aboutsummaryrefslogtreecommitdiffstats
path: root/src/gallium/auxiliary/gallivm/lp_bld_format_aos.c
blob: b52acca1b3e44447bfc54c117ca7f6787cb01186 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
/**************************************************************************
 *
 * Copyright 2009 VMware, Inc.
 * All Rights Reserved.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the
 * "Software"), to deal in the Software without restriction, including
 * without limitation the rights to use, copy, modify, merge, publish,
 * distribute, sub license, and/or sell copies of the Software, and to
 * permit persons to whom the Software is furnished to do so, subject to
 * the following conditions:
 *
 * The above copyright notice and this permission notice (including the
 * next paragraph) shall be included in all copies or substantial portions
 * of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 *
 **************************************************************************/

/**
 * @file
 * AoS pixel format manipulation.
 *
 * @author Jose Fonseca <jfonseca@vmware.com>
 */


#include "util/u_format.h"
#include "util/u_memory.h"
#include "util/u_math.h"
#include "util/u_pointer.h"
#include "util/u_string.h"
#include "util/u_cpu_detect.h"

#include "lp_bld_arit.h"
#include "lp_bld_init.h"
#include "lp_bld_type.h"
#include "lp_bld_flow.h"
#include "lp_bld_const.h"
#include "lp_bld_conv.h"
#include "lp_bld_swizzle.h"
#include "lp_bld_gather.h"
#include "lp_bld_debug.h"
#include "lp_bld_format.h"
#include "lp_bld_pack.h"
#include "lp_bld_intr.h"
#include "lp_bld_logic.h"
#include "lp_bld_bitarit.h"


/**
 * Basic swizzling.  Rearrange the order of the unswizzled array elements
 * according to the format description.  PIPE_SWIZZLE_0/ONE are supported
 * too.
 * Ex: if unswizzled[4] = {B, G, R, x}, then swizzled_out[4] = {R, G, B, 1}.
 */
LLVMValueRef
lp_build_format_swizzle_aos(const struct util_format_description *desc,
                            struct lp_build_context *bld,
                            LLVMValueRef unswizzled)
{
   unsigned char swizzles[4];
   unsigned chan;

   assert(bld->type.length % 4 == 0);

   for (chan = 0; chan < 4; ++chan) {
      enum pipe_swizzle swizzle;

      if (desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) {
         /*
          * For ZS formats do RGBA = ZZZ1
          */
         if (chan == 3) {
            swizzle = PIPE_SWIZZLE_1;
         } else if (desc->swizzle[0] == PIPE_SWIZZLE_NONE) {
            swizzle = PIPE_SWIZZLE_0;
         } else {
            swizzle = desc->swizzle[0];
         }
      } else {
         swizzle = desc->swizzle[chan];
      }
      swizzles[chan] = swizzle;
   }

   return lp_build_swizzle_aos(bld, unswizzled, swizzles);
}


/**
 * Whether the format matches the vector type, apart of swizzles.
 */
static inline boolean
format_matches_type(const struct util_format_description *desc,
                    struct lp_type type)
{
   enum util_format_type chan_type;
   unsigned chan;

   assert(type.length % 4 == 0);

   if (desc->layout != UTIL_FORMAT_LAYOUT_PLAIN ||
       desc->colorspace != UTIL_FORMAT_COLORSPACE_RGB ||
       desc->block.width != 1 ||
       desc->block.height != 1) {
      return FALSE;
   }

   if (type.floating) {
      chan_type = UTIL_FORMAT_TYPE_FLOAT;
   } else if (type.fixed) {
      chan_type = UTIL_FORMAT_TYPE_FIXED;
   } else if (type.sign) {
      chan_type = UTIL_FORMAT_TYPE_SIGNED;
   } else {
      chan_type = UTIL_FORMAT_TYPE_UNSIGNED;
   }

   for (chan = 0; chan < desc->nr_channels; ++chan) {
      if (desc->channel[chan].size != type.width) {
         return FALSE;
      }

      if (desc->channel[chan].type != UTIL_FORMAT_TYPE_VOID) {
         if (desc->channel[chan].type != chan_type ||
             desc->channel[chan].normalized != type.norm) {
            return FALSE;
         }
      }
   }

