-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathindex.html
1093 lines (1005 loc) · 157 KB
/
index.html
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
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width, initial-scale=1.0" />
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
<meta http-equiv="x-ua-compatible" content="ie=edge">
<title>5. Integration with Machine Learning Frameworks — Machine Learing Compiler 0.0.1 documentation</title>
<link rel="stylesheet" href="../_static/material-design-lite-1.3.0/material.blue-deep_orange.min.css" type="text/css" />
<link rel="stylesheet" href="../_static/sphinx_materialdesign_theme.css" type="text/css" />
<link rel="stylesheet" href="../_static/fontawesome/all.css" type="text/css" />
<link rel="stylesheet" href="../_static/fonts.css" type="text/css" />
<link rel="stylesheet" type="text/css" href="../_static/pygments.css" />
<link rel="stylesheet" type="text/css" href="../_static/basic.css" />
<link rel="stylesheet" type="text/css" href="../_static/d2l.css" />
<script data-url_root="../" id="documentation_options" src="../_static/documentation_options.js"></script>
<script src="../_static/jquery.js"></script>
<script src="../_static/underscore.js"></script>
<script src="../_static/_sphinx_javascript_frameworks_compat.js"></script>
<script src="../_static/doctools.js"></script>
<script src="../_static/sphinx_highlight.js"></script>
<script src="../_static/d2l.js"></script>
<link rel="shortcut icon" href="../_static/mlc-favicon.ico"/>
<link rel="index" title="Index" href="../genindex.html" />
<link rel="search" title="Search" href="../search.html" />
<link rel="next" title="6. GPU and Hardware Acceleration" href="../chapter_gpu_acceleration/index.html" />
<link rel="prev" title="4. Automatic Program Optimization" href="../chapter_auto_program_optimization/index.html" />
</head>
<body>
<div class="mdl-layout mdl-js-layout mdl-layout--fixed-header mdl-layout--fixed-drawer"><header class="mdl-layout__header mdl-layout__header--waterfall ">
<div class="mdl-layout__header-row">
<nav class="mdl-navigation breadcrumb">
<a class="mdl-navigation__link is-active"><span class="section-number">5. </span>Integration with Machine Learning Frameworks</a>
</nav>
<div class="mdl-layout-spacer"></div>
<nav class="mdl-navigation">
<form class="form-inline pull-sm-right" action="../search.html" method="get">
<div class="mdl-textfield mdl-js-textfield mdl-textfield--expandable mdl-textfield--floating-label mdl-textfield--align-right">
<label id="quick-search-icon" class="mdl-button mdl-js-button mdl-button--icon" for="waterfall-exp">
<i class="material-icons">search</i>
</label>
<div class="mdl-textfield__expandable-holder">
<input class="mdl-textfield__input" type="text" name="q" id="waterfall-exp" placeholder="Search" />
<input type="hidden" name="check_keywords" value="yes" />
<input type="hidden" name="area" value="default" />
</div>
</div>
<div class="mdl-tooltip" data-mdl-for="quick-search-icon">
Quick search
</div>
</form>
<a id="button-show-source"
class="mdl-button mdl-js-button mdl-button--icon"
href="../_sources/chapter_integration/index.rst.txt" rel="nofollow">
<i class="material-icons">code</i>
</a>
<div class="mdl-tooltip" data-mdl-for="button-show-source">
Show Source
</div>
</nav>
</div>
<div class="mdl-layout__header-row header-links">
<div class="mdl-layout-spacer"></div>
<nav class="mdl-navigation">
<a class="mdl-navigation__link" href="https://mlc.ai/summer22">
<i class="fas fa-user-graduate"></i>
Course
</a>
<a class="mdl-navigation__link" href="https://github.com/mlc-ai/mlc-en">
<i class="fab fa-github"></i>
GitHub
</a>
<a class="mdl-navigation__link" href="https://mlc.ai/zh">
<i class="fas fa-external-link-alt"></i>
中文版
</a>
</nav>
</div>
</header><header class="mdl-layout__drawer">
<!-- Title -->
<span class="mdl-layout-title">
<a class="title" href="../index.html">
<img class="logo" src="../_static/mlc-logo-with-text-landscape.svg" alt="Machine Learing Compiler"/>
</a>
</span>
<div class="globaltoc">
<span class="mdl-layout-title toc">Table Of Contents</span>
<nav class="mdl-navigation">
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="../chapter_introduction/index.html">1. Introduction</a></li>
<li class="toctree-l1"><a class="reference internal" href="../chapter_tensor_program/index.html">2. Tensor Program Abstraction</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../chapter_tensor_program/tensor_program.html">2.1. Primitive Tensor Function</a></li>
<li class="toctree-l2"><a class="reference internal" href="../chapter_tensor_program/tensor_program.html#tensor-program-abstraction">2.2. Tensor Program Abstraction</a></li>
<li class="toctree-l2"><a class="reference internal" href="../chapter_tensor_program/tensor_program.html#summary">2.3. Summary</a></li>
<li class="toctree-l2"><a class="reference internal" href="../chapter_tensor_program/case_study.html">2.4. TensorIR: Tensor Program Abstraction Case Study</a></li>
<li class="toctree-l2"><a class="reference internal" href="../chapter_tensor_program/tensorir_exercises.html">2.5. Exercises for TensorIR</a></li>
</ul>
</li>
<li class="toctree-l1"><a class="reference internal" href="../chapter_end_to_end/index.html">3. End to End Model Execution</a></li>
<li class="toctree-l1"><a class="reference internal" href="../chapter_auto_program_optimization/index.html">4. Automatic Program Optimization</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">5. Integration with Machine Learning Frameworks</a></li>
<li class="toctree-l1"><a class="reference internal" href="../chapter_gpu_acceleration/index.html">6. GPU and Hardware Acceleration</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../chapter_gpu_acceleration/part1.html">6.1. Part 1</a></li>
<li class="toctree-l2"><a class="reference internal" href="../chapter_gpu_acceleration/part2.html">6.2. Part 2</a></li>
</ul>
</li>
<li class="toctree-l1"><a class="reference internal" href="../chapter_graph_optimization/index.html">7. Computational Graph Optimization</a></li>
</ul>
</nav>
</div>
</header>
<main class="mdl-layout__content" tabIndex="0">
<script type="text/javascript" src="../_static/sphinx_materialdesign_theme.js "></script>
<header class="mdl-layout__drawer">
<!-- Title -->
<span class="mdl-layout-title">
<a class="title" href="../index.html">
<img class="logo" src="../_static/mlc-logo-with-text-landscape.svg" alt="Machine Learing Compiler"/>
</a>
</span>
<div class="globaltoc">
<span class="mdl-layout-title toc">Table Of Contents</span>
<nav class="mdl-navigation">
<ul class="current">
<li class="toctree-l1"><a class="reference internal" href="../chapter_introduction/index.html">1. Introduction</a></li>
<li class="toctree-l1"><a class="reference internal" href="../chapter_tensor_program/index.html">2. Tensor Program Abstraction</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../chapter_tensor_program/tensor_program.html">2.1. Primitive Tensor Function</a></li>
<li class="toctree-l2"><a class="reference internal" href="../chapter_tensor_program/tensor_program.html#tensor-program-abstraction">2.2. Tensor Program Abstraction</a></li>
<li class="toctree-l2"><a class="reference internal" href="../chapter_tensor_program/tensor_program.html#summary">2.3. Summary</a></li>
<li class="toctree-l2"><a class="reference internal" href="../chapter_tensor_program/case_study.html">2.4. TensorIR: Tensor Program Abstraction Case Study</a></li>
<li class="toctree-l2"><a class="reference internal" href="../chapter_tensor_program/tensorir_exercises.html">2.5. Exercises for TensorIR</a></li>
</ul>
</li>
<li class="toctree-l1"><a class="reference internal" href="../chapter_end_to_end/index.html">3. End to End Model Execution</a></li>
<li class="toctree-l1"><a class="reference internal" href="../chapter_auto_program_optimization/index.html">4. Automatic Program Optimization</a></li>
<li class="toctree-l1 current"><a class="current reference internal" href="#">5. Integration with Machine Learning Frameworks</a></li>
<li class="toctree-l1"><a class="reference internal" href="../chapter_gpu_acceleration/index.html">6. GPU and Hardware Acceleration</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../chapter_gpu_acceleration/part1.html">6.1. Part 1</a></li>
<li class="toctree-l2"><a class="reference internal" href="../chapter_gpu_acceleration/part2.html">6.2. Part 2</a></li>
</ul>
</li>
<li class="toctree-l1"><a class="reference internal" href="../chapter_graph_optimization/index.html">7. Computational Graph Optimization</a></li>
</ul>
</nav>
</div>
</header>
<div class="document">
<div class="page-content" role="main">
<div class="section" id="integration-with-machine-learning-frameworks">
<h1><span class="section-number">5. </span>Integration with Machine Learning Frameworks<a class="headerlink" href="#integration-with-machine-learning-frameworks" title="Permalink to this heading">¶</a></h1>
<div class="section" id="prelude">
<h2><span class="section-number">5.1. </span>Prelude<a class="headerlink" href="#prelude" title="Permalink to this heading">¶</a></h2>
<p>In the past chapters, we have learned about abstractions for machine
learning compilation and transformations among tensor functions.</p>
<p>This chapter will discuss how to bring machine learning models from the
existing ML framework into an MLC flow.</p>
</div>
<div class="section" id="preparations">
<h2><span class="section-number">5.2. </span>Preparations<a class="headerlink" href="#preparations" title="Permalink to this heading">¶</a></h2>
<p>To begin with, we will import necessary dependencies.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">numpy</span> <span class="k">as</span> <span class="nn">np</span>
<span class="kn">import</span> <span class="nn">tvm</span>
<span class="kn">from</span> <span class="nn">tvm</span> <span class="kn">import</span> <span class="n">relax</span>
<span class="kn">from</span> <span class="nn">tvm.ir.module</span> <span class="kn">import</span> <span class="n">IRModule</span>
<span class="kn">from</span> <span class="nn">tvm.script</span> <span class="kn">import</span> <span class="n">relax</span> <span class="k">as</span> <span class="n">R</span>
<span class="kn">from</span> <span class="nn">tvm.script</span> <span class="kn">import</span> <span class="n">tir</span> <span class="k">as</span> <span class="n">T</span>
</pre></div>
</div>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="kn">import</span> <span class="nn">torch.nn</span> <span class="k">as</span> <span class="nn">nn</span>
<span class="kn">from</span> <span class="nn">torch</span> <span class="kn">import</span> <span class="n">fx</span>
<span class="kn">from</span> <span class="nn">torch.nn</span> <span class="kn">import</span> <span class="n">functional</span> <span class="k">as</span> <span class="n">F</span>
</pre></div>
</div>
</div>
<div class="section" id="build-an-irmodule-through-a-builder">
<h2><span class="section-number">5.3. </span>Build an IRModule Through a Builder<a class="headerlink" href="#build-an-irmodule-through-a-builder" title="Permalink to this heading">¶</a></h2>
<p>In the past chapters, we have been building IRModule by directly writing
TVMScript. As the model gets larger, we need a programmatical way to
build up an IRModule. In this section, let us review some of the tools
to support that process.</p>
<div class="section" id="tensor-expression-for-tensorir-creation">
<h3><span class="section-number">5.3.1. </span>Tensor Expression for TensorIR Creation<a class="headerlink" href="#tensor-expression-for-tensorir-creation" title="Permalink to this heading">¶</a></h3>
<p>First, we review the tensor expression domain-specific language to build
TensorIR functions.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">from</span> <span class="nn">tvm</span> <span class="kn">import</span> <span class="n">te</span>
</pre></div>
</div>
<p>We begin by creating a placeholder object, which represents an input to
a TensorIR function.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">A</span> <span class="o">=</span> <span class="n">te</span><span class="o">.</span><span class="n">placeholder</span><span class="p">((</span><span class="mi">128</span><span class="p">,</span> <span class="mi">128</span><span class="p">),</span> <span class="n">name</span><span class="o">=</span><span class="s2">"A"</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="s2">"float32"</span><span class="p">)</span>
<span class="n">B</span> <span class="o">=</span> <span class="n">te</span><span class="o">.</span><span class="n">placeholder</span><span class="p">((</span><span class="mi">128</span><span class="p">,</span> <span class="mi">128</span><span class="p">),</span> <span class="n">name</span><span class="o">=</span><span class="s2">"B"</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="s2">"float32"</span><span class="p">)</span>
</pre></div>
</div>
<p>Each input and intermediate result here are represented as a
<code class="docutils literal notranslate"><span class="pre">te.Tensor</span></code> object.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="nb">type</span><span class="p">(</span><span class="n">A</span><span class="p">)</span>
</pre></div>
</div>
<div class="output highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">tvm</span><span class="o">.</span><span class="n">te</span><span class="o">.</span><span class="n">tensor</span><span class="o">.</span><span class="n">Tensor</span>
</pre></div>
</div>
<p>Each <code class="docutils literal notranslate"><span class="pre">te.Tensor</span></code> has a shape field and dtype field that tracks the
shape and data type of the computation.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">A</span><span class="o">.</span><span class="n">shape</span>
</pre></div>
</div>
<div class="output highlight-default notranslate"><div class="highlight"><pre><span></span><span class="p">[</span><span class="mi">128</span><span class="p">,</span> <span class="mi">128</span><span class="p">]</span>
</pre></div>
</div>
<p>We can describe computations through a sequence of tensor expression
computation, Here <code class="docutils literal notranslate"><span class="pre">te.compute</span></code> takes the signature
<code class="docutils literal notranslate"><span class="pre">te.compute(output_shape,</span> <span class="pre">fcompute)</span></code>. And the fcompute function
describes how we want to compute the value of each element <code class="docutils literal notranslate"><span class="pre">[i,</span> <span class="pre">j]</span></code>
for a given index.</p>
<p>The <code class="docutils literal notranslate"><span class="pre">te_matmul</span></code> function takes in an object with type <code class="docutils literal notranslate"><span class="pre">te.Tensor</span></code>,
and returns the matrix multiplication result. Note how we build up
computations depending on A and B’s input shape. The <code class="docutils literal notranslate"><span class="pre">te_matmul</span></code> works
for A and B with different input shapes.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">te_matmul</span><span class="p">(</span><span class="n">A</span><span class="p">:</span> <span class="n">te</span><span class="o">.</span><span class="n">Tensor</span><span class="p">,</span> <span class="n">B</span><span class="p">:</span> <span class="n">te</span><span class="o">.</span><span class="n">Tensor</span><span class="p">)</span> <span class="o">-></span> <span class="n">te</span><span class="o">.</span><span class="n">Tensor</span><span class="p">:</span>
<span class="k">assert</span> <span class="n">A</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">1</span><span class="p">]</span> <span class="o">==</span> <span class="n">B</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span>
<span class="n">n</span> <span class="o">=</span> <span class="n">A</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">0</span><span class="p">]</span>
<span class="n">m</span> <span class="o">=</span> <span class="n">B</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">1</span><span class="p">]</span>
