summaryrefslogtreecommitdiffstats
path: root/src/gallium/auxiliary/gallivm/lp_bld_conv.c
blob: eb2d09638fa0e96bd423bdd3d9f1d93d17a19761 (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
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
/**************************************************************************
 *
 * 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
 * Helper functions for type conversions.
 *
 * We want to use the fastest type for a given computation whenever feasible.
 * The other side of this is that we need to be able convert between several
 * types accurately and efficiently.
 *
 * Conversion between types of different bit width is quite complex since a 
 *
 * To remember there are a few invariants in type conversions:
 *
 * - register width must remain constant:
 *
 *     src_type.width * src_type.length == dst_type.width * dst_type.length
 *
 * - total number of elements must remain constant:
 *
 *     src_type.length * num_srcs == dst_type.length * num_dsts
 *
 * It is not always possible to do the conversion both accurately and
 * efficiently, usually due to lack of adequate machine instructions. In these
 * cases it is important not to cut shortcuts here and sacrifice accuracy, as
 * there this functions can be used anywhere. In the future we might have a
 * precision parameter which can gauge the accuracy vs efficiency compromise,
 * but for now if the data conversion between two stages happens to be the
 * bottleneck, then most likely should just avoid converting at all and run
 * both stages with the same type.
 *
 * Make sure to run lp_test_conv unit test after any change to this file.
 *
 * @author Jose Fonseca <jfonseca@vmware.com>
 */


#include "util/u_debug.h"
#include "util/u_math.h"
#include "util/u_half.h"
#include "util/u_cpu_detect.h"

#include "lp_bld_type.h"
#include "lp_bld_const.h"
#include "lp_bld_arit.h"
#include "lp_bld_bitarit.h"
#include "lp_bld_pack.h"
#include "lp_bld_conv.h"
#include "lp_bld_logic.h"
#include "lp_bld_intr.h"
#include "lp_bld_printf.h"
#include "lp_bld_format.h"



/**
 * Byte swap on element. It will construct a call to intrinsic llvm.bswap
 * based on the type.
 *
 * @param res           element to byte swap.
 * @param type          int16_t, int32_t, int64_t, float or double
 * @param 
 */
LLVMValueRef
lp_build_bswap(struct gallivm_state *gallivm,
               LLVMValueRef res,
               struct lp_type type)
{
   LLVMTypeRef int_type = LLVMIntTypeInContext(gallivm->context,
                                               type.width);
   const char *intrinsic = NULL;
   if (type.width == 8)
      return res;
   if (type.width == 16)
      intrinsic = "llvm.bswap.i16";
   else if (type.width == 32)
     intrinsic = "llvm.bswap.i32";
   else if (type.width == 64)
      intrinsic = "llvm.bswap.i64";

   assert (intrinsic != NULL);

   /* In case of a floating-point type cast to a int of same size and then
    * cast back to fp type.
    */
   if (type.floating)
      res = LLVMBuildBitCast(gallivm->builder, res, int_type, "");
   res = lp_build_intrinsic_unary(gallivm->builder, intrinsic, int_type, res);
   if (type.floating)
      res = LLVMBuildBitCast(gallivm->builder, res,
                             lp_build_elem_type(gallivm, type), "");
   return res;
}


/**
 * Byte swap every element in the vector.
 *
 * @param packed        <vector> to convert
 * @param src_type      <vector> type of int16_t, int32_t, int64_t, float or
 *                      double
 * @param dst_type      <vector> type to return
 */
LLVMValueRef
lp_build_bswap_vec(struct gallivm_state *gallivm,
                   LLVMValueRef packed,
                   struct lp_type src_type_vec,
                   struct lp_type dst_type_vec)
{
   LLVMBuilderRef builder = gallivm->builder;
   LLVMTypeRef dst_type = lp_build_elem_type(gallivm, dst_type_vec);
   LLVMValueRef res;

   if (src_type_vec.length == 1) {
      res = lp_build_bswap(gallivm, packed, src_type_vec);
      res = LLVMBuildBitCast(gallivm->builder, res, dst_type, "");
   } else {
      unsigned i;
      res = LLVMGetUndef(lp_build_vec_type(gallivm, dst_type_vec));
      for (i = 0; i < src_type_vec.length; ++i) {
         LLVMValueRef index = lp_build_const_int32(gallivm, i);
         LLVMValueRef elem = LLVMBuildExtractElement(builder, packed, index, "");
         elem = lp_build_bswap(gallivm, elem, src_type_vec);
         elem = LLVMBuildBitCast(gallivm->builder, elem, dst_type, "");
         res = LLVMBuildInsertElement(gallivm->builder, res, elem, index, "");
      }
   }
   return res;
}


