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 RGBA 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 RGBA 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 shift;
 163    unsigned i;
 164
 165    /* TODO: Support more formats */
 166    assert(desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
 167    assert(desc->block.width == 1);
 168    assert(desc->block.height == 1);
 169    assert(desc->block.bits <= 32);
 170
 171    /* Do the intermediate integer computations with 32bit integers since it
 172     * matches floating point size */
 173    assert (LLVMTypeOf(packed) == LLVMInt32TypeInContext(gallivm->context));
 174
 175 #ifdef PIPE_ARCH_BIG_ENDIAN
 176    packed = lp_build_bswap(gallivm, packed, lp_type_uint(32));
 177 #endif
 178
 179    /* Broadcast the packed value to all four channels
 180     * before: packed = BGRA
 181     * after: packed = {BGRA, BGRA, BGRA, BGRA}
 182     */
 183    packed = LLVMBuildInsertElement(builder,
 184                                    LLVMGetUndef(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4)),
 185                                    packed,
 186                                    LLVMConstNull(LLVMInt32TypeInContext(gallivm->context)),
 187                                    "");
 188    packed = LLVMBuildShuffleVector(builder,
 189                                    packed,
 190                                    LLVMGetUndef(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4)),
 191                                    LLVMConstNull(LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4)),
 192                                    "");
 193
 194    /* Initialize vector constants */
 195    normalized = FALSE;
 196    needs_uitofp = FALSE;
 197    shift = 0;
 198
 199    /* Loop over 4 color components */
 200    for (i = 0; i < 4; ++i) {
 201       unsigned bits = desc->channel[i].size;
 202
 203       if (desc->channel[i].type == UTIL_FORMAT_TYPE_VOID) {
 204          shifts[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
 205          masks[i] = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
 206          scales[i] =  LLVMConstNull(LLVMFloatTypeInContext(gallivm->context));
 207       }
 208       else {
 209          unsigned long long mask = (1ULL << bits) - 1;
 210
 211          assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED);
 212
 213          if (bits == 32) {
 214             needs_uitofp = TRUE;
 215          }
 216
 217          shifts[i] = lp_build_const_int32(gallivm, shift);
 218          masks[i] = lp_build_const_int32(gallivm, mask);
 219
 220          if (desc->channel[i].normalized) {
 221             scales[i] = lp_build_const_float(gallivm, 1.0 / mask);
 222             normalized = TRUE;
 223          }
 224          else
 225             scales[i] =  lp_build_const_float(gallivm, 1.0);
 226       }
 227
 228       shift += bits;
 229    }
 230
 231    /* Ex: convert packed = {BGRA, BGRA, BGRA, BGRA}
 232     * into masked = {B, G, R, A}
 233     */
 234    shifted = LLVMBuildLShr(builder, packed, LLVMConstVector(shifts, 4), "");
 235    masked = LLVMBuildAnd(builder, shifted, LLVMConstVector(masks, 4), "");
 236
 237
 238    if (!needs_uitofp) {
 239       /* UIToFP can't be expressed in SSE2 */
 240       casted = LLVMBuildSIToFP(builder, masked, LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), "");
 241    } else {
 242       casted = LLVMBuildUIToFP(builder, masked, LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), "");
 243    }
 244
 245    /* At this point 'casted' may be a vector of floats such as
 246     * {255.0, 255.0, 255.0, 255.0}.  Next, if the pixel values are normalized
 247     * we'll scale this to {1.0, 1.0, 1.0, 1.0}.
 248     */
 249
 250    if (normalized)
 251       scaled = LLVMBuildFMul(builder, casted, LLVMConstVector(scales, 4), "");
 252    else
 253       scaled = casted;
 254
 255    return scaled;
 256 }
 257
 258
 259 /**
 260  * Pack a single pixel.
 261  *
 262  * @param rgba 4 float vector with the unpacked components.
 263  *
 264  * XXX: This is mostly for reference and testing -- operating a single pixel at
 265  * a time is rarely if ever needed.
