apps/codecs/libtremor/asm_arm.h

   1 /********************************************************************
   2  *                                                                  *
   3  * THIS FILE IS PART OF THE OggVorbis 'TREMOR' CODEC SOURCE CODE.   *
   4  *                                                                  *
   5  * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
   6  * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
   7  * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
   8  *                                                                  *
   9  * THE OggVorbis 'TREMOR' SOURCE CODE IS (C) COPYRIGHT 1994-2002    *
  10  * BY THE Xiph.Org FOUNDATION http://www.xiph.org/                  *
  11  *                                                                  *
  12  ********************************************************************
  13
  14  function: arm7 and later wide math functions
  15
  16  ********************************************************************/
  17
  18 #ifdef _ARM_ASSEM_
  19
  20 #if !defined(_V_WIDE_MATH) && !defined(_LOW_ACCURACY_)
  21 #define _V_WIDE_MATH
  22
  23 static inline ogg_int32_t MULT32(ogg_int32_t x, ogg_int32_t y) {
  24   int lo,hi;
  25   asm volatile("smull\t%0, %1, %2, %3"
  26                : "=&r"(lo),"=&r"(hi)
  27                : "%r"(x),"r"(y) );
  28   return(hi);
  29 }
  30
  31 static inline ogg_int32_t MULT31(ogg_int32_t x, ogg_int32_t y) {
  32   return MULT32(x,y)<<1;
  33 }
  34
  35 static inline ogg_int32_t MULT31_SHIFT15(ogg_int32_t x, ogg_int32_t y) {
  36   int lo,hi;
  37   asm volatile("smull   %0, %1, %2, %3\n\t"
  38                "movs    %0, %0, lsr #15\n\t"
  39                "adc     %1, %0, %1, lsl #17\n\t"
  40                : "=&r"(lo),"=&r"(hi)
  41                : "%r"(x),"r"(y)
  42                : "cc");
  43   return(hi);
  44 }
  45
  46 #ifndef _V_VECT_OPS
  47 #define _V_VECT_OPS
  48
  49 /* asm versions of vector operations for block.c, window.c */
  50 /* SOME IMPORTANT NOTES: this implementation of vect_mult_bw does
  51    NOT do a final shift, meaning that the result of vect_mult_bw is
  52    only 31 bits not 32.  This is so that we can do the shift in-place
  53    in vect_add_xxxx instead to save one instruction for each mult on arm */
  54 static inline
  55 void vect_add_right_left(ogg_int32_t *x, const ogg_int32_t *y, int n)
  56 {
  57   /* first arg is right subframe of previous frame and second arg
  58      is left subframe of current frame.  overlap left onto right overwriting
  59      the right subframe */
  60
  61   do{
  62     asm volatile (
  63                   "ldmia %[x], {r0, r1, r2, r3};"
  64                   "ldmia %[y]!, {r4, r5, r6, r7};"
  65                   "add r0, r4, r0, lsl #1;"
  66                   "add r1, r5, r1, lsl #1;"
  67                   "add r2, r6, r2, lsl #1;"
  68                   "add r3, r7, r3, lsl #1;"
  69                   "stmia %[x]!, {r0, r1, r2, r3};"
  70                   "ldmia %[x], {r0, r1, r2, r3};"
  71                   "ldmia %[y]!, {r4, r5, r6, r7};"
  72                   "add r0, r4, r0, lsl #1;"
  73                   "add r1, r5, r1, lsl #1;"
  74                   "add r2, r6, r2, lsl #1;"
  75                   "add r3, r7, r3, lsl #1;"
  76                   "stmia %[x]!, {r0, r1, r2, r3};"
  77                   : [x] "+r" (x), [y] "+r" (y)