   return TRUE;
}

/*
 * Do rounding when converting small unorm values to larger ones.
 * Not quite 100% accurate, as it's done by appending MSBs, but
 * should be good enough.
 */

static inline LLVMValueRef
scale_bits_up(struct gallivm_state *gallivm,
              int src_bits,
              int dst_bits,
              LLVMValueRef src,
              struct lp_type src_type)
{
   LLVMBuilderRef builder = gallivm->builder;
   LLVMValueRef result = src;

   if (src_bits == 1 && dst_bits > 1) {
      /*
       * Useful for a1 - we'd need quite some repeated copies otherwise.
       */
      struct lp_build_context bld;
      LLVMValueRef dst_mask;
      lp_build_context_init(&bld, gallivm, src_type);
      dst_mask = lp_build_const_int_vec(gallivm, src_type,
                                        (1 << dst_bits) - 1),
      result = lp_build_cmp(&bld, PIPE_FUNC_EQUAL, src,
                            lp_build_const_int_vec(gallivm, src_type, 0));
      result = lp_build_andnot(&bld, dst_mask, result);
   }
   else if (dst_bits > src_bits) {
      /* Scale up bits */
      int db = dst_bits - src_bits;

      /* Shift left by difference in bits */
      result = LLVMBuildShl(builder,
                            src,
                            lp_build_const_int_vec(gallivm, src_type, db),
                            "");

      if (db <= src_bits) {
         /* Enough bits in src to fill the remainder */
         LLVMValueRef lower = LLVMBuildLShr(builder,
                                            src,
                                            lp_build_const_int_vec(gallivm, src_type,
                                                                   src_bits - db),
                                            "");

         result = LLVMBuildOr(builder, result, lower, "");
      } else if (db > src_bits) {
         /* Need to repeatedly copy src bits to fill remainder in dst */
         unsigned n;

         for (n = src_bits; n < dst_bits; n *= 2) {
            LLVMValueRef shuv = lp_build_const_int_vec(gallivm, src_type, n);

            result = LLVMBuildOr(builder,
                                 result,
                                 LLVMBuildLShr(builder, result, shuv, ""),
                                 "");
         }
      }
   } else {
      assert (dst_bits == src_bits);
   }

   return result;
}

/**
 * Unpack a single pixel into its XYZW components.
 *
 * @param desc  the pixel format for the packed pixel value
 * @param packed integer pixel in a format such as PIPE_FORMAT_B8G8R8A8_UNORM
 *
 * @return XYZW in a float[4] or ubyte[4] or ushort[4] vector.
 */
static inline LLVMValueRef
lp_build_unpack_arith_rgba_aos(struct gallivm_state *gallivm,
                               const struct util_format_description *desc,
                               LLVMValueRef packed)
{
   LLVMBuilderRef builder = gallivm->builder;
   LLVMValueRef shifted, casted, scaled, masked;
   LLVMValueRef shifts[4];
   LLVMValueRef masks[4];
   LLVMValueRef scales[4];
   LLVMTypeRef vec32_type;

   boolean normalized;
   boolean needs_uitofp;
   unsigned i;

   /* TODO: Support more formats */
   assert(desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
   assert(desc->block.width == 1);
   assert(desc->block.height == 1);
   assert(desc->block.bits <= 32);

   /* Do the intermediate integer computations with 32bit integers since it
    * matches floating point size */
   assert (LLVMTypeOf(packed) == LLVMInt32TypeInContext(gallivm->context));

   vec32_type = LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4);

   /* Broadcast the packed value to all four channels
    * before: packed = BGRA
    * after: packed = {BGRA, BGRA, BGRA, BGRA}
    */
   packed = LLVMBuildInsertElement(builder, LLVMGetUndef(vec32_type), packed,
                                   LLVMConstNull(LLVMInt32TypeInContext(gallivm->context)),
                                   "");
   packed = LLVMBuildShuffleVector(builder, packed, LLVMGetUndef(vec32_type),
                                   LLVMConstNull(vec32_type),
                                   "");