/**
 * Converts int16 half-float to float32
 * Note this can be performed in 1 instruction if vcvtph2ps exists (f16c/cvt16)
 * [llvm.x86.vcvtph2ps / _mm_cvtph_ps]
 *
 * @param src           value to convert
 *
 */
LLVMValueRef
lp_build_half_to_float(struct gallivm_state *gallivm,
                       LLVMValueRef src)
{
   LLVMBuilderRef builder = gallivm->builder;
   LLVMTypeRef src_type = LLVMTypeOf(src);
   unsigned src_length = LLVMGetTypeKind(src_type) == LLVMVectorTypeKind ?
                            LLVMGetVectorSize(src_type) : 1;

   struct lp_type f32_type = lp_type_float_vec(32, 32 * src_length);
   struct lp_type i32_type = lp_type_int_vec(32, 32 * src_length);
   LLVMTypeRef int_vec_type = lp_build_vec_type(gallivm, i32_type);
   LLVMValueRef h;

   if (util_cpu_caps.has_f16c && HAVE_LLVM >= 0x0301 &&
       (src_length == 4 || src_length == 8)) {
      const char *intrinsic = NULL;
      if (src_length == 4) {
         src = lp_build_pad_vector(gallivm, src, 8);
         intrinsic = "llvm.x86.vcvtph2ps.128";
      }
      else {
         intrinsic = "llvm.x86.vcvtph2ps.256";
      }
      return lp_build_intrinsic_unary(builder, intrinsic,
                                      lp_build_vec_type(gallivm, f32_type), src);
   }

   /* Convert int16 vector to int32 vector by zero ext (might generate bad code) */
   h = LLVMBuildZExt(builder, src, int_vec_type, "");
   return lp_build_smallfloat_to_float(gallivm, f32_type, h, 10, 5, 0, true);
}


/**
 * Converts float32 to int16 half-float
 * Note this can be performed in 1 instruction if vcvtps2ph exists (f16c/cvt16)
 * [llvm.x86.vcvtps2ph / _mm_cvtps_ph]
 *
 * @param src           value to convert
 *
 * Convert float32 to half floats, preserving Infs and NaNs,
 * with rounding towards zero (trunc).
 */
LLVMValueRef
lp_build_float_to_half(struct gallivm_state *gallivm,
                       LLVMValueRef src)
{
   LLVMBuilderRef builder = gallivm->builder;
   LLVMTypeRef f32_vec_type = LLVMTypeOf(src);
   unsigned length = LLVMGetTypeKind(f32_vec_type) == LLVMVectorTypeKind
                   ? LLVMGetVectorSize(f32_vec_type) : 1;
   struct lp_type i32_type = lp_type_int_vec(32, 32 * length);
   struct lp_type i16_type = lp_type_int_vec(16, 16 * length);
   LLVMValueRef result;

   if (util_cpu_caps.has_f16c && HAVE_LLVM >= 0x0301 &&
       (length == 4 || length == 8)) {
      struct lp_type i168_type = lp_type_int_vec(16, 16 * 8);
      unsigned mode = 3; /* same as LP_BUILD_ROUND_TRUNCATE */
      LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
      const char *intrinsic = NULL;
      if (length == 4) {
         intrinsic = "llvm.x86.vcvtps2ph.128";
      }
      else {
         intrinsic = "llvm.x86.vcvtps2ph.256";
      }
      result = lp_build_intrinsic_binary(builder, intrinsic,
                                         lp_build_vec_type(gallivm, i168_type),
                                         src, LLVMConstInt(i32t, mode, 0));
      if (length == 4) {
         result = lp_build_extract_range(gallivm, result, 0, 4);
      }
   }

   else {
      result = lp_build_float_to_smallfloat(gallivm, i32_type, src, 10, 5, 0, true);
      /* Convert int32 vector to int16 vector by trunc (might generate bad code) */
      result = LLVMBuildTrunc(builder, result, lp_build_vec_type(gallivm, i16_type), "");
   }