  78                   : : "r0", "r1", "r2", "r3",
  79                   "r4", "r5", "r6", "r7",
  80                   "memory");
  81     n -= 8;
  82   } while (n);
  83 }
  84
  85 static inline
  86 void vect_add_left_right(ogg_int32_t *x, const ogg_int32_t *y, int n)
  87 {
  88   /* first arg is left subframe of current frame and second arg
  89      is right subframe of previous frame.  overlap right onto left overwriting
  90      the LEFT subframe */
  91   do{
  92     asm volatile (
  93                   "ldmia %[x], {r0, r1, r2, r3};"
  94                   "ldmia %[y]!, {r4, r5, r6, r7};"
  95                   "add r0, r0, r4, lsl #1;"
  96                   "add r1, r1, r5, lsl #1;"
  97                   "add r2, r2, r6, lsl #1;"
  98                   "add r3, r3, r7, lsl #1;"
  99                   "stmia %[x]!, {r0, r1, r2, r3};"
 100                   "ldmia %[x], {r0, r1, r2, r3};"
 101                   "ldmia %[y]!, {r4, r5, r6, r7};"
 102                   "add r0, r0, r4, lsl #1;"
 103                   "add r1, r1, r5, lsl #1;"
 104                   "add r2, r2, r6, lsl #1;"
 105                   "add r3, r3, r7, lsl #1;"
 106                   "stmia %[x]!, {r0, r1, r2, r3};"
 107                   : [x] "+r" (x), [y] "+r" (y)
 108                   : : "r0", "r1", "r2", "r3",
 109                   "r4", "r5", "r6", "r7",
 110                   "memory");
 111     n -= 8;
 112   } while (n);
 113 }
 114
 115 #if ARM_ARCH >= 6
 116 static inline
 117 void vect_mult_fw(ogg_int32_t *data, LOOKUP_T *window, int n)
 118 {
 119   /* Note, mult_fw uses MULT31 */
 120   do{
 121     asm volatile (
 122                   "ldmia %[d], {r0, r1, r2, r3};"
 123                   "ldmia %[w]!, {r4, r5, r6, r7};"
 124                   "smmul r0, r4, r0;"
 125                   "smmul r1, r5, r1;"
 126                   "smmul r2, r6, r2;"
 127                   "smmul r3, r7, r3;"
 128                   "mov   r0, r0, lsl #1;"
 129                   "mov   r1, r1, lsl #1;"
 130                   "mov   r2, r2, lsl #1;"
 131                   "mov   r3, r3, lsl #1;"
 132                   "stmia %[d]!, {r0, r1, r2, r3};"
 133                   : [d] "+r" (data), [w] "+r" (window)
 134                   : : "r0", "r1", "r2", "r3",
 135                   "r4", "r5", "r6", "r7",
 136                   "memory" );
 137     n -= 4;
 138   } while (n);
 139 }
 140 #else
 141 static inline
 142 void vect_mult_fw(ogg_int32_t *data, LOOKUP_T *window, int n)
 143 {
 144   /* Note, mult_fw uses MULT31 */
 145   do{
 146     asm volatile (
 147                   "ldmia %[d], {r0, r1, r2, r3};"
 148                   "ldmia %[w]!, {r4, r5, r6, r7};"
 149                   "smull r8, r0, r4, r0;"
 150                   "mov   r0, r0, lsl #1;"
 151                   "smull r8, r1, r5, r1;"
 152                   "mov   r1, r1, lsl #1;"
 153                   "smull r8, r2, r6, r2;"
 154                   "mov   r2, r2, lsl #1;"
 155                   "smull r8, r3, r7, r3;"
 156                   "mov   r3, r3, lsl #1;"
 157                   "stmia %[d]!, {r0, r1, r2, r3};"
 158                   : [d] "+r" (data), [w] "+r" (window)
 159                   : : "r0", "r1", "r2", "r3",
 160                   "r4", "r5", "r6", "r7", "r8",
 161                   "memory" );
 162     n -= 4;
 163   } while (n);
 164 }
