src/gallium/auxiliary/gallivm/lp_bld_swizzle.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  * @file
  30  * Helper functions for swizzling/shuffling.
  31  *
  32  * @author Jose Fonseca <jfonseca@vmware.com>
  33  */
  34
  35
  36 #include "util/u_debug.h"
  37
  38 #include "lp_bld_type.h"
  39 #include "lp_bld_const.h"
  40 #include "lp_bld_logic.h"
  41 #include "lp_bld_swizzle.h"
  42
  43
  44 LLVMValueRef
  45 lp_build_broadcast(LLVMBuilderRef builder,
  46                    LLVMTypeRef vec_type,
  47                    LLVMValueRef scalar)
  48 {
  49    const unsigned n = LLVMGetVectorSize(vec_type);
  50    LLVMValueRef res;
  51    unsigned i;
  52
  53    res = LLVMGetUndef(vec_type);
  54    for(i = 0; i < n; ++i) {
  55       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
  56       res = LLVMBuildInsertElement(builder, res, scalar, index, "");
  57    }
  58
  59    return res;
  60 }
  61
  62
  63 /**
  64  * Broadcast
  65  */
  66 LLVMValueRef
  67 lp_build_broadcast_scalar(struct lp_build_context *bld,
  68                           LLVMValueRef scalar)
  69 {
  70    const struct lp_type type = bld->type;
  71
  72    assert(lp_check_elem_type(type, LLVMTypeOf(scalar)));
  73
  74    if (type.length == 1) {
  75       return scalar;
  76    }
  77    else {
  78       LLVMValueRef res;
  79 #if HAVE_LLVM >= 0x207
  80       res = LLVMBuildInsertElement(bld->builder, bld->undef, scalar,
  81                                    LLVMConstInt(LLVMInt32Type(), 0, 0), "");
  82       res = LLVMBuildShuffleVector(bld->builder, res, bld->undef,
  83                                    lp_build_const_int_vec(type, 0), "");
  84 #else
  85       /* XXX: The above path provokes a bug in LLVM 2.6 */
  86       unsigned i;
  87       res = bld->undef;
  88       for(i = 0; i < type.length; ++i) {
  89          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
  90          res = LLVMBuildInsertElement(bld->builder, res, scalar, index, "");
  91       }
  92 #endif
  93       return res;
  94    }
  95 }
  96
  97
  98 /**
  99  * Swizzle one channel into all other three channels.
 100  */
 101 LLVMValueRef
 102 lp_build_swizzle_scalar_aos(struct lp_build_context *bld,
 103                             LLVMValueRef a,
 104                             unsigned channel)
 105 {
 106    const struct lp_type type = bld->type;
 107    const unsigned n = type.length;
 108    unsigned i, j;
 109
 110    if(a == bld->undef || a == bld->zero || a == bld->one)
 111       return a;
 112
 113    /* XXX: SSE3 has PSHUFB which should be better than bitmasks, but forcing
 114     * using shuffles here actually causes worst results. More investigation is
 115     * needed. */
 116    if (type.width >= 16) {
 117       /*
 118        * Shuffle.
 119        */
 120       LLVMTypeRef elem_type = LLVMInt32Type();
 121       LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
 122
 123       for(j = 0; j < n; j += 4)
 124          for(i = 0; i < 4; ++i)
 125             shuffles[j + i] = LLVMConstInt(elem_type, j + channel, 0);
 126
 127       return LLVMBuildShuffleVector(bld->builder, a, bld->undef, LLVMConstVector(shuffles, n), "");
 128    }
 129    else {
 130       /*
 131        * Bit mask and recursive shifts
 132        *
 133        *   XYZW XYZW .... XYZW  <= input
 134        *   0Y00 0Y00 .... 0Y00
 135        *   YY00 YY00 .... YY00
 136        *   YYYY YYYY .... YYYY  <= output
 137        */
 138       struct lp_type type4;
 139       const char shifts[4][2] = {
 140          { 1,  2},
 141          {-1,  2},
 142          { 1, -2},
 143          {-1, -2}
 144       };
 145       unsigned i;
 146
 147       a = LLVMBuildAnd(bld->builder, a,
 148                        lp_build_const_mask_aos(type, 1 << channel), "");
 149
 150       /*
 151        * Build a type where each element is an integer that cover the four
 152        * channels.
