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
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
|
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<title>User Guide for NVPTX Back-end — LLVM 13 documentation</title>
<link rel="stylesheet" href="_static/pygments.css" type="text/css" />
<link rel="stylesheet" href="_static/llvm-theme.css" type="text/css" />
<script id="documentation_options" data-url_root="./" src="_static/documentation_options.js"></script>
<script src="_static/jquery.js"></script>
<script src="_static/underscore.js"></script>
<script src="_static/doctools.js"></script>
<link rel="index" title="Index" href="genindex.html" />
<link rel="search" title="Search" href="search.html" />
<link rel="next" title="Code Reviews with Phabricator" href="Phabricator.html" />
<link rel="prev" title="Using the New Pass Manager" href="NewPassManager.html" />
<style type="text/css">
table.right { float: right; margin-left: 20px; }
table.right td { border: 1px solid #ccc; }
</style>
</head><body>
<div class="logo">
<a href="index.html">
<img src="_static/logo.png"
alt="LLVM Logo" width="250" height="88"/></a>
</div>
<div class="related" role="navigation" aria-label="related navigation">
<h3>Navigation</h3>
<ul>
<li class="right" style="margin-right: 10px">
<a href="genindex.html" title="General Index"
accesskey="I">index</a></li>
<li class="right" >
<a href="Phabricator.html" title="Code Reviews with Phabricator"
accesskey="N">next</a> |</li>
<li class="right" >
<a href="NewPassManager.html" title="Using the New Pass Manager"
accesskey="P">previous</a> |</li>
<li><a href="https://llvm.org/">LLVM Home</a> | </li>
<li><a href="index.html">Documentation</a>»</li>
<li class="nav-item nav-item-1"><a href="UserGuides.html" accesskey="U">User Guides</a> »</li>
<li class="nav-item nav-item-this"><a href="">User Guide for NVPTX Back-end</a></li>
</ul>
</div>
<div class="sphinxsidebar" role="navigation" aria-label="main navigation">
<div class="sphinxsidebarwrapper">
<h3>Documentation</h3>
<ul class="want-points">
<li><a href="https://llvm.org/docs/GettingStartedTutorials.html">Getting Started/Tutorials</a></li>
<li><a href="https://llvm.org/docs/UserGuides.html">User Guides</a></li>
<li><a href="https://llvm.org/docs/Reference.html">Reference</a></li>
</ul>
<h3>Getting Involved</h3>
<ul class="want-points">
<li><a href="https://llvm.org/docs/Contributing.html">Contributing to LLVM</a></li>
<li><a href="https://llvm.org/docs/HowToSubmitABug.html">Submitting Bug Reports</a></li>
<li><a href="https://llvm.org/docs/GettingInvolved.html#mailing-lists">Mailing Lists</a></li>
<li><a href="https://llvm.org/docs/GettingInvolved.html#irc">IRC</a></li>
<li><a href="https://llvm.org/docs/GettingInvolved.html#meetups-and-social-events">Meetups and Social Events</a></li>
</ul>
<h3>Additional Links</h3>
<ul class="want-points">
<li><a href="https://llvm.org/docs/FAQ.html">FAQ</a></li>
<li><a href="https://llvm.org/docs/Lexicon.html">Glossary</a></li>
<li><a href="https://llvm.org/pubs">Publications</a></li>
<li><a href="https://github.com/llvm/llvm-project//">Github Repository</a></li>
</ul>
<div role="note" aria-label="source link">
<h3>This Page</h3>
<ul class="this-page-menu">
<li><a href="_sources/NVPTXUsage.rst.txt"
rel="nofollow">Show Source</a></li>
</ul>
</div>
<div id="searchbox" style="display: none" role="search">
<h3 id="searchlabel">Quick search</h3>
<div class="searchformwrapper">
<form class="search" action="search.html" method="get">
<input type="text" name="q" aria-labelledby="searchlabel" />
<input type="submit" value="Go" />
</form>
</div>
</div>
<script>$('#searchbox').show(0);</script>
</div>
</div>
<div class="document">
<div class="documentwrapper">
<div class="bodywrapper">
<div class="body" role="main">
<div class="section" id="user-guide-for-nvptx-back-end">
<h1>User Guide for NVPTX Back-end<a class="headerlink" href="#user-guide-for-nvptx-back-end" title="Permalink to this headline">¶</a></h1>
<div class="contents local topic" id="contents">
<ul class="simple">
<li><p><a class="reference internal" href="#introduction" id="id11">Introduction</a></p></li>
<li><p><a class="reference internal" href="#conventions" id="id12">Conventions</a></p>
<ul>
<li><p><a class="reference internal" href="#marking-functions-as-kernels" id="id13">Marking Functions as Kernels</a></p></li>
<li><p><a class="reference internal" href="#address-spaces" id="id14">Address Spaces</a></p></li>
<li><p><a class="reference internal" href="#triples" id="id15">Triples</a></p></li>
</ul>
</li>
<li><p><a class="reference internal" href="#nvptx-intrinsics" id="id16">NVPTX Intrinsics</a></p>
<ul>
<li><p><a class="reference internal" href="#address-space-conversion" id="id17">Address Space Conversion</a></p>
<ul>
<li><p><a class="reference internal" href="#llvm-nvvm-ptr-to-gen-intrinsics" id="id18">‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.ptr.*.to.gen</span></code>’ Intrinsics</a></p></li>
<li><p><a class="reference internal" href="#llvm-nvvm-ptr-gen-to-intrinsics" id="id19">‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.ptr.gen.to.*</span></code>’ Intrinsics</a></p></li>
</ul>
</li>
<li><p><a class="reference internal" href="#reading-ptx-special-registers" id="id20">Reading PTX Special Registers</a></p>
<ul>
<li><p><a class="reference internal" href="#llvm-nvvm-read-ptx-sreg" id="id21">‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.read.ptx.sreg.*</span></code>’</a></p></li>
</ul>
</li>
<li><p><a class="reference internal" href="#barriers" id="id22">Barriers</a></p>
<ul>
<li><p><a class="reference internal" href="#llvm-nvvm-barrier0" id="id23">‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.barrier0</span></code>’</a></p></li>
</ul>
</li>
<li><p><a class="reference internal" href="#other-intrinsics" id="id24">Other Intrinsics</a></p></li>
</ul>
</li>
<li><p><a class="reference internal" href="#linking-with-libdevice" id="id25">Linking with Libdevice</a></p>
<ul>
<li><p><a class="reference internal" href="#reflection-parameters" id="id26">Reflection Parameters</a></p></li>
</ul>
</li>
<li><p><a class="reference internal" href="#executing-ptx" id="id27">Executing PTX</a></p></li>
<li><p><a class="reference internal" href="#common-issues" id="id28">Common Issues</a></p>
<ul>
<li><p><a class="reference internal" href="#ptxas-complains-of-undefined-function-nvvm-reflect" id="id29">ptxas complains of undefined function: __nvvm_reflect</a></p></li>
</ul>
</li>
<li><p><a class="reference internal" href="#tutorial-a-simple-compute-kernel" id="id30">Tutorial: A Simple Compute Kernel</a></p>
<ul>
<li><p><a class="reference internal" href="#the-kernel" id="id31">The Kernel</a></p></li>
<li><p><a class="reference internal" href="#dissecting-the-kernel" id="id32">Dissecting the Kernel</a></p>
<ul>
<li><p><a class="reference internal" href="#data-layout" id="id33">Data Layout</a></p></li>
<li><p><a class="reference internal" href="#target-intrinsics" id="id34">Target Intrinsics</a></p></li>
<li><p><a class="reference internal" href="#id10" id="id35">Address Spaces</a></p></li>
<li><p><a class="reference internal" href="#kernel-metadata" id="id36">Kernel Metadata</a></p></li>
</ul>
</li>
<li><p><a class="reference internal" href="#running-the-kernel" id="id37">Running the Kernel</a></p></li>
</ul>
</li>
<li><p><a class="reference internal" href="#tutorial-linking-with-libdevice" id="id38">Tutorial: Linking with Libdevice</a></p></li>
</ul>
</div>
<div class="section" id="introduction">
<h2><a class="toc-backref" href="#id11">Introduction</a><a class="headerlink" href="#introduction" title="Permalink to this headline">¶</a></h2>
<p>To support GPU programming, the NVPTX back-end supports a subset of LLVM IR
along with a defined set of conventions used to represent GPU programming
concepts. This document provides an overview of the general usage of the back-
end, including a description of the conventions used and the set of accepted
LLVM IR.</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>This document assumes a basic familiarity with CUDA and the PTX
assembly language. Information about the CUDA Driver API and the PTX assembly
language can be found in the <a class="reference external" href="http://docs.nvidia.com/cuda/index.html">CUDA documentation</a>.</p>
</div>
</div>
<div class="section" id="conventions">
<h2><a class="toc-backref" href="#id12">Conventions</a><a class="headerlink" href="#conventions" title="Permalink to this headline">¶</a></h2>
<div class="section" id="marking-functions-as-kernels">
<h3><a class="toc-backref" href="#id13">Marking Functions as Kernels</a><a class="headerlink" href="#marking-functions-as-kernels" title="Permalink to this headline">¶</a></h3>
<p>In PTX, there are two types of functions: <em>device functions</em>, which are only
callable by device code, and <em>kernel functions</em>, which are callable by host
code. By default, the back-end will emit device functions. Metadata is used to