<span class="n">k</span> <span class="o">=</span> <span class="n">te</span><span class="o">.</span><span class="n">reduce_axis</span><span class="p">((</span><span class="mi">0</span><span class="p">,</span> <span class="n">A</span><span class="o">.</span><span class="n">shape</span><span class="p">[</span><span class="mi">1</span><span class="p">]),</span> <span class="n">name</span><span class="o">=</span><span class="s2">"k"</span><span class="p">)</span>
<span class="k">return</span> <span class="n">te</span><span class="o">.</span><span class="n">compute</span><span class="p">((</span><span class="n">n</span><span class="p">,</span> <span class="n">m</span><span class="p">),</span> <span class="k">lambda</span> <span class="n">i</span><span class="p">,</span> <span class="n">j</span><span class="p">:</span> <span class="n">te</span><span class="o">.</span><span class="n">sum</span><span class="p">(</span><span class="n">A</span><span class="p">[</span><span class="n">i</span><span class="p">,</span> <span class="n">k</span><span class="p">]</span> <span class="o">*</span> <span class="n">B</span><span class="p">[</span><span class="n">k</span><span class="p">,</span> <span class="n">j</span><span class="p">],</span> <span class="n">axis</span><span class="o">=</span><span class="n">k</span><span class="p">),</span> <span class="n">name</span><span class="o">=</span><span class="s2">"matmul"</span><span class="p">)</span>
</pre></div>
</div>
<p>We can create the result of matmul calling <code class="docutils literal notranslate"><span class="pre">te_matmul</span></code> with A and B.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">C</span> <span class="o">=</span> <span class="n">te_matmul</span><span class="p">(</span><span class="n">A</span><span class="p">,</span> <span class="n">B</span><span class="p">)</span>
</pre></div>
</div>
<p>To create a TensorIR function, we can call <code class="docutils literal notranslate"><span class="pre">te.create_prim_func</span></code> and
pass in the input and output values.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">te</span><span class="o">.</span><span class="n">create_prim_func</span><span class="p">([</span><span class="n">A</span><span class="p">,</span> <span class="n">B</span><span class="p">,</span> <span class="n">C</span><span class="p">])</span><span class="o">.</span><span class="n">show</span><span class="p">()</span>
</pre></div>
</div>
<div class="highlight" style="background: "><pre style="line-height: 125%;"><span></span><span style="color: #007979; font-style: italic"># from tvm.script import tir as T</span>
<span style="color: #AA22FF">@T</span><span style="color: #AA22FF; font-weight: bold">.</span>prim_func
<span style="color: #008000; font-weight: bold">def</span> <span style="color: #0000FF">main</span>(A: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((<span style="color: #008000">128</span>, <span style="color: #008000">128</span>), <span style="color: #BA2121">"float32"</span>), B: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((<span style="color: #008000">128</span>, <span style="color: #008000">128</span>), <span style="color: #BA2121">"float32"</span>), matmul: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((<span style="color: #008000">128</span>, <span style="color: #008000">128</span>), <span style="color: #BA2121">"float32"</span>)):
T<span style="color: #AA22FF; font-weight: bold">.</span>func_attr({<span style="color: #BA2121">"tir.noalias"</span>: T<span style="color: #AA22FF; font-weight: bold">.</span>bool(<span style="color: #008000; font-weight: bold">True</span>)})
<span style="color: #007979; font-style: italic"># with T.block("root"):</span>
<span style="color: #008000; font-weight: bold">for</span> i, j, k <span style="color: #008000; font-weight: bold">in</span> T<span style="color: #AA22FF; font-weight: bold">.</span>grid(<span style="color: #008000">128</span>, <span style="color: #008000">128</span>, <span style="color: #008000">128</span>):
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>block(<span style="color: #BA2121">"matmul"</span>):
v_i, v_j, v_k <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>axis<span style="color: #AA22FF; font-weight: bold">.</span>remap(<span style="color: #BA2121">"SSR"</span>, [i, j, k])
T<span style="color: #AA22FF; font-weight: bold">.</span>reads(A[v_i, v_k], B[v_k, v_j])
T<span style="color: #AA22FF; font-weight: bold">.</span>writes(matmul[v_i, v_j])
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>init():
matmul[v_i, v_j] <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>float32(<span style="color: #008000">0.0</span>)
matmul[v_i, v_j] <span style="color: #AA22FF; font-weight: bold">=</span> matmul[v_i, v_j] <span style="color: #AA22FF; font-weight: bold">+</span> A[v_i, v_k] <span style="color: #AA22FF; font-weight: bold">*</span> B[v_k, v_j]
</pre></div><p>We can create a tensor expression for relu computation in a similar
fashion. Here we write it in a way so that <code class="docutils literal notranslate"><span class="pre">te_relu</span></code> function can work
for <code class="docutils literal notranslate"><span class="pre">te.Tensor</span></code> with any dimension and shape.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">te_relu</span><span class="p">(</span><span class="n">A</span><span class="p">:</span> <span class="n">te</span><span class="o">.</span><span class="n">Tensor</span><span class="p">)</span> <span class="o">-></span> <span class="n">te</span><span class="o">.</span><span class="n">Tensor</span><span class="p">:</span>
<span class="k">return</span> <span class="n">te</span><span class="o">.</span><span class="n">compute</span><span class="p">(</span><span class="n">A</span><span class="o">.</span><span class="n">shape</span><span class="p">,</span> <span class="k">lambda</span> <span class="o">*</span><span class="n">i</span><span class="p">:</span> <span class="n">te</span><span class="o">.</span><span class="n">max</span><span class="p">(</span><span class="n">A</span><span class="p">(</span><span class="o">*</span><span class="n">i</span><span class="p">),</span> <span class="mi">0</span><span class="p">),</span> <span class="n">name</span><span class="o">=</span><span class="s2">"relu"</span><span class="p">)</span>
</pre></div>
</div>
<p>Let us try out <code class="docutils literal notranslate"><span class="pre">te_relu</span></code> on two different input shapes and dimensions.
First <code class="docutils literal notranslate"><span class="pre">X1</span></code> with shape <code class="docutils literal notranslate"><span class="pre">(10,)</span></code>.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">X1</span> <span class="o">=</span> <span class="n">te</span><span class="o">.</span><span class="n">placeholder</span><span class="p">((</span><span class="mi">10</span><span class="p">,),</span> <span class="n">name</span><span class="o">=</span><span class="s2">"X1"</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="s2">"float32"</span><span class="p">)</span>
<span class="n">Y1</span> <span class="o">=</span> <span class="n">te_relu</span><span class="p">(</span><span class="n">X1</span><span class="p">)</span>
<span class="n">te</span><span class="o">.</span><span class="n">create_prim_func</span><span class="p">([</span><span class="n">X1</span><span class="p">,</span> <span class="n">Y1</span><span class="p">])</span><span class="o">.</span><span class="n">show</span><span class="p">()</span>
</pre></div>
</div>
<div class="highlight" style="background: "><pre style="line-height: 125%;"><span></span><span style="color: #007979; font-style: italic"># from tvm.script import tir as T</span>
<span style="color: #AA22FF">@T</span><span style="color: #AA22FF; font-weight: bold">.</span>prim_func
<span style="color: #008000; font-weight: bold">def</span> <span style="color: #0000FF">main</span>(X1: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((<span style="color: #008000">10</span>,), <span style="color: #BA2121">"float32"</span>), relu: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((<span style="color: #008000">10</span>,), <span style="color: #BA2121">"float32"</span>)):
T<span style="color: #AA22FF; font-weight: bold">.</span>func_attr({<span style="color: #BA2121">"tir.noalias"</span>: T<span style="color: #AA22FF; font-weight: bold">.</span>bool(<span style="color: #008000; font-weight: bold">True</span>)})
<span style="color: #007979; font-style: italic"># with T.block("root"):</span>
<span style="color: #008000; font-weight: bold">for</span> i0 <span style="color: #008000; font-weight: bold">in</span> range(<span style="color: #008000">10</span>):
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>block(<span style="color: #BA2121">"relu"</span>):
v_i0 <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>axis<span style="color: #AA22FF; font-weight: bold">.</span>spatial(<span style="color: #008000">10</span>, i0)
T<span style="color: #AA22FF; font-weight: bold">.</span>reads(X1[v_i0])
T<span style="color: #AA22FF; font-weight: bold">.</span>writes(relu[v_i0])
relu[v_i0] <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>max(X1[v_i0], T<span style="color: #AA22FF; font-weight: bold">.</span>float32(<span style="color: #008000">0.0</span>))
</pre></div><p>Then <code class="docutils literal notranslate"><span class="pre">X2</span></code> with shape <code class="docutils literal notranslate"><span class="pre">(10,</span> <span class="pre">20)</span></code>.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">X2</span> <span class="o">=</span> <span class="n">te</span><span class="o">.</span><span class="n">placeholder</span><span class="p">((</span><span class="mi">10</span><span class="p">,</span> <span class="mi">20</span><span class="p">),</span> <span class="n">name</span><span class="o">=</span><span class="s2">"X1"</span><span class="p">,</span> <span class="n">dtype</span><span class="o">=</span><span class="s2">"float32"</span><span class="p">)</span>
<span class="n">Y2</span> <span class="o">=</span> <span class="n">te_relu</span><span class="p">(</span><span class="n">X2</span><span class="p">)</span>
<span class="n">te</span><span class="o">.</span><span class="n">create_prim_func</span><span class="p">([</span><span class="n">X2</span><span class="p">,</span> <span class="n">Y2</span><span class="p">])</span><span class="o">.</span><span class="n">show</span><span class="p">()</span>
</pre></div>
</div>
<div class="highlight" style="background: "><pre style="line-height: 125%;"><span></span><span style="color: #007979; font-style: italic"># from tvm.script import tir as T</span>
<span style="color: #AA22FF">@T</span><span style="color: #AA22FF; font-weight: bold">.</span>prim_func
<span style="color: #008000; font-weight: bold">def</span> <span style="color: #0000FF">main</span>(X1: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((<span style="color: #008000">10</span>, <span style="color: #008000">20</span>), <span style="color: #BA2121">"float32"</span>), relu: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((<span style="color: #008000">10</span>, <span style="color: #008000">20</span>), <span style="color: #BA2121">"float32"</span>)):
T<span style="color: #AA22FF; font-weight: bold">.</span>func_attr({<span style="color: #BA2121">"tir.noalias"</span>: T<span style="color: #AA22FF; font-weight: bold">.</span>bool(<span style="color: #008000; font-weight: bold">True</span>)})
<span style="color: #007979; font-style: italic"># with T.block("root"):</span>
<span style="color: #008000; font-weight: bold">for</span> i0, i1 <span style="color: #008000; font-weight: bold">in</span> T<span style="color: #AA22FF; font-weight: bold">.</span>grid(<span style="color: #008000">10</span>, <span style="color: #008000">20</span>):
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>block(<span style="color: #BA2121">"relu"</span>):
v_i0, v_i1 <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>axis<span style="color: #AA22FF; font-weight: bold">.</span>remap(<span style="color: #BA2121">"SS"</span>, [i0, i1])
T<span style="color: #AA22FF; font-weight: bold">.</span>reads(X1[v_i0, v_i1])
T<span style="color: #AA22FF; font-weight: bold">.</span>writes(relu[v_i0, v_i1])
relu[v_i0, v_i1] <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>max(X1[v_i0, v_i1], T<span style="color: #AA22FF; font-weight: bold">.</span>float32(<span style="color: #008000">0.0</span>))
</pre></div><p>One final thing that <code class="docutils literal notranslate"><span class="pre">te</span></code> API allows us to do is to compose operations
and create “fused” operators. For example, we can take the result of
matmul and apply relu again.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">C</span> <span class="o">=</span> <span class="n">te_matmul</span><span class="p">(</span><span class="n">A</span><span class="p">,</span> <span class="n">B</span><span class="p">)</span>
<span class="n">D</span> <span class="o">=</span> <span class="n">te_relu</span><span class="p">(</span><span class="n">C</span><span class="p">)</span>
</pre></div>
</div>
<p>We can create a TensorIR function by only passing the input and output
values of interest, skipping intermediate values. This will cause the
result of matmul being allocated as a temp space in the TensorIR
function.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">te</span><span class="o">.</span><span class="n">create_prim_func</span><span class="p">([</span><span class="n">A</span><span class="p">,</span> <span class="n">B</span><span class="p">,</span> <span class="n">D</span><span class="p">])</span><span class="o">.</span><span class="n">show</span><span class="p">()</span>
</pre></div>
</div>
<div class="highlight" style="background: "><pre style="line-height: 125%;"><span></span><span style="color: #007979; font-style: italic"># from tvm.script import tir as T</span>
<span style="color: #AA22FF">@T</span><span style="color: #AA22FF; font-weight: bold">.</span>prim_func
<span style="color: #008000; font-weight: bold">def</span> <span style="color: #0000FF">main</span>(A: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((<span style="color: #008000">128</span>, <span style="color: #008000">128</span>), <span style="color: #BA2121">"float32"</span>), B: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((<span style="color: #008000">128</span>, <span style="color: #008000">128</span>), <span style="color: #BA2121">"float32"</span>), relu: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((<span style="color: #008000">128</span>, <span style="color: #008000">128</span>), <span style="color: #BA2121">"float32"</span>)):
T<span style="color: #AA22FF; font-weight: bold">.</span>func_attr({<span style="color: #BA2121">"tir.noalias"</span>: T<span style="color: #AA22FF; font-weight: bold">.</span>bool(<span style="color: #008000; font-weight: bold">True</span>)})
<span style="color: #007979; font-style: italic"># with T.block("root"):</span>
matmul <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>alloc_buffer((<span style="color: #008000">128</span>, <span style="color: #008000">128</span>))
<span style="color: #008000; font-weight: bold">for</span> i, j, k <span style="color: #008000; font-weight: bold">in</span> T<span style="color: #AA22FF; font-weight: bold">.</span>grid(<span style="color: #008000">128</span>, <span style="color: #008000">128</span>, <span style="color: #008000">128</span>):
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>block(<span style="color: #BA2121">"matmul"</span>):
v_i, v_j, v_k <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>axis<span style="color: #AA22FF; font-weight: bold">.</span>remap(<span style="color: #BA2121">"SSR"</span>, [i, j, k])
T<span style="color: #AA22FF; font-weight: bold">.</span>reads(A[v_i, v_k], B[v_k, v_j])
T<span style="color: #AA22FF; font-weight: bold">.</span>writes(matmul[v_i, v_j])
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>init():
matmul[v_i, v_j] <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>float32(<span style="color: #008000">0.0</span>)
matmul[v_i, v_j] <span style="color: #AA22FF; font-weight: bold">=</span> matmul[v_i, v_j] <span style="color: #AA22FF; font-weight: bold">+</span> A[v_i, v_k] <span style="color: #AA22FF; font-weight: bold">*</span> B[v_k, v_j]
<span style="color: #008000; font-weight: bold">for</span> i0, i1 <span style="color: #008000; font-weight: bold">in</span> T<span style="color: #AA22FF; font-weight: bold">.</span>grid(<span style="color: #008000">128</span>, <span style="color: #008000">128</span>):
