[r2], %[r1] \n\t" 63 "muleq_s.w.phl %[r8], %[r2], %[r0] \n\t" 64 "muleq_s.w.phl %[r0], %[r3], %[r0] \n\t" 65 "muleq_s.w.phl %[r3], %[r3], %[r1] \n\t" 66 "muleq_s.w.phl %[r1], %[r2], %[r1] \n\t" 67 "addiu %[k], %[k], -2 \n\t" 68 "addu %[r4], %[r4], %[r6] \n\t" 69 "subu %[r5], %[r7], %[r5] \n\t" 70 "sra %[r4], %[r4], 8 \n\t" 71 "sra %[r5], %[r5], 8 \n\t" 72 "mult $ac0, %[factor], %[r4] \n\t" 73 "mult $ac1, %[factor], %[r5] \n\t" 74 "addu %[r3], %[r8], %[r3] \n\t" 75 "subu %[r0], %[r1], %[r0] \n\t" 76 "sra %[r3], %[r3], 8 \n\t" 77 "sra %[r0], %[r0], 8 \n\t" 78 "mult $ac2, %[factor], %[r3] \n\t" 79 "mult $ac3, %[factor], %[r0] \n\t" 80 "extr_r.w %[r4], $ac0, 16 \n\t" 81 "extr_r.w %[r5], $ac1, 16 \n\t" 82 "addiu %[inre1], %[inre1], 4 \n\t" 83 "addiu %[inre2], %[inre2], 4 \n\t" 84 "extr_r.w %[r6], $ac2, 16 \n\t" 85 "extr_r.w %[r7], $ac3, 16 \n\t" 86 "addiu %[cosptr], %[cosptr], 4 \n\t" 87 "addiu %[sinptr], %[sinptr], 4 \n\t" 88 "shra_r.w %[r4], %[r4], 3 \n\t" 89 "shra_r.w %[r5], %[r5], 3 \n\t" 90 "sw %[r4], 0(%[tmpre]) \n\t" 91 "absq_s.w %[r4], %[r4] \n\t" 92 "sw %[r5], 0(%[tmpim]) \n\t" 93 "absq_s.w %[r5], %[r5] \n\t" 94 "shra_r.w %[r6], %[r6], 3 \n\t" 95 "shra_r.w %[r7], %[r7], 3 \n\t" 96 "sw %[r6], 4(%[tmpre]) \n\t" 97 "absq_s.w %[r6], %[r6] \n\t" 98 "sw %[r7], 4(%[tmpim]) \n\t" 99 "absq_s.w %[r7], %[r7] \n\t" 100 "slt %[r0], %[r4], %[r5] \n\t" 101 "movn %[r4], %[r5], %[r0] \n\t" 102 "slt %[r1], %[r6], %[r7] \n\t" 103 "movn %[r6], %[r7], %[r1] \n\t" 104 "slt %[r0], %[max], %[r4] \n\t" 105 "movn %[max], %[r4], %[r0] \n\t" 106 "slt %[r1], %[max], %[r6] \n\t" 107 "movn %[max], %[r6], %[r1] \n\t" 108 "addiu %[tmpre], %[tmpre], 8 \n\t" 109 "bgtz %[k], 1b \n\t" 110 " addiu %[tmpim], %[tmpim], 8 \n\t" 111#else // #if defined(MIPS_DSP_R2_LE) 112 "lh %[r0], 0(%[inre1]) \n\t" 113 "lh %[r1], 0(%[inre2]) \n\t" 114 "lh %[r2], 0(%[cosptr]) \n\t" 115 "lh %[r3], 0(%[sinptr]) \n\t" 116 "addiu %[k], %[k], -1 \n\t" 117 "mul %[r4], %[r0], %[r2] \n\t" 118 "mul %[r5], %[r1], %[r3] \n\t" 119 "mul %[r0], %[r0], %[r3] \n\t" 120 "mul %[r2], %[r1], %[r2] \n\t" 121 "addiu %[inre1], %[inre1], 2 \n\t" 122 "addiu %[inre2], %[inre2], 2 \n\t" 123 "addiu %[cosptr], %[cosptr], 2 \n\t" 124 "addiu %[sinptr], %[sinptr], 2 \n\t" 125 "addu %[r1], %[r4], %[r5] \n\t" 126 "sra %[r1], %[r1], 7 \n\t" 127 "sra %[r3], %[r1], 16 \n\t" 128 "andi %[r1], %[r1], 0xFFFF \n\t" 129 "sra %[r1], %[r1], 1 \n\t" 130 "mul %[r1], %[factor], %[r1] \n\t" 131 "mul %[r3], %[factor], %[r3] \n\t" 132 "subu %[r0], %[r2], %[r0] \n\t" 133 "sra %[r0], %[r0], 7 \n\t" 134 "sra %[r2], %[r0], 16 \n\t" 135 "andi %[r0], %[r0], 0xFFFF \n\t" 136 "sra %[r0], %[r0], 1 \n\t" 137 "mul %[r0], %[factor], %[r0] \n\t" 138 "mul %[r2], %[factor], %[r2] \n\t" 139#if defined(MIPS_DSP_R1_LE) 140 "shra_r.w %[r1], %[r1], 15 \n\t" 141#else // #if defined(MIPS_DSP_R1_LE) 142 "addiu %[r1], %[r1], 0x4000 \n\t" 143 "sra %[r1], %[r1], 15 \n\t" 144#endif // #if defined(MIPS_DSP_R1_LE) 145 "addu %[r1], %[r3], %[r1] \n\t" 146#if defined(MIPS_DSP_R1_LE) 147 "shra_r.w %[r1], %[r1], 3 \n\t" 148#else // #if defined(MIPS_DSP_R1_LE) 149 "addiu %[r1], %[r1], 4 \n\t" 150 "sra %[r1], %[r1], 3 \n\t" 151#endif // #if defined(MIPS_DSP_R1_LE) 152 "sw %[r1], 0(%[tmpre]) \n\t" 153 "addiu %[tmpre], %[tmpre], 4 \n\t" 154#if defined(MIPS_DSP_R1_LE) 155 "absq_s.w %[r1], %[r1] \n\t" 156 "shra_r.w %[r0], %[r0], 15 \n\t" 157#else // #if defined(MIPS_DSP_R1_LE) 158 "negu %[r4], %[r1] \n\t" 159 "slt %[r3], %[r1], $zero \n\t" 160 "movn %[r1], %[r4], %[r3] \n\t" 161 "addiu %[r0], %[r0], 0x4000 \n\t" 162 "sra %[r0], %[r0], 15 \n\t" 163#endif // #if defined(MIPS_DSP_R1_LE) 164 "addu %[r0], %[r0], %[r2] \n\t" 165#if defined(MIPS_DSP_R1_LE) 166 "shra_r.w %[r0], %[r0], 3 \n\t" 167 "sw %[r0], 0(%[tmpim]) \n\t" 168 "absq_s.w %[r0], %[r0] \n\t" 169#else // #if defined(MIPS_DSP_R1_LE) 170 "addiu %[r0], %[r0], 4 \n\t" 171 "sra %[r0], %[r0], 3 \n\t" 172 "sw %[r0], 0(%[tmpim]) \n\t" 173 "negu %[r2], %[r0] \n\t" 174 "slt %[r3], %[r0], $zero \n\t" 175 "movn %[r0], %[r2], %[r3] \n\t" 176#endif // #if defined(MIPS_DSP_R1_LE) 177 "slt %[r2], %[max], %[r1] \n\t" 178 "movn %[max], %[r1], %[r2] \n\t" 179 "slt %[r2], %[max], %[r0] \n\t" 180 "movn %[max], %[r0], %[r2] \n\t" 181 "bgtz %[k], 1b \n\t" 182 " addiu %[tmpim], %[tmpim], 4 \n\t" 183#endif // #if defined(MIPS_DSP_R2_LE) 184 // Calculate WebRtcSpl_NormW32(max). 185 // If max gets value >=0, we should shift max steps to the left, and the 186 // domain will be Q(16+shift). If max gets value <0, we should shift -max 187 // steps to the right, and the domain will be Q(16+max) 188 "clz %[max], %[max] \n\t" 189 "addiu %[max], %[max], -25 \n\t" 190 ".set pop \n\t" 191 : [k] "+r" (k), [inre1] "=&r" (inre1), [inre2] "=&r" (inre2), 192 [r0] "=&r" (r0), [r1] "=&r" (r1), [r2] "=&r" (r2), 193 [r3] "=&r" (r3), [r4] "=&r" (r4), [tmpre] "=&r" (tmpre), 194 [tmpim] "=&r" (tmpim), [max] "=&r" (max), [factor] "=&r" (factor), 195#if defined(MIPS_DSP_R2_LE) 196 [r6] "=&r" (r6), [r7] "=&r" (r7), [r8] "=&r" (r8), 197#endif // #if defined(MIPS_DSP_R2_LE) 198 [r5] "=&r" (r5) 199 : [inre1Q9] "r" (inre1Q9), [inre2Q9] "r" (inre2Q9), 200 [tmpreQ16] "r" (tmpreQ16), [tmpimQ16] "r" (tmpimQ16), 201 [cosptr] "r" (cosptr), [sinptr] "r" (sinptr) 202 : "hi", "lo", "memory" 203#if defined(MIPS_DSP_R2_LE) 204 , "$ac1hi", "$ac1lo", "$ac2hi", "$ac2lo", "$ac3hi", "$ac3lo" 205#endif // #if defined(MIPS_DSP_R2_LE) 206 ); 207 208 // "Fastest" vectors 209 k = FRAMESAMPLES / 4; 210 __asm __volatile ( 211 ".set push \n\t" 212 ".set noreorder \n\t" 213 "addiu %[tmpre], %[tmpreQ16], 