diff --git "a/parse/train/ryG6xZ-RZ/ryG6xZ-RZ_middle.json" "b/parse/train/ryG6xZ-RZ/ryG6xZ-RZ_middle.json"
new file mode 100644--- /dev/null
+++ "b/parse/train/ryG6xZ-RZ/ryG6xZ-RZ_middle.json"
@@ -0,0 +1,22480 @@
+{
+ "pdf_info": [
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 79,
+ 504,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 79,
+ 506,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 79,
+ 506,
+ 100
+ ],
+ "score": 1.0,
+ "content": "DLVM: A MODERN COMPILER INFRASTRUCTURE FOR",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 102,
+ 300,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 102,
+ 300,
+ 118
+ ],
+ "score": 1.0,
+ "content": "DEEP LEARNING SYSTEMS",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 112,
+ 134,
+ 308,
+ 190
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 135,
+ 169,
+ 146
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 135,
+ 169,
+ 146
+ ],
+ "score": 1.0,
+ "content": "Richard Wei",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 111,
+ 146,
+ 309,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 146,
+ 309,
+ 159
+ ],
+ "score": 1.0,
+ "content": "Departments of Computer Science & Linguistics",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 111,
+ 156,
+ 289,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 156,
+ 289,
+ 169
+ ],
+ "score": 1.0,
+ "content": "University of Illinois at Urbana-Champaign",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 112,
+ 168,
+ 185,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 168,
+ 185,
+ 178
+ ],
+ "score": 1.0,
+ "content": "Urbana, IL 61801",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 111,
+ 179,
+ 228,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 179,
+ 228,
+ 190
+ ],
+ "score": 1.0,
+ "content": "xwei12@illinois.edu",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 322,
+ 135,
+ 498,
+ 190
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 322,
+ 135,
+ 389,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 322,
+ 135,
+ 389,
+ 147
+ ],
+ "score": 1.0,
+ "content": "Lane Schwartz",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 322,
+ 147,
+ 430,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 322,
+ 147,
+ 430,
+ 158
+ ],
+ "score": 1.0,
+ "content": "Department of Linguistics",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 321,
+ 156,
+ 499,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 321,
+ 156,
+ 499,
+ 169
+ ],
+ "score": 1.0,
+ "content": "University of Illinois at Urbana-Champaign",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 323,
+ 168,
+ 396,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 323,
+ 168,
+ 396,
+ 178
+ ],
+ "score": 1.0,
+ "content": "Urbana, IL 61801",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 322,
+ 180,
+ 432,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 322,
+ 180,
+ 432,
+ 190
+ ],
+ "score": 1.0,
+ "content": "lanes@illinois.edu",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 113,
+ 207,
+ 171,
+ 218
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 205,
+ 173,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 205,
+ 173,
+ 219
+ ],
+ "score": 1.0,
+ "content": "Vikram Adve",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 113,
+ 219,
+ 288,
+ 262
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 218,
+ 248,
+ 230
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 218,
+ 248,
+ 230
+ ],
+ "score": 1.0,
+ "content": "Department of Computer Science",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 111,
+ 228,
+ 289,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 228,
+ 289,
+ 241
+ ],
+ "score": 1.0,
+ "content": "University of Illinois at Urbana-Champaign",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 113,
+ 240,
+ 185,
+ 251
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 240,
+ 185,
+ 251
+ ],
+ "score": 1.0,
+ "content": "Urbana, IL 61801",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 112,
+ 252,
+ 222,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 252,
+ 222,
+ 262
+ ],
+ "score": 1.0,
+ "content": "vadve@illinois.edu",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 14.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 291,
+ 333,
+ 303
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 290,
+ 336,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 290,
+ 336,
+ 306
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 318,
+ 469,
+ 440
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 319,
+ 470,
+ 332
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 319,
+ 470,
+ 332
+ ],
+ "score": 1.0,
+ "content": "Deep learning software demands reliability and performance. However, many of",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 330,
+ 469,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 330,
+ 469,
+ 343
+ ],
+ "score": 1.0,
+ "content": "the existing deep learning frameworks are software libraries that act as an unsafe",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 340,
+ 469,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 340,
+ 469,
+ 354
+ ],
+ "score": 1.0,
+ "content": "DSL in Python and a computation graph interpreter. We present DLVM, a design",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 141,
+ 352,
+ 469,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 352,
+ 469,
+ 364
+ ],
+ "score": 1.0,
+ "content": "and implementation of a compiler infrastructure with a linear algebra intermediate",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 363,
+ 470,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 363,
+ 470,
+ 375
+ ],
+ "score": 1.0,
+ "content": "representation, algorithmic differentiation by adjoint code generation, domain-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 141,
+ 374,
+ 470,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 374,
+ 470,
+ 386
+ ],
+ "score": 1.0,
+ "content": "specific optimizations and a code generator targeting GPU via LLVM. Designed",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 141,
+ 385,
+ 470,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 385,
+ 470,
+ 398
+ ],
+ "score": 1.0,
+ "content": "as a modern compiler infrastructure inspired by LLVM, DLVM is more modular",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 141,
+ 396,
+ 470,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 396,
+ 470,
+ 408
+ ],
+ "score": 1.0,
+ "content": "and more generic than existing deep learning compiler frameworks, and supports",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 141,
+ 407,
+ 470,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 407,
+ 470,
+ 420
+ ],
+ "score": 1.0,
+ "content": "tensor DSLs with high expressivity. With our prototypical staged DSL embedded",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 142,
+ 418,
+ 469,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 418,
+ 469,
+ 429
+ ],
+ "score": 1.0,
+ "content": "in Swift, we argue that the DLVM system enables a form of modular, safe and",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 141,
+ 427,
+ 312,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 427,
+ 312,
+ 443
+ ],
+ "score": 1.0,
+ "content": "performant frameworks for deep learning.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 468,
+ 206,
+ 480
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 208,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 208,
+ 483
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 495,
+ 505,
+ 649
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 506,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 506,
+ 509
+ ],
+ "score": 1.0,
+ "content": "Within the deep learning community, most current approaches to neural networks make use of",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 507,
+ 506,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 507,
+ 506,
+ 518
+ ],
+ "score": 1.0,
+ "content": "high-level frameworks with a tensor domain-specific language (DSL) such as Torch (Collobert et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 516,
+ 506,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 516,
+ 506,
+ 530
+ ],
+ "score": 1.0,
+ "content": "2011), TensorFlow (Abadi et al., 2016), PyTorch (PyTorch Development Team, 2016), and MXNet",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 528,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 528,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": "(Chen et al., 2015). Traditionally, developers would build a computation graph (or dynamically",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 539,
+ 505,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 539,
+ 505,
+ 552
+ ],
+ "score": 1.0,
+ "content": "generate graph nodes) using a DSL and let the framework interpret the computation graph on parallel",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 550,
+ 505,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 550,
+ 505,
+ 563
+ ],
+ "score": 1.0,
+ "content": "architectures such as NVIDIA GPUs. While using hand-tuned GPU subroutines usually yields the",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 560,
+ 506,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 506,
+ 575
+ ],
+ "score": 1.0,
+ "content": "best performance for complex operators, advanced compiler techniques can be applied to simplify",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "computation, merge high-level operators based on shaping conditions, and fuse compatible element-",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 583,
+ 506,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 583,
+ 506,
+ 596
+ ],
+ "score": 1.0,
+ "content": "wise operators to a single kernel to minimize the latency between kernel launches. Recent projects, the",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 104,
+ 592,
+ 506,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 592,
+ 506,
+ 608
+ ],
+ "score": 1.0,
+ "content": "TensorFlow XLA compiler (Leary & Wang, 2017) and the NNVM compiler (NNVM, 2017) including",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 104,
+ 604,
+ 506,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 604,
+ 506,
+ 619
+ ],
+ "score": 1.0,
+ "content": "TVM (Chen et al., 2017), have begun to apply compiler techniques to deep learning systems, targeting",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "LLVM (Lattner & Adve, 2004) and various back-ends to achieve good performance. However, their",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 627,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "design and implementation have not entirely followed established best practices in widely-used",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 638,
+ 256,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 638,
+ 256,
+ 651
+ ],
+ "score": 1.0,
+ "content": "compiler frameworks in the industry.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 36.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 655,
+ 505,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 505,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 505,
+ 668
+ ],
+ "score": 1.0,
+ "content": "Moreover, some frameworks use operator-overloading algorithmic differentiation (AD) to compute",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "gradients, leaving the gradient computation unoptimizable. The other approach to AD, source code",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "transformation, can produce more efficient code. While frameworks such as TensorFlow already",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "perform AD as a graph transformation and apply various optimizations, their AD transformation is",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "not designed as a transformation pass in the pipeline of their compiler framework, but as part of the",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "DSL library. Making AD part of the compiler framework would greatly simplify the development of",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 270,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 270,
+ 734
+ ],
+ "score": 1.0,
+ "content": "DSLs, achieving separation of concerns.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "page_idx": 0,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 224,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 225,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 225,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Workshop track - ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 761
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 761
+ ],
+ "score": 1.0,
+ "content": "1",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 79,
+ 504,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 79,
+ 506,
+ 100
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 79,
+ 506,
+ 100
+ ],
+ "score": 1.0,
+ "content": "DLVM: A MODERN COMPILER INFRASTRUCTURE FOR",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 102,
+ 300,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 102,
