src/gallium/auxiliary/gallivm/lp_bld_format_aos.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  * AoS pixel format manipulation.
  31  *
  32  * @author Jose Fonseca <jfonseca@vmware.com>
  33  */
  34
  35
  36 #include "util/u_format.h"
  37 #include "util/u_memory.h"
  38 #include "util/u_math.h"
  39 #include "util/u_pointer.h"
  40 #include "util/u_string.h"
  41
  42 #include "lp_bld_arit.h"
  43 #include "lp_bld_init.h"
  44 #include "lp_bld_type.h"
  45 #include "lp_bld_flow.h"
  46 #include "lp_bld_const.h"
  47 #include "lp_bld_conv.h"
  48 #include "lp_bld_swizzle.h"
  49 #include "lp_bld_gather.h"
  50 #include "lp_bld_debug.h"
  51 #include "lp_bld_format.h"
  52 #include "lp_bld_intr.h"
  53
  54
  55 /**
  56  * Basic swizzling.  Rearrange the order of the unswizzled array elements
  57  * according to the format description.  PIPE_SWIZZLE_ZERO/ONE are supported
  58  * too.
  59  * Ex: if unswizzled[4] = {B, G, R, x}, then swizzled_out[4] = {R, G, B, 1}.
  60  */
  61 LLVMValueRef
  62 lp_build_format_swizzle_aos(const struct util_format_description *desc,
  63                             struct lp_build_context *bld,
  64                             LLVMValueRef unswizzled)
  65 {
  66    unsigned char swizzles[4];
  67    unsigned chan;
  68
  69    assert(bld->type.length % 4 == 0);
  70
  71    for (chan = 0; chan < 4; ++chan) {
  72       enum util_format_swizzle swizzle;
  73
  74       if (desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) {
  75          /*
  76           * For ZS formats do RGBA = ZZZ1
  77           */
  78          if (chan == 3) {
  79             swizzle = UTIL_FORMAT_SWIZZLE_1;
  80          } else if (desc->swizzle[0] == UTIL_FORMAT_SWIZZLE_NONE) {
  81             swizzle = UTIL_FORMAT_SWIZZLE_0;
  82          } else {
  83             swizzle = desc->swizzle[0];
  84          }
  85       } else {
  86          swizzle = desc->swizzle[chan];
  87       }
  88       swizzles[chan] = swizzle;
  89    }
  90
  91    return lp_build_swizzle_aos(bld, unswizzled, swizzles);
  92 }
  93
  94
  95 /**
  96  * Whether the format matches the vector type, apart of swizzles.
  97  */
  98 static INLINE boolean
  99 format_matches_type(const struct util_format_description *desc,
 100                     struct lp_type type)
 101 {
 102    enum util_format_type chan_type;
 103    unsigned chan;
 104
 105    assert(type.length % 4 == 0);
 106
 107    if (desc->layout != UTIL_FORMAT_LAYOUT_PLAIN ||
 108        desc->colorspace != UTIL_FORMAT_COLORSPACE_RGB ||
 109        desc->block.width != 1 ||
 110        desc->block.height != 1) {
 111       return FALSE;
 112    }
 113
 114    if (type.floating) {
 115       chan_type = UTIL_FORMAT_TYPE_FLOAT;
 116    } else if (type.fixed) {
 117       chan_type = UTIL_FORMAT_TYPE_FIXED;
 118    } else if (type.sign) {
 119       chan_type = UTIL_FORMAT_TYPE_SIGNED;
 120    } else {
 121       chan_type = UTIL_FORMAT_TYPE_UNSIGNED;
 122    }
 123
 124    for (chan = 0; chan < desc->nr_channels; ++chan) {
 125       if (desc->channel[chan].size != type.width) {
 126          return FALSE;
 127       }
 128
 129       if (desc->channel[chan].type != UTIL_FORMAT_TYPE_VOID) {
 130          if (desc->channel[chan].type != chan_type ||
 131              desc->channel[chan].normalized != type.norm) {
 132             return FALSE;
 133          }
 134       }
 135    }
 136
 137    return TRUE;
 138 }
 139
 140
 141 /**
 142  * Unpack a single pixel into its XYZW components.