   /*
    * Debugging code.
    */
   if (0) {
     LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
     LLVMTypeRef i16t = LLVMInt16TypeInContext(gallivm->context);
     LLVMTypeRef f32t = LLVMFloatTypeInContext(gallivm->context);
     LLVMValueRef ref_result = LLVMGetUndef(LLVMVectorType(i16t, length));
     unsigned i;

     LLVMTypeRef func_type = LLVMFunctionType(i16t, &f32t, 1, 0);
     LLVMValueRef func = lp_build_const_int_pointer(gallivm, func_to_pointer((func_pointer)util_float_to_half));
     func = LLVMBuildBitCast(builder, func, LLVMPointerType(func_type, 0), "util_float_to_half");

     for (i = 0; i < length; ++i) {
        LLVMValueRef index = LLVMConstInt(i32t, i, 0);
        LLVMValueRef f32 = LLVMBuildExtractElement(builder, src, index, "");
#if 0
        /* XXX: not really supported by backends */
        LLVMValueRef f16 = lp_build_intrinsic_unary(builder, "llvm.convert.to.fp16", i16t, f32);
#else
        LLVMValueRef f16 = LLVMBuildCall(builder, func, &f32, 1, "");
#endif
        ref_result = LLVMBuildInsertElement(builder, ref_result, f16, index, "");
     }

     lp_build_print_value(gallivm, "src  = ", src);
     lp_build_print_value(gallivm, "llvm = ", result);
     lp_build_print_value(gallivm, "util = ", ref_result);
     lp_build_printf(gallivm, "\n");
  }

   return result;
}


/**
 * Special case for converting clamped IEEE-754 floats to unsigned norms.
 *
 * The mathematical voodoo below may seem excessive but it is actually
 * paramount we do it this way for several reasons. First, there is no single
 * precision FP to unsigned integer conversion Intel SSE instruction. Second,
 * secondly, even if there was, since the FP's mantissa takes only a fraction
 * of register bits the typically scale and cast approach would require double
 * precision for accurate results, and therefore half the throughput
 *
 * Although the result values can be scaled to an arbitrary bit width specified
 * by dst_width, the actual result type will have the same width.
 *
 * Ex: src = { float, float, float, float }
 * return { i32, i32, i32, i32 } where each value is in [0, 2^dst_width-1].
 */
LLVMValueRef
lp_build_clamped_float_to_unsigned_norm(struct gallivm_state *gallivm,
                                        struct lp_type src_type,
                                        unsigned dst_width,
                                        LLVMValueRef src)
{
   LLVMBuilderRef builder = gallivm->builder;
   LLVMTypeRef int_vec_type = lp_build_int_vec_type(gallivm, src_type);
   LLVMValueRef res;
   unsigned mantissa;

   assert(src_type.floating);
   assert(dst_width <= src_type.width);
   src_type.sign = FALSE;

   mantissa = lp_mantissa(src_type);

   if (dst_width <= mantissa) {
      /*
       * Apply magic coefficients that will make the desired result to appear
       * in the lowest significant bits of the mantissa, with correct rounding.
       *
       * This only works if the destination width fits in the mantissa.
       */

      unsigned long long ubound;
      unsigned long long mask;
      double scale;
      double bias;

      ubound = (1ULL << dst_width);
      mask = ubound - 1;
      scale = (double)mask/ubound;
      bias = (double)(1ULL << (mantissa - dst_width));

      res = LLVMBuildFMul(builder, src, lp_build_const_vec(gallivm, src_type, scale), "");
      res = LLVMBuildFAdd(builder, res, lp_build_const_vec(gallivm, src_type, bias), "");
      res = LLVMBuildBitCast(builder, res, int_vec_type, "");
      res = LLVMBuildAnd(builder, res,
                         lp_build_const_int_vec(gallivm, src_type, mask), "");
   }
   else if (dst_width == (mantissa + 1)) {
      /*
       * The destination width matches exactly what can be represented in
       * floating point (i.e., mantissa + 1 bits). So do a straight
       * multiplication followed by casting. No further rounding is necessary.
       */

      double scale;

      scale = (double)((1ULL << dst_width) - 1);