 165 #endif
 166
 167 #if ARM_ARCH >= 6
 168 static inline
 169 void vect_mult_bw(ogg_int32_t *data, LOOKUP_T *window, int n)
 170 {
 171   /* NOTE mult_bw uses MULT_32 i.e. doesn't shift result left at end */
 172   /* On ARM, we can do the shift at the same time as the overlap-add */
 173   do{
 174     asm volatile ("ldmia %[d], {r0, r1, r2, r3};"
 175                   "ldmda %[w]!, {r4, r5, r6, r7};"
 176                   "smmul r0, r7, r0;"
 177                   "smmul r1, r6, r1;"
 178                   "smmul r2, r5, r2;"
 179                   "smmul r3, r4, r3;"
 180                   "stmia %[d]!, {r0, r1, r2, r3};"
 181                   : [d] "+r" (data), [w] "+r" (window)
 182                   : : "r0", "r1", "r2", "r3",
 183                   "r4", "r5", "r6", "r7",
 184                   "memory" );
 185     n -= 4;
 186   } while (n);
 187 }
 188 #else
 189 static inline
 190 void vect_mult_bw(ogg_int32_t *data, LOOKUP_T *window, int n)
 191 {
 192   /* NOTE mult_bw uses MULT_32 i.e. doesn't shift result left at end */
 193   /* On ARM, we can do the shift at the same time as the overlap-add */
 194   do{
 195     asm volatile ("ldmia %[d], {r0, r1, r2, r3};"
 196                   "ldmda %[w]!, {r4, r5, r6, r7};"
 197                   "smull r8, r0, r7, r0;"
 198                   "smull r7, r1, r6, r1;"
 199                   "smull r6, r2, r5, r2;"
 200                   "smull r5, r3, r4, r3;"
 201                   "stmia %[d]!, {r0, r1, r2, r3};"
 202                   : [d] "+r" (data), [w] "+r" (window)
 203                   : : "r0", "r1", "r2", "r3",
 204                   "r4", "r5", "r6", "r7", "r8",
 205                   "memory" );
 206     n -= 4;
 207   } while (n);
 208 }
 209 #endif
 210
 211 static inline void vect_copy(ogg_int32_t *x, const ogg_int32_t *y, int n)
 212 {
 213   memcpy(x,y,n*sizeof(ogg_int32_t));
 214 }
 215
 216 #endif
 217
 218 #endif
 219
 220 #ifndef _V_CLIP_MATH
 221 #define _V_CLIP_MATH
 222
 223 static inline ogg_int32_t CLIP_TO_15(ogg_int32_t x) {
 224   int tmp;
 225   asm volatile("subs    %1, %0, #32768\n\t"
 226                "movpl   %0, #0x7f00\n\t"
 227                "orrpl   %0, %0, #0xff\n"
 228                "adds    %1, %0, #32768\n\t"
 229                "movmi   %0, #0x8000"
 230                : "+r"(x),"=r"(tmp)
 231                :
 232                : "cc");
 233   return(x);
 234 }
 235
 236 #endif
 237
 238 #ifndef _V_LSP_MATH_ASM
 239 #define _V_LSP_MATH_ASM
 240
 241 static inline void lsp_loop_asm(ogg_uint32_t *qip,ogg_uint32_t *pip,
 242                                 ogg_int32_t *qexpp,
 243                                 ogg_int32_t *ilsp,ogg_int32_t wi,
 244                                 ogg_int32_t m){
 245
 246   ogg_uint32_t qi=*qip,pi=*pip;
 247   ogg_int32_t qexp=*qexpp;
 248
 249   asm("mov     r0,%3;"
 250       "mov     r1,%5,asr#1;"
 251       "add     r0,r0,r1,lsl#3;"
 252       "1:"
 253
 254       "ldmdb   r0!,{r1,r3};"
 255       "subs    r1,r1,%4;"          //ilsp[j]-wi
 256       "rsbmi   r1,r1,#0;"          //labs(ilsp[j]-wi)
 257       "umull   %0,r2,r1,%0;"       //qi*=labs(ilsp[j]-wi)
 258
 259       "subs    r1,r3,%4;"          //ilsp[j+1]-wi
 260       "rsbmi   r1,r1,#0;"          //labs(ilsp[j+1]-wi)
 261       "umull   %1,r3,r1,%1;"       //pi*=labs(ilsp[j+1]-wi)
 262
 263       "cmn     r2,r3;"             // shift down 16?