declare a function as a kernel function. This metadata is attached to the
<code class="docutils literal notranslate"><span class="pre">nvvm.annotations</span></code> named metadata object, and has the following format:</p>
<div class="highlight-text notranslate"><div class="highlight"><pre><span></span>!0 = !{<function-ref>, metadata !"kernel", i32 1}
</pre></div>
</div>
<p>The first parameter is a reference to the kernel function. The following
example shows a kernel function calling a device function in LLVM IR. The
function <code class="docutils literal notranslate"><span class="pre">@my_kernel</span></code> is callable from host code, but <code class="docutils literal notranslate"><span class="pre">@my_fmad</span></code> is not.</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="k">define</span> <span class="k">float</span> <span class="vg">@my_fmad</span><span class="p">(</span><span class="k">float</span> <span class="nv">%x</span><span class="p">,</span> <span class="k">float</span> <span class="nv">%y</span><span class="p">,</span> <span class="k">float</span> <span class="nv">%z</span><span class="p">)</span> <span class="p">{</span>
<span class="nv">%mul</span> <span class="p">=</span> <span class="k">fmul</span> <span class="k">float</span> <span class="nv">%x</span><span class="p">,</span> <span class="nv">%y</span>
<span class="nv">%add</span> <span class="p">=</span> <span class="k">fadd</span> <span class="k">float</span> <span class="nv">%mul</span><span class="p">,</span> <span class="nv">%z</span>
<span class="k">ret</span> <span class="k">float</span> <span class="nv">%add</span>
<span class="p">}</span>
<span class="k">define</span> <span class="k">void</span> <span class="vg">@my_kernel</span><span class="p">(</span><span class="k">float</span><span class="p">*</span> <span class="nv">%ptr</span><span class="p">)</span> <span class="p">{</span>
<span class="nv">%val</span> <span class="p">=</span> <span class="k">load</span> <span class="k">float</span><span class="p">,</span> <span class="k">float</span><span class="p">*</span> <span class="nv">%ptr</span>
<span class="nv">%ret</span> <span class="p">=</span> <span class="k">call</span> <span class="k">float</span> <span class="vg">@my_fmad</span><span class="p">(</span><span class="k">float</span> <span class="nv">%val</span><span class="p">,</span> <span class="k">float</span> <span class="nv">%val</span><span class="p">,</span> <span class="k">float</span> <span class="nv">%val</span><span class="p">)</span>
<span class="k">store</span> <span class="k">float</span> <span class="nv">%ret</span><span class="p">,</span> <span class="k">float</span><span class="p">*</span> <span class="nv">%ptr</span>
<span class="k">ret</span> <span class="k">void</span>
<span class="p">}</span>
<span class="nv">!nvvm.annotations</span> <span class="p">=</span> <span class="p">!{</span><span class="nv nv-Anonymous">!1</span><span class="p">}</span>
<span class="nv nv-Anonymous">!1</span> <span class="p">=</span> <span class="p">!{</span><span class="k">void</span> <span class="p">(</span><span class="k">float</span><span class="p">*)*</span> <span class="vg">@my_kernel</span><span class="p">,</span> <span class="nv">!"kernel"</span><span class="p">,</span> <span class="k">i32</span> <span class="m">1</span><span class="p">}</span>
</pre></div>
</div>
<p>When compiled, the PTX kernel functions are callable by host-side code.</p>
</div>
<div class="section" id="address-spaces">
<span id="id1"></span><h3><a class="toc-backref" href="#id14">Address Spaces</a><a class="headerlink" href="#address-spaces" title="Permalink to this headline">¶</a></h3>
<p>The NVPTX back-end uses the following address space mapping:</p>
<blockquote>
<div><table class="docutils align-default">
<colgroup>
<col style="width: 37%" />
<col style="width: 63%" />
</colgroup>
<thead>
<tr class="row-odd"><th class="head"><p>Address Space</p></th>
<th class="head"><p>Memory Space</p></th>
</tr>
</thead>
<tbody>
<tr class="row-even"><td><p>0</p></td>
<td><p>Generic</p></td>
</tr>
<tr class="row-odd"><td><p>1</p></td>
<td><p>Global</p></td>
</tr>
<tr class="row-even"><td><p>2</p></td>
<td><p>Internal Use</p></td>
</tr>
<tr class="row-odd"><td><p>3</p></td>
<td><p>Shared</p></td>
</tr>
<tr class="row-even"><td><p>4</p></td>
<td><p>Constant</p></td>
</tr>
<tr class="row-odd"><td><p>5</p></td>
<td><p>Local</p></td>
</tr>
</tbody>
</table>
</div></blockquote>
<p>Every global variable and pointer type is assigned to one of these address
spaces, with 0 being the default address space. Intrinsics are provided which
can be used to convert pointers between the generic and non-generic address
spaces.</p>
<p>As an example, the following IR will define an array <code class="docutils literal notranslate"><span class="pre">@g</span></code> that resides in
global device memory.</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="vg">@g</span> <span class="p">=</span> <span class="k">internal</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)</span> <span class="k">global</span> <span class="p">[</span><span class="m">4</span> <span class="k">x</span> <span class="k">i32</span><span class="p">]</span> <span class="p">[</span> <span class="k">i32</span> <span class="m">0</span><span class="p">,</span> <span class="k">i32</span> <span class="m">1</span><span class="p">,</span> <span class="k">i32</span> <span class="m">2</span><span class="p">,</span> <span class="k">i32</span> <span class="m">3</span> <span class="p">]</span>
</pre></div>
</div>
<p>LLVM IR functions can read and write to this array, and host-side code can
copy data to it by name with the CUDA Driver API.</p>
<p>Note that since address space 0 is the generic space, it is illegal to have
global variables in address space 0. Address space 0 is the default address
space in LLVM, so the <code class="docutils literal notranslate"><span class="pre">addrspace(N)</span></code> annotation is <em>required</em> for global
variables.</p>
</div>
<div class="section" id="triples">
<h3><a class="toc-backref" href="#id15">Triples</a><a class="headerlink" href="#triples" title="Permalink to this headline">¶</a></h3>
<p>The NVPTX target uses the module triple to select between 32/64-bit code
generation and the driver-compiler interface to use. The triple architecture
can be one of <code class="docutils literal notranslate"><span class="pre">nvptx</span></code> (32-bit PTX) or <code class="docutils literal notranslate"><span class="pre">nvptx64</span></code> (64-bit PTX). The
operating system should be one of <code class="docutils literal notranslate"><span class="pre">cuda</span></code> or <code class="docutils literal notranslate"><span class="pre">nvcl</span></code>, which determines the
interface used by the generated code to communicate with the driver. Most
users will want to use <code class="docutils literal notranslate"><span class="pre">cuda</span></code> as the operating system, which makes the
generated PTX compatible with the CUDA Driver API.</p>
<p>Example: 32-bit PTX for CUDA Driver API: <code class="docutils literal notranslate"><span class="pre">nvptx-nvidia-cuda</span></code></p>
<p>Example: 64-bit PTX for CUDA Driver API: <code class="docutils literal notranslate"><span class="pre">nvptx64-nvidia-cuda</span></code></p>
</div>
</div>
<div class="section" id="nvptx-intrinsics">
<span id="id2"></span><h2><a class="toc-backref" href="#id16">NVPTX Intrinsics</a><a class="headerlink" href="#nvptx-intrinsics" title="Permalink to this headline">¶</a></h2>
<div class="section" id="address-space-conversion">
<h3><a class="toc-backref" href="#id17">Address Space Conversion</a><a class="headerlink" href="#address-space-conversion" title="Permalink to this headline">¶</a></h3>
<div class="section" id="llvm-nvvm-ptr-to-gen-intrinsics">
<h4><a class="toc-backref" href="#id18">‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.ptr.*.to.gen</span></code>’ Intrinsics</a><a class="headerlink" href="#llvm-nvvm-ptr-to-gen-intrinsics" title="Permalink to this headline">¶</a></h4>
<div class="section" id="syntax">
<h5>Syntax:<a class="headerlink" href="#syntax" title="Permalink to this headline">¶</a></h5>
<p>These are overloaded intrinsics. You can use these on any pointer types.</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="k">declare</span> <span class="k">i8</span><span class="p">*</span> <span class="vg">@llvm.nvvm.ptr.global.to.gen.p0i8.p1i8</span><span class="p">(</span><span class="k">i8</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*)</span>
<span class="k">declare</span> <span class="k">i8</span><span class="p">*</span> <span class="vg">@llvm.nvvm.ptr.shared.to.gen.p0i8.p3i8</span><span class="p">(</span><span class="k">i8</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">3</span><span class="p">)*)</span>