 143  *
 144  * @param desc  the pixel format for the packed pixel value
 145  * @param packed integer pixel in a format such as PIPE_FORMAT_B8G8R8A8_UNORM
 146  *
 147  * @return XYZW in a float[4] or ubyte[4] or ushort[4] vector.
 148  */
 149 static INLINE LLVMValueRef
 150 lp_build_unpack_arith_rgba_aos(struct gallivm_state *gallivm,
 151                                const struct util_format_description *desc,
 152                                LLVMValueRef packed)
 153 {
 154    LLVMBuilderRef builder = gallivm->builder;
 155    LLVMValueRef shifted, casted, scaled, masked;
 156    LLVMValueRef shifts[4];
 157    LLVMValueRef masks[4];
 158    LLVMValueRef scales[4];
 159
 160    boolean normalized;
 161    boolean needs_uitofp;
 162    unsigned i;
 163
 164    /* TODO: Support more formats */
 165    assert(desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
 166    assert(desc->block.width == 1);
 167    assert(desc->block.height == 1);
 168    assert(desc->block.bits <= 32);
 169
 170    /* Do the intermediate integer computations with 32bit integers since it
 171     * matches floating point size */
 172    assert (LLVMTypeOf(packed) == LLVMInt32TypeInContext(gallivm->context));
 173
 174    /* Broadcast the packed value to all four channels
 175     * before: packed = BGRA
 176     * after: packed = {BGRA, BGRA, BGRA, BGRA}
 177     */
 178    packed = LLVMBuildInsertElement(builder,
 179                                    LLVMGetUndef(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4)),
 180                                    packed,
 181                                    LLVMConstNull(LLVMInt32TypeInContext(gallivm->context)),
 182                                    "");
 183    packed = LLVMBuildShuffleVector(builder,
 184                                    packed,
 185                                    LLVMGetUndef(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4)),
 186                                    LLVMConstNull(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4)),
 187                                    "");
 188
 189    /* Initialize vector constants */
 190    normalized = FALSE;
 191    needs_uitofp = FALSE;
 192
 193    /* Loop over 4 color components */
 194    for (i = 0; i < 4; ++i) {
 195       unsigned bits = desc->channel[i].size;
 196       unsigned shift = desc->channel[i].shift;
 197
 198       if (desc->channel[i].type == UTIL_FORMAT_TYPE_VOID) {
 199          shifts[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
 200          masks[i] = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
 201          scales[i] =  LLVMConstNull(LLVMFloatTypeInContext(gallivm->context));
 202       }
 203       else {
 204          unsigned long long mask = (1ULL << bits) - 1;
 205
 206          assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED);
 207
 208          if (bits == 32) {
 209             needs_uitofp = TRUE;
 210          }
 211
 212          shifts[i] = lp_build_const_int32(gallivm, shift);
 213          masks[i] = lp_build_const_int32(gallivm, mask);
 214
 215          if (desc->channel[i].normalized) {
 216             scales[i] = lp_build_const_float(gallivm, 1.0 / mask);
 217             normalized = TRUE;
 218          }
 219          else
 220             scales[i] =  lp_build_const_float(gallivm, 1.0);
 221       }
 222    }
 223
 224    /* Ex: convert packed = {XYZW, XYZW, XYZW, XYZW}
 225     * into masked = {X, Y, Z, W}
 226     */
 227    shifted = LLVMBuildLShr(builder, packed, LLVMConstVector(shifts, 4), "");
 228    masked = LLVMBuildAnd(builder, shifted, LLVMConstVector(masks, 4), "");
 229
 230    if (!needs_uitofp) {
 231       /* UIToFP can't be expressed in SSE2 */
 232       casted = LLVMBuildSIToFP(builder, masked, LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), "");
 233    } else {
 234       casted = LLVMBuildUIToFP(builder, masked, LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), "");
 235    }
 236
 237    /* At this point 'casted' may be a vector of floats such as
 238     * {255.0, 255.0, 255.0, 255.0}.  Next, if the pixel values are normalized
 239     * we'll scale this to {1.0, 1.0, 1.0, 1.0}.
 240     */
 241
 242    if (normalized)
 243       scaled = LLVMBuildFMul(builder, casted, LLVMConstVector(scales, 4), "");
 244    else
 245       scaled = casted;
 246
 247    return scaled;
 248 }
 249
 250
 251 /**
 252  * Pack a single pixel.