      res = LLVMBuildFMul(builder, src,
                          lp_build_const_vec(gallivm, src_type, scale), "");
      res = LLVMBuildFPToSI(builder, res, int_vec_type, "");
   }
   else {
      /*
       * The destination exceeds what can be represented in the floating point.
       * So multiply by the largest power two we get away with, and when
       * subtract the most significant bit to rescale to normalized values.
       *
       * The largest power of two factor we can get away is
       * (1 << (src_type.width - 1)), because we need to use signed . In theory it
       * should be (1 << (src_type.width - 2)), but IEEE 754 rules states
       * INT_MIN should be returned in FPToSI, which is the correct result for
       * values near 1.0!
       *
       * This means we get (src_type.width - 1) correct bits for values near 0.0,
       * and (mantissa + 1) correct bits for values near 1.0. Equally or more
       * important, we also get exact results for 0.0 and 1.0.
       */

      unsigned n = MIN2(src_type.width - 1, dst_width);

      double scale = (double)(1ULL << n);
      unsigned lshift = dst_width - n;
      unsigned rshift = n;
      LLVMValueRef lshifted;
      LLVMValueRef rshifted;

      res = LLVMBuildFMul(builder, src,
                          lp_build_const_vec(gallivm, src_type, scale), "");
      res = LLVMBuildFPToSI(builder, res, int_vec_type, "");

      /*
       * Align the most significant bit to its final place.
       *
       * This will cause 1.0 to overflow to 0, but the later adjustment will
       * get it right.
       */
      if (lshift) {
         lshifted = LLVMBuildShl(builder, res,
                                 lp_build_const_int_vec(gallivm, src_type,
                                                        lshift), "");
      } else {
         lshifted = res;
      }

      /*
       * Align the most significant bit to the right.
       */
      rshifted =  LLVMBuildLShr(builder, res,
                                lp_build_const_int_vec(gallivm, src_type, rshift),
                                "");

      /*
       * Subtract the MSB to the LSB, therefore re-scaling from
       * (1 << dst_width) to ((1 << dst_width) - 1).
       */

      res = LLVMBuildSub(builder, lshifted, rshifted, "");
   }

   return res;
}


/**
 * Inverse of lp_build_clamped_float_to_unsigned_norm above.
 * Ex: src = { i32, i32, i32, i32 } with values in range [0, 2^src_width-1]
 * return {float, float, float, float} with values in range [0, 1].
 */
LLVMValueRef
lp_build_unsigned_norm_to_float(struct gallivm_state *gallivm,
                                unsigned src_width,
                                struct lp_type dst_type,
                                LLVMValueRef src)
{
   LLVMBuilderRef builder = gallivm->builder;
   LLVMTypeRef vec_type = lp_build_vec_type(gallivm, dst_type);
   LLVMTypeRef int_vec_type = lp_build_int_vec_type(gallivm, dst_type);
   LLVMValueRef bias_;
   LLVMValueRef res;
   unsigned mantissa;
   unsigned n;
   unsigned long long ubound;
   unsigned long long mask;
   double scale;
   double bias;

   assert(dst_type.floating);

   mantissa = lp_mantissa(dst_type);

   if (src_width <= (mantissa + 1)) {
      /*
       * The source width matches fits what can be represented in floating
       * point (i.e., mantissa + 1 bits). So do a straight multiplication
       * followed by casting. No further rounding is necessary.
       */

      scale = 1.0/(double)((1ULL << src_width) - 1);
      res = LLVMBuildSIToFP(builder, src, vec_type, "");
      res = LLVMBuildFMul(builder, res,
                          lp_build_const_vec(gallivm, dst_type, scale), "");
      return res;
   }
   else {
      /*
       * The source width exceeds what can be represented in floating
       * point. So truncate the incoming values.
       */

      n = MIN2(mantissa, src_width);

      ubound = ((unsigned long long)1 << n);
      mask = ubound - 1;
      scale = (double)ubound/mask;
      bias = (double)((unsigned long long)1 << (mantissa - n));

      res = src;

      if (src_width > mantissa) {
         int shift = src_width - mantissa;
         res = LLVMBuildLShr(builder, res,
                             lp_build_const_int_vec(gallivm, dst_type, shift), "");
      }

      bias_ = lp_build_const_vec(gallivm, dst_type, bias);