<span class="k">declare</span> <span class="k">i8</span><span class="p">*</span> <span class="vg">@llvm.nvvm.ptr.constant.to.gen.p0i8.p4i8</span><span class="p">(</span><span class="k">i8</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">4</span><span class="p">)*)</span>
<span class="k">declare</span> <span class="k">i8</span><span class="p">*</span> <span class="vg">@llvm.nvvm.ptr.local.to.gen.p0i8.p5i8</span><span class="p">(</span><span class="k">i8</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">5</span><span class="p">)*)</span>
</pre></div>
</div>
</div>
<div class="section" id="overview">
<h5>Overview:<a class="headerlink" href="#overview" title="Permalink to this headline">¶</a></h5>
<p>The ‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.ptr.*.to.gen</span></code>’ intrinsics convert a pointer in a non-generic
address space to a generic address space pointer.</p>
</div>
<div class="section" id="semantics">
<h5>Semantics:<a class="headerlink" href="#semantics" title="Permalink to this headline">¶</a></h5>
<p>These intrinsics modify the pointer value to be a valid generic address space
pointer.</p>
</div>
</div>
<div class="section" id="llvm-nvvm-ptr-gen-to-intrinsics">
<h4><a class="toc-backref" href="#id19">‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.ptr.gen.to.*</span></code>’ Intrinsics</a><a class="headerlink" href="#llvm-nvvm-ptr-gen-to-intrinsics" title="Permalink to this headline">¶</a></h4>
<div class="section" id="id3">
<h5>Syntax:<a class="headerlink" href="#id3" title="Permalink to this headline">¶</a></h5>
<p>These are overloaded intrinsics. You can use these on any pointer types.</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="k">declare</span> <span class="k">i8</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="vg">@llvm.nvvm.ptr.gen.to.global.p1i8.p0i8</span><span class="p">(</span><span class="k">i8</span><span class="p">*)</span>
<span class="k">declare</span> <span class="k">i8</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">3</span><span class="p">)*</span> <span class="vg">@llvm.nvvm.ptr.gen.to.shared.p3i8.p0i8</span><span class="p">(</span><span class="k">i8</span><span class="p">*)</span>
<span class="k">declare</span> <span class="k">i8</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">4</span><span class="p">)*</span> <span class="vg">@llvm.nvvm.ptr.gen.to.constant.p4i8.p0i8</span><span class="p">(</span><span class="k">i8</span><span class="p">*)</span>
<span class="k">declare</span> <span class="k">i8</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">5</span><span class="p">)*</span> <span class="vg">@llvm.nvvm.ptr.gen.to.local.p5i8.p0i8</span><span class="p">(</span><span class="k">i8</span><span class="p">*)</span>
</pre></div>
</div>
</div>
<div class="section" id="id4">
<h5>Overview:<a class="headerlink" href="#id4" title="Permalink to this headline">¶</a></h5>
<p>The ‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.ptr.gen.to.*</span></code>’ intrinsics convert a pointer in the generic
address space to a pointer in the target address space. Note that these
intrinsics are only useful if the address space of the target address space of
the pointer is known. It is not legal to use address space conversion
intrinsics to convert a pointer from one non-generic address space to another
non-generic address space.</p>
</div>
<div class="section" id="id5">
<h5>Semantics:<a class="headerlink" href="#id5" title="Permalink to this headline">¶</a></h5>
<p>These intrinsics modify the pointer value to be a valid pointer in the target
non-generic address space.</p>
</div>
</div>
</div>
<div class="section" id="reading-ptx-special-registers">
<h3><a class="toc-backref" href="#id20">Reading PTX Special Registers</a><a class="headerlink" href="#reading-ptx-special-registers" title="Permalink to this headline">¶</a></h3>
<div class="section" id="llvm-nvvm-read-ptx-sreg">
<h4><a class="toc-backref" href="#id21">‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.read.ptx.sreg.*</span></code>’</a><a class="headerlink" href="#llvm-nvvm-read-ptx-sreg" title="Permalink to this headline">¶</a></h4>
<div class="section" id="id6">
<h5>Syntax:<a class="headerlink" href="#id6" title="Permalink to this headline">¶</a></h5>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.tid.x</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.tid.y</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.tid.z</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.ntid.x</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.ntid.y</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.ntid.z</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.ctaid.x</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.ctaid.y</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.ctaid.z</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.nctaid.x</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.nctaid.y</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.nctaid.z</span><span class="p">()</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.warpsize</span><span class="p">()</span>
</pre></div>
</div>
</div>
<div class="section" id="id7">
<h5>Overview:<a class="headerlink" href="#id7" title="Permalink to this headline">¶</a></h5>
<p>The ‘<code class="docutils literal notranslate"><span class="pre">@llvm.nvvm.read.ptx.sreg.*</span></code>’ intrinsics provide access to the PTX
special registers, in particular the kernel launch bounds. These registers
map in the following way to CUDA builtins:</p>
<blockquote>
<div><table class="docutils align-default">
<colgroup>
<col style="width: 24%" />
<col style="width: 76%" />
</colgroup>
<thead>
<tr class="row-odd"><th class="head"><p>CUDA Builtin</p></th>
<th class="head"><p>PTX Special Register Intrinsic</p></th>
</tr>
</thead>
<tbody>
<tr class="row-even"><td><p><code class="docutils literal notranslate"><span class="pre">threadId</span></code></p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">@llvm.nvvm.read.ptx.sreg.tid.*</span></code></p></td>
</tr>
<tr class="row-odd"><td><p><code class="docutils literal notranslate"><span class="pre">blockIdx</span></code></p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">@llvm.nvvm.read.ptx.sreg.ctaid.*</span></code></p></td>
</tr>
<tr class="row-even"><td><p><code class="docutils literal notranslate"><span class="pre">blockDim</span></code></p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">@llvm.nvvm.read.ptx.sreg.ntid.*</span></code></p></td>
</tr>
<tr class="row-odd"><td><p><code class="docutils literal notranslate"><span class="pre">gridDim</span></code></p></td>
<td><p><code class="docutils literal notranslate"><span class="pre">@llvm.nvvm.read.ptx.sreg.nctaid.*</span></code></p></td>
</tr>
</tbody>
</table>
</div></blockquote>
</div>
</div>
</div>
<div class="section" id="barriers">
<h3><a class="toc-backref" href="#id22">Barriers</a><a class="headerlink" href="#barriers" title="Permalink to this headline">¶</a></h3>
<div class="section" id="llvm-nvvm-barrier0">
<h4><a class="toc-backref" href="#id23">‘<code class="docutils literal notranslate"><span class="pre">llvm.nvvm.barrier0</span></code>’</a><a class="headerlink" href="#llvm-nvvm-barrier0" title="Permalink to this headline">¶</a></h4>
<div class="section" id="id8">
<h5>Syntax:<a class="headerlink" href="#id8" title="Permalink to this headline">¶</a></h5>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="k">declare</span> <span class="k">void</span> <span class="vg">@llvm.nvvm.barrier0</span><span class="p">()</span>
</pre></div>
</div>
</div>
<div class="section" id="id9">
<h5>Overview:<a class="headerlink" href="#id9" title="Permalink to this headline">¶</a></h5>
<p>The ‘<code class="docutils literal notranslate"><span class="pre">@llvm.nvvm.barrier0()</span></code>’ intrinsic emits a PTX <code class="docutils literal notranslate"><span class="pre">bar.sync</span> <span class="pre">0</span></code>
instruction, equivalent to the <code class="docutils literal notranslate"><span class="pre">__syncthreads()</span></code> call in CUDA.</p>
</div>
</div>
</div>
<div class="section" id="other-intrinsics">
<h3><a class="toc-backref" href="#id24">Other Intrinsics</a><a class="headerlink" href="#other-intrinsics" title="Permalink to this headline">¶</a></h3>
<p>For the full set of NVPTX intrinsics, please see the
<code class="docutils literal notranslate"><span class="pre">include/llvm/IR/IntrinsicsNVVM.td</span></code> file in the LLVM source tree.</p>
</div>
</div>
<div class="section" id="linking-with-libdevice">
<span id="libdevice"></span><h2><a class="toc-backref" href="#id25">Linking with Libdevice</a><a class="headerlink" href="#linking-with-libdevice" title="Permalink to this headline">¶</a></h2>
<p>The CUDA Toolkit comes with an LLVM bitcode library called <code class="docutils literal notranslate"><span class="pre">libdevice</span></code> that
implements many common mathematical functions. This library can be used as a
high-performance math library for any compilers using the LLVM NVPTX target.