   /* Initialize vector constants */
   normalized = FALSE;
   needs_uitofp = FALSE;

   /* Loop over 4 color components */
   for (i = 0; i < 4; ++i) {
      unsigned bits = desc->channel[i].size;
      unsigned shift = desc->channel[i].shift;

      if (desc->channel[i].type == UTIL_FORMAT_TYPE_VOID) {
         shifts[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
         masks[i] = LLVMConstNull(LLVMInt32TypeInContext(gallivm->context));
         scales[i] =  LLVMConstNull(LLVMFloatTypeInContext(gallivm->context));
      }
      else {
         unsigned long long mask = (1ULL << bits) - 1;

         assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED);

         if (bits == 32) {
            needs_uitofp = TRUE;
         }

         shifts[i] = lp_build_const_int32(gallivm, shift);
         masks[i] = lp_build_const_int32(gallivm, mask);

         if (desc->channel[i].normalized) {
            scales[i] = lp_build_const_float(gallivm, 1.0 / mask);
            normalized = TRUE;
         }
         else
            scales[i] =  lp_build_const_float(gallivm, 1.0);
      }
   }

   /* Ex: convert packed = {XYZW, XYZW, XYZW, XYZW}
    * into masked = {X, Y, Z, W}
    */
   if (desc->block.bits < 32 && normalized) {
      /*
       * Note: we cannot do the shift below on x86 natively until AVX2.
       *
       * Old llvm versions will resort to scalar extract/shift insert,
       * which is definitely terrible, new versions will just do
       * several vector shifts and shuffle/blend results together.
       * We could turn this into a variable left shift plus a constant
       * right shift, and llvm would then turn the variable left shift
       * into a mul for us (albeit without sse41 the mul needs emulation
       * too...). However, since we're going to do a float mul
       * anyway, we just adjust that mul instead (plus the mask), skipping
       * the shift completely.
       * We could also use a extra mul when the format isn't normalized and
       * we don't have AVX2 support, but don't bother for now. Unfortunately,
       * this strategy doesn't work for 32bit formats (such as rgb10a2 or even
       * rgba8 if it ends up here), as that would require UIToFP, albeit that
       * would be fixable with easy 16bit shuffle (unless there's channels
       * crossing 16bit boundaries).
       */
      for (i = 0; i < 4; ++i) {
         if (desc->channel[i].type != UTIL_FORMAT_TYPE_VOID) {
            unsigned bits = desc->channel[i].size;
            unsigned shift = desc->channel[i].shift;
            unsigned long long mask = ((1ULL << bits) - 1) << shift;
            scales[i] = lp_build_const_float(gallivm, 1.0 / mask);
            masks[i] = lp_build_const_int32(gallivm, mask);
         }
      }
      masked = LLVMBuildAnd(builder, packed, LLVMConstVector(masks, 4), "");
   } else {
      shifted = LLVMBuildLShr(builder, packed, LLVMConstVector(shifts, 4), "");
      masked = LLVMBuildAnd(builder, shifted, LLVMConstVector(masks, 4), "");
   }

   if (!needs_uitofp) {
      /* UIToFP can't be expressed in SSE2 */
      casted = LLVMBuildSIToFP(builder, masked, LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), "");
   } else {
      casted = LLVMBuildUIToFP(builder, masked, LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4), "");
   }

   /*
    * At this point 'casted' may be a vector of floats such as
    * {255.0, 255.0, 255.0, 255.0}. (Normalized values may be multiplied
    * by powers of two). Next, if the pixel values are normalized
    * we'll scale this to {1.0, 1.0, 1.0, 1.0}.
    */

   if (normalized)
      scaled = LLVMBuildFMul(builder, casted, LLVMConstVector(scales, 4), "");
   else
      scaled = casted;