 153        */
 154
 155       type4 = type;
 156       type4.floating = FALSE;
 157       type4.width *= 4;
 158       type4.length /= 4;
 159
 160       a = LLVMBuildBitCast(bld->builder, a, lp_build_vec_type(type4), "");
 161
 162       for(i = 0; i < 2; ++i) {
 163          LLVMValueRef tmp = NULL;
 164          int shift = shifts[channel][i];
 165
 166 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 167          shift = -shift;
 168 #endif
 169
 170          if(shift > 0)
 171             tmp = LLVMBuildLShr(bld->builder, a, lp_build_const_int_vec(type4, shift*type.width), "");
 172          if(shift < 0)
 173             tmp = LLVMBuildShl(bld->builder, a, lp_build_const_int_vec(type4, -shift*type.width), "");
 174
 175          assert(tmp);
 176          if(tmp)
 177             a = LLVMBuildOr(bld->builder, a, tmp, "");
 178       }
 179
 180       return LLVMBuildBitCast(bld->builder, a, lp_build_vec_type(type), "");
 181    }
 182 }
 183
 184
 185 LLVMValueRef
 186 lp_build_swizzle_aos(struct lp_build_context *bld,
 187                      LLVMValueRef a,
 188                      const unsigned char swizzles[4])
 189 {
 190    const struct lp_type type = bld->type;
 191    const unsigned n = type.length;
 192    unsigned i, j;
 193
 194    if (swizzles[0] == PIPE_SWIZZLE_RED &&
 195        swizzles[1] == PIPE_SWIZZLE_GREEN &&
 196        swizzles[2] == PIPE_SWIZZLE_BLUE &&
 197        swizzles[3] == PIPE_SWIZZLE_ALPHA) {
 198       return a;
 199    }
 200
 201    if (swizzles[0] == swizzles[1] &&
 202        swizzles[1] == swizzles[2] &&
 203        swizzles[2] == swizzles[3]) {
 204       switch (swizzles[0]) {
 205       case PIPE_SWIZZLE_RED:
 206       case PIPE_SWIZZLE_GREEN:
 207       case PIPE_SWIZZLE_BLUE:
 208       case PIPE_SWIZZLE_ALPHA:
 209          return lp_build_swizzle_scalar_aos(bld, a, swizzles[0]);
 210       case PIPE_SWIZZLE_ZERO:
 211          return bld->zero;
 212       case PIPE_SWIZZLE_ONE:
 213          return bld->one;
 214       default:
 215          assert(0);
 216          return bld->undef;
 217       }
 218    }
 219
 220    if (type.width >= 16) {
 221       /*
 222        * Shuffle.
 223        */
 224       LLVMValueRef undef = LLVMGetUndef(lp_build_elem_type(type));
 225       LLVMTypeRef i32t = LLVMInt32Type();
 226       LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
 227       LLVMValueRef aux[LP_MAX_VECTOR_LENGTH];
 228
 229       memset(aux, 0, sizeof aux);
 230
 231       for(j = 0; j < n; j += 4) {
 232          for(i = 0; i < 4; ++i) {
 233             unsigned shuffle;
 234             switch (swizzles[i]) {
 235             default:
 236                assert(0);
 237                /* fall through */
 238             case PIPE_SWIZZLE_RED:
 239             case PIPE_SWIZZLE_GREEN:
 240             case PIPE_SWIZZLE_BLUE:
 241             case PIPE_SWIZZLE_ALPHA:
 242                shuffle = j + swizzles[i];
 243                break;
 244             case PIPE_SWIZZLE_ZERO:
 245                shuffle = type.length + 0;
 246                if (!aux[0]) {
 247                   aux[0] = lp_build_const_elem(type, 0.0);
 248                }
 249                break;
 250             case PIPE_SWIZZLE_ONE:
 251                shuffle = type.length + 1;
 252                if (!aux[1]) {
 253                   aux[1] = lp_build_const_elem(type, 1.0);
 254                }
 255                break;
 256             }
 257             shuffles[j + i] = LLVMConstInt(i32t, shuffle, 0);
 258          }
 259       }
 260
 261       for (i = 0; i < n; ++i) {
 262          if (!aux[i]) {
 263             aux[i] = undef;
 264          }
 265       }
 266
 267       return LLVMBuildShuffleVector(bld->builder, a,
 268                                     LLVMConstVector(aux, n),
 269                                     LLVMConstVector(shuffles, n), "");
 270    } else {
 271       /*
 272        * Bit mask and shifts.
 273        *
 274        * For example, this will convert BGRA to RGBA by doing
 275        *
 276        *   rgba = (bgra & 0x00ff0000) >> 16
 277        *        | (bgra & 0xff00ff00)
 278        *        | (bgra & 0x000000ff) << 16
 279        *
 280        * This is necessary not only for faster cause, but because X86 backend
 281        * will refuse shuffles of <4 x i8> vectors
 282        */
 283       LLVMValueRef res;
 284       struct lp_type type4;
 285       unsigned cond = 0;
 286       unsigned chan;
 287       int shift;
 288
 289       /*
 290        * Start with a mixture of 1 and 0.
 291        */
 292       for (chan = 0; chan < 4; ++chan) {
 293          if (swizzles[chan] == PIPE_SWIZZLE_ONE) {
 294             cond |= 1 << chan;
 295          }
 296       }
 297       res = lp_build_select_aos(bld, cond, bld->one, bld->zero);
 298
 299       /*
 300        * Build a type where each element is an integer that cover the four
 301        * channels.