The library can be found under <code class="docutils literal notranslate"><span class="pre">nvvm/libdevice/</span></code> in the CUDA Toolkit and
there is a separate version for each compute architecture.</p>
<p>For a list of all math functions implemented in libdevice, see
<a class="reference external" href="http://docs.nvidia.com/cuda/libdevice-users-guide/index.html">libdevice Users Guide</a>.</p>
<p>To accommodate various math-related compiler flags that can affect code
generation of libdevice code, the library code depends on a special LLVM IR
pass (<code class="docutils literal notranslate"><span class="pre">NVVMReflect</span></code>) to handle conditional compilation within LLVM IR. This
pass looks for calls to the <code class="docutils literal notranslate"><span class="pre">@__nvvm_reflect</span></code> function and replaces them
with constants based on the defined reflection parameters. Such conditional
code often follows a pattern:</p>
<div class="highlight-c++ notranslate"><div class="highlight"><pre><span></span><span class="kt">float</span> <span class="nf">my_function</span><span class="p">(</span><span class="kt">float</span> <span class="n">a</span><span class="p">)</span> <span class="p">{</span>
<span class="k">if</span> <span class="p">(</span><span class="n">__nvvm_reflect</span><span class="p">(</span><span class="s">"FASTMATH"</span><span class="p">))</span>
<span class="k">return</span> <span class="n">my_function_fast</span><span class="p">(</span><span class="n">a</span><span class="p">);</span>
<span class="k">else</span>
<span class="k">return</span> <span class="n">my_function_precise</span><span class="p">(</span><span class="n">a</span><span class="p">);</span>
<span class="p">}</span>
</pre></div>
</div>
<p>The default value for all unspecified reflection parameters is zero.</p>
<p>The <code class="docutils literal notranslate"><span class="pre">NVVMReflect</span></code> pass should be executed early in the optimization
pipeline, immediately after the link stage. The <code class="docutils literal notranslate"><span class="pre">internalize</span></code> pass is also
recommended to remove unused math functions from the resulting PTX. For an
input IR module <code class="docutils literal notranslate"><span class="pre">module.bc</span></code>, the following compilation flow is recommended:</p>
<ol class="arabic simple">
<li><p>Save list of external functions in <code class="docutils literal notranslate"><span class="pre">module.bc</span></code></p></li>
<li><p>Link <code class="docutils literal notranslate"><span class="pre">module.bc</span></code> with <code class="docutils literal notranslate"><span class="pre">libdevice.compute_XX.YY.bc</span></code></p></li>
<li><p>Internalize all functions not in list from (1)</p></li>
<li><p>Eliminate all unused internal functions</p></li>
<li><p>Run <code class="docutils literal notranslate"><span class="pre">NVVMReflect</span></code> pass</p></li>
<li><p>Run standard optimization pipeline</p></li>
</ol>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p><code class="docutils literal notranslate"><span class="pre">linkonce</span></code> and <code class="docutils literal notranslate"><span class="pre">linkonce_odr</span></code> linkage types are not suitable for the
libdevice functions. It is possible to link two IR modules that have been
linked against libdevice using different reflection variables.</p>
</div>
<p>Since the <code class="docutils literal notranslate"><span class="pre">NVVMReflect</span></code> pass replaces conditionals with constants, it will
often leave behind dead code of the form:</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="nl">entry:</span>
<span class="p">..</span>
<span class="k">br</span> <span class="k">i1</span> <span class="k">true</span><span class="p">,</span> <span class="k">label</span> <span class="nv">%foo</span><span class="p">,</span> <span class="k">label</span> <span class="nv">%bar</span>
<span class="nl">foo:</span>
<span class="p">..</span>
<span class="nl">bar:</span>
<span class="c">; Dead code</span>
<span class="p">..</span>
</pre></div>
</div>
<p>Therefore, it is recommended that <code class="docutils literal notranslate"><span class="pre">NVVMReflect</span></code> is executed early in the
optimization pipeline before dead-code elimination.</p>
<p>The NVPTX TargetMachine knows how to schedule <code class="docutils literal notranslate"><span class="pre">NVVMReflect</span></code> at the beginning
of your pass manager; just use the following code when setting up your pass
manager:</p>
<div class="highlight-c++ notranslate"><div class="highlight"><pre><span></span><span class="n">std</span><span class="o">::</span><span class="n">unique_ptr</span><span class="o"><</span><span class="n">TargetMachine</span><span class="o">></span> <span class="n">TM</span> <span class="o">=</span> <span class="p">...;</span>
<span class="n">PassManagerBuilder</span> <span class="nf">PMBuilder</span><span class="p">(...);</span>
<span class="k">if</span> <span class="p">(</span><span class="n">TM</span><span class="p">)</span>
<span class="n">TM</span><span class="o">-></span><span class="n">adjustPassManager</span><span class="p">(</span><span class="n">PMBuilder</span><span class="p">);</span>
</pre></div>
</div>
<div class="section" id="reflection-parameters">
<h3><a class="toc-backref" href="#id26">Reflection Parameters</a><a class="headerlink" href="#reflection-parameters" title="Permalink to this headline">¶</a></h3>
<p>The libdevice library currently uses the following reflection parameters to
control code generation:</p>
<table class="docutils align-default">
<colgroup>
<col style="width: 27%" />
<col style="width: 73%" />
</colgroup>
<thead>
<tr class="row-odd"><th class="head"><p>Flag</p></th>
<th class="head"><p>Description</p></th>
</tr>
</thead>
<tbody>
<tr class="row-even"><td><p><code class="docutils literal notranslate"><span class="pre">__CUDA_FTZ=[0,1]</span></code></p></td>
<td><p>Use optimized code paths that flush subnormals to zero</p></td>
</tr>
</tbody>
</table>
<p>The value of this flag is determined by the “nvvm-reflect-ftz” module flag.
The following sets the ftz flag to 1.</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="nv">!llvm.module.flag</span> <span class="p">=</span> <span class="p">!{</span><span class="nv nv-Anonymous">!0</span><span class="p">}</span>
<span class="nv nv-Anonymous">!0</span> <span class="p">=</span> <span class="p">!{</span><span class="k">i32</span> <span class="m">4</span><span class="p">,</span> <span class="nv">!"nvvm-reflect-ftz"</span><span class="p">,</span> <span class="k">i32</span> <span class="m">1</span><span class="p">}</span>
</pre></div>
</div>
<p>(<code class="docutils literal notranslate"><span class="pre">i32</span> <span class="pre">4</span></code> indicates that the value set here overrides the value in another
module we link with. See the <cite>LangRef <LangRef.html#module-flags-metadata></cite>
for details.)</p>
</div>
</div>
<div class="section" id="executing-ptx">
<h2><a class="toc-backref" href="#id27">Executing PTX</a><a class="headerlink" href="#executing-ptx" title="Permalink to this headline">¶</a></h2>
<p>The most common way to execute PTX assembly on a GPU device is to use the CUDA
Driver API. This API is a low-level interface to the GPU driver and allows for
JIT compilation of PTX code to native GPU machine code.</p>
<p>Initializing the Driver API:</p>
<div class="highlight-c++ notranslate"><div class="highlight"><pre><span></span><span class="n">CUdevice</span> <span class="n">device</span><span class="p">;</span>
<span class="n">CUcontext</span> <span class="n">context</span><span class="p">;</span>
<span class="c1">// Initialize the driver API</span>
<span class="n">cuInit</span><span class="p">(</span><span class="mi">0</span><span class="p">);</span>
<span class="c1">// Get a handle to the first compute device</span>
<span class="n">cuDeviceGet</span><span class="p">(</span><span class="o">&</span><span class="n">device</span><span class="p">,</span> <span class="mi">0</span><span class="p">);</span>
<span class="c1">// Create a compute device context</span>
<span class="n">cuCtxCreate</span><span class="p">(</span><span class="o">&</span><span class="n">context</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="n">device</span><span class="p">);</span>
</pre></div>
</div>
<p>JIT compiling a PTX string to a device binary:</p>
<div class="highlight-c++ notranslate"><div class="highlight"><pre><span></span><span class="n">CUmodule</span> <span class="k">module</span><span class="p">;</span>
<span class="n">CUfunction</span> <span class="n">function</span><span class="p">;</span>
<span class="c1">// JIT compile a null-terminated PTX string</span>
<span class="n">cuModuleLoadData</span><span class="p">(</span><span class="o">&</span><span class="k">module</span><span class="p">,</span> <span class="p">(</span><span class="kt">void</span><span class="o">*</span><span class="p">)</span><span class="n">PTXString</span><span class="p">);</span>
<span class="c1">// Get a handle to the "myfunction" kernel function</span>
<span class="n">cuModuleGetFunction</span><span class="p">(</span><span class="o">&</span><span class="n">function</span><span class="p">,</span> <span class="k">module</span><span class="p">,</span> <span class="s">"myfunction"</span><span class="p">);</span>
</pre></div>
</div>
<p>For full examples of executing PTX assembly, please see the <a class="reference external" href="https://developer.nvidia.com/cuda-downloads">CUDA Samples</a> distribution.</p>
</div>
<div class="section" id="common-issues">
<h2><a class="toc-backref" href="#id28">Common Issues</a><a class="headerlink" href="#common-issues" title="Permalink to this headline">¶</a></h2>
<div class="section" id="ptxas-complains-of-undefined-function-nvvm-reflect">
<h3><a class="toc-backref" href="#id29">ptxas complains of undefined function: __nvvm_reflect</a><a class="headerlink" href="#ptxas-complains-of-undefined-function-nvvm-reflect" title="Permalink to this headline">¶</a></h3>
<p>When linking with libdevice, the <code class="docutils literal notranslate"><span class="pre">NVVMReflect</span></code> pass must be used. See
<a class="reference internal" href="#libdevice"><span class="std std-ref">Linking with Libdevice</span></a> for more information.</p>
</div>
</div>
<div class="section" id="tutorial-a-simple-compute-kernel">
<h2><a class="toc-backref" href="#id30">Tutorial: A Simple Compute Kernel</a><a class="headerlink" href="#tutorial-a-simple-compute-kernel" title="Permalink to this headline">¶</a></h2>
<p>To start, let us take a look at a simple compute kernel written directly in
LLVM IR. The kernel implements vector addition, where each thread computes one
element of the output vector C from the input vectors A and B. To make this
easier, we also assume that only a single CTA (thread block) will be launched,
and that it will be one dimensional.</p>
<div class="section" id="the-kernel">
<h3><a class="toc-backref" href="#id31">The Kernel</a><a class="headerlink" href="#the-kernel" title="Permalink to this headline">¶</a></h3>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="k">target</span> <span class="k">datalayout</span> <span class="p">=</span> <span class="s">"e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"</span>
<span class="k">target</span> <span class="k">triple</span> <span class="p">=</span> <span class="s">"nvptx64-nvidia-cuda"</span>
<span class="c">; Intrinsic to read X component of thread ID</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.tid.x</span><span class="p">()</span> <span class="k">readnone</span> <span class="k">nounwind</span>
<span class="k">define</span> <span class="k">void</span> <span class="vg">@kernel</span><span class="p">(</span><span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%A</span><span class="p">,</span>
<span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%B</span><span class="p">,</span>
<span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%C</span><span class="p">)</span> <span class="p">{</span>
