1/************************************************************************** 2 * 3 * Copyright 2009 VMware, Inc. 4 * All Rights Reserved. 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the 8 * "Software"), to deal in the Software without restriction, including 9 * without limitation the rights to use, copy, modify, merge, publish, 10 * distribute, sub license, and/or sell copies of the Software, and to 11 * permit persons to whom the Software is furnished to do so, subject to 12 * the following conditions: 13 * 14 * The above copyright notice and this permission notice (including the 15 * next paragraph) shall be included in all copies or substantial portions 16 * of the Software. 17 * 18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS 19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. 21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR 22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, 23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE 24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. 25 * 26 **************************************************************************/ 27 28 29/** 30 * @file 31 * Helper functions for constant building. 32 * 33 * @author Jose Fonseca <jfonseca@vmware.com> 34 */ 35 36#include <float.h> 37 38#include "util/u_debug.h" 39#include "util/u_math.h" 40#include "util/half_float.h" 41 42#include "lp_bld_type.h" 43#include "lp_bld_const.h" 44#include "lp_bld_init.h" 45#include "lp_bld_limits.h" 46 47 48unsigned 49lp_mantissa(struct lp_type type) 50{ 51 assert(type.floating); 52 53 if (type.floating) { 54 switch (type.width) { 55 case 16: 56 return 10; 57 case 32: 58 return 23; 59 case 64: 60 return 52; 61 default: 62 assert(0); 63 return 0; 64 } 65 } else { 66 if (type.sign) 67 return type.width - 1; 68 else 69 return type.width; 70 } 71} 72 73 74/** 75 * Shift of the unity. 76 * 77 * Same as lp_const_scale(), but in terms of shifts. 78 */ 79unsigned 80lp_const_shift(struct lp_type type) 81{ 82 if (type.floating) 83 return 0; 84 else if (type.fixed) 85 return type.width/2; 86 else if (type.norm) 87 return type.sign ? type.width - 1 : type.width; 88 else 89 return 0; 90} 91 92 93unsigned 94lp_const_offset(struct lp_type type) 95{ 96 if (type.floating || type.fixed) 97 return 0; 98 else if (type.norm) 99 return 1; 100 else 101 return 0; 102} 103 104 105/** 106 * Scaling factor between the LLVM native value and its interpretation. 107 * 108 * This is 1.0 for all floating types and unnormalized integers, and something 109 * else for the fixed points types and normalized integers. 110 */ 111double 112lp_const_scale(struct lp_type type) 113{ 114 unsigned long long llscale; 115 double dscale; 116 117 llscale = (unsigned long long)1 << lp_const_shift(type); 118 llscale -= lp_const_offset(type); 119 dscale = (double)llscale; 120 assert((unsigned long long)dscale == llscale); 121 122 return dscale; 123} 124 125 126/** 127 * Minimum value representable by the type. 128 */ 129double 130lp_const_min(struct lp_type type) 131{ 132 if (!type.sign) 133 return 0.0; 134 135 if (type.norm) 136 return -1.0; 137 138 if (type.floating) { 139 switch (type.width) { 140 case 16: 141 return -65504; 142 case 32: 143 return -FLT_MAX; 144 case 64: 145 return -DBL_MAX; 146 default: 147 assert(0); 148 return 0.0; 149 } 150 } 151 152 unsigned bits; 153 if (type.fixed) 154 /* FIXME: consider the fractional bits? */ 155 bits = type.width / 2 - 1; 156 else 157 bits = type.width - 1; 158 159 return (double)-((long long)1 << bits); 160} 161 162 163/** 164 * Maximum value representable by the type. 165 */ 166double 167lp_const_max(struct lp_type type) 168{ 169 if (type.norm) 170 return 1.0; 171 172 if (type.floating) { 173 switch (type.width) { 174 case 16: 175 return 65504; 176 case 32: 177 return FLT_MAX; 178 case 64: 179 return DBL_MAX; 180 default: 181 assert(0); 182 return 0.0; 183 } 184 } 185 186 unsigned bits; 187 if (type.fixed) 188 bits = type.width / 2; 189 else 190 bits = type.width; 191 192 if (type.sign) 193 bits -= 1; 194 195 return (double)(((unsigned long long)1 << bits) - 1); 196} 197 198 199double 200lp_const_eps(struct lp_type type) 201{ 202 if (type.floating) { 203 switch (type.width) { 204 case 16: 205 return 2E-10; 206 case 32: 207 return FLT_EPSILON; 208 case 64: 209 return DBL_EPSILON; 210 default: 211 assert(0); 212 return 0.0; 213 } 214 } else { 215 double scale = lp_const_scale(type); 216 return 1.0/scale; 217 } 218} 219 220 221LLVMValueRef 222lp_build_undef(struct gallivm_state *gallivm, struct lp_type type) 223{ 224 LLVMTypeRef vec_type = lp_build_vec_type(gallivm, type); 225 return LLVMGetUndef(vec_type); 226} 227 228 229LLVMValueRef 230lp_build_zero(struct gallivm_state *gallivm, struct lp_type type) 231{ 232 if (type.length == 1) { 233 if (type.floating) 234 return lp_build_const_float(gallivm, 0.0); 235 else 236 return LLVMConstInt(LLVMIntTypeInContext(gallivm->context, type.width), 0, 0); 237 } else { 238 LLVMTypeRef vec_type = lp_build_vec_type(gallivm, type); 239 return LLVMConstNull(vec_type); 240 } 241} 242 243 244LLVMValueRef 245lp_build_one(struct gallivm_state *gallivm, struct lp_type type) 246{ 247 LLVMTypeRef elem_type; 248 LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; 249 250 assert(type.length <= LP_MAX_VECTOR_LENGTH); 251 252 elem_type = lp_build_elem_type(gallivm, type); 253 254 if (!lp_has_fp16() && type.floating && type.width == 16) 255 elems[0] = LLVMConstInt(elem_type, _mesa_float_to_half(1.0f), 0); 256 else if (type.floating) 257 elems[0] = LLVMConstReal(elem_type, 1.0); 258 else if (type.fixed) 259 elems[0] = LLVMConstInt(elem_type, 1LL << (type.width/2), 0); 260 else if (!type.norm) 261 elems[0] = LLVMConstInt(elem_type, 1, 0); 262 else if (type.sign) 263 elems[0] = LLVMConstInt(elem_type, (1LL << (type.width - 1)) - 1, 0); 264 else { 265 /* special case' -- 1.0 for normalized types is more easily attained if 266 * we start with a vector consisting of all bits set */ 267 LLVMTypeRef vec_type = lp_build_vec_type(gallivm, type); 268 LLVMValueRef vec = LLVMConstAllOnes(vec_type); 269 270#if 0 271 if (type.sign) 272 /* TODO: Unfortunately this caused "Tried to create a shift operation 273 * on a non-integer type!" */ 274 vec = LLVMConstLShr(vec, lp_build_const_int_vec(type, 1)); 275#endif 276 277 return vec; 278 } 279 280 for (unsigned i = 1; i < type.length; ++i) 281 elems[i] = elems[0]; 282 283 if (type.length == 1) 284 return elems[0]; 285 else 286 return LLVMConstVector(elems, type.length); 287} 288 289 290/** 291 * Build constant-valued element from a scalar value. 292 */ 293LLVMValueRef 294lp_build_const_elem(struct gallivm_state *gallivm, 295 struct lp_type type, 296 double val) 297{ 298 LLVMTypeRef elem_type = lp_build_elem_type(gallivm, type); 299 LLVMValueRef elem; 300 301 if (!lp_has_fp16() && type.floating && type.width == 16) { 302 elem = LLVMConstInt(elem_type, _mesa_float_to_half((float)val), 0); 303 } else if (type.floating) { 304 elem = LLVMConstReal(elem_type, val); 305 } else { 306 double dscale = lp_const_scale(type); 307 308 elem = LLVMConstInt(elem_type, (long long) round(val*dscale), 0); 309 } 310 311 return elem; 312} 313 314 315/** 316 * Build constant-valued vector from a scalar value. 317 */ 318LLVMValueRef 319lp_build_const_vec(struct gallivm_state *gallivm, struct lp_type type, 320 double val) 321{ 322 if (type.length == 1) { 323 return lp_build_const_elem(gallivm, type, val); 324 } else { 325 LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; 326 elems[0] = lp_build_const_elem(gallivm, type, val); 327 for (unsigned i = 1; i < type.length; ++i) 328 elems[i] = elems[0]; 329 return LLVMConstVector(elems, type.length); 330 } 331} 332 333 334LLVMValueRef 335lp_build_const_int_vec(struct gallivm_state *gallivm, struct lp_type type, 336 long long val) 337{ 338 LLVMTypeRef elem_type = lp_build_int_elem_type(gallivm, type); 339 LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; 340 341 assert(type.length <= LP_MAX_VECTOR_LENGTH); 342 343 for (unsigned i = 0; i < type.length; ++i) 344 elems[i] = LLVMConstInt(elem_type, val, type.sign ? 