rs_cl.c revision 54cd5d1771ea5c95e181befc66ef8e2a2c1b78cd
1#include "rs_core.rsh" 2 3extern float2 __attribute__((overloadable)) convert_float2(int2 c); 4extern float3 __attribute__((overloadable)) convert_float3(int3 c); 5extern float4 __attribute__((overloadable)) convert_float4(int4 c); 6 7extern int2 __attribute__((overloadable)) convert_int2(float2 c); 8extern int3 __attribute__((overloadable)) convert_int3(float3 c); 9extern int4 __attribute__((overloadable)) convert_int4(float4 c); 10 11 12extern float __attribute__((overloadable)) fmin(float v, float v2); 13extern float2 __attribute__((overloadable)) fmin(float2 v, float v2); 14extern float3 __attribute__((overloadable)) fmin(float3 v, float v2); 15extern float4 __attribute__((overloadable)) fmin(float4 v, float v2); 16 17extern float __attribute__((overloadable)) fmax(float v, float v2); 18extern float2 __attribute__((overloadable)) fmax(float2 v, float v2); 19extern float3 __attribute__((overloadable)) fmax(float3 v, float v2); 20extern float4 __attribute__((overloadable)) fmax(float4 v, float v2); 21 22// Float ops, 6.11.2 23 24#define FN_FUNC_FN(fnc) \ 25extern float2 __attribute__((overloadable)) fnc(float2 v) { \ 26 float2 r; \ 27 r.x = fnc(v.x); \ 28 r.y = fnc(v.y); \ 29 return r; \ 30} \ 31extern float3 __attribute__((overloadable)) fnc(float3 v) { \ 32 float3 r; \ 33 r.x = fnc(v.x); \ 34 r.y = fnc(v.y); \ 35 r.z = fnc(v.z); \ 36 return r; \ 37} \ 38extern float4 __attribute__((overloadable)) fnc(float4 v) { \ 39 float4 r; \ 40 r.x = fnc(v.x); \ 41 r.y = fnc(v.y); \ 42 r.z = fnc(v.z); \ 43 r.w = fnc(v.w); \ 44 return r; \ 45} 46 47#define IN_FUNC_FN(fnc) \ 48extern int2 __attribute__((overloadable)) fnc(float2 v) { \ 49 int2 r; \ 50 r.x = fnc(v.x); \ 51 r.y = fnc(v.y); \ 52 return r; \ 53} \ 54extern int3 __attribute__((overloadable)) fnc(float3 v) { \ 55 int3 r; \ 56 r.x = fnc(v.x); \ 57 r.y = fnc(v.y); \ 58 r.z = fnc(v.z); \ 59 return r; \ 60} \ 61extern int4 __attribute__((overloadable)) fnc(float4 v) { \ 62 int4 r; \ 63 r.x = fnc(v.x); \ 64 r.y = fnc(v.y); \ 65 r.z = fnc(v.z); \ 66 r.w = fnc(v.w); \ 67 return r; \ 68} 69 70#define FN_FUNC_FN_FN(fnc) \ 71extern float2 __attribute__((overloadable)) fnc(float2 v1, float2 v2) { \ 72 float2 r; \ 73 r.x = fnc(v1.x, v2.x); \ 74 r.y = fnc(v1.y, v2.y); \ 75 return r; \ 76} \ 77extern float3 __attribute__((overloadable)) fnc(float3 v1, float3 v2) { \ 78 float3 r; \ 79 r.x = fnc(v1.x, v2.x); \ 80 r.y = fnc(v1.y, v2.y); \ 81 r.z = fnc(v1.z, v2.z); \ 82 return r; \ 83} \ 84extern float4 __attribute__((overloadable)) fnc(float4 v1, float4 v2) { \ 85 float4 r; \ 86 r.x = fnc(v1.x, v2.x); \ 87 r.y = fnc(v1.y, v2.y); \ 88 r.z = fnc(v1.z, v2.z); \ 89 r.w = fnc(v1.w, v2.w); \ 90 return r; \ 91} 92 93#define FN_FUNC_FN_F(fnc) \ 94extern float2 __attribute__((overloadable)) fnc(float2 v1, float v2) { \ 95 float2 r; \ 96 r.x = fnc(v1.x, v2); \ 97 r.y = fnc(v1.y, v2); \ 98 return r; \ 99} \ 100extern float3 __attribute__((overloadable)) fnc(float3 v1, float v2) { \ 101 float3 r; \ 102 r.x = fnc(v1.x, v2); \ 103 r.y = fnc(v1.y, v2); \ 104 r.z = fnc(v1.z, v2); \ 105 return r; \ 106} \ 107extern float4 __attribute__((overloadable)) fnc(float4 v1, float v2) { \ 108 float4 r; \ 109 r.x = fnc(v1.x, v2); \ 110 r.y = fnc(v1.y, v2); \ 111 r.z = fnc(v1.z, v2); \ 112 r.w = fnc(v1.w, v2); \ 113 return r; \ 114} 115 116#define FN_FUNC_FN_IN(fnc) \ 117extern float2 __attribute__((overloadable)) fnc(float2 v1, int2 v2) { \ 118 float2 r; \ 119 r.x = fnc(v1.x, v2.x); \ 120 r.y = fnc(v1.y, v2.y); \ 121 return r; \ 122} \ 123extern float3 __attribute__((overloadable)) fnc(float3 v1, int3 v2) { \ 124 float3 r; \ 125 r.x = fnc(v1.x, v2.x); \ 126 r.y = fnc(v1.y, v2.y); \ 127 r.z = fnc(v1.z, v2.z); \ 128 return r; \ 129} \ 130extern float4 __attribute__((overloadable)) fnc(float4 v1, int4 v2) { \ 131 float4 r; \ 132 r.x = fnc(v1.x, v2.x); \ 133 r.y = fnc(v1.y, v2.y); \ 134 r.z = fnc(v1.z, v2.z); \ 135 r.w = fnc(v1.w, v2.w); \ 136 return r; \ 137} 138 139#define FN_FUNC_FN_I(fnc) \ 140extern float2 __attribute__((overloadable)) fnc(float2 v1, int v2) { \ 141 float2 r; \ 142 r.x = fnc(v1.x, v2); \ 143 r.y = fnc(v1.y, v2); \ 144 return r; \ 145} \ 146extern float3 __attribute__((overloadable)) fnc(float3 v1, int v2) { \ 147 float3 r; \ 148 r.x = fnc(v1.x, v2); \ 149 r.y = fnc(v1.y, v2); \ 150 r.z = fnc(v1.z, v2); \ 151 return r; \ 152} \ 153extern float4 __attribute__((overloadable)) fnc(float4 v1, int v2) { \ 154 float4 r; \ 155 r.x = fnc(v1.x, v2); \ 156 r.y = fnc(v1.y, v2); \ 157 r.z = fnc(v1.z, v2); \ 158 r.w = fnc(v1.w, v2); \ 159 return r; \ 160} 161 162#define FN_FUNC_FN_PFN(fnc) \ 163extern float2 __attribute__((overloadable)) \ 164 fnc(float2 v1, float2 *v2) { \ 165 float2 r; \ 166 float t[2]; \ 167 r.x = fnc(v1.x, &t[0]); \ 168 r.y = fnc(v1.y, &t[1]); \ 169 v2->x = t[0]; \ 170 v2->y = t[1]; \ 171 return r; \ 172} \ 173extern float3 __attribute__((overloadable)) \ 174 fnc(float3 v1, float3 *v2) { \ 175 float3 r; \ 176 float t[3]; \ 177 r.x = fnc(v1.x, &t[0]); \ 178 r.y = fnc(v1.y, &t[1]); \ 179 r.z = fnc(v1.z, &t[2]); \ 180 v2->x = t[0]; \ 181 v2->y = t[1]; \ 182 v2->z = t[2]; \ 183 return r; \ 184} \ 185extern float4 __attribute__((overloadable)) \ 186 fnc(float4 v1, float4 *v2) { \ 187 float4 r; \ 188 float t[4]; \ 189 r.x = fnc(v1.x, &t[0]); \ 190 r.y = fnc(v1.y, &t[1]); \ 191 r.z = fnc(v1.z, &t[2]); \ 192 r.w = fnc(v1.w, &t[3]); \ 193 v2->x = t[0]; \ 194 v2->y = t[1]; \ 195 v2->z = t[2]; \ 196 v2->w = t[3]; \ 197 return r; \ 198} 199 200#define FN_FUNC_FN_PIN(fnc) \ 201extern float2 __attribute__((overloadable)) fnc(float2 v1, int2 *v2) { \ 202 float2 r; \ 203 int t[2]; \ 204 r.x = fnc(v1.x, &t[0]); \ 205 r.y = fnc(v1.y, &t[1]); \ 206 v2->x = t[0]; \ 207 v2->y = t[1]; \ 208 return r; \ 209} \ 210extern float3 __attribute__((overloadable)) fnc(float3 v1, int3 *v2) { \ 211 float3 r; \ 212 int t[3]; \ 213 r.x = fnc(v1.x, &t[0]); \ 214 r.y = fnc(v1.y, &t[1]); \ 215 r.z = fnc(v1.z, &t[2]); \ 216 v2->x = t[0]; \ 217 v2->y = t[1]; \ 218 v2->z = t[2]; \ 219 return r; \ 220} \ 221extern float4 __attribute__((overloadable)) fnc(float4 v1, int4 *v2) { \ 222 float4 r; \ 223 int t[4]; \ 224 r.x = fnc(v1.x, &t[0]); \ 225 r.y = fnc(v1.y, &t[1]); \ 226 r.z = fnc(v1.z, &t[2]); \ 227 r.w = fnc(v1.w, &t[3]); \ 228 v2->x = t[0]; \ 229 v2->y = t[1]; \ 230 v2->z = t[2]; \ 231 v2->w = t[3]; \ 232 return r; \ 233} 234 235#define FN_FUNC_FN_FN_FN(fnc) \ 236extern float2 __attribute__((overloadable)) \ 237 fnc(float2 v1, float2 v2, float2 v3) { \ 238 float2 r; \ 239 r.x = fnc(v1.x, v2.x, v3.x); \ 240 r.y = fnc(v1.y, v2.y, v3.y); \ 241 return r; \ 242} \ 243extern float3 __attribute__((overloadable)) \ 244 fnc(float3 v1, float3 v2, float3 v3) { \ 245 float3 r; \ 246 r.x = fnc(v1.x, v2.x, v3.x); \ 247 r.y = fnc(v1.y, v2.y, v3.y); \ 248 r.z = fnc(v1.z, v2.z, v3.z); \ 249 return r; \ 250} \ 251extern float4 __attribute__((overloadable)) \ 252 fnc(float4 v1, float4 v2, float4 v3) { \ 253 float4 r; \ 254 r.x = fnc(v1.x, v2.x, v3.x); \ 255 r.y = fnc(v1.y, v2.y, v3.y); \ 256 r.z = fnc(v1.z, v2.z, v3.z); \ 257 r.w = fnc(v1.w, v2.w, v3.w); \ 258 return r; \ 259} 260 261#define FN_FUNC_FN_FN_PIN(fnc) \ 262extern float2 __attribute__((overloadable)) \ 263 fnc(float2 v1, float2 v2, int2 *v3) { \ 264 float2 r; \ 265 int t[2]; \ 266 r.x = fnc(v1.x, v2.x, &t[0]); \ 267 r.y = fnc(v1.y, v2.y, &t[1]); \ 268 v3->x = t[0]; \ 269 v3->y = t[1]; \ 270 return r; \ 271} \ 272extern float3 __attribute__((overloadable)) \ 273 fnc(float3 v1, float3 v2, int3 *v3) { \ 274 float3 r; \ 275 int t[3]; \ 276 r.x = fnc(v1.x, v2.x, &t[0]); \ 277 r.y = fnc(v1.y, v2.y, &t[1]); \ 278 r.z = fnc(v1.z, v2.z, &t[2]); \ 279 v3->x = t[0]; \ 280 v3->y = t[1]; \ 281 v3->z = t[2]; \ 282 return r; \ 283} \ 284extern float4 __attribute__((overloadable)) \ 285 fnc(float4 v1, float4 v2, int4 *v3) { \ 286 float4 r; \ 287 int t[4]; \ 288 r.x = fnc(v1.x, v2.x, &t[0]); \ 289 r.y = fnc(v1.y, v2.y, &t[1]); \ 290 r.z = fnc(v1.z, v2.z, &t[2]); \ 291 r.w = fnc(v1.w, v2.w, &t[3]); \ 292 v3->x = t[0]; \ 293 v3->y = t[1]; \ 294 v3->z = t[2]; \ 295 v3->w = t[3]; \ 296 return r; \ 297} 298 299static const int iposinf = 0x7f800000; 300static const int ineginf = 0xff800000; 301 302static const float posinf() { 303 float f = *((float*)&iposinf); 304 return f; 305} 306 307static const float neginf() { 308 float f = *((float*)&ineginf); 309 return f; 310} 311 312static bool isinf(float f) { 313 int i = *((int*)(void*)&f); 314 return (i == iposinf) || (i == ineginf); 315} 316 317static bool isnan(float f) { 318 int i = *((int*)(void*)&f); 319 return (((i & 0x7f800000) == 0x7f800000) && (i & 0x007fffff)); 320} 321 322static bool isposzero(float f) { 323 int i = *((int*)(void*)&f); 324 return (i == 0x00000000); 325} 326 327static bool isnegzero(float f) { 328 int i = *((int*)(void*)&f); 329 return (i == 0x80000000); 330} 331 332static bool iszero(float f) { 333 return isposzero(f) || isnegzero(f); 334} 335 336 337extern float __attribute__((overloadable)) acos(float); 338FN_FUNC_FN(acos) 339 340extern float __attribute__((overloadable)) acosh(float); 341FN_FUNC_FN(acosh) 342 343 344extern float __attribute__((overloadable)) acospi(float v) { 345 return acos(v) / M_PI; 346} 347FN_FUNC_FN(acospi) 348 349extern float __attribute__((overloadable)) asin(float); 350FN_FUNC_FN(asin) 351 352extern float __attribute__((overloadable)) asinh(float); 353FN_FUNC_FN(asinh) 354 355extern float __attribute__((overloadable)) asinpi(float v) { 356 return asin(v) / M_PI; 357} 358FN_FUNC_FN(asinpi) 359 360extern float __attribute__((overloadable)) atan(float); 361FN_FUNC_FN(atan) 362 363extern float __attribute__((overloadable)) atan2(float, float); 364FN_FUNC_FN_FN(atan2) 365 366extern float __attribute__((overloadable)) atanh(float); 367FN_FUNC_FN(atanh) 368 369extern float __attribute__((overloadable)) atanpi(float v) { 370 return atan(v) / M_PI; 371} 372FN_FUNC_FN(atanpi) 373 374 375extern float __attribute__((overloadable)) atan2pi(float y, float x) { 376 return atan2(y, x) / M_PI; 377} 378FN_FUNC_FN_FN(atan2pi) 379 380extern float __attribute__((overloadable)) cbrt(float); 381FN_FUNC_FN(cbrt) 382 383extern float __attribute__((overloadable)) ceil(float); 384FN_FUNC_FN(ceil) 385 386extern float __attribute__((overloadable)) copysign(float, float); 387FN_FUNC_FN_FN(copysign) 388 389extern float __attribute__((overloadable)) cos(float); 390FN_FUNC_FN(cos) 391 392extern float __attribute__((overloadable)) cosh(float); 393FN_FUNC_FN(cosh) 394 395extern float __attribute__((overloadable)) cospi(float v) { 396 return cos(v * M_PI); 397} 398FN_FUNC_FN(cospi) 399 400extern float __attribute__((overloadable)) erfc(float); 401FN_FUNC_FN(erfc) 402 403extern float __attribute__((overloadable)) erf(float); 404FN_FUNC_FN(erf) 405 406extern float __attribute__((overloadable)) exp(float); 407FN_FUNC_FN(exp) 408 409extern float __attribute__((overloadable)) exp2(float); 410FN_FUNC_FN(exp2) 411 412extern float __attribute__((overloadable)) pow(float, float); 413 414extern float __attribute__((overloadable)) exp10(float v) { 415 return exp2(v * 3.321928095f); 416} 417FN_FUNC_FN(exp10) 418 419extern float __attribute__((overloadable)) expm1(float); 420FN_FUNC_FN(expm1) 421 422extern float __attribute__((overloadable)) fabs(float v) { 423 int i = *((int*)(void*)&v) & 0x7fffffff; 424 return *((float*)(void*)&i); 425} 426FN_FUNC_FN(fabs) 427 428extern float __attribute__((overloadable)) fdim(float, float); 429FN_FUNC_FN_FN(fdim) 430 431extern float __attribute__((overloadable)) floor(float); 432FN_FUNC_FN(floor) 433 434extern float __attribute__((overloadable)) fma(float, float, float); 435FN_FUNC_FN_FN_FN(fma) 436 437extern float __attribute__((overloadable)) fmin(float, float); 438 439extern float __attribute__((overloadable)) fmod(float, float); 440FN_FUNC_FN_FN(fmod) 441 442extern float __attribute__((overloadable)) fract(float v, float *iptr) { 443 int i = (int)floor(v); 444 if (iptr) { 445 iptr[0] = i; 446 } 447 return fmin(v - i, 0x1.fffffep-1f); 448} 449FN_FUNC_FN_PFN(fract) 450 451extern float __attribute__((overloadable)) frexp(float, int *); 452FN_FUNC_FN_PIN(frexp) 453 454extern float __attribute__((overloadable)) hypot(float, float); 455FN_FUNC_FN_FN(hypot) 456 457extern int __attribute__((overloadable)) ilogb(float); 458IN_FUNC_FN(ilogb) 459 460extern float __attribute__((overloadable)) ldexp(float, int); 461FN_FUNC_FN_IN(ldexp) 462FN_FUNC_FN_I(ldexp) 463 464extern float __attribute__((overloadable)) lgamma(float); 465FN_FUNC_FN(lgamma) 466extern float __attribute__((overloadable)) lgamma(float, int*); 467FN_FUNC_FN_PIN(lgamma) 468 469extern float __attribute__((overloadable)) log(float); 470FN_FUNC_FN(log) 471 472extern float __attribute__((overloadable)) log10(float); 473FN_FUNC_FN(log10) 474 475 476extern float __attribute__((overloadable)) log2(float v) { 477 return log10(v) * 3.321928095f; 478} 479FN_FUNC_FN(log2) 480 481extern float __attribute__((overloadable)) log1p(float); 482FN_FUNC_FN(log1p) 483 484extern float __attribute__((overloadable)) logb(float); 485FN_FUNC_FN(logb) 486 487extern float __attribute__((overloadable)) mad(float a, float b, float c) { 488 return a * b + c; 489} 490extern float2 __attribute__((overloadable)) mad(float2 a, float2 b, float2 c) { 491 return a * b + c; 492} 493extern float3 __attribute__((overloadable)) mad(float3 a, float3 b, float3 c) { 494 return a * b + c; 495} 496extern float4 __attribute__((overloadable)) mad(float4 a, float4 b, float4 c) { 497 return a * b + c; 498} 499 500extern float __attribute__((overloadable)) modf(float, float *); 501FN_FUNC_FN_PFN(modf); 502 503extern float __attribute__((overloadable)) nan(uint v) { 504 float f[1]; 505 uint32_t *ip = (uint32_t *)f; 506 *ip = v | 0x7fc00000; 507 return f[0]; 508} 509 510extern float __attribute__((overloadable)) nextafter(float, float); 511FN_FUNC_FN_FN(nextafter) 512 513FN_FUNC_FN_FN(pow) 514 515extern float __attribute__((overloadable)) pown(float v, int p) { 516 /* The mantissa of a float has fewer bits than an int (24 effective vs. 31). 517 * For very large ints, we'll lose whether the exponent is even or odd, making 518 * the selection of a correct sign incorrect. We correct this. Use copysign 519 * to handle the negative zero case. 520 */ 521 float sign = (p & 0x1) ? copysign(1.f, v) : 1.f; 522 float f = pow(v, (float)p); 523 return copysign(f, sign); 524} 525FN_FUNC_FN_IN(pown) 526 527extern float __attribute__((overloadable)) powr(float v, float p) { 528 return pow(v, p); 529} 530extern float2 __attribute__((overloadable)) powr(float2 v, float2 p) { 531 return pow(v, p); 532} 533extern float3 __attribute__((overloadable)) powr(float3 v, float3 p) { 534 return pow(v, p); 535} 536extern float4 __attribute__((overloadable)) powr(float4 v, float4 p) { 537 return pow(v, p); 538} 539 540extern float __attribute__((overloadable)) remainder(float, float); 541FN_FUNC_FN_FN(remainder) 542 543extern float __attribute__((overloadable)) remquo(float, float, int *); 544FN_FUNC_FN_FN_PIN(remquo) 545 546extern float __attribute__((overloadable)) rint(float); 547FN_FUNC_FN(rint) 548 549extern float __attribute__((overloadable)) rootn(float v, int r) { 550 if (r == 0) { 551 return posinf(); 552 } 553 554 if (iszero(v)) { 555 if (r < 0) { 556 if (r & 1) { 557 return copysign(posinf(), v); 558 } else { 559 return posinf(); 560 } 561 } else { 562 if (r & 1) { 563 return copysign(0.f, v); 564 } else { 565 return 0.f; 566 } 567 } 568 } 569 570 if (!isinf(v) && !isnan(v) && (v < 0.f)) { 571 if (r & 1) { 572 return (-1.f * pow(-1.f * v, 1.f / r)); 573 } else { 574 return nan(0); 575 } 576 } 577 578 return pow(v, 1.f / r); 579} 580FN_FUNC_FN_IN(rootn); 581 582extern float __attribute__((overloadable)) round(float); 583FN_FUNC_FN(round) 584 585 586extern float __attribute__((overloadable)) sqrt(float); 587extern float __attribute__((overloadable)) rsqrt(float v) { 588 return 1.f / sqrt(v); 589} 590 591#if !defined(ARCH_X86_HAVE_SSSE3) || defined(RS_DEBUG_RUNTIME) 592// These functions must be defined here if we are not using the SSE 593// implementation, which includes when we are built as part of the 594// debug runtime (libclcore_debug.bc). 