      res = LLVMBuildOr(builder,
                        res,
                        LLVMBuildBitCast(builder, bias_, int_vec_type, ""), "");

      res = LLVMBuildBitCast(builder, res, vec_type, "");

      res = LLVMBuildFSub(builder, res, bias_, "");
      res = LLVMBuildFMul(builder, res, lp_build_const_vec(gallivm, dst_type, scale), "");
   }

   return res;
}


/**
 * Pick a suitable num_dsts for lp_build_conv to ensure optimal cases are used.
 *
 * Returns the number of dsts created from src
 */
int lp_build_conv_auto(struct gallivm_state *gallivm,
                       struct lp_type src_type,
                       struct lp_type* dst_type,
                       const LLVMValueRef *src,
                       unsigned num_srcs,
                       LLVMValueRef *dst)
{
   int i;
   int num_dsts = num_srcs;

   if (src_type.floating == dst_type->floating &&
       src_type.width == dst_type->width &&
       src_type.length == dst_type->length &&
       src_type.fixed == dst_type->fixed &&
       src_type.norm == dst_type->norm &&
       src_type.sign == dst_type->sign)
      return num_dsts;

   /* Special case 4x4f -> 1x16ub or 2x8f -> 1x16ub
    */
   if (src_type.floating == 1 &&
       src_type.fixed    == 0 &&
       src_type.sign     == 1 &&
       src_type.norm     == 0 &&
       src_type.width    == 32 &&

       dst_type->floating == 0 &&
       dst_type->fixed    == 0 &&
       dst_type->sign     == 0 &&
       dst_type->norm     == 1 &&
       dst_type->width    == 8)
   {
      /* Special case 4x4f --> 1x16ub */
      if (src_type.length == 4 && util_cpu_caps.has_sse2)
      {
         assert((num_srcs % 4) == 0);

         num_dsts = num_srcs / 4;
         dst_type->length = 16;

         lp_build_conv(gallivm, src_type, *dst_type, src, num_srcs, dst, num_dsts);
         return num_dsts;
      }

      /* Special case 2x8f --> 1x16ub */
      if (src_type.length == 8 && util_cpu_caps.has_avx)
      {
         assert((num_srcs % 2) == 0);

         num_dsts = num_srcs / 2;
         dst_type->length = 16;

         lp_build_conv(gallivm, src_type, *dst_type, src, num_srcs, dst, num_dsts);
         return num_dsts;
      }
   }

   /* lp_build_resize does not support M:N */
   if (src_type.width == dst_type->width) {
      lp_build_conv(gallivm, src_type, *dst_type, src, num_srcs, dst, num_dsts);
   } else {
      for (i = 0; i < num_srcs; ++i) {
         lp_build_conv(gallivm, src_type, *dst_type, &src[i], 1, &dst[i], 1);
      }
   }

   return num_dsts;
}


/**
 * Generic type conversion.
 *
 * TODO: Take a precision argument, or even better, add a new precision member
 * to the lp_type union.
 */
void
lp_build_conv(struct gallivm_state *gallivm,
              struct lp_type src_type,
              struct lp_type dst_type,
              const LLVMValueRef *src, unsigned num_srcs,
              LLVMValueRef *dst, unsigned num_dsts)
{
   LLVMBuilderRef builder = gallivm->builder;
   struct lp_type tmp_type;
   LLVMValueRef tmp[LP_MAX_VECTOR_LENGTH];
   unsigned num_tmps;
   unsigned i;

   /* We must not loose or gain channels. Only precision */
   assert(src_type.length * num_srcs == dst_type.length * num_dsts);

   assert(src_type.length <= LP_MAX_VECTOR_LENGTH);
   assert(dst_type.length <= LP_MAX_VECTOR_LENGTH);
   assert(num_srcs <= LP_MAX_VECTOR_LENGTH);
   assert(num_dsts <= LP_MAX_VECTOR_LENGTH);

   tmp_type = src_type;
   for(i = 0; i < num_srcs; ++i) {
      assert(lp_check_value(src_type, src[i]));
      tmp[i] = src[i];
   }
   num_tmps = num_srcs;