<span class="nl">entry:</span>
<span class="c">; What is my ID?</span>
<span class="nv">%id</span> <span class="p">=</span> <span class="k">tail</span> <span class="k">call</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.tid.x</span><span class="p">()</span> <span class="k">readnone</span> <span class="k">nounwind</span>
<span class="c">; Compute pointers into A, B, and C</span>
<span class="nv">%ptrA</span> <span class="p">=</span> <span class="k">getelementptr</span> <span class="k">float</span><span class="p">,</span> <span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%A</span><span class="p">,</span> <span class="k">i32</span> <span class="nv">%id</span>
<span class="nv">%ptrB</span> <span class="p">=</span> <span class="k">getelementptr</span> <span class="k">float</span><span class="p">,</span> <span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%B</span><span class="p">,</span> <span class="k">i32</span> <span class="nv">%id</span>
<span class="nv">%ptrC</span> <span class="p">=</span> <span class="k">getelementptr</span> <span class="k">float</span><span class="p">,</span> <span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%C</span><span class="p">,</span> <span class="k">i32</span> <span class="nv">%id</span>
<span class="c">; Read A, B</span>
<span class="nv">%valA</span> <span class="p">=</span> <span class="k">load</span> <span class="k">float</span><span class="p">,</span> <span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%ptrA</span><span class="p">,</span> <span class="k">align</span> <span class="m">4</span>
<span class="nv">%valB</span> <span class="p">=</span> <span class="k">load</span> <span class="k">float</span><span class="p">,</span> <span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%ptrB</span><span class="p">,</span> <span class="k">align</span> <span class="m">4</span>
<span class="c">; Compute C = A + B</span>
<span class="nv">%valC</span> <span class="p">=</span> <span class="k">fadd</span> <span class="k">float</span> <span class="nv">%valA</span><span class="p">,</span> <span class="nv">%valB</span>
<span class="c">; Store back to C</span>
<span class="k">store</span> <span class="k">float</span> <span class="nv">%valC</span><span class="p">,</span> <span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%ptrC</span><span class="p">,</span> <span class="k">align</span> <span class="m">4</span>
<span class="k">ret</span> <span class="k">void</span>
<span class="p">}</span>
<span class="nv">!nvvm.annotations</span> <span class="p">=</span> <span class="p">!{</span><span class="nv nv-Anonymous">!0</span><span class="p">}</span>
<span class="nv nv-Anonymous">!0</span> <span class="p">=</span> <span class="p">!{</span><span class="k">void</span> <span class="p">(</span><span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*,</span>
<span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*,</span>
<span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*)*</span> <span class="vg">@kernel</span><span class="p">,</span> <span class="nv">!"kernel"</span><span class="p">,</span> <span class="k">i32</span> <span class="m">1</span><span class="p">}</span>
</pre></div>
</div>
<p>We can use the LLVM <code class="docutils literal notranslate"><span class="pre">llc</span></code> tool to directly run the NVPTX code generator:</p>
<div class="highlight-text notranslate"><div class="highlight"><pre><span></span># llc -mcpu=sm_20 kernel.ll -o kernel.ptx
</pre></div>
</div>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>If you want to generate 32-bit code, change <code class="docutils literal notranslate"><span class="pre">p:64:64:64</span></code> to <code class="docutils literal notranslate"><span class="pre">p:32:32:32</span></code>
in the module data layout string and use <code class="docutils literal notranslate"><span class="pre">nvptx-nvidia-cuda</span></code> as the
target triple.</p>
</div>
<p>The output we get from <code class="docutils literal notranslate"><span class="pre">llc</span></code> (as of LLVM 3.4):</p>
<div class="highlight-text notranslate"><div class="highlight"><pre><span></span>//
// Generated by LLVM NVPTX Back-End
//
.version 3.1
.target sm_20
.address_size 64
// .globl kernel
// @kernel
.visible .entry kernel(
.param .u64 kernel_param_0,
.param .u64 kernel_param_1,
.param .u64 kernel_param_2
)
{
.reg .f32 %f<4>;
.reg .s32 %r<2>;
.reg .s64 %rl<8>;
// %bb.0: // %entry
ld.param.u64 %rl1, [kernel_param_0];
mov.u32 %r1, %tid.x;
mul.wide.s32 %rl2, %r1, 4;
add.s64 %rl3, %rl1, %rl2;
ld.param.u64 %rl4, [kernel_param_1];
add.s64 %rl5, %rl4, %rl2;
ld.param.u64 %rl6, [kernel_param_2];
add.s64 %rl7, %rl6, %rl2;
ld.global.f32 %f1, [%rl3];
ld.global.f32 %f2, [%rl5];
add.f32 %f3, %f1, %f2;
st.global.f32 [%rl7], %f3;
ret;
}
</pre></div>
</div>
</div>
<div class="section" id="dissecting-the-kernel">
<h3><a class="toc-backref" href="#id32">Dissecting the Kernel</a><a class="headerlink" href="#dissecting-the-kernel" title="Permalink to this headline">¶</a></h3>
<p>Now let us dissect the LLVM IR that makes up this kernel.</p>
<div class="section" id="data-layout">
<h4><a class="toc-backref" href="#id33">Data Layout</a><a class="headerlink" href="#data-layout" title="Permalink to this headline">¶</a></h4>
<p>The data layout string determines the size in bits of common data types, their
ABI alignment, and their storage size. For NVPTX, you should use one of the
following:</p>
<p>32-bit PTX:</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="k">target</span> <span class="k">datalayout</span> <span class="p">=</span> <span class="s">"e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"</span>
</pre></div>
</div>
<p>64-bit PTX:</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="k">target</span> <span class="k">datalayout</span> <span class="p">=</span> <span class="s">"e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"</span>
</pre></div>
</div>
</div>
<div class="section" id="target-intrinsics">
<h4><a class="toc-backref" href="#id34">Target Intrinsics</a><a class="headerlink" href="#target-intrinsics" title="Permalink to this headline">¶</a></h4>
<p>In this example, we use the <code class="docutils literal notranslate"><span class="pre">@llvm.nvvm.read.ptx.sreg.tid.x</span></code> intrinsic to
read the X component of the current thread’s ID, which corresponds to a read
of register <code class="docutils literal notranslate"><span class="pre">%tid.x</span></code> in PTX. The NVPTX back-end supports a large set of
intrinsics. A short list is shown below; please see
<code class="docutils literal notranslate"><span class="pre">include/llvm/IR/IntrinsicsNVVM.td</span></code> for the full list.</p>
<table class="docutils align-default">
<colgroup>
<col style="width: 71%" />
<col style="width: 29%" />
</colgroup>
<thead>
<tr class="row-odd"><th class="head"><p>Intrinsic</p></th>
<th class="head"><p>CUDA Equivalent</p></th>
</tr>
</thead>
<tbody>
<tr class="row-even"><td><p><code class="docutils literal notranslate"><span class="pre">i32</span> <span class="pre">@llvm.nvvm.read.ptx.sreg.tid.{x,y,z}</span></code></p></td>
<td><p>threadIdx.{x,y,z}</p></td>
</tr>
<tr class="row-odd"><td><p><code class="docutils literal notranslate"><span class="pre">i32</span> <span class="pre">@llvm.nvvm.read.ptx.sreg.ctaid.{x,y,z}</span></code></p></td>
<td><p>blockIdx.{x,y,z}</p></td>
</tr>
<tr class="row-even"><td><p><code class="docutils literal notranslate"><span class="pre">i32</span> <span class="pre">@llvm.nvvm.read.ptx.sreg.ntid.{x,y,z}</span></code></p></td>
<td><p>blockDim.{x,y,z}</p></td>
</tr>
<tr class="row-odd"><td><p><code class="docutils literal notranslate"><span class="pre">i32</span> <span class="pre">@llvm.nvvm.read.ptx.sreg.nctaid.{x,y,z}</span></code></p></td>
<td><p>gridDim.{x,y,z}</p></td>
</tr>
<tr class="row-even"><td><p><code class="docutils literal notranslate"><span class="pre">void</span> <span class="pre">@llvm.nvvm.barrier0()</span></code></p></td>
<td><p>__syncthreads()</p></td>
</tr>
</tbody>
</table>
</div>
<div class="section" id="id10">
<h4><a class="toc-backref" href="#id35">Address Spaces</a><a class="headerlink" href="#id10" title="Permalink to this headline">¶</a></h4>
<p>You may have noticed that all of the pointer types in the LLVM IR example had
an explicit address space specifier. What is address space 1? NVIDIA GPU
devices (generally) have four types of memory:</p>
<ul class="simple">
<li><p>Global: Large, off-chip memory</p></li>
<li><p>Shared: Small, on-chip memory shared among all threads in a CTA</p></li>
<li><p>Local: Per-thread, private memory</p></li>
<li><p>Constant: Read-only memory shared across all threads</p></li>
</ul>
<p>These different types of memory are represented in LLVM IR as address spaces.
There is also a fifth address space used by the NVPTX code generator that
corresponds to the “generic” address space. This address space can represent
addresses in any other address space (with a few exceptions). This allows
users to write IR functions that can load/store memory using the same
instructions. Intrinsics are provided to convert pointers between the generic
and non-generic address spaces.</p>
<p>See <a class="reference internal" href="#address-spaces"><span class="std std-ref">Address Spaces</span></a> and <a class="reference internal" href="#nvptx-intrinsics"><span class="std std-ref">NVPTX Intrinsics</span></a> for more information.</p>
</div>
<div class="section" id="kernel-metadata">
<h4><a class="toc-backref" href="#id36">Kernel Metadata</a><a class="headerlink" href="#kernel-metadata" title="Permalink to this headline">¶</a></h4>
<p>In PTX, a function can be either a <cite>kernel</cite> function (callable from the host
program), or a <cite>device</cite> function (callable only from GPU code). You can think
of <cite>kernel</cite> functions as entry-points in the GPU program. To mark an LLVM IR
function as a <cite>kernel</cite> function, we make use of special LLVM metadata. The
NVPTX back-end will look for a named metadata node called
<code class="docutils literal notranslate"><span class="pre">nvvm.annotations</span></code>. This named metadata must contain a list of metadata that
describe the IR. For our purposes, we need to declare a metadata node that
assigns the “kernel” attribute to the LLVM IR function that should be emitted
as a PTX <cite>kernel</cite> function. These metadata nodes take the form:</p>
<div class="highlight-text notranslate"><div class="highlight"><pre><span></span>!{<function ref>, metadata !"kernel", i32 1}
</pre></div>
</div>
<p>For the previous example, we have:</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="nv">!nvvm.annotations</span> <span class="p">=</span> <span class="p">!{</span><span class="nv nv-Anonymous">!0</span><span class="p">}</span>
<span class="nv nv-Anonymous">!0</span> <span class="p">=</span> <span class="p">!{</span><span class="k">void</span> <span class="p">(</span><span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*,</span>
<span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*,</span>
<span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*)*</span> <span class="vg">@kernel</span><span class="p">,</span> <span class="nv">!"kernel"</span><span class="p">,</span> <span class="k">i32</span> <span class="m">1</span><span class="p">}</span>
</pre></div>
</div>
<p>Here, we have a single metadata declaration in <code class="docutils literal notranslate"><span class="pre">nvvm.annotations</span></code>. This
metadata annotates our <code class="docutils literal notranslate"><span class="pre">@kernel</span></code> function with the <code class="docutils literal notranslate"><span class="pre">kernel</span></code> attribute.</p>
</div>
</div>
<div class="section" id="running-the-kernel">
<h3><a class="toc-backref" href="#id37">Running the Kernel</a><a class="headerlink" href="#running-the-kernel" title="Permalink to this headline">¶</a></h3>
<p>Generating PTX from LLVM IR is all well and good, but how do we execute it on
a real GPU device? The CUDA Driver API provides a convenient mechanism for
loading and JIT compiling PTX to a native GPU device, and launching a kernel.