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>block(<span style="color: #BA2121">"relu"</span>):
v_i0, v_i1 <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>axis<span style="color: #AA22FF; font-weight: bold">.</span>remap(<span style="color: #BA2121">"SS"</span>, [i0, i1])
T<span style="color: #AA22FF; font-weight: bold">.</span>reads(matmul[v_i0, v_i1])
T<span style="color: #AA22FF; font-weight: bold">.</span>writes(relu[v_i0, v_i1])
relu[v_i0, v_i1] <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>max(matmul[v_i0, v_i1], T<span style="color: #AA22FF; font-weight: bold">.</span>float32(<span style="color: #008000">0.0</span>))
</pre></div><p>We can also pass the intermediate result C into the argument list. In
this case, the TensorIR function expects us to also pass in the buffer
of C from the caller side. Normally we recommend only passing in the
input/output so we can have more advanced fusion inside.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">te</span><span class="o">.</span><span class="n">create_prim_func</span><span class="p">([</span><span class="n">A</span><span class="p">,</span> <span class="n">B</span><span class="p">,</span> <span class="n">C</span><span class="p">,</span> <span class="n">D</span><span class="p">])</span><span class="o">.</span><span class="n">show</span><span class="p">()</span>
</pre></div>
</div>
<div class="highlight" style="background: "><pre style="line-height: 125%;"><span></span><span style="color: #007979; font-style: italic"># from tvm.script import tir as T</span>
<span style="color: #AA22FF">@T</span><span style="color: #AA22FF; font-weight: bold">.</span>prim_func
<span style="color: #008000; font-weight: bold">def</span> <span style="color: #0000FF">main</span>(A: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((<span style="color: #008000">128</span>, <span style="color: #008000">128</span>), <span style="color: #BA2121">"float32"</span>), B: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((<span style="color: #008000">128</span>, <span style="color: #008000">128</span>), <span style="color: #BA2121">"float32"</span>), matmul: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((<span style="color: #008000">128</span>, <span style="color: #008000">128</span>), <span style="color: #BA2121">"float32"</span>), relu: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((<span style="color: #008000">128</span>, <span style="color: #008000">128</span>), <span style="color: #BA2121">"float32"</span>)):
T<span style="color: #AA22FF; font-weight: bold">.</span>func_attr({<span style="color: #BA2121">"tir.noalias"</span>: T<span style="color: #AA22FF; font-weight: bold">.</span>bool(<span style="color: #008000; font-weight: bold">True</span>)})
<span style="color: #007979; font-style: italic"># with T.block("root"):</span>
<span style="color: #008000; font-weight: bold">for</span> i, j, k <span style="color: #008000; font-weight: bold">in</span> T<span style="color: #AA22FF; font-weight: bold">.</span>grid(<span style="color: #008000">128</span>, <span style="color: #008000">128</span>, <span style="color: #008000">128</span>):
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>block(<span style="color: #BA2121">"matmul"</span>):
v_i, v_j, v_k <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>axis<span style="color: #AA22FF; font-weight: bold">.</span>remap(<span style="color: #BA2121">"SSR"</span>, [i, j, k])
T<span style="color: #AA22FF; font-weight: bold">.</span>reads(A[v_i, v_k], B[v_k, v_j])
T<span style="color: #AA22FF; font-weight: bold">.</span>writes(matmul[v_i, v_j])
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>init():
matmul[v_i, v_j] <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>float32(<span style="color: #008000">0.0</span>)
matmul[v_i, v_j] <span style="color: #AA22FF; font-weight: bold">=</span> matmul[v_i, v_j] <span style="color: #AA22FF; font-weight: bold">+</span> A[v_i, v_k] <span style="color: #AA22FF; font-weight: bold">*</span> B[v_k, v_j]
<span style="color: #008000; font-weight: bold">for</span> i0, i1 <span style="color: #008000; font-weight: bold">in</span> T<span style="color: #AA22FF; font-weight: bold">.</span>grid(<span style="color: #008000">128</span>, <span style="color: #008000">128</span>):
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>block(<span style="color: #BA2121">"relu"</span>):
v_i0, v_i1 <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>axis<span style="color: #AA22FF; font-weight: bold">.</span>remap(<span style="color: #BA2121">"SS"</span>, [i0, i1])
T<span style="color: #AA22FF; font-weight: bold">.</span>reads(matmul[v_i0, v_i1])
T<span style="color: #AA22FF; font-weight: bold">.</span>writes(relu[v_i0, v_i1])
relu[v_i0, v_i1] <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>max(matmul[v_i0, v_i1], T<span style="color: #AA22FF; font-weight: bold">.</span>float32(<span style="color: #008000">0.0</span>))
</pre></div></div>
<div class="section" id="use-blockbuilder-to-create-an-irmodule">
<h3><span class="section-number">5.3.2. </span>Use BlockBuilder to Create an IRModule<a class="headerlink" href="#use-blockbuilder-to-create-an-irmodule" title="Permalink to this heading">¶</a></h3>
<p>So far, we have created a single TensorIR function. In order to build
end-to-end model execution, we also need to be able to connect multiple
TensorIR functions through a computational graph.</p>
<p>Let us first create a block builder, which helps us incrementally build
a <code class="docutils literal notranslate"><span class="pre">relax.Function</span></code>.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">A</span> <span class="o">=</span> <span class="n">relax</span><span class="o">.</span><span class="n">Var</span><span class="p">(</span><span class="s2">"A"</span><span class="p">,</span> <span class="n">relax</span><span class="o">.</span><span class="n">TensorStructInfo</span><span class="p">((</span><span class="mi">128</span><span class="p">,</span> <span class="mi">128</span><span class="p">),</span> <span class="s2">"float32"</span><span class="p">))</span>
<span class="n">B</span> <span class="o">=</span> <span class="n">relax</span><span class="o">.</span><span class="n">Var</span><span class="p">(</span><span class="s2">"B"</span><span class="p">,</span> <span class="n">relax</span><span class="o">.</span><span class="n">TensorStructInfo</span><span class="p">((</span><span class="mi">128</span><span class="p">,</span> <span class="mi">128</span><span class="p">),</span> <span class="s2">"float32"</span><span class="p">))</span>
</pre></div>
</div>
<p>We construct the relax function by creating a block builder and then a
sequence of primitive tensor operations.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">bb</span> <span class="o">=</span> <span class="n">relax</span><span class="o">.</span><span class="n">BlockBuilder</span><span class="p">()</span>
<span class="k">with</span> <span class="n">bb</span><span class="o">.</span><span class="n">function</span><span class="p">(</span><span class="s2">"main"</span><span class="p">):</span>
<span class="k">with</span> <span class="n">bb</span><span class="o">.</span><span class="n">dataflow</span><span class="p">():</span>
<span class="n">C</span> <span class="o">=</span> <span class="n">bb</span><span class="o">.</span><span class="n">emit_te</span><span class="p">(</span><span class="n">te_matmul</span><span class="p">,</span> <span class="n">A</span><span class="p">,</span> <span class="n">B</span><span class="p">)</span>
<span class="n">D</span> <span class="o">=</span> <span class="n">bb</span><span class="o">.</span><span class="n">emit_te</span><span class="p">(</span><span class="n">te_relu</span><span class="p">,</span> <span class="n">C</span><span class="p">)</span>
<span class="n">R</span> <span class="o">=</span> <span class="n">bb</span><span class="o">.</span><span class="n">emit_output</span><span class="p">(</span><span class="n">D</span><span class="p">)</span>
<span class="n">bb</span><span class="o">.</span><span class="n">emit_func_output</span><span class="p">(</span><span class="n">R</span><span class="p">,</span> <span class="n">params</span><span class="o">=</span><span class="p">[</span><span class="n">A</span><span class="p">,</span> <span class="n">B</span><span class="p">])</span>
<span class="n">MyModule</span> <span class="o">=</span> <span class="n">bb</span><span class="o">.</span><span class="n">get</span><span class="p">()</span>
<span class="n">MyModule</span><span class="o">.</span><span class="n">show</span><span class="p">()</span>
</pre></div>
</div>
<div class="highlight" style="background: "><pre style="line-height: 125%;"><span></span><span style="color: #007979; font-style: italic"># from tvm.script import ir as I</span>
<span style="color: #007979; font-style: italic"># from tvm.script import tir as T</span>
<span style="color: #007979; font-style: italic"># from tvm.script import relax as R</span>
<span style="color: #AA22FF">@I</span><span style="color: #AA22FF; font-weight: bold">.</span>ir_module
<span style="color: #008000; font-weight: bold">class</span> <span style="color: #0000FF; font-weight: bold">Module</span>:
<span style="color: #AA22FF">@T</span><span style="color: #AA22FF; font-weight: bold">.</span>prim_func(private<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #008000; font-weight: bold">True</span>)
<span style="color: #008000; font-weight: bold">def</span> <span style="color: #0000FF">te_matmul</span>(A: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)), <span style="color: #BA2121">"float32"</span>), B: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)), <span style="color: #BA2121">"float32"</span>), matmul: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)), <span style="color: #BA2121">"float32"</span>)):
T<span style="color: #AA22FF; font-weight: bold">.</span>func_attr({<span style="color: #BA2121">"tir.noalias"</span>: T<span style="color: #AA22FF; font-weight: bold">.</span>bool(<span style="color: #008000; font-weight: bold">True</span>)})
<span style="color: #007979; font-style: italic"># with T.block("root"):</span>
<span style="color: #008000; font-weight: bold">for</span> i, j, k <span style="color: #008000; font-weight: bold">in</span> T<span style="color: #AA22FF; font-weight: bold">.</span>grid(T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)):
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>block(<span style="color: #BA2121">"matmul"</span>):
v_i, v_j, v_k <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>axis<span style="color: #AA22FF; font-weight: bold">.</span>remap(<span style="color: #BA2121">"SSR"</span>, [i, j, k])
T<span style="color: #AA22FF; font-weight: bold">.</span>reads(A[v_i, v_k], B[v_k, v_j])
T<span style="color: #AA22FF; font-weight: bold">.</span>writes(matmul[v_i, v_j])
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>init():
matmul[v_i, v_j] <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>float32(<span style="color: #008000">0.0</span>)
matmul[v_i, v_j] <span style="color: #AA22FF; font-weight: bold">=</span> matmul[v_i, v_j] <span style="color: #AA22FF; font-weight: bold">+</span> A[v_i, v_k] <span style="color: #AA22FF; font-weight: bold">*</span> B[v_k, v_j]
<span style="color: #AA22FF">@T</span><span style="color: #AA22FF; font-weight: bold">.</span>prim_func(private<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #008000; font-weight: bold">True</span>)
<span style="color: #008000; font-weight: bold">def</span> <span style="color: #0000FF">te_relu</span>(lv: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)), <span style="color: #BA2121">"float32"</span>), relu: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)), <span style="color: #BA2121">"float32"</span>)):
T<span style="color: #AA22FF; font-weight: bold">.</span>func_attr({<span style="color: #BA2121">"tir.noalias"</span>: T<span style="color: #AA22FF; font-weight: bold">.</span>bool(<span style="color: #008000; font-weight: bold">True</span>)})
<span style="color: #007979; font-style: italic"># with T.block("root"):</span>
<span style="color: #008000; font-weight: bold">for</span> i0, i1 <span style="color: #008000; font-weight: bold">in</span> T<span style="color: #AA22FF; font-weight: bold">.</span>grid(T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)):
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>block(<span style="color: #BA2121">"relu"</span>):
v_i0, v_i1 <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>axis<span style="color: #AA22FF; font-weight: bold">.</span>remap(<span style="color: #BA2121">"SS"</span>, [i0, i1])
T<span style="color: #AA22FF; font-weight: bold">.</span>reads(lv[v_i0, v_i1])
T<span style="color: #AA22FF; font-weight: bold">.</span>writes(relu[v_i0, v_i1])
relu[v_i0, v_i1] <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>max(lv[v_i0, v_i1], T<span style="color: #AA22FF; font-weight: bold">.</span>float32(<span style="color: #008000">0.0</span>))
<span style="color: #AA22FF">@R</span><span style="color: #AA22FF; font-weight: bold">.</span>function
<span style="color: #008000; font-weight: bold">def</span> <span style="color: #0000FF">main</span>(A: R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">128</span>, <span style="color: #008000">128</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>), B: R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">128</span>, <span style="color: #008000">128</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>)) <span style="color: #AA22FF; font-weight: bold">-></span> R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">128</span>, <span style="color: #008000">128</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>):
cls <span style="color: #AA22FF; font-weight: bold">=</span> Module
<span style="color: #008000; font-weight: bold">with</span> R<span style="color: #AA22FF; font-weight: bold">.</span>dataflow():
lv <span style="color: #AA22FF; font-weight: bold">=</span> R<span style="color: #AA22FF; font-weight: bold">.</span>call_tir(cls<span style="color: #AA22FF; font-weight: bold">.</span>te_matmul, (A, B), out_sinfo<span style="color: #AA22FF; font-weight: bold">=</span>R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">128</span>, <span style="color: #008000">128</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>))
lv1 <span style="color: #AA22FF; font-weight: bold">=</span> R<span style="color: #AA22FF; font-weight: bold">.</span>call_tir(cls<span style="color: #AA22FF; font-weight: bold">.</span>te_relu, (lv,), out_sinfo<span style="color: #AA22FF; font-weight: bold">=</span>R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">128</span>, <span style="color: #008000">128</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>))
gv: R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">128</span>, <span style="color: #008000">128</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>) <span style="color: #AA22FF; font-weight: bold">=</span> lv1
R<span style="color: #AA22FF; font-weight: bold">.</span>output(gv)
<span style="color: #008000; font-weight: bold">return</span> gv
</pre></div></div>
<div class="section" id="deep-dive-into-block-builder-apis">
<h3><span class="section-number">5.3.3. </span>Deep Dive into Block Builder APIs<a class="headerlink" href="#deep-dive-into-block-builder-apis" title="Permalink to this heading">¶</a></h3>
<p>Now let us do a deep dive into each block builder API. It is helpful to
put the block builder code and the resulting module side by side.</p>
<div class="figure align-default">
<img alt="../_images/integration_block_builder.png" src="../_images/integration_block_builder.png" />
</div>
<p>The block builder comes with scopes that correspond to the scopes in the
relax function. For example, <code class="docutils literal notranslate"><span class="pre">bb.dataflow()</span></code> creates a dataflow block
where all the block builder calls inside the scope belonging to the
dataflow scope.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">with</span> <span class="n">bb</span><span class="o">.</span><span class="n">function</span><span class="p">(</span><span class="s2">"main"</span><span class="p">):</span>