 266  */
 267 LLVMValueRef
 268 lp_build_pack_rgba_aos(struct gallivm_state *gallivm,
 269                        const struct util_format_description *desc,
 270                        LLVMValueRef rgba)
 271 {
 272    LLVMBuilderRef builder = gallivm->builder;
 273    LLVMTypeRef type;
 274    LLVMValueRef packed = NULL;
 275    LLVMValueRef swizzles[4];
 276    LLVMValueRef shifted, casted, scaled, unswizzled;
 277    LLVMValueRef shifts[4];
 278    LLVMValueRef scales[4];
 279    boolean normalized;
 280    unsigned shift;
 281    unsigned i, j;
 282
 283    assert(desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
 284    assert(desc->block.width == 1);
 285    assert(desc->block.height == 1);
 286
 287    type = LLVMIntTypeInContext(gallivm->context, desc->block.bits);
 288
 289    /* Unswizzle the color components into the source vector. */
 290    for (i = 0; i < 4; ++i) {
 291       for (j = 0; j < 4; ++j) {
 292          if (desc->swizzle[j] == i)
 293             break;
 294       }
 295       if (j < 4)
 296          swizzles[i] = lp_build_const_int32(gallivm, j);
 297       else
 298          swizzles[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
 299    }
 300
 301    unswizzled = LLVMBuildShuffleVector(builder, rgba,
 302                                        LLVMGetUndef(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4)),
 303                                        LLVMConstVector(swizzles, 4), "");
 304
 305    normalized = FALSE;
 306    shift = 0;
 307    for (i = 0; i < 4; ++i) {
 308       unsigned bits = desc->channel[i].size;
 309
 310       if (desc->channel[i].type == UTIL_FORMAT_TYPE_VOID) {
 311          shifts[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
 312          scales[i] =  LLVMGetUndef(LLVMFloatTypeInContext(gallivm->context));
 313       }
 314       else {
 315          unsigned mask = (1 << bits) - 1;
 316
 317          assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED);
 318          assert(bits < 32);
 319
 320          shifts[i] = lp_build_const_int32(gallivm, shift);
 321
 322          if (desc->channel[i].normalized) {
 323             scales[i] = lp_build_const_float(gallivm, mask);
 324             normalized = TRUE;
 325          }
 326          else
 327             scales[i] = lp_build_const_float(gallivm, 1.0);
 328       }
 329
 330       shift += bits;
 331    }
 332
 333    if (normalized)
 334       scaled = LLVMBuildFMul(builder, unswizzled, LLVMConstVector(scales, 4), "");
 335    else
 336       scaled = unswizzled;
 337
 338    casted = LLVMBuildFPToSI(builder, scaled, LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4), "");
 339
 340    shifted = LLVMBuildShl(builder, casted, LLVMConstVector(shifts, 4), "");
 341
 342    /* Bitwise or all components */
 343    for (i = 0; i < 4; ++i) {
 344       if (desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED) {
 345          LLVMValueRef component = LLVMBuildExtractElement(builder, shifted,
 346                                                lp_build_const_int32(gallivm, i), "");
 347          if (packed)
 348             packed = LLVMBuildOr(builder, packed, component, "");
 349          else
 350             packed = component;
 351       }
 352    }
 353
 354    if (!packed)
 355       packed = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
 356
 357    if (desc->block.bits < 32)
 358       packed = LLVMBuildTrunc(builder, packed, type, "");
 359
 360    return packed;
 361 }
 362
 363
 364
 365
 366 /**
 367  * Fetch a pixel into a 4 float AoS.
 368  *
 369  * \param format_desc  describes format of the image we're fetching from
 370  * \param ptr  address of the pixel block (or the texel if uncompressed)
 371  * \param i, j  the sub-block pixel coordinates.  For non-compressed formats
 372  *              these will always be (0, 0).
 373  * \return  a 4 element vector with the pixel's RGBA values.
