src/gallium/auxiliary/gallivm/lp_bld_pack.c

   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 packing/unpacking.
  32  *
  33  * Pack/unpacking is necessary for conversion between types of different
  34  * bit width.
  35  *
  36  * They are also commonly used when an computation needs higher
  37  * precision for the intermediate values. For example, if one needs the
  38  * function:
  39  *
  40  *   c = compute(a, b);
  41  *
  42  * to use more precision for intermediate results then one should implement it
  43  * as:
  44  *
  45  *   LLVMValueRef
  46  *   compute(LLVMBuilderRef builder struct lp_type type, LLVMValueRef a, LLVMValueRef b)
  47  *   {
  48  *      struct lp_type wide_type = lp_wider_type(type);
  49  *      LLVMValueRef al, ah, bl, bh, cl, ch, c;
  50  *
  51  *      lp_build_unpack2(builder, type, wide_type, a, &al, &ah);
  52  *      lp_build_unpack2(builder, type, wide_type, b, &bl, &bh);
  53  *
  54  *      cl = compute_half(al, bl);
  55  *      ch = compute_half(ah, bh);
  56  *
  57  *      c = lp_build_pack2(bld->builder, wide_type, type, cl, ch);
  58  *
  59  *      return c;
  60  *   }
  61  *
  62  * where compute_half() would do the computation for half the elements with
  63  * twice the precision.
  64  *
  65  * @author Jose Fonseca <jfonseca@vmware.com>
  66  */
  67
  68
  69 #include "util/u_debug.h"
  70 #include "util/u_math.h"
  71 #include "util/u_cpu_detect.h"
  72
  73 #include "lp_bld_type.h"
  74 #include "lp_bld_const.h"
  75 #include "lp_bld_intr.h"
  76 #include "lp_bld_arit.h"
  77 #include "lp_bld_pack.h"
  78
  79
  80 /**
  81  * Build shuffle vectors that match PUNPCKLxx and PUNPCKHxx instructions.
  82  */
  83 static LLVMValueRef
  84 lp_build_const_unpack_shuffle(unsigned n, unsigned lo_hi)
  85 {
  86    LLVMValueRef elems[LP_MAX_VECTOR_LENGTH];
  87    unsigned i, j;
  88
  89    assert(n <= LP_MAX_VECTOR_LENGTH);
  90    assert(lo_hi < 2);
  91
  92    /* TODO: cache results in a static table */
  93
  94    for(i = 0, j = lo_hi*n/2; i < n; i += 2, ++j) {
  95       elems[i + 0] = LLVMConstInt(LLVMInt32Type(), 0 + j, 0);
  96       elems[i + 1] = LLVMConstInt(LLVMInt32Type(), n + j, 0);
  97    }
  98
  99    return LLVMConstVector(elems, n);
 100 }
 101
 102
 103 /**
 104  * Build shuffle vectors that match PACKxx instructions.
 105  */
 106 static LLVMValueRef
 107 lp_build_const_pack_shuffle(unsigned n)
 108 {
 109    LLVMValueRef elems[LP_MAX_VECTOR_LENGTH];
 110    unsigned i;
 111
 112    assert(n <= LP_MAX_VECTOR_LENGTH);
 113
 114    /* TODO: cache results in a static table */
 115
 116    for(i = 0; i < n; ++i)
 117       elems[i] = LLVMConstInt(LLVMInt32Type(), 2*i, 0);
 118
 119    return LLVMConstVector(elems, n);
 120 }
 121
 122
 123 /**
 124  * Interleave vector elements.
 125  *
 126  * Matches the PUNPCKLxx and PUNPCKHxx SSE instructions.
 127  */
 128 LLVMValueRef
 129 lp_build_interleave2(LLVMBuilderRef builder,
 130                      struct lp_type type,
 131                      LLVMValueRef a,
 132                      LLVMValueRef b,
 133                      unsigned lo_hi)
 134 {
 135    LLVMValueRef shuffle;
 136
 137    shuffle = lp_build_const_unpack_shuffle(type.length, lo_hi);
 138
 139    return LLVMBuildShuffleVector(builder, a, b, shuffle, "");
 140 }
 141
 142
 143 /**
 144  * Double the bit width.
 145  *
 146  * This will only change the number of bits the values are represented, not the
 147  * values themselves.