<span class="k">with</span> <span class="n">bb</span><span class="o">.</span><span class="n">dataflow</span><span class="p">():</span>
<span class="c1"># every emit call generates a variable inside a dataflow block.</span>
</pre></div>
</div>
<p>Each intermediate result is a <code class="docutils literal notranslate"><span class="pre">relax.Var</span></code> corresponding to a variable
that stores the result of the computation. <code class="docutils literal notranslate"><span class="pre">DataflowVar</span></code> indicates
that the var is an intermediate step inside a dataflow block
(computational graph).</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="nb">type</span><span class="p">(</span><span class="n">C</span><span class="p">)</span>
</pre></div>
</div>
<div class="output highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">tvm</span><span class="o">.</span><span class="n">relax</span><span class="o">.</span><span class="n">expr</span><span class="o">.</span><span class="n">DataflowVar</span>
</pre></div>
</div>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="nb">isinstance</span><span class="p">(</span><span class="n">C</span><span class="p">,</span> <span class="n">relax</span><span class="o">.</span><span class="n">Var</span><span class="p">)</span>
</pre></div>
</div>
<div class="output highlight-default notranslate"><div class="highlight"><pre><span></span><span class="kc">True</span>
</pre></div>
</div>
<p>Each line in the relax function is generated by an <code class="docutils literal notranslate"><span class="pre">emit_te</span></code> call. For
example,</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">lv</span> <span class="o">=</span> <span class="n">R</span><span class="o">.</span><span class="n">call_dps_packed</span><span class="p">(</span><span class="n">te_matmul</span><span class="p">,</span> <span class="p">(</span><span class="n">A</span><span class="p">,</span> <span class="n">B</span><span class="p">),</span> <span class="p">(</span><span class="mi">128</span><span class="p">,</span> <span class="mi">128</span><span class="p">),</span> <span class="n">dtype</span><span class="o">=</span><span class="s2">"float32"</span><span class="p">)</span>
</pre></div>
</div>
<p>is generated by</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">C</span> <span class="o">=</span> <span class="n">bb</span><span class="o">.</span><span class="n">emit_te</span><span class="p">(</span><span class="n">te_matmul</span><span class="p">,</span> <span class="n">A</span><span class="p">,</span> <span class="n">B</span><span class="p">)</span><span class="o">.</span>
</pre></div>
</div>
<p>Under the hood, the bb.emit_te does the following things:</p>
<ul class="simple">
<li><p>Create an input <code class="docutils literal notranslate"><span class="pre">te.placeholder</span></code> for A and B</p></li>
<li><p>Run them through <code class="docutils literal notranslate"><span class="pre">te_matmul</span></code> function.</p></li>
<li><p>Call into <code class="docutils literal notranslate"><span class="pre">te.create_prim_func</span></code> to create a TensorIR function.</p></li>
<li><p>Generate a call into the function via <code class="docutils literal notranslate"><span class="pre">call_dps_packed</span></code>.</p></li>
</ul>
<p>We can find that the result is a computational graph with two
intermediate values, with one node corresponding to the te_matmul
operation and another one corresponding to <code class="docutils literal notranslate"><span class="pre">te_relu</span></code>.</p>
<p>We can create output variable of each dataflow block through
<code class="docutils literal notranslate"><span class="pre">bb.emit_output</span></code>.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">with</span> <span class="n">bb</span><span class="o">.</span><span class="n">dataflow</span><span class="p">():</span>
<span class="o">...</span>
<span class="n">R</span> <span class="o">=</span> <span class="n">bb</span><span class="o">.</span><span class="n">emit_output</span><span class="p">(</span><span class="n">D</span><span class="p">)</span>
</pre></div>
</div>
<p>The above code marks that D is a variable that can be referred to
outside of the dataflow block.</p>
<p>Finally, the function output is marked by <code class="docutils literal notranslate"><span class="pre">bb.emit_func_output</span></code>. We
can only call <code class="docutils literal notranslate"><span class="pre">emit_func_output</span></code> once in each function scope.</p>
<p>Notably, we can specify the list of parameters of the function in the
output emission stage. Doing so helps us in cases where we collect the
list of parameters on the fly.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">with</span> <span class="n">bb</span><span class="o">.</span><span class="n">function</span><span class="p">(</span><span class="s2">"main"</span><span class="p">):</span>
<span class="o">...</span>
<span class="c1"># specify parameters in the end</span>
<span class="n">bb</span><span class="o">.</span><span class="n">emit_func_output</span><span class="p">(</span><span class="n">R</span><span class="p">,</span> <span class="n">params</span><span class="o">=</span><span class="p">[</span><span class="n">A</span><span class="p">,</span> <span class="n">B</span><span class="p">])</span>
</pre></div>
</div>
<p>Alternatively, we can specify the list of parameters at the beginning of
the function scope.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="c1"># specify parameters in the beginning.</span>
<span class="k">with</span> <span class="n">bb</span><span class="o">.</span><span class="n">function</span><span class="p">(</span><span class="s2">"main"</span><span class="p">,</span> <span class="n">params</span><span class="o">=</span><span class="p">[</span><span class="n">A</span><span class="p">,</span> <span class="n">B</span><span class="p">]):</span>
<span class="o">...</span>
<span class="n">bb</span><span class="o">.</span><span class="n">emit_func_output</span><span class="p">(</span><span class="n">R</span><span class="p">)</span>
</pre></div>
</div>
</div>
</div>
<div class="section" id="import-model-from-pytorch">
<h2><span class="section-number">5.4. </span>Import Model From PyTorch<a class="headerlink" href="#import-model-from-pytorch" title="Permalink to this heading">¶</a></h2>
<p>Now that we have learned the tools to construct an IRModule
programmatically. Let us use them to bring a model from PyTorch into the
IRModule format.</p>
<p>Most machine learning framework comes with computational graph
abstractions, where each node corresponds to an operation, and the edges
correspond to the dependency among them. We will take a PyTorch model,
obtain a computational graph in PyTorch’s native format, and translate
that into IRModule.</p>
<p>Let us begin by defining a model in PyTorch. To keep the example
consistent, we will use matmul relu example.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">MyModel</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="nb">super</span><span class="p">(</span><span class="n">MyModel</span><span class="p">,</span> <span class="bp">self</span><span class="p">)</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">weight</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Parameter</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">randn</span><span class="p">(</span><span class="mi">128</span><span class="p">,</span> <span class="mi">128</span><span class="p">))</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">matmul</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="bp">self</span><span class="o">.</span><span class="n">weight</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="k">return</span> <span class="n">x</span>
</pre></div>
</div>
<div class="section" id="create-torchfx-graphmodule">
<h3><span class="section-number">5.4.1. </span>Create TorchFX GraphModule<a class="headerlink" href="#create-torchfx-graphmodule" title="Permalink to this heading">¶</a></h3>
<p>We use TorchFX to trace a graph from the PyTorch module.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">model</span> <span class="o">=</span> <span class="n">MyModel</span><span class="p">()</span>
<span class="n">fx_module</span> <span class="o">=</span> <span class="n">fx</span><span class="o">.</span><span class="n">symbolic_trace</span><span class="p">(</span><span class="n">model</span><span class="p">)</span>
<span class="nb">type</span><span class="p">(</span><span class="n">fx_module</span><span class="p">)</span>
</pre></div>
</div>
<div class="output highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">torch</span><span class="o">.</span><span class="n">fx</span><span class="o">.</span><span class="n">graph_module</span><span class="o">.</span><span class="n">GraphModule</span><span class="o">.</span><span class="fm">__new__</span><span class="o">.<</span><span class="nb">locals</span><span class="o">>.</span><span class="n">GraphModuleImpl</span>
</pre></div>
</div>
<p>The <code class="docutils literal notranslate"><span class="pre">fx_module</span></code> contains a simple computation graph view that can be
printed as tabular data. Our goal is to translate this graph into an
IRModule.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">fx_module</span><span class="o">.</span><span class="n">graph</span><span class="o">.</span><span class="n">print_tabular</span><span class="p">()</span>
</pre></div>
</div>
<div class="output highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">opcode</span> <span class="n">name</span> <span class="n">target</span> <span class="n">args</span> <span class="n">kwargs</span>
<span class="o">-------------</span> <span class="o">------</span> <span class="o">---------------------------------------------------------</span> <span class="o">-----------</span> <span class="o">--------</span>
<span class="n">placeholder</span> <span class="n">x</span> <span class="n">x</span> <span class="p">()</span> <span class="p">{}</span>
<span class="n">get_attr</span> <span class="n">weight</span> <span class="n">weight</span> <span class="p">()</span> <span class="p">{}</span>
<span class="n">call_function</span> <span class="n">matmul</span> <span class="o"><</span><span class="n">built</span><span class="o">-</span><span class="ow">in</span> <span class="n">method</span> <span class="n">matmul</span> <span class="n">of</span> <span class="nb">type</span> <span class="nb">object</span> <span class="n">at</span> <span class="mh">0x7f28e7c5c480</span><span class="o">></span> <span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">weight</span><span class="p">)</span> <span class="p">{}</span>
<span class="n">call_function</span> <span class="n">relu</span> <span class="o"><</span><span class="n">built</span><span class="o">-</span><span class="ow">in</span> <span class="n">method</span> <span class="n">relu</span> <span class="n">of</span> <span class="nb">type</span> <span class="nb">object</span> <span class="n">at</span> <span class="mh">0x7f28e7c5c480</span><span class="o">></span> <span class="p">(</span><span class="n">matmul</span><span class="p">,)</span> <span class="p">{}</span>
<span class="n">output</span> <span class="n">output</span> <span class="n">output</span> <span class="p">(</span><span class="n">relu</span><span class="p">,)</span> <span class="p">{}</span>
</pre></div>
</div>
</div>
<div class="section" id="create-map-function">
<h3><span class="section-number">5.4.2. </span>Create Map Function<a class="headerlink" href="#create-map-function" title="Permalink to this heading">¶</a></h3>
<p>Let us define the overall high-level translation logic. The main flow is
as follows:</p>
<ul class="simple">
<li><p>Create a <code class="docutils literal notranslate"><span class="pre">node_map</span></code> that maps <code class="docutils literal notranslate"><span class="pre">fx.Node</span></code> to the corresponding
<code class="docutils literal notranslate"><span class="pre">relax.Var</span></code> that represents the translated node in IRModule.</p></li>
<li><p>Iterate over the nodes in the fx graph in topological order.</p></li>
<li><p>Compute the mapped output of the node given the mapped inputs.</p></li>
</ul>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">map_param</span><span class="p">(</span><span class="n">param</span><span class="p">:</span> <span class="n">nn</span><span class="o">.</span><span class="n">Parameter</span><span class="p">):</span>
<span class="k">return</span> <span class="n">relax</span><span class="o">.</span><span class="n">const</span><span class="p">(</span>
<span class="n">param</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">cpu</span><span class="p">()</span><span class="o">.</span><span class="n">numpy</span><span class="p">(),</span> <span class="n">relax</span><span class="o">.</span><span class="n">TensorStructInfo</span><span class="p">(</span><span class="n">param</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">shape</span><span class="p">,</span> <span class="s2">"float32"</span><span class="p">)</span>
<span class="p">)</span>
<span class="k">def</span> <span class="nf">fetch_attr</span><span class="p">(</span><span class="n">fx_mod</span><span class="p">,</span> <span class="n">target</span><span class="p">:</span> <span class="nb">str</span><span class="p">):</span>
<span class="w"> </span><span class="sd">"""Helper function to fetch an attr"""</span>
<span class="n">target_atoms</span> <span class="o">=</span> <span class="n">target</span><span class="o">.</span><span class="n">split</span><span class="p">(</span><span class="s1">'.'</span><span class="p">)</span>
<span class="n">attr_itr</span> <span class="o">=</span> <span class="n">fx_mod</span>
<span class="k">for</span> <span class="n">i</span><span class="p">,</span> <span class="n">atom</span> <span class="ow">in</span> <span class="nb">enumerate</span><span class="p">(</span><span class="n">target_atoms</span><span class="p">):</span>
<span class="k">if</span> <span class="ow">not</span> <span class="nb">hasattr</span><span class="p">(</span><span class="n">attr_itr</span><span class="p">,</span> <span class="n">atom</span><span class="p">):</span>
<span class="k">raise</span> <span class="ne">RuntimeError</span><span class="p">(</span><span class="sa">f</span><span class="s2">"Node referenced nonexistant target </span><span class="si">{</span><span class="s1">'.'</span><span class="o">.</span><span class="n">join</span><span class="p">(</span><span class="n">target_atoms</span><span class="p">[:</span><span class="n">i</span><span class="p">])</span><span class="si">}</span><span class="s2">"</span><span class="p">)</span>
<span class="n">attr_itr</span> <span class="o">=</span> <span class="nb">getattr</span><span class="p">(</span><span class="n">attr_itr</span><span class="p">,</span> <span class="n">atom</span><span class="p">)</span>
<span class="k">return</span> <span class="n">attr_itr</span>
<span class="k">def</span> <span class="nf">from_fx</span><span class="p">(</span><span class="n">fx_mod</span><span class="p">,</span> <span class="n">input_shapes</span><span class="p">,</span> <span class="n">call_function_map</span><span class="p">,</span> <span class="n">call_module_map</span><span class="p">):</span>
<span class="n">input_index</span> <span class="o">=</span> <span class="mi">0</span>
<span class="n">node_map</span> <span class="o">=</span> <span class="p">{}</span>
<span class="n">named_modules</span> <span class="o">=</span> <span class="nb">dict</span><span class="p">(</span><span class="n">fx_mod</span><span class="o">.</span><span class="n">named_modules</span><span class="p">())</span>
<span class="n">bb</span> <span class="o">=</span> <span class="n">relax</span><span class="o">.</span><span class="n">BlockBuilder</span><span class="p">()</span>
<span class="n">fn_inputs</span> <span class="o">=</span> <span class="p">[]</span>
<span class="n">fn_output</span> <span class="o">=</span> <span class="kc">None</span>
<span class="k">with</span> <span class="n">bb</span><span class="o">.</span><span class="n">function</span><span class="p">(</span><span class="s2">"main"</span><span class="p">):</span>