+ 300,
+ 118
+ ],
+ "score": 1.0,
+ "content": "DEEP LEARNING SYSTEMS",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 112,
+ 134,
+ 308,
+ 190
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 135,
+ 169,
+ 146
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 135,
+ 169,
+ 146
+ ],
+ "score": 1.0,
+ "content": "Richard Wei",
+ "type": "text"
+ }
+ ],
+ "index": 2,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 111,
+ 146,
+ 309,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 146,
+ 309,
+ 159
+ ],
+ "score": 1.0,
+ "content": "Departments of Computer Science & Linguistics",
+ "type": "text"
+ }
+ ],
+ "index": 4,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 111,
+ 156,
+ 289,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 156,
+ 289,
+ 169
+ ],
+ "score": 1.0,
+ "content": "University of Illinois at Urbana-Champaign",
+ "type": "text"
+ }
+ ],
+ "index": 6,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 112,
+ 168,
+ 185,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 168,
+ 185,
+ 178
+ ],
+ "score": 1.0,
+ "content": "Urbana, IL 61801",
+ "type": "text"
+ }
+ ],
+ "index": 8,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 111,
+ 179,
+ 228,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 179,
+ 228,
+ 190
+ ],
+ "score": 1.0,
+ "content": "xwei12@illinois.edu",
+ "type": "text"
+ }
+ ],
+ "index": 10,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 322,
+ 135,
+ 389,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 322,
+ 135,
+ 389,
+ 147
+ ],
+ "score": 1.0,
+ "content": "Lane Schwartz",
+ "type": "text"
+ }
+ ],
+ "index": 3,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 322,
+ 147,
+ 430,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 322,
+ 147,
+ 430,
+ 158
+ ],
+ "score": 1.0,
+ "content": "Department of Linguistics",
+ "type": "text"
+ }
+ ],
+ "index": 5,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 321,
+ 156,
+ 499,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 321,
+ 156,
+ 499,
+ 169
+ ],
+ "score": 1.0,
+ "content": "University of Illinois at Urbana-Champaign",
+ "type": "text"
+ }
+ ],
+ "index": 7,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 323,
+ 168,
+ 396,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 323,
+ 168,
+ 396,
+ 178
+ ],
+ "score": 1.0,
+ "content": "Urbana, IL 61801",
+ "type": "text"
+ }
+ ],
+ "index": 9,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 322,
+ 180,
+ 432,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 322,
+ 180,
+ 432,
+ 190
+ ],
+ "score": 1.0,
+ "content": "lanes@illinois.edu",
+ "type": "text"
+ }
+ ],
+ "index": 11,
+ "is_list_start_line": true
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 111,
+ 135,
+ 309,
+ 190
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 322,
+ 135,
+ 498,
+ 190
+ ],
+ "lines": [],
+ "index": 7,
+ "bbox_fs": [
+ 321,
+ 135,
+ 499,
+ 190
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 113,
+ 207,
+ 171,
+ 218
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 205,
+ 173,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 205,
+ 173,
+ 219
+ ],
+ "score": 1.0,
+ "content": "Vikram Adve",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 113,
+ 219,
+ 288,
+ 262
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 218,
+ 248,
+ 230
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 218,
+ 248,
+ 230
+ ],
+ "score": 1.0,
+ "content": "Department of Computer Science",
+ "type": "text"
+ }
+ ],
+ "index": 13,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 111,
+ 228,
+ 289,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 228,
+ 289,
+ 241
+ ],
+ "score": 1.0,
+ "content": "University of Illinois at Urbana-Champaign",
+ "type": "text"
+ }
+ ],
+ "index": 14,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 113,
+ 240,
+ 185,
+ 251
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 240,
+ 185,
+ 251
+ ],
+ "score": 1.0,
+ "content": "Urbana, IL 61801",
+ "type": "text"
+ }
+ ],
+ "index": 15,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 112,
+ 252,
+ 222,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 252,
+ 222,
+ 262
+ ],
+ "score": 1.0,
+ "content": "vadve@illinois.edu",
+ "type": "text"
+ }
+ ],
+ "index": 16,
+ "is_list_start_line": true
+ }
+ ],
+ "index": 14.5,
+ "bbox_fs": [
+ 111,
+ 218,
+ 289,
+ 262
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 278,
+ 291,
+ 333,
+ 303
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 276,
+ 290,
+ 336,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 276,
+ 290,
+ 336,
+ 306
+ ],
+ "score": 1.0,
+ "content": "ABSTRACT",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 143,
+ 318,
+ 469,
+ 440
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 319,
+ 470,
+ 332
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 319,
+ 470,
+ 332
+ ],
+ "score": 1.0,
+ "content": "Deep learning software demands reliability and performance. However, many of",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 330,
+ 469,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 330,
+ 469,
+ 343
+ ],
+ "score": 1.0,
+ "content": "the existing deep learning frameworks are software libraries that act as an unsafe",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 340,
+ 469,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 340,
+ 469,
+ 354
+ ],
+ "score": 1.0,
+ "content": "DSL in Python and a computation graph interpreter. We present DLVM, a design",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 141,
+ 352,
+ 469,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 352,
+ 469,
+ 364
+ ],
+ "score": 1.0,
+ "content": "and implementation of a compiler infrastructure with a linear algebra intermediate",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 363,
+ 470,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 363,
+ 470,
+ 375
+ ],
+ "score": 1.0,
+ "content": "representation, algorithmic differentiation by adjoint code generation, domain-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 141,
+ 374,
+ 470,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 374,
+ 470,
+ 386
+ ],
+ "score": 1.0,
+ "content": "specific optimizations and a code generator targeting GPU via LLVM. Designed",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 141,
+ 385,
+ 470,
+ 398
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 385,
+ 470,
+ 398
+ ],
+ "score": 1.0,
+ "content": "as a modern compiler infrastructure inspired by LLVM, DLVM is more modular",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 141,
+ 396,
+ 470,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 396,
+ 470,
+ 408
+ ],
+ "score": 1.0,
+ "content": "and more generic than existing deep learning compiler frameworks, and supports",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 141,
+ 407,
+ 470,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 407,
+ 470,
+ 420
+ ],
+ "score": 1.0,
+ "content": "tensor DSLs with high expressivity. With our prototypical staged DSL embedded",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 142,
+ 418,
+ 469,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 142,
+ 418,
+ 469,
+ 429
+ ],
+ "score": 1.0,
+ "content": "in Swift, we argue that the DLVM system enables a form of modular, safe and",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 141,
+ 427,
+ 312,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 427,
+ 312,
+ 443
+ ],
+ "score": 1.0,
+ "content": "performant frameworks for deep learning.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 23,
+ "bbox_fs": [
+ 141,
+ 319,
+ 470,
+ 443
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 468,
+ 206,
+ 480
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 208,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 208,
+ 483
+ ],
+ "score": 1.0,
+ "content": "1 INTRODUCTION",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 495,
+ 505,
+ 649
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 506,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 506,
+ 509
+ ],
+ "score": 1.0,
+ "content": "Within the deep learning community, most current approaches to neural networks make use of",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 507,
+ 506,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 507,
+ 506,
+ 518
+ ],
+ "score": 1.0,
+ "content": "high-level frameworks with a tensor domain-specific language (DSL) such as Torch (Collobert et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 516,
+ 506,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 516,
+ 506,
+ 530
+ ],
+ "score": 1.0,
+ "content": "2011), TensorFlow (Abadi et al., 2016), PyTorch (PyTorch Development Team, 2016), and MXNet",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 528,
+ 505,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 528,
+ 505,
+ 541
+ ],
+ "score": 1.0,
+ "content": "(Chen et al., 2015). Traditionally, developers would build a computation graph (or dynamically",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 539,
+ 505,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 539,
+ 505,
+ 552
+ ],
+ "score": 1.0,
+ "content": "generate graph nodes) using a DSL and let the framework interpret the computation graph on parallel",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 550,
+ 505,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 550,
+ 505,
+ 563
+ ],
+ "score": 1.0,
+ "content": "architectures such as NVIDIA GPUs. While using hand-tuned GPU subroutines usually yields the",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 560,
+ 506,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 560,
+ 506,
+ 575
+ ],
+ "score": 1.0,
+ "content": "best performance for complex operators, advanced compiler techniques can be applied to simplify",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "computation, merge high-level operators based on shaping conditions, and fuse compatible element-",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 583,
+ 506,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 583,
+ 506,
+ 596
+ ],
+ "score": 1.0,
+ "content": "wise operators to a single kernel to minimize the latency between kernel launches. Recent projects, the",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 104,
+ 592,
+ 506,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 592,
+ 506,
+ 608
+ ],
+ "score": 1.0,
+ "content": "TensorFlow XLA compiler (Leary & Wang, 2017) and the NNVM compiler (NNVM, 2017) including",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 104,
+ 604,
+ 506,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 604,
+ 506,
+ 619
+ ],
+ "score": 1.0,
+ "content": "TVM (Chen et al., 2017), have begun to apply compiler techniques to deep learning systems, targeting",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "LLVM (Lattner & Adve, 2004) and various back-ends to achieve good performance. However, their",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 627,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 627,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "design and implementation have not entirely followed established best practices in widely-used",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 638,
+ 256,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 638,
+ 256,
+ 651
+ ],
+ "score": 1.0,
+ "content": "compiler frameworks in the industry.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 104,
+ 494,
+ 506,
+ 651
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 655,
+ 505,
+ 731
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 505,
+ 668
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 505,
+ 668
+ ],
+ "score": 1.0,
+ "content": "Moreover, some frameworks use operator-overloading algorithmic differentiation (AD) to compute",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "gradients, leaving the gradient computation unoptimizable. The other approach to AD, source code",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "transformation, can produce more efficient code. While frameworks such as TensorFlow already",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "perform AD as a graph transformation and apply various optimizations, their AD transformation is",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "not designed as a transformation pass in the pipeline of their compiler framework, but as part of the",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "DSL library. Making AD part of the compiler framework would greatly simplify the development of",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 270,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 270,
+ 734
+ ],
+ "score": 1.0,
+ "content": "DSLs, achieving separation of concerns.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 47,
+ "bbox_fs": [
+ 105,
+ 654,
+ 506,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 107,
+ 79,
+ 501,
+ 281
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 107,
+ 79,
+ 501,
+ 281
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 79,
+ 501,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 79,