 148  */
 149 void
 150 lp_build_unpack2(LLVMBuilderRef builder,
 151                  struct lp_type src_type,
 152                  struct lp_type dst_type,
 153                  LLVMValueRef src,
 154                  LLVMValueRef *dst_lo,
 155                  LLVMValueRef *dst_hi)
 156 {
 157    LLVMValueRef msb;
 158    LLVMTypeRef dst_vec_type;
 159
 160    assert(!src_type.floating);
 161    assert(!dst_type.floating);
 162    assert(dst_type.width == src_type.width * 2);
 163    assert(dst_type.length * 2 == src_type.length);
 164
 165    if(dst_type.sign && src_type.sign) {
 166       /* Replicate the sign bit in the most significant bits */
 167       msb = LLVMBuildAShr(builder, src, lp_build_const_int_vec(src_type, src_type.width - 1), "");
 168    }
 169    else
 170       /* Most significant bits always zero */
 171       msb = lp_build_zero(src_type);
 172
 173    /* Interleave bits */
 174 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 175       *dst_lo = lp_build_interleave2(builder, src_type, src, msb, 0);
 176       *dst_hi = lp_build_interleave2(builder, src_type, src, msb, 1);
 177 #else
 178       *dst_lo = lp_build_interleave2(builder, src_type, msb, src, 0);
 179       *dst_hi = lp_build_interleave2(builder, src_type, msb, src, 1);
 180 #endif
 181
 182    /* Cast the result into the new type (twice as wide) */
 183
 184    dst_vec_type = lp_build_vec_type(dst_type);
 185
 186    *dst_lo = LLVMBuildBitCast(builder, *dst_lo, dst_vec_type, "");
 187    *dst_hi = LLVMBuildBitCast(builder, *dst_hi, dst_vec_type, "");
 188 }
 189
 190
 191 /**
 192  * Expand the bit width.
 193  *
 194  * This will only change the number of bits the values are represented, not the
 195  * values themselves.
 196  */
 197 void
 198 lp_build_unpack(LLVMBuilderRef builder,
 199                 struct lp_type src_type,
 200                 struct lp_type dst_type,
 201                 LLVMValueRef src,
 202                 LLVMValueRef *dst, unsigned num_dsts)
 203 {
 204    unsigned num_tmps;
 205    unsigned i;
 206
 207    /* Register width must remain constant */
 208    assert(src_type.width * src_type.length == dst_type.width * dst_type.length);
 209
 210    /* We must not loose or gain channels. Only precision */
 211    assert(src_type.length == dst_type.length * num_dsts);
 212
 213    num_tmps = 1;
 214    dst[0] = src;
 215
 216    while(src_type.width < dst_type.width) {
 217       struct lp_type tmp_type = src_type;
 218
 219       tmp_type.width *= 2;
 220       tmp_type.length /= 2;
 221
 222       for(i = num_tmps; i--; ) {
 223          lp_build_unpack2(builder, src_type, tmp_type, dst[i], &dst[2*i + 0], &dst[2*i + 1]);
 224       }
 225
 226       src_type = tmp_type;
 227
 228       num_tmps *= 2;
 229    }
 230
 231    assert(num_tmps == num_dsts);
 232 }
 233
 234
 235 /**
 236  * Non-interleaved pack.
 237  *
 238  * This will move values as
 239  *
 240  *   lo =   __ l0 __ l1 __ l2 __..  __ ln
 241  *   hi =   __ h0 __ h1 __ h2 __..  __ hn
 242  *   res =  l0 l1 l2 .. ln h0 h1 h2 .. hn
 243  *
 244  * This will only change the number of bits the values are represented, not the
 245  * values themselves.
 246  *
 247  * It is assumed the values are already clamped into the destination type range.
 248  * Values outside that range will produce undefined results. Use
 249  * lp_build_packs2 instead.