<span class="k">with</span> <span class="n">bb</span><span class="o">.</span><span class="n">dataflow</span><span class="p">():</span>
<span class="k">for</span> <span class="n">node</span> <span class="ow">in</span> <span class="n">fx_mod</span><span class="o">.</span><span class="n">graph</span><span class="o">.</span><span class="n">nodes</span><span class="p">:</span>
<span class="k">if</span> <span class="n">node</span><span class="o">.</span><span class="n">op</span> <span class="o">==</span> <span class="s2">"placeholder"</span><span class="p">:</span>
<span class="c1"># create input placeholder</span>
<span class="n">shape</span> <span class="o">=</span> <span class="n">input_shapes</span><span class="p">[</span><span class="n">input_index</span><span class="p">]</span>
<span class="n">input_index</span> <span class="o">+=</span> <span class="mi">1</span>
<span class="n">input_var</span> <span class="o">=</span> <span class="n">relax</span><span class="o">.</span><span class="n">Var</span><span class="p">(</span>
<span class="n">node</span><span class="o">.</span><span class="n">target</span><span class="p">,</span> <span class="n">relax</span><span class="o">.</span><span class="n">TensorStructInfo</span><span class="p">(</span><span class="n">shape</span><span class="p">,</span> <span class="s2">"float32"</span><span class="p">)</span>
<span class="p">)</span>
<span class="n">fn_inputs</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">input_var</span><span class="p">)</span>
<span class="n">node_map</span><span class="p">[</span><span class="n">node</span><span class="p">]</span> <span class="o">=</span> <span class="n">input_var</span>
<span class="k">elif</span> <span class="n">node</span><span class="o">.</span><span class="n">op</span> <span class="o">==</span> <span class="s2">"get_attr"</span><span class="p">:</span>
<span class="n">node_map</span><span class="p">[</span><span class="n">node</span><span class="p">]</span> <span class="o">=</span> <span class="n">map_param</span><span class="p">(</span><span class="n">fetch_attr</span><span class="p">(</span><span class="n">fx_mod</span><span class="p">,</span> <span class="n">node</span><span class="o">.</span><span class="n">target</span><span class="p">))</span>
<span class="k">elif</span> <span class="n">node</span><span class="o">.</span><span class="n">op</span> <span class="o">==</span> <span class="s2">"call_function"</span><span class="p">:</span>
<span class="n">node_map</span><span class="p">[</span><span class="n">node</span><span class="p">]</span> <span class="o">=</span> <span class="n">call_function_map</span><span class="p">[</span><span class="n">node</span><span class="o">.</span><span class="n">target</span><span class="p">](</span><span class="n">bb</span><span class="p">,</span> <span class="n">node_map</span><span class="p">,</span> <span class="n">node</span><span class="p">)</span>
<span class="k">elif</span> <span class="n">node</span><span class="o">.</span><span class="n">op</span> <span class="o">==</span> <span class="s2">"call_module"</span><span class="p">:</span>
<span class="n">named_module</span> <span class="o">=</span> <span class="n">named_modules</span><span class="p">[</span><span class="n">node</span><span class="o">.</span><span class="n">target</span><span class="p">]</span>
<span class="n">node_map</span><span class="p">[</span><span class="n">node</span><span class="p">]</span> <span class="o">=</span> <span class="n">call_module_map</span><span class="p">[</span><span class="nb">type</span><span class="p">(</span><span class="n">named_module</span><span class="p">)](</span><span class="n">bb</span><span class="p">,</span> <span class="n">node_map</span><span class="p">,</span> <span class="n">node</span><span class="p">,</span> <span class="n">named_module</span><span class="p">)</span>
<span class="k">elif</span> <span class="n">node</span><span class="o">.</span><span class="n">op</span> <span class="o">==</span> <span class="s2">"output"</span><span class="p">:</span>
<span class="n">output</span> <span class="o">=</span> <span class="n">node_map</span><span class="p">[</span><span class="n">node</span><span class="o">.</span><span class="n">args</span><span class="p">[</span><span class="mi">0</span><span class="p">]]</span>
<span class="k">assert</span> <span class="n">fn_output</span> <span class="ow">is</span> <span class="kc">None</span>
<span class="n">fn_output</span> <span class="o">=</span> <span class="n">bb</span><span class="o">.</span><span class="n">emit_output</span><span class="p">(</span><span class="n">output</span><span class="p">)</span>
<span class="c1"># output and finalize the function</span>
<span class="n">bb</span><span class="o">.</span><span class="n">emit_func_output</span><span class="p">(</span><span class="n">output</span><span class="p">,</span> <span class="n">fn_inputs</span><span class="p">)</span>
<span class="k">return</span> <span class="n">bb</span><span class="o">.</span><span class="n">get</span><span class="p">()</span>
</pre></div>
</div>
<p>We did not define the function map in the <code class="docutils literal notranslate"><span class="pre">from_fx</span></code> function. We will
supply the translation rule of each torch function via a map.
Specifically, the following code block shows how we can do that through
the <code class="docutils literal notranslate"><span class="pre">emit_te</span></code> API.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">map_matmul</span><span class="p">(</span><span class="n">bb</span><span class="p">,</span> <span class="n">node_map</span><span class="p">,</span> <span class="n">node</span><span class="p">:</span> <span class="n">fx</span><span class="o">.</span><span class="n">Node</span><span class="p">):</span>
<span class="n">A</span> <span class="o">=</span> <span class="n">node_map</span><span class="p">[</span><span class="n">node</span><span class="o">.</span><span class="n">args</span><span class="p">[</span><span class="mi">0</span><span class="p">]]</span>
<span class="n">B</span> <span class="o">=</span> <span class="n">node_map</span><span class="p">[</span><span class="n">node</span><span class="o">.</span><span class="n">args</span><span class="p">[</span><span class="mi">1</span><span class="p">]]</span>
<span class="k">return</span> <span class="n">bb</span><span class="o">.</span><span class="n">emit_te</span><span class="p">(</span><span class="n">te_matmul</span><span class="p">,</span> <span class="n">A</span><span class="p">,</span> <span class="n">B</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">map_relu</span><span class="p">(</span><span class="n">bb</span><span class="p">,</span> <span class="n">node_map</span><span class="p">,</span> <span class="n">node</span><span class="p">:</span> <span class="n">fx</span><span class="o">.</span><span class="n">Node</span><span class="p">):</span>
<span class="n">A</span> <span class="o">=</span> <span class="n">node_map</span><span class="p">[</span><span class="n">node</span><span class="o">.</span><span class="n">args</span><span class="p">[</span><span class="mi">0</span><span class="p">]]</span>
<span class="k">return</span> <span class="n">bb</span><span class="o">.</span><span class="n">emit_te</span><span class="p">(</span><span class="n">te_relu</span><span class="p">,</span> <span class="n">A</span><span class="p">)</span>
<span class="n">MyModule</span> <span class="o">=</span> <span class="n">from_fx</span><span class="p">(</span>
<span class="n">fx_module</span><span class="p">,</span>
<span class="n">input_shapes</span> <span class="o">=</span> <span class="p">[(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">128</span><span class="p">)],</span>
<span class="n">call_function_map</span> <span class="o">=</span> <span class="p">{</span>
<span class="n">torch</span><span class="o">.</span><span class="n">matmul</span><span class="p">:</span> <span class="n">map_matmul</span><span class="p">,</span>
<span class="n">torch</span><span class="o">.</span><span class="n">relu</span><span class="p">:</span> <span class="n">map_relu</span><span class="p">,</span>
<span class="p">},</span>
<span class="n">call_module_map</span><span class="o">=</span><span class="p">{},</span>
<span class="p">)</span>
<span class="n">MyModule</span><span class="o">.</span><span class="n">show</span><span class="p">()</span>
</pre></div>
</div>
<div class="highlight" style="background: "><pre style="line-height: 125%;"><span></span><span style="color: #007979; font-style: italic"># from tvm.script import ir as I</span>
<span style="color: #007979; font-style: italic"># from tvm.script import tir as T</span>
<span style="color: #007979; font-style: italic"># from tvm.script import relax as R</span>
<span style="color: #AA22FF">@I</span><span style="color: #AA22FF; font-weight: bold">.</span>ir_module
<span style="color: #008000; font-weight: bold">class</span> <span style="color: #0000FF; font-weight: bold">Module</span>:
<span style="color: #AA22FF">@T</span><span style="color: #AA22FF; font-weight: bold">.</span>prim_func(private<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #008000; font-weight: bold">True</span>)
<span style="color: #008000; font-weight: bold">def</span> <span style="color: #0000FF">te_matmul</span>(x: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)), <span style="color: #BA2121">"float32"</span>), B: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)), <span style="color: #BA2121">"float32"</span>), matmul: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)), <span style="color: #BA2121">"float32"</span>)):
T<span style="color: #AA22FF; font-weight: bold">.</span>func_attr({<span style="color: #BA2121">"tir.noalias"</span>: T<span style="color: #AA22FF; font-weight: bold">.</span>bool(<span style="color: #008000; font-weight: bold">True</span>)})
<span style="color: #007979; font-style: italic"># with T.block("root"):</span>
<span style="color: #008000; font-weight: bold">for</span> i, j, k <span style="color: #008000; font-weight: bold">in</span> T<span style="color: #AA22FF; font-weight: bold">.</span>grid(T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)):
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>block(<span style="color: #BA2121">"matmul"</span>):
v_i, v_j, v_k <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>axis<span style="color: #AA22FF; font-weight: bold">.</span>remap(<span style="color: #BA2121">"SSR"</span>, [i, j, k])
T<span style="color: #AA22FF; font-weight: bold">.</span>reads(x[v_i, v_k], B[v_k, v_j])
T<span style="color: #AA22FF; font-weight: bold">.</span>writes(matmul[v_i, v_j])
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>init():
matmul[v_i, v_j] <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>float32(<span style="color: #008000">0.0</span>)
matmul[v_i, v_j] <span style="color: #AA22FF; font-weight: bold">=</span> matmul[v_i, v_j] <span style="color: #AA22FF; font-weight: bold">+</span> x[v_i, v_k] <span style="color: #AA22FF; font-weight: bold">*</span> B[v_k, v_j]
<span style="color: #AA22FF">@T</span><span style="color: #AA22FF; font-weight: bold">.</span>prim_func(private<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #008000; font-weight: bold">True</span>)
<span style="color: #008000; font-weight: bold">def</span> <span style="color: #0000FF">te_relu</span>(lv: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)), <span style="color: #BA2121">"float32"</span>), relu: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)), <span style="color: #BA2121">"float32"</span>)):
T<span style="color: #AA22FF; font-weight: bold">.</span>func_attr({<span style="color: #BA2121">"tir.noalias"</span>: T<span style="color: #AA22FF; font-weight: bold">.</span>bool(<span style="color: #008000; font-weight: bold">True</span>)})
<span style="color: #007979; font-style: italic"># with T.block("root"):</span>
<span style="color: #008000; font-weight: bold">for</span> i0, i1 <span style="color: #008000; font-weight: bold">in</span> T<span style="color: #AA22FF; font-weight: bold">.</span>grid(T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)):
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>block(<span style="color: #BA2121">"relu"</span>):
v_i0, v_i1 <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>axis<span style="color: #AA22FF; font-weight: bold">.</span>remap(<span style="color: #BA2121">"SS"</span>, [i0, i1])
T<span style="color: #AA22FF; font-weight: bold">.</span>reads(lv[v_i0, v_i1])
T<span style="color: #AA22FF; font-weight: bold">.</span>writes(relu[v_i0, v_i1])
relu[v_i0, v_i1] <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>max(lv[v_i0, v_i1], T<span style="color: #AA22FF; font-weight: bold">.</span>float32(<span style="color: #008000">0.0</span>))
<span style="color: #AA22FF">@R</span><span style="color: #AA22FF; font-weight: bold">.</span>function
<span style="color: #008000; font-weight: bold">def</span> <span style="color: #0000FF">main</span>(x: R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">128</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>)) <span style="color: #AA22FF; font-weight: bold">-></span> R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">128</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>):
cls <span style="color: #AA22FF; font-weight: bold">=</span> Module
<span style="color: #008000; font-weight: bold">with</span> R<span style="color: #AA22FF; font-weight: bold">.</span>dataflow():
lv <span style="color: #AA22FF; font-weight: bold">=</span> R<span style="color: #AA22FF; font-weight: bold">.</span>call_tir(cls<span style="color: #AA22FF; font-weight: bold">.</span>te_matmul, (x, metadata[<span style="color: #BA2121">"relax.expr.Constant"</span>][<span style="color: #008000">0</span>]), out_sinfo<span style="color: #AA22FF; font-weight: bold">=</span>R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">128</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>))
lv1 <span style="color: #AA22FF; font-weight: bold">=</span> R<span style="color: #AA22FF; font-weight: bold">.</span>call_tir(cls<span style="color: #AA22FF; font-weight: bold">.</span>te_relu, (lv,), out_sinfo<span style="color: #AA22FF; font-weight: bold">=</span>R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">128</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>))
gv: R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">128</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>) <span style="color: #AA22FF; font-weight: bold">=</span> lv1
R<span style="color: #AA22FF; font-weight: bold">.</span>output(gv)
<span style="color: #008000; font-weight: bold">return</span> lv1
<span style="color: #007979; font-style: italic"># Metadata omitted. Use show_meta=True in script() method to show it.</span>
</pre></div></div>
</div>
<div class="section" id="coming-back-to-fashionmnist-example">
<h2><span class="section-number">5.5. </span>Coming back to FashionMNIST Example<a class="headerlink" href="#coming-back-to-fashionmnist-example" title="Permalink to this heading">¶</a></h2>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">torch</span>
<span class="kn">import</span> <span class="nn">torchvision</span>
<span class="n">test_data</span> <span class="o">=</span> <span class="n">torchvision</span><span class="o">.</span><span class="n">datasets</span><span class="o">.</span><span class="n">FashionMNIST</span><span class="p">(</span>
<span class="n">root</span><span class="o">=</span><span class="s2">"data"</span><span class="p">,</span>
<span class="n">train</span><span class="o">=</span><span class="kc">False</span><span class="p">,</span>
<span class="n">download</span><span class="o">=</span><span class="kc">True</span><span class="p">,</span>
<span class="n">transform</span><span class="o">=</span><span class="n">torchvision</span><span class="o">.</span><span class="n">transforms</span><span class="o">.</span><span class="n">ToTensor</span><span class="p">()</span>
<span class="p">)</span>