+ 501,
+ 281
+ ],
+ "score": 0.972,
+ "type": "image",
+ "image_path": "849b709b69e101ea7ef26d6864d1b1e0eafec6758703952e4749d0cbe5bb8354.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 107,
+ 79,
+ 501,
+ 146.33333333333331
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 107,
+ 146.33333333333331,
+ 501,
+ 213.66666666666663
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 107,
+ 213.66666666666663,
+ 501,
+ 280.99999999999994
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 104,
+ 291,
+ 501,
+ 303
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 290,
+ 501,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 290,
+ 501,
+ 304
+ ],
+ "score": 1.0,
+ "content": "Figure 1: Software stack of the DLVM infrastructure. Blue components are the compiler framework.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 325,
+ 506,
+ 424
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 325,
+ 506,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 325,
+ 506,
+ 338
+ ],
+ "score": 1.0,
+ "content": "We introduce DLVM, a new compiler infrastructure for deep learning systems that addresses shortcom-",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 336,
+ 506,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 336,
+ 506,
+ 348
+ ],
+ "score": 1.0,
+ "content": "ings of existing deep learning frameworks. Our solution includes (1) a domain-specific intermediate",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 104,
+ 346,
+ 506,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 346,
+ 506,
+ 360
+ ],
+ "score": 1.0,
+ "content": "representation specifically designed for tensor computation, (2) principled use of modern compiler",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 357,
+ 506,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 357,
+ 506,
+ 371
+ ],
+ "score": 1.0,
+ "content": "optimization techniques to substantially simplify neural network computation, including algebra sim-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 369,
+ 506,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 369,
+ 506,
+ 382
+ ],
+ "score": 1.0,
+ "content": "plification, AD checkpointing, compute kernel fusion, and various traditional compiler optimizations,",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 380,
+ 506,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 380,
+ 506,
+ 393
+ ],
+ "score": 1.0,
+ "content": "(3) code generation through a mature compiler infrastructure that allows for transparent targeting of",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 391,
+ 506,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 391,
+ 506,
+ 404
+ ],
+ "score": 1.0,
+ "content": "various hardware, and (4) an embedded DSL that supports static analysis, type safety, and natural",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 401,
+ 506,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 401,
+ 506,
+ 416
+ ],
+ "score": 1.0,
+ "content": "expression of tensor computation, and has a just-in-time (JIT) compiler targeting DLVM for AD,",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 413,
+ 250,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 413,
+ 250,
+ 426
+ ],
+ "score": 1.0,
+ "content": "optimizations, and code generation.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 441,
+ 211,
+ 454
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 440,
+ 213,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 440,
+ 213,
+ 457
+ ],
+ "score": 1.0,
+ "content": "2 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 467,
+ 505,
+ 555
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 505,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 505,
+ 480
+ ],
+ "score": 1.0,
+ "content": "Numerous existing projects provide specialized systems for machine learning but are not closely",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 477,
+ 505,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 477,
+ 505,
+ 492
+ ],
+ "score": 1.0,
+ "content": "related to our work. These include Apache SystemML (Ghoting et al., 2011), a high-level language",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "and framework for writing and executing machine learning problems targeting Apache Spark, and",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 505,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 241,
+ 513
+ ],
+ "score": 1.0,
+ "content": "TACO (Kjolstad et al., 2017), a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 241,
+ 501,
+ 262,
+ 511
+ ],
+ "score": 0.85,
+ "content": "\\mathrm { C } { + } { + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 500,
+ 505,
+ 513
+ ],
+ "score": 1.0,
+ "content": "library for compiling and optimizing kernels that is more",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 524
+ ],
+ "score": 1.0,
+ "content": "similar to Halide (Ragan-Kelley et al., 2013) than to our work. Our work treats the creation of neural",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 535
+ ],
+ "score": 1.0,
+ "content": "networks as a compilers problem to be addressed using mature compiler techniques. SystemML does",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 533,
+ 506,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 533,
+ 506,
+ 547
+ ],
+ "score": 1.0,
+ "content": "not consider this issue at all; TACO does use compiler optimization, but only at a very low level to",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 545,
+ 220,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 545,
+ 220,
+ 556
+ ],
+ "score": 1.0,
+ "content": "generate individual kernels.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 17.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 561,
+ 505,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 561,
+ 506,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 561,
+ 506,
+ 573
+ ],
+ "score": 1.0,
+ "content": "Two projects closely related to this work are the TensorFlow XLA compiler and the NNVM compiler.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "The code representation in these frameworks is a “sea of nodes” representation, embedding control",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 583,
+ 506,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 583,
+ 506,
+ 596
+ ],
+ "score": 1.0,
+ "content": "flow nodes and composite nodes in a data flow graph. To apply algorithmic differentiation on this",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 594,
+ 505,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 594,
+ 505,
+ 607
+ ],
+ "score": 1.0,
+ "content": "IR requires non-standard processing. In contrast, our approach is designed from the start around the",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 604,
+ 506,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 506,
+ 619
+ ],
+ "score": 1.0,
+ "content": "idea that a neural network (and its associated tensor computations) is itself a program, which is best",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 629
+ ],
+ "score": 1.0,
+ "content": "optimized through robust application of mature techniques in a principled compilation pipeline. We",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "represent tensor computation in static single assignment (SSA) form with control flow graph, and",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 507,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 507,
+ 652
+ ],
+ "score": 1.0,
+ "content": "perform algorithmic differentiation, domain-specific optimizations, general-purpose optimizations,",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 289,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 289,
+ 661
+ ],
+ "score": 1.0,
+ "content": "low-level optimizations, and code generation.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 666,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "XLA takes a similar approach to ours, transforming TensorFlow sub-graphs to XLA’s HLO graph",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "and performing optimizations. Our intermediate representation is much more expressive than XLA’s",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "by including modular IR components and general-purpose instructions; this enables our approach",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "to support full-fledged DSLs including standalone compiled DSLs and perform more extensive",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "optimizations such as inlining and interprocedual optimizations. Our approach also differs from XLA",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "score": 1.0,
+ "content": "by representing composite functions such as min and max directly through primitive instructions",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 33.5
+ }
+ ],
+ "page_idx": 1,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 224,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 225,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 225,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Workshop track - ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 310,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 310,
+ 763
+ ],
+ "score": 1.0,
+ "content": "2",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 107,
+ 79,
+ 501,
+ 281
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 107,
+ 79,
+ 501,
+ 281
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 79,
+ 501,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 79,
+ 501,
+ 281
+ ],
+ "score": 0.972,
+ "type": "image",
+ "image_path": "849b709b69e101ea7ef26d6864d1b1e0eafec6758703952e4749d0cbe5bb8354.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 107,
+ 79,
+ 501,
+ 146.33333333333331
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 107,
+ 146.33333333333331,
+ 501,
+ 213.66666666666663
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 107,
+ 213.66666666666663,
+ 501,
+ 280.99999999999994
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 104,
+ 291,
+ 501,
+ 303
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 290,
+ 501,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 290,
+ 501,
+ 304
+ ],
+ "score": 1.0,
+ "content": "Figure 1: Software stack of the DLVM infrastructure. Blue components are the compiler framework.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 325,
+ 506,
+ 424
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 325,
+ 506,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 325,
+ 506,
+ 338
+ ],
+ "score": 1.0,
+ "content": "We introduce DLVM, a new compiler infrastructure for deep learning systems that addresses shortcom-",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 336,
+ 506,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 336,
+ 506,
+ 348
+ ],
+ "score": 1.0,
+ "content": "ings of existing deep learning frameworks. Our solution includes (1) a domain-specific intermediate",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 104,
+ 346,
+ 506,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 346,
+ 506,
+ 360
+ ],
+ "score": 1.0,
+ "content": "representation specifically designed for tensor computation, (2) principled use of modern compiler",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 357,
+ 506,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 357,
+ 506,
+ 371
+ ],
+ "score": 1.0,
+ "content": "optimization techniques to substantially simplify neural network computation, including algebra sim-",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 369,
+ 506,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 369,
+ 506,
+ 382
+ ],
+ "score": 1.0,
+ "content": "plification, AD checkpointing, compute kernel fusion, and various traditional compiler optimizations,",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 380,
+ 506,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 380,
+ 506,
+ 393
+ ],
+ "score": 1.0,
+ "content": "(3) code generation through a mature compiler infrastructure that allows for transparent targeting of",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 391,
+ 506,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 391,
+ 506,
+ 404
+ ],
+ "score": 1.0,
+ "content": "various hardware, and (4) an embedded DSL that supports static analysis, type safety, and natural",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 401,
+ 506,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 401,
+ 506,
+ 416
+ ],
+ "score": 1.0,
+ "content": "expression of tensor computation, and has a just-in-time (JIT) compiler targeting DLVM for AD,",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 413,
+ 250,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 413,
+ 250,
+ 426
+ ],
+ "score": 1.0,
+ "content": "optimizations, and code generation.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 8,
+ "bbox_fs": [
+ 104,
+ 325,
+ 506,
+ 426
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 441,
+ 211,
+ 454
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 440,
+ 213,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 440,