 250  */
 251 LLVMValueRef
 252 lp_build_pack2(LLVMBuilderRef builder,
 253                struct lp_type src_type,
 254                struct lp_type dst_type,
 255                LLVMValueRef lo,
 256                LLVMValueRef hi)
 257 {
 258 #if HAVE_LLVM < 0x0207
 259    LLVMTypeRef src_vec_type = lp_build_vec_type(src_type);
 260 #endif
 261    LLVMTypeRef dst_vec_type = lp_build_vec_type(dst_type);
 262    LLVMValueRef shuffle;
 263    LLVMValueRef res = NULL;
 264
 265    assert(!src_type.floating);
 266    assert(!dst_type.floating);
 267    assert(src_type.width == dst_type.width * 2);
 268    assert(src_type.length * 2 == dst_type.length);
 269
 270    /* Check for special cases first */
 271    if(util_cpu_caps.has_sse2 && src_type.width * src_type.length == 128) {
 272       switch(src_type.width) {
 273       case 32:
 274          if(dst_type.sign) {
 275 #if HAVE_LLVM >= 0x0207
 276             res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packssdw.128", dst_vec_type, lo, hi);
 277 #else
 278             res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packssdw.128", src_vec_type, lo, hi);
 279 #endif
 280          }
 281          else {
 282             if (util_cpu_caps.has_sse4_1) {
 283                return lp_build_intrinsic_binary(builder, "llvm.x86.sse41.packusdw", dst_vec_type, lo, hi);
 284             }
 285             else {
 286                /* use generic shuffle below */
 287                res = NULL;
 288             }
 289          }
 290          break;
 291
 292       case 16:
 293          if(dst_type.sign)
 294 #if HAVE_LLVM >= 0x0207
 295             res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packsswb.128", dst_vec_type, lo, hi);
 296 #else
 297             res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packsswb.128", src_vec_type, lo, hi);
 298 #endif
 299          else
 300 #if HAVE_LLVM >= 0x0207
 301             res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packuswb.128", dst_vec_type, lo, hi);
 302 #else
 303             res = lp_build_intrinsic_binary(builder, "llvm.x86.sse2.packuswb.128", src_vec_type, lo, hi);
 304 #endif
 305          break;
 306
 307       default:
 308          assert(0);
 309          return LLVMGetUndef(dst_vec_type);
 310          break;
 311       }
 312
 313       if (res) {
 314          res = LLVMBuildBitCast(builder, res, dst_vec_type, "");
 315          return res;
 316       }
 317    }
 318
 319    /* generic shuffle */
 320    lo = LLVMBuildBitCast(builder, lo, dst_vec_type, "");
 321    hi = LLVMBuildBitCast(builder, hi, dst_vec_type, "");
 322
 323    shuffle = lp_build_const_pack_shuffle(dst_type.length);
 324
 325    res = LLVMBuildShuffleVector(builder, lo, hi, shuffle, "");
 326
 327    return res;
 328 }
 329
 330
 331
 332 /**
 333  * Non-interleaved pack and saturate.
 334  *
 335  * Same as lp_build_pack2 but will saturate values so that they fit into the
 336  * destination type.
 337  */
 338 LLVMValueRef
 339 lp_build_packs2(LLVMBuilderRef builder,
 340                 struct lp_type src_type,
 341                 struct lp_type dst_type,
 342                 LLVMValueRef lo,
 343                 LLVMValueRef hi)
 344 {
 345    boolean clamp;
 346
 347    assert(!src_type.floating);
 348    assert(!dst_type.floating);
 349    assert(src_type.sign == dst_type.sign);
 350    assert(src_type.width == dst_type.width * 2);
 351    assert(src_type.length * 2 == dst_type.length);
 352
 353    clamp = TRUE;
 354
 355    /* All X86 SSE non-interleaved pack instructions take signed inputs and
 356     * saturate them, so no need to clamp for those cases. */
 357    if(util_cpu_caps.has_sse2 &&
 358       src_type.width * src_type.length == 128 &&
 359       src_type.sign)
 360       clamp = FALSE;
 361
 362    if(clamp) {
 363       struct lp_build_context bld;
 364       unsigned dst_bits = dst_type.sign ? dst_type.width - 1 : dst_type.width;
 365       LLVMValueRef dst_max = lp_build_const_int_vec(src_type, ((unsigned long long)1 << dst_bits) - 1);
 366       lp_build_context_init(&bld, builder, src_type);
 367       lo = lp_build_min(&bld, lo, dst_max);
 368       hi = lp_build_min(&bld, hi, dst_max);
 369       /* FIXME: What about lower bound? */
 370    }
 371
 372    return lp_build_pack2(builder, src_type, dst_type, lo, hi);
 373 }
 374
 375
 376 /**
 377  * Truncate the bit width.
 378  *
 379  * TODO: Handle saturation consistently.
 380  */
 381 LLVMValueRef
 382 lp_build_pack(LLVMBuilderRef builder,
 383               struct lp_type src_type,
 384               struct lp_type dst_type,
 385               boolean clamped,
 386               const LLVMValueRef *src, unsigned num_srcs)
 387 {
 388    LLVMValueRef (*pack2)(LLVMBuilderRef builder,
 389                          struct lp_type src_type,
 390                          struct lp_type dst_type,
 391                          LLVMValueRef lo,
 392                          LLVMValueRef hi);
 393    LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH];
 394    unsigned i;
 395
 396
 397    /* Register width must remain constant */
 398    assert(src_type.width * src_type.length == dst_type.width * dst_type.length);
 399
 400    /* We must not loose or gain channels. Only precision */
 401    assert(src_type.length * num_srcs == dst_type.length);
 402
 403    if(clamped)
 404       pack2 = &lp_build_pack2;
 405    else
 406       pack2 = &lp_build_packs2;
 407
 408    for(i = 0; i < num_srcs; ++i)
 409       tmp[i] = src[i];
 410
 411    while(src_type.width > dst_type.width) {
 412       struct lp_type tmp_type = src_type;
 413
 414       tmp_type.width /= 2;
 415       tmp_type.length *= 2;
 416
 417       /* Take in consideration the sign changes only in the last step */
 418       if(tmp_type.width == dst_type.width)
 419          tmp_type.sign = dst_type.sign;
 420
 421       num_srcs /= 2;
 422
 423       for(i = 0; i < num_srcs; ++i)
 424          tmp[i] = pack2(builder, src_type, tmp_type, tmp[2*i + 0], tmp[2*i + 1]);
 425
 426       src_type = tmp_type;
 427    }
 428
 429    assert(num_srcs == 1);
 430
 431    return tmp[0];
 432 }
 433
 434
 435 /**
 436  * Truncate or expand the bitwidth.