The API is similar to OpenCL. A simple example showing how to load and
execute our vector addition code is shown below. Note that for brevity this
code does not perform much error checking!</p>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>You can also use the <code class="docutils literal notranslate"><span class="pre">ptxas</span></code> tool provided by the CUDA Toolkit to offline
compile PTX to machine code (SASS) for a specific GPU architecture. Such
binaries can be loaded by the CUDA Driver API in the same way as PTX. This
can be useful for reducing startup time by precompiling the PTX kernels.</p>
</div>
<div class="highlight-c++ notranslate"><div class="highlight"><pre><span></span><span class="cp">#include</span> <span class="cpf"><iostream></span><span class="cp"></span>
<span class="cp">#include</span> <span class="cpf"><fstream></span><span class="cp"></span>
<span class="cp">#include</span> <span class="cpf"><cassert></span><span class="cp"></span>
<span class="cp">#include</span> <span class="cpf">"cuda.h"</span><span class="cp"></span>
<span class="kt">void</span> <span class="nf">checkCudaErrors</span><span class="p">(</span><span class="n">CUresult</span> <span class="n">err</span><span class="p">)</span> <span class="p">{</span>
<span class="n">assert</span><span class="p">(</span><span class="n">err</span> <span class="o">==</span> <span class="n">CUDA_SUCCESS</span><span class="p">);</span>
<span class="p">}</span>
<span class="c1">/// main - Program entry point</span>
<span class="kt">int</span> <span class="nf">main</span><span class="p">(</span><span class="kt">int</span> <span class="n">argc</span><span class="p">,</span> <span class="kt">char</span> <span class="o">**</span><span class="n">argv</span><span class="p">)</span> <span class="p">{</span>
<span class="n">CUdevice</span> <span class="n">device</span><span class="p">;</span>
<span class="n">CUmodule</span> <span class="n">cudaModule</span><span class="p">;</span>
<span class="n">CUcontext</span> <span class="n">context</span><span class="p">;</span>
<span class="n">CUfunction</span> <span class="n">function</span><span class="p">;</span>
<span class="n">CUlinkState</span> <span class="n">linker</span><span class="p">;</span>
<span class="kt">int</span> <span class="n">devCount</span><span class="p">;</span>
<span class="c1">// CUDA initialization</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuInit</span><span class="p">(</span><span class="mi">0</span><span class="p">));</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuDeviceGetCount</span><span class="p">(</span><span class="o">&</span><span class="n">devCount</span><span class="p">));</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuDeviceGet</span><span class="p">(</span><span class="o">&</span><span class="n">device</span><span class="p">,</span> <span class="mi">0</span><span class="p">));</span>
<span class="kt">char</span> <span class="n">name</span><span class="p">[</span><span class="mi">128</span><span class="p">];</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuDeviceGetName</span><span class="p">(</span><span class="n">name</span><span class="p">,</span> <span class="mi">128</span><span class="p">,</span> <span class="n">device</span><span class="p">));</span>
<span class="n">std</span><span class="o">::</span><span class="n">cout</span> <span class="o"><<</span> <span class="s">"Using CUDA Device [0]: "</span> <span class="o"><<</span> <span class="n">name</span> <span class="o"><<</span> <span class="s">"</span><span class="se">\n</span><span class="s">"</span><span class="p">;</span>
<span class="kt">int</span> <span class="n">devMajor</span><span class="p">,</span> <span class="n">devMinor</span><span class="p">;</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuDeviceComputeCapability</span><span class="p">(</span><span class="o">&</span><span class="n">devMajor</span><span class="p">,</span> <span class="o">&</span><span class="n">devMinor</span><span class="p">,</span> <span class="n">device</span><span class="p">));</span>
<span class="n">std</span><span class="o">::</span><span class="n">cout</span> <span class="o"><<</span> <span class="s">"Device Compute Capability: "</span>
<span class="o"><<</span> <span class="n">devMajor</span> <span class="o"><<</span> <span class="s">"."</span> <span class="o"><<</span> <span class="n">devMinor</span> <span class="o"><<</span> <span class="s">"</span><span class="se">\n</span><span class="s">"</span><span class="p">;</span>
<span class="k">if</span> <span class="p">(</span><span class="n">devMajor</span> <span class="o"><</span> <span class="mi">2</span><span class="p">)</span> <span class="p">{</span>
<span class="n">std</span><span class="o">::</span><span class="n">cerr</span> <span class="o"><<</span> <span class="s">"ERROR: Device 0 is not SM 2.0 or greater</span><span class="se">\n</span><span class="s">"</span><span class="p">;</span>
<span class="k">return</span> <span class="mi">1</span><span class="p">;</span>
<span class="p">}</span>
<span class="n">std</span><span class="o">::</span><span class="n">ifstream</span> <span class="n">t</span><span class="p">(</span><span class="s">"kernel.ptx"</span><span class="p">);</span>
<span class="k">if</span> <span class="p">(</span><span class="o">!</span><span class="n">t</span><span class="p">.</span><span class="n">is_open</span><span class="p">())</span> <span class="p">{</span>
<span class="n">std</span><span class="o">::</span><span class="n">cerr</span> <span class="o"><<</span> <span class="s">"kernel.ptx not found</span><span class="se">\n</span><span class="s">"</span><span class="p">;</span>
<span class="k">return</span> <span class="mi">1</span><span class="p">;</span>
<span class="p">}</span>
<span class="n">std</span><span class="o">::</span><span class="n">string</span> <span class="n">str</span><span class="p">((</span><span class="n">std</span><span class="o">::</span><span class="n">istreambuf_iterator</span><span class="o"><</span><span class="kt">char</span><span class="o">></span><span class="p">(</span><span class="n">t</span><span class="p">)),</span>
<span class="n">std</span><span class="o">::</span><span class="n">istreambuf_iterator</span><span class="o"><</span><span class="kt">char</span><span class="o">></span><span class="p">());</span>
<span class="c1">// Create driver context</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuCtxCreate</span><span class="p">(</span><span class="o">&</span><span class="n">context</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="n">device</span><span class="p">));</span>
<span class="c1">// Create module for object</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuModuleLoadDataEx</span><span class="p">(</span><span class="o">&</span><span class="n">cudaModule</span><span class="p">,</span> <span class="n">str</span><span class="p">.</span><span class="n">c_str</span><span class="p">(),</span> <span class="mi">0</span><span class="p">,</span> <span class="mi">0</span><span class="p">,</span> <span class="mi">0</span><span class="p">));</span>
<span class="c1">// Get kernel function</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuModuleGetFunction</span><span class="p">(</span><span class="o">&</span><span class="n">function</span><span class="p">,</span> <span class="n">cudaModule</span><span class="p">,</span> <span class="s">"kernel"</span><span class="p">));</span>
<span class="c1">// Device data</span>
<span class="n">CUdeviceptr</span> <span class="n">devBufferA</span><span class="p">;</span>
<span class="n">CUdeviceptr</span> <span class="n">devBufferB</span><span class="p">;</span>
<span class="n">CUdeviceptr</span> <span class="n">devBufferC</span><span class="p">;</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuMemAlloc</span><span class="p">(</span><span class="o">&</span><span class="n">devBufferA</span><span class="p">,</span> <span class="k">sizeof</span><span class="p">(</span><span class="kt">float</span><span class="p">)</span><span class="o">*</span><span class="mi">16</span><span class="p">));</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuMemAlloc</span><span class="p">(</span><span class="o">&</span><span class="n">devBufferB</span><span class="p">,</span> <span class="k">sizeof</span><span class="p">(</span><span class="kt">float</span><span class="p">)</span><span class="o">*</span><span class="mi">16</span><span class="p">));</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuMemAlloc</span><span class="p">(</span><span class="o">&</span><span class="n">devBufferC</span><span class="p">,</span> <span class="k">sizeof</span><span class="p">(</span><span class="kt">float</span><span class="p">)</span><span class="o">*</span><span class="mi">16</span><span class="p">));</span>
<span class="kt">float</span><span class="o">*</span> <span class="n">hostA</span> <span class="o">=</span> <span class="k">new</span> <span class="kt">float</span><span class="p">[</span><span class="mi">16</span><span class="p">];</span>
<span class="kt">float</span><span class="o">*</span> <span class="n">hostB</span> <span class="o">=</span> <span class="k">new</span> <span class="kt">float</span><span class="p">[</span><span class="mi">16</span><span class="p">];</span>
<span class="kt">float</span><span class="o">*</span> <span class="n">hostC</span> <span class="o">=</span> <span class="k">new</span> <span class="kt">float</span><span class="p">[</span><span class="mi">16</span><span class="p">];</span>