+ 213,
+ 457
+ ],
+ "score": 1.0,
+ "content": "2 RELATED WORK",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 467,
+ 505,
+ 555
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 505,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 505,
+ 480
+ ],
+ "score": 1.0,
+ "content": "Numerous existing projects provide specialized systems for machine learning but are not closely",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 477,
+ 505,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 477,
+ 505,
+ 492
+ ],
+ "score": 1.0,
+ "content": "related to our work. These include Apache SystemML (Ghoting et al., 2011), a high-level language",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 489,
+ 505,
+ 502
+ ],
+ "score": 1.0,
+ "content": "and framework for writing and executing machine learning problems targeting Apache Spark, and",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 505,
+ 513
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 241,
+ 513
+ ],
+ "score": 1.0,
+ "content": "TACO (Kjolstad et al., 2017), a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 241,
+ 501,
+ 262,
+ 511
+ ],
+ "score": 0.85,
+ "content": "\\mathrm { C } { + } { + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 263,
+ 500,
+ 505,
+ 513
+ ],
+ "score": 1.0,
+ "content": "library for compiling and optimizing kernels that is more",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 505,
+ 524
+ ],
+ "score": 1.0,
+ "content": "similar to Halide (Ragan-Kelley et al., 2013) than to our work. Our work treats the creation of neural",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 535
+ ],
+ "score": 1.0,
+ "content": "networks as a compilers problem to be addressed using mature compiler techniques. SystemML does",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 533,
+ 506,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 533,
+ 506,
+ 547
+ ],
+ "score": 1.0,
+ "content": "not consider this issue at all; TACO does use compiler optimization, but only at a very low level to",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 545,
+ 220,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 545,
+ 220,
+ 556
+ ],
+ "score": 1.0,
+ "content": "generate individual kernels.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 17.5,
+ "bbox_fs": [
+ 105,
+ 466,
+ 506,
+ 556
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 561,
+ 505,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 561,
+ 506,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 561,
+ 506,
+ 573
+ ],
+ "score": 1.0,
+ "content": "Two projects closely related to this work are the TensorFlow XLA compiler and the NNVM compiler.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "The code representation in these frameworks is a “sea of nodes” representation, embedding control",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 583,
+ 506,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 583,
+ 506,
+ 596
+ ],
+ "score": 1.0,
+ "content": "flow nodes and composite nodes in a data flow graph. To apply algorithmic differentiation on this",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 594,
+ 505,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 594,
+ 505,
+ 607
+ ],
+ "score": 1.0,
+ "content": "IR requires non-standard processing. In contrast, our approach is designed from the start around the",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 604,
+ 506,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 506,
+ 619
+ ],
+ "score": 1.0,
+ "content": "idea that a neural network (and its associated tensor computations) is itself a program, which is best",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 629
+ ],
+ "score": 1.0,
+ "content": "optimized through robust application of mature techniques in a principled compilation pipeline. We",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "represent tensor computation in static single assignment (SSA) form with control flow graph, and",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 507,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 507,
+ 652
+ ],
+ "score": 1.0,
+ "content": "perform algorithmic differentiation, domain-specific optimizations, general-purpose optimizations,",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 289,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 289,
+ 661
+ ],
+ "score": 1.0,
+ "content": "low-level optimizations, and code generation.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 26,
+ "bbox_fs": [
+ 105,
+ 561,
+ 507,
+ 661
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 666,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "XLA takes a similar approach to ours, transforming TensorFlow sub-graphs to XLA’s HLO graph",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 690
+ ],
+ "score": 1.0,
+ "content": "and performing optimizations. Our intermediate representation is much more expressive than XLA’s",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "by including modular IR components and general-purpose instructions; this enables our approach",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "to support full-fledged DSLs including standalone compiled DSLs and perform more extensive",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "optimizations such as inlining and interprocedual optimizations. Our approach also differs from XLA",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 506,
+ 734
+ ],
+ "score": 1.0,
+ "content": "by representing composite functions such as min and max directly through primitive instructions",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 246,
+ 505,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 246,
+ 505,
+ 259
+ ],
+ "score": 1.0,
+ "content": "such as compare and select, which enables us to apply generic AD, and by using SSA form",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 257,
+ 505,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 257,
+ 505,
+ 270
+ ],
+ "score": 1.0,
+ "content": "with control flow graph, which allows for reusing battle-tested SSA optimization algorithms in the",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 267,
+ 506,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 267,
+ 506,
+ 281
+ ],
+ "score": 1.0,
+ "content": "LLVM community. Importantly, our entire infrastructure was designed from the start around a",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 104,
+ 278,
+ 506,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 278,
+ 506,
+ 293
+ ],
+ "score": 1.0,
+ "content": "robust compile-time framework for tensor DSLs, whereas XLA has been adapted around the existing",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 288,
+ 506,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 288,
+ 506,
+ 304
+ ],
+ "score": 1.0,
+ "content": "TensorFlow infrastructure with a particular focus on hardware support for Google’s Tensor Processing",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 302,
+ 215,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 302,
+ 215,
+ 313
+ ],
+ "score": 1.0,
+ "content": "Units (Jouppi et al., 2017).",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 33.5,
+ "bbox_fs": [
+ 105,
+ 665,
+ 506,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 105,
+ 78,
+ 509,
+ 203
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 105,
+ 78,
+ 509,
+ 203
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 78,
+ 509,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 78,
+ 509,
+ 203
+ ],
+ "score": 0.959,
+ "type": "image",
+ "image_path": "3c33c1e1b4e49d79cea6b0b96005ddd908e5cf9752dd02376466dc1917156439.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 105,
+ 78,
+ 509,
+ 119.66666666666666
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 119.66666666666666,
+ 509,
+ 161.33333333333331
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 161.33333333333331,
+ 509,
+ 202.99999999999997
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 172,
+ 213,
+ 434,
+ 225
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 172,
+ 210,
+ 435,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 172,
+ 210,
+ 435,
+ 228
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Compilation stages in the DLVM compilation pipeline.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 246,
+ 505,
+ 313
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 246,
+ 505,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 246,
+ 505,
+ 259
+ ],
+ "score": 1.0,
+ "content": "such as compare and select, which enables us to apply generic AD, and by using SSA form",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 257,
+ 505,
+ 270
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 257,
+ 505,
+ 270
+ ],
+ "score": 1.0,
+ "content": "with control flow graph, which allows for reusing battle-tested SSA optimization algorithms in the",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 267,
+ 506,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 267,
+ 506,
+ 281
+ ],
+ "score": 1.0,
+ "content": "LLVM community. Importantly, our entire infrastructure was designed from the start around a",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 104,
+ 278,
+ 506,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 278,
+ 506,
+ 293
+ ],
+ "score": 1.0,
+ "content": "robust compile-time framework for tensor DSLs, whereas XLA has been adapted around the existing",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 288,
+ 506,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 288,
+ 506,
+ 304
+ ],
+ "score": 1.0,
+ "content": "TensorFlow infrastructure with a particular focus on hardware support for Google’s Tensor Processing",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 302,
+ 215,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 302,
+ 215,
+ 313
+ ],
+ "score": 1.0,
+ "content": "Units (Jouppi et al., 2017).",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 318,
+ 505,
+ 417
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 317,
+ 506,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 317,
+ 420,
+ 331
+ ],
+ "score": 1.0,
+ "content": "Where TVM and NNVM are built as a DSL and a graph library in Python with a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 420,
+ 319,
+ 439,
+ 329
+ ],
+ "score": 0.85,
+ "content": "\\mathrm { C } { + } { + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 440,
+ 317,
+ 506,
+ 331
+ ],
+ "score": 1.0,
+ "content": "implementation,",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 104,
+ 327,
+ 507,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 327,
+ 507,
+ 343
+ ],
+ "score": 1.0,
+ "content": "DLVM’s architecture is closer to LLVM and the Swift Intermediate Language (Groff & Lattner,",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 339,
+ 505,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 339,
+ 505,
+ 353
+ ],
+ "score": 1.0,
+ "content": "2015), having an IR file format and a full-fledged command line toolchain. More specifically, our",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 350,
+ 506,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 350,
+ 506,
+ 364
+ ],
+ "score": 1.0,
+ "content": "work differs from NNVM and XLA in the design and presence of an IR that has a textual parsable",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 361,
+ 506,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 506,
+ 375
+ ],
+ "score": 1.0,
+ "content": "format, a module/function/basic block hierarchy, custom type declarations and memory semantics.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 372,
+ 506,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 372,
+ 506,
+ 386
+ ],
+ "score": 1.0,
+ "content": "The textual IR enables robust unit testing via FileCheck, which is used extensively in LLVM and",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 397
+ ],
+ "score": 1.0,
+ "content": "most LLVM-based compilers. Moreover, DLVM and its associated DSLs are implemented entirely",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 104,
+ 394,
+ 506,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 394,
+ 506,
+ 409
+ ],
+ "score": 1.0,
+ "content": "in Swift, a safe systems programming language, and thus have an elegantly compact codebase and",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 407,
+ 170,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 407,
+ 170,
+ 417
+ ],
+ "score": 1.0,
+ "content": "type-safe APIs.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 434,
+ 163,
+ 447
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 433,
+ 165,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 433,
+ 165,
+ 450
+ ],
+ "score": 1.0,
+ "content": "3 DLVM",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 461,
+ 505,
+ 527
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 475
+ ],
+ "score": 1.0,
+ "content": "Deep Learning Virtual Machine (DLVM) is a compiler infrastructure designed for modern deep",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "learning systems.1 DLVM is designed to apply a multi-stage compiler optimization strategy to both",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 483,