   /* Special case 4x4f --> 1x16ub 
    */
   if (src_type.floating == 1 &&
       src_type.fixed    == 0 &&
       src_type.sign     == 1 &&
       src_type.norm     == 0 &&
       src_type.width    == 32 &&
       src_type.length   == 4 &&

       dst_type.floating == 0 &&
       dst_type.fixed    == 0 &&
       dst_type.sign     == 0 &&
       dst_type.norm     == 1 &&
       dst_type.width    == 8 &&
       dst_type.length   == 16 &&

       4 * num_dsts      == num_srcs &&

       util_cpu_caps.has_sse2)
   {
      struct lp_build_context bld;
      struct lp_type int16_type = dst_type;
      struct lp_type int32_type = dst_type;
      LLVMValueRef const_255f;
      unsigned i, j;

      lp_build_context_init(&bld, gallivm, src_type);

      int16_type.width *= 2;
      int16_type.length /= 2;
      int16_type.sign = 1;

      int32_type.width *= 4;
      int32_type.length /= 4;
      int32_type.sign = 1;

      const_255f = lp_build_const_vec(gallivm, src_type, 255.0f);

      for (i = 0; i < num_dsts; ++i, src += 4) {
         LLVMValueRef lo, hi;

         for (j = 0; j < 4; ++j) {
            tmp[j] = LLVMBuildFMul(builder, src[j], const_255f, "");
            tmp[j] = lp_build_iround(&bld, tmp[j]);
         }

         /* relying on clamping behavior of sse2 intrinsics here */
         lo = lp_build_pack2(gallivm, int32_type, int16_type, tmp[0], tmp[1]);
         hi = lp_build_pack2(gallivm, int32_type, int16_type, tmp[2], tmp[3]);
         dst[i] = lp_build_pack2(gallivm, int16_type, dst_type, lo, hi);
      }

      return; 
   }

   /* Special case 2x8f --> 1x16ub
    */
   else if (src_type.floating == 1 &&
      src_type.fixed    == 0 &&
      src_type.sign     == 1 &&
      src_type.norm     == 0 &&
      src_type.width    == 32 &&
      src_type.length   == 8 &&

      dst_type.floating == 0 &&
      dst_type.fixed    == 0 &&
      dst_type.sign     == 0 &&
      dst_type.norm     == 1 &&
      dst_type.width    == 8 &&
      dst_type.length   == 16 &&

      2 * num_dsts      == num_srcs &&

      util_cpu_caps.has_avx) {

      struct lp_build_context bld;
      struct lp_type int16_type = dst_type;
      struct lp_type int32_type = dst_type;
      LLVMValueRef const_255f;
      unsigned i;

      lp_build_context_init(&bld, gallivm, src_type);

      int16_type.width *= 2;
      int16_type.length /= 2;
      int16_type.sign = 1;

      int32_type.width *= 4;
      int32_type.length /= 4;
      int32_type.sign = 1;

      const_255f = lp_build_const_vec(gallivm, src_type, 255.0f);

      for (i = 0; i < num_dsts; ++i, src += 2) {
         LLVMValueRef lo, hi, a, b;

         a = LLVMBuildFMul(builder, src[0], const_255f, "");
         b = LLVMBuildFMul(builder, src[1], const_255f, "");

         a = lp_build_iround(&bld, a);
         b = lp_build_iround(&bld, b);

         tmp[0] = lp_build_extract_range(gallivm, a, 0, 4);
         tmp[1] = lp_build_extract_range(gallivm, a, 4, 4);
         tmp[2] = lp_build_extract_range(gallivm, b, 0, 4);
         tmp[3] = lp_build_extract_range(gallivm, b, 4, 4);

         /* relying on clamping behavior of sse2 intrinsics here */
         lo = lp_build_pack2(gallivm, int32_type, int16_type, tmp[0], tmp[1]);
         hi = lp_build_pack2(gallivm, int32_type, int16_type, tmp[2], tmp[3]);
         dst[i] = lp_build_pack2(gallivm, int16_type, dst_type, lo, hi);
      }
      return;
   }

