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                : "cc");
  29   return(hi);
  30 }
  31
  32 static inline ogg_int32_t MULT31(ogg_int32_t x, ogg_int32_t y) {
  33   return MULT32(x,y)<<1;
  34 }
  35
  36 static inline ogg_int32_t MULT31_SHIFT15(ogg_int32_t x, ogg_int32_t y) {
  37   int lo,hi;
  38   asm volatile("smull   %0, %1, %2, %3\n\t"
  39                "movs    %0, %0, lsr #15\n\t"
  40                "adc     %1, %0, %1, lsl #17\n\t"
  41                : "=&r"(lo),"=&r"(hi)
  42                : "%r"(x),"r"(y)
  43                : "cc");
  44   return(hi);
  45 }
  46
  47 #define MB() asm volatile ("" : : : "memory")
  48
  49 #define XPROD32(a, b, t, v, x, y) \
  50 { \
  51   long l; \
  52   asm(  "smull  %0, %1, %4, %6\n\t" \
  53         "smlal  %0, %1, %5, %7\n\t" \
  54         "rsb    %3, %4, #0\n\t" \
  55         "smull  %0, %2, %5, %6\n\t" \
  56         "smlal  %0, %2, %3, %7" \
  57         : "=&r" (l), "=&r" (x), "=&r" (y), "=r" ((a)) \
  58         : "3" ((a)), "r" ((b)), "r" ((t)), "r" ((v)) \
  59         : "cc" ); \
  60 }
  61
  62 static inline void XPROD31(ogg_int32_t  a, ogg_int32_t  b,
  63                            ogg_int32_t  t, ogg_int32_t  v,
  64                            ogg_int32_t *x, ogg_int32_t *y)
  65 {
  66   int x1, y1, l;
  67   asm(  "smull  %0, %1, %4, %6\n\t"
  68         "smlal  %0, %1, %5, %7\n\t"
  69         "rsb    %3, %4, #0\n\t"
  70         "smull  %0, %2, %5, %6\n\t"
  71         "smlal  %0, %2, %3, %7"
  72         : "=&r" (l), "=&r" (x1), "=&r" (y1), "=r" (a)
  73         : "3" (a), "r" (b), "r" (t), "r" (v)
  74         : "cc" );
  75   *x = x1 << 1;
  76   MB();
  77   *y = y1 << 1;
  78 }
  79
  80 static inline void XNPROD31(ogg_int32_t  a, ogg_int32_t  b,
  81                             ogg_int32_t  t, ogg_int32_t  v,
  82                             ogg_int32_t *x, ogg_int32_t *y)
  83 {
  84   int x1, y1, l;
  85   asm(  "rsb    %2, %4, #0\n\t"
  86         "smull  %0, %1, %3, %5\n\t"
  87         "smlal  %0, %1, %2, %6\n\t"
  88         "smull  %0, %2, %4, %5\n\t"
  89         "smlal  %0, %2, %3, %6"
  90         : "=&r" (l), "=&r" (x1), "=&r" (y1)
  91         : "r" (a), "r" (b), "r" (t), "r" (v)
  92         : "cc" );
  93   *x = x1 << 1;
  94   MB();
  95   *y = y1 << 1;
  96 }
  97
  98 #ifndef _V_VECT_OPS
  99 #define _V_VECT_OPS
 100
 101 /* asm versions of vector operations for block.c, window.c */
 102 /* SOME IMPORTANT NOTES: this implementation of vect_mult_bw does
 103    NOT do a final shift, meaning that the result of vect_mult_bw is
 104    only 31 bits not 32.  This is so that we can do the shift in-place
 105    in vect_add_xxxx instead to save one instruction for each mult on arm */
 106 static inline
 107 void vect_add_right_left(ogg_int32_t *x, const ogg_int32_t *y, int n)
 108 {
 109   /* first arg is right subframe of previous frame and second arg
 110      is left subframe of current frame.  overlap left onto right overwriting
 111      the right subframe */
 112
 113   do{
 114     asm volatile (
 115                   "ldmia %[x], {r0, r1, r2, r3};"
 116                   "ldmia %[y]!, {r4, r5, r6, r7};"
 117                   "add r0, r4, r0, lsl #1;"
 118                   "add r1, r5, r1, lsl #1;"
 119                   "add r2, r6, r2, lsl #1;"
 120                   "add r3, r7, r3, lsl #1;"
 121                   "stmia %[x]!, {r0, r1, r2, r3};"
 122                   "ldmia %[x], {r0, r1, r2, r3};"
 123                   "ldmia %[y]!, {r4, r5, r6, r7};"
 124                   "add r0, r4, r0, lsl #1;"
 125                   "add r1, r5, r1, lsl #1;"
 126                   "add r2, r6, r2, lsl #1;"
 127                   "add r3, r7, r3, lsl #1;"
 128                   "stmia %[x]!, {r0, r1, r2, r3};"
 129                   : [x] "+r" (x), [y] "+r" (y)
 130                   : : "r0", "r1", "r2", "r3",
 131                   "r4", "r5", "r6", "r7",
 132                   "memory");
 133     n -= 8;
 134   } while (n);
 135 }
 136
 137 static inline
 138 void vect_add_left_right(ogg_int32_t *x, const ogg_int32_t *y, int n)
 139 {
 140   /* first arg is left subframe of current frame and second arg
 141      is right subframe of previous frame.  overlap right onto left overwriting
 142      the LEFT subframe */
 143   do{
 144     asm volatile (
 145                   "ldmia %[x], {r0, r1, r2, r3};"