+ 506,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 483,
+ 506,
+ 496
+ ],
+ "score": 1.0,
+ "content": "high-level linear algebra and low-level parallelism, perform domain-specific transformations, relieve",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 494,
+ 506,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 506,
+ 507
+ ],
+ "score": 1.0,
+ "content": "the overhead in front-end languages, and serve as the host for research and development of DSLs for",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 504,
+ 507,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 504,
+ 507,
+ 519
+ ],
+ "score": 1.0,
+ "content": "deep learning. The complete DLVM software stack, including sample front-end deep learning DSLs,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 515,
+ 283,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 283,
+ 530
+ ],
+ "score": 1.0,
+ "content": "is shown in Figure 1 on the preceding page.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 532,
+ 505,
+ 599
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 546
+ ],
+ "score": 1.0,
+ "content": "Figure 2 illustrates the major stages in the DLVM compilation pipeline. The DLVM compilation stages",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 543,
+ 506,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 543,
+ 506,
+ 557
+ ],
+ "score": 1.0,
+ "content": "address algorithmic differentiation, domain-specific optimizations, general-purpose optimizations,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 506,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 506,
+ 568
+ ],
+ "score": 1.0,
+ "content": "and static code generation targeting a variety of compute architectures. The raw and optimizable",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": "stages allow constructs for high-level tensor operations and various high-level optimizations. The",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 590
+ ],
+ "score": 1.0,
+ "content": "compute and schedule stages allow constructs for low-level array operations lowered from tensor",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 588,
+ 484,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 588,
+ 484,
+ 601
+ ],
+ "score": 1.0,
+ "content": "operations in high-level stages, borrowing the design from Halide (Ragan-Kelley et al., 2013).",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 28.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 604,
+ 505,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "The DLVM Intermediate Representation (IR) is the core language of the system. It uses static single",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "assignment (SSA) form, control flow graphs, high-level types including a first-class tensor type, and",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "a set of linear algebra operators combined with a general-purpose instruction set (see Table 1). The",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "system enables a wide variety of domain-specific analyses and transformations, such as reverse-mode",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 396,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 396,
+ 660
+ ],
+ "score": 1.0,
+ "content": "AD, AD checkpointing, algebra simplification and linear algebra fusion.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 665,
+ 504,
+ 709
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "To demonstrate how DLVM helps the development of domain-specific languages (DSLs), in Sec-",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 689
+ ],
+ "score": 1.0,
+ "content": "tion 3.4 we present one prototypical staged DSL: NNKit. NNKit features safe and natural expression",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "of tensor computation alongside the host program, and targets DLVM for differentiation, optimizations",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 214,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 214,
+ 711
+ ],
+ "score": 1.0,
+ "content": "and static code generation.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 38.5
+ }
+ ],
+ "page_idx": 2,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 119,
+ 722,
+ 389,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 119,
+ 720,
+ 390,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 119,
+ 720,
+ 390,
+ 733
+ ],
+ "score": 1.0,
+ "content": "1Code for DLVM is available at https://github.com/dlvm-team",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 224,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 225,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 225,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Workshop track - ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "score": 1.0,
+ "content": "3",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 105,
+ 78,
+ 509,
+ 203
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 105,
+ 78,
+ 509,
+ 203
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 78,
+ 509,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 78,
+ 509,
+ 203
+ ],
+ "score": 0.959,
+ "type": "image",
+ "image_path": "3c33c1e1b4e49d79cea6b0b96005ddd908e5cf9752dd02376466dc1917156439.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 105,
+ 78,
+ 509,
+ 119.66666666666666
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 119.66666666666666,
+ 509,
+ 161.33333333333331
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 161.33333333333331,
+ 509,
+ 202.99999999999997
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 172,
+ 213,
+ 434,
+ 225
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 172,
+ 210,
+ 435,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 172,
+ 210,
+ 435,
+ 228
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Compilation stages in the DLVM compilation pipeline.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 246,
+ 505,
+ 313
+ ],
+ "lines": [],
+ "index": 6.5,
+ "bbox_fs": [
+ 104,
+ 246,
+ 506,
+ 313
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 318,
+ 505,
+ 417
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 317,
+ 506,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 317,
+ 420,
+ 331
+ ],
+ "score": 1.0,
+ "content": "Where TVM and NNVM are built as a DSL and a graph library in Python with a",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 420,
+ 319,
+ 439,
+ 329
+ ],
+ "score": 0.85,
+ "content": "\\mathrm { C } { + } { + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 440,
+ 317,
+ 506,
+ 331
+ ],
+ "score": 1.0,
+ "content": "implementation,",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 104,
+ 327,
+ 507,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 327,
+ 507,
+ 343
+ ],
+ "score": 1.0,
+ "content": "DLVM’s architecture is closer to LLVM and the Swift Intermediate Language (Groff & Lattner,",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 339,
+ 505,
+ 353
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 339,
+ 505,
+ 353
+ ],
+ "score": 1.0,
+ "content": "2015), having an IR file format and a full-fledged command line toolchain. More specifically, our",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 350,
+ 506,
+ 364
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 350,
+ 506,
+ 364
+ ],
+ "score": 1.0,
+ "content": "work differs from NNVM and XLA in the design and presence of an IR that has a textual parsable",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 361,
+ 506,
+ 375
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 361,
+ 506,
+ 375
+ ],
+ "score": 1.0,
+ "content": "format, a module/function/basic block hierarchy, custom type declarations and memory semantics.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 372,
+ 506,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 372,
+ 506,
+ 386
+ ],
+ "score": 1.0,
+ "content": "The textual IR enables robust unit testing via FileCheck, which is used extensively in LLVM and",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 505,
+ 397
+ ],
+ "score": 1.0,
+ "content": "most LLVM-based compilers. Moreover, DLVM and its associated DSLs are implemented entirely",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 104,
+ 394,
+ 506,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 394,
+ 506,
+ 409
+ ],
+ "score": 1.0,
+ "content": "in Swift, a safe systems programming language, and thus have an elegantly compact codebase and",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 407,
+ 170,
+ 417
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 407,
+ 170,
+ 417
+ ],
+ "score": 1.0,
+ "content": "type-safe APIs.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 14,
+ "bbox_fs": [
+ 104,
+ 317,
+ 507,
+ 417
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 434,
+ 163,
+ 447
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 433,
+ 165,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 433,
+ 165,
+ 450
+ ],
+ "score": 1.0,
+ "content": "3 DLVM",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 461,
+ 505,
+ 527
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 460,
+ 505,
+ 475
+ ],
+ "score": 1.0,
+ "content": "Deep Learning Virtual Machine (DLVM) is a compiler infrastructure designed for modern deep",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "learning systems.1 DLVM is designed to apply a multi-stage compiler optimization strategy to both",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 483,
+ 506,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 483,
+ 506,
+ 496
+ ],
+ "score": 1.0,
+ "content": "high-level linear algebra and low-level parallelism, perform domain-specific transformations, relieve",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 494,
+ 506,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 506,
+ 507
+ ],
+ "score": 1.0,
+ "content": "the overhead in front-end languages, and serve as the host for research and development of DSLs for",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 504,
+ 507,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 504,
+ 507,
+ 519
+ ],
+ "score": 1.0,
+ "content": "deep learning. The complete DLVM software stack, including sample front-end deep learning DSLs,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 515,
+ 283,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 515,
+ 283,
+ 530
+ ],
+ "score": 1.0,
+ "content": "is shown in Figure 1 on the preceding page.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22.5,
+ "bbox_fs": [
+ 105,
+ 460,
+ 507,
+ 530
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 532,
+ 505,
+ 599
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 546
+ ],
+ "score": 1.0,
+ "content": "Figure 2 illustrates the major stages in the DLVM compilation pipeline. The DLVM compilation stages",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 543,
+ 506,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 543,
+ 506,
+ 557
+ ],
+ "score": 1.0,
+ "content": "address algorithmic differentiation, domain-specific optimizations, general-purpose optimizations,",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 555,
+ 506,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 555,
+ 506,
+ 568
+ ],
+ "score": 1.0,
+ "content": "and static code generation targeting a variety of compute architectures. The raw and optimizable",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 505,
+ 579
+ ],
+ "score": 1.0,
+ "content": "stages allow constructs for high-level tensor operations and various high-level optimizations. The",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 577,
+ 506,
+ 590
+ ],
+ "score": 1.0,
+ "content": "compute and schedule stages allow constructs for low-level array operations lowered from tensor",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 588,
+ 484,
+ 601
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 588,
+ 484,
+ 601
+ ],
+ "score": 1.0,
+ "content": "operations in high-level stages, borrowing the design from Halide (Ragan-Kelley et al., 2013).",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 28.5,
+ "bbox_fs": [
+ 105,
+ 532,
+ 506,
+ 601
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 604,
+ 505,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 505,
+ 618
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 505,
+ 618
+ ],
+ "score": 1.0,
+ "content": "The DLVM Intermediate Representation (IR) is the core language of the system. It uses static single",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 505,
+ 628
+ ],
+ "score": 1.0,
+ "content": "assignment (SSA) form, control flow graphs, high-level types including a first-class tensor type, and",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "a set of linear algebra operators combined with a general-purpose instruction set (see Table 1). The",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 637,
+ 505,
+ 650
+ ],
+ "score": 1.0,
+ "content": "system enables a wide variety of domain-specific analyses and transformations, such as reverse-mode",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 649,