   return scaled;
}


/**
 * Pack a single pixel.
 *
 * @param rgba 4 float vector with the unpacked components.
 *
 * XXX: This is mostly for reference and testing -- operating a single pixel at
 * a time is rarely if ever needed.
 */
LLVMValueRef
lp_build_pack_rgba_aos(struct gallivm_state *gallivm,
                       const struct util_format_description *desc,
                       LLVMValueRef rgba)
{
   LLVMBuilderRef builder = gallivm->builder;
   LLVMTypeRef type;
   LLVMValueRef packed = NULL;
   LLVMValueRef swizzles[4];
   LLVMValueRef shifted, casted, scaled, unswizzled;
   LLVMValueRef shifts[4];
   LLVMValueRef scales[4];
   boolean normalized;
   unsigned i, j;

   assert(desc->layout == UTIL_FORMAT_LAYOUT_PLAIN);
   assert(desc->block.width == 1);
   assert(desc->block.height == 1);

   type = LLVMIntTypeInContext(gallivm->context, desc->block.bits);

   /* Unswizzle the color components into the source vector. */
   for (i = 0; i < 4; ++i) {
      for (j = 0; j < 4; ++j) {
         if (desc->swizzle[j] == i)
            break;
      }
      if (j < 4)
         swizzles[i] = lp_build_const_int32(gallivm, j);
      else
         swizzles[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
   }

   unswizzled = LLVMBuildShuffleVector(builder, rgba,
                                       LLVMGetUndef(LLVMVectorType(LLVMFloatTypeInContext(gallivm->context), 4)),
                                       LLVMConstVector(swizzles, 4), "");

   normalized = FALSE;
   for (i = 0; i < 4; ++i) {
      unsigned bits = desc->channel[i].size;
      unsigned shift = desc->channel[i].shift;

      if (desc->channel[i].type == UTIL_FORMAT_TYPE_VOID) {
         shifts[i] = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));
         scales[i] =  LLVMGetUndef(LLVMFloatTypeInContext(gallivm->context));
      }
      else {
         unsigned mask = (1 << bits) - 1;

         assert(desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED);
         assert(bits < 32);

         shifts[i] = lp_build_const_int32(gallivm, shift);

         if (desc->channel[i].normalized) {
            scales[i] = lp_build_const_float(gallivm, mask);
            normalized = TRUE;
         }
         else
            scales[i] = lp_build_const_float(gallivm, 1.0);
      }
   }

   if (normalized)
      scaled = LLVMBuildFMul(builder, unswizzled, LLVMConstVector(scales, 4), "");
   else
      scaled = unswizzled;

   casted = LLVMBuildFPToSI(builder, scaled, LLVMVectorType(LLVMInt32TypeInContext(gallivm->context), 4), "");

   shifted = LLVMBuildShl(builder, casted, LLVMConstVector(shifts, 4), "");
   
   /* Bitwise or all components */
   for (i = 0; i < 4; ++i) {
      if (desc->channel[i].type == UTIL_FORMAT_TYPE_UNSIGNED) {
         LLVMValueRef component = LLVMBuildExtractElement(builder, shifted,
                                               lp_build_const_int32(gallivm, i), "");
         if (packed)
            packed = LLVMBuildOr(builder, packed, component, "");
         else
            packed = component;
      }
   }

   if (!packed)
      packed = LLVMGetUndef(LLVMInt32TypeInContext(gallivm->context));

   if (desc->block.bits < 32)
      packed = LLVMBuildTrunc(builder, packed, type, "");

   return packed;
}




/**
 * Fetch a pixel into a 4 float AoS.
 *
 * \param format_desc  describes format of the image we're fetching from
 * \param aligned  whether the data is guaranteed to be aligned
 * \param ptr  address of the pixel block (or the texel if uncompressed)
 * \param i, j  the sub-block pixel coordinates.  For non-compressed formats
 *              these will always be (0, 0).
 * \return  a 4 element vector with the pixel's RGBA values.
 */
LLVMValueRef
lp_build_fetch_rgba_aos(struct gallivm_state *gallivm,
                        const struct util_format_description *format_desc,
                        struct lp_type type,
                        boolean aligned,
                        LLVMValueRef base_ptr,
                        LLVMValueRef offset,
                        LLVMValueRef i,
                        LLVMValueRef j,
                        LLVMValueRef cache)
{
   LLVMBuilderRef builder = gallivm->builder;
   unsigned num_pixels = type.length / 4;
   struct lp_build_context bld;

   assert(type.length <= LP_MAX_VECTOR_LENGTH);
   assert(type.length % 4 == 0);

   lp_build_context_init(&bld, gallivm, type);