   /* Special case -> 16bit half-float
    */
   else if (dst_type.floating && dst_type.width == 16)
   {
      /* Only support src as 32bit float currently */
      assert(src_type.floating && src_type.width == 32);

      for(i = 0; i < num_tmps; ++i)
         dst[i] = lp_build_float_to_half(gallivm, tmp[i]);

      return;
   }

   /* Pre convert half-floats to floats
    */
   else if (src_type.floating && src_type.width == 16)
   {
      for(i = 0; i < num_tmps; ++i)
         tmp[i] = lp_build_half_to_float(gallivm, tmp[i]);

      tmp_type.width = 32;
   }

   /*
    * Clamp if necessary
    */

   if(memcmp(&src_type, &dst_type, sizeof src_type) != 0) {
      struct lp_build_context bld;
      double src_min = lp_const_min(src_type);
      double dst_min = lp_const_min(dst_type);
      double src_max = lp_const_max(src_type);
      double dst_max = lp_const_max(dst_type);
      LLVMValueRef thres;

      lp_build_context_init(&bld, gallivm, tmp_type);

      if(src_min < dst_min) {
         if(dst_min == 0.0)
            thres = bld.zero;
         else
            thres = lp_build_const_vec(gallivm, src_type, dst_min);
         for(i = 0; i < num_tmps; ++i)
            tmp[i] = lp_build_max(&bld, tmp[i], thres);
      }

      if(src_max > dst_max) {
         if(dst_max == 1.0)
            thres = bld.one;
         else
            thres = lp_build_const_vec(gallivm, src_type, dst_max);
         for(i = 0; i < num_tmps; ++i)
            tmp[i] = lp_build_min(&bld, tmp[i], thres);
      }
   }

   /*
    * Scale to the narrowest range
    */

   if(dst_type.floating) {
      /* Nothing to do */
   }
   else if(tmp_type.floating) {
      if(!dst_type.fixed && !dst_type.sign && dst_type.norm) {
         for(i = 0; i < num_tmps; ++i) {
            tmp[i] = lp_build_clamped_float_to_unsigned_norm(gallivm,
                                                             tmp_type,
                                                             dst_type.width,
                                                             tmp[i]);
         }
         tmp_type.floating = FALSE;
      }
      else {
         double dst_scale = lp_const_scale(dst_type);
         LLVMTypeRef tmp_vec_type;

         if (dst_scale != 1.0) {
            LLVMValueRef scale = lp_build_const_vec(gallivm, tmp_type, dst_scale);
            for(i = 0; i < num_tmps; ++i)
               tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, "");
         }

         /* Use an equally sized integer for intermediate computations */
         tmp_type.floating = FALSE;
         tmp_vec_type = lp_build_vec_type(gallivm, tmp_type);
         for(i = 0; i < num_tmps; ++i) {
#if 0
            if(dst_type.sign)
               tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, "");
            else
               tmp[i] = LLVMBuildFPToUI(builder, tmp[i], tmp_vec_type, "");
#else
           /* FIXME: there is no SSE counterpart for LLVMBuildFPToUI */
            tmp[i] = LLVMBuildFPToSI(builder, tmp[i], tmp_vec_type, "");
#endif
         }
      }
   }
   else {
      unsigned src_shift = lp_const_shift(src_type);
      unsigned dst_shift = lp_const_shift(dst_type);
      unsigned src_offset = lp_const_offset(src_type);
      unsigned dst_offset = lp_const_offset(dst_type);

      /* Compensate for different offsets */
      if (dst_offset > src_offset && src_type.width > dst_type.width) {
         for (i = 0; i < num_tmps; ++i) {
            LLVMValueRef shifted;
            LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type, src_shift - 1);
            if(src_type.sign)
               shifted = LLVMBuildAShr(builder, tmp[i], shift, "");
            else
               shifted = LLVMBuildLShr(builder, tmp[i], shift, "");

            tmp[i] = LLVMBuildSub(builder, tmp[i], shifted, "");
         }
      }

      if(src_shift > dst_shift) {
         LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type,
                                                     src_shift - dst_shift);
         for(i = 0; i < num_tmps; ++i)
            if(src_type.sign)
               tmp[i] = LLVMBuildAShr(builder, tmp[i], shift, "");
            else
               tmp[i] = LLVMBuildLShr(builder, tmp[i], shift, "");
      }
   }

   /*
    * Truncate or expand bit width
    *
    * No data conversion should happen here, although the sign bits are
    * crucial to avoid bad clamping.
    */