+ 396,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 649,
+ 396,
+ 660
+ ],
+ "score": 1.0,
+ "content": "AD, AD checkpointing, algebra simplification and linear algebra fusion.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 105,
+ 604,
+ 505,
+ 660
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 665,
+ 504,
+ 709
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 664,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "To demonstrate how DLVM helps the development of domain-specific languages (DSLs), in Sec-",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 689
+ ],
+ "score": 1.0,
+ "content": "tion 3.4 we present one prototypical staged DSL: NNKit. NNKit features safe and natural expression",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "of tensor computation alongside the host program, and targets DLVM for differentiation, optimizations",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 214,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 214,
+ 711
+ ],
+ "score": 1.0,
+ "content": "and static code generation.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 38.5,
+ "bbox_fs": [
+ 105,
+ 664,
+ 506,
+ 711
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 106,
+ 81,
+ 507,
+ 227
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 106,
+ 81,
+ 507,
+ 227
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 81,
+ 507,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 81,
+ 507,
+ 227
+ ],
+ "score": 0.983,
+ "html": "
| Kind | Example |
| Element-wise unaryElement-wise binaryDot | | |
| dot %a:<10 x | |
| ReduceTransposeSliceCompareData type cast | | |
| slice %a: | |
| Function applicationBranchConditional branchShape cast | arpl | ly% |
| bran | nch |
",
+ "type": "table",
+ "image_path": "20d98aab7a123e6c09753bfb3cc8dfb7035594771e0bb772d661ba37e0d7df18.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 81,
+ 507,
+ 129.66666666666666
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 129.66666666666666,
+ 507,
+ 178.33333333333331
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 178.33333333333331,
+ 507,
+ 226.99999999999997
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 239,
+ 506,
+ 272
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 238,
+ 505,
+ 251
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 238,
+ 505,
+ 251
+ ],
+ "score": 1.0,
+ "content": "Table 1: This table illustrates a selection of the instructions in the DLVM virtual instruction set. The",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 250,
+ 504,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 250,
+ 504,
+ 261
+ ],
+ "score": 1.0,
+ "content": "instruction set includes linear algebra operations such as tanh and dot in addition to control flow",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 261,
+ 227,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 261,
+ 227,
+ 272
+ ],
+ "score": 1.0,
+ "content": "instructions such as branch.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 298,
+ 189,
+ 310
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 297,
+ 191,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 297,
+ 191,
+ 311
+ ],
+ "score": 1.0,
+ "content": "3.1 DLVM CORE",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 320,
+ 505,
+ 375
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 319,
+ 505,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 319,
+ 505,
+ 333
+ ],
+ "score": 1.0,
+ "content": "DLVM Core contains essential components for an optimizing compiler: IR, pass manager, and passes",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 331,
+ 505,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 331,
+ 505,
+ 344
+ ],
+ "score": 1.0,
+ "content": "(see Figure 1 on page 2). The DLVM IR consists of a virtual instruction set, control flow graph",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 342,
+ 505,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 342,
+ 505,
+ 354
+ ],
+ "score": 1.0,
+ "content": "and data flow representation. Passes are functions that traverse the intermediate representation of a",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 352,
+ 506,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 352,
+ 506,
+ 366
+ ],
+ "score": 1.0,
+ "content": "program, either producing useful results as analyses of the program (analysis passes), or mutating the",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 363,
+ 368,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 363,
+ 368,
+ 377
+ ],
+ "score": 1.0,
+ "content": "program for differentiation and optimizations (transform passes).",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 389,
+ 474,
+ 401
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 389,
+ 477,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 389,
+ 477,
+ 402
+ ],
+ "score": 1.0,
+ "content": "3.1.1 A DOMAIN-SPECIFIC COMPILER INTERMEDIATE REPRESENTATION FOR DLVM",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 409,
+ 505,
+ 498
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 505,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 505,
+ 422
+ ],
+ "score": 1.0,
+ "content": "Inspired by the LLVM IR (Lattner & Adve, 2004) and the Swift Intermediate Language (Groff &",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 104,
+ 420,
+ 506,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 420,
+ 506,
+ 435
+ ],
+ "score": 1.0,
+ "content": "Lattner, 2015), DLVM IR is a graph-based, modular code representation, with both an in-memory",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 433,
+ 506,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 433,
+ 506,
+ 444
+ ],
+ "score": 1.0,
+ "content": "format and a textual format. The code representation has a hierarchy of abstractions: module,",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 456
+ ],
+ "score": 1.0,
+ "content": "function, basic block, and instruction. An instruction is the minimal unit of code that operates",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 453,
+ 506,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 506,
+ 467
+ ],
+ "score": 1.0,
+ "content": "on values, which can be globals, function arguments or temporary virtual registers produced by",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 465,
+ 506,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 506,
+ 477
+ ],
+ "score": 1.0,
+ "content": "instructions. Each module contains a collection of type definitions, global values and functions. Each",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 476,
+ 506,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 476,
+ 506,
+ 487
+ ],
+ "score": 1.0,
+ "content": "function has a control flow graph formed by basic blocks and control flow edges. Each basic block",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 487,
+ 490,
+ 499
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 487,
+ 490,
+ 499
+ ],
+ "score": 1.0,
+ "content": "contains an ordered list of instructions with data dependencies forming a directed acyclic graph.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 16.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 504,
+ 505,
+ 581
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 504,
+ 505,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 504,
+ 505,
+ 516
+ ],
+ "score": 1.0,
+ "content": "The DLVM IR has a high-level type system with tensor as a first-class type. The DLVM virtual",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 515,
+ 506,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 515,
+ 506,
+ 528
+ ],
+ "score": 1.0,
+ "content": "instruction set includes domain-specific primitive math operators, as well as general-purpose instruc-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 526,
+ 505,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 526,
+ 505,
+ 538
+ ],
+ "score": 1.0,
+ "content": "tions for memory management, control flow and function application. Domain-specific instructions",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 506,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 506,
+ 549
+ ],
+ "score": 1.0,
+ "content": "include element-wise unary operators, such as tanh and negate, element-wise binary operators,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 549,
+ 505,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 549,
+ 505,
+ 559
+ ],
+ "score": 1.0,
+ "content": "such as add and power, and complex operators such as dot, transpose, and convolve. All",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 558,
+ 505,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 558,
+ 505,
+ 570
+ ],
+ "score": 1.0,
+ "content": "element-wise binary operators support broadcasting. A sample of DLVM IR code is shown in Figure 3",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 569,
+ 177,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 569,
+ 177,
+ 583
+ ],
+ "score": 1.0,
+ "content": "on the next page.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 586,
+ 505,
+ 631
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 586,
+ 507,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 586,
+ 507,
+ 599
+ ],
+ "score": 1.0,
+ "content": "The DLVM instruction set does not include composite math functions such as softmax, sigmoid,",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 597,
+ 505,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 597,
+ 505,
+ 609
+ ],
+ "score": 1.0,
+ "content": "min or max. All of these functions can be composed of primitive math instructions and control flow",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 609,
+ 505,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 609,
+ 505,
+ 621
+ ],
+ "score": 1.0,
+ "content": "constructs. This design allows for the standard AD algorithm to be applied to any differentiable",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 619,
+ 362,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 619,
+ 362,
+ 632
+ ],
+ "score": 1.0,
+ "content": "program, with no need for special handling of composite cases.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 645,
+ 353,
+ 657
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 645,
+ 354,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 645,
+ 354,
+ 658
+ ],
+ "score": 1.0,
+ "content": "3.1.2 DOMAIN-SPECIFIC COMPILER PASSES FOR DLVM",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 666,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "DLVM has a full-fledged pass infrastructure, performing various analyses and two kinds of trans-",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "formations: differentiation and optimization. Differentiation constructs function definitions from",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "gradient declarations using adjoint code generation (see Section 3.1.3 below). Optimization is then",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "performed on the resulting IR, maximizing the code performance. Optimizations include domain-",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "specific optimizations, such as algebra simplification, linear algebra fusion, matrix multiplication",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 405,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 405,
+ 733
+ ],
+ "score": 1.0,
+ "content": "reordering, and AD checkpointing, and traditional compiler optimizations.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 35.5
+ }
+ ],
+ "page_idx": 3,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 224,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 225,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 225,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Workshop track - ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 759
+ ],
+ "lines": []
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 106,
+ 81,
+ 507,
+ 227
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 106,
+ 81,
+ 507,
+ 227
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 81,
+ 507,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 81,
+ 507,
+ 227
+ ],
+ "score": 0.983,
+ "html": "| Kind | Example |
| Element-wise unaryElement-wise binaryDot | | |
| dot %a:<10 x | |
| ReduceTransposeSliceCompareData type cast | | |
| slice %a: | |
| Function applicationBranchConditional branchShape cast | arpl | ly% |
| bran | nch |
",
+ "type": "table",
+ "image_path": "20d98aab7a123e6c09753bfb3cc8dfb7035594771e0bb772d661ba37e0d7df18.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 106,
+ 81,
+ 507,
+ 129.66666666666666
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 106,
+ 129.66666666666666,
+ 507,
+ 178.33333333333331
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 106,
+ 178.33333333333331,
+ 507,
+ 226.99999999999997
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_caption",
+ "bbox": [
+ 106,
+ 239,
+ 506,
+ 272
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 238,
+ 505,
+ 251