   /*
    * Trivial case
    *
    * The format matches the type (apart of a swizzle) so no need for
    * scaling or converting.
    */

   if (format_matches_type(format_desc, type) &&
       format_desc->block.bits <= type.width * 4 &&
       /* XXX this shouldn't be needed */
       util_is_power_of_two_or_zero(format_desc->block.bits)) {
      LLVMValueRef packed;
      LLVMTypeRef dst_vec_type = lp_build_vec_type(gallivm, type);
      struct lp_type fetch_type;
      unsigned vec_len = type.width * type.length;

      /*
       * The format matches the type (apart of a swizzle) so no need for
       * scaling or converting.
       */

      fetch_type = lp_type_uint(type.width*4);
      packed = lp_build_gather(gallivm, type.length/4,
                               format_desc->block.bits, fetch_type,
                               aligned, base_ptr, offset, TRUE);

      assert(format_desc->block.bits <= vec_len);
      (void) vec_len; /* silence unused var warning for non-debug build */

      packed = LLVMBuildBitCast(gallivm->builder, packed, dst_vec_type, "");
      return lp_build_format_swizzle_aos(format_desc, &bld, packed);
   }

   /*
    * Bit arithmetic for converting small_unorm to unorm8.
    *
    * This misses some opportunities for optimizations (like skipping mask
    * for the highest channel for instance, or doing bit scaling in parallel
    * for channels with the same bit width) but it should be passable for
    * all arithmetic formats.
    */
   if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN &&
       format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB &&
       util_format_fits_8unorm(format_desc) &&
       type.width == 8 && type.norm == 1 && type.sign == 0 &&
       type.fixed == 0 && type.floating == 0) {
      LLVMValueRef packed, res = NULL, chans[4], rgba[4];
      LLVMTypeRef dst_vec_type, conv_vec_type;
      struct lp_type fetch_type, conv_type;
      struct lp_build_context bld_conv;
      unsigned j;

      fetch_type = lp_type_uint(type.width*4);
      conv_type = lp_type_int_vec(type.width*4, type.width * type.length);
      dst_vec_type = lp_build_vec_type(gallivm, type);
      conv_vec_type = lp_build_vec_type(gallivm, conv_type);
      lp_build_context_init(&bld_conv, gallivm, conv_type);

      packed = lp_build_gather(gallivm, type.length/4,
                               format_desc->block.bits, fetch_type,
                               aligned, base_ptr, offset, TRUE);

      assert(format_desc->block.bits * type.length / 4 <=
             type.width * type.length);

      packed = LLVMBuildBitCast(gallivm->builder, packed, conv_vec_type, "");

      for (j = 0; j < format_desc->nr_channels; ++j) {
         unsigned mask = 0;
         unsigned sa = format_desc->channel[j].shift;

         mask = (1 << format_desc->channel[j].size) - 1;

         /* Extract bits from source */
         chans[j] = LLVMBuildLShr(builder, packed,
                                  lp_build_const_int_vec(gallivm, conv_type, sa),
                                  "");

         chans[j] = LLVMBuildAnd(builder, chans[j],
                                 lp_build_const_int_vec(gallivm, conv_type, mask),
                                 "");

         /* Scale bits */
         if (type.norm) {
            chans[j] = scale_bits_up(gallivm, format_desc->channel[j].size,
                                     type.width, chans[j], conv_type);
         }
      }
      /*
       * This is a hacked lp_build_format_swizzle_soa() since we need a
       * normalized 1 but only 8 bits in a 32bit vector...
       */
      for (j = 0; j < 4; ++j) {
         enum pipe_swizzle swizzle = format_desc->swizzle[j];
         if (swizzle == PIPE_SWIZZLE_1) {
            rgba[j] = lp_build_const_int_vec(gallivm, conv_type, (1 << type.width) - 1);
         } else {
            rgba[j] = lp_build_swizzle_soa_channel(&bld_conv, chans, swizzle);
         }
         if (j == 0) {
            res = rgba[j];
         } else {
            rgba[j] = LLVMBuildShl(builder, rgba[j],
                                   lp_build_const_int_vec(gallivm, conv_type,
                                                          j * type.width), "");
            res = LLVMBuildOr(builder, res, rgba[j], "");
         }
      }
      res = LLVMBuildBitCast(gallivm->builder, res, dst_vec_type, "");