 253  *
 254  * @param rgba 4 float vector with the unpacked components.
 255  *
 256  * XXX: This is mostly for reference and testing -- operating a single pixel at
 257  * a time is rarely if ever needed.
 258  */
 259 LLVMValueRef
 260 lp_build_pack_rgba_aos(struct gallivm_state *gallivm,
 261                        const struct util_format_description *desc,
 262                        LLVMValueRef rgba)
 263 {
 264    LLVMBuilderRef builder = gallivm->builder;
 265    LLVMTypeRef type;
 266    LLVMValueRef packed = NULL;
 267    LLVMValueRef swizzles[4];
 268    LLVMValueRef shifted, casted, scaled, unswizzled;
 269    LLVMValueRef shifts[4];
 270    LLVMValueRef scales[4];
 271    boolean normalized;
 272    unsigned i, j;
 273
 274    assert(desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
 275    assert(desc->block.width == 1);
 276    assert(desc->block.height == 1);
 277
 278    type = LLVMIntTypeInContext(gallivm->context, desc->block.bits);
 279
 280    /* Unswizzle the color components into the source vector. */
 281    for (i = 0; i < 4; ++i) {
 282       for (j = 0; j < 4; ++j) {
 283          if (desc->swizzle[j] == i)
 284             break;
 285       }
 286       if (j < 4)
 287          swizzles[i] = lp_build_const_int32(gallivm, j);
 288       else
 289          swizzles[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
 290    }
 291
 292    unswizzled = LLVMBuildShuffleVector(builder, rgba,
 293                                        LLVMGetUndef(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4)),
 294                                        LLVMConstVector(swizzles, 4), "");
 295
 296    normalized = FALSE;
 297    for (i = 0; i < 4; ++i) {
 298       unsigned bits = desc->channel[i].size;
 299       unsigned shift = desc->channel[i].shift;
 300
 301       if (desc->channel[i].type == UTIL_FORMAT_TYPE_VOID) {
 302          shifts[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
 303          scales[i] =  LLVMGetUndef(LLVMFloatTypeInContext(gallivm->context));
 304       }
 305       else {
 306          unsigned mask = (1 << bits) - 1;
 307
 308          assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED);
 309          assert(bits < 32);
 310
 311          shifts[i] = lp_build_const_int32(gallivm, shift);
 312
 313          if (desc->channel[i].normalized) {
 314             scales[i] = lp_build_const_float(gallivm, mask);
 315             normalized = TRUE;
 316          }
 317          else
 318             scales[i] = lp_build_const_float(gallivm, 1.0);
 319       }
 320    }
 321
 322    if (normalized)
 323       scaled = LLVMBuildFMul(builder, unswizzled, LLVMConstVector(scales, 4), "");
 324    else
 325       scaled = unswizzled;
 326
 327    casted = LLVMBuildFPToSI(builder, scaled, LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4), "");
 328
 329    shifted = LLVMBuildShl(builder, casted, LLVMConstVector(shifts, 4), "");
 330
 331    /* Bitwise or all components */
 332    for (i = 0; i < 4; ++i) {
 333       if (desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED) {
 334          LLVMValueRef component = LLVMBuildExtractElement(builder, shifted,
 335                                                lp_build_const_int32(gallivm, i), "");
 336          if (packed)
 337             packed = LLVMBuildOr(builder, packed, component, "");
 338          else
 339             packed = component;
 340       }
 341    }
 342
 343    if (!packed)
 344       packed = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
 345
 346    if (desc->block.bits < 32)
 347       packed = LLVMBuildTrunc(builder, packed, type, "");
 348
 349    return packed;
 350 }
 351
 352
 353
 354
 355 /**
 356  * Fetch a pixel into a 4 float AoS.
 357  *
 358  * \param format_desc  describes format of the image we're fetching from
 359  * \param ptr  address of the pixel block (or the texel if uncompressed)
 360  * \param i, j  the sub-block pixel coordinates.  For non-compressed formats
 361  *              these will always be (0, 0).
 362  * \return  a 4 element vector with the pixel's RGBA values.