0 \n\t" 214 "addiu %[tmpim], %[tmpimQ16], 0 \n\t" 215 "addiu %[inre1], %[inre1Q9], 0 \n\t" 216 "addiu %[inre2], %[inre2Q9], 0 \n\t" 217 "blez %[max], 2f \n\t" 218 " subu %[max1], $zero, %[max] \n\t" 219 "1: \n\t" 220 "lw %[r0], 0(%[tmpre]) \n\t" 221 "lw %[r1], 0(%[tmpim]) \n\t" 222 "lw %[r2], 4(%[tmpre]) \n\t" 223 "lw %[r3], 4(%[tmpim]) \n\t" 224 "addiu %[k], %[k], -1 \n\t" 225 "sllv %[r0], %[r0], %[max] \n\t" 226 "sllv %[r1], %[r1], %[max] \n\t" 227 "sllv %[r2], %[r2], %[max] \n\t" 228 "sllv %[r3], %[r3], %[max] \n\t" 229 "addiu %[tmpre], %[tmpre], 8 \n\t" 230 "addiu %[tmpim], %[tmpim], 8 \n\t" 231 "sh %[r0], 0(%[inre1]) \n\t" 232 "sh %[r1], 0(%[inre2]) \n\t" 233 "sh %[r2], 2(%[inre1]) \n\t" 234 "sh %[r3], 2(%[inre2]) \n\t" 235 "addiu %[inre1], %[inre1], 4 \n\t" 236 "bgtz %[k], 1b \n\t" 237 " addiu %[inre2], %[inre2], 4 \n\t" 238 "b 4f \n\t" 239 " nop \n\t" 240 "2: \n\t" 241#if !defined(MIPS_DSP_R1_LE) 242 "addiu %[r4], %[max1], -1 \n\t" 243 "addiu %[r5], $zero, 1 \n\t" 244 "sllv %[r4], %[r5], %[r4] \n\t" 245#endif // #if !defined(MIPS_DSP_R1_LE) 246 "3: \n\t" 247 "lw %[r0], 0(%[tmpre]) \n\t" 248 "lw %[r1], 0(%[tmpim]) \n\t" 249 "lw %[r2], 4(%[tmpre]) \n\t" 250 "lw %[r3], 4(%[tmpim]) \n\t" 251 "addiu %[k], %[k], -1 \n\t" 252#if defined(MIPS_DSP_R1_LE) 253 "shrav_r.w %[r0], %[r0], %[max1] \n\t" 254 "shrav_r.w %[r1], %[r1], %[max1] \n\t" 255 "shrav_r.w %[r2], %[r2], %[max1] \n\t" 256 "shrav_r.w %[r3], %[r3], %[max1] \n\t" 257#else // #if !defined(MIPS_DSP_R1_LE) 258 "addu %[r0], %[r0], %[r4] \n\t" 259 "addu %[r1], %[r1], %[r4] \n\t" 260 "addu %[r2], %[r2], %[r4] \n\t" 261 "addu %[r3], %[r3], %[r4] \n\t" 262 "srav %[r0], %[r0], %[max1] \n\t" 263 "srav %[r1], %[r1], %[max1] \n\t" 264 "srav %[r2], %[r2], %[max1] \n\t" 265 "srav %[r3], %[r3], %[max1] \n\t" 266#endif // #if !defined(MIPS_DSP_R1_LE) 267 "addiu %[tmpre], %[tmpre], 8 \n\t" 268 "addiu %[tmpim], %[tmpim], 8 \n\t" 269 "sh %[r0], 0(%[inre1]) \n\t" 270 "sh %[r1], 0(%[inre2]) \n\t" 271 "sh %[r2], 2(%[inre1]) \n\t" 272 "sh %[r3], 2(%[inre2]) \n\t" 273 "addiu %[inre1], %[inre1], 4 \n\t" 274 "bgtz %[k], 3b \n\t" 275 " addiu %[inre2], %[inre2], 4 \n\t" 276 "4: \n\t" 277 ".set pop \n\t" 278 : [tmpre] "=&r" (tmpre), [tmpim] "=&r" (tmpim), [inre1] "=&r" (inre1), 279 [inre2] "=&r" (inre2), [k] "+r" (k), [max1] "=&r" (max1), 280#if !defined(MIPS_DSP_R1_LE) 281 [r4] "=&r" (r4), [r5] "=&r" (r5), 282#endif // #if !defined(MIPS_DSP_R1_LE) 283 [r0] "=&r" (r0), [r1] "=&r" (r1), [r2] "=&r" (r2), [r3] "=&r" (r3) 284 : [tmpreQ16] "r" (tmpreQ16), [tmpimQ16] "r" (tmpimQ16), 285 [inre1Q9] "r" (inre1Q9), [inre2Q9] "r" (inre2Q9), [max] "r" (max) 286 : "memory" 287 ); 288 289 // Get DFT 290 WebRtcIsacfix_FftRadix16Fastest(inre1Q9, inre2Q9, -1); // real call 291 292 // "Fastest" vectors and 293 // Use symmetry to separate into two complex vectors 294 // and center frames in time around zero 295 // merged into one loop 296 cosptr = (int16_t*)WebRtcIsacfix_kCosTab2; 297 sinptr = (int16_t*)WebRtcIsacfix_kSinTab2; 298 k = FRAMESAMPLES / 4; 299 factor = FRAMESAMPLES - 2; // offset for FRAMESAMPLES / 2 - 1 array member 300 301 __asm __volatile ( 302 ".set push \n\t" 303 ".set noreorder \n\t" 304 "addiu %[inre1], %[inre1Q9], 0 \n\t" 305 "addiu %[inre2], %[inre2Q9], 0 \n\t" 306 "addiu %[tmpre], %[outreQ7], 0 \n\t" 307 "addiu %[tmpim], %[outimQ7], 0 \n\t" 308 "bltz %[max], 2f \n\t" 309 " subu %[max1], $zero, %[max] \n\t" 310 "1: \n\t" 311#if !defined(MIPS_DSP_R1_LE) 312 "addu %[r4], %[inre1], %[offset] \n\t" 313 "addu %[r5], %[inre2], %[offset] \n\t" 314#endif // #if !defined(MIPS_DSP_R1_LE) 315 "lh %[r0], 0(%[inre1]) \n\t" 316 "lh %[r1], 0(%[inre2]) \n\t" 317#if defined(MIPS_DSP_R1_LE) 318 "lhx %[r2], %[offset](%[inre1]) \n\t" 319 "lhx %[r3], %[offset](%[inre2]) \n\t" 320#else // #if defined(MIPS_DSP_R1_LE) 321 "lh %[r2], 0(%[r4]) \n\t" 322 "lh %[r3], 0(%[r5]) \n\t" 323#endif // #if defined(MIPS_DSP_R1_LE) 324 "srav %[r0], %[r0], %[max] \n\t" 325 "srav %[r1], %[r1], %[max] \n\t" 326 "srav %[r2], %[r2], %[max] \n\t" 327 "srav %[r3], %[r3], %[max] \n\t" 328 "addu %[r4], %[r0], %[r2] \n\t" 329 "subu %[r0], %[r2], %[r0] \n\t" 330 "subu %[r2], %[r1], %[r3] \n\t" 331 "addu %[r1], %[r1], %[r3] \n\t" 332 "lh %[r3], 0(%[cosptr]) \n\t" 333 "lh %[r5], 0(%[sinptr]) \n\t" 334 "andi %[r6], %[r4], 0xFFFF \n\t" 335 "sra %[r4], %[r4], 16 \n\t" 336 "mul %[r7], %[r3], %[r6] \n\t" 337 "mul %[r8], %[r3], %[r4] \n\t" 338 "mul %[r6], %[r5], %[r6] \n\t" 339 "mul %[r4], %[r5], %[r4] \n\t" 340 "addiu %[k], %[k], -1 \n\t" 341 "addiu %[inre1], %[inre1], 2 \n\t" 342 "addiu %[inre2], %[inre2], 2 \n\t" 343#if defined(MIPS_DSP_R1_LE) 344 "shra_r.w %[r7], %[r7], 14 \n\t" 345#else // #if defined(MIPS_DSP_R1_LE) 346 "addiu %[r7], %[r7], 0x2000 \n\t" 347 "sra %[r7], %[r7], 14 \n\t" 348#endif // #if defined(MIPS_DSP_R1_LE) 349 "sll %[r8], %[r8], 2 \n\t" 350 "addu %[r8], %[r8], %[r7] \n\t" 351#if defined(MIPS_DSP_R1_LE) 352 "shra_r.w %[r6], %[r6], 14 \n\t" 353#else // #if defined(MIPS_DSP_R1_LE) 354 "addiu %[r6], %[r6], 0x2000 \n\t" 355 "sra %[r6], %[r6], 14 \n\t" 356#endif // #if defined(MIPS_DSP_R1_LE) 357 "sll %[r4], %[r4], 2 \n\t" 358 "addu %[r4], %[r4], %[r6] \n\t" 359 "andi %[r6], %[r2], 0xFFFF \n\t" 360 "sra %[r2], %[r2], 16 \n\t" 361 "mul %[r7], %[r5], %[r6] \n\t" 362 "mul %[r9], %[r5], %[r2] \n\t" 363 "mul %[r6], %[r3], %[r6] \n\t" 364 "mul %[r2], %[r3], %[r2] \n\t" 365 "addiu %[cosptr], %[cosptr], 2 \n\t" 366 "addiu %[sinptr], %[sinptr], 2 \n\t" 367#if defined(MIPS_DSP_R1_LE) 368 "shra_r.w %[r7], %[r7], 14 \n\t" 369#else // #if