<span class="n">test_loader</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">utils</span><span class="o">.</span><span class="n">data</span><span class="o">.</span><span class="n">DataLoader</span><span class="p">(</span><span class="n">test_data</span><span class="p">,</span> <span class="n">batch_size</span><span class="o">=</span><span class="mi">1</span><span class="p">,</span> <span class="n">shuffle</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="n">class_names</span> <span class="o">=</span> <span class="p">[</span><span class="s1">'T-shirt/top'</span><span class="p">,</span> <span class="s1">'Trouser'</span><span class="p">,</span> <span class="s1">'Pullover'</span><span class="p">,</span> <span class="s1">'Dress'</span><span class="p">,</span> <span class="s1">'Coat'</span><span class="p">,</span>
<span class="s1">'Sandal'</span><span class="p">,</span> <span class="s1">'Shirt'</span><span class="p">,</span> <span class="s1">'Sneaker'</span><span class="p">,</span> <span class="s1">'Bag'</span><span class="p">,</span> <span class="s1">'Ankle boot'</span><span class="p">]</span>
<span class="n">img</span><span class="p">,</span> <span class="n">label</span> <span class="o">=</span> <span class="nb">next</span><span class="p">(</span><span class="nb">iter</span><span class="p">(</span><span class="n">test_loader</span><span class="p">))</span>
<span class="n">img</span> <span class="o">=</span> <span class="n">img</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">28</span><span class="p">,</span> <span class="mi">28</span><span class="p">)</span><span class="o">.</span><span class="n">numpy</span><span class="p">()</span>
</pre></div>
</div>
<div class="output highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">Downloading</span> <span class="n">http</span><span class="p">:</span><span class="o">//</span><span class="n">fashion</span><span class="o">-</span><span class="n">mnist</span><span class="o">.</span><span class="n">s3</span><span class="o">-</span><span class="n">website</span><span class="o">.</span><span class="n">eu</span><span class="o">-</span><span class="n">central</span><span class="o">-</span><span class="mf">1.</span><span class="n">amazonaws</span><span class="o">.</span><span class="n">com</span><span class="o">/</span><span class="n">train</span><span class="o">-</span><span class="n">images</span><span class="o">-</span><span class="n">idx3</span><span class="o">-</span><span class="n">ubyte</span><span class="o">.</span><span class="n">gz</span>
<span class="n">Downloading</span> <span class="n">http</span><span class="p">:</span><span class="o">//</span><span class="n">fashion</span><span class="o">-</span><span class="n">mnist</span><span class="o">.</span><span class="n">s3</span><span class="o">-</span><span class="n">website</span><span class="o">.</span><span class="n">eu</span><span class="o">-</span><span class="n">central</span><span class="o">-</span><span class="mf">1.</span><span class="n">amazonaws</span><span class="o">.</span><span class="n">com</span><span class="o">/</span><span class="n">train</span><span class="o">-</span><span class="n">images</span><span class="o">-</span><span class="n">idx3</span><span class="o">-</span><span class="n">ubyte</span><span class="o">.</span><span class="n">gz</span> <span class="n">to</span> <span class="n">data</span><span class="o">/</span><span class="n">FashionMNIST</span><span class="o">/</span><span class="n">raw</span><span class="o">/</span><span class="n">train</span><span class="o">-</span><span class="n">images</span><span class="o">-</span><span class="n">idx3</span><span class="o">-</span><span class="n">ubyte</span><span class="o">.</span><span class="n">gz</span>
<span class="mf">100.0</span><span class="o">%</span>
<span class="n">Extracting</span> <span class="n">data</span><span class="o">/</span><span class="n">FashionMNIST</span><span class="o">/</span><span class="n">raw</span><span class="o">/</span><span class="n">train</span><span class="o">-</span><span class="n">images</span><span class="o">-</span><span class="n">idx3</span><span class="o">-</span><span class="n">ubyte</span><span class="o">.</span><span class="n">gz</span> <span class="n">to</span> <span class="n">data</span><span class="o">/</span><span class="n">FashionMNIST</span><span class="o">/</span><span class="n">raw</span>
<span class="n">Downloading</span> <span class="n">http</span><span class="p">:</span><span class="o">//</span><span class="n">fashion</span><span class="o">-</span><span class="n">mnist</span><span class="o">.</span><span class="n">s3</span><span class="o">-</span><span class="n">website</span><span class="o">.</span><span class="n">eu</span><span class="o">-</span><span class="n">central</span><span class="o">-</span><span class="mf">1.</span><span class="n">amazonaws</span><span class="o">.</span><span class="n">com</span><span class="o">/</span><span class="n">train</span><span class="o">-</span><span class="n">labels</span><span class="o">-</span><span class="n">idx1</span><span class="o">-</span><span class="n">ubyte</span><span class="o">.</span><span class="n">gz</span>
<span class="n">Downloading</span> <span class="n">http</span><span class="p">:</span><span class="o">//</span><span class="n">fashion</span><span class="o">-</span><span class="n">mnist</span><span class="o">.</span><span class="n">s3</span><span class="o">-</span><span class="n">website</span><span class="o">.</span><span class="n">eu</span><span class="o">-</span><span class="n">central</span><span class="o">-</span><span class="mf">1.</span><span class="n">amazonaws</span><span class="o">.</span><span class="n">com</span><span class="o">/</span><span class="n">train</span><span class="o">-</span><span class="n">labels</span><span class="o">-</span><span class="n">idx1</span><span class="o">-</span><span class="n">ubyte</span><span class="o">.</span><span class="n">gz</span> <span class="n">to</span> <span class="n">data</span><span class="o">/</span><span class="n">FashionMNIST</span><span class="o">/</span><span class="n">raw</span><span class="o">/</span><span class="n">train</span><span class="o">-</span><span class="n">labels</span><span class="o">-</span><span class="n">idx1</span><span class="o">-</span><span class="n">ubyte</span><span class="o">.</span><span class="n">gz</span>
<span class="mf">100.0</span><span class="o">%</span>
<span class="n">Extracting</span> <span class="n">data</span><span class="o">/</span><span class="n">FashionMNIST</span><span class="o">/</span><span class="n">raw</span><span class="o">/</span><span class="n">train</span><span class="o">-</span><span class="n">labels</span><span class="o">-</span><span class="n">idx1</span><span class="o">-</span><span class="n">ubyte</span><span class="o">.</span><span class="n">gz</span> <span class="n">to</span> <span class="n">data</span><span class="o">/</span><span class="n">FashionMNIST</span><span class="o">/</span><span class="n">raw</span>
<span class="n">Downloading</span> <span class="n">http</span><span class="p">:</span><span class="o">//</span><span class="n">fashion</span><span class="o">-</span><span class="n">mnist</span><span class="o">.</span><span class="n">s3</span><span class="o">-</span><span class="n">website</span><span class="o">.</span><span class="n">eu</span><span class="o">-</span><span class="n">central</span><span class="o">-</span><span class="mf">1.</span><span class="n">amazonaws</span><span class="o">.</span><span class="n">com</span><span class="o">/</span><span class="n">t10k</span><span class="o">-</span><span class="n">images</span><span class="o">-</span><span class="n">idx3</span><span class="o">-</span><span class="n">ubyte</span><span class="o">.</span><span class="n">gz</span>
<span class="n">Downloading</span> <span class="n">http</span><span class="p">:</span><span class="o">//</span><span class="n">fashion</span><span class="o">-</span><span class="n">mnist</span><span class="o">.</span><span class="n">s3</span><span class="o">-</span><span class="n">website</span><span class="o">.</span><span class="n">eu</span><span class="o">-</span><span class="n">central</span><span class="o">-</span><span class="mf">1.</span><span class="n">amazonaws</span><span class="o">.</span><span class="n">com</span><span class="o">/</span><span class="n">t10k</span><span class="o">-</span><span class="n">images</span><span class="o">-</span><span class="n">idx3</span><span class="o">-</span><span class="n">ubyte</span><span class="o">.</span><span class="n">gz</span> <span class="n">to</span> <span class="n">data</span><span class="o">/</span><span class="n">FashionMNIST</span><span class="o">/</span><span class="n">raw</span><span class="o">/</span><span class="n">t10k</span><span class="o">-</span><span class="n">images</span><span class="o">-</span><span class="n">idx3</span><span class="o">-</span><span class="n">ubyte</span><span class="o">.</span><span class="n">gz</span>
<span class="mf">100.0</span><span class="o">%</span>
<span class="n">Extracting</span> <span class="n">data</span><span class="o">/</span><span class="n">FashionMNIST</span><span class="o">/</span><span class="n">raw</span><span class="o">/</span><span class="n">t10k</span><span class="o">-</span><span class="n">images</span><span class="o">-</span><span class="n">idx3</span><span class="o">-</span><span class="n">ubyte</span><span class="o">.</span><span class="n">gz</span> <span class="n">to</span> <span class="n">data</span><span class="o">/</span><span class="n">FashionMNIST</span><span class="o">/</span><span class="n">raw</span>
<span class="n">Downloading</span> <span class="n">http</span><span class="p">:</span><span class="o">//</span><span class="n">fashion</span><span class="o">-</span><span class="n">mnist</span><span class="o">.</span><span class="n">s3</span><span class="o">-</span><span class="n">website</span><span class="o">.</span><span class="n">eu</span><span class="o">-</span><span class="n">central</span><span class="o">-</span><span class="mf">1.</span><span class="n">amazonaws</span><span class="o">.</span><span class="n">com</span><span class="o">/</span><span class="n">t10k</span><span class="o">-</span><span class="n">labels</span><span class="o">-</span><span class="n">idx1</span><span class="o">-</span><span class="n">ubyte</span><span class="o">.</span><span class="n">gz</span>
<span class="n">Downloading</span> <span class="n">http</span><span class="p">:</span><span class="o">//</span><span class="n">fashion</span><span class="o">-</span><span class="n">mnist</span><span class="o">.</span><span class="n">s3</span><span class="o">-</span><span class="n">website</span><span class="o">.</span><span class="n">eu</span><span class="o">-</span><span class="n">central</span><span class="o">-</span><span class="mf">1.</span><span class="n">amazonaws</span><span class="o">.</span><span class="n">com</span><span class="o">/</span><span class="n">t10k</span><span class="o">-</span><span class="n">labels</span><span class="o">-</span><span class="n">idx1</span><span class="o">-</span><span class="n">ubyte</span><span class="o">.</span><span class="n">gz</span> <span class="n">to</span> <span class="n">data</span><span class="o">/</span><span class="n">FashionMNIST</span><span class="o">/</span><span class="n">raw</span><span class="o">/</span><span class="n">t10k</span><span class="o">-</span><span class="n">labels</span><span class="o">-</span><span class="n">idx1</span><span class="o">-</span><span class="n">ubyte</span><span class="o">.</span><span class="n">gz</span>
<span class="mf">100.0</span><span class="o">%</span><span class="n">Extracting</span> <span class="n">data</span><span class="o">/</span><span class="n">FashionMNIST</span><span class="o">/</span><span class="n">raw</span><span class="o">/</span><span class="n">t10k</span><span class="o">-</span><span class="n">labels</span><span class="o">-</span><span class="n">idx1</span><span class="o">-</span><span class="n">ubyte</span><span class="o">.</span><span class="n">gz</span> <span class="n">to</span> <span class="n">data</span><span class="o">/</span><span class="n">FashionMNIST</span><span class="o">/</span><span class="n">raw</span>
</pre></div>
</div>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">matplotlib.pyplot</span> <span class="k">as</span> <span class="nn">plt</span>
<span class="n">plt</span><span class="o">.</span><span class="n">figure</span><span class="p">()</span>
<span class="n">plt</span><span class="o">.</span><span class="n">imshow</span><span class="p">(</span><span class="n">img</span><span class="p">[</span><span class="mi">0</span><span class="p">])</span>
<span class="n">plt</span><span class="o">.</span><span class="n">colorbar</span><span class="p">()</span>
<span class="n">plt</span><span class="o">.</span><span class="n">grid</span><span class="p">(</span><span class="kc">False</span><span class="p">)</span>
<span class="n">plt</span><span class="o">.</span><span class="n">show</span><span class="p">()</span>
<span class="nb">print</span><span class="p">(</span><span class="s2">"Class:"</span><span class="p">,</span> <span class="n">class_names</span><span class="p">[</span><span class="n">label</span><span class="p">[</span><span class="mi">0</span><span class="p">]])</span>
</pre></div>
</div>
<div class="figure align-default">
<img alt="../_images/output_index_4f28a7_60_0.png" src="../_images/output_index_4f28a7_60_0.png" />
</div>
<div class="output highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">Class</span><span class="p">:</span> <span class="n">Ankle</span> <span class="n">boot</span>
</pre></div>
</div>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span># Hide outputs
!wget -nc https://github.com/mlc-ai/web-data/raw/main/models/fasionmnist_mlp_params.pkl
</pre></div>
</div>
<div class="figure align-default">
<img alt="../_images/e2e_fashionmnist_mlp_model.png" src="../_images/e2e_fashionmnist_mlp_model.png" />
</div>
<p>The above is our model of interest, we can build the PyTorch model as
follows.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">class</span> <span class="nc">MLP</span><span class="p">(</span><span class="n">nn</span><span class="o">.</span><span class="n">Module</span><span class="p">):</span>
<span class="k">def</span> <span class="fm">__init__</span><span class="p">(</span><span class="bp">self</span><span class="p">):</span>
<span class="nb">super</span><span class="p">(</span><span class="n">MLP</span><span class="p">,</span> <span class="bp">self</span><span class="p">)</span><span class="o">.</span><span class="fm">__init__</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">linear0</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">784</span><span class="p">,</span> <span class="mi">128</span><span class="p">,</span> <span class="n">bias</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="bp">self</span><span class="o">.</span><span class="n">relu</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">ReLU</span><span class="p">()</span>
<span class="bp">self</span><span class="o">.</span><span class="n">linear1</span> <span class="o">=</span> <span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">(</span><span class="mi">128</span><span class="p">,</span> <span class="mi">10</span><span class="p">,</span> <span class="n">bias</span><span class="o">=</span><span class="kc">True</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">forward</span><span class="p">(</span><span class="bp">self</span><span class="p">,</span> <span class="n">x</span><span class="p">):</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">linear0</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="n">x</span> <span class="o">=</span> <span class="bp">self</span><span class="o">.</span><span class="n">linear1</span><span class="p">(</span><span class="n">x</span><span class="p">)</span>
<span class="k">return</span> <span class="n">x</span>
</pre></div>
</div>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">import</span> <span class="nn">pickle</span> <span class="k">as</span> <span class="nn">pkl</span>
<span class="n">mlp_model</span> <span class="o">=</span> <span class="n">MLP</span><span class="p">()</span>
<span class="n">mlp_params</span> <span class="o">=</span> <span class="n">pkl</span><span class="o">.</span><span class="n">load</span><span class="p">(</span><span class="nb">open</span><span class="p">(</span><span class="s2">"fasionmnist_mlp_params.pkl"</span><span class="p">,</span> <span class="s2">"rb"</span><span class="p">))</span>