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 238,
+ 505,
+ 251
+ ],
+ "score": 1.0,
+ "content": "Table 1: This table illustrates a selection of the instructions in the DLVM virtual instruction set. The",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 250,
+ 504,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 250,
+ 504,
+ 261
+ ],
+ "score": 1.0,
+ "content": "instruction set includes linear algebra operations such as tanh and dot in addition to control flow",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 261,
+ 227,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 261,
+ 227,
+ 272
+ ],
+ "score": 1.0,
+ "content": "instructions such as branch.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 298,
+ 189,
+ 310
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 297,
+ 191,
+ 311
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 297,
+ 191,
+ 311
+ ],
+ "score": 1.0,
+ "content": "3.1 DLVM CORE",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 320,
+ 505,
+ 375
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 319,
+ 505,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 319,
+ 505,
+ 333
+ ],
+ "score": 1.0,
+ "content": "DLVM Core contains essential components for an optimizing compiler: IR, pass manager, and passes",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 331,
+ 505,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 331,
+ 505,
+ 344
+ ],
+ "score": 1.0,
+ "content": "(see Figure 1 on page 2). The DLVM IR consists of a virtual instruction set, control flow graph",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 342,
+ 505,
+ 354
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 342,
+ 505,
+ 354
+ ],
+ "score": 1.0,
+ "content": "and data flow representation. Passes are functions that traverse the intermediate representation of a",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 352,
+ 506,
+ 366
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 352,
+ 506,
+ 366
+ ],
+ "score": 1.0,
+ "content": "program, either producing useful results as analyses of the program (analysis passes), or mutating the",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 363,
+ 368,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 363,
+ 368,
+ 377
+ ],
+ "score": 1.0,
+ "content": "program for differentiation and optimizations (transform passes).",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9,
+ "bbox_fs": [
+ 105,
+ 319,
+ 506,
+ 377
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 389,
+ 474,
+ 401
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 389,
+ 477,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 389,
+ 477,
+ 402
+ ],
+ "score": 1.0,
+ "content": "3.1.1 A DOMAIN-SPECIFIC COMPILER INTERMEDIATE REPRESENTATION FOR DLVM",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 12
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 409,
+ 505,
+ 498
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 505,
+ 422
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 410,
+ 505,
+ 422
+ ],
+ "score": 1.0,
+ "content": "Inspired by the LLVM IR (Lattner & Adve, 2004) and the Swift Intermediate Language (Groff &",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 104,
+ 420,
+ 506,
+ 435
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 420,
+ 506,
+ 435
+ ],
+ "score": 1.0,
+ "content": "Lattner, 2015), DLVM IR is a graph-based, modular code representation, with both an in-memory",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 433,
+ 506,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 433,
+ 506,
+ 444
+ ],
+ "score": 1.0,
+ "content": "format and a textual format. The code representation has a hierarchy of abstractions: module,",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 456
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 456
+ ],
+ "score": 1.0,
+ "content": "function, basic block, and instruction. An instruction is the minimal unit of code that operates",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 453,
+ 506,
+ 467
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 506,
+ 467
+ ],
+ "score": 1.0,
+ "content": "on values, which can be globals, function arguments or temporary virtual registers produced by",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 465,
+ 506,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 506,
+ 477
+ ],
+ "score": 1.0,
+ "content": "instructions. Each module contains a collection of type definitions, global values and functions. Each",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 476,
+ 506,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 476,
+ 506,
+ 487
+ ],
+ "score": 1.0,
+ "content": "function has a control flow graph formed by basic blocks and control flow edges. Each basic block",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 487,
+ 490,
+ 499
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 487,
+ 490,
+ 499
+ ],
+ "score": 1.0,
+ "content": "contains an ordered list of instructions with data dependencies forming a directed acyclic graph.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 16.5,
+ "bbox_fs": [
+ 104,
+ 410,
+ 506,
+ 499
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 504,
+ 505,
+ 581
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 504,
+ 505,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 504,
+ 505,
+ 516
+ ],
+ "score": 1.0,
+ "content": "The DLVM IR has a high-level type system with tensor as a first-class type. The DLVM virtual",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 515,
+ 506,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 515,
+ 506,
+ 528
+ ],
+ "score": 1.0,
+ "content": "instruction set includes domain-specific primitive math operators, as well as general-purpose instruc-",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 526,
+ 505,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 526,
+ 505,
+ 538
+ ],
+ "score": 1.0,
+ "content": "tions for memory management, control flow and function application. Domain-specific instructions",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 536,
+ 506,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 536,
+ 506,
+ 549
+ ],
+ "score": 1.0,
+ "content": "include element-wise unary operators, such as tanh and negate, element-wise binary operators,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 549,
+ 505,
+ 559
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 549,
+ 505,
+ 559
+ ],
+ "score": 1.0,
+ "content": "such as add and power, and complex operators such as dot, transpose, and convolve. All",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 558,
+ 505,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 558,
+ 505,
+ 570
+ ],
+ "score": 1.0,
+ "content": "element-wise binary operators support broadcasting. A sample of DLVM IR code is shown in Figure 3",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 569,
+ 177,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 569,
+ 177,
+ 583
+ ],
+ "score": 1.0,
+ "content": "on the next page.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 24,
+ "bbox_fs": [
+ 105,
+ 504,
+ 506,
+ 583
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 586,
+ 505,
+ 631
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 586,
+ 507,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 586,
+ 507,
+ 599
+ ],
+ "score": 1.0,
+ "content": "The DLVM instruction set does not include composite math functions such as softmax, sigmoid,",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 597,
+ 505,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 597,
+ 505,
+ 609
+ ],
+ "score": 1.0,
+ "content": "min or max. All of these functions can be composed of primitive math instructions and control flow",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 609,
+ 505,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 609,
+ 505,
+ 621
+ ],
+ "score": 1.0,
+ "content": "constructs. This design allows for the standard AD algorithm to be applied to any differentiable",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 619,
+ 362,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 619,
+ 362,
+ 632
+ ],
+ "score": 1.0,
+ "content": "program, with no need for special handling of composite cases.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 29.5,
+ "bbox_fs": [
+ 105,
+ 586,
+ 507,
+ 632
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 645,
+ 353,
+ 657
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 645,
+ 354,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 645,
+ 354,
+ 658
+ ],
+ "score": 1.0,
+ "content": "3.1.2 DOMAIN-SPECIFIC COMPILER PASSES FOR DLVM",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 666,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "DLVM has a full-fledged pass infrastructure, performing various analyses and two kinds of trans-",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "formations: differentiation and optimization. Differentiation constructs function definitions from",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "gradient declarations using adjoint code generation (see Section 3.1.3 below). Optimization is then",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "performed on the resulting IR, maximizing the code performance. Optimizations include domain-",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "specific optimizations, such as algebra simplification, linear algebra fusion, matrix multiplication",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 405,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 405,
+ 733
+ ],
+ "score": 1.0,
+ "content": "reordering, and AD checkpointing, and traditional compiler optimizations.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 35.5,
+ "bbox_fs": [
+ 105,
+ 665,
+ 506,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 108,
+ 81,
+ 501,
+ 327
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 108,
+ 81,
+ 501,
+ 327
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 81,
+ 501,
+ 327
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 81,
+ 501,
+ 327
+ ],
+ "score": 0.952,
+ "type": "image",
+ "image_path": "d717d7aa13932d973e32e917d093ebca23c6520956de73c54d87b779344714ea.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 108,
+ 81,
+ 501,
+ 163.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 108,
+ 163.0,
+ 501,
+ 245.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 108,
+ 245.0,
+ 501,
+ 327.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 344,
+ 505,
+ 378
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 343,
+ 506,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 343,
+ 506,
+ 357
+ ],
+ "score": 1.0,
+ "content": "Figure 3: Example code in DLVM intermediate representation. Note that some functions are annotated",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 355,
+ 505,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 355,
+ 505,
+ 367
+ ],
+ "score": 1.0,
+ "content": "as defining the gradient of another function with respect some or all arguments. The body of",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 366,
+ 334,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 366,
+ 334,
+ 378
+ ],
+ "score": 1.0,
+ "content": "these gradient functions will be automatically generated.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 402,
+ 505,
+ 468
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 402,
+ 506,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 402,
+ 506,
+ 415
+ ],
+ "score": 1.0,
+ "content": "Since DLVM IR is aware of mathematical operators such as tanh and power, the algebra simplifi-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 414,
+ 506,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 414,
+ 506,
+ 425
+ ],
+ "score": 1.0,
+ "content": "cation pass can find and simplify certain mathematical operations that are expensive or redundant.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 423,
+ 504,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 423,
+ 164,
+ 437
+ ],
+ "score": 1.0,
+ "content": "For example,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 164,
+ 424,
+ 176,
+ 434
+ ],
+ "score": 0.84,
+ "content": "x ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 176,
+ 423,
+ 263,
+ 437
+ ],
+ "score": 1.0,
+ "content": "can be simplified to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 263,
+ 425,
+ 290,
+ 435
+ ],
+ "score": 0.76,
+ "content": "x \\odot x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 290,
+ 423,
+ 295,
+ 437
+ ],
+ "score": 0.0,
+ "content": "",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 296,
+ 425,
+ 304,
+ 435
+ ],
+ "score": 0.75,
+ "content": "\\odot",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 305,
+ 423,
+ 492,
+ 437
+ ],
+ "score": 1.0,
+ "content": "stands for element-wise multiplication), and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 492,
+ 424,
+ 504,
+ 434
+ ],
+ "score": 0.85,
+ "content": "x ^ { 0 }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 435,
+ 505,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 505,
+ 448
+ ],
+ "score": 1.0,