1 : 0); 345 346 if (type.length == 1) 347 return elems[0]; 348 349 return LLVMConstVector(elems, type.length); 350} 351 352 353LLVMValueRef 354lp_build_const_aos(struct gallivm_state *gallivm, 355 struct lp_type type, 356 double r, double g, double b, double a, 357 const unsigned char *swizzle) 358{ 359 const unsigned char default_swizzle[4] = {0, 1, 2, 3}; 360 LLVMValueRef elems[LP_MAX_VECTOR_LENGTH]; 361 362 assert(type.length % 4 == 0); 363 assert(type.length <= LP_MAX_VECTOR_LENGTH); 364 365 lp_build_elem_type(gallivm, type); 366 367 if (!swizzle) 368 swizzle = default_swizzle; 369 370 elems[swizzle[0]] = lp_build_const_elem(gallivm, type, r); 371 elems[swizzle[1]] = lp_build_const_elem(gallivm, type, g); 372 elems[swizzle[2]] = lp_build_const_elem(gallivm, type, b); 373 elems[swizzle[3]] = lp_build_const_elem(gallivm, type, a); 374 375 for (unsigned i = 4; i < type.length; ++i) 376 elems[i] = elems[i % 4]; 377 378 return LLVMConstVector(elems, type.length); 379} 380 381 382/** 383 * @param mask TGSI_WRITEMASK_xxx 384 */ 385LLVMValueRef 386lp_build_const_mask_aos(struct gallivm_state *gallivm, 387 struct lp_type type, 388 unsigned mask, 389 unsigned channels) 390{ 391 LLVMTypeRef elem_type = LLVMIntTypeInContext(gallivm->context, type.width); 392 LLVMValueRef masks[LP_MAX_VECTOR_LENGTH]; 393 394 assert(type.length <= LP_MAX_VECTOR_LENGTH); 395 396 for (unsigned j = 0; j < type.length; j += channels) { 397 for (unsigned i = 0; i < channels; ++i) { 398 masks[j + i] = LLVMConstInt(elem_type, 399 mask & (1 << i) ? ~0ULL : 0, 400 1); 401 } 402 } 403 404 return LLVMConstVector(masks, type.length); 405} 406 407 408/** 409 * Performs lp_build_const_mask_aos, but first swizzles the mask 410 */ 411LLVMValueRef 412lp_build_const_mask_aos_swizzled(struct gallivm_state *gallivm, 413 struct lp_type type, 414 unsigned mask, 415 unsigned channels, 416 const unsigned char *swizzle) 417{ 418 unsigned mask_swizzled = 0; 419 420 for (unsigned i = 0; i < channels; ++i) { 421 if (swizzle[i] < 4) { 422 mask_swizzled |= ((mask & (1 << swizzle[i])) >> swizzle[i]) << i; 423 } 424 } 425 426 return lp_build_const_mask_aos(gallivm, type, mask_swizzled, channels); 427} 428 429 430/** 431 * Build a zero-terminated constant string. 432 */ 433LLVMValueRef 434lp_build_const_string(struct gallivm_state *gallivm, 435 const char *str) 436{ 437 unsigned len = strlen(str) + 1; 438 LLVMTypeRef i8 = LLVMInt8TypeInContext(gallivm->context); 439 LLVMValueRef string = LLVMAddGlobal(gallivm->module, LLVMArrayType(i8, len), ""); 440 LLVMSetGlobalConstant(string, TRUE); 441 LLVMSetLinkage(string, LLVMInternalLinkage); 442 LLVMSetInitializer(string, LLVMConstStringInContext(gallivm->context, str, len, TRUE)); 443 string = LLVMConstBitCast(string, LLVMPointerType(i8, 0)); 444 return string; 445} 446 447 448LLVMValueRef 449lp_build_const_func_pointer_from_type(struct gallivm_state *gallivm, 450 const void *ptr, 451 LLVMTypeRef function_type, 452 const char *name) 453{ 454 return LLVMBuildBitCast(gallivm->builder, 455 lp_build_const_int_pointer(gallivm, ptr), 456 LLVMPointerType(function_type, 0), 457 name); 458} 459 460 461/** 462 * Build a callable function pointer. 463 * 464 * We use function pointer constants instead of LLVMAddGlobalMapping() 465 * to work around a bug in LLVM 2.6, and for efficiency/simplicity. 466 */ 467LLVMValueRef 468lp_build_const_func_pointer(struct gallivm_state *gallivm, 469 const void *ptr, 470 LLVMTypeRef ret_type, 471 LLVMTypeRef *arg_types, 472 unsigned num_args, 473 const char *name) 474{ 475 LLVMTypeRef function_type = LLVMFunctionType(ret_type, arg_types, num_args, 0); 476 return lp_build_const_func_pointer_from_type(gallivm, ptr, function_type, name); 477} 478