<span class="n">mlp_model</span><span class="o">.</span><span class="n">linear0</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">data</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">from_numpy</span><span class="p">(</span><span class="n">mlp_params</span><span class="p">[</span><span class="s2">"w0"</span><span class="p">])</span>
<span class="n">mlp_model</span><span class="o">.</span><span class="n">linear0</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">data</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">from_numpy</span><span class="p">(</span><span class="n">mlp_params</span><span class="p">[</span><span class="s2">"b0"</span><span class="p">])</span>
<span class="n">mlp_model</span><span class="o">.</span><span class="n">linear1</span><span class="o">.</span><span class="n">weight</span><span class="o">.</span><span class="n">data</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">from_numpy</span><span class="p">(</span><span class="n">mlp_params</span><span class="p">[</span><span class="s2">"w1"</span><span class="p">])</span>
<span class="n">mlp_model</span><span class="o">.</span><span class="n">linear1</span><span class="o">.</span><span class="n">bias</span><span class="o">.</span><span class="n">data</span> <span class="o">=</span> <span class="n">torch</span><span class="o">.</span><span class="n">from_numpy</span><span class="p">(</span><span class="n">mlp_params</span><span class="p">[</span><span class="s2">"b1"</span><span class="p">])</span>
</pre></div>
</div>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">torch_res</span> <span class="o">=</span> <span class="n">mlp_model</span><span class="p">(</span><span class="n">torch</span><span class="o">.</span><span class="n">from_numpy</span><span class="p">(</span><span class="n">img</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">784</span><span class="p">)))</span>
<span class="n">pred_kind</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">argmax</span><span class="p">(</span><span class="n">torch_res</span><span class="o">.</span><span class="n">detach</span><span class="p">()</span><span class="o">.</span><span class="n">numpy</span><span class="p">(),</span> <span class="n">axis</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="s2">"Torch Prediction:"</span><span class="p">,</span> <span class="n">class_names</span><span class="p">[</span><span class="n">pred_kind</span><span class="p">[</span><span class="mi">0</span><span class="p">]])</span>
</pre></div>
</div>
<div class="output highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">Torch</span> <span class="n">Prediction</span><span class="p">:</span> <span class="n">Ankle</span> <span class="n">boot</span>
</pre></div>
</div>
<p>Let us try to translate from fx by defining mapping functions for the
corresponding <code class="docutils literal notranslate"><span class="pre">nn.Module</span></code>. Here we are reusing pre-defined TE
libraries from TVM <code class="docutils literal notranslate"><span class="pre">topi</span></code> instead of defining our own tensor
expression.</p>
<ul class="simple">
<li><p><code class="docutils literal notranslate"><span class="pre">topi.nn.dense(x,</span> <span class="pre">w)</span></code> performs transposed matrix multiplication
<code class="docutils literal notranslate"><span class="pre">x</span> <span class="pre">@</span> <span class="pre">w.T</span></code></p></li>
<li><p><code class="docutils literal notranslate"><span class="pre">topi.add</span></code> performs broadcast add.</p></li>
</ul>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="kn">from</span> <span class="nn">tvm</span> <span class="kn">import</span> <span class="n">topi</span>
<span class="k">def</span> <span class="nf">map_nn_linear</span><span class="p">(</span><span class="n">bb</span><span class="p">,</span> <span class="n">node_map</span><span class="p">,</span> <span class="n">node</span><span class="p">,</span> <span class="n">nn_mod</span><span class="p">):</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">node_map</span><span class="p">[</span><span class="n">node</span><span class="o">.</span><span class="n">args</span><span class="p">[</span><span class="mi">0</span><span class="p">]]</span>
<span class="n">w</span> <span class="o">=</span> <span class="n">map_param</span><span class="p">(</span><span class="n">nn_mod</span><span class="o">.</span><span class="n">weight</span><span class="p">)</span>
<span class="k">if</span> <span class="n">nn_mod</span><span class="o">.</span><span class="n">bias</span> <span class="ow">is</span> <span class="ow">not</span> <span class="kc">None</span><span class="p">:</span>
<span class="n">b</span> <span class="o">=</span> <span class="n">map_param</span><span class="p">(</span><span class="n">nn_mod</span><span class="o">.</span><span class="n">bias</span><span class="p">)</span>
<span class="n">y</span> <span class="o">=</span> <span class="n">bb</span><span class="o">.</span><span class="n">emit_te</span><span class="p">(</span><span class="n">topi</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">dense</span><span class="p">,</span> <span class="n">x</span><span class="p">,</span> <span class="n">w</span><span class="p">)</span>
<span class="k">return</span> <span class="n">bb</span><span class="o">.</span><span class="n">emit_te</span><span class="p">(</span><span class="n">topi</span><span class="o">.</span><span class="n">add</span><span class="p">,</span> <span class="n">y</span><span class="p">,</span> <span class="n">b</span><span class="p">)</span>
<span class="k">def</span> <span class="nf">map_nn_relu</span><span class="p">(</span><span class="n">bb</span><span class="p">,</span> <span class="n">node_map</span><span class="p">,</span> <span class="n">node</span><span class="p">,</span> <span class="n">nn_mod</span><span class="p">):</span>
<span class="k">return</span> <span class="n">map_relu</span><span class="p">(</span><span class="n">bb</span><span class="p">,</span> <span class="n">node_map</span><span class="p">,</span> <span class="n">node</span><span class="p">)</span>
<span class="n">MLPModule</span> <span class="o">=</span> <span class="n">from_fx</span><span class="p">(</span>
<span class="n">fx</span><span class="o">.</span><span class="n">symbolic_trace</span><span class="p">(</span><span class="n">mlp_model</span><span class="p">),</span>
<span class="n">input_shapes</span> <span class="o">=</span> <span class="p">[(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">784</span><span class="p">)],</span>
<span class="n">call_function_map</span><span class="o">=</span><span class="p">{</span>
<span class="p">},</span>
<span class="n">call_module_map</span><span class="o">=</span><span class="p">{</span>
<span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">:</span> <span class="n">map_nn_linear</span><span class="p">,</span>
<span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">ReLU</span><span class="p">:</span> <span class="n">map_nn_relu</span><span class="p">,</span>
<span class="p">},</span>
<span class="p">)</span>
<span class="n">MLPModule</span><span class="o">.</span><span class="n">show</span><span class="p">()</span>
</pre></div>
</div>
<div class="highlight" style="background: "><pre style="line-height: 125%;"><span></span><span style="color: #007979; font-style: italic"># from tvm.script import ir as I</span>
<span style="color: #007979; font-style: italic"># from tvm.script import tir as T</span>
<span style="color: #007979; font-style: italic"># from tvm.script import relax as R</span>
<span style="color: #AA22FF">@I</span><span style="color: #AA22FF; font-weight: bold">.</span>ir_module
<span style="color: #008000; font-weight: bold">class</span> <span style="color: #0000FF; font-weight: bold">Module</span>:
<span style="color: #AA22FF">@T</span><span style="color: #AA22FF; font-weight: bold">.</span>prim_func(private<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #008000; font-weight: bold">True</span>)
<span style="color: #008000; font-weight: bold">def</span> <span style="color: #0000FF">add</span>(lv: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)), <span style="color: #BA2121">"float32"</span>), B: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>),), <span style="color: #BA2121">"float32"</span>), T_add: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)), <span style="color: #BA2121">"float32"</span>)):
T<span style="color: #AA22FF; font-weight: bold">.</span>func_attr({<span style="color: #BA2121">"tir.noalias"</span>: T<span style="color: #AA22FF; font-weight: bold">.</span>bool(<span style="color: #008000; font-weight: bold">True</span>)})
<span style="color: #007979; font-style: italic"># with T.block("root"):</span>
<span style="color: #008000; font-weight: bold">for</span> ax0, ax1 <span style="color: #008000; font-weight: bold">in</span> T<span style="color: #AA22FF; font-weight: bold">.</span>grid(T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)):
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>block(<span style="color: #BA2121">"T_add"</span>):
v_ax0, v_ax1 <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>axis<span style="color: #AA22FF; font-weight: bold">.</span>remap(<span style="color: #BA2121">"SS"</span>, [ax0, ax1])
T<span style="color: #AA22FF; font-weight: bold">.</span>reads(lv[v_ax0, v_ax1], B[v_ax1])
T<span style="color: #AA22FF; font-weight: bold">.</span>writes(T_add[v_ax0, v_ax1])
T_add[v_ax0, v_ax1] <span style="color: #AA22FF; font-weight: bold">=</span> lv[v_ax0, v_ax1] <span style="color: #AA22FF; font-weight: bold">+</span> B[v_ax1]
<span style="color: #AA22FF">@T</span><span style="color: #AA22FF; font-weight: bold">.</span>prim_func(private<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #008000; font-weight: bold">True</span>)
<span style="color: #008000; font-weight: bold">def</span> <span style="color: #0000FF">add1</span>(lv3: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">10</span>)), <span style="color: #BA2121">"float32"</span>), B: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">10</span>),), <span style="color: #BA2121">"float32"</span>), T_add: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">10</span>)), <span style="color: #BA2121">"float32"</span>)):
T<span style="color: #AA22FF; font-weight: bold">.</span>func_attr({<span style="color: #BA2121">"tir.noalias"</span>: T<span style="color: #AA22FF; font-weight: bold">.</span>bool(<span style="color: #008000; font-weight: bold">True</span>)})
<span style="color: #007979; font-style: italic"># with T.block("root"):</span>
<span style="color: #008000; font-weight: bold">for</span> ax0, ax1 <span style="color: #008000; font-weight: bold">in</span> T<span style="color: #AA22FF; font-weight: bold">.</span>grid(T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">10</span>)):
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>block(<span style="color: #BA2121">"T_add"</span>):
v_ax0, v_ax1 <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>axis<span style="color: #AA22FF; font-weight: bold">.</span>remap(<span style="color: #BA2121">"SS"</span>, [ax0, ax1])
T<span style="color: #AA22FF; font-weight: bold">.</span>reads(lv3[v_ax0, v_ax1], B[v_ax1])
T<span style="color: #AA22FF; font-weight: bold">.</span>writes(T_add[v_ax0, v_ax1])
T_add[v_ax0, v_ax1] <span style="color: #AA22FF; font-weight: bold">=</span> lv3[v_ax0, v_ax1] <span style="color: #AA22FF; font-weight: bold">+</span> B[v_ax1]
<span style="color: #AA22FF">@T</span><span style="color: #AA22FF; font-weight: bold">.</span>prim_func(private<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #008000; font-weight: bold">True</span>)
<span style="color: #008000; font-weight: bold">def</span> <span style="color: #0000FF">dense</span>(x: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">784</span>)), <span style="color: #BA2121">"float32"</span>), B: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">784</span>)), <span style="color: #BA2121">"float32"</span>), T_matmul_NT: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)), <span style="color: #BA2121">"float32"</span>)):
T<span style="color: #AA22FF; font-weight: bold">.</span>func_attr({<span style="color: #BA2121">"layout_free_buffers"</span>: [<span style="color: #008000">1</span>], <span style="color: #BA2121">"tir.noalias"</span>: T<span style="color: #AA22FF; font-weight: bold">.</span>bool(<span style="color: #008000; font-weight: bold">True</span>)})
<span style="color: #007979; font-style: italic"># with T.block("root"):</span>
<span style="color: #008000; font-weight: bold">for</span> i0, i1, k <span style="color: #008000; font-weight: bold">in</span> T<span style="color: #AA22FF; font-weight: bold">.</span>grid(T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">784</span>)):
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>block(<span style="color: #BA2121">"T_matmul_NT"</span>):
v_i0, v_i1, v_k <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>axis<span style="color: #AA22FF; font-weight: bold">.</span>remap(<span style="color: #BA2121">"SSR"</span>, [i0, i1, k])
T<span style="color: #AA22FF; font-weight: bold">.</span>reads(x[v_i0, v_k], B[v_i1, v_k])
T<span style="color: #AA22FF; font-weight: bold">.</span>writes(T_matmul_NT[v_i0, v_i1])
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>init():
T_matmul_NT[v_i0, v_i1] <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>float32(<span style="color: #008000">0.0</span>)
T_matmul_NT[v_i0, v_i1] <span style="color: #AA22FF; font-weight: bold">=</span> T_matmul_NT[v_i0, v_i1] <span style="color: #AA22FF; font-weight: bold">+</span> x[v_i0, v_k] <span style="color: #AA22FF; font-weight: bold">*</span> B[v_i1, v_k]
<span style="color: #AA22FF">@T</span><span style="color: #AA22FF; font-weight: bold">.</span>prim_func(private<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #008000; font-weight: bold">True</span>)
<span style="color: #008000; font-weight: bold">def</span> <span style="color: #0000FF">dense1</span>(lv2: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)), <span style="color: #BA2121">"float32"</span>), B: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">10</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)), <span style="color: #BA2121">"float32"</span>), T_matmul_NT: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">10</span>)), <span style="color: #BA2121">"float32"</span>)):
T<span style="color: #AA22FF; font-weight: bold">.</span>func_attr({<span style="color: #BA2121">"layout_free_buffers"</span>: [<span style="color: #008000">1</span>], <span style="color: #BA2121">"tir.noalias"</span>: T<span style="color: #AA22FF; font-weight: bold">.</span>bool(<span style="color: #008000; font-weight: bold">True</span>)})
<span style="color: #007979; font-style: italic"># with T.block("root"):</span>
<span style="color: #008000; font-weight: bold">for</span> i0, i1, k <span style="color: #008000; font-weight: bold">in</span> T<span style="color: #AA22FF; font-weight: bold">.</span>grid(T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">10</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)):
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>block(<span style="color: #BA2121">"T_matmul_NT"</span>):
v_i0, v_i1, v_k <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>axis<span style="color: #AA22FF; font-weight: bold">.</span>remap(<span style="color: #BA2121">"SSR"</span>, [i0, i1, k])
T<span style="color: #AA22FF; font-weight: bold">.</span>reads(lv2[v_i0, v_k], B[v_i1, v_k])
T<span style="color: #AA22FF; font-weight: bold">.</span>writes(T_matmul_NT[v_i0, v_i1])
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>init():
T_matmul_NT[v_i0, v_i1] <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>float32(<span style="color: #008000">0.0</span>)
T_matmul_NT[v_i0, v_i1] <span style="color: #AA22FF; font-weight: bold">=</span> T_matmul_NT[v_i0, v_i1] <span style="color: #AA22FF; font-weight: bold">+</span> lv2[v_i0, v_k] <span style="color: #AA22FF; font-weight: bold">*</span> B[v_i1, v_k]
<span style="color: #AA22FF">@T</span><span style="color: #AA22FF; font-weight: bold">.</span>prim_func(private<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #008000; font-weight: bold">True</span>)