<span class="c1">// Populate input</span>
<span class="k">for</span> <span class="p">(</span><span class="kt">unsigned</span> <span class="n">i</span> <span class="o">=</span> <span class="mi">0</span><span class="p">;</span> <span class="n">i</span> <span class="o">!=</span> <span class="mi">16</span><span class="p">;</span> <span class="o">++</span><span class="n">i</span><span class="p">)</span> <span class="p">{</span>
<span class="n">hostA</span><span class="p">[</span><span class="n">i</span><span class="p">]</span> <span class="o">=</span> <span class="p">(</span><span class="kt">float</span><span class="p">)</span><span class="n">i</span><span class="p">;</span>
<span class="n">hostB</span><span class="p">[</span><span class="n">i</span><span class="p">]</span> <span class="o">=</span> <span class="p">(</span><span class="kt">float</span><span class="p">)(</span><span class="mi">2</span><span class="o">*</span><span class="n">i</span><span class="p">);</span>
<span class="n">hostC</span><span class="p">[</span><span class="n">i</span><span class="p">]</span> <span class="o">=</span> <span class="mf">0.0f</span><span class="p">;</span>
<span class="p">}</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuMemcpyHtoD</span><span class="p">(</span><span class="n">devBufferA</span><span class="p">,</span> <span class="o">&</span><span class="n">hostA</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span> <span class="k">sizeof</span><span class="p">(</span><span class="kt">float</span><span class="p">)</span><span class="o">*</span><span class="mi">16</span><span class="p">));</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuMemcpyHtoD</span><span class="p">(</span><span class="n">devBufferB</span><span class="p">,</span> <span class="o">&</span><span class="n">hostB</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span> <span class="k">sizeof</span><span class="p">(</span><span class="kt">float</span><span class="p">)</span><span class="o">*</span><span class="mi">16</span><span class="p">));</span>
<span class="kt">unsigned</span> <span class="n">blockSizeX</span> <span class="o">=</span> <span class="mi">16</span><span class="p">;</span>
<span class="kt">unsigned</span> <span class="n">blockSizeY</span> <span class="o">=</span> <span class="mi">1</span><span class="p">;</span>
<span class="kt">unsigned</span> <span class="n">blockSizeZ</span> <span class="o">=</span> <span class="mi">1</span><span class="p">;</span>
<span class="kt">unsigned</span> <span class="n">gridSizeX</span> <span class="o">=</span> <span class="mi">1</span><span class="p">;</span>
<span class="kt">unsigned</span> <span class="n">gridSizeY</span> <span class="o">=</span> <span class="mi">1</span><span class="p">;</span>
<span class="kt">unsigned</span> <span class="n">gridSizeZ</span> <span class="o">=</span> <span class="mi">1</span><span class="p">;</span>
<span class="c1">// Kernel parameters</span>
<span class="kt">void</span> <span class="o">*</span><span class="n">KernelParams</span><span class="p">[]</span> <span class="o">=</span> <span class="p">{</span> <span class="o">&</span><span class="n">devBufferA</span><span class="p">,</span> <span class="o">&</span><span class="n">devBufferB</span><span class="p">,</span> <span class="o">&</span><span class="n">devBufferC</span> <span class="p">};</span>
<span class="n">std</span><span class="o">::</span><span class="n">cout</span> <span class="o"><<</span> <span class="s">"Launching kernel</span><span class="se">\n</span><span class="s">"</span><span class="p">;</span>
<span class="c1">// Kernel launch</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuLaunchKernel</span><span class="p">(</span><span class="n">function</span><span class="p">,</span> <span class="n">gridSizeX</span><span class="p">,</span> <span class="n">gridSizeY</span><span class="p">,</span> <span class="n">gridSizeZ</span><span class="p">,</span>
<span class="n">blockSizeX</span><span class="p">,</span> <span class="n">blockSizeY</span><span class="p">,</span> <span class="n">blockSizeZ</span><span class="p">,</span>
<span class="mi">0</span><span class="p">,</span> <span class="nb">NULL</span><span class="p">,</span> <span class="n">KernelParams</span><span class="p">,</span> <span class="nb">NULL</span><span class="p">));</span>
<span class="c1">// Retrieve device data</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuMemcpyDtoH</span><span class="p">(</span><span class="o">&</span><span class="n">hostC</span><span class="p">[</span><span class="mi">0</span><span class="p">],</span> <span class="n">devBufferC</span><span class="p">,</span> <span class="k">sizeof</span><span class="p">(</span><span class="kt">float</span><span class="p">)</span><span class="o">*</span><span class="mi">16</span><span class="p">));</span>
<span class="n">std</span><span class="o">::</span><span class="n">cout</span> <span class="o"><<</span> <span class="s">"Results:</span><span class="se">\n</span><span class="s">"</span><span class="p">;</span>
<span class="k">for</span> <span class="p">(</span><span class="kt">unsigned</span> <span class="n">i</span> <span class="o">=</span> <span class="mi">0</span><span class="p">;</span> <span class="n">i</span> <span class="o">!=</span> <span class="mi">16</span><span class="p">;</span> <span class="o">++</span><span class="n">i</span><span class="p">)</span> <span class="p">{</span>
<span class="n">std</span><span class="o">::</span><span class="n">cout</span> <span class="o"><<</span> <span class="n">hostA</span><span class="p">[</span><span class="n">i</span><span class="p">]</span> <span class="o"><<</span> <span class="s">" + "</span> <span class="o"><<</span> <span class="n">hostB</span><span class="p">[</span><span class="n">i</span><span class="p">]</span> <span class="o"><<</span> <span class="s">" = "</span> <span class="o"><<</span> <span class="n">hostC</span><span class="p">[</span><span class="n">i</span><span class="p">]</span> <span class="o"><<</span> <span class="s">"</span><span class="se">\n</span><span class="s">"</span><span class="p">;</span>
<span class="p">}</span>
<span class="c1">// Clean up after ourselves</span>
<span class="k">delete</span> <span class="p">[]</span> <span class="n">hostA</span><span class="p">;</span>
<span class="k">delete</span> <span class="p">[]</span> <span class="n">hostB</span><span class="p">;</span>
<span class="k">delete</span> <span class="p">[]</span> <span class="n">hostC</span><span class="p">;</span>
<span class="c1">// Clean-up</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuMemFree</span><span class="p">(</span><span class="n">devBufferA</span><span class="p">));</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuMemFree</span><span class="p">(</span><span class="n">devBufferB</span><span class="p">));</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuMemFree</span><span class="p">(</span><span class="n">devBufferC</span><span class="p">));</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuModuleUnload</span><span class="p">(</span><span class="n">cudaModule</span><span class="p">));</span>
<span class="n">checkCudaErrors</span><span class="p">(</span><span class="n">cuCtxDestroy</span><span class="p">(</span><span class="n">context</span><span class="p">));</span>
<span class="k">return</span> <span class="mi">0</span><span class="p">;</span>
<span class="p">}</span>
</pre></div>
</div>
<p>You will need to link with the CUDA driver and specify the path to cuda.h.</p>
<div class="highlight-text notranslate"><div class="highlight"><pre><span></span># clang++ sample.cpp -o sample -O2 -g -I/usr/local/cuda-5.5/include -lcuda
</pre></div>
</div>
<p>We don’t need to specify a path to <code class="docutils literal notranslate"><span class="pre">libcuda.so</span></code> since this is installed in a
system location by the driver, not the CUDA toolkit.</p>
<p>If everything goes as planned, you should see the following output when
running the compiled program:</p>
<div class="highlight-text notranslate"><div class="highlight"><pre><span></span>Using CUDA Device [0]: GeForce GTX 680
Device Compute Capability: 3.0
Launching kernel
Results:
0 + 0 = 0
1 + 2 = 3
2 + 4 = 6
3 + 6 = 9
4 + 8 = 12
5 + 10 = 15
6 + 12 = 18
7 + 14 = 21
8 + 16 = 24
9 + 18 = 27
10 + 20 = 30
11 + 22 = 33
12 + 24 = 36
13 + 26 = 39
14 + 28 = 42
15 + 30 = 45
</pre></div>
</div>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>You will likely see a different device identifier based on your hardware</p>
</div>
</div>
</div>
<div class="section" id="tutorial-linking-with-libdevice">
<h2><a class="toc-backref" href="#id38">Tutorial: Linking with Libdevice</a><a class="headerlink" href="#tutorial-linking-with-libdevice" title="Permalink to this headline">¶</a></h2>
<p>In this tutorial, we show a simple example of linking LLVM IR with the
libdevice library. We will use the same kernel as the previous tutorial,
except that we will compute <code class="docutils literal notranslate"><span class="pre">C</span> <span class="pre">=</span> <span class="pre">pow(A,</span> <span class="pre">B)</span></code> instead of <code class="docutils literal notranslate"><span class="pre">C</span> <span class="pre">=</span> <span class="pre">A</span> <span class="pre">+</span> <span class="pre">B</span></code>.