 374  */
 375 LLVMValueRef
 376 lp_build_fetch_rgba_aos(struct gallivm_state *gallivm,
 377                         const struct util_format_description *format_desc,
 378                         struct lp_type type,
 379                         LLVMValueRef base_ptr,
 380                         LLVMValueRef offset,
 381                         LLVMValueRef i,
 382                         LLVMValueRef j)
 383 {
 384    LLVMBuilderRef builder = gallivm->builder;
 385    unsigned num_pixels = type.length / 4;
 386    struct lp_build_context bld;
 387
 388    assert(type.length <= LP_MAX_VECTOR_LENGTH);
 389    assert(type.length % 4 == 0);
 390
 391    lp_build_context_init(&bld, gallivm, type);
 392
 393    /*
 394     * Trivial case
 395     *
 396     * The format matches the type (apart of a swizzle) so no need for
 397     * scaling or converting.
 398     */
 399
 400    if (format_matches_type(format_desc, type) &&
 401        format_desc->block.bits <= type.width * 4 &&
 402        util_is_power_of_two(format_desc->block.bits)) {
 403       LLVMValueRef packed;
 404       LLVMTypeRef dst_vec_type = lp_build_vec_type(gallivm, type);
 405       unsigned vec_len = type.width * type.length;
 406
 407       /*
 408        * The format matches the type (apart of a swizzle) so no need for
 409        * scaling or converting.
 410        */
 411
 412       packed = lp_build_gather(gallivm, type.length/4,
 413                                format_desc->block.bits, type.width*4,
 414                                base_ptr, offset);
 415
 416       assert(format_desc->block.bits <= vec_len);
 417
 418       packed = LLVMBuildBitCast(gallivm->builder, packed, dst_vec_type, "");
 419 #ifdef PIPE_ARCH_BIG_ENDIAN
 420       if (type.floating)
 421          packed = lp_build_bswap_vec(gallivm, packed, type,
 422                                     lp_type_float_vec(type.width, vec_len));
 423 #endif
 424       return lp_build_format_swizzle_aos(format_desc, &bld, packed);
 425    }
 426
 427    /*
 428     * Bit arithmetic
 429     */
 430
 431    if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN &&
 432        (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
 433         format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) &&
 434        format_desc->block.width == 1 &&
 435        format_desc->block.height == 1 &&
 436        util_is_power_of_two(format_desc->block.bits) &&
 437        format_desc->block.bits <= 32 &&
 438        format_desc->is_bitmask &&
 439        !format_desc->is_mixed &&
 440        (format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED ||
 441         format_desc->channel[1].type == UTIL_FORMAT_TYPE_UNSIGNED)) {
 442
 443       LLVMValueRef tmps[LP_MAX_VECTOR_LENGTH/4];
 444       LLVMValueRef res;
 445       unsigned k;
 446
 447       /*
 448        * Unpack a pixel at a time into a <4 x float> RGBA vector
 449        */
 450
 451       for (k = 0; k < num_pixels; ++k) {
 452          LLVMValueRef packed;
 453
 454          packed = lp_build_gather_elem(gallivm, num_pixels,
 455                                        format_desc->block.bits, 32,
 456                                        base_ptr, offset, k);
 457
 458          tmps[k] = lp_build_unpack_arith_rgba_aos(gallivm,
 459                                                   format_desc,
 460                                                   packed);
 461       }
 462
 463       /*
 464        * Type conversion.
 465        *
 466        * TODO: We could avoid floating conversion for integer to
 467        * integer conversions.
 468        */
 469
 470       if (gallivm_debug & GALLIVM_DEBUG_PERF && !type.floating) {
 471          debug_printf("%s: unpacking %s with floating point\n",
 472                       __FUNCTION__, format_desc->short_name);
 473       }
 474
 475       lp_build_conv(gallivm,
 476                     lp_float32_vec4_type(),
 477                     type,
 478                     tmps, num_pixels, &res, 1);
 479
 480       return lp_build_format_swizzle_aos(format_desc, &bld, res);
 481    }
 482
 483    /* If all channels are of same type and we are not using half-floats */
 484    if (util_format_is_array(format_desc)) {
 485       return lp_build_fetch_rgba_aos_array(gallivm, format_desc, type, base_ptr, offset);
 486    }
 487
 488    /*
 489     * YUV / subsampled formats
 490     */
 491
 492    if (format_desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
 493       struct lp_type tmp_type;
 494       LLVMValueRef tmp;
 495
 496       memset(&tmp_type, 0, sizeof tmp_type);
 497       tmp_type.width = 8;
 498       tmp_type.length = num_pixels * 4;
 499       tmp_type.norm = TRUE;
 500
 501       tmp = lp_build_fetch_subsampled_rgba_aos(gallivm,
 502                                                format_desc,
 503                                                num_pixels,
 504                                                base_ptr,
 505                                                offset,
 506                                                i, j);
 507
 508       lp_build_conv(gallivm,
 509                     tmp_type, type,
 510                     &tmp, 1, &tmp, 1);
 511
 512       return tmp;
 513    }
 514
 515    /*
 516     * Fallback to util_format_description::fetch_rgba_8unorm().