<span style="color: #008000; font-weight: bold">def</span> <span style="color: #0000FF">te_relu</span>(lv1: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)), <span style="color: #BA2121">"float32"</span>), relu: T<span style="color: #AA22FF; font-weight: bold">.</span>Buffer((T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)), <span style="color: #BA2121">"float32"</span>)):
T<span style="color: #AA22FF; font-weight: bold">.</span>func_attr({<span style="color: #BA2121">"tir.noalias"</span>: T<span style="color: #AA22FF; font-weight: bold">.</span>bool(<span style="color: #008000; font-weight: bold">True</span>)})
<span style="color: #007979; font-style: italic"># with T.block("root"):</span>
<span style="color: #008000; font-weight: bold">for</span> i0, i1 <span style="color: #008000; font-weight: bold">in</span> T<span style="color: #AA22FF; font-weight: bold">.</span>grid(T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">1</span>), T<span style="color: #AA22FF; font-weight: bold">.</span>int64(<span style="color: #008000">128</span>)):
<span style="color: #008000; font-weight: bold">with</span> T<span style="color: #AA22FF; font-weight: bold">.</span>block(<span style="color: #BA2121">"relu"</span>):
v_i0, v_i1 <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>axis<span style="color: #AA22FF; font-weight: bold">.</span>remap(<span style="color: #BA2121">"SS"</span>, [i0, i1])
T<span style="color: #AA22FF; font-weight: bold">.</span>reads(lv1[v_i0, v_i1])
T<span style="color: #AA22FF; font-weight: bold">.</span>writes(relu[v_i0, v_i1])
relu[v_i0, v_i1] <span style="color: #AA22FF; font-weight: bold">=</span> T<span style="color: #AA22FF; font-weight: bold">.</span>max(lv1[v_i0, v_i1], T<span style="color: #AA22FF; font-weight: bold">.</span>float32(<span style="color: #008000">0.0</span>))
<span style="color: #AA22FF">@R</span><span style="color: #AA22FF; font-weight: bold">.</span>function
<span style="color: #008000; font-weight: bold">def</span> <span style="color: #0000FF">main</span>(x: R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">784</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>)) <span style="color: #AA22FF; font-weight: bold">-></span> R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">10</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>):
cls <span style="color: #AA22FF; font-weight: bold">=</span> Module
<span style="color: #008000; font-weight: bold">with</span> R<span style="color: #AA22FF; font-weight: bold">.</span>dataflow():
lv <span style="color: #AA22FF; font-weight: bold">=</span> R<span style="color: #AA22FF; font-weight: bold">.</span>call_tir(cls<span style="color: #AA22FF; font-weight: bold">.</span>dense, (x, metadata[<span style="color: #BA2121">"relax.expr.Constant"</span>][<span style="color: #008000">0</span>]), out_sinfo<span style="color: #AA22FF; font-weight: bold">=</span>R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">128</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>))
lv1 <span style="color: #AA22FF; font-weight: bold">=</span> R<span style="color: #AA22FF; font-weight: bold">.</span>call_tir(cls<span style="color: #AA22FF; font-weight: bold">.</span>add, (lv, metadata[<span style="color: #BA2121">"relax.expr.Constant"</span>][<span style="color: #008000">1</span>]), out_sinfo<span style="color: #AA22FF; font-weight: bold">=</span>R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">128</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>))
lv2 <span style="color: #AA22FF; font-weight: bold">=</span> R<span style="color: #AA22FF; font-weight: bold">.</span>call_tir(cls<span style="color: #AA22FF; font-weight: bold">.</span>te_relu, (lv1,), out_sinfo<span style="color: #AA22FF; font-weight: bold">=</span>R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">128</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>))
lv3 <span style="color: #AA22FF; font-weight: bold">=</span> R<span style="color: #AA22FF; font-weight: bold">.</span>call_tir(cls<span style="color: #AA22FF; font-weight: bold">.</span>dense1, (lv2, metadata[<span style="color: #BA2121">"relax.expr.Constant"</span>][<span style="color: #008000">2</span>]), out_sinfo<span style="color: #AA22FF; font-weight: bold">=</span>R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">10</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>))
lv4 <span style="color: #AA22FF; font-weight: bold">=</span> R<span style="color: #AA22FF; font-weight: bold">.</span>call_tir(cls<span style="color: #AA22FF; font-weight: bold">.</span>add1, (lv3, metadata[<span style="color: #BA2121">"relax.expr.Constant"</span>][<span style="color: #008000">3</span>]), out_sinfo<span style="color: #AA22FF; font-weight: bold">=</span>R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">10</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>))
gv: R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">10</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>) <span style="color: #AA22FF; font-weight: bold">=</span> lv4
R<span style="color: #AA22FF; font-weight: bold">.</span>output(gv)
<span style="color: #008000; font-weight: bold">return</span> lv4
<span style="color: #007979; font-style: italic"># Metadata omitted. Use show_meta=True in script() method to show it.</span>
</pre></div><div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="n">ex</span> <span class="o">=</span> <span class="n">relax</span><span class="o">.</span><span class="n">build</span><span class="p">(</span><span class="n">MLPModule</span><span class="p">,</span> <span class="n">target</span><span class="o">=</span><span class="s2">"llvm"</span><span class="p">)</span>
<span class="n">vm</span> <span class="o">=</span> <span class="n">relax</span><span class="o">.</span><span class="n">VirtualMachine</span><span class="p">(</span><span class="n">ex</span><span class="p">,</span> <span class="n">tvm</span><span class="o">.</span><span class="n">cpu</span><span class="p">())</span>
<span class="n">data_nd</span> <span class="o">=</span> <span class="n">tvm</span><span class="o">.</span><span class="n">nd</span><span class="o">.</span><span class="n">array</span><span class="p">(</span><span class="n">img</span><span class="o">.</span><span class="n">reshape</span><span class="p">(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">784</span><span class="p">))</span>
<span class="n">nd_res</span> <span class="o">=</span> <span class="n">vm</span><span class="p">[</span><span class="s2">"main"</span><span class="p">](</span><span class="n">data_nd</span><span class="p">)</span>
<span class="n">pred_kind</span> <span class="o">=</span> <span class="n">np</span><span class="o">.</span><span class="n">argmax</span><span class="p">(</span><span class="n">nd_res</span><span class="o">.</span><span class="n">numpy</span><span class="p">(),</span> <span class="n">axis</span><span class="o">=</span><span class="mi">1</span><span class="p">)</span>
<span class="nb">print</span><span class="p">(</span><span class="s2">"MLPModule Prediction:"</span><span class="p">,</span> <span class="n">class_names</span><span class="p">[</span><span class="n">pred_kind</span><span class="p">[</span><span class="mi">0</span><span class="p">]])</span>
</pre></div>
</div>
<div class="output highlight-default notranslate"><div class="highlight"><pre><span></span><span class="n">MLPModule</span> <span class="n">Prediction</span><span class="p">:</span> <span class="n">Ankle</span> <span class="n">boot</span>
</pre></div>
</div>
</div>
<div class="section" id="remark-translating-into-high-level-operators">
<h2><span class="section-number">5.6. </span>Remark: Translating into High-level Operators<a class="headerlink" href="#remark-translating-into-high-level-operators" title="Permalink to this heading">¶</a></h2>
<p>In most machine learning frameworks, it is sometimes helpful to first
translate into high-level built-in primitive operators. The following
code block gives an example to do that.</p>
<div class="highlight-python notranslate"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">map_nn_relu_op</span><span class="p">(</span><span class="n">bb</span><span class="p">,</span> <span class="n">node_map</span><span class="p">,</span> <span class="n">node</span><span class="p">,</span> <span class="n">nn_mod</span><span class="p">):</span>
<span class="n">A</span> <span class="o">=</span> <span class="n">node_map</span><span class="p">[</span><span class="n">node</span><span class="o">.</span><span class="n">args</span><span class="p">[</span><span class="mi">0</span><span class="p">]]</span>
<span class="k">return</span> <span class="n">bb</span><span class="o">.</span><span class="n">emit</span><span class="p">(</span><span class="n">relax</span><span class="o">.</span><span class="n">op</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">relu</span><span class="p">(</span><span class="n">A</span><span class="p">))</span>
<span class="k">def</span> <span class="nf">map_nn_linear_op</span><span class="p">(</span><span class="n">bb</span><span class="p">,</span> <span class="n">node_map</span><span class="p">,</span> <span class="n">node</span><span class="p">,</span> <span class="n">nn_mod</span><span class="p">):</span>
<span class="n">x</span> <span class="o">=</span> <span class="n">node_map</span><span class="p">[</span><span class="n">node</span><span class="o">.</span><span class="n">args</span><span class="p">[</span><span class="mi">0</span><span class="p">]]</span>
<span class="n">w</span> <span class="o">=</span> <span class="n">map_param</span><span class="p">(</span><span class="n">nn_mod</span><span class="o">.</span><span class="n">weight</span><span class="p">)</span>
<span class="n">b</span> <span class="o">=</span> <span class="n">map_param</span><span class="p">(</span><span class="n">nn_mod</span><span class="o">.</span><span class="n">bias</span><span class="p">)</span>
<span class="k">return</span> <span class="n">bb</span><span class="o">.</span><span class="n">emit</span><span class="p">(</span><span class="n">relax</span><span class="o">.</span><span class="n">op</span><span class="o">.</span><span class="n">linear</span><span class="p">(</span><span class="n">x</span><span class="p">,</span> <span class="n">w</span><span class="p">,</span> <span class="n">b</span><span class="p">))</span>
<span class="n">MLPModuleHighLevel</span> <span class="o">=</span> <span class="n">from_fx</span><span class="p">(</span>
<span class="n">fx</span><span class="o">.</span><span class="n">symbolic_trace</span><span class="p">(</span><span class="n">mlp_model</span><span class="p">),</span>
<span class="n">input_shapes</span> <span class="o">=</span> <span class="p">[(</span><span class="mi">1</span><span class="p">,</span> <span class="mi">784</span><span class="p">)],</span>
<span class="n">call_function_map</span><span class="o">=</span><span class="p">{</span>
<span class="p">},</span>
<span class="n">call_module_map</span><span class="o">=</span><span class="p">{</span>
<span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">Linear</span><span class="p">:</span> <span class="n">map_nn_linear_op</span><span class="p">,</span>
<span class="n">torch</span><span class="o">.</span><span class="n">nn</span><span class="o">.</span><span class="n">ReLU</span><span class="p">:</span> <span class="n">map_nn_relu_op</span><span class="p">,</span>
<span class="p">},</span>
<span class="p">)</span>
<span class="n">MLPModuleHighLevel</span><span class="o">.</span><span class="n">show</span><span class="p">()</span>
</pre></div>
</div>
<div class="highlight" style="background: "><pre style="line-height: 125%;"><span></span><span style="color: #007979; font-style: italic"># from tvm.script import ir as I</span>
<span style="color: #007979; font-style: italic"># from tvm.script import relax as R</span>
<span style="color: #AA22FF">@I</span><span style="color: #AA22FF; font-weight: bold">.</span>ir_module
<span style="color: #008000; font-weight: bold">class</span> <span style="color: #0000FF; font-weight: bold">Module</span>:
<span style="color: #AA22FF">@R</span><span style="color: #AA22FF; font-weight: bold">.</span>function
<span style="color: #008000; font-weight: bold">def</span> <span style="color: #0000FF">main</span>(x: R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">784</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>)) <span style="color: #AA22FF; font-weight: bold">-></span> R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">10</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>):
<span style="color: #008000; font-weight: bold">with</span> R<span style="color: #AA22FF; font-weight: bold">.</span>dataflow():
lv: R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">784</span>, <span style="color: #008000">128</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>) <span style="color: #AA22FF; font-weight: bold">=</span> R<span style="color: #AA22FF; font-weight: bold">.</span>permute_dims(metadata[<span style="color: #BA2121">"relax.expr.Constant"</span>][<span style="color: #008000">0</span>], axes<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #008000; font-weight: bold">None</span>)
lv1: R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">128</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>) <span style="color: #AA22FF; font-weight: bold">=</span> R<span style="color: #AA22FF; font-weight: bold">.</span>matmul(x, lv, out_dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"void"</span>)
lv2: R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">128</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>) <span style="color: #AA22FF; font-weight: bold">=</span> R<span style="color: #AA22FF; font-weight: bold">.</span>add(lv1, metadata[<span style="color: #BA2121">"relax.expr.Constant"</span>][<span style="color: #008000">1</span>])
lv3: R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">128</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>) <span style="color: #AA22FF; font-weight: bold">=</span> R<span style="color: #AA22FF; font-weight: bold">.</span>nn<span style="color: #AA22FF; font-weight: bold">.</span>relu(lv2)
lv4: R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">128</span>, <span style="color: #008000">10</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>) <span style="color: #AA22FF; font-weight: bold">=</span> R<span style="color: #AA22FF; font-weight: bold">.</span>permute_dims(metadata[<span style="color: #BA2121">"relax.expr.Constant"</span>][<span style="color: #008000">2</span>], axes<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #008000; font-weight: bold">None</span>)
lv5: R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">10</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>) <span style="color: #AA22FF; font-weight: bold">=</span> R<span style="color: #AA22FF; font-weight: bold">.</span>matmul(lv3, lv4, out_dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"void"</span>)
lv6: R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">10</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>) <span style="color: #AA22FF; font-weight: bold">=</span> R<span style="color: #AA22FF; font-weight: bold">.</span>add(lv5, metadata[<span style="color: #BA2121">"relax.expr.Constant"</span>][<span style="color: #008000">3</span>])
gv: R<span style="color: #AA22FF; font-weight: bold">.</span>Tensor((<span style="color: #008000">1</span>, <span style="color: #008000">10</span>), dtype<span style="color: #AA22FF; font-weight: bold">=</span><span style="color: #BA2121">"float32"</span>) <span style="color: #AA22FF; font-weight: bold">=</span> lv6
R<span style="color: #AA22FF; font-weight: bold">.</span>output(gv)
<span style="color: #008000; font-weight: bold">return</span> lv6
<span style="color: #007979; font-style: italic"># Metadata omitted. Use show_meta=True in script() method to show it.</span>
</pre></div><p>After we get the model into IRModule with those built-in operator calls.
These built-in operators are <strong>higher-level abstraction</strong> than the
TensorIR functions. There can be different opportunities to further
translate these primitive operators into either library or TensorIR
functions.</p>