Libdevice provides an <code class="docutils literal notranslate"><span class="pre">__nv_powf</span></code> function that we will use.</p>
<div class="highlight-llvm notranslate"><div class="highlight"><pre><span></span><span class="k">target</span> <span class="k">datalayout</span> <span class="p">=</span> <span class="s">"e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"</span>
<span class="k">target</span> <span class="k">triple</span> <span class="p">=</span> <span class="s">"nvptx64-nvidia-cuda"</span>
<span class="c">; Intrinsic to read X component of thread ID</span>
<span class="k">declare</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.tid.x</span><span class="p">()</span> <span class="k">readnone</span> <span class="k">nounwind</span>
<span class="c">; libdevice function</span>
<span class="k">declare</span> <span class="k">float</span> <span class="vg">@__nv_powf</span><span class="p">(</span><span class="k">float</span><span class="p">,</span> <span class="k">float</span><span class="p">)</span>
<span class="k">define</span> <span class="k">void</span> <span class="vg">@kernel</span><span class="p">(</span><span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%A</span><span class="p">,</span>
<span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%B</span><span class="p">,</span>
<span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%C</span><span class="p">)</span> <span class="p">{</span>
<span class="nl">entry:</span>
<span class="c">; What is my ID?</span>
<span class="nv">%id</span> <span class="p">=</span> <span class="k">tail</span> <span class="k">call</span> <span class="k">i32</span> <span class="vg">@llvm.nvvm.read.ptx.sreg.tid.x</span><span class="p">()</span> <span class="k">readnone</span> <span class="k">nounwind</span>
<span class="c">; Compute pointers into A, B, and C</span>
<span class="nv">%ptrA</span> <span class="p">=</span> <span class="k">getelementptr</span> <span class="k">float</span><span class="p">,</span> <span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%A</span><span class="p">,</span> <span class="k">i32</span> <span class="nv">%id</span>
<span class="nv">%ptrB</span> <span class="p">=</span> <span class="k">getelementptr</span> <span class="k">float</span><span class="p">,</span> <span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%B</span><span class="p">,</span> <span class="k">i32</span> <span class="nv">%id</span>
<span class="nv">%ptrC</span> <span class="p">=</span> <span class="k">getelementptr</span> <span class="k">float</span><span class="p">,</span> <span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%C</span><span class="p">,</span> <span class="k">i32</span> <span class="nv">%id</span>
<span class="c">; Read A, B</span>
<span class="nv">%valA</span> <span class="p">=</span> <span class="k">load</span> <span class="k">float</span><span class="p">,</span> <span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%ptrA</span><span class="p">,</span> <span class="k">align</span> <span class="m">4</span>
<span class="nv">%valB</span> <span class="p">=</span> <span class="k">load</span> <span class="k">float</span><span class="p">,</span> <span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%ptrB</span><span class="p">,</span> <span class="k">align</span> <span class="m">4</span>
<span class="c">; Compute C = pow(A, B)</span>
<span class="nv">%valC</span> <span class="p">=</span> <span class="k">call</span> <span class="k">float</span> <span class="vg">@__nv_powf</span><span class="p">(</span><span class="k">float</span> <span class="nv">%valA</span><span class="p">,</span> <span class="k">float</span> <span class="nv">%valB</span><span class="p">)</span>
<span class="c">; Store back to C</span>
<span class="k">store</span> <span class="k">float</span> <span class="nv">%valC</span><span class="p">,</span> <span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*</span> <span class="nv">%ptrC</span><span class="p">,</span> <span class="k">align</span> <span class="m">4</span>
<span class="k">ret</span> <span class="k">void</span>
<span class="p">}</span>
<span class="nv">!nvvm.annotations</span> <span class="p">=</span> <span class="p">!{</span><span class="nv nv-Anonymous">!0</span><span class="p">}</span>
<span class="nv nv-Anonymous">!0</span> <span class="p">=</span> <span class="p">!{</span><span class="k">void</span> <span class="p">(</span><span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*,</span>
<span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*,</span>
<span class="k">float</span> <span class="k">addrspace</span><span class="p">(</span><span class="m">1</span><span class="p">)*)*</span> <span class="vg">@kernel</span><span class="p">,</span> <span class="nv">!"kernel"</span><span class="p">,</span> <span class="k">i32</span> <span class="m">1</span><span class="p">}</span>
</pre></div>
</div>
<p>To compile this kernel, we perform the following steps:</p>
<ol class="arabic simple">
<li><p>Link with libdevice</p></li>
<li><p>Internalize all but the public kernel function</p></li>
<li><p>Run <code class="docutils literal notranslate"><span class="pre">NVVMReflect</span></code> and set <code class="docutils literal notranslate"><span class="pre">__CUDA_FTZ</span></code> to 0</p></li>
<li><p>Optimize the linked module</p></li>
<li><p>Codegen the module</p></li>
</ol>
<p>These steps can be performed by the LLVM <code class="docutils literal notranslate"><span class="pre">llvm-link</span></code>, <code class="docutils literal notranslate"><span class="pre">opt</span></code>, and <code class="docutils literal notranslate"><span class="pre">llc</span></code>
tools. In a complete compiler, these steps can also be performed entirely
programmatically by setting up an appropriate pass configuration (see
<a class="reference internal" href="#libdevice"><span class="std std-ref">Linking with Libdevice</span></a>).</p>
<div class="highlight-text notranslate"><div class="highlight"><pre><span></span># llvm-link t2.bc libdevice.compute_20.10.bc -o t2.linked.bc
# opt -internalize -internalize-public-api-list=kernel -nvvm-reflect-list=__CUDA_FTZ=0 -nvvm-reflect -O3 t2.linked.bc -o t2.opt.bc
# llc -mcpu=sm_20 t2.opt.bc -o t2.ptx
</pre></div>
</div>
<div class="admonition note">
<p class="admonition-title">Note</p>
<p>The <code class="docutils literal notranslate"><span class="pre">-nvvm-reflect-list=_CUDA_FTZ=0</span></code> is not strictly required, as any
undefined variables will default to zero. It is shown here for evaluation
purposes.</p>
</div>
<p>This gives us the following PTX (excerpt):</p>
<div class="highlight-text notranslate"><div class="highlight"><pre><span></span>//
// Generated by LLVM NVPTX Back-End
//
.version 3.1
.target sm_20
.address_size 64
// .globl kernel
// @kernel
.visible .entry kernel(
.param .u64 kernel_param_0,
.param .u64 kernel_param_1,
.param .u64 kernel_param_2
)
{
.reg .pred %p<30>;
.reg .f32 %f<111>;
.reg .s32 %r<21>;
.reg .s64 %rl<8>;
// %bb.0: // %entry
ld.param.u64 %rl2, [kernel_param_0];
mov.u32 %r3, %tid.x;
ld.param.u64 %rl3, [kernel_param_1];
mul.wide.s32 %rl4, %r3, 4;
add.s64 %rl5, %rl2, %rl4;
ld.param.u64 %rl6, [kernel_param_2];
add.s64 %rl7, %rl3, %rl4;
add.s64 %rl1, %rl6, %rl4;
ld.global.f32 %f1, [%rl5];
ld.global.f32 %f2, [%rl7];
setp.eq.f32 %p1, %f1, 0f3F800000;
setp.eq.f32 %p2, %f2, 0f00000000;
or.pred %p3, %p1, %p2;
@%p3 bra BB0_1;
bra.uni BB0_2;
BB0_1:
mov.f32 %f110, 0f3F800000;
st.global.f32 [%rl1], %f110;
ret;
BB0_2: // %__nv_isnanf.exit.i
abs.f32 %f4, %f1;
setp.gtu.f32 %p4, %f4, 0f7F800000;
@%p4 bra BB0_4;
// %bb.3: // %__nv_isnanf.exit5.i
abs.f32 %f5, %f2;
setp.le.f32 %p5, %f5, 0f7F800000;
@%p5 bra BB0_5;
BB0_4: // %.critedge1.i
add.f32 %f110, %f1, %f2;
st.global.f32 [%rl1], %f110;
ret;
BB0_5: // %__nv_isinff.exit.i
...
BB0_26: // %__nv_truncf.exit.i.i.i.i.i
mul.f32 %f90, %f107, 0f3FB8AA3B;
cvt.rzi.f32.f32 %f91, %f90;
mov.f32 %f92, 0fBF317200;
fma.rn.f32 %f93, %f91, %f92, %f107;
mov.f32 %f94, 0fB5BFBE8E;
fma.rn.f32 %f95, %f91, %f94, %f93;
mul.f32 %f89, %f95, 0f3FB8AA3B;
// inline asm
ex2.approx.ftz.f32 %f88,%f89;
// inline asm
add.f32 %f96, %f91, 0f00000000;
ex2.approx.f32 %f97, %f96;
mul.f32 %f98, %f88, %f97;
setp.lt.f32 %p15, %f107, 0fC2D20000;
selp.f32 %f99, 0f00000000, %f98, %p15;
setp.gt.f32 %p16, %f107, 0f42D20000;
selp.f32 %f110, 0f7F800000, %f99, %p16;
setp.eq.f32 %p17, %f110, 0f7F800000;
@%p17 bra BB0_28;
// %bb.27:
fma.rn.f32 %f110, %f110, %f108, %f110;
BB0_28: // %__internal_accurate_powf.exit.i
setp.lt.f32 %p18, %f1, 0f00000000;
setp.eq.f32 %p19, %f3, 0f3F800000;
and.pred %p20, %p18, %p19;
@!%p20 bra BB0_30;
bra.uni BB0_29;
BB0_29:
mov.b32 %r9, %f110;
xor.b32 %r10, %r9, -2147483648;
mov.b32 %f110, %r10;
BB0_30: // %__nv_powf.exit
st.global.f32 [%rl1], %f110;
ret;
}
</pre></div>
</div>
</div>
</div>
<div class="clearer"></div>
</div>
</div>
</div>
<div class="clearer"></div>
</div>
<div class="related" role="navigation" aria-label="related navigation">
<h3>Navigation</h3>
<ul>
<li class="right" style="margin-right: 10px">
<a href="genindex.html" title="General Index"
>index</a></li>
<li class="right" >
<a href="Phabricator.html" title="Code Reviews with Phabricator"
>next</a> |</li>
<li class="right" >
<a href="NewPassManager.html" title="Using the New Pass Manager"
>previous</a> |</li>
<li><a href="https://llvm.org/">LLVM Home</a> | </li>
<li><a href="index.html">Documentation</a>»</li>
<li class="nav-item nav-item-1"><a href="UserGuides.html" >User Guides</a> »</li>
<li class="nav-item nav-item-this"><a href="">User Guide for NVPTX Back-end</a></li>
</ul>
</div>
<div class="footer" role="contentinfo">
© Copyright 2003-2021, LLVM Project.
Last updated on 2021-09-18.
Created using <a href="https://www.sphinx-doc.org/">Sphinx</a> 3.5.4.
</div>
</body>
</html>
|