defined(MIPS_DSP_R1_LE) 370 "addiu %[r7], %[r7], 0x2000 \n\t" 371 "sra %[r7], %[r7], 14 \n\t" 372#endif // #if defined(MIPS_DSP_R1_LE) 373 "sll %[r9], %[r9], 2 \n\t" 374 "addu %[r9], %[r7], %[r9] \n\t" 375#if defined(MIPS_DSP_R1_LE) 376 "shra_r.w %[r6], %[r6], 14 \n\t" 377#else // #if defined(MIPS_DSP_R1_LE) 378 "addiu %[r6], %[r6], 0x2000 \n\t" 379 "sra %[r6], %[r6], 14 \n\t" 380#endif // #if defined(MIPS_DSP_R1_LE) 381 "sll %[r2], %[r2], 2 \n\t" 382 "addu %[r2], %[r6], %[r2] \n\t" 383 "subu %[r8], %[r8], %[r9] \n\t" 384 "sra %[r8], %[r8], 9 \n\t" 385 "addu %[r2], %[r4], %[r2] \n\t" 386 "sra %[r2], %[r2], 9 \n\t" 387 "sh %[r8], 0(%[tmpre]) \n\t" 388 "sh %[r2], 0(%[tmpim]) \n\t" 389 390 "andi %[r4], %[r1], 0xFFFF \n\t" 391 "sra %[r1], %[r1], 16 \n\t" 392 "andi %[r6], %[r0], 0xFFFF \n\t" 393 "sra %[r0], %[r0], 16 \n\t" 394 "mul %[r7], %[r5], %[r4] \n\t" 395 "mul %[r9], %[r5], %[r1] \n\t" 396 "mul %[r4], %[r3], %[r4] \n\t" 397 "mul %[r1], %[r3], %[r1] \n\t" 398 "mul %[r8], %[r3], %[r0] \n\t" 399 "mul %[r3], %[r3], %[r6] \n\t" 400 "mul %[r6], %[r5], %[r6] \n\t" 401 "mul %[r0], %[r5], %[r0] \n\t" 402#if defined(MIPS_DSP_R1_LE) 403 "shra_r.w %[r7], %[r7], 14 \n\t" 404#else // #if defined(MIPS_DSP_R1_LE) 405 "addiu %[r7], %[r7], 0x2000 \n\t" 406 "sra %[r7], %[r7], 14 \n\t" 407#endif // #if defined(MIPS_DSP_R1_LE) 408 "sll %[r9], %[r9], 2 \n\t" 409 "addu %[r9], %[r9], %[r7] \n\t" 410#if defined(MIPS_DSP_R1_LE) 411 "shra_r.w %[r4], %[r4], 14 \n\t" 412#else // #if defined(MIPS_DSP_R1_LE) 413 "addiu %[r4], %[r4], 0x2000 \n\t" 414 "sra %[r4], %[r4], 14 \n\t" 415#endif // #if defined(MIPS_DSP_R1_LE) 416 "sll %[r1], %[r1], 2 \n\t" 417 "addu %[r1], %[r1], %[r4] \n\t" 418#if defined(MIPS_DSP_R1_LE) 419 "shra_r.w %[r3], %[r3], 14 \n\t" 420#else // #if defined(MIPS_DSP_R1_LE) 421 "addiu %[r3], %[r3], 0x2000 \n\t" 422 "sra %[r3], %[r3], 14 \n\t" 423#endif // #if defined(MIPS_DSP_R1_LE) 424 "sll %[r8], %[r8], 2 \n\t" 425 "addu %[r8], %[r8], %[r3] \n\t" 426#if defined(MIPS_DSP_R1_LE) 427 "shra_r.w %[r6], %[r6], 14 \n\t" 428#else // #if defined(MIPS_DSP_R1_LE) 429 "addiu %[r6], %[r6], 0x2000 \n\t" 430 "sra %[r6], %[r6], 14 \n\t" 431#endif // #if defined(MIPS_DSP_R1_LE) 432 "sll %[r0], %[r0], 2 \n\t" 433 "addu %[r0], %[r0], %[r6] \n\t" 434 "addu %[r3], %[tmpre], %[offset] \n\t" 435 "addu %[r2], %[tmpim], %[offset] \n\t" 436 "addu %[r9], %[r9], %[r8] \n\t" 437 "negu %[r9], %[r9] \n\t" 438 "sra %[r9], %[r9], 9 \n\t" 439 "subu %[r0], %[r0], %[r1] \n\t" 440 "addiu %[offset], %[offset], -4 \n\t" 441 "sh %[r9], 0(%[r3]) \n\t" 442 "sh %[r0], 0(%[r2]) \n\t" 443 "addiu %[tmpre], %[tmpre], 2 \n\t" 444 "bgtz %[k], 1b \n\t" 445 " addiu %[tmpim], %[tmpim], 2 \n\t" 446 "b 3f \n\t" 447 " nop \n\t" 448 "2: \n\t" 449#if !defined(MIPS_DSP_R1_LE) 450 "addu %[r4], %[inre1], %[offset] \n\t" 451 "addu %[r5], %[inre2], %[offset] \n\t" 452#endif // #if !defined(MIPS_DSP_R1_LE) 453 "lh %[r0], 0(%[inre1]) \n\t" 454 "lh %[r1], 0(%[inre2]) \n\t" 455#if defined(MIPS_DSP_R1_LE) 456 "lhx %[r2], %[offset](%[inre1]) \n\t" 457 "lhx %[r3], %[offset](%[inre2]) \n\t" 458#else // #if defined(MIPS_DSP_R1_LE) 459 "lh %[r2], 0(%[r4]) \n\t" 460 "lh %[r3], 0(%[r5]) \n\t" 461#endif // #if defined(MIPS_DSP_R1_LE) 462 "sllv %[r0], %[r0], %[max1] \n\t" 463 "sllv %[r1], %[r1], %[max1] \n\t" 464 "sllv %[r2], %[r2], %[max1] \n\t" 465 "sllv %[r3], %[r3], %[max1] \n\t" 466 "addu %[r4], %[r0], %[r2] \n\t" 467 "subu %[r0], %[r2], %[r0] \n\t" 468 "subu %[r2], %[r1], %[r3] \n\t" 469 "addu %[r1], %[r1], %[r3] \n\t" 470 "lh %[r3], 0(%[cosptr]) \n\t" 471 "lh %[r5], 0(%[sinptr]) \n\t" 472 "andi %[r6], %[r4], 0xFFFF \n\t" 473 "sra %[r4], %[r4], 16 \n\t" 474 "mul %[r7], %[r3], %[r6] \n\t" 475 "mul %[r8], %[r3], %[r4] \n\t" 476 "mul %[r6], %[r5], %[r6] \n\t" 477 "mul %[r4], %[r5], %[r4] \n\t" 478 "addiu %[k], %[k], -1 \n\t" 479 "addiu %[inre1], %[inre1], 2 \n\t" 480 "addiu %[inre2], %[inre2], 2 \n\t" 481#if defined(MIPS_DSP_R1_LE) 482 "shra_r.w %[r7], %[r7], 14 \n\t" 483#else // #if defined(MIPS_DSP_R1_LE) 484 "addiu %[r7], %[r7], 0x2000 \n\t" 485 "sra %[r7], %[r7], 14 \n\t" 486#endif // #if defined(MIPS_DSP_R1_LE) 487 "sll %[r8], %[r8], 2 \n\t" 488 "addu %[r8], %[r8], %[r7] \n\t" 489#if defined(MIPS_DSP_R1_LE) 490 "shra_r.w %[r6], %[r6], 14 \n\t" 491#else // #if defined(MIPS_DSP_R1_LE) 492 "addiu %[r6], %[r6], 0x2000 \n\t" 493 "sra %[r6], %[r6], 14 \n\t" 494#endif // #if defined(MIPS_DSP_R1_LE) 495 "sll %[r4], %[r4], 2 \n\t" 496 "addu %[r4], %[r4], %[r6] \n\t" 497 "andi %[r6], %[r2], 0xFFFF \n\t" 498 "sra %[r2], %[r2], 16 \n\t" 499 "mul %[r7], %[r5], %[r6] \n\t" 500 "mul %[r9], %[r5], %[r2] \n\t" 501 "mul %[r6], %[r3], %[r6] \n\t" 502 "mul %[r2], %[r3], %[r2] \n\t" 503 "addiu %[cosptr], %[cosptr], 2 \n\t" 504 "addiu %[sinptr], %[sinptr], 2 \n\t" 505#if defined(MIPS_DSP_R1_LE) 506 "shra_r.w %[r7], %[r7], 14 \n\t" 507#else // #if defined(MIPS_DSP_R1_LE) 508 "addiu %[r7], %[r7], 0x2000 \n\t" 509 "sra %[r7], %[r7], 14 \n\t" 510#endif // #if defined(MIPS_DSP_R1_LE) 511 "sll %[r9], %[r9], 2 \n\t" 512 "addu %[r9], %[r7], %[r9] \n\t" 513#if defined(MIPS_DSP_R1_LE) 514 "shra_r.w %[r6], %[r6], 14 \n\t" 515#else // #if defined(MIPS_DSP_R1_LE) 516 "addiu %[r6], %[r6], 0x2000 \n\t" 517 "sra %[r6], %[r6], 14 \n\t" 518#endif // #if defined(MIPS_DSP_R1_LE) 519 "sll %[r2], %[r2], 2 \n\t" 520 "addu %[r2], %[r6], %[r2] \n\t" 521 "subu %[r8], %[r8], %[r9] \n\t" 522 "sra %[r8], %[r8], 9 \n\t" 523 "addu %[r2], %[r4], %[r2] \n\t" 524 "sra %[r2], %[r2], 9 \n\t" 525 "sh %[r8], 0(%[tmpre]) \n\t" 526 "sh %[r2], 0(%[tmpim]) \n\t" 527 "andi %[r4], %[r1], 