 363  */
 364 LLVMValueRef
 365 lp_build_fetch_rgba_aos(struct gallivm_state *gallivm,
 366                         const struct util_format_description *format_desc,
 367                         struct lp_type type,
 368                         LLVMValueRef base_ptr,
 369                         LLVMValueRef offset,
 370                         LLVMValueRef i,
 371                         LLVMValueRef j)
 372 {
 373    LLVMBuilderRef builder = gallivm->builder;
 374    unsigned num_pixels = type.length / 4;
 375    struct lp_build_context bld;
 376
 377    assert(type.length <= LP_MAX_VECTOR_LENGTH);
 378    assert(type.length % 4 == 0);
 379
 380    lp_build_context_init(&bld, gallivm, type);
 381
 382    /*
 383     * Trivial case
 384     *
 385     * The format matches the type (apart of a swizzle) so no need for
 386     * scaling or converting.
 387     */
 388
 389    if (format_matches_type(format_desc, type) &&
 390        format_desc->block.bits <= type.width * 4 &&
 391        util_is_power_of_two(format_desc->block.bits)) {
 392       LLVMValueRef packed;
 393       LLVMTypeRef dst_vec_type = lp_build_vec_type(gallivm, type);
 394       unsigned vec_len = type.width * type.length;
 395
 396       /*
 397        * The format matches the type (apart of a swizzle) so no need for
 398        * scaling or converting.
 399        */
 400
 401       packed = lp_build_gather(gallivm, type.length/4,
 402                                format_desc->block.bits, type.width*4,
 403                                base_ptr, offset, TRUE);
 404
 405       assert(format_desc->block.bits <= vec_len);
 406
 407       packed = LLVMBuildBitCast(gallivm->builder, packed, dst_vec_type, "");
 408       return lp_build_format_swizzle_aos(format_desc, &bld, packed);
 409    }
 410
 411    /*
 412     * Bit arithmetic
 413     */
 414
 415    if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN &&
 416        (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
 417         format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) &&
 418        format_desc->block.width == 1 &&
 419        format_desc->block.height == 1 &&
 420        util_is_power_of_two(format_desc->block.bits) &&
 421        format_desc->block.bits <= 32 &&
 422        format_desc->is_bitmask &&
 423        !format_desc->is_mixed &&
 424        (format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED ||
 425         format_desc->channel[1].type == UTIL_FORMAT_TYPE_UNSIGNED) &&
 426        !format_desc->channel[0].pure_integer) {
 427
 428       LLVMValueRef tmps[LP_MAX_VECTOR_LENGTH/4];
 429       LLVMValueRef res;
 430       unsigned k;
 431
 432       /*
 433        * Unpack a pixel at a time into a <4 x float> RGBA vector
 434        */
 435
 436       for (k = 0; k < num_pixels; ++k) {
 437          LLVMValueRef packed;
 438
 439          packed = lp_build_gather_elem(gallivm, num_pixels,
 440                                        format_desc->block.bits, 32,
 441                                        base_ptr, offset, k, FALSE);
 442
 443          tmps[k] = lp_build_unpack_arith_rgba_aos(gallivm,
 444                                                   format_desc,
 445                                                   packed);
 446       }
 447
 448       /*
 449        * Type conversion.
 450        *
 451        * TODO: We could avoid floating conversion for integer to
 452        * integer conversions.
 453        */
 454
 455       if (gallivm_debug & GALLIVM_DEBUG_PERF && !type.floating) {
 456          debug_printf("%s: unpacking %s with floating point\n",
 457                       __FUNCTION__, format_desc->short_name);
 458       }
 459
 460       lp_build_conv(gallivm,
 461                     lp_float32_vec4_type(),
 462                     type,
 463                     tmps, num_pixels, &res, 1);
 464
 465       return lp_build_format_swizzle_aos(format_desc, &bld, res);
 466    }
 467
 468    /* If all channels are of same type and we are not using half-floats */
 469    if (format_desc->is_array &&
 470        format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB) {
 471       assert(!format_desc->is_mixed);
 472       return lp_build_fetch_rgba_aos_array(gallivm, format_desc, type, base_ptr, offset);
 473    }
 474
 475    /*
 476     * YUV / subsampled formats
 477     */
 478
 479    if (format_desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
 480       struct lp_type tmp_type;
 481       LLVMValueRef tmp;
 482
 483       memset(&tmp_type, 0, sizeof tmp_type);
 484       tmp_type.width = 8;
 485       tmp_type.length = num_pixels * 4;
 486       tmp_type.norm = TRUE;
 487
 488       tmp = lp_build_fetch_subsampled_rgba_aos(gallivm,
 489                                                format_desc,
 490                                                num_pixels,
 491                                                base_ptr,
 492                                                offset,
 493                                                i, j);
 494
 495       lp_build_conv(gallivm,
 496                     tmp_type, type,
 497                     &tmp, 1, &tmp, 1);
 498
 499       return tmp;
 500    }
 501
 502    /*
 503     * Fallback to util_format_description::fetch_rgba_8unorm().