   {
      struct lp_type new_type;

      new_type = tmp_type;
      new_type.sign   = dst_type.sign;
      new_type.width  = dst_type.width;
      new_type.length = dst_type.length;

      lp_build_resize(gallivm, tmp_type, new_type, tmp, num_srcs, tmp, num_dsts);

      tmp_type = new_type;
      num_tmps = num_dsts;
   }

   /*
    * Scale to the widest range
    */

   if(src_type.floating) {
      /* Nothing to do */
   }
   else if(!src_type.floating && dst_type.floating) {
      if(!src_type.fixed && !src_type.sign && src_type.norm) {
         for(i = 0; i < num_tmps; ++i) {
            tmp[i] = lp_build_unsigned_norm_to_float(gallivm,
                                                     src_type.width,
                                                     dst_type,
                                                     tmp[i]);
         }
         tmp_type.floating = TRUE;
      }
      else {
         double src_scale = lp_const_scale(src_type);
         LLVMTypeRef tmp_vec_type;

         /* Use an equally sized integer for intermediate computations */
         tmp_type.floating = TRUE;
         tmp_type.sign = TRUE;
         tmp_vec_type = lp_build_vec_type(gallivm, tmp_type);
         for(i = 0; i < num_tmps; ++i) {
#if 0
            if(dst_type.sign)
               tmp[i] = LLVMBuildSIToFP(builder, tmp[i], tmp_vec_type, "");
            else
               tmp[i] = LLVMBuildUIToFP(builder, tmp[i], tmp_vec_type, "");
#else
            /* FIXME: there is no SSE counterpart for LLVMBuildUIToFP */
            tmp[i] = LLVMBuildSIToFP(builder, tmp[i], tmp_vec_type, "");
#endif
          }

          if (src_scale != 1.0) {
             LLVMValueRef scale = lp_build_const_vec(gallivm, tmp_type, 1.0/src_scale);
             for(i = 0; i < num_tmps; ++i)
                tmp[i] = LLVMBuildFMul(builder, tmp[i], scale, "");
          }
      }
    }
    else {
       unsigned src_shift = lp_const_shift(src_type);
       unsigned dst_shift = lp_const_shift(dst_type);
       unsigned src_offset = lp_const_offset(src_type);
       unsigned dst_offset = lp_const_offset(dst_type);

       if (src_shift < dst_shift) {
          LLVMValueRef pre_shift[LP_MAX_VECTOR_LENGTH];
          LLVMValueRef shift = lp_build_const_int_vec(gallivm, tmp_type, dst_shift - src_shift);

          for (i = 0; i < num_tmps; ++i) {
             pre_shift[i] = tmp[i];
             tmp[i] = LLVMBuildShl(builder, tmp[i], shift, "");
          }

          /* Compensate for different offsets */
          if (dst_offset > src_offset) {
             for (i = 0; i < num_tmps; ++i) {
                tmp[i] = LLVMBuildSub(builder, tmp[i], pre_shift[i], "");
             }
          }
       }
    }

   for(i = 0; i < num_dsts; ++i) {
      dst[i] = tmp[i];
      assert(lp_check_value(dst_type, dst[i]));
   }
}


/**
 * Bit mask conversion.
 *
 * This will convert the integer masks that match the given types.
 *
 * The mask values should 0 or -1, i.e., all bits either set to zero or one.
 * Any other value will likely cause unpredictable results.
 *
 * This is basically a very trimmed down version of lp_build_conv.
 */
void
lp_build_conv_mask(struct gallivm_state *gallivm,
                   struct lp_type src_type,
                   struct lp_type dst_type,
                   const LLVMValueRef *src, unsigned num_srcs,
                   LLVMValueRef *dst, unsigned num_dsts)
{

   /* We must not loose or gain channels. Only precision */
   assert(src_type.length * num_srcs == dst_type.length * num_dsts);

   /*
    * Drop
    *
    * We assume all values are 0 or -1
    */

   src_type.floating = FALSE;
   src_type.fixed = FALSE;
   src_type.sign = TRUE;
   src_type.norm = FALSE;

   dst_type.floating = FALSE;
   dst_type.fixed = FALSE;
   dst_type.sign = TRUE;
   dst_type.norm = FALSE;

   /*
    * Truncate or expand bit width
    */

   lp_build_resize(gallivm, src_type, dst_type, src, num_srcs, dst, num_dsts);
}