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 LLVMValueRef
  99 lp_build_broadcast_aos(struct lp_build_context *bld,
 100                        LLVMValueRef a,
 101                        unsigned channel)
 102 {
 103    const struct lp_type type = bld->type;
 104    const unsigned n = type.length;
 105    unsigned i, j;
 106
 107    if(a == bld->undef || a == bld->zero || a == bld->one)
 108       return a;
 109
 110    /* XXX: SSE3 has PSHUFB which should be better than bitmasks, but forcing
 111     * using shuffles here actually causes worst results. More investigation is
 112     * needed. */
 113    if (type.width >= 16) {
 114       /*
 115        * Shuffle.
 116        */
 117       LLVMTypeRef elem_type = LLVMInt32Type();
 118       LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
 119
 120       for(j = 0; j < n; j += 4)
 121          for(i = 0; i < 4; ++i)
 122             shuffles[j + i] = LLVMConstInt(elem_type, j + channel, 0);
 123
 124       return LLVMBuildShuffleVector(bld->builder, a, bld->undef, LLVMConstVector(shuffles, n), "");
 125    }
 126    else {
 127       /*
 128        * Bit mask and recursive shifts
 129        *
 130        *   XYZW XYZW .... XYZW  <= input
 131        *   0Y00 0Y00 .... 0Y00
 132        *   YY00 YY00 .... YY00
 133        *   YYYY YYYY .... YYYY  <= output
 134        */
 135       struct lp_type type4;
 136       const char shifts[4][2] = {
 137          { 1,  2},
 138          {-1,  2},
 139          { 1, -2},
 140          {-1, -2}
 141       };
 142       boolean cond[4];
 143       unsigned i;
 144
 145       memset(cond, 0, sizeof cond);
 146       cond[channel] = 1;
 147
 148       a = LLVMBuildAnd(bld->builder, a, lp_build_const_mask_aos(type, cond), "");
 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_broadcast_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       boolean cond[4];
 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          cond[chan] = swizzles[chan] == PIPE_SWIZZLE_ONE ? TRUE : FALSE;
 294       }
 295       res = lp_build_select_aos(bld, bld->one, bld->zero, cond);
 296
 297       /*
 298        * Build a type where each element is an integer that cover the four
 299        * channels.
 300        */
 301       type4 = type;
 302       type4.floating = FALSE;
 303       type4.width *= 4;
 304       type4.length /= 4;
 305
 306       a = LLVMBuildBitCast(bld->builder, a, lp_build_vec_type(type4), "");
 307       res = LLVMBuildBitCast(bld->builder, res, lp_build_vec_type(type4), "");
 308
 309       /*
 310        * Mask and shift the channels, trying to group as many channels in the
 311        * same shift as possible
 312        */
 313       for (shift = -3; shift <= 3; ++shift) {
 314          unsigned long long mask = 0;
 315
 316          assert(type4.width <= sizeof(mask)*8);
 317
 318          for (chan = 0; chan < 4; ++chan) {
 319             /* FIXME: big endian */
 320             if (swizzles[chan] < 4 &&
 321                 chan - swizzles[chan] == shift) {
 322                mask |= ((1ULL << type.width) - 1) << (swizzles[chan] * type.width);
 323             }
 324          }
 325
 326          if (mask) {
 327             LLVMValueRef masked;
 328             LLVMValueRef shifted;
 329
 330             if (0)
 331                debug_printf("shift = %i, mask = 0x%08llx\n", shift, mask);
 332
 333             masked = LLVMBuildAnd(bld->builder, a,
 334                                   lp_build_const_int_vec(type4, mask), "");
 335             if (shift > 0) {
 336                shifted = LLVMBuildShl(bld->builder, masked,
 337                                       lp_build_const_int_vec(type4, shift*type.width), "");
 338             } else if (shift < 0) {
 339                shifted = LLVMBuildLShr(bld->builder, masked,
 340                                        lp_build_const_int_vec(type4, -shift*type.width), "");
 341             } else {
 342                shifted = masked;
 343             }
 344
 345             res = LLVMBuildOr(bld->builder, res, shifted, "");
 346          }
 347       }
 348
 349       return LLVMBuildBitCast(bld->builder, res, lp_build_vec_type(type), "");
 350    }
 351 }
 352
 353
 354 /**
 355  * Extended swizzle of a single channel of a SoA vector.
 356  *
 357  * @param bld         building context
 358  * @param unswizzled  array with the 4 unswizzled values
 359  * @param swizzle     one of the PIPE_SWIZZLE_*
 360  *
 361  * @return  the swizzled value.
 362  */
 363 LLVMValueRef
 364 lp_build_swizzle_soa_channel(struct lp_build_context *bld,
 365                              const LLVMValueRef *unswizzled,
 366                              unsigned swizzle)
 367 {
 368    switch (swizzle) {
 369    case PIPE_SWIZZLE_RED:
 370    case PIPE_SWIZZLE_GREEN:
 371    case PIPE_SWIZZLE_BLUE:
 372    case PIPE_SWIZZLE_ALPHA:
 373       return unswizzled[swizzle];
 374    case PIPE_SWIZZLE_ZERO:
 375       return bld->zero;
 376    case PIPE_SWIZZLE_ONE:
 377       return bld->one;
 378    default:
 379       assert(0);
 380       return bld->undef;
 381    }
 382 }
 383
 384
 385 /**
 386  * Extended swizzle of a SoA vector.
 387  *
 388  * @param bld         building context
 389  * @param unswizzled  array with the 4 unswizzled values
 390  * @param swizzles    array of PIPE_SWIZZLE_*
 391  * @param swizzled    output swizzled values
 392  */
 393 void
 394 lp_build_swizzle_soa(struct lp_build_context *bld,
 395                      const LLVMValueRef *unswizzled,
 396                      const unsigned char swizzles[4],
 397                      LLVMValueRef *swizzled)
 398 {
 399    unsigned chan;
 400
 401    for (chan = 0; chan < 4; ++chan) {
 402       swizzled[chan] = lp_build_swizzle_soa_channel(bld, unswizzled,
 403                                                     swizzles[chan]);
 404    }
 405 }
 406
 407
 408 /**
 409  * Do an extended swizzle of a SoA vector inplace.
 410  *
 411  * @param bld         building context
 412  * @param values      intput/output array with the 4 values
 413  * @param swizzles    array of PIPE_SWIZZLE_*
 414  */
 415 void
 416 lp_build_swizzle_soa_inplace(struct lp_build_context *bld,
 417                              LLVMValueRef *values,
 418                              const unsigned char swizzles[4])
 419 {
 420    LLVMValueRef unswizzled[4];
 421    unsigned chan;
 422
 423    for (chan = 0; chan < 4; ++chan) {
 424       unswizzled[chan] = values[chan];
 425    }
 426
 427    lp_build_swizzle_soa(bld, unswizzled, swizzles, values);
 428 }