 504     */
 505
 506    if (format_desc->fetch_rgba_8unorm &&
 507        !type.floating && type.width == 8 && !type.sign && type.norm) {
 508       /*
 509        * Fallback to calling util_format_description::fetch_rgba_8unorm.
 510        *
 511        * This is definitely not the most efficient way of fetching pixels, as
 512        * we miss the opportunity to do vectorization, but this it is a
 513        * convenient for formats or scenarios for which there was no opportunity
 514        * or incentive to optimize.
 515        */
 516
 517       LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context);
 518       LLVMTypeRef pi8t = LLVMPointerType(i8t, 0);
 519       LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
 520       LLVMValueRef function;
 521       LLVMValueRef tmp_ptr;
 522       LLVMValueRef tmp;
 523       LLVMValueRef res;
 524       unsigned k;
 525
 526       if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 527          debug_printf("%s: falling back to util_format_%s_fetch_rgba_8unorm\n",
 528                       __FUNCTION__, format_desc->short_name);
 529       }
 530
 531       /*
 532        * Declare and bind format_desc->fetch_rgba_8unorm().
 533        */
 534
 535       {
 536          /*
 537           * Function to call looks like:
 538           *   fetch(uint8_t *dst, const uint8_t *src, unsigned i, unsigned j)
 539           */
 540          LLVMTypeRef ret_type;
 541          LLVMTypeRef arg_types[4];
 542          LLVMTypeRef function_type;
 543
 544          ret_type = LLVMVoidTypeInContext(gallivm->context);
 545          arg_types[0] = pi8t;
 546          arg_types[1] = pi8t;
 547          arg_types[2] = i32t;
 548          arg_types[3] = i32t;
 549          function_type = LLVMFunctionType(ret_type, arg_types,
 550                                           Elements(arg_types), 0);
 551
 552          /* make const pointer for the C fetch_rgba_8unorm function */
 553          function = lp_build_const_int_pointer(gallivm,
 554             func_to_pointer((func_pointer) format_desc->fetch_rgba_8unorm));
 555
 556          /* cast the callee pointer to the function's type */
 557          function = LLVMBuildBitCast(builder, function,
 558                                      LLVMPointerType(function_type, 0),
 559                                      "cast callee");
 560       }
 561
 562       tmp_ptr = lp_build_alloca(gallivm, i32t, "");
 563
 564       res = LLVMGetUndef(LLVMVectorType(i32t, num_pixels));
 565
 566       /*
 567        * Invoke format_desc->fetch_rgba_8unorm() for each pixel and insert the result
 568        * in the SoA vectors.
 569        */
 570
 571       for (k = 0; k < num_pixels; ++k) {
 572          LLVMValueRef index = lp_build_const_int32(gallivm, k);
 573          LLVMValueRef args[4];
 574
 575          args[0] = LLVMBuildBitCast(builder, tmp_ptr, pi8t, "");
 576          args[1] = lp_build_gather_elem_ptr(gallivm, num_pixels,
 577                                             base_ptr, offset, k);
 578
 579          if (num_pixels == 1) {
 580             args[2] = i;
 581             args[3] = j;
 582          }
 583          else {
 584             args[2] = LLVMBuildExtractElement(builder, i, index, "");
 585             args[3] = LLVMBuildExtractElement(builder, j, index, "");
 586          }
 587
 588          LLVMBuildCall(builder, function, args, Elements(args), "");
 589
 590          tmp = LLVMBuildLoad(builder, tmp_ptr, "");
 591
 592          if (num_pixels == 1) {
 593             res = tmp;
 594          }
 595          else {
 596             res = LLVMBuildInsertElement(builder, res, tmp, index, "");
 597          }
 598       }
 599
 600       /* Bitcast from <n x i32> to <4n x i8> */
 601       res = LLVMBuildBitCast(builder, res, bld.vec_type, "");
 602
 603       return res;
 604    }
 605
 606    /*
 607     * Fallback to util_format_description::fetch_rgba_float().