0xFFFF \n\t" 528 "sra %[r1], %[r1], 16 \n\t" 529 "andi %[r6], %[r0], 0xFFFF \n\t" 530 "sra %[r0], %[r0], 16 \n\t" 531 "mul %[r7], %[r5], %[r4] \n\t" 532 "mul %[r9], %[r5], %[r1] \n\t" 533 "mul %[r4], %[r3], %[r4] \n\t" 534 "mul %[r1], %[r3], %[r1] \n\t" 535 "mul %[r8], %[r3], %[r0] \n\t" 536 "mul %[r3], %[r3], %[r6] \n\t" 537 "mul %[r6], %[r5], %[r6] \n\t" 538 "mul %[r0], %[r5], %[r0] \n\t" 539#if defined(MIPS_DSP_R1_LE) 540 "shra_r.w %[r7], %[r7], 14 \n\t" 541#else // #if defined(MIPS_DSP_R1_LE) 542 "addiu %[r7], %[r7], 0x2000 \n\t" 543 "sra %[r7], %[r7], 14 \n\t" 544#endif // #if defined(MIPS_DSP_R1_LE) 545 "sll %[r9], %[r9], 2 \n\t" 546 "addu %[r9], %[r9], %[r7] \n\t" 547#if defined(MIPS_DSP_R1_LE) 548 "shra_r.w %[r4], %[r4], 14 \n\t" 549#else // #if defined(MIPS_DSP_R1_LE) 550 "addiu %[r4], %[r4], 0x2000 \n\t" 551 "sra %[r4], %[r4], 14 \n\t" 552#endif 553 "sll %[r1], %[r1], 2 \n\t" 554 "addu %[r1], %[r1], %[r4] \n\t" 555#if defined(MIPS_DSP_R1_LE) 556 "shra_r.w %[r3], %[r3], 14 \n\t" 557#else // #if defined(MIPS_DSP_R1_LE) 558 "addiu %[r3], %[r3], 0x2000 \n\t" 559 "sra %[r3], %[r3], 14 \n\t" 560#endif // #if defined(MIPS_DSP_R1_LE) 561 "sll %[r8], %[r8], 2 \n\t" 562 "addu %[r8], %[r8], %[r3] \n\t" 563#if defined(MIPS_DSP_R1_LE) 564 "shra_r.w %[r6], %[r6], 14 \n\t" 565#else // #if defined(MIPS_DSP_R1_LE) 566 "addiu %[r6], %[r6], 0x2000 \n\t" 567 "sra %[r6], %[r6], 14 \n\t" 568#endif // #if defined(MIPS_DSP_R1_LE) 569 "sll %[r0], %[r0], 2 \n\t" 570 "addu %[r0], %[r0], %[r6] \n\t" 571 "addu %[r3], %[tmpre], %[offset] \n\t" 572 "addu %[r2], %[tmpim], %[offset] \n\t" 573 "addu %[r9], %[r9], %[r8] \n\t" 574 "negu %[r9], %[r9] \n\t" 575 "sra %[r9], %[r9], 9 \n\t" 576 "subu %[r0], %[r0], %[r1] \n\t" 577 "sra %[r0], %[r0], 9 \n\t" 578 "addiu %[offset], %[offset], -4 \n\t" 579 "sh %[r9], 0(%[r3]) \n\t" 580 "sh %[r0], 0(%[r2]) \n\t" 581 "addiu %[tmpre], %[tmpre], 2 \n\t" 582 "bgtz %[k], 2b \n\t" 583 " addiu %[tmpim], %[tmpim], 2 \n\t" 584 "3: \n\t" 585 ".set pop \n\t" 586 : [inre1] "=&r" (inre1), [inre2] "=&r" (inre2), [tmpre] "=&r" (tmpre), 587 [tmpim] "=&r" (tmpim), [offset] "+r" (factor), [k] "+r" (k), 588 [r0] "=&r" (r0), [r1] "=&r" (r1), [r2] "=&r" (r2), [r3] "=&r" (r3), 589 [r4] "=&r" (r4), [r5] "=&r" (r5), [r6] "=&r" (r6), [r7] "=&r" (r7), 590 [r8] "=&r" (r8), [r9] "=&r" (r9), [max1] "=&r" (max1) 591 : [inre1Q9] "r" (inre1Q9), [inre2Q9] "r" (inre2Q9), 592 [outreQ7] "r" (outreQ7), [outimQ7] "r" (outimQ7), 593 [max] "r" (max), [cosptr] "r" (cosptr), [sinptr] "r" (sinptr) 594 : "hi", "lo", "memory" 595 ); 596} 597 598void WebRtcIsacfix_Spec2TimeMIPS(int16_t *inreQ7, 599 int16_t *inimQ7, 600 int32_t *outre1Q16, 601 int32_t *outre2Q16) { 602 int k = FRAMESAMPLES / 4; 603 int16_t* inre; 604 int16_t* inim; 605 int32_t* outre1; 606 int32_t* outre2; 607 int16_t* cosptr = (int16_t*)WebRtcIsacfix_kCosTab2; 608 int16_t* sinptr = (int16_t*)WebRtcIsacfix_kSinTab2; 609 int32_t r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, max, max1; 610#if defined(MIPS_DSP_R1_LE) 611 int32_t offset = FRAMESAMPLES - 4; 612#else // #if defined(MIPS_DSP_R1_LE) 613 int32_t offset = FRAMESAMPLES - 2; 614#endif // #if defined(MIPS_DSP_R1_LE) 615 616 __asm __volatile ( 617 ".set push \n\t" 618 ".set noreorder \n\t" 619 "addiu %[inre], %[inreQ7], 0 \n\t" 620 "addiu %[inim] , %[inimQ7], 0 \n\t" 621 "addiu %[outre1], %[outre1Q16], 0 \n\t" 622 "addiu %[outre2], %[outre2Q16], 0 \n\t" 623 "mul %[max], $zero, $zero \n\t" 624 "1: \n\t" 625#if defined(MIPS_DSP_R1_LE) 626 // Process two samples in one iteration avoiding left shift before 627 // multiplication. MaxAbsValueW32 function inlined into the loop. 628 "addu %[r8], %[inre], %[offset] \n\t" 629 "addu %[r9], %[inim], %[offset] \n\t" 630 "lwl %[r4], 0(%[r8]) \n\t" 631 "lwl %[r5], 0(%[r9]) \n\t" 632 "lwl %[r0], 0(%[inre]) \n\t" 633 "lwl %[r1], 0(%[inim]) \n\t" 634 "lwl %[r2], 0(%[cosptr]) \n\t" 635 "lwl %[r3], 0(%[sinptr]) \n\t" 636 "lwr %[r4], 0(%[r8]) \n\t" 637 "lwr %[r5], 0(%[r9]) \n\t" 638 "lwr %[r0], 0(%[inre]) \n\t" 639 "lwr %[r1], 0(%[inim]) \n\t" 640 "lwr %[r2], 0(%[cosptr]) \n\t" 641 "lwr %[r3], 0(%[sinptr]) \n\t" 642 "packrl.ph %[r4], %[r4], %[r4] \n\t" 643 "packrl.ph %[r5], %[r5], %[r5] \n\t" 644 "muleq_s.w.phr %[r6], %[r0], %[r2] \n\t" 645 "muleq_s.w.phr %[r7], %[r1], %[r3] \n\t" 646 "muleq_s.w.phr %[r8], %[r4], %[r2] \n\t" 647 "muleq_s.w.phr %[r9], %[r5], %[r3] \n\t" 648 "addiu %[k], %[k], -2 \n\t" 649 "addiu %[cosptr], %[cosptr], 4 \n\t" 650 "addiu %[sinptr], %[sinptr], 4 \n\t" 651 "addiu %[inre], %[inre], 4 \n\t" 652 "addiu %[inim], %[inim], 4 \n\t" 653 "shra_r.w %[r6], %[r6], 6 \n\t" 654 "shra_r.w %[r7], %[r7], 6 \n\t" 655 "shra_r.w %[r8], %[r8], 6 \n\t" 656 "shra_r.w %[r9], %[r9], 6 \n\t" 657 "addu %[r6], %[r6], %[r7] \n\t" 658 "subu %[r9], %[r9], %[r8] \n\t" 659 "subu %[r7], %[r6], %[r9] \n\t" 660 "addu %[r6], %[r6], %[r9] \n\t" 661 "sw %[r7], 0(%[outre1]) \n\t" 662 "absq_s.w %[r7], %[r7] \n\t" 663 "slt %[r8], %[max], %[r7] \n\t" 664 "movn %[max], %[r7], %[r8] \n\t" 665 "sll %[r7], %[offset], 1 \n\t" 666 "addu %[r7], %[outre1], %[r7] \n\t" 667 "sw %[r6], 4(%[r7]) \n\t" 668 "absq_s.w %[r6], %[r6] \n\t" 669 "slt %[r8], %[max], %[r6] \n\t" 670 "movn %[max], %[r6], %[r8] \n\t" 671 "muleq_s.w.phl %[r6], %[r0], %[r2] \n\t" 672 "muleq_s.w.phl %[r7], %[r1], %[r3] \n\t" 673 "muleq_s.w.phl %[r8], %[r4], %[r2] \n\t" 674 "muleq_s.w.phl %[r9], %[r5], %[r3] \n\t" 675 "shra_r.w %[r6], %[r6], 6 \n\t" 676 "shra_r.w %[r7], %[r7], 6 \n\t" 677 "shra_r.w %[r8], %[r8], 6 \n\t" 678 "shra_r.w %[r9], %[r9], 6 \n\t" 679 "addu %[r6], %[r6], %[r7] \n\t" 680 "subu %[r9], %[r9], %[r8] \n\t" 681 "subu %[r7], %[r6], %[r9] \n\t" 682 "addu %[r6], %[r6], %[r9] \n\t" 683 "sw %[r7], 4(%[outre1]) \n\t" 684 "absq_s.w %[r7], %[r7] \n\t" 685 "slt %[r8], %[max], %[r7] \n\t" 686 "movn %[max], %[r7], %[r8] \n\t" 687 "sll %[r7], %[offset], 1 \n\t" 688 "addu %[r7], %[outre1], %[r7] \n\t" 689 "sw %[r6], 0(%[r7]) \n\t" 690 "absq_s.w %[r6], %[r6] \n\t" 691 "slt %[r8], %[max], %[r6] \n\t" 692 "movn %[max], %[r6], %[r8] \n\t" 693 "muleq_s.w.phr %[r6], %[r1], %[r2] \n\t" 694 "muleq_s.w.phr %[r7], %[r0], %[r3] \n\t" 695 "muleq_s.w.phr %[r8], %[r5], %[r2] \n\t" 696 "muleq_s.w.phr %[r9], %[r4], %[r3] \n\t" 697 "addiu %[outre1], %[outre1], 8 \n\t" 698 "shra_r.w %[r6], %[r6], 6 \n\t" 699 "shra_r.w %[r7], %[r7], 6 \n\t" 700 "shra_r.w %[r8], %[r8], 6 \n\t" 701 "shra_r.w %[r9], %[r9], 6 \n\t" 702 "subu %[r6], %[r6], %[r7] \n\t" 703 "addu %[r9], %[r9], %[r8] \n\t" 704 "subu %[r7], %[r6], %[r9] \n\t" 705 "addu %[r6], %[r9], %[r6] \n\t" 706 "negu %[r6], %[r6] \n\t" 707 "sw %[r7], 0(%[outre2]) \n\t" 708 "absq_s.w %[r7], %[r7] \n\t" 709 "slt %[r8], %[max], %[r7] \n\t" 710 "movn %[max], %[r7], %[r8] \n\t" 711 "sll %[r7], %[offset], 1 \n\t" 712 "addu %[r7], %[outre2], %[r7] \n\t" 713 "sw %[r6], 4(%[r7]) \n\t" 714 "absq_s.w %[r6], %[r6] \n\t" 715 "slt %[r8], %[max], %[r6] \n\t" 716 "movn %[max], %[r6], %[r8] \n\t" 717 "muleq_s.w.phl %[r6], %[r1], %[r2] \n\t" 718 "muleq_s.w.phl %[r7], %[r0], %[r3] \n\t" 719 "muleq_s.w.phl %[r8], %[r5], %[r2] \n\t" 720 "muleq_s.w.phl %[r9], %[r4], %[r3] \n\t" 721 "addiu %[offset], %[offset], -8 \n\t" 722 "shra_r.w %[r6], %[r6], 6 \n\t" 723 "shra_r.w %[r7], %[r7], 6 \n\t" 724 "shra_r.w %[r8], %[r8], 6 \n\t" 725 "shra_r.w %[r9], %[r9], 6 \n\t" 726 "subu %[r6], %[r6], %[r7] \n\t" 727 "addu %[r9], %[r9], %[r8] \n\t" 728 "subu %[r7], %[r6], %[r9] \n\t" 729 "addu %[r6], %[r9], %[r6] \n\t" 730 "negu %[r6], %[r6] \n\t" 731 "sw %[r7], 4(%[outre2]) \n\t" 732 "absq_s.w %[r7], %[r7] \n\t" 733 "slt %[r8], %[max], %[r7] \n\t" 734 "movn %[max], %[r7], %[r8] \n\t" 735 "sll %[r7], %[offset], 1 \n\t" 736 "addu %[r7], %[outre2], %[r7] \n\t" 737 "sw %[r6], 0(%[r7]) \n\t" 738 "absq_s.w %[r6], %[r6] \n\t" 739 "slt %[r8], %[max], %[r6] \n\t" 740 "movn %[max], %[r6], %[r8] \n\t" 741 "bgtz %[k], 1b \n\t" 742 " addiu %[outre2], %[outre2], 8 \n\t" 743#else // #if defined(MIPS_DSP_R1_LE) 744 "lh %[r0], 0(%[inre]) \n\t" 745 "lh %[r1], 0(%[inim]) \n\t" 746 "lh %[r4], 0(%[cosptr]) \n\t" 747 "lh %[r5], 0(%[sinptr]) \n\t" 748 "addiu %[k], %[k], -1 \n\t" 749 "mul %[r2], %[r0], %[r4] \n\t" 750 "mul %[r0], %[r0], %[r5] \n\t" 751 "mul %[r3], %[r1], %[r5] \n\t" 752 "mul %[r1], %[r1], %[r4] \n\t" 753 "addiu %[cosptr], %[cosptr], 2 \n\t" 754 "addiu %[sinptr], %[sinptr], 2 \n\t" 755 "addu %[r8], %[inre], %[offset] \n\t" 756 "addu %[r9], %[inim], %[offset] \n\t" 757 "addiu %[r2], %[r2], 16 \n\t" 758 "sra %[r2], %[r2], 5 \n\t" 759 "addiu %[r0], %[r0], 16 \n\t" 760 "sra %[r0], %[r0], 5 \n\t" 761 "addiu %[r3], %[r3], 16 \n\t" 762 "sra %[r3], %[r3], 5 \n\t" 763 "lh %[r6], 0(%[r8]) \n\t" 764 "lh %[r7], 0(%[r9]) \n\t" 765 "addiu %[r1], %[r1], 16 \n\t" 766 "sra %[r1], %[r1], 5 \n\t" 767 "mul %[r8], %[r7], %[r4] \n\t" 768 "mul %[r7], %[r7], %[r5] \n\t" 769 "mul %[r9], %[r6], %[r4] \n\t" 770 "mul %[r6], %[r6], %[r5] \n\t" 771 "addu %[r2], %[r2], %[r3] \n\t" 772 "subu %[r1], %[r1], %[r0] \n\t" 773 "sll %[r0], %[offset], 1 \n\t" 774 "addu %[r4], %[outre1], %[r0] \n\t" 775 "addu %[r5], %[outre2], %[r0] \n\t" 776 "addiu %[r8], %[r8], 16 \n\t" 777 "sra %[r8], %[r8], 5 \n\t" 778 "addiu %[r7], %[r7], 16 \n\t" 779 "sra %[r7], %[r7], 5 \n\t" 780 "addiu %[r6], %[r6], 16 \n\t" 781 "sra %[r6], %[r6], 5 \n\t" 782 "addiu %[r9], %[r9], 16 \n\t" 783 "sra %[r9], %[r9], 5 \n\t" 784 "addu %[r8], %[r8], %[r6] \n\t" 785 "negu %[r8], %[r8] \n\t" 786 "subu %[r7], %[r7], %[r9] \n\t" 787 "subu %[r6], %[r2], %[r7] \n\t" 788 "addu %[r0], %[r2], %[r7] \n\t" 789 "addu %[r3], %[r1], %[r8] \n\t" 790 "subu %[r1], %[r8], %[r1] \n\t" 791 "sw %[r6], 0(%[outre1]) \n\t" 792 "sw %[r0], 0(%[r4]) \n\t" 793 "sw %[r3], 0(%[outre2]) \n\t" 794 "sw %[r1], 0(%[r5]) \n\t" 795 "addiu %[outre1], %[outre1], 4 \n\t" 796 "addiu %[offset], %[offset], -4 \n\t" 797 "addiu %[inre], %[inre], 2 \n\t" 798 "addiu %[inim], %[inim], 2 \n\t" 799 // Inlined WebRtcSpl_MaxAbsValueW32 800 "negu %[r5], %[r6] \n\t" 801 "slt %[r2], %[r6], $zero \n\t" 802 "movn %[r6], %[r5], %[r2] \n\t" 803 "negu %[r5], %[r0] \n\t" 804 "slt %[r2], %[r0], $zero \n\t" 805 "movn %[r0], %[r5], %[r2] \n\t" 806 "negu %[r5], %[r3] \n\t" 807 "slt %[r2], %[r3], $zero \n\t" 808 "movn %[r3], %[r5], %[r2] \n\t" 809 "negu %[r5], %[r1] \n\t" 810 "slt %[r2], %[r1], $zero \n\t" 811 "movn %[r1], %[r5], %[r2] \n\t" 812 "slt %[r2], %[r6], %[r0] \n\t" 813 "slt %[r5], %[r3], %[r1] \n\t" 814 "movn %[r6], %[r0], %[r2] \n\t" 815 "movn %[r3], %[r1], %[r5] \n\t" 816 "slt %[r2], %[r6], %[r3] \n\t" 817 "movn %[r6], %[r3], %[r2] \n\t" 818 "slt %[r2], %[max], %[r6] \n\t" 819 "movn %[max], %[r6], %[r2] \n\t" 820 "bgtz %[k], 1b \n\t" 821 " addiu %[outre2], %[outre2], 4 \n\t" 822#endif // #if defined(MIPS_DSP_R1_LE) 823 "clz %[max], %[max] \n\t" 824 "addiu %[max], %[max], -25 \n\t" 825 ".set pop \n\t" 826 : [inre] "=&r" (inre), [inim] "=&r" (inim), 827 [outre1] "=&r" (outre1), [outre2] "=&r" (outre2), 828 [offset] "+r" (offset), [k] "+r" (k), [r0] "=&r" (r0), 829 [r1] "=&r" (r1), [r2] "=&r" (r2), [r3] "=&r" (r3), 830 [r4] "=&r" (r4), [r5] "=&r" (r5), [r6] "=&r" (r6), 831 [r7] "=&r" (r7), [r8] "=&r" (r8), [r9] "=&r" (r9), 832 [max] "=&r" (max) 833 : [inreQ7] "r" (inreQ7), [inimQ7] "r" (inimQ7), 834 [cosptr] "r" (cosptr), [sinptr] "r" (sinptr), 835 [outre1Q16] "r" (outre1Q16), [outre2Q16] "r" (outre2Q16) 836 : "hi", "lo", "memory" 837 ); 838 839 // "Fastest" vectors 840 k = FRAMESAMPLES / 4; 841 __asm __volatile ( 842 ".set push \n\t" 843 ".set noreorder \n\t" 844 "addiu %[inre], %[inreQ7], 0 \n\t" 845 "addiu %[inim], %[inimQ7], 0 \n\t" 846 "addiu %[outre1], %[outre1Q16], 0 \n\t" 847 "addiu %[outre2], %[outre2Q16], 0 \n\t" 848 "bltz %[max], 2f \n\t" 849 " subu %[max1], $zero, %[max] \n\t" 850 "1: \n\t" 851 "lw %[r0], 0(%[outre1]) \n\t" 852 "lw %[r1], 0(%[outre2]) \n\t" 853 "lw %[r2], 4(%[outre1]) \n\t" 854 "lw %[r3], 4(%[outre2]) \n\t" 855 "sllv %[r0], %[r0], %[max] \n\t" 856 "sllv %[r1], %[r1], %[max] \n\t" 857 "sllv %[r2], %[r2], %[max] \n\t" 858 "sllv %[r3], %[r3], %[max] \n\t" 859 "addiu %[k], %[k], -1 \n\t" 860 "addiu %[outre1], %[outre1], 8 \n\t" 861 "addiu %[outre2], %[outre2], 8 \n\t" 862 "sh %[r0], 0(%[inre]) \n\t" 863 "sh %[r1], 0(%[inim]) \n\t" 864 "sh %[r2], 2(%[inre]) \n\t" 865 "sh %[r3], 2(%[inim]) \n\t" 866 "addiu %[inre], %[inre], 4 \n\t" 867 "bgtz %[k], 1b \n\t" 868 " addiu %[inim], %[inim], 4 \n\t" 869 "b 4f \n\t" 870 " nop \n\t" 871 "2: \n\t" 872#if !defined(MIPS_DSP_R1_LE) 873 "addiu %[r4], $zero, 1 \n\t" 874 "addiu %[r5], %[max1], -1 \n\t" 875 "sllv %[r4], %[r4], %[r5] \n\t" 876#endif // #if !defined(MIPS_DSP_R1_LE) 877 "3: \n\t" 878 "lw %[r0], 0(%[outre1]) \n\t" 879 "lw %[r1], 0(%[outre2]) \n\t" 880 "lw %[r2], 4(%[outre1]) \n\t" 881 "lw %[r3], 4(%[outre2]) \n\t" 882#if defined(MIPS_DSP_R1_LE) 883 "shrav_r.w %[r0], %[r0], %[max1] \n\t" 884 "shrav_r.w %[r1], %[r1], %[max1] \n\t" 885 "shrav_r.w %[r2], %[r2], %[max1] \n\t" 886 "shrav_r.w %[r3], %[r3], %[max1] \n\t" 887#else // #if defined(MIPS_DSP_R1_LE) 888 "addu %[r0], %[r0], %[r4] \n\t" 889 "addu %[r1], %[r1], %[r4] \n\t" 890 "addu %[r2], %[r2], %[r4] \n\t" 891 "addu %[r3], %[r3], %[r4] \n\t" 892 "srav %[r0], %[r0], %[max1] \n\t" 893 "srav %[r1], %[r1], %[max1] \n\t" 894 "srav %[r2], %[r2], %[max1] \n\t" 895 "srav %[r3], %[r3], %[max1] \n\t" 896#endif // #if defined(MIPS_DSP_R1_LE) 897 "addiu %[outre1], %[outre1], 8 \n\t" 898 "addiu %[outre2], %[outre2], 8 \n\t" 899 "sh %[r0], 0(%[inre]) \n\t" 900 "sh %[r1], 0(%[inim]) \n\t" 901 "sh %[r2], 2(%[inre]) \n\t" 902 "sh %[r3], 2(%[inim]) \n\t" 903 "addiu %[k], %[k], -1 \n\t" 904 "addiu %[inre], %[inre], 4 \n\t" 905 "bgtz %[k], 3b \n\t" 906 " addiu %[inim], %[inim], 4 \n\t" 907 "4: \n\t" 908 ".set pop \n\t" 909 : [k] "+r" (k), [max1] "=&r" (max1), [r0] "=&r" (r0), 910 [inre] "=&r" (inre), [inim] "=&r" (inim), 911 [outre1] "=&r" (outre1), [outre2] "=&r" (outre2), 912#if !defined(MIPS_DSP_R1_LE) 913 [r4] "=&r" (r4), [r5] "=&r" (r5), 914#endif // #if !defined(MIPS_DSP_R1_LE) 915 [r1] "=&r" (r1), [r2] "=&r" (r2), [r3] "=&r" (r3) 916 : [max] "r" (max), [inreQ7] "r" (inreQ7), 917 [inimQ7] "r" (inimQ7), [outre1Q16] "r" (outre1Q16), 918 [outre2Q16] "r" (outre2Q16) 919 : "memory" 920 ); 921 922 WebRtcIsacfix_FftRadix16Fastest(inreQ7, inimQ7, 1); // real call 923 924 // All the remaining processing is done inside a single loop to avoid 925 // unnecessary memory accesses. MIPS DSPr2 version processes two samples 926 // at a time. 927 cosptr = (int16_t*)WebRtcIsacfix_kCosTab1; 928 sinptr = (int16_t*)WebRtcIsacfix_kSinTab1; 929 k = FRAMESAMPLES / 2; 930 __asm __volatile ( 931 ".set push \n\t" 932 ".set noreorder \n\t" 933 "addiu %[inre], %[inreQ7], 0 \n\t" 934 "addiu %[inim], %[inimQ7], 0 \n\t" 935 "addiu %[outre1], %[outre1Q16], 0 \n\t" 936 "addiu %[outre2], %[outre2Q16], 0 \n\t" 937 "addiu %[r4], $zero, 273 \n\t" 938 "addiu %[r5], $zero, 31727 \n\t" 939#if defined(MIPS_DSP_R2_LE) 940 "addiu %[max], %[max], 16 \n\t" 941 "replv.ph %[r4], %[r4] \n\t" 942#endif // #if defined(MIPS_DSP_R2_LE) 943 "bltz %[max], 2f \n\t" 944 " subu %[max1], $zero, %[max] \n\t" 945#if defined(MIPS_DSP_R2_LE) 946 "addiu %[max], %[max], 1 \n\t" 947#endif // #if defined(MIPS_DSP_R2_LE) 948 "1: \n\t" 949#if defined(MIPS_DSP_R2_LE) 950 "lwl %[r0], 0(%[inre]) \n\t" 951 "lwl %[r1], 0(%[inim]) \n\t" 952 "lh %[r2], 0(%[cosptr]) \n\t" 953 "lwr %[r0], 0(%[inre]) \n\t" 954 "lwr %[r1], 0(%[inim]) \n\t" 955 "lh %[r3], 0(%[sinptr]) \n\t" 956 "muleq_s.w.phr %[r6], %[r0], %[r4] \n\t" 957 "muleq_s.w.phr %[r7], %[r1], %[r4] \n\t" 958 "muleq_s.w.phl %[r0], %[r0], %[r4] \n\t" 959 "muleq_s.w.phl %[r1], %[r1], %[r4] \n\t" 960 "addiu %[k], %[k], -2 \n\t" 961 "addiu %[inre], %[inre], 4 \n\t" 962 "addiu %[inim], %[inim], 4 \n\t" 963 "shrav_r.w %[r6], %[r6], %[max] \n\t" 964 "shrav_r.w %[r7], %[r7], %[max] \n\t" 965 "mult $ac0, %[r2], %[r6] \n\t" 966 "mult $ac1, %[r3], %[r7] \n\t" 967 "mult $ac2, %[r2], %[r7] \n\t" 968 "mult $ac3, %[r3], %[r6] \n\t" 969 "lh %[r2], 2(%[cosptr]) \n\t" 970 "lh %[r3], 2(%[sinptr]) \n\t" 971 "extr_r.w %[r6], $ac0, 14 \n\t" 972 "extr_r.w %[r7], $ac1, 14 \n\t" 973 "extr_r.w %[r8], $ac2, 14 \n\t" 974 "extr_r.w %[r9], $ac3, 14 \n\t" 975 "shrav_r.w %[r0], %[r0], %[max] \n\t" 976 "shrav_r.w %[r1], %[r1], %[max] \n\t" 977 "mult $ac0, %[r2], %[r0] \n\t" 978 "mult $ac1, %[r3], %[r1] \n\t" 979 "mult $ac2, %[r2], %[r1] \n\t" 980 "mult $ac3, %[r3], %[r0] \n\t" 981 "addiu %[cosptr], %[cosptr], 4 \n\t" 982 "extr_r.w %[r0], $ac0, 14 \n\t" 983 "extr_r.w %[r1], $ac1, 14 \n\t" 984 "extr_r.w %[r2], $ac2, 14 \n\t" 985 "extr_r.w %[r3], $ac3, 14 \n\t" 986 "subu %[r6], %[r6], %[r7] \n\t" 987 "addu %[r8], %[r8], %[r9] \n\t" 988 "mult $ac0, %[r5], %[r6] \n\t" 989 "mult $ac1, %[r5], %[r8] \n\t" 990 "addiu %[sinptr], %[sinptr], 4 \n\t" 991 "subu %[r0], %[r0], %[r1] \n\t" 992 "addu %[r2], %[r2], %[r3] \n\t" 993 "extr_r.w %[r1], $ac0, 11 \n\t" 994 "extr_r.w %[r3], $ac1, 11 \n\t" 995 "mult $ac2, %[r5], %[r0] \n\t" 996 "mult $ac3, %[r5], %[r2] \n\t" 997 "sw %[r1], 0(%[outre1]) \n\t" 998 "sw %[r3], 0(%[outre2]) \n\t" 999 "addiu %[outre1], %[outre1], 8 \n\t" 1000 "extr_r.w %[r0], $ac2, 11 \n\t" 1001 "extr_r.w %[r2], $ac3, 11 \n\t" 1002 "sw %[r0], -4(%[outre1]) \n\t" 1003 "sw %[r2], 4(%[outre2]) \n\t" 1004 "bgtz %[k], 1b \n\t" 1005 " addiu %[outre2], %[outre2], 8 \n\t" 1006 "b 3f \n\t" 1007#else // #if defined(MIPS_DSP_R2_LE) 1008 "lh %[r0], 0(%[inre]) \n\t" 1009 "lh %[r1], 0(%[inim]) \n\t" 1010 "addiu %[k], %[k], -1 \n\t" 1011 "srav %[r0], %[r0], %[max] \n\t" 1012 "srav %[r1], %[r1], %[max] \n\t" 1013 "sra %[r2], %[r0], 16 \n\t" 1014 "andi %[r0], %[r0], 0xFFFF \n\t" 1015 "sra %[r0], %[r0], 1 \n\t" 1016 "sra %[r3], %[r1], 16 \n\t" 1017 "andi %[r1], %[r1], 0xFFFF \n\t" 1018 "sra %[r1], %[r1], 1 \n\t" 1019 "mul %[r2], %[r2], %[r4] \n\t" 1020 "mul %[r0], %[r0], %[r4] \n\t" 1021 "mul %[r3], %[r3], %[r4] \n\t" 1022 "mul %[r1], %[r1], %[r4] \n\t" 1023 "addiu %[inre], %[inre], 2 \n\t" 1024 "addiu %[inim], %[inim], 2 \n\t" 1025 "lh %[r6], 0(%[cosptr]) \n\t" 1026 "lh %[r7], 0(%[sinptr]) \n\t" 1027#if defined(MIPS_DSP_R1_LE) 1028 "shra_r.w %[r0], %[r0], 15 \n\t" 1029 "shra_r.w %[r1], %[r1], 15 \n\t" 1030#else // #if defined(MIPS_DSP_R1_LE) 1031 "addiu %[r0], %[r0], 0x4000 \n\t" 1032 "addiu %[r1], %[r1], 0x4000 \n\t" 1033 "sra %[r0], %[r0], 15 \n\t" 1034 "sra %[r1], %[r1], 15 \n\t" 1035#endif // #if defined(MIPS_DSP_R1_LE) 1036 "addu %[r0], %[r2], %[r0] \n\t" 1037 "addu %[r1], %[r3], %[r1] \n\t" 1038 "sra %[r2], %[r0], 16 \n\t" 1039 "andi %[r0], %[r0], 0xFFFF \n\t" 1040 "mul %[r9], %[r2], %[r6] \n\t" 1041 "mul %[r2], %[r2], %[r7] \n\t" 1042 "mul %[r8], %[r0], %[r6] \n\t" 1043 "mul %[r0], %[r0], %[r7] \n\t" 1044 "sra %[r3], %[r3], 16 \n\t" 1045 "andi %[r1], %[r1], 0xFFFF \n\t" 1046 "sll %[r9], %[r9], 2 \n\t" 1047 "sll %[r2], %[r2], 2 \n\t" 1048#if defined(MIPS_DSP_R1_LE) 1049 "shra_r.w %[r8], %[r8], 14 \n\t" 1050 "shra_r.w %[r0], %[r0], 14 \n\t" 1051#else // #if defined(MIPS_DSP_R1_LE) 1052 "addiu %[r8], %[r8], 0x2000 \n\t" 1053 "addiu %[r0], %[r0], 0x2000 \n\t" 1054 "sra %[r8], %[r8], 14 \n\t" 1055 "sra %[r0], %[r0], 14 \n\t" 1056#endif // #if defined(MIPS_DSP_R1_LE) 1057 "addu %[r9], %[r9], %[r8] \n\t" 1058 "addu %[r2], %[r2], %[r0] \n\t" 1059 "mul %[r0], %[r3], %[r6] \n\t" 1060 "mul %[r3], %[r3], %[r7] \n\t" 1061 "mul %[r8], %[r1], %[r6] \n\t" 1062 "mul %[r1], %[r1], %[r8] \n\t" 1063 "addiu %[cosptr], %[cosptr], 2 \n\t" 1064 "addiu %[sinptr], %[sinptr], 2 \n\t" 1065 "sll %[r0], %[r0], 2 \n\t" 1066 "sll %[r3], %[r3], 2 \n\t" 1067#if defined(MIPS_DSP_R1_LE) 1068 "shra_r.w %[r8], %[r8], 14 \n\t" 1069 "shra_r.w %[r1], %[r1], 14 \n\t" 1070#else // #if defined(MIPS_DSP_R1_LE) 1071 "addiu %[r8], %[r8], 0x2000 \n\t" 1072 "addiu %[r1], %[r1], 0x2000 \n\t" 1073 "sra %[r8], %[r8], 14 \n\t" 1074 "sra %[r1], %[r1], 14 \n\t" 1075#endif // #if defined(MIPS_DSP_R1_LE) 1076 "addu %[r0], %[r0], %[r8] \n\t" 1077 "addu %[r3], %[r3], %[r1] \n\t" 1078 "subu %[r9], %[r9], %[r3] \n\t" 1079 "addu %[r0], %[r0], %[r2] \n\t" 1080 "sra %[r1], %[r9], 16 \n\t" 1081 "andi %[r9], %[r9], 0xFFFF \n\t" 1082 "mul %[r1], %[r1], %[r5] \n\t" 1083 "mul %[r9], %[r9], %[r5] \n\t" 1084 "sra %[r2], %[r0], 16 \n\t" 1085 "andi %[r0], %[r0], 0xFFFF \n\t" 1086 "mul %[r2], %[r2], %[r5] \n\t" 1087 "mul %[r0], %[r0], %[r5] \n\t" 1088 "sll %[r1], %[r1], 5 \n\t" 1089#if defined(MIPS_DSP_R1_LE) 1090 "shra_r.w %[r9], %[r9], 11 \n\t" 1091#else // #if defined(MIPS_DSP_R1_LE) 1092 "addiu %[r9], %[r9], 0x400 \n\t" 1093 "sra %[r9], %[r9], 11 \n\t" 1094#endif // #if defined(MIPS_DSP_R1_LE) 1095 "addu %[r1], %[r1], %[r9] \n\t" 1096 "sll %[r2], %[r2], 5 \n\t" 1097#if defined(MIPS_DSP_R1_LE) 1098 "shra_r.w %[r0], %[r0], 11 \n\t" 1099#else // #if defined(MIPS_DSP_R1_LE) 1100 "addiu %[r0], %[r0], 0x400 \n\t" 1101 "sra %[r0], %[r0], 11 \n\t" 1102#endif // #if defined(MIPS_DSP_R1_LE) 1103 "addu %[r0], %[r0], %[r2] \n\t" 1104 "sw %[r1], 0(%[outre1]) \n\t" 1105 "addiu %[outre1], %[outre1], 4 \n\t" 1106 "sw %[r0], 0(%[outre2]) \n\t" 1107 "bgtz %[k], 1b \n\t" 1108 " addiu %[outre2], %[outre2], 4 \n\t" 1109 "b 3f \n\t" 1110 " nop \n\t" 1111#endif // #if defined(MIPS_DSP_R2_LE) 1112 "2: \n\t" 1113#if defined(MIPS_DSP_R2_LE) 1114 "addiu %[max1], %[max1], -1 \n\t" 1115 "21: \n\t" 1116 "lwl %[r0], 0(%[inre]) \n\t" 1117 "lwl %[r1], 0(%[inim]) \n\t" 1118 "lh %[r2], 0(%[cosptr]) \n\t" 1119 "lwr %[r0], 0(%[inre]) \n\t" 1120 "lwr %[r1], 0(%[inim]) \n\t" 1121 "lh %[r3], 0(%[sinptr]) \n\t" 1122 "muleq_s.w.phr %[r6], %[r0], %[r4] \n\t" 1123 "muleq_s.w.phr %[r7], %[r1], %[r4] \n\t" 1124 "muleq_s.w.phl %[r0], %[r0], %[r4] \n\t" 1125 "muleq_s.w.phl %[r1], %[r1], %[r4] \n\t" 1126 "addiu %[k], %[k], -2 \n\t" 1127 "addiu %[inre], %[inre], 4 \n\t" 1128 "addiu %[inim], %[inim], 4 \n\t" 1129 "sllv %[r6], %[r6], %[max1] \n\t" 1130 "sllv %[r7], %[r7], %[max1] \n\t" 1131 "mult $ac0, %[r2], %[r6] \n\t" 1132 "mult $ac1, %[r3], %[r7] \n\t" 1133 "mult $ac2, %[r2], %[r7] \n\t" 1134 "mult $ac3, %[r3], %[r6] \n\t" 1135 "lh %[r2], 2(%[cosptr]) \n\t" 1136 "lh %[r3], 2(%[sinptr]) \n\t" 1137 "extr_r.w %[r6], $ac0, 14 \n\t" 1138 "extr_r.w %[r7], $ac1, 14 \n\t" 1139 "extr_r.w %[r8], $ac2, 14 \n\t" 1140 "extr_r.w %[r9], $ac3, 14 \n\t" 1141 "sllv %[r0], %[r0], %[max1] \n\t" 1142 "sllv %[r1], %[r1], %[max1] \n\t" 1143 "mult $ac0, %[r2], %[r0] \n\t" 1144 "mult $ac1, %[r3], %[r1] \n\t" 1145 "mult $ac2, %[r2], %[r1] \n\t" 1146 "mult $ac3, %[r3], %[r0] \n\t" 1147 "addiu %[cosptr], %[cosptr], 4 \n\t" 1148 "extr_r.w %[r0], $ac0, 14 \n\t" 1149 "extr_r.w %[r1], $ac1, 14 \n\t" 1150 "extr_r.w %[r2], $ac2, 14 \n\t" 1151 "extr_r.w %[r3], $ac3, 14 \n\t" 1152 "subu %[r6], %[r6], %[r7] \n\t" 1153 "addu %[r8], %[r8], %[r9] \n\t" 1154 "mult $ac0, %[r5], %[r6] \n\t" 1155 "mult $ac1, %[r5], %[r8] \n\t" 1156 "addiu %[sinptr], %[sinptr], 4 \n\t" 1157 "subu %[r0], %[r0], %[r1] \n\t" 1158 "addu %[r2], %[r2], %[r3] \n\t" 1159 "extr_r.w %[r1], $ac0, 11 \n\t" 1160 "extr_r.w %[r3], $ac1, 11 \n\t" 1161 "mult $ac2, %[r5], %[r0] \n\t" 1162 "mult $ac3, %[r5], %[r2] \n\t" 1163 "sw %[r1], 0(%[outre1]) \n\t" 1164 "sw %[r3], 0(%[outre2]) \n\t" 1165 "addiu %[outre1], %[outre1], 8 \n\t" 1166 "extr_r.w %[r0], $ac2, 11 \n\t" 1167 "extr_r.w %[r2], $ac3, 11 \n\t" 1168 "sw %[r0], -4(%[outre1]) \n\t" 1169 "sw %[r2], 4(%[outre2]) \n\t" 1170 "bgtz %[k], 21b \n\t" 1171 " addiu %[outre2], %[outre2], 8 \n\t" 1172 "b 3f \n\t" 1173 " nop \n\t" 1174#else // #if defined(MIPS_DSP_R2_LE) 1175 "lh %[r0], 0(%[inre]) \n\t" 1176 "lh %[r1], 0(%[inim]) \n\t" 1177 "addiu %[k], %[k], -1 \n\t" 1178 "sllv %[r0], %[r0], %[max1] \n\t" 1179 "sllv %[r1], %[r1], %[max1] \n\t" 1180 "sra %[r2], %[r0], 16 \n\t" 1181 "andi %[r0], %[r0], 0xFFFF \n\t" 1182 "sra %[r0], %[r0], 1 \n\t" 1183 "sra %[r3], %[r1], 16 \n\t" 1184 "andi %[r1], %[r1], 0xFFFF \n\t" 1185 "sra %[r1], %[r1], 1 \n\t" 1186 "mul %[r2], %[r2], %[r4] \n\t" 1187 "mul %[r0], %[r0], %[r4] \n\t" 1188 "mul %[r3], %[r3], %[r4] \n\t" 1189 "mul %[r1], %[r1], %[r4] \n\t" 1190 "addiu %[inre], %[inre], 2 \n\t" 1191 "addiu %[inim], %[inim], 2 \n\t" 1192 "lh %[r6], 0(%[cosptr]) \n\t" 1193 "lh %[r7], 0(%[sinptr]) \n\t" 1194#if defined(MIPS_DSP_R1_LE) 1195 "shra_r.w %[r0], %[r0], 15 \n\t" 1196 "shra_r.w %[r1], %[r1], 15 \n\t" 1197#else // #if defined(MIPS_DSP_R1_LE) 1198 "addiu %[r0], %[r0], 0x4000 \n\t" 1199 "addiu %[r1], %[r1], 0x4000 \n\t" 1200 "sra %[r0], %[r0], 15 \n\t" 1201 "sra %[r1], %[r1], 15 \n\t" 1202#endif // #if defined(MIPS_DSP_R1_LE) 1203 "addu %[r0], %[r2], %[r0] \n\t" 1204 "addu %[r1], %[r3], %[r1] \n\t" 1205 "sra %[r2], %[r0], 16 \n\t" 1206 "andi %[r0], %[r0], 0xFFFF \n\t" 1207 "mul %[r9], %[r2], %[r6] \n\t" 1208 "mul %[r2], %[r2], %[r7] \n\t" 1209 "mul %[r8], %[r0], %[r6] \n\t" 1210 "mul %[r0], %[r0], %[r7] \n\t" 1211 "sra %[r3], %[r1], 16 \n\t" 1212 "andi %[r1], %[r1], 0xFFFF \n\t" 1213 "sll %[r9], %[r9], 2 \n\t" 1214 "sll %[r2], %[r2], 2 \n\t" 1215#if defined(MIPS_DSP_R1_LE) 1216 "shra_r.w %[r8], %[r8], 14 \n\t" 1217 "shra_r.w %[r0], %[r0], 14 \n\t" 1218#else // #if defined(MIPS_DSP_R1_LE) 1219 "addiu %[r8], %[r8], 0x2000 \n\t" 1220 "addiu %[r0], %[r0], 0x2000 \n\t" 1221 "sra %[r8], %[r8], 14 \n\t" 1222 "sra %[r0], %[r0], 14 \n\t" 1223#endif // #if defined(MIPS_DSP_R1_LE) 1224 "addu %[r9], %[r9], %[r8] \n\t" 1225 "addu %[r2], %[r2], %[r0] \n\t" 1226 "mul %[r0], %[r3], %[r6] \n\t" 1227 "mul %[r3], %[r3], %[r7] \n\t" 1228 "mul %[r8], %[r1], %[r6] \n\t" 1229 "mul %[r1], %[r1], %[r7] \n\t" 1230 "addiu %[cosptr], %[cosptr], 2 \n\t" 1231 "addiu %[sinptr], %[sinptr], 2 \n\t" 1232 "sll %[r0], %[r0], 2 \n\t" 1233 "sll %[r3], %[r3], 2 \n\t" 1234#if defined(MIPS_DSP_R1_LE) 1235 "shra_r.w %[r8], %[r8], 14 \n\t" 1236 "shra_r.w %[r1], %[r1], 14 \n\t" 1237#else // #if defined(MIPS_DSP_R1_LE) 1238 "addiu %[r8], %[r8], 0x2000 \n\t" 1239 "addiu %[r1], %[r1], 0x2000 \n\t" 1240 "sra %[r8], %[r8], 14 \n\t" 1241 "sra %[r1], %[r1], 14 \n\t" 1242#endif // #if defined(MIPS_DSP_R1_LE) 1243 "addu %[r0], %[r0], %[r8] \n\t" 1244 "addu %[r3], %[r3], %[r1] \n\t" 1245 "subu %[r9], %[r9], %[r3] \n\t" 1246 "addu %[r0], %[r0], %[r2] \n\t" 1247 "sra %[r1], %[r9], 16 \n\t" 1248 "andi %[r9], %[r9], 0xFFFF \n\t" 1249 "mul %[r1], %[r1], %[r5] \n\t" 1250 "mul %[r9], %[r9], %[r5] \n\t" 1251 "sra %[r2], %[r0], 16 \n\t" 1252 "andi %[r0], %[r0], 0xFFFF \n\t" 1253 "mul %[r2], %[r2], %[r5] \n\t" 1254 "mul %[r0], %[r0], %[r5] \n\t" 1255 "sll %[r1], %[r1], 5 \n\t" 1256#if defined(MIPS_DSP_R1_LE) 1257 "shra_r.w %[r9], %[r9], 11 \n\t" 1258#else // #if defined(MIPS_DSP_R1_LE) 1259 "addiu %[r9], %[r9], 0x400 \n\t" 1260 "sra %[r9], %[r9], 11 \n\t" 1261#endif // #if defined(MIPS_DSP_R1_LE) 1262 "addu %[r1], %[r1], %[r9] \n\t" 1263 "sll %[r2], %[r2], 5 \n\t" 1264#if defined(MIPS_DSP_R1_LE) 1265 "shra_r.w %[r0], %[r0], 11 \n\t" 1266#else // #if defined(MIPS_DSP_R1_LE) 1267 "addiu %[r0], %[r0], 0x400 \n\t" 1268 "sra %[r0], %[r0], 11 \n\t" 1269#endif // #if defined(MIPS_DSP_R1_LE) 1270 "addu %[r0], %[r0], %[r2] \n\t" 1271 "sw %[r1], 0(%[outre1]) \n\t" 1272 "addiu %[outre1], %[outre1], 4 \n\t" 1273 "sw %[r0], 0(%[outre2]) \n\t" 1274 "bgtz %[k], 2b \n\t" 1275 " addiu %[outre2], %[outre2], 4 \n\t" 1276#endif // #if defined(MIPS_DSP_R2_LE) 1277 "3: \n\t" 1278 ".set pop \n\t" 1279 : [k] "+r" (k), [r0] "=&r" (r0), [r1] "=&r" (r1), 1280 [r2] "=&r" (r2), [r3] "=&r" (r3), [r4] "=&r" (r4), 1281 [r5] "=&r" (r5), [r6] "=&r" (r6), [r7] "=&r" (r7), 1282 [r8] "=&r" (r8), [r9] "=&r" (r9), [max1] "=&r" (max1), 1283 [inre] "=&r" (inre), [inim] "=&r" (inim), 1284 [outre1] "=&r" (outre1), [outre2] "=&r" (outre2) 1285 : [max] "r" (max), [inreQ7] "r" (inreQ7), 1286 [inimQ7] "r" (inimQ7), [cosptr] "r" (cosptr), 1287 [sinptr] "r" (sinptr), [outre1Q16] "r" (outre1Q16), 1288 [outre2Q16] "r" (outre2Q16) 1289 : "hi", "lo", "memory" 1290#if defined(MIPS_DSP_R2_LE) 1291 , "$ac1hi", "$ac1lo", "$ac2hi", "$ac2lo", "$ac3hi", "$ac3lo" 1292#endif // #if defined(MIPS_DSP_R2_LE) 1293 ); 1294} 1295