 517     */
 518
 519    if (format_desc->fetch_rgba_8unorm &&
 520        !type.floating && type.width == 8 && !type.sign && type.norm) {
 521       /*
 522        * Fallback to calling util_format_description::fetch_rgba_8unorm.
 523        *
 524        * This is definitely not the most efficient way of fetching pixels, as
 525        * we miss the opportunity to do vectorization, but this it is a
 526        * convenient for formats or scenarios for which there was no opportunity
 527        * or incentive to optimize.
 528        */
 529
 530       LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context);
 531       LLVMTypeRef pi8t = LLVMPointerType(i8t, 0);
 532       LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
 533       LLVMValueRef function;
 534       LLVMValueRef tmp_ptr;
 535       LLVMValueRef tmp;
 536       LLVMValueRef res;
 537       unsigned k;
 538
 539       if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 540          debug_printf("%s: falling back to util_format_%s_fetch_rgba_8unorm\n",
 541                       __FUNCTION__, format_desc->short_name);
 542       }
 543
 544       /*
 545        * Declare and bind format_desc->fetch_rgba_8unorm().
 546        */
 547
 548       {
 549          /*
 550           * Function to call looks like:
 551           *   fetch(uint8_t *dst, const uint8_t *src, unsigned i, unsigned j)
 552           */
 553          LLVMTypeRef ret_type;
 554          LLVMTypeRef arg_types[4];
 555          LLVMTypeRef function_type;
 556
 557          ret_type = LLVMVoidTypeInContext(gallivm->context);
 558          arg_types[0] = pi8t;
 559          arg_types[1] = pi8t;
 560          arg_types[2] = i32t;
 561          arg_types[3] = i32t;
 562          function_type = LLVMFunctionType(ret_type, arg_types,
 563                                           Elements(arg_types), 0);
 564
 565          /* make const pointer for the C fetch_rgba_8unorm function */
 566          function = lp_build_const_int_pointer(gallivm,
 567             func_to_pointer((func_pointer) format_desc->fetch_rgba_8unorm));
 568
 569          /* cast the callee pointer to the function's type */
 570          function = LLVMBuildBitCast(builder, function,
 571                                      LLVMPointerType(function_type, 0),
 572                                      "cast callee");
 573       }
 574
 575       tmp_ptr = lp_build_alloca(gallivm, i32t, "");
 576
 577       res = LLVMGetUndef(LLVMVectorType(i32t, num_pixels));
 578
 579       /*
 580        * Invoke format_desc->fetch_rgba_8unorm() for each pixel and insert the result
 581        * in the SoA vectors.
 582        */
 583
 584       for (k = 0; k < num_pixels; ++k) {
 585          LLVMValueRef index = lp_build_const_int32(gallivm, k);
 586          LLVMValueRef args[4];
 587
 588          args[0] = LLVMBuildBitCast(builder, tmp_ptr, pi8t, "");
 589          args[1] = lp_build_gather_elem_ptr(gallivm, num_pixels,
 590                                             base_ptr, offset, k);
 591
 592          if (num_pixels == 1) {
 593             args[2] = i;
 594             args[3] = j;
 595          }
 596          else {
 597             args[2] = LLVMBuildExtractElement(builder, i, index, "");
 598             args[3] = LLVMBuildExtractElement(builder, j, index, "");
 599          }
 600
 601          LLVMBuildCall(builder, function, args, Elements(args), "");
 602
 603          tmp = LLVMBuildLoad(builder, tmp_ptr, "");
 604
 605          if (num_pixels == 1) {
 606             res = tmp;
 607          }
 608          else {
 609             res = LLVMBuildInsertElement(builder, res, tmp, index, "");
 610          }
 611       }
 612
 613       /* Bitcast from <n x i32> to <4n x i8> */
 614       res = LLVMBuildBitCast(builder, res, bld.vec_type, "");
 615
 616       return res;
 617    }
 618
 619    /*
 620     * Fallback to util_format_description::fetch_rgba_float().
