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
  80 #if HAVE_LLVM >= 0x207
  81       /* The shuffle vector is always made of int32 elements */
  82       struct lp_type i32_vec_type = lp_type_int_vec(32);
  83       i32_vec_type.length = type.length;
  84
  85       res = LLVMBuildInsertElement(bld->builder, bld->undef, scalar,
  86                                    LLVMConstInt(LLVMInt32Type(), 0, 0), "");
  87       res = LLVMBuildShuffleVector(bld->builder, res, bld->undef,
  88                                    lp_build_const_int_vec(i32_vec_type, 0), "");
  89 #else
  90       /* XXX: The above path provokes a bug in LLVM 2.6 */
  91       unsigned i;
  92       res = bld->undef;
  93       for(i = 0; i < type.length; ++i) {
  94          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
  95          res = LLVMBuildInsertElement(bld->builder, res, scalar, index, "");
  96       }
  97 #endif
  98       return res;
  99    }
 100 }
 101
 102
 103 /**
 104  * Swizzle one channel into all other three channels.
 105  */
 106 LLVMValueRef
 107 lp_build_swizzle_scalar_aos(struct lp_build_context *bld,
 108                             LLVMValueRef a,
 109                             unsigned channel)
 110 {
 111    const struct lp_type type = bld->type;
 112    const unsigned n = type.length;
 113    unsigned i, j;
 114
 115    if(a == bld->undef || a == bld->zero || a == bld->one)
 116       return a;
 117
 118    /* XXX: SSE3 has PSHUFB which should be better than bitmasks, but forcing
 119     * using shuffles here actually causes worst results. More investigation is
 120     * needed. */
 121    if (type.width >= 16) {
 122       /*
 123        * Shuffle.
 124        */
 125       LLVMTypeRef elem_type = LLVMInt32Type();
 126       LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
 127
 128       for(j = 0; j < n; j += 4)
 129          for(i = 0; i < 4; ++i)
 130             shuffles[j + i] = LLVMConstInt(elem_type, j + channel, 0);
 131
 132       return LLVMBuildShuffleVector(bld->builder, a, bld->undef, LLVMConstVector(shuffles, n), "");
 133    }
 134    else {
 135       /*
 136        * Bit mask and recursive shifts
 137        *
 138        *   XYZW XYZW .... XYZW  <= input
 139        *   0Y00 0Y00 .... 0Y00
 140        *   YY00 YY00 .... YY00
 141        *   YYYY YYYY .... YYYY  <= output
 142        */
 143       struct lp_type type4;
 144       const char shifts[4][2] = {
 145          { 1,  2},
 146          {-1,  2},
 147          { 1, -2},
 148          {-1, -2}
 149       };
 150       unsigned i;
 151
 152       a = LLVMBuildAnd(bld->builder, a,
 153                        lp_build_const_mask_aos(type, 1 << channel), "");
 154
 155       /*
 156        * Build a type where each element is an integer that cover the four
 157        * channels.
 158        */
 159
 160       type4 = type;
 161       type4.floating = FALSE;
 162       type4.width *= 4;
 163       type4.length /= 4;
 164
 165       a = LLVMBuildBitCast(bld->builder, a, lp_build_vec_type(type4), "");
 166
 167       for(i = 0; i < 2; ++i) {
 168          LLVMValueRef tmp = NULL;
 169          int shift = shifts[channel][i];
 170
 171 #ifdef PIPE_ARCH_LITTLE_ENDIAN
 172          shift = -shift;
 173 #endif
 174
 175          if(shift > 0)
 176             tmp = LLVMBuildLShr(bld->builder, a, lp_build_const_int_vec(type4, shift*type.width), "");
 177          if(shift < 0)
 178             tmp = LLVMBuildShl(bld->builder, a, lp_build_const_int_vec(type4, -shift*type.width), "");
 179
 180          assert(tmp);
 181          if(tmp)
 182             a = LLVMBuildOr(bld->builder, a, tmp, "");
 183       }
 184
 185       return LLVMBuildBitCast(bld->builder, a, lp_build_vec_type(type), "");
 186    }
 187 }
 188
 189
 190 LLVMValueRef
 191 lp_build_swizzle_aos(struct lp_build_context *bld,
 192                      LLVMValueRef a,
 193                      const unsigned char swizzles[4])
 194 {
 195    const struct lp_type type = bld->type;
 196    const unsigned n = type.length;
 197    unsigned i, j;
 198
 199    if (swizzles[0] == PIPE_SWIZZLE_RED &&
 200        swizzles[1] == PIPE_SWIZZLE_GREEN &&
 201        swizzles[2] == PIPE_SWIZZLE_BLUE &&
 202        swizzles[3] == PIPE_SWIZZLE_ALPHA) {
 203       return a;
 204    }
 205
 206    if (swizzles[0] == swizzles[1] &&
 207        swizzles[1] == swizzles[2] &&
 208        swizzles[2] == swizzles[3]) {
 209       switch (swizzles[0]) {
 210       case PIPE_SWIZZLE_RED:
 211       case PIPE_SWIZZLE_GREEN:
 212       case PIPE_SWIZZLE_BLUE:
 213       case PIPE_SWIZZLE_ALPHA:
 214          return lp_build_swizzle_scalar_aos(bld, a, swizzles[0]);
 215       case PIPE_SWIZZLE_ZERO:
 216          return bld->zero;
 217       case PIPE_SWIZZLE_ONE:
 218          return bld->one;
 219       default:
 220          assert(0);
 221          return bld->undef;
 222       }
 223    }
 224
 225    if (type.width >= 16) {
 226       /*
 227        * Shuffle.