      return res;
   }

   /*
    * Bit arithmetic
    */

   if (format_desc->layout == UTIL_FORMAT_LAYOUT_PLAIN &&
       (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
        format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS) &&
       format_desc->block.width == 1 &&
       format_desc->block.height == 1 &&
       /* XXX this shouldn't be needed */
       util_is_power_of_two_or_zero(format_desc->block.bits) &&
       format_desc->block.bits <= 32 &&
       format_desc->is_bitmask &&
       !format_desc->is_mixed &&
       (format_desc->channel[0].type == UTIL_FORMAT_TYPE_UNSIGNED ||
        format_desc->channel[1].type == UTIL_FORMAT_TYPE_UNSIGNED) &&
       !format_desc->channel[0].pure_integer) {

      LLVMValueRef tmps[LP_MAX_VECTOR_LENGTH/4];
      LLVMValueRef res[LP_MAX_VECTOR_WIDTH / 128];
      struct lp_type conv_type;
      unsigned k, num_conv_src, num_conv_dst;

      /*
       * Note this path is generally terrible for fetching multiple pixels.
       * We should make sure we cannot hit this code path for anything but
       * single pixels.
       */

      /*
       * Unpack a pixel at a time into a <4 x float> RGBA vector
       */

      for (k = 0; k < num_pixels; ++k) {
         LLVMValueRef packed;

         packed = lp_build_gather_elem(gallivm, num_pixels,
                                       format_desc->block.bits, 32, aligned,
                                       base_ptr, offset, k, FALSE);

         tmps[k] = lp_build_unpack_arith_rgba_aos(gallivm,
                                                  format_desc,
                                                  packed);
      }

      /*
       * Type conversion.
       *
       * TODO: We could avoid floating conversion for integer to
       * integer conversions.
       */

      if (gallivm_debug & GALLIVM_DEBUG_PERF && !type.floating) {
         debug_printf("%s: unpacking %s with floating point\n",
                      __FUNCTION__, format_desc->short_name);
      }

      conv_type = lp_float32_vec4_type();
      num_conv_src = num_pixels;
      num_conv_dst = 1;

      if (num_pixels % 8 == 0) {
         lp_build_concat_n(gallivm, lp_float32_vec4_type(),
                           tmps, num_pixels, tmps, num_pixels / 2);
         conv_type.length *= num_pixels / 4;
         num_conv_src = 4 * num_pixels / 8;
         if (type.width == 8 && type.floating == 0 && type.fixed == 0) {
            /*
             * FIXME: The fast float->unorm path (which is basically
             * skipping the MIN/MAX which are extremely pointless in any
             * case) requires that there's 2 destinations...
             * In any case, we really should make sure we don't hit this
             * code with multiple pixels for unorm8 dst types, it's
             * completely hopeless even if we do hit the right conversion.
             */
            type.length /= num_pixels / 4;
            num_conv_dst = num_pixels / 4;
         }
      }

      lp_build_conv(gallivm, conv_type, type,
                    tmps, num_conv_src, res, num_conv_dst);

      if (num_pixels % 8 == 0 &&
          (type.width == 8 && type.floating == 0 && type.fixed == 0)) {
         lp_build_concat_n(gallivm, type, res, num_conv_dst, res, 1);
      }

      return lp_build_format_swizzle_aos(format_desc, &bld, res[0]);
   }

   /* If all channels are of same type and we are not using half-floats */
   if (format_desc->is_array &&
       format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB) {
      assert(!format_desc->is_mixed);
      return lp_build_fetch_rgba_aos_array(gallivm, format_desc, type, base_ptr, offset);
   }