 302        */
 303       type4 = type;
 304       type4.floating = FALSE;
 305       type4.width *= 4;
 306       type4.length /= 4;
 307
 308       a = LLVMBuildBitCast(bld->builder, a, lp_build_vec_type(type4), "");
 309       res = LLVMBuildBitCast(bld->builder, res, lp_build_vec_type(type4), "");
 310
 311       /*
 312        * Mask and shift the channels, trying to group as many channels in the
 313        * same shift as possible
 314        */
 315       for (shift = -3; shift <= 3; ++shift) {
 316          unsigned long long mask = 0;
 317
 318          assert(type4.width <= sizeof(mask)*8);
 319
 320          for (chan = 0; chan < 4; ++chan) {
 321             /* FIXME: big endian */
 322             if (swizzles[chan] < 4 &&
 323                 chan - swizzles[chan] == shift) {
 324                mask |= ((1ULL << type.width) - 1) << (swizzles[chan] * type.width);
 325             }
 326          }
 327
 328          if (mask) {
 329             LLVMValueRef masked;
 330             LLVMValueRef shifted;
 331
 332             if (0)
 333                debug_printf("shift = %i, mask = 0x%08llx\n", shift, mask);
 334
 335             masked = LLVMBuildAnd(bld->builder, a,
 336                                   lp_build_const_int_vec(type4, mask), "");
 337             if (shift > 0) {
 338                shifted = LLVMBuildShl(bld->builder, masked,
 339                                       lp_build_const_int_vec(type4, shift*type.width), "");
 340             } else if (shift < 0) {
 341                shifted = LLVMBuildLShr(bld->builder, masked,
 342                                        lp_build_const_int_vec(type4, -shift*type.width), "");
 343             } else {
 344                shifted = masked;
 345             }
 346
 347             res = LLVMBuildOr(bld->builder, res, shifted, "");
 348          }
 349       }
 350
 351       return LLVMBuildBitCast(bld->builder, res, lp_build_vec_type(type), "");
 352    }
 353 }
 354
 355
 356 /**
 357  * Extended swizzle of a single channel of a SoA vector.
 358  *
 359  * @param bld         building context
 360  * @param unswizzled  array with the 4 unswizzled values
 361  * @param swizzle     one of the PIPE_SWIZZLE_*
 362  *
 363  * @return  the swizzled value.
 364  */
 365 LLVMValueRef
 366 lp_build_swizzle_soa_channel(struct lp_build_context *bld,
 367                              const LLVMValueRef *unswizzled,
 368                              unsigned swizzle)
 369 {
 370    switch (swizzle) {
 371    case PIPE_SWIZZLE_RED:
 372    case PIPE_SWIZZLE_GREEN:
 373    case PIPE_SWIZZLE_BLUE:
 374    case PIPE_SWIZZLE_ALPHA:
 375       return unswizzled[swizzle];
 376    case PIPE_SWIZZLE_ZERO:
 377       return bld->zero;
 378    case PIPE_SWIZZLE_ONE:
 379       return bld->one;
 380    default:
 381       assert(0);
 382       return bld->undef;
 383    }
 384 }
 385
 386
 387 /**
 388  * Extended swizzle of a SoA vector.
 389  *
 390  * @param bld         building context
 391  * @param unswizzled  array with the 4 unswizzled values
 392  * @param swizzles    array of PIPE_SWIZZLE_*
 393  * @param swizzled    output swizzled values
 394  */
 395 void
 396 lp_build_swizzle_soa(struct lp_build_context *bld,
 397                      const LLVMValueRef *unswizzled,
 398                      const unsigned char swizzles[4],
 399                      LLVMValueRef *swizzled)
 400 {
 401    unsigned chan;
 402
 403    for (chan = 0; chan < 4; ++chan) {
 404       swizzled[chan] = lp_build_swizzle_soa_channel(bld, unswizzled,
 405                                                     swizzles[chan]);
 406    }
 407 }
 408
 409
 410 /**
 411  * Do an extended swizzle of a SoA vector inplace.
 412  *
 413  * @param bld         building context
 414  * @param values      intput/output array with the 4 values
 415  * @param swizzles    array of PIPE_SWIZZLE_*
 416  */
 417 void
 418 lp_build_swizzle_soa_inplace(struct lp_build_context *bld,
 419                              LLVMValueRef *values,
 420                              const unsigned char swizzles[4])
 421 {
 422    LLVMValueRef unswizzled[4];
 423    unsigned chan;
 424
 425    for (chan = 0; chan < 4; ++chan) {
 426       unswizzled[chan] = values[chan];
 427    }
 428
 429    lp_build_swizzle_soa(bld, unswizzled, swizzles, values);
 430 }