595FN_FUNC_FN(sqrt) 596#else 597extern float2 __attribute__((overloadable)) sqrt(float2); 598extern float3 __attribute__((overloadable)) sqrt(float3); 599extern float4 __attribute__((overloadable)) sqrt(float4); 600#endif // !defined(ARCH_X86_HAVE_SSSE3) || defined(RS_DEBUG_RUNTIME) 601 602FN_FUNC_FN(rsqrt) 603 604extern float __attribute__((overloadable)) sin(float); 605FN_FUNC_FN(sin) 606 607extern float __attribute__((overloadable)) sincos(float v, float *cosptr) { 608 *cosptr = cos(v); 609 return sin(v); 610} 611extern float2 __attribute__((overloadable)) sincos(float2 v, float2 *cosptr) { 612 *cosptr = cos(v); 613 return sin(v); 614} 615extern float3 __attribute__((overloadable)) sincos(float3 v, float3 *cosptr) { 616 *cosptr = cos(v); 617 return sin(v); 618} 619extern float4 __attribute__((overloadable)) sincos(float4 v, float4 *cosptr) { 620 *cosptr = cos(v); 621 return sin(v); 622} 623 624extern float __attribute__((overloadable)) sinh(float); 625FN_FUNC_FN(sinh) 626 627extern float __attribute__((overloadable)) sinpi(float v) { 628 return sin(v * M_PI); 629} 630FN_FUNC_FN(sinpi) 631 632extern float __attribute__((overloadable)) tan(float); 633FN_FUNC_FN(tan) 634 635extern float __attribute__((overloadable)) tanh(float); 636FN_FUNC_FN(tanh) 637 638extern float __attribute__((overloadable)) tanpi(float v) { 639 return tan(v * M_PI); 640} 641FN_FUNC_FN(tanpi) 642 643 644extern float __attribute__((overloadable)) tgamma(float); 645FN_FUNC_FN(tgamma) 646 647extern float __attribute__((overloadable)) trunc(float); 648FN_FUNC_FN(trunc) 649 650// Int ops (partial), 6.11.3 651 652#define XN_FUNC_YN(typeout, fnc, typein) \ 653extern typeout __attribute__((overloadable)) fnc(typein); \ 654extern typeout##2 __attribute__((overloadable)) fnc(typein##2 v) { \ 655 typeout##2 r; \ 656 r.x = fnc(v.x); \ 657 r.y = fnc(v.y); \ 658 return r; \ 659} \ 660extern typeout##3 __attribute__((overloadable)) fnc(typein##3 v) { \ 661 typeout##3 r; \ 662 r.x = fnc(v.x); \ 663 r.y = fnc(v.y); \ 664 r.z = fnc(v.z); \ 665 return r; \ 666} \ 667extern typeout##4 __attribute__((overloadable)) fnc(typein##4 v) { \ 668 typeout##4 r; \ 669 r.x = fnc(v.x); \ 670 r.y = fnc(v.y); \ 671 r.z = fnc(v.z); \ 672 r.w = fnc(v.w); \ 673 return r; \ 674} 675 676 677#define UIN_FUNC_IN(fnc) \ 678XN_FUNC_YN(uchar, fnc, char) \ 679XN_FUNC_YN(ushort, fnc, short) \ 680XN_FUNC_YN(uint, fnc, int) 681 682#define IN_FUNC_IN(fnc) \ 683XN_FUNC_YN(uchar, fnc, uchar) \ 684XN_FUNC_YN(char, fnc, char) \ 685XN_FUNC_YN(ushort, fnc, ushort) \ 686XN_FUNC_YN(short, fnc, short) \ 687XN_FUNC_YN(uint, fnc, uint) \ 688XN_FUNC_YN(int, fnc, int) 689 690 691#define XN_FUNC_XN_XN_BODY(type, fnc, body) \ 692extern type __attribute__((overloadable)) \ 693 fnc(type v1, type v2) { \ 694 return body; \ 695} \ 696extern type##2 __attribute__((overloadable)) \ 697 fnc(type##2 v1, type##2 v2) { \ 698 type##2 r; \ 699 r.x = fnc(v1.x, v2.x); \ 700 r.y = fnc(v1.y, v2.y); \ 701 return r; \ 702} \ 703extern type##3 __attribute__((overloadable)) \ 704 fnc(type##3 v1, type##3 v2) { \ 705 type##3 r; \ 706 r.x = fnc(v1.x, v2.x); \ 707 r.y = fnc(v1.y, v2.y); \ 708 r.z = fnc(v1.z, v2.z); \ 709 return r; \ 710} \ 711extern type##4 __attribute__((overloadable)) \ 712 fnc(type##4 v1, type##4 v2) { \ 713 type##4 r; \ 714 r.x = fnc(v1.x, v2.x); \ 715 r.y = fnc(v1.y, v2.y); \ 716 r.z = fnc(v1.z, v2.z); \ 717 r.w = fnc(v1.w, v2.w); \ 718 return r; \ 719} 720 721#define IN_FUNC_IN_IN_BODY(fnc, body) \ 722XN_FUNC_XN_XN_BODY(uchar, fnc, body) \ 723XN_FUNC_XN_XN_BODY(char, fnc, body) \ 724XN_FUNC_XN_XN_BODY(ushort, fnc, body) \ 725XN_FUNC_XN_XN_BODY(short, fnc, body) \ 726XN_FUNC_XN_XN_BODY(uint, fnc, body) \ 727XN_FUNC_XN_XN_BODY(int, fnc, body) \ 728XN_FUNC_XN_XN_BODY(float, fnc, body) 729 730 731/** 732 * abs 733 */ 734extern uint32_t __attribute__((overloadable)) abs(int32_t v) { 735 if (v < 0) 736 return -v; 737 return v; 738} 739extern uint16_t __attribute__((overloadable)) abs(int16_t v) { 740 if (v < 0) 741 return -v; 742 return v; 743} 744extern uint8_t __attribute__((overloadable)) abs(int8_t v) { 745 if (v < 0) 746 return -v; 747 return v; 748} 749 750/** 751 * clz 752 * __builtin_clz only accepts a 32-bit unsigned int, so every input will be 753 * expanded to 32 bits. For our smaller data types, we need to subtract off 754 * these unused top bits (that will be always be composed of zeros). 755 */ 756extern uint32_t __attribute__((overloadable)) clz(uint32_t v) { 757 return __builtin_clz(v); 758} 759extern uint16_t __attribute__((overloadable)) clz(uint16_t v) { 760 return __builtin_clz(v) - 16; 761} 762extern uint8_t __attribute__((overloadable)) clz(uint8_t v) { 763 return __builtin_clz(v) - 24; 764} 765extern int32_t __attribute__((overloadable)) clz(int32_t v) { 766 return __builtin_clz(v); 767} 768extern int16_t __attribute__((overloadable)) clz(int16_t v) { 769 return __builtin_clz(((uint32_t)v) & 0x0000ffff) - 16; 770} 771extern int8_t __attribute__((overloadable)) clz(int8_t v) { 772 return __builtin_clz(((uint32_t)v) & 0x000000ff) - 24; 773} 774 775 776UIN_FUNC_IN(abs) 777IN_FUNC_IN(clz) 778 779 780// 6.11.4 781 782 783extern float __attribute__((overloadable)) degrees(float radians) { 784 return radians * (180.f / M_PI); 785} 786extern float2 __attribute__((overloadable)) degrees(float2 radians) { 787 return radians * (180.f / M_PI); 788} 789extern float3 __attribute__((overloadable)) degrees(float3 radians) { 790 return radians * (180.f / M_PI); 791} 792extern float4 __attribute__((overloadable)) degrees(float4 radians) { 793 return radians * (180.f / M_PI); 794} 795 796extern float __attribute__((overloadable)) mix(float start, float stop, float amount) { 797 return start + (stop - start) * amount; 798} 799extern float2 __attribute__((overloadable)) mix(float2 start, float2 stop, float2 amount) { 800 return start + (stop - start) * amount; 801} 802extern float3 __attribute__((overloadable)) mix(float3 start, float3 stop, float3 amount) { 803 return start + (stop - start) * amount; 804} 805extern float4 __attribute__((overloadable)) mix(float4 start, float4 stop, float4 amount) { 806 return start + (stop - start) * amount; 807} 808extern float2 __attribute__((overloadable)) mix(float2 start, float2 stop, float amount) { 809 return start + (stop - start) * amount; 810} 811extern float3 __attribute__((overloadable)) mix(float3 start, float3 stop, float amount) { 812 return start + (stop - start) * amount; 813} 814extern float4 __attribute__((overloadable)) mix(float4 start, float4 stop, float amount) { 815 return start + (stop - start) * amount; 816} 817 818extern float __attribute__((overloadable)) radians(float degrees) { 819 return degrees * (M_PI / 180.f); 820} 821extern float2 __attribute__((overloadable)) radians(float2 degrees) { 822 return degrees * (M_PI / 180.f); 823} 824extern float3 __attribute__((overloadable)) radians(float3 degrees) { 825 return degrees * (M_PI / 180.f); 826} 827extern float4 __attribute__((overloadable)) radians(float4 degrees) { 828 return degrees * (M_PI / 180.f); 829} 830 831extern float __attribute__((overloadable)) step(float edge, float v) { 832 return (v < edge) ? 0.f : 1.f; 833} 834extern float2 __attribute__((overloadable)) step(float2 edge, float2 v) { 835 float2 r; 836 r.x = (v.x < edge.x) ? 0.f : 1.f; 837 r.y = (v.y < edge.y) ? 0.f : 1.f; 838 return r; 839} 840extern float3 __attribute__((overloadable)) step(float3 edge, float3 v) { 841 float3 r; 842 r.x = (v.x < edge.x) ? 0.f : 1.f; 843 r.y = (v.y < edge.y) ? 0.f : 1.f; 844 r.z = (v.z < edge.z) ? 0.f : 1.f; 845 return r; 846} 847extern float4 __attribute__((overloadable)) step(float4 edge, float4 v) { 848 float4 r; 849 r.x = (v.x < edge.x) ? 0.f : 1.f; 850 r.y = (v.y < edge.y) ? 0.f : 1.f; 851 r.z = (v.z < edge.z) ? 0.f : 1.f; 852 r.w = (v.w < edge.w) ? 