 146                   "ldmia %[y]!, {r4, r5, r6, r7};"
 147                   "add r0, r0, r4, lsl #1;"
 148                   "add r1, r1, r5, lsl #1;"
 149                   "add r2, r2, r6, lsl #1;"
 150                   "add r3, r3, r7, lsl #1;"
 151                   "stmia %[x]!, {r0, r1, r2, r3};"
 152                   "ldmia %[x], {r0, r1, r2, r3};"
 153                   "ldmia %[y]!, {r4, r5, r6, r7};"
 154                   "add r0, r0, r4, lsl #1;"
 155                   "add r1, r1, r5, lsl #1;"
 156                   "add r2, r2, r6, lsl #1;"
 157                   "add r3, r3, r7, lsl #1;"
 158                   "stmia %[x]!, {r0, r1, r2, r3};"
 159                   : [x] "+r" (x), [y] "+r" (y)
 160                   : : "r0", "r1", "r2", "r3",
 161                   "r4", "r5", "r6", "r7",
 162                   "memory");
 163     n -= 8;
 164   } while (n);
 165 }
 166
 167 static inline
 168 void vect_mult_fw(ogg_int32_t *data, LOOKUP_T *window, int n)
 169 {
 170   /* Note, mult_fw uses MULT31 */
 171   do{
 172     asm volatile (
 173                   "ldmia %[d], {r0, r1, r2, r3};"
 174                   "ldmia %[w]!, {r4, r5, r6, r7};"
 175                   "smull r8, r0, r4, r0;"
 176                   "mov   r0, r0, lsl #1;"
 177                   "smull r8, r1, r5, r1;"
 178                   "mov   r1, r1, lsl #1;"
 179                   "smull r8, r2, r6, r2;"
 180                   "mov   r2, r2, lsl #1;"
 181                   "smull r8, r3, r7, r3;"
 182                   "mov   r3, r3, lsl #1;"
 183                   "stmia %[d]!, {r0, r1, r2, r3};"
 184                   : [d] "+r" (data), [w] "+r" (window)
 185                   : : "r0", "r1", "r2", "r3",
 186                   "r4", "r5", "r6", "r7", "r8",
 187                   "memory", "cc");
 188     n -= 4;
 189   } while (n);
 190 }
 191
 192 static inline
 193 void vect_mult_bw(ogg_int32_t *data, LOOKUP_T *window, int n)
 194 {
 195   /* NOTE mult_bw uses MULT_32 i.e. doesn't shift result left at end */
 196   /* On ARM, we can do the shift at the same time as the overlap-add */
 197   do{
 198     asm volatile ("ldmia %[d], {r0, r1, r2, r3};"
 199                   "ldmda %[w]!, {r4, r5, r6, r7};"
 200                   "smull r8, r0, r7, r0;"
 201                   "smull r7, r1, r6, r1;"
 202                   "smull r6, r2, r5, r2;"
 203                   "smull r5, r3, r4, r3;"
 204                   "stmia %[d]!, {r0, r1, r2, r3};"
 205                   : [d] "+r" (data), [w] "+r" (window)
 206                   : : "r0", "r1", "r2", "r3",
 207                   "r4", "r5", "r6", "r7", "r8",
 208                   "memory", "cc");
 209     n -= 4;
 210   } while (n);
 211 }
 212
 213 static inline void vect_copy(ogg_int32_t *x, const ogg_int32_t *y, int n)
 214 {
 215   memcpy(x,y,n*sizeof(ogg_int32_t));
 216 }
 217
 218 #endif
 219
 220 #endif
 221
 222 #ifndef _V_CLIP_MATH
 223 #define _V_CLIP_MATH
 224
 225 static inline ogg_int32_t CLIP_TO_15(ogg_int32_t x) {
 226   int tmp;
 227   asm volatile("subs    %1, %0, #32768\n\t"
 228                "movpl   %0, #0x7f00\n\t"
 229                "orrpl   %0, %0, #0xff\n"
 230                "adds    %1, %0, #32768\n\t"
 231                "movmi   %0, #0x8000"
 232                : "+r"(x),"=r"(tmp)
 233                :
 234                : "cc");
 235   return(x);
 236 }
 237
 238 #endif
 239
 240 #ifndef _V_LSP_MATH_ASM
 241 #define _V_LSP_MATH_ASM
 242
 243 static inline void lsp_loop_asm(ogg_uint32_t *qip,ogg_uint32_t *pip,
 244                                 ogg_int32_t *qexpp,
 245                                 ogg_int32_t *ilsp,ogg_int32_t wi,
 246                                 ogg_int32_t m){
 247
 248   ogg_uint32_t qi=*qip,pi=*pip;
 249   ogg_int32_t qexp=*qexpp;
 250
 251   asm("mov     r0,%3;"
 252       "mov     r1,%5,asr#1;"
 253       "add     r0,r0,r1,lsl#3;"
 254       "1:"
 255
 256       "ldmdb   r0!,{r1,r3};"
 257       "subs    r1,r1,%4;"          //ilsp[j]-wi
 258       "rsbmi   r1,r1,#0;"          //labs(ilsp[j]-wi)
 259       "umull   %0,r2,r1,%0;"       //qi*=labs(ilsp[j]-wi)
 260
 261       "subs    r1,r3,%4;"          //ilsp[j+1]-wi
 262       "rsbmi   r1,r1,#0;"          //labs(ilsp[j+1]-wi)
 263       "umull   %1,r3,r1,%1;"       //pi*=labs(ilsp[j+1]-wi)
 264
 265       "cmn     r2,r3;"             // shift down 16?