 264       "beq     0f;"
 265       "add     %2,%2,#16;"
 266       "mov     %0,%0,lsr #16;"
 267       "orr     %0,%0,r2,lsl #16;"
 268       "mov     %1,%1,lsr #16;"
 269       "orr     %1,%1,r3,lsl #16;"
 270       "0:"
 271       "cmp     r0,%3;\n"
 272       "bhi     1b;\n"
 273
 274       // odd filter assymetry
 275       "ands    r0,%5,#1;\n"
 276       "beq     2f;\n"
 277       "add     r0,%3,%5,lsl#2;\n"
 278
 279       "ldr     r1,[r0,#-4];\n"
 280       "mov     r0,#0x4000;\n"
 281
 282       "subs    r1,r1,%4;\n"          //ilsp[j]-wi
 283       "rsbmi   r1,r1,#0;\n"          //labs(ilsp[j]-wi)
 284       "umull   %0,r2,r1,%0;\n"       //qi*=labs(ilsp[j]-wi)
 285       "umull   %1,r3,r0,%1;\n"       //pi*=labs(ilsp[j+1]-wi)
 286
 287       "cmn     r2,r3;\n"             // shift down 16?
 288       "beq     2f;\n"
 289       "add     %2,%2,#16;\n"
 290       "mov     %0,%0,lsr #16;\n"
 291       "orr     %0,%0,r2,lsl #16;\n"
 292       "mov     %1,%1,lsr #16;\n"
 293       "orr     %1,%1,r3,lsl #16;\n"
 294
 295       //qi=(pi>>shift)*labs(ilsp[j]-wi);
 296       //pi=(qi>>shift)*labs(ilsp[j+1]-wi);
 297       //qexp+=shift;
 298
 299       //}
 300
 301       /* normalize to max 16 sig figs */
 302       "2:"
 303       "mov     r2,#0;"
 304       "orr     r1,%0,%1;"
 305       "tst     r1,#0xff000000;"
 306       "addne   r2,r2,#8;"
 307       "movne   r1,r1,lsr #8;"
 308       "tst     r1,#0x00f00000;"
 309       "addne   r2,r2,#4;"
 310       "movne   r1,r1,lsr #4;"
 311       "tst     r1,#0x000c0000;"
 312       "addne   r2,r2,#2;"
 313       "movne   r1,r1,lsr #2;"
 314       "tst     r1,#0x00020000;"
 315       "addne   r2,r2,#1;"
 316       "movne   r1,r1,lsr #1;"
 317       "tst     r1,#0x00010000;"
 318       "addne   r2,r2,#1;"
 319       "mov     %0,%0,lsr r2;"
 320       "mov     %1,%1,lsr r2;"
 321       "add     %2,%2,r2;"
 322
 323       : "+r"(qi),"+r"(pi),"+r"(qexp)
 324       : "r"(ilsp),"r"(wi),"r"(m)
 325       : "r0","r1","r2","r3","cc");
 326
 327   *qip=qi;
 328   *pip=pi;
 329   *qexpp=qexp;
 330 }
 331
 332 static inline void lsp_norm_asm(ogg_uint32_t *qip,ogg_int32_t *qexpp){
 333
 334   ogg_uint32_t qi=*qip;
 335   ogg_int32_t qexp=*qexpp;
 336
 337   asm("tst     %0,#0x0000ff00;"
 338       "moveq   %0,%0,lsl #8;"
 339       "subeq   %1,%1,#8;"
 340       "tst     %0,#0x0000f000;"
 341       "moveq   %0,%0,lsl #4;"
 342       "subeq   %1,%1,#4;"
 343       "tst     %0,#0x0000c000;"
 344       "moveq   %0,%0,lsl #2;"
 345       "subeq   %1,%1,#2;"
 346       "tst     %0,#0x00008000;"
 347       "moveq   %0,%0,lsl #1;"
 348       "subeq   %1,%1,#1;"
 349       : "+r"(qi),"+r"(qexp)
 350       :
 351       : "cc");
 352   *qip=qi;
 353   *qexpp=qexp;
 354 }
 355
 356 #endif
 357 #endif
 358