   /*
    * YUV / subsampled formats
    */

   if (format_desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
      struct lp_type tmp_type;
      LLVMValueRef tmp;

      memset(&tmp_type, 0, sizeof tmp_type);
      tmp_type.width = 8;
      tmp_type.length = num_pixels * 4;
      tmp_type.norm = TRUE;

      tmp = lp_build_fetch_subsampled_rgba_aos(gallivm,
                                               format_desc,
                                               num_pixels,
                                               base_ptr,
                                               offset,
                                               i, j);

      lp_build_conv(gallivm,
                    tmp_type, type,
                    &tmp, 1, &tmp, 1);

      return tmp;
   }

   /*
    * s3tc rgb formats
    */

   if (format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC && cache) {
      struct lp_type tmp_type;
      LLVMValueRef tmp;

      memset(&tmp_type, 0, sizeof tmp_type);
      tmp_type.width = 8;
      tmp_type.length = num_pixels * 4;
      tmp_type.norm = TRUE;

      tmp = lp_build_fetch_cached_texels(gallivm,
                                         format_desc,
                                         num_pixels,
                                         base_ptr,
                                         offset,
                                         i, j,
                                         cache);

      lp_build_conv(gallivm,
                    tmp_type, type,
                    &tmp, 1, &tmp, 1);

       return tmp;
   }

   /*
    * Fallback to util_format_description::fetch_rgba_8unorm().
    */

   if (format_desc->fetch_rgba_8unorm &&
       !type.floating && type.width == 8 && !type.sign && type.norm) {
      /*
       * Fallback to calling util_format_description::fetch_rgba_8unorm.
       *
       * This is definitely not the most efficient way of fetching pixels, as
       * we miss the opportunity to do vectorization, but this it is a
       * convenient for formats or scenarios for which there was no opportunity
       * or incentive to optimize.
       */

      LLVMTypeRef i8t = LLVMInt8TypeInContext(gallivm->context);
      LLVMTypeRef pi8t = LLVMPointerType(i8t, 0);
      LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
      LLVMValueRef function;
      LLVMValueRef tmp_ptr;
      LLVMValueRef tmp;
      LLVMValueRef res;
      unsigned k;

      if (gallivm_debug & GALLIVM_DEBUG_PERF) {
         debug_printf("%s: falling back to util_format_%s_fetch_rgba_8unorm\n",
                      __FUNCTION__, format_desc->short_name);
      }

      /*
       * Declare and bind format_desc->fetch_rgba_8unorm().
       */

      {
         /*
          * Function to call looks like:
          *   fetch(uint8_t *dst, const uint8_t *src, unsigned i, unsigned j)
          */
         LLVMTypeRef ret_type;
         LLVMTypeRef arg_types[4];
         LLVMTypeRef function_type;

         ret_type = LLVMVoidTypeInContext(gallivm->context);
         arg_types[0] = pi8t;
         arg_types[1] = pi8t;
         arg_types[2] = i32t;
         arg_types[3] = i32t;
         function_type = LLVMFunctionType(ret_type, arg_types,
                                          ARRAY_SIZE(arg_types), 0);

         /* make const pointer for the C fetch_rgba_8unorm function */
         function = lp_build_const_int_pointer(gallivm,
            func_to_pointer((func_pointer) format_desc->fetch_rgba_8unorm));

         /* cast the callee pointer to the function's type */
         function = LLVMBuildBitCast(builder, function,
                                     LLVMPointerType(function_type, 0),
                                     "cast callee");
      }

      tmp_ptr = lp_build_alloca(gallivm, i32t, "");

      res = LLVMGetUndef(LLVMVectorType(i32t, num_pixels));

      /*
       * Invoke format_desc->fetch_rgba_8unorm() for each pixel and insert the result
       * in the SoA vectors.
       */

      for (k = 0; k < num_pixels; ++k) {
         LLVMValueRef index = lp_build_const_int32(gallivm, k);
         LLVMValueRef args[4];

         args[0] = LLVMBuildBitCast(builder, tmp_ptr, pi8t, "");
         args[1] = lp_build_gather_elem_ptr(gallivm, num_pixels,
                                            base_ptr, offset, k);

         if (num_pixels == 1) {
            args[2] = i;
            args[3] = j;
         }
         else {
            args[2] = LLVMBuildExtractElement(builder, i, index, "");
            args[3] = LLVMBuildExtractElement(builder, j, index, "");
         }