0.f : 1.f; 853 return r; 854} 855extern float2 __attribute__((overloadable)) step(float2 edge, float v) { 856 float2 r; 857 r.x = (v < edge.x) ? 0.f : 1.f; 858 r.y = (v < edge.y) ? 0.f : 1.f; 859 return r; 860} 861extern float3 __attribute__((overloadable)) step(float3 edge, float v) { 862 float3 r; 863 r.x = (v < edge.x) ? 0.f : 1.f; 864 r.y = (v < edge.y) ? 0.f : 1.f; 865 r.z = (v < edge.z) ? 0.f : 1.f; 866 return r; 867} 868extern float4 __attribute__((overloadable)) step(float4 edge, float v) { 869 float4 r; 870 r.x = (v < edge.x) ? 0.f : 1.f; 871 r.y = (v < edge.y) ? 0.f : 1.f; 872 r.z = (v < edge.z) ? 0.f : 1.f; 873 r.w = (v < edge.w) ? 0.f : 1.f; 874 return r; 875} 876extern float2 __attribute__((overloadable)) step(float edge, float2 v) { 877 float2 r; 878 r.x = (v.x < edge) ? 0.f : 1.f; 879 r.y = (v.y < edge) ? 0.f : 1.f; 880 return r; 881} 882extern float3 __attribute__((overloadable)) step(float edge, float3 v) { 883 float3 r; 884 r.x = (v.x < edge) ? 0.f : 1.f; 885 r.y = (v.y < edge) ? 0.f : 1.f; 886 r.z = (v.z < edge) ? 0.f : 1.f; 887 return r; 888} 889extern float4 __attribute__((overloadable)) step(float edge, float4 v) { 890 float4 r; 891 r.x = (v.x < edge) ? 0.f : 1.f; 892 r.y = (v.y < edge) ? 0.f : 1.f; 893 r.z = (v.z < edge) ? 0.f : 1.f; 894 r.w = (v.w < edge) ? 0.f : 1.f; 895 return r; 896} 897 898extern float __attribute__((overloadable)) sign(float v) { 899 if (v > 0) return 1.f; 900 if (v < 0) return -1.f; 901 return v; 902} 903FN_FUNC_FN(sign) 904 905 906// 6.11.5 907extern float3 __attribute__((overloadable)) cross(float3 lhs, float3 rhs) { 908 float3 r; 909 r.x = lhs.y * rhs.z - lhs.z * rhs.y; 910 r.y = lhs.z * rhs.x - lhs.x * rhs.z; 911 r.z = lhs.x * rhs.y - lhs.y * rhs.x; 912 return r; 913} 914 915extern float4 __attribute__((overloadable)) cross(float4 lhs, float4 rhs) { 916 float4 r; 917 r.x = lhs.y * rhs.z - lhs.z * rhs.y; 918 r.y = lhs.z * rhs.x - lhs.x * rhs.z; 919 r.z = lhs.x * rhs.y - lhs.y * rhs.x; 920 r.w = 0.f; 921 return r; 922} 923 924#if !defined(ARCH_X86_HAVE_SSSE3) || defined(RS_DEBUG_RUNTIME) 925// These functions must be defined here if we are not using the SSE 926// implementation, which includes when we are built as part of the 927// debug runtime (libclcore_debug.bc). 928 929extern float __attribute__((overloadable)) dot(float lhs, float rhs) { 930 return lhs * rhs; 931} 932extern float __attribute__((overloadable)) dot(float2 lhs, float2 rhs) { 933 return lhs.x*rhs.x + lhs.y*rhs.y; 934} 935extern float __attribute__((overloadable)) dot(float3 lhs, float3 rhs) { 936 return lhs.x*rhs.x + lhs.y*rhs.y + lhs.z*rhs.z; 937} 938extern float __attribute__((overloadable)) dot(float4 lhs, float4 rhs) { 939 return lhs.x*rhs.x + lhs.y*rhs.y + lhs.z*rhs.z + lhs.w*rhs.w; 940} 941 942extern float __attribute__((overloadable)) length(float v) { 943 return fabs(v); 944} 945extern float __attribute__((overloadable)) length(float2 v) { 946 return sqrt(v.x*v.x + v.y*v.y); 947} 948extern float __attribute__((overloadable)) length(float3 v) { 949 return sqrt(v.x*v.x + v.y*v.y + v.z*v.z); 950} 951extern float __attribute__((overloadable)) length(float4 v) { 952 return sqrt(v.x*v.x + v.y*v.y + v.z*v.z + v.w*v.w); 953} 954 955#else 956 957extern float __attribute__((overloadable)) length(float v); 958extern float __attribute__((overloadable)) length(float2 v); 959extern float __attribute__((overloadable)) length(float3 v); 960extern float __attribute__((overloadable)) length(float4 v); 961 962#endif // !defined(ARCH_X86_HAVE_SSSE3) || defined(RS_DEBUG_RUNTIME) 963 964extern float __attribute__((overloadable)) distance(float lhs, float rhs) { 965 return length(lhs - rhs); 966} 967extern float __attribute__((overloadable)) distance(float2 lhs, float2 rhs) { 968 return length(lhs - rhs); 969} 970extern float __attribute__((overloadable)) distance(float3 lhs, float3 rhs) { 971 return length(lhs - rhs); 972} 973extern float __attribute__((overloadable)) distance(float4 lhs, float4 rhs) { 974 return length(lhs - rhs); 975} 976 977/* For the normalization functions, vectors of length 0 should simply be 978 * returned (i.e. all the components of that vector are 0). 979 */ 980extern float __attribute__((overloadable)) normalize(float v) { 981 if (v == 0.0f) { 982 return 0.0f; 983 } else if (v < 0.0f) { 984 return -1.0f; 985 } else { 986 return 1.0f; 987 } 988} 989extern float2 __attribute__((overloadable)) normalize(float2 v) { 990 float l = length(v); 991 return l == 0.0f ? v : v / l; 992} 993extern float3 __attribute__((overloadable)) normalize(float3 v) { 994 float l = length(v); 995 return l == 0.0f ? v : v / l; 996} 997extern float4 __attribute__((overloadable)) normalize(float4 v) { 998 float l = length(v); 999 return l == 0.0f ? v : v / l; 1000} 1001 1002extern float __attribute__((overloadable)) half_sqrt(float v) { 1003 return sqrt(v); 1004} 1005FN_FUNC_FN(half_sqrt) 1006 1007extern float __attribute__((overloadable)) fast_length(float v) { 1008 return fabs(v); 1009} 1010extern float __attribute__((overloadable)) fast_length(float2 v) { 1011 return half_sqrt(v.x*v.x + v.y*v.y); 1012} 1013extern float __attribute__((overloadable)) fast_length(float3 v) { 1014 return half_sqrt(v.x*v.x + v.y*v.y + v.z*v.z); 1015} 1016extern float __attribute__((overloadable)) fast_length(float4 v) { 1017 return half_sqrt(v.x*v.x + v.y*v.y + v.z*v.z + v.w*v.w); 1018} 1019 1020extern float __attribute__((overloadable)) fast_distance(float lhs, float rhs) { 1021 return fast_length(lhs - rhs); 1022} 1023extern float __attribute__((overloadable)) fast_distance(float2 lhs, float2 rhs) { 1024 return fast_length(lhs - rhs); 1025} 1026extern float __attribute__((overloadable)) fast_distance(float3 lhs, float3 rhs) { 1027 return fast_length(lhs - rhs); 1028} 1029extern float __attribute__((overloadable)) fast_distance(float4 lhs, float4 rhs) { 1030 return fast_length(lhs - rhs); 1031} 1032 1033extern float __attribute__((overloadable)) half_rsqrt(float); 1034 1035/* For the normalization functions, vectors of length 0 should simply be 1036 * returned (i.e. all the components of that vector are 0). 1037 */ 1038extern float __attribute__((overloadable)) fast_normalize(float v) { 1039 if (v == 0.0f) { 1040 return 0.0f; 1041 } else if (v < 0.0f) { 1042 return -1.0f; 1043 } else { 1044 return 1.0f; 1045 } 1046} 1047// If the length is 0, then rlength should be NaN. 1048extern float2 __attribute__((overloadable)) fast_normalize(float2 v) { 1049 float rlength = half_rsqrt(v.x*v.x + v.y*v.y); 1050 return (rlength == rlength) ? v * rlength : v; 1051} 1052extern float3 __attribute__((overloadable)) fast_normalize(float3 v) { 1053 float rlength = half_rsqrt(v.x*v.x + v.y*v.y + v.z*v.z); 1054 return (rlength == rlength) ? v * rlength : v; 1055} 1056extern float4 __attribute__((overloadable)) fast_normalize(float4 v) { 1057 float rlength = half_rsqrt(v.x*v.x + v.y*v.y + v.z*v.z + v.w*v.w); 1058 return (rlength == rlength) ? v * rlength : v; 1059} 1060 1061extern float __attribute__((overloadable)) half_recip(float v) { 1062 return 1.f / v; 1063} 1064 1065/* 1066extern float __attribute__((overloadable)) approx_atan(float x) { 1067 if (x == 0.f) 1068 return 0.f; 1069 if (x < 0.f) 1070 return -1.f * approx_atan(-1.f * x); 1071 if (x > 1.f) 1072 return M_PI_2 - approx_atan(approx_recip(x)); 1073 return x * approx_recip(1.f + 0.28f * x*x); 1074} 1075FN_FUNC_FN(approx_atan) 1076*/ 1077 1078typedef union 1079{ 1080 float fv; 1081 int32_t iv; 1082} ieee_float_shape_type; 1083 1084/* Get a 32 bit int from a float. */ 1085 1086#define GET_FLOAT_WORD(i,d) \ 1087do { \ 1088 ieee_float_shape_type gf_u; \ 1089 gf_u.fv = (d); \ 1090 (i) = gf_u.iv; \ 1091} while (0) 1092 1093/* Set a float from a 32 bit int. */ 1094 1095#define SET_FLOAT_WORD(d,i) \ 1096do { \ 1097 ieee_float_shape_type sf_u; \ 1098 sf_u.iv = (i); \ 1099 (d) = sf_u.fv; \ 1100} while (0) 1101 1102 1103 1104// Valid -125 to 125 1105extern float __attribute__((overloadable)) native_exp2(float v) { 1106 int32_t iv = (int)v; 1107 int32_t