 608     */
 609
 610    if (format_desc->fetch_rgba_float) {
 611       /*
 612        * Fallback to calling util_format_description::fetch_rgba_float.
 613        *
 614        * This is definitely not the most efficient way of fetching pixels, as
 615        * we miss the opportunity to do vectorization, but this it is a
 616        * convenient for formats or scenarios for which there was no opportunity
 617        * or incentive to optimize.
 618        */
 619
 620       LLVMTypeRef f32t = LLVMFloatTypeInContext(gallivm->context);
 621       LLVMTypeRef f32x4t = LLVMVectorType(f32t, 4);
 622       LLVMTypeRef pf32t = LLVMPointerType(f32t, 0);
 623       LLVMTypeRef pi8t = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
 624       LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
 625       LLVMValueRef function;
 626       LLVMValueRef tmp_ptr;
 627       LLVMValueRef tmps[LP_MAX_VECTOR_LENGTH/4];
 628       LLVMValueRef res;
 629       unsigned k;
 630
 631       if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 632          debug_printf("%s: falling back to util_format_%s_fetch_rgba_float\n",
 633                       __FUNCTION__, format_desc->short_name);
 634       }
 635
 636       /*
 637        * Declare and bind format_desc->fetch_rgba_float().
 638        */
 639
 640       {
 641          /*
 642           * Function to call looks like:
 643           *   fetch(float *dst, const uint8_t *src, unsigned i, unsigned j)
 644           */
 645          LLVMTypeRef ret_type;
 646          LLVMTypeRef arg_types[4];
 647
 648          ret_type = LLVMVoidTypeInContext(gallivm->context);
 649          arg_types[0] = pf32t;
 650          arg_types[1] = pi8t;
 651          arg_types[2] = i32t;
 652          arg_types[3] = i32t;
 653
 654          function = lp_build_const_func_pointer(gallivm,
 655                                                 func_to_pointer((func_pointer) format_desc->fetch_rgba_float),
 656                                                 ret_type,
 657                                                 arg_types, Elements(arg_types),
 658                                                 format_desc->short_name);
 659       }
 660
 661       tmp_ptr = lp_build_alloca(gallivm, f32x4t, "");
 662
 663       /*
 664        * Invoke format_desc->fetch_rgba_float() for each pixel and insert the result
 665        * in the SoA vectors.
 666        */
 667
 668       for (k = 0; k < num_pixels; ++k) {
 669          LLVMValueRef args[4];
 670
 671          args[0] = LLVMBuildBitCast(builder, tmp_ptr, pf32t, "");
 672          args[1] = lp_build_gather_elem_ptr(gallivm, num_pixels,
 673                                             base_ptr, offset, k);
 674
 675          if (num_pixels == 1) {
 676             args[2] = i;
 677             args[3] = j;
 678          }
 679          else {
 680             LLVMValueRef index = lp_build_const_int32(gallivm, k);
 681             args[2] = LLVMBuildExtractElement(builder, i, index, "");
 682             args[3] = LLVMBuildExtractElement(builder, j, index, "");
 683          }
 684
 685          LLVMBuildCall(builder, function, args, Elements(args), "");
 686
 687          tmps[k] = LLVMBuildLoad(builder, tmp_ptr, "");
 688       }
 689
 690       lp_build_conv(gallivm,
 691                     lp_float32_vec4_type(),
 692                     type,
 693                     tmps, num_pixels, &res, 1);
 694
 695       return res;
 696    }
 697
 698    assert(!util_format_is_pure_integer(format_desc->format));
 699
 700    assert(0);
 701    return lp_build_undef(gallivm, type);
 702 }