         LLVMBuildCall(builder, function, args, ARRAY_SIZE(args), "");

         tmp = LLVMBuildLoad(builder, tmp_ptr, "");

         if (num_pixels == 1) {
            res = tmp;
         }
         else {
            res = LLVMBuildInsertElement(builder, res, tmp, index, "");
         }
      }

      /* Bitcast from <n x i32> to <4n x i8> */
      res = LLVMBuildBitCast(builder, res, bld.vec_type, "");

      return res;
   }

   /*
    * Fallback to util_format_description::fetch_rgba_float().
    */

   if (format_desc->fetch_rgba_float) {
      /*
       * Fallback to calling util_format_description::fetch_rgba_float.
       *
       * This is definitely not the most efficient way of fetching pixels, as
       * we miss the opportunity to do vectorization, but this it is a
       * convenient for formats or scenarios for which there was no opportunity
       * or incentive to optimize.
       */

      LLVMTypeRef f32t = LLVMFloatTypeInContext(gallivm->context);
      LLVMTypeRef f32x4t = LLVMVectorType(f32t, 4);
      LLVMTypeRef pf32t = LLVMPointerType(f32t, 0);
      LLVMTypeRef pi8t = LLVMPointerType(LLVMInt8TypeInContext(gallivm->context), 0);
      LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
      LLVMValueRef function;
      LLVMValueRef tmp_ptr;
      LLVMValueRef tmps[LP_MAX_VECTOR_LENGTH/4];
      LLVMValueRef res;
      unsigned k;

      if (gallivm_debug & GALLIVM_DEBUG_PERF) {
         debug_printf("%s: falling back to util_format_%s_fetch_rgba_float\n",
                      __FUNCTION__, format_desc->short_name);
      }

      /*
       * Declare and bind format_desc->fetch_rgba_float().
       */

      {
         /*
          * Function to call looks like:
          *   fetch(float *dst, const uint8_t *src, unsigned i, unsigned j)
          */
         LLVMTypeRef ret_type;
         LLVMTypeRef arg_types[4];

         ret_type = LLVMVoidTypeInContext(gallivm->context);
         arg_types[0] = pf32t;
         arg_types[1] = pi8t;
         arg_types[2] = i32t;
         arg_types[3] = i32t;

         function = lp_build_const_func_pointer(gallivm,
                                                func_to_pointer((func_pointer) format_desc->fetch_rgba_float),
                                                ret_type,
                                                arg_types, ARRAY_SIZE(arg_types),
                                                format_desc->short_name);
      }

      tmp_ptr = lp_build_alloca(gallivm, f32x4t, "");

      /*
       * Invoke format_desc->fetch_rgba_float() for each pixel and insert the result
       * in the SoA vectors.
       */

      for (k = 0; k < num_pixels; ++k) {
         LLVMValueRef args[4];

         args[0] = LLVMBuildBitCast(builder, tmp_ptr, pf32t, "");
         args[1] = lp_build_gather_elem_ptr(gallivm, num_pixels,
                                            base_ptr, offset, k);

         if (num_pixels == 1) {
            args[2] = i;
            args[3] = j;
         }
         else {
            LLVMValueRef index = lp_build_const_int32(gallivm, k);
            args[2] = LLVMBuildExtractElement(builder, i, index, "");
            args[3] = LLVMBuildExtractElement(builder, j, index, "");
         }

         LLVMBuildCall(builder, function, args, ARRAY_SIZE(args), "");

         tmps[k] = LLVMBuildLoad(builder, tmp_ptr, "");
      }

      lp_build_conv(gallivm,
                    lp_float32_vec4_type(),
                    type,
                    tmps, num_pixels, &res, 1);

      return res;
   }

   assert(!util_format_is_pure_integer(format_desc->format));

   assert(0);
   return lp_build_undef(gallivm, type);
}