x = iv + (iv >> 31); // ~floor(v) 1108 float r = (v - x); 1109 1110 float fo; 1111 SET_FLOAT_WORD(fo, (x + 127) << 23); 1112 1113 r *= 0.694f; // ~ log(e) / log(2) 1114 float r2 = r*r; 1115 float adj = 1.f + r + (r2 * 0.5f) + (r2*r * 0.166666f) + (r2*r2 * 0.0416666f); 1116 return fo * adj; 1117} 1118 1119extern float2 __attribute__((overloadable)) native_exp2(float2 v) { 1120 int2 iv = convert_int2(v); 1121 int2 x = iv + (iv >> (int2)31);//floor(v); 1122 float2 r = (v - convert_float2(x)); 1123 1124 x += 127; 1125 1126 float2 fo = (float2)(x << (int2)23); 1127 1128 r *= 0.694f; // ~ log(e) / log(2) 1129 float2 r2 = r*r; 1130 float2 adj = 1.f + r + (r2 * 0.5f) + (r2*r * 0.166666f) + (r2*r2 * 0.0416666f); 1131 return fo * adj; 1132} 1133 1134extern float4 __attribute__((overloadable)) native_exp2(float4 v) { 1135 int4 iv = convert_int4(v); 1136 int4 x = iv + (iv >> (int4)31);//floor(v); 1137 float4 r = (v - convert_float4(x)); 1138 1139 x += 127; 1140 1141 float4 fo = (float4)(x << (int4)23); 1142 1143 r *= 0.694f; // ~ log(e) / log(2) 1144 float4 r2 = r*r; 1145 float4 adj = 1.f + r + (r2 * 0.5f) + (r2*r * 0.166666f) + (r2*r2 * 0.0416666f); 1146 return fo * adj; 1147} 1148 1149extern float3 __attribute__((overloadable)) native_exp2(float3 v) { 1150 float4 t = 1.f; 1151 t.xyz = v; 1152 return native_exp2(t).xyz; 1153} 1154 1155 1156extern float __attribute__((overloadable)) native_exp(float v) { 1157 return native_exp2(v * 1.442695041f); 1158} 1159extern float2 __attribute__((overloadable)) native_exp(float2 v) { 1160 return native_exp2(v * 1.442695041f); 1161} 1162extern float3 __attribute__((overloadable)) native_exp(float3 v) { 1163 return native_exp2(v * 1.442695041f); 1164} 1165extern float4 __attribute__((overloadable)) native_exp(float4 v) { 1166 return native_exp2(v * 1.442695041f); 1167} 1168 1169extern float __attribute__((overloadable)) native_exp10(float v) { 1170 return native_exp2(v * 3.321928095f); 1171} 1172extern float2 __attribute__((overloadable)) native_exp10(float2 v) { 1173 return native_exp2(v * 3.321928095f); 1174} 1175extern float3 __attribute__((overloadable)) native_exp10(float3 v) { 1176 return native_exp2(v * 3.321928095f); 1177} 1178extern float4 __attribute__((overloadable)) native_exp10(float4 v) { 1179 return native_exp2(v * 3.321928095f); 1180} 1181 1182extern float __attribute__((overloadable)) native_log2(float v) { 1183 int32_t ibits; 1184 GET_FLOAT_WORD(ibits, v); 1185 1186 int32_t e = (ibits >> 23) & 0xff; 1187 1188 ibits &= 0x7fffff; 1189 ibits |= 127 << 23; 1190 1191 float ir; 1192 SET_FLOAT_WORD(ir, ibits); 1193 ir -= 1.5f; 1194 float ir2 = ir*ir; 1195 float adj2 = (0.405465108f / 0.693147181f) + 1196 ((0.666666667f / 0.693147181f) * ir) - 1197 ((0.222222222f / 0.693147181f) * ir2) + 1198 ((0.098765432f / 0.693147181f) * ir*ir2) - 1199 ((0.049382716f / 0.693147181f) * ir2*ir2) + 1200 ((0.026337449f / 0.693147181f) * ir*ir2*ir2) - 1201 ((0.014631916f / 0.693147181f) * ir2*ir2*ir2); 1202 return (float)(e - 127) + adj2; 1203} 1204extern float2 __attribute__((overloadable)) native_log2(float2 v) { 1205 float2 v2 = {native_log2(v.x), native_log2(v.y)}; 1206 return v2; 1207} 1208extern float3 __attribute__((overloadable)) native_log2(float3 v) { 1209 float3 v2 = {native_log2(v.x), native_log2(v.y), native_log2(v.z)}; 1210 return v2; 1211} 1212extern float4 __attribute__((overloadable)) native_log2(float4 v) { 1213 float4 v2 = {native_log2(v.x), native_log2(v.y), native_log2(v.z), native_log2(v.w)}; 1214 return v2; 1215} 1216 1217extern float __attribute__((overloadable)) native_log(float v) { 1218 return native_log2(v) * (1.f / 1.442695041f); 1219} 1220extern float2 __attribute__((overloadable)) native_log(float2 v) { 1221 return native_log2(v) * (1.f / 1.442695041f); 1222} 1223extern float3 __attribute__((overloadable)) native_log(float3 v) { 1224 return native_log2(v) * (1.f / 1.442695041f); 1225} 1226extern float4 __attribute__((overloadable)) native_log(float4 v) { 1227 return native_log2(v) * (1.f / 1.442695041f); 1228} 1229 1230extern float __attribute__((overloadable)) native_log10(float v) { 1231 return native_log2(v) * (1.f / 3.321928095f); 1232} 1233extern float2 __attribute__((overloadable)) native_log10(float2 v) { 1234 return native_log2(v) * (1.f / 3.321928095f); 1235} 1236extern float3 __attribute__((overloadable)) native_log10(float3 v) { 1237 return native_log2(v) * (1.f / 3.321928095f); 1238} 1239extern float4 __attribute__((overloadable)) native_log10(float4 v) { 1240 return native_log2(v) * (1.f / 3.321928095f); 1241} 1242 1243 1244extern float __attribute__((overloadable)) native_powr(float v, float y) { 1245 float v2 = native_log2(v); 1246 v2 = fmax(v2 * y, -125.f); 1247 return native_exp2(v2); 1248} 1249extern float2 __attribute__((overloadable)) native_powr(float2 v, float2 y) { 1250 float2 v2 = native_log2(v); 1251 v2 = fmax(v2 * y, -125.f); 1252 return native_exp2(v2); 1253} 1254extern float3 __attribute__((overloadable)) native_powr(float3 v, float3 y) { 1255 float3 v2 = native_log2(v); 1256 v2 = fmax(v2 * y, -125.f); 1257 return native_exp2(v2); 1258} 1259extern float4 __attribute__((overloadable)) native_powr(float4 v, float4 y) { 1260 float4 v2 = native_log2(v); 1261 v2 = fmax(v2 * y, -125.f); 1262 return native_exp2(v2); 1263} 1264 1265extern double __attribute__((overloadable)) min(double v1, double v2) { 1266 return v1 < v2 ? v1 : v2; 1267} 1268 1269extern double2 __attribute__((overloadable)) min(double2 v1, double2 v2) { 1270 double2 r; 1271 r.x = v1.x < v2.x ? v1.x : v2.x; 1272 r.y = v1.y < v2.y ? v1.y : v2.y; 1273 return r; 1274} 1275 1276extern double3 __attribute__((overloadable)) min(double3 v1, double3 v2) { 1277 double3 r; 1278 r.x = v1.x < v2.x ? v1.x : v2.x; 1279 r.y = v1.y < v2.y ? v1.y : v2.y; 1280 r.z = v1.z < v2.z ? v1.z : v2.z; 1281 return r; 1282} 1283 1284extern double4 __attribute__((overloadable)) min(double4 v1, double4 v2) { 1285 double4 r; 1286 r.x = v1.x < v2.x ? v1.x : v2.x; 1287 r.y = v1.y < v2.y ? v1.y : v2.y; 1288 r.z = v1.z < v2.z ? v1.z : v2.z; 1289 r.w = v1.w < v2.w ? v1.w : v2.w; 1290 return r; 1291} 1292 1293extern long __attribute__((overloadable)) min(long v1, long v2) { 1294 return v1 < v2 ? v1 : v2; 1295} 1296extern long2 __attribute__((overloadable)) min(long2 v1, long2 v2) { 1297 long2 r; 1298 r.x = v1.x < v2.x ? v1.x : v2.x; 1299 r.y = v1.y < v2.y ? v1.y : v2.y; 1300 return r; 1301} 1302extern long3 __attribute__((overloadable)) min(long3 v1, long3 v2) { 1303 long3 r; 1304 r.x = v1.x < v2.x ? v1.x : v2.x; 1305 r.y = v1.y < v2.y ? v1.y : v2.y; 1306 r.z = v1.z < v2.z ? v1.z : v2.z; 1307 return r; 1308} 1309extern long4 __attribute__((overloadable)) min(long4 v1, long4 v2) { 1310 long4 r; 1311 r.x = v1.x < v2.x ? v1.x : v2.x; 1312 r.y = v1.y < v2.y ? v1.y : v2.y; 1313 r.z = v1.z < v2.z ? v1.z : v2.z; 1314 r.w = v1.w < v2.w ? v1.w : v2.w; 1315 return r; 1316} 1317 1318extern ulong __attribute__((overloadable)) min(ulong v1, ulong v2) { 1319 return v1 < v2 ? v1 : v2; 1320} 1321extern ulong2 __attribute__((overloadable)) min(ulong2 v1, ulong2 v2) { 1322 ulong2 r; 1323 r.x = v1.x < v2.x ? v1.x : v2.x; 1324 r.y = v1.y < v2.y ? v1.y : v2.y; 1325 return r; 1326} 1327extern ulong3 __attribute__((overloadable)) min(ulong3 v1, ulong3 v2) { 1328 ulong3 r; 1329 r.x = v1.x < v2.x ? v1.x : v2.x; 1330 r.y = v1.y < v2.y ? v1.y : v2.y; 1331 r.z = v1.z < v2.z ? v1.z : v2.z; 1332 return r; 1333} 1334extern ulong4 __attribute__((overloadable)) min(ulong4 v1, ulong4 v2) { 1335 ulong4 r; 1336 r.x = v1.x < v2.x ? v1.x : v2.x; 1337 r.y = v1.y < v2.y ? v1.y : v2.y; 1338 r.z = v1.z < v2.z ? v1.z : v2.z; 1339 r.w = v1.w < v2.w ? v1.w : v2.w; 1340 return r; 1341} 1342 1343extern double __attribute__((overloadable)) max(double v1, double v2) { 1344 return v1 > v2 ? v1 : v2; 1345} 1346 1347extern double2 __attribute__((overloadable)) max(double2 v1, double2 v2) { 1348 double2 r; 1349 r.x = v1.x > v2.x ? v1.x : v2.x; 1350 r.y = v1.y > v2.y ? v1.y : v2.y; 1351 return r; 1352} 1353 1354extern