 437  *
 438  * NOTE: Getting the right sign flags is crucial here, as we employ some
 439  * intrinsics that do saturation.
 440  */
 441 void
 442 lp_build_resize(LLVMBuilderRef builder,
 443                 struct lp_type src_type,
 444                 struct lp_type dst_type,
 445                 const LLVMValueRef *src, unsigned num_srcs,
 446                 LLVMValueRef *dst, unsigned num_dsts)
 447 {
 448    LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH];
 449    unsigned i;
 450
 451    /*
 452     * We don't support float <-> int conversion here. That must be done
 453     * before/after calling this function.
 454     */
 455    assert(src_type.floating == dst_type.floating);
 456
 457    /*
 458     * We don't support double <-> float conversion yet, although it could be
 459     * added with little effort.
 460     */
 461    assert((!src_type.floating && !dst_type.floating) ||
 462           src_type.width == dst_type.width);
 463
 464    /* We must not loose or gain channels. Only precision */
 465    assert(src_type.length * num_srcs == dst_type.length * num_dsts);
 466
 467    /* We don't support M:N conversion, only 1:N, M:1, or 1:1 */
 468    assert(num_srcs == 1 || num_dsts == 1);
 469
 470    assert(src_type.length <= LP_MAX_VECTOR_LENGTH);
 471    assert(dst_type.length <= LP_MAX_VECTOR_LENGTH);
 472    assert(num_srcs <= LP_MAX_VECTOR_LENGTH);
 473    assert(num_dsts <= LP_MAX_VECTOR_LENGTH);
 474
 475    if (src_type.width > dst_type.width) {
 476       /*
 477        * Truncate bit width.
 478        */
 479
 480       assert(num_dsts == 1);
 481
 482       if (src_type.width * src_type.length == dst_type.width * dst_type.length) {
 483         /*
 484          * Register width remains constant -- use vector packing intrinsics
 485          */
 486
 487          tmp[0] = lp_build_pack(builder, src_type, dst_type, TRUE, src, num_srcs);
 488       }
 489       else {
 490          /*
 491           * Do it element-wise.
 492           */
 493
 494          assert(src_type.length == dst_type.length);
 495          tmp[0] = lp_build_undef(dst_type);
 496          for (i = 0; i < dst_type.length; ++i) {
 497             LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 498             LLVMValueRef val = LLVMBuildExtractElement(builder, src[0], index, "");
 499             val = LLVMBuildTrunc(builder, val, lp_build_elem_type(dst_type), "");
 500             tmp[0] = LLVMBuildInsertElement(builder, tmp[0], val, index, "");
 501          }
 502       }
 503    }
 504    else if (src_type.width < dst_type.width) {
 505       /*
 506        * Expand bit width.
 507        */
 508
 509       assert(num_srcs == 1);
 510
 511       if (src_type.width * src_type.length == dst_type.width * dst_type.length) {
 512          /*
 513           * Register width remains constant -- use vector unpack intrinsics
 514           */
 515          lp_build_unpack(builder, src_type, dst_type, src[0], tmp, num_dsts);
 516       }
 517       else {
 518          /*
 519           * Do it element-wise.
 520           */
 521
 522          assert(src_type.length == dst_type.length);
 523          tmp[0] = lp_build_undef(dst_type);
 524          for (i = 0; i < dst_type.length; ++i) {
 525             LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 526             LLVMValueRef val = LLVMBuildExtractElement(builder, src[0], index, "");
 527
 528             if (src_type.sign && dst_type.sign) {
 529                val = LLVMBuildSExt(builder, val, lp_build_elem_type(dst_type), "");
 530             } else {
 531                val = LLVMBuildZExt(builder, val, lp_build_elem_type(dst_type), "");
 532             }
 533             tmp[0] = LLVMBuildInsertElement(builder, tmp[0], val, index, "");
 534          }
 535       }
 536    }
 537    else {
 538       /*
 539        * No-op
 540        */
 541
 542       assert(num_srcs == 1);
 543       assert(num_dsts == 1);
 544
 545       tmp[0] = src[0];
 546    }
 547
 548    for(i = 0; i < num_dsts; ++i)
 549       dst[i] = tmp[i];
 550 }
 551
 552