 621     */
 622
 623    if (format_desc->fetch_rgba_float) {
 624       /*
 625        * Fallback to calling util_format_description::fetch_rgba_float.
 626        *
 627        * This is definitely not the most efficient way of fetching pixels, as
 628        * we miss the opportunity to do vectorization, but this it is a
 629        * convenient for formats or scenarios for which there was no opportunity
 630        * or incentive to optimize.
 631        */
 632
 633       LLVMTypeRef f32t = LLVMFloatTypeInContext(gallivm->context);
 634       LLVMTypeRef f32x4t = LLVMVectorType(f32t, 4);
 635       LLVMTypeRef pf32t = LLVMPointerType(f32t, 0);
 636       LLVMTypeRef pi8t = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
 637       LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
 638       LLVMValueRef function;
 639       LLVMValueRef tmp_ptr;
 640       LLVMValueRef tmps[LP_MAX_VECTOR_LENGTH/4];
 641       LLVMValueRef res;
 642       unsigned k;
 643
 644       if (gallivm_debug & GALLIVM_DEBUG_PERF) {
 645          debug_printf("%s: falling back to util_format_%s_fetch_rgba_float\n",
 646                       __FUNCTION__, format_desc->short_name);
 647       }
 648
 649       /*
 650        * Declare and bind format_desc->fetch_rgba_float().
 651        */
 652
 653       {
 654          /*
 655           * Function to call looks like:
 656           *   fetch(float *dst, const uint8_t *src, unsigned i, unsigned j)
 657           */
 658          LLVMTypeRef ret_type;
 659          LLVMTypeRef arg_types[4];
 660
 661          ret_type = LLVMVoidTypeInContext(gallivm->context);
 662          arg_types[0] = pf32t;
 663          arg_types[1] = pi8t;
 664          arg_types[2] = i32t;
 665          arg_types[3] = i32t;
 666
 667          function = lp_build_const_func_pointer(gallivm,
 668                                                 func_to_pointer((func_pointer) format_desc->fetch_rgba_float),
 669                                                 ret_type,
 670                                                 arg_types, Elements(arg_types),
 671                                                 format_desc->short_name);
 672       }
 673
 674       tmp_ptr = lp_build_alloca(gallivm, f32x4t, "");
 675
 676       /*
 677        * Invoke format_desc->fetch_rgba_float() for each pixel and insert the result
 678        * in the SoA vectors.
 679        */
 680
 681       for (k = 0; k < num_pixels; ++k) {
 682          LLVMValueRef args[4];
 683
 684          args[0] = LLVMBuildBitCast(builder, tmp_ptr, pf32t, "");
 685          args[1] = lp_build_gather_elem_ptr(gallivm, num_pixels,
 686                                             base_ptr, offset, k);
 687
 688          if (num_pixels == 1) {
 689             args[2] = i;
 690             args[3] = j;
 691          }
 692          else {
 693             LLVMValueRef index = lp_build_const_int32(gallivm, k);
 694             args[2] = LLVMBuildExtractElement(builder, i, index, "");
 695             args[3] = LLVMBuildExtractElement(builder, j, index, "");
 696          }
 697
 698          LLVMBuildCall(builder, function, args, Elements(args), "");
 699
 700          tmps[k] = LLVMBuildLoad(builder, tmp_ptr, "");
 701       }
 702
 703       lp_build_conv(gallivm,
 704                     lp_float32_vec4_type(),
 705                     type,
 706                     tmps, num_pixels, &res, 1);
 707
 708       return res;
 709    }
 710
 711    assert(0);
 712    return lp_build_undef(gallivm, type);
 713 }