double3 __attribute__((overloadable)) max(double3 v1, double3 v2) { 1355 double3 r; 1356 r.x = v1.x > v2.x ? v1.x : v2.x; 1357 r.y = v1.y > v2.y ? v1.y : v2.y; 1358 r.z = v1.z > v2.z ? v1.z : v2.z; 1359 return r; 1360} 1361 1362extern double4 __attribute__((overloadable)) max(double4 v1, double4 v2) { 1363 double4 r; 1364 r.x = v1.x > v2.x ? v1.x : v2.x; 1365 r.y = v1.y > v2.y ? v1.y : v2.y; 1366 r.z = v1.z > v2.z ? v1.z : v2.z; 1367 r.w = v1.w > v2.w ? v1.w : v2.w; 1368 return r; 1369} 1370 1371extern long __attribute__((overloadable)) max(long v1, long v2) { 1372 return v1 > v2 ? v1 : v2; 1373} 1374extern long2 __attribute__((overloadable)) max(long2 v1, long2 v2) { 1375 long2 r; 1376 r.x = v1.x > v2.x ? v1.x : v2.x; 1377 r.y = v1.y > v2.y ? v1.y : v2.y; 1378 return r; 1379} 1380extern long3 __attribute__((overloadable)) max(long3 v1, long3 v2) { 1381 long3 r; 1382 r.x = v1.x > v2.x ? v1.x : v2.x; 1383 r.y = v1.y > v2.y ? v1.y : v2.y; 1384 r.z = v1.z > v2.z ? v1.z : v2.z; 1385 return r; 1386} 1387extern long4 __attribute__((overloadable)) max(long4 v1, long4 v2) { 1388 long4 r; 1389 r.x = v1.x > v2.x ? v1.x : v2.x; 1390 r.y = v1.y > v2.y ? v1.y : v2.y; 1391 r.z = v1.z > v2.z ? v1.z : v2.z; 1392 r.w = v1.w > v2.w ? v1.w : v2.w; 1393 return r; 1394} 1395 1396extern ulong __attribute__((overloadable)) max(ulong v1, ulong v2) { 1397 return v1 > v2 ? v1 : v2; 1398} 1399extern ulong2 __attribute__((overloadable)) max(ulong2 v1, ulong2 v2) { 1400 ulong2 r; 1401 r.x = v1.x > v2.x ? v1.x : v2.x; 1402 r.y = v1.y > v2.y ? v1.y : v2.y; 1403 return r; 1404} 1405extern ulong3 __attribute__((overloadable)) max(ulong3 v1, ulong3 v2) { 1406 ulong3 r; 1407 r.x = v1.x > v2.x ? v1.x : v2.x; 1408 r.y = v1.y > v2.y ? v1.y : v2.y; 1409 r.z = v1.z > v2.z ? v1.z : v2.z; 1410 return r; 1411} 1412extern ulong4 __attribute__((overloadable)) max(ulong4 v1, ulong4 v2) { 1413 ulong4 r; 1414 r.x = v1.x > v2.x ? v1.x : v2.x; 1415 r.y = v1.y > v2.y ? v1.y : v2.y; 1416 r.z = v1.z > v2.z ? v1.z : v2.z; 1417 r.w = v1.w > v2.w ? v1.w : v2.w; 1418 return r; 1419} 1420 1421#define THUNK_NATIVE_F(fn) \ 1422 float __attribute__((overloadable)) native_##fn(float v) { return fn(v);} \ 1423 float2 __attribute__((overloadable)) native_##fn(float2 v) { return fn(v);} \ 1424 float3 __attribute__((overloadable)) native_##fn(float3 v) { return fn(v);} \ 1425 float4 __attribute__((overloadable)) native_##fn(float4 v) { return fn(v);} 1426 1427#define THUNK_NATIVE_F_F(fn) \ 1428 float __attribute__((overloadable)) native_##fn(float v1, float v2) { return fn(v1, v2);} \ 1429 float2 __attribute__((overloadable)) native_##fn(float2 v1, float2 v2) { return fn(v1, v2);} \ 1430 float3 __attribute__((overloadable)) native_##fn(float3 v1, float3 v2) { return fn(v1, v2);} \ 1431 float4 __attribute__((overloadable)) native_##fn(float4 v1, float4 v2) { return fn(v1, v2);} 1432 1433#define THUNK_NATIVE_F_FP(fn) \ 1434 float __attribute__((overloadable)) native_##fn(float v1, float *v2) { return fn(v1, v2);} \ 1435 float2 __attribute__((overloadable)) native_##fn(float2 v1, float2 *v2) { return fn(v1, v2);} \ 1436 float3 __attribute__((overloadable)) native_##fn(float3 v1, float3 *v2) { return fn(v1, v2);} \ 1437 float4 __attribute__((overloadable)) native_##fn(float4 v1, float4 *v2) { return fn(v1, v2);} 1438 1439#define THUNK_NATIVE_F_I(fn) \ 1440 float __attribute__((overloadable)) native_##fn(float v1, int v2) { return fn(v1, v2);} \ 1441 float2 __attribute__((overloadable)) native_##fn(float2 v1, int2 v2) { return fn(v1, v2);} \ 1442 float3 __attribute__((overloadable)) native_##fn(float3 v1, int3 v2) { return fn(v1, v2);} \ 1443 float4 __attribute__((overloadable)) native_##fn(float4 v1, int4 v2) { return fn(v1, v2);} 1444 1445THUNK_NATIVE_F(acos) 1446THUNK_NATIVE_F(acosh) 1447THUNK_NATIVE_F(acospi) 1448THUNK_NATIVE_F(asin) 1449THUNK_NATIVE_F(asinh) 1450THUNK_NATIVE_F(asinpi) 1451THUNK_NATIVE_F(atan) 1452THUNK_NATIVE_F_F(atan2) 1453THUNK_NATIVE_F(atanh) 1454THUNK_NATIVE_F(atanpi) 1455THUNK_NATIVE_F_F(atan2pi) 1456THUNK_NATIVE_F(cbrt) 1457THUNK_NATIVE_F(cos) 1458THUNK_NATIVE_F(cosh) 1459THUNK_NATIVE_F(cospi) 1460THUNK_NATIVE_F(expm1) 1461THUNK_NATIVE_F_F(hypot) 1462THUNK_NATIVE_F(log1p) 1463THUNK_NATIVE_F_I(rootn) 1464THUNK_NATIVE_F(rsqrt) 1465THUNK_NATIVE_F(sqrt) 1466THUNK_NATIVE_F(sin) 1467THUNK_NATIVE_F_FP(sincos) 1468THUNK_NATIVE_F(sinh) 1469THUNK_NATIVE_F(sinpi) 1470THUNK_NATIVE_F(tan) 1471THUNK_NATIVE_F(tanh) 1472THUNK_NATIVE_F(tanpi) 1473 1474#undef THUNK_NATIVE_F 1475#undef THUNK_NATIVE_F_F 1476#undef THUNK_NATIVE_F_I 1477#undef THUNK_NATIVE_F_FP 1478 1479float __attribute__((overloadable)) native_normalize(float v) { return fast_normalize(v);} 1480float2 __attribute__((overloadable)) native_normalize(float2 v) { return fast_normalize(v);} 1481float3 __attribute__((overloadable)) native_normalize(float3 v) { return fast_normalize(v);} 1482float4 __attribute__((overloadable)) native_normalize(float4 v) { return fast_normalize(v);} 1483 1484float __attribute__((overloadable)) native_distance(float v1, float v2) { return fast_distance(v1, v2);} 1485float __attribute__((overloadable)) native_distance(float2 v1, float2 v2) { return fast_distance(v1, v2);} 1486float __attribute__((overloadable)) native_distance(float3 v1, float3 v2) { return fast_distance(v1, v2);} 1487float __attribute__((overloadable)) native_distance(float4 v1, float4 v2) { return fast_distance(v1, v2);} 1488 1489float __attribute__((overloadable)) native_length(float v) { return fast_length(v);} 1490float __attribute__((overloadable)) native_length(float2 v) { return fast_length(v);} 1491float __attribute__((overloadable)) native_length(float3 v) { return fast_length(v);} 1492float __attribute__((overloadable)) native_length(float4 v) { return fast_length(v);} 1493 1494float __attribute__((overloadable)) native_divide(float v1, float v2) { return v1 / v2;} 1495float2 __attribute__((overloadable)) native_divide(float2 v1, float2 v2) { return v1 / v2;} 1496float3 __attribute__((overloadable)) native_divide(float3 v1, float3 v2) { return v1 / v2;} 1497float4 __attribute__((overloadable)) native_divide(float4 v1, float4 v2) { return v1 / v2;} 1498 1499float __attribute__((overloadable)) native_recip(float v) { return 1.f / v;} 1500float2 __attribute__((overloadable)) native_recip(float2 v) { return ((float2)1.f) / v;} 1501float3 __attribute__((overloadable)) native_recip(float3 v) { return ((float3)1.f) / v;} 1502float4 __attribute__((overloadable)) native_recip(float4 v) { return ((float4)1.f) / v;} 1503 1504 1505 1506 1507 1508#undef FN_FUNC_FN 1509#undef IN_FUNC_FN 1510#undef FN_FUNC_FN_FN 1511#undef FN_FUNC_FN_F 1512#undef FN_FUNC_FN_IN 1513#undef FN_FUNC_FN_I 1514#undef FN_FUNC_FN_PFN 1515#undef FN_FUNC_FN_PIN 1516#undef FN_FUNC_FN_FN_FN 1517#undef FN_FUNC_FN_FN_PIN 1518#undef XN_FUNC_YN 1519#undef UIN_FUNC_IN 1520#undef IN_FUNC_IN 1521#undef XN_FUNC_XN_XN_BODY 1522#undef IN_FUNC_IN_IN_BODY 1523 1524typedef union { 1525 half hval; 1526 short sval; 1527} fp16_shape_type; 1528 1529/* half h = unsigned short s; */ 1530#define SET_HALF_WORD(h, s) \ 1531do { \ 1532 fp16_shape_type fp16_u; \ 1533 fp16_u.sval = (s); \ 1534 (h) = fp16_u.hval; \ 1535} while (0) 1536 1537static const unsigned short kHalfPositiveInfinity = 0x7c00; 1538 1539/* Define f16 functions of the form 1540 * HN output = fn(HN input) 1541 * where HN is scalar or vector half type 1542 */ 1543#define HN_FUNC_HN(fn) \ 1544extern half __attribute__((overloadable)) fn(half h) { \ 1545 return (half) fn((float) h); \ 1546} \ 1547extern half2 __attribute__((overloadable)) fn(half2 v) { \ 1548 return convert_half2(fn(convert_float2(v))); \ 1549} \ 1550extern half3 __attribute__((overloadable)) fn(half3 v) { \ 1551 return convert_half3(fn(convert_float3(v))); \ 1552} \ 1553extern half4 __attribute__((overloadable)) fn(half4 v) { \ 1554 return convert_half4(fn(convert_float4(v))); \ 1555} 1556 1557/* Define f16 functions of the form 1558 * HN output = fn(HN input1, HN input2) 1559 * where