+ "content": "can be simplified to constant 1. Matrix multiplication reordering is another classic optimization",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 447,
+ 505,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 447,
+ 505,
+ 458
+ ],
+ "score": 1.0,
+ "content": "that minimizes the number of sub-operations in a chain of matrix multiplications with different",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 457,
+ 354,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 457,
+ 354,
+ 470
+ ],
+ "score": 1.0,
+ "content": "dimensionality, based on matrix multiplication’s associativity.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 8.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 474,
+ 505,
+ 649
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 474,
+ 506,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 474,
+ 506,
+ 487
+ ],
+ "score": 1.0,
+ "content": "Since the DLVM optimizer is aware of linear algebra operations with static dimensionality, maxi-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 485,
+ 506,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 485,
+ 506,
+ 497
+ ],
+ "score": 1.0,
+ "content": "mizing the performance by fusing verbose linear operations into a single matrix multiplication is",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 496,
+ 503,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 496,
+ 469,
+ 509
+ ],
+ "score": 1.0,
+ "content": "beneficial as well. For example, it is very common to encounter expressions of the form",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 469,
+ 496,
+ 503,
+ 507
+ ],
+ "score": 0.89,
+ "content": "\\mathbf { W } \\mathbf { X } + \\mathbf { b }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 506,
+ 505,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 506,
+ 505,
+ 520
+ ],
+ "score": 1.0,
+ "content": "When unoptimized, the matrix multiplication and the addition will be parallelized separately. Since",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 518,
+ 505,
+ 531
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 518,
+ 505,
+ 531
+ ],
+ "score": 1.0,
+ "content": "launching compute kernels separately can be expensive, DLVM performs linear algebra fusion, which",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 529,
+ 506,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 529,
+ 506,
+ 542
+ ],
+ "score": 1.0,
+ "content": "transforms subexpressions involving both matrix multiplication and element-wise operations into a",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 540,
+ 505,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 505,
+ 552
+ ],
+ "score": 1.0,
+ "content": "single matrix multiplication instruction on padded tensors. Besides the simple pattern like an addition",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 551,
+ 505,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 551,
+ 505,
+ 563
+ ],
+ "score": 1.0,
+ "content": "of matrix multiplication and a vector, we can apply the same approach to a polynomial of multiple",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 561,
+ 506,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 506,
+ 575
+ ],
+ "score": 1.0,
+ "content": "matrix multiplications, turning the polynomial into a single matrix multiplication. For example, in a",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 573,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 573,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "simple recurrent neural network (RNN), each cell of the recurrence is a feed forward neural network",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 104,
+ 583,
+ 506,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 583,
+ 191,
+ 597
+ ],
+ "score": 1.0,
+ "content": "that takes two inputs:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 585,
+ 203,
+ 595
+ ],
+ "score": 0.85,
+ "content": "\\mathbf { x } _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 583,
+ 372,
+ 597
+ ],
+ "score": 1.0,
+ "content": ", the input local to the current timestep, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 373,
+ 584,
+ 384,
+ 595
+ ],
+ "score": 0.87,
+ "content": "\\mathbf { h } _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 384,
+ 583,
+ 506,
+ 597
+ ],
+ "score": 1.0,
+ "content": ", the hidden state carried along",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 388,
+ 608
+ ],
+ "score": 1.0,
+ "content": "the recurrence. The linear algebra fusion pass can simplify operations in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 388,
+ 595,
+ 505,
+ 606
+ ],
+ "score": 0.92,
+ "content": "\\mathbf { h } _ { t } = f ( \\mathbf { W } \\mathbf { x } _ { t - 1 } + \\mathbf { U } \\mathbf { h } _ { t - 1 } + \\mathbf { b } )",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 506,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 506,
+ 619
+ ],
+ "score": 1.0,
+ "content": "from two matrix multiplications and two additions into a single matrix multiplication. A more aggres-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 104,
+ 615,
+ 505,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 615,
+ 505,
+ 630
+ ],
+ "score": 1.0,
+ "content": "sive, interprocedural version of linear algebra fusion can optimize parameter passing and memory",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 628,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 628,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "allocation, so that the entire concatenated matrix can be created and passed around in the first place",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 639,
+ 191,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 639,
+ 191,
+ 650
+ ],
+ "score": 1.0,
+ "content": "without reallocation.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 19.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 666,
+ 455,
+ 678
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 457,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 457,
+ 679
+ ],
+ "score": 1.0,
+ "content": "3.1.3 ALGORITHMIC DIFFERENTIATION THROUGH ADJOINT CODE GENERATION",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 28
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Algorithmic differentiation (AD), also known as automatic differentiation, encompasses a family of",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 445,
+ 711
+ ],
+ "score": 1.0,
+ "content": "a well-known techniques for algorithmically obtaining the derivatives of a function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 445,
+ 699,
+ 505,
+ 710
+ ],
+ "score": 0.91,
+ "content": "f : \\mathbf { x } \\in \\mathbb { R } ^ { n } ",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 142,
+ 721
+ ],
+ "score": 0.89,
+ "content": "\\mathbf { y } \\in \\mathbb { R } ^ { m }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 142,
+ 709,
+ 278,
+ 722
+ ],
+ "score": 1.0,
+ "content": "(Naumann, 2011). The function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 278,
+ 710,
+ 285,
+ 721
+ ],
+ "score": 0.85,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 285,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "can be represented as a directed acyclic computation",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 506,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 506,
+ 733
+ ],
+ "score": 1.0,
+ "content": "graph representing the composition of elementary computational operations for which the respective",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 30.5
+ }
+ ],
+ "page_idx": 4,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 224,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 225,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 225,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Workshop track - ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 763
+ ],
+ "score": 1.0,
+ "content": "5",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 108,
+ 81,
+ 501,
+ 327
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 108,
+ 81,
+ 501,
+ 327
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 81,
+ 501,
+ 327
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 81,
+ 501,
+ 327
+ ],
+ "score": 0.952,
+ "type": "image",
+ "image_path": "d717d7aa13932d973e32e917d093ebca23c6520956de73c54d87b779344714ea.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 108,
+ 81,
+ 501,
+ 163.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 108,
+ 163.0,
+ 501,
+ 245.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 108,
+ 245.0,
+ 501,
+ 327.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 107,
+ 344,
+ 505,
+ 378
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 343,
+ 506,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 343,
+ 506,
+ 357
+ ],
+ "score": 1.0,
+ "content": "Figure 3: Example code in DLVM intermediate representation. Note that some functions are annotated",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 355,
+ 505,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 355,
+ 505,
+ 367
+ ],
+ "score": 1.0,
+ "content": "as defining the gradient of another function with respect some or all arguments. The body of",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 366,
+ 334,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 366,
+ 334,
+ 378
+ ],
+ "score": 1.0,
+ "content": "these gradient functions will be automatically generated.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 402,
+ 505,
+ 468
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 402,
+ 506,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 402,
+ 506,
+ 415
+ ],
+ "score": 1.0,
+ "content": "Since DLVM IR is aware of mathematical operators such as tanh and power, the algebra simplifi-",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 414,
+ 506,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 414,
+ 506,
+ 425
+ ],
+ "score": 1.0,
+ "content": "cation pass can find and simplify certain mathematical operations that are expensive or redundant.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 423,
+ 504,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 423,
+ 164,
+ 437
+ ],
+ "score": 1.0,
+ "content": "For example,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 164,
+ 424,
+ 176,
+ 434
+ ],
+ "score": 0.84,
+ "content": "x ^ { 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 176,
+ 423,
+ 263,
+ 437
+ ],
+ "score": 1.0,
+ "content": "can be simplified to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 263,
+ 425,
+ 290,
+ 435
+ ],
+ "score": 0.76,
+ "content": "x \\odot x",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 290,
+ 423,
+ 295,
+ 437
+ ],
+ "score": 0.0,
+ "content": "",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 296,
+ 425,
+ 304,
+ 435
+ ],
+ "score": 0.75,
+ "content": "\\odot",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 305,
+ 423,
+ 492,
+ 437
+ ],
+ "score": 1.0,
+ "content": "stands for element-wise multiplication), and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 492,
+ 424,
+ 504,
+ 434
+ ],
+ "score": 0.85,
+ "content": "x ^ { 0 }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 435,
+ 505,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 435,
+ 505,
+ 448
+ ],
+ "score": 1.0,
+ "content": "can be simplified to constant 1. Matrix multiplication reordering is another classic optimization",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 447,
+ 505,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 447,
+ 505,
+ 458
+ ],
+ "score": 1.0,
+ "content": "that minimizes the number of sub-operations in a chain of matrix multiplications with different",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 457,
+ 354,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 457,
+ 354,
+ 470
+ ],
+ "score": 1.0,
+ "content": "dimensionality, based on matrix multiplication’s associativity.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 8.5,
+ "bbox_fs": [
+ 105,
+ 402,
+ 506,
+ 470
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 474,
+ 505,
+ 649
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 474,
+ 506,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 474,
+ 506,
+ 487
+ ],
+ "score": 1.0,
+ "content": "Since the DLVM optimizer is aware of linear algebra operations with static dimensionality, maxi-",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 485,
+ 506,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 485,
+ 506,
+ 497
+ ],
+ "score": 1.0,
+ "content": "mizing the performance by fusing verbose linear operations into a single matrix multiplication is",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 496,
+ 503,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 496,
+ 469,
+ 509
+ ],
+ "score": 1.0,
+ "content": "beneficial as well. For example, it is very common to encounter expressions of the form",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 469,
+ 496,
+ 503,
+ 507
+ ],
+ "score": 0.89,
+ "content": "\\mathbf { W } \\mathbf { X } + \\mathbf { b }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 506,