 266       "beq     0f;"
 267       "add     %2,%2,#16;"
 268       "mov     %0,%0,lsr #16;"
 269       "orr     %0,%0,r2,lsl #16;"
 270       "mov     %1,%1,lsr #16;"
 271       "orr     %1,%1,r3,lsl #16;"
 272       "0:"
 273       "cmp     r0,%3;\n"
 274       "bhi     1b;\n"
 275
 276       // odd filter assymetry
 277       "ands    r0,%5,#1;\n"
 278       "beq     2f;\n"
 279       "add     r0,%3,%5,lsl#2;\n"
 280
 281       "ldr     r1,[r0,#-4];\n"
 282       "mov     r0,#0x4000;\n"
 283
 284       "subs    r1,r1,%4;\n"          //ilsp[j]-wi
 285       "rsbmi   r1,r1,#0;\n"          //labs(ilsp[j]-wi)
 286       "umull   %0,r2,r1,%0;\n"       //qi*=labs(ilsp[j]-wi)
 287       "umull   %1,r3,r0,%1;\n"       //pi*=labs(ilsp[j+1]-wi)
 288
 289       "cmn     r2,r3;\n"             // shift down 16?
 290       "beq     2f;\n"
 291       "add     %2,%2,#16;\n"
 292       "mov     %0,%0,lsr #16;\n"
 293       "orr     %0,%0,r2,lsl #16;\n"
 294       "mov     %1,%1,lsr #16;\n"
 295       "orr     %1,%1,r3,lsl #16;\n"
 296
 297       //qi=(pi>>shift)*labs(ilsp[j]-wi);
 298       //pi=(qi>>shift)*labs(ilsp[j+1]-wi);
 299       //qexp+=shift;
 300
 301       //}
 302
 303       /* normalize to max 16 sig figs */
 304       "2:"
 305       "mov     r2,#0;"
 306       "orr     r1,%0,%1;"
 307       "tst     r1,#0xff000000;"
 308       "addne   r2,r2,#8;"
 309       "movne   r1,r1,lsr #8;"
 310       "tst     r1,#0x00f00000;"
 311       "addne   r2,r2,#4;"
 312       "movne   r1,r1,lsr #4;"
 313       "tst     r1,#0x000c0000;"
 314       "addne   r2,r2,#2;"
 315       "movne   r1,r1,lsr #2;"
 316       "tst     r1,#0x00020000;"
 317       "addne   r2,r2,#1;"
 318       "movne   r1,r1,lsr #1;"
 319       "tst     r1,#0x00010000;"
 320       "addne   r2,r2,#1;"
 321       "mov     %0,%0,lsr r2;"
 322       "mov     %1,%1,lsr r2;"
 323       "add     %2,%2,r2;"
 324
 325       : "+r"(qi),"+r"(pi),"+r"(qexp)
 326       : "r"(ilsp),"r"(wi),"r"(m)
 327       : "r0","r1","r2","r3","cc");
 328
 329   *qip=qi;
 330   *pip=pi;
 331   *qexpp=qexp;
 332 }
 333
 334 static inline void lsp_norm_asm(ogg_uint32_t *qip,ogg_int32_t *qexpp){
 335
 336   ogg_uint32_t qi=*qip;
 337   ogg_int32_t qexp=*qexpp;
 338
 339   asm("tst     %0,#0x0000ff00;"
 340       "moveq   %0,%0,lsl #8;"
 341       "subeq   %1,%1,#8;"
 342       "tst     %0,#0x0000f000;"
 343       "moveq   %0,%0,lsl #4;"
 344       "subeq   %1,%1,#4;"
 345       "tst     %0,#0x0000c000;"
 346       "moveq   %0,%0,lsl #2;"
 347       "subeq   %1,%1,#2;"
 348       "tst     %0,#0x00008000;"
 349       "moveq   %0,%0,lsl #1;"
 350       "subeq   %1,%1,#1;"
 351       : "+r"(qi),"+r"(qexp)
 352       :
 353       : "cc");
 354   *qip=qi;
 355   *qexpp=qexp;
 356 }
 357
 358 #endif
 359 #endif
 360