HN is scalar or vector half type 1560 */ 1561#define HN_FUNC_HN_HN(fn) \ 1562extern half __attribute__((overloadable)) fn(half h1, half h2) { \ 1563 return (half) fn((float) h1, (float) h2); \ 1564} \ 1565extern half2 __attribute__((overloadable)) fn(half2 v1, half2 v2) { \ 1566 return convert_half2(fn(convert_float2(v1), \ 1567 convert_float2(v2))); \ 1568} \ 1569extern half3 __attribute__((overloadable)) fn(half3 v1, half3 v2) { \ 1570 return convert_half3(fn(convert_float3(v1), \ 1571 convert_float3(v2))); \ 1572} \ 1573extern half4 __attribute__((overloadable)) fn(half4 v1, half4 v2) { \ 1574 return convert_half4(fn(convert_float4(v1), \ 1575 convert_float4(v2))); \ 1576} 1577 1578/* Define f16 functions of the form 1579 * HN output = fn(HN input1, half input2) 1580 * where HN is scalar or vector half type 1581 */ 1582#define HN_FUNC_HN_H(fn) \ 1583extern half2 __attribute__((overloadable)) fn(half2 v1, half v2) { \ 1584 return convert_half2(fn(convert_float2(v1), (float) v2)); \ 1585} \ 1586extern half3 __attribute__((overloadable)) fn(half3 v1, half v2) { \ 1587 return convert_half3(fn(convert_float3(v1), (float) v2)); \ 1588} \ 1589extern half4 __attribute__((overloadable)) fn(half4 v1, half v2) { \ 1590 return convert_half4(fn(convert_float4(v1), (float) v2)); \ 1591} 1592 1593/* Define f16 functions of the form 1594 * HN output = fn(HN input1, HN input2, HN input3) 1595 * where HN is scalar or vector half type 1596 */ 1597#define HN_FUNC_HN_HN_HN(fn) \ 1598extern half __attribute__((overloadable)) fn(half h1, half h2, half h3) { \ 1599 return (half) fn((float) h1, (float) h2, (float) h3); \ 1600} \ 1601extern half2 __attribute__((overloadable)) fn(half2 v1, half2 v2, half2 v3) { \ 1602 return convert_half2(fn(convert_float2(v1), \ 1603 convert_float2(v2), \ 1604 convert_float2(v3))); \ 1605} \ 1606extern half3 __attribute__((overloadable)) fn(half3 v1, half3 v2, half3 v3) { \ 1607 return convert_half3(fn(convert_float3(v1), \ 1608 convert_float3(v2), \ 1609 convert_float3(v3))); \ 1610} \ 1611extern half4 __attribute__((overloadable)) fn(half4 v1, half4 v2, half4 v3) { \ 1612 return convert_half4(fn(convert_float4(v1), \ 1613 convert_float4(v2), \ 1614 convert_float4(v3))); \ 1615} 1616 1617/* Define f16 functions of the form 1618 * HN output = fn(HN input1, IN input2) 1619 * where HN is scalar or vector half type and IN the equivalent integer type 1620 * of same vector length. 1621 */ 1622#define HN_FUNC_HN_IN(fn) \ 1623extern half __attribute__((overloadable)) fn(half h1, int v) { \ 1624 return (half) fn((float) h1, v); \ 1625} \ 1626extern half2 __attribute__((overloadable)) fn(half2 v1, int2 v2) { \ 1627 return convert_half2(fn(convert_float2(v1), v2)); \ 1628} \ 1629extern half3 __attribute__((overloadable)) fn(half3 v1, int3 v2) { \ 1630 return convert_half3(fn(convert_float3(v1), v2)); \ 1631} \ 1632extern half4 __attribute__((overloadable)) fn(half4 v1, int4 v2) { \ 1633 return convert_half4(fn(convert_float4(v1), v2)); \ 1634} 1635 1636/* Define f16 functions of the form 1637 * half output = fn(HN input1) 1638 * where HN is a scalar or vector half type. 1639 */ 1640#define H_FUNC_HN(fn) \ 1641extern half __attribute__((overloadable)) fn(half h) { \ 1642 return (half) fn((float) h); \ 1643} \ 1644extern half __attribute__((overloadable)) fn(half2 v) { \ 1645 return fn(convert_float2(v)); \ 1646} \ 1647extern half __attribute__((overloadable)) fn(half3 v) { \ 1648 return fn(convert_float3(v)); \ 1649} \ 1650extern half __attribute__((overloadable)) fn(half4 v) { \ 1651 return fn(convert_float4(v)); \ 1652} 1653 1654/* Define f16 functions of the form 1655 * half output = fn(HN input1, HN input2) 1656 * where HN is a scalar or vector half type. 1657 */ 1658#define H_FUNC_HN_HN(fn) \ 1659extern half __attribute__((overloadable)) fn(half h1, half h2) { \ 1660 return (half) fn((float) h1, (float) h2); \ 1661} \ 1662extern half __attribute__((overloadable)) fn(half2 v1, half2 v2) { \ 1663 return fn(convert_float2(v1), convert_float2(v2)); \ 1664} \ 1665extern half __attribute__((overloadable)) fn(half3 v1, half3 v2) { \ 1666 return fn(convert_float3(v1), convert_float3(v2)); \ 1667} \ 1668extern half __attribute__((overloadable)) fn(half4 v1, half4 v2) { \ 1669 return fn(convert_float4(v1), convert_float4(v2)); \ 1670} 1671 1672/* Define f16 functions of the form 1673 * HN output = fn(HN input1, HN input2) 1674 * where HN is a vector half type. The functions are defined to call the 1675 * scalar function of the same name. 1676 */ 1677#define SCALARIZE_HN_FUNC_HN_HN(fn) \ 1678extern half2 __attribute__((overloadable)) fn(half2 v1, half2 v2) { \ 1679 half2 ret; \ 1680 ret.x = fn(v1.x, v2.x); \ 1681 ret.y = fn(v1.y, v2.y); \ 1682 return ret; \ 1683} \ 1684extern half3 __attribute__((overloadable)) fn(half3 v1, half3 v2) { \ 1685 half3 ret; \ 1686 ret.x = fn(v1.x, v2.x); \ 1687 ret.y = fn(v1.y, v2.y); \ 1688 ret.z = fn(v1.z, v2.z); \ 1689 return ret; \ 1690} \ 1691extern half4 __attribute__((overloadable)) fn(half4 v1, half4 v2) { \ 1692 half4 ret; \ 1693 ret.x = fn(v1.x, v2.x); \ 1694 ret.y = fn(v1.y, v2.y); \ 1695 ret.z = fn(v1.z, v2.z); \ 1696 ret.w = fn(v1.w, v2.w); \ 1697 return ret; \ 1698} \ 1699 1700HN_FUNC_HN(acos); 1701HN_FUNC_HN(acosh); 1702HN_FUNC_HN(acospi); 1703HN_FUNC_HN(asin); 1704HN_FUNC_HN(asinh); 1705HN_FUNC_HN(asinpi); 1706HN_FUNC_HN(atan); 1707HN_FUNC_HN(atanh); 1708HN_FUNC_HN(atanpi); 1709HN_FUNC_HN_HN(atan2); 1710HN_FUNC_HN_HN(atan2pi); 1711 1712HN_FUNC_HN(cbrt); 1713HN_FUNC_HN(ceil); 1714 1715// TODO Add copysign 1716 1717HN_FUNC_HN(cos); 1718HN_FUNC_HN(cosh); 1719HN_FUNC_HN(cospi); 1720 1721extern half3 __attribute__((overloadable)) cross(half3 lhs, half3 rhs) { 1722 half3 r; 1723 r.x = lhs.y * rhs.z - lhs.z * rhs.y; 1724 r.y = lhs.z * rhs.x - lhs.x * rhs.z; 1725 r.z = lhs.x * rhs.y - lhs.y * rhs.x; 1726 return r; 1727} 1728 1729extern half4 __attribute__((overloadable)) cross(half4 lhs, half4 rhs) { 1730 half4 r; 1731 r.x = lhs.y * rhs.z - lhs.z * rhs.y; 1732 r.y = lhs.z * rhs.x - lhs.x * rhs.z; 1733 r.z = lhs.x * rhs.y - lhs.y * rhs.x; 1734 r.w = 0.f; 1735 return r; 1736} 1737 1738HN_FUNC_HN(degrees); 1739H_FUNC_HN_HN(distance); 1740H_FUNC_HN_HN(dot); 1741 1742HN_FUNC_HN(erf); 1743HN_FUNC_HN(erfc); 1744HN_FUNC_HN(exp); 1745HN_FUNC_HN(exp10); 1746HN_FUNC_HN(exp2); 1747HN_FUNC_HN(expm1); 1748 1749HN_FUNC_HN(fabs); 1750HN_FUNC_HN_HN(fdim); 1751HN_FUNC_HN(floor); 1752HN_FUNC_HN_HN_HN(fma); 1753HN_FUNC_HN_HN(fmax); 1754HN_FUNC_HN_H(fmax); 1755HN_FUNC_HN_HN(fmin); 1756HN_FUNC_HN_H(fmin); 1757HN_FUNC_HN_HN(fmod); 1758 1759// TODO Add (both variants) of fract 1760// TODO Add frexp 1761 1762HN_FUNC_HN_HN(hypot); 1763 1764// TODO Add ilogb 1765 1766HN_FUNC_HN_IN(ldexp); 1767extern half2 __attribute__((overloadable)) ldexp(half2 v, int exponent) { 1768 return convert_half2(ldexp(convert_float2(v), exponent)); 1769} 1770extern half3 __attribute__((overloadable)) ldexp(half3 v, int exponent) { 1771 return convert_half3(ldexp(convert_float3(v), exponent)); 1772} 1773extern half4 __attribute__((overloadable)) ldexp(half4 v, int exponent) { 1774 return convert_half4(ldexp(convert_float4(v), exponent)); 1775} 1776 1777H_FUNC_HN(length); 1778HN_FUNC_HN(lgamma); 1779 1780extern half __attribute__((overloadable)) lgamma(half h, int *signp) { 1781 return (half) lgamma((float) h, signp); 1782} 1783extern half2 __attribute__((overloadable)) lgamma(half2 v, int2 *signp) { 1784 return convert_half2(lgamma(convert_float2(v), signp)); 1785} 1786extern half3 __attribute__((overloadable)) lgamma(half3 v, int3 *signp) { 1787 return convert_half3(lgamma(convert_float3(v), signp)); 1788} 1789extern half4 __attribute__((overloadable)) lgamma(half4 v, int4 *signp) { 1790 return convert_half4(lgamma(convert_float4(v), signp)); 1791} 1792 1793HN_FUNC_HN(log); 1794HN_FUNC_HN(log10); 1795HN_FUNC_HN(log1p); 1796HN_FUNC_HN(log2); 1797HN_FUNC_HN(logb); 1798 1799HN_FUNC_HN_HN_HN(mad); 1800HN_FUNC_HN_HN(max); 1801HN_FUNC_HN_H(max); // TODO can this be arch-specific similar to _Z3maxDv2_ff? 1802HN_FUNC_HN_HN(min); 1803HN_FUNC_HN_H(min); // TODO can this be arch-specific similar to _Z3minDv2_ff? 1804 1805extern half __attribute__((overloadable)) mix(half start, half stop, half amount) { 1806 return start + (stop - start) * amount; 1807} 1808extern half2 __attribute__((overloadable)) mix(half2 start, half2 stop, half2 amount) { 1809 return start + (stop - start) * amount; 1810} 1811extern half3 __attribute__((overloadable)) mix(half3 start, half3 stop, half3 amount) { 1812 return start + (stop - start) * amount; 1813} 1814extern half4 __attribute__((overloadable)) mix(half4 start, half4 stop, half4 amount) { 1815 return start + (stop - start) * amount; 1816} 1817extern half2 __attribute__((overloadable)) mix(half2 start, half2 stop, half amount) { 1818 return start + (stop - start) * amount; 1819} 1820extern half3 __attribute__((overloadable)) mix(half3 start, half3 stop, half amount) { 1821 return start + (stop - start) * amount; 1822} 1823extern half4 __attribute__((overloadable)) mix(half4 start, half4 stop, half amount) { 1824 return start + (stop - start) * amount; 1825} 1826 1827// TODO Define modf. Does it make sense to delegate to the float? 1828 1829half __attribute__((overloadable)) nan_half() { 1830 unsigned short nan_short = kHalfPositiveInfinity | 0x0200; 1831 half nan; 1832 SET_HALF_WORD(nan, nan_short); 1833 return nan; 1834} 1835 1836// TODO Add nextafter 1837 1838HN_FUNC_HN(normalize); 1839 1840HN_FUNC_HN_HN(pow); 1841HN_FUNC_HN_IN(pown); 1842HN_FUNC_HN_HN(powr); 1843HN_FUNC_HN(radians); 1844HN_FUNC_HN_HN(remainder); 1845 1846extern half __attribute__((overloadable)) remquo(half n, half d, int *quo) { 1847 return (float) remquo((float) n, (float) d, quo); 1848} 1849extern half2 __attribute__((overloadable)) remquo(half2 n, half2 d, int2 *quo) { 1850 return convert_half2(remquo(convert_float2(d), convert_float2(n), quo)); 1851} 1852extern half3 __attribute__((overloadable)) remquo(half3 n, half3 d, int3 *quo) { 1853 return convert_half3(remquo(convert_float3(d), convert_float3(n), quo)); 1854} 1855extern half4 __attribute__((overloadable)) remquo(half4 n, half4 d, int4 *quo) { 1856 return convert_half4(remquo(convert_float4(d), convert_float4(n), quo)); 1857} 1858 1859HN_FUNC_HN(rint); 1860HN_FUNC_HN_IN(rootn); 1861HN_FUNC_HN(round); 1862HN_FUNC_HN(rsqrt); 1863 1864extern half __attribute__((overloadable)) sign(half h) { 1865 if (h > 0) return (half) 1.f; 1866 if (h < 0) return (half) -1.f; 1867 return h; 1868} 1869extern half2 __attribute__((overloadable)) sign(half2 v) { 1870 half2 ret; 1871 ret.x = sign(v.x); 1872 ret.y = sign(v.y); 1873 return ret; 1874} 1875extern half3 __attribute__((overloadable)) sign(half3 v) { 1876 half3 ret; 1877 ret.x = sign(v.x); 1878 ret.y = sign(v.y); 1879 ret.z = sign(v.z); 1880 return ret; 1881} 1882extern half4 __attribute__((overloadable)) sign(half4 v) { 1883 half4 ret; 1884 ret.x = sign(v.x); 1885 ret.y = sign(v.y); 1886 ret.z = sign(v.z); 1887 ret.w = sign(v.w); 1888 return ret; 1889} 1890 1891HN_FUNC_HN(sin); 1892 1893extern half __attribute__((overloadable)) sincos(half v, half *cosptr) { 1894 *cosptr = cos(v); 1895 return sin(v); 1896} 1897// TODO verify if LLVM eliminates the duplicate convert_float2 1898extern half2 __attribute__((overloadable)) sincos(half2 v, half2 *cosptr) { 1899 *cosptr = cos(v); 1900 return sin(v); 1901} 1902extern half3 __attribute__((overloadable)) sincos(half3 v, half3 *cosptr) { 1903 *cosptr = cos(v); 1904 return sin(v); 1905} 1906extern half4 __attribute__((overloadable)) sincos(half4 v, half4 *cosptr) { 1907 *cosptr = cos(v); 1908 return sin(v); 1909} 1910 1911HN_FUNC_HN(sinh); 1912HN_FUNC_HN(sinpi); 1913HN_FUNC_HN(sqrt); 1914 1915extern half __attribute__((overloadable)) step(half edge, half v) { 1916 return (v < edge) ? 0.f : 1.f; 1917} 1918extern half2 __attribute__((overloadable)) step(half2 edge, half2 v) { 1919 half2 r; 1920 r.x = (v.x < edge.x) ? 0.f : 1.f; 1921 r.y = (v.y < edge.y) ? 0.f : 1.f; 1922 return r; 1923} 1924extern half3 __attribute__((overloadable)) step(half3 edge, half3 v) { 1925 half3 r; 1926 r.x = (v.x < edge.x) ? 0.f : 1.f; 1927 r.y = (v.y < edge.y) ? 0.f : 1.f; 1928 r.z = (v.z < edge.z) ? 0.f : 1.f; 1929 return r; 1930} 1931extern half4 __attribute__((overloadable)) step(half4 edge, half4 v) { 1932 half4 r; 1933 r.x = (v.x < edge.x) ? 0.f : 1.f; 1934 r.y = (v.y < edge.y) ? 0.f : 1.f; 1935 r.z = (v.z < edge.z) ? 0.f : 1.f; 1936 r.w = (v.w < edge.w) ? 0.f : 1.f; 1937 return r; 1938} 1939extern half2 __attribute__((overloadable)) step(half2 edge, half v) { 1940 half2 r; 1941 r.x = (v < edge.x) ? 0.f : 1.f; 1942 r.y = (v < edge.y) ? 0.f : 1.f; 1943 return r; 1944} 1945extern half3 __attribute__((overloadable)) step(half3 edge, half v) { 1946 half3 r; 1947 r.x = (v < edge.x) ? 0.f : 1.f; 1948 r.y = (v < edge.y) ? 0.f : 1.f; 1949 r.z = (v < edge.z) ? 0.f : 1.f; 1950 return r; 1951} 1952extern half4 __attribute__((overloadable)) step(half4 edge, half v) { 1953 half4 r; 1954 r.x = (v < edge.x) ? 0.f : 1.f; 1955 r.y = (v < edge.y) ? 0.f : 1.f; 1956 r.z = (v < edge.z) ? 0.f : 1.f; 1957 r.w = (v < edge.w) ? 0.f : 1.f; 1958 return r; 1959} 1960extern half2 __attribute__((overloadable)) step(half edge, half2 v) { 1961 half2 r; 1962 r.x = (v.x < edge) ? 0.f : 1.f; 1963 r.y = (v.y < edge) ? 0.f : 1.f; 1964 return r; 1965} 1966extern half3 __attribute__((overloadable)) step(half edge, half3 v) { 1967 half3 r; 1968 r.x = (v.x < edge) ? 0.f : 1.f; 1969 r.y = (v.y < edge) ? 0.f : 1.f; 1970 r.z = (v.z < edge) ? 0.f : 1.f; 1971 return r; 1972} 1973extern half4 __attribute__((overloadable)) step(half edge, half4 v) { 1974 half4 r; 1975 r.x = (v.x < edge) ? 0.f : 1.f; 1976 r.y = (v.y < edge) ? 0.f : 1.f; 1977 r.z = (v.z < edge) ? 0.f : 1.f; 1978 r.w = (v.w < edge) ? 0.f : 1.f; 1979 return r; 1980} 1981 1982HN_FUNC_HN(tan); 1983HN_FUNC_HN(tanh); 1984HN_FUNC_HN(tanpi); 1985HN_FUNC_HN(tgamma); 1986HN_FUNC_HN(trunc); // TODO: rethink: needs half-specific implementation? 1987 1988HN_FUNC_HN(native_acos); 1989HN_FUNC_HN(native_acosh); 1990HN_FUNC_HN(native_acospi); 1991HN_FUNC_HN(native_asin); 1992HN_FUNC_HN(native_asinh); 1993HN_FUNC_HN(native_asinpi); 1994HN_FUNC_HN(native_atan); 1995HN_FUNC_HN(native_atanh); 1996HN_FUNC_HN(native_atanpi); 1997HN_FUNC_HN_HN(native_atan2); 1998HN_FUNC_HN_HN(native_atan2pi); 1999 2000HN_FUNC_HN(native_cbrt); 2001HN_FUNC_HN(native_cos); 2002HN_FUNC_HN(native_cosh); 2003HN_FUNC_HN(native_cospi); 2004 2005H_FUNC_HN_HN(native_distance); 2006HN_FUNC_HN_HN(native_divide); 2007 2008HN_FUNC_HN(native_exp); 2009HN_FUNC_HN(native_exp10); 2010HN_FUNC_HN(native_exp2); 2011HN_FUNC_HN(native_expm1); 2012 2013HN_FUNC_HN_HN(native_hypot); 2014H_FUNC_HN(native_length); 2015 2016HN_FUNC_HN(native_log); 2017HN_FUNC_HN(native_log10); 2018HN_FUNC_HN(native_log1p); 2019HN_FUNC_HN(native_log2); 2020 2021HN_FUNC_HN(native_normalize); 2022 2023HN_FUNC_HN_HN(native_powr); // TODO are parameter limits different for half? 2024 2025HN_FUNC_HN(native_recip); 2026HN_FUNC_HN_IN(native_rootn); 2027HN_FUNC_HN(native_rsqrt); 2028 2029HN_FUNC_HN(native_sin); 2030 2031extern half __attribute__((overloadable)) native_sincos(half v, half *cosptr) { 2032 return sincos(v, cosptr); 2033} 2034extern half2 __attribute__((overloadable)) native_sincos(half2 v, half2 *cosptr) { 2035 return sincos(v, cosptr); 2036} 2037extern half3 __attribute__((overloadable)) native_sincos(half3 v, half3 *cosptr) { 2038 return sincos(v, cosptr); 2039} 2040extern half4 __attribute__((overloadable)) native_sincos(half4 v, half4 *cosptr) { 2041 return sincos(v, cosptr); 2042} 2043 2044HN_FUNC_HN(native_sinh); 2045HN_FUNC_HN(native_sinpi); 2046HN_FUNC_HN(native_sqrt); 2047 2048HN_FUNC_HN(native_tan); 2049HN_FUNC_HN(native_tanh); 2050HN_FUNC_HN(native_tanpi); 2051 2052#undef HN_FUNC_HN 2053#undef HN_FUNC_HN_HN 2054#undef HN_FUNC_HN_H 2055#undef HN_FUNC_HN_HN_HN 2056#undef HN_FUNC_HN_IN 2057#undef H_FUNC_HN 2058#undef H_FUNC_HN_HN 2059#undef SCALARIZE_HN_FUNC_HN_HN 2060 2061