 228        */
 229       LLVMValueRef undef = LLVMGetUndef(lp_build_elem_type(type));
 230       LLVMTypeRef i32t = LLVMInt32Type();
 231       LLVMValueRef shuffles[LP_MAX_VECTOR_LENGTH];
 232       LLVMValueRef aux[LP_MAX_VECTOR_LENGTH];
 233
 234       memset(aux, 0, sizeof aux);
 235
 236       for(j = 0; j < n; j += 4) {
 237          for(i = 0; i < 4; ++i) {
 238             unsigned shuffle;
 239             switch (swizzles[i]) {
 240             default:
 241                assert(0);
 242                /* fall through */
 243             case PIPE_SWIZZLE_RED:
 244             case PIPE_SWIZZLE_GREEN:
 245             case PIPE_SWIZZLE_BLUE:
 246             case PIPE_SWIZZLE_ALPHA:
 247                shuffle = j + swizzles[i];
 248                break;
 249             case PIPE_SWIZZLE_ZERO:
 250                shuffle = type.length + 0;
 251                if (!aux[0]) {
 252                   aux[0] = lp_build_const_elem(type, 0.0);
 253                }
 254                break;
 255             case PIPE_SWIZZLE_ONE:
 256                shuffle = type.length + 1;
 257                if (!aux[1]) {
 258                   aux[1] = lp_build_const_elem(type, 1.0);
 259                }
 260                break;
 261             }
 262             shuffles[j + i] = LLVMConstInt(i32t, shuffle, 0);
 263          }
 264       }
 265
 266       for (i = 0; i < n; ++i) {
 267          if (!aux[i]) {
 268             aux[i] = undef;
 269          }
 270       }
 271
 272       return LLVMBuildShuffleVector(bld->builder, a,
 273                                     LLVMConstVector(aux, n),
 274                                     LLVMConstVector(shuffles, n), "");
 275    } else {
 276       /*
 277        * Bit mask and shifts.
 278        *
 279        * For example, this will convert BGRA to RGBA by doing
 280        *
 281        *   rgba = (bgra & 0x00ff0000) >> 16
 282        *        | (bgra & 0xff00ff00)
 283        *        | (bgra & 0x000000ff) << 16
 284        *
 285        * This is necessary not only for faster cause, but because X86 backend
 286        * will refuse shuffles of <4 x i8> vectors
 287        */
 288       LLVMValueRef res;
 289       struct lp_type type4;
 290       unsigned cond = 0;
 291       unsigned chan;
 292       int shift;
 293
 294       /*
 295        * Start with a mixture of 1 and 0.
 296        */
 297       for (chan = 0; chan < 4; ++chan) {
 298          if (swizzles[chan] == PIPE_SWIZZLE_ONE) {
 299             cond |= 1 << chan;
 300          }
 301       }
 302       res = lp_build_select_aos(bld, cond, bld->one, bld->zero);
 303
 304       /*
 305        * Build a type where each element is an integer that cover the four
 306        * channels.
 307        */
 308       type4 = type;
 309       type4.floating = FALSE;
 310       type4.width *= 4;
 311       type4.length /= 4;
 312
 313       a = LLVMBuildBitCast(bld->builder, a, lp_build_vec_type(type4), "");
 314       res = LLVMBuildBitCast(bld->builder, res, lp_build_vec_type(type4), "");
 315
 316       /*
 317        * Mask and shift the channels, trying to group as many channels in the
 318        * same shift as possible
 319        */
 320       for (shift = -3; shift <= 3; ++shift) {
 321          unsigned long long mask = 0;
 322
 323          assert(type4.width <= sizeof(mask)*8);
 324
 325          for (chan = 0; chan < 4; ++chan) {
 326             /* FIXME: big endian */
 327             if (swizzles[chan] < 4 &&
 328                 chan - swizzles[chan] == shift) {
 329                mask |= ((1ULL << type.width) - 1) << (swizzles[chan] * type.width);
 330             }
 331          }
 332
 333          if (mask) {
 334             LLVMValueRef masked;
 335             LLVMValueRef shifted;
 336
 337             if (0)
 338                debug_printf("shift = %i, mask = 0x%08llx\n", shift, mask);
 339
 340             masked = LLVMBuildAnd(bld->builder, a,
 341                                   lp_build_const_int_vec(type4, mask), "");
 342             if (shift > 0) {
 343                shifted = LLVMBuildShl(bld->builder, masked,
 344                                       lp_build_const_int_vec(type4, shift*type.width), "");
 345             } else if (shift < 0) {
 346                shifted = LLVMBuildLShr(bld->builder, masked,
 347                                        lp_build_const_int_vec(type4, -shift*type.width), "");
 348             } else {
 349                shifted = masked;
 350             }
 351
 352             res = LLVMBuildOr(bld->builder, res, shifted, "");
 353          }
 354       }
 355
 356       return LLVMBuildBitCast(bld->builder, res, lp_build_vec_type(type), "");
 357    }
 358 }
 359
 360
 361 /**
 362  * Extended swizzle of a single channel of a SoA vector.
 363  *
 364  * @param bld         building context
 365  * @param unswizzled  array with the 4 unswizzled values
 366  * @param swizzle     one of the PIPE_SWIZZLE_*
 367  *
 368  * @return  the swizzled value.
 369  */
 370 LLVMValueRef
 371 lp_build_swizzle_soa_channel(struct lp_build_context *bld,
 372                              const LLVMValueRef *unswizzled,
 373                              unsigned swizzle)
 374 {
 375    switch (swizzle) {
 376    case PIPE_SWIZZLE_RED:
 377    case PIPE_SWIZZLE_GREEN:
 378    case PIPE_SWIZZLE_BLUE:
 379    case PIPE_SWIZZLE_ALPHA:
 380       return unswizzled[swizzle];
 381    case PIPE_SWIZZLE_ZERO:
 382       return bld->zero;
 383    case PIPE_SWIZZLE_ONE:
 384       return bld->one;
 385    default:
 386       assert(0);
 387       return bld->undef;
 388    }
 389 }
 390
 391
 392 /**
 393  * Extended swizzle of a SoA vector.
 394  *
 395  * @param bld         building context
 396  * @param unswizzled  array with the 4 unswizzled values
 397  * @param swizzles    array of PIPE_SWIZZLE_*
 398  * @param swizzled    output swizzled values
 399  */
 400 void
 401 lp_build_swizzle_soa(struct lp_build_context *bld,
 402                      const LLVMValueRef *unswizzled,
 403                      const unsigned char swizzles[4],
 404                      LLVMValueRef *swizzled)
 405 {
 406    unsigned chan;
 407
 408    for (chan = 0; chan < 4; ++chan) {
 409       swizzled[chan] = lp_build_swizzle_soa_channel(bld, unswizzled,
 410                                                     swizzles[chan]);
 411    }
 412 }
 413
 414
 415 /**
 416  * Do an extended swizzle of a SoA vector inplace.
 417  *
 418  * @param bld         building context
 419  * @param values      intput/output array with the 4 values
 420  * @param swizzles    array of PIPE_SWIZZLE_*
 421  */
 422 void
 423 lp_build_swizzle_soa_inplace(struct lp_build_context *bld,
 424                              LLVMValueRef *values,
 425                              const unsigned char swizzles[4])
 426 {
 427    LLVMValueRef unswizzled[4];
 428    unsigned chan;
 429
 430    for (chan = 0; chan < 4; ++chan) {
 431       unswizzled[chan] = values[chan];
 432    }
 433
 434    lp_build_swizzle_soa(bld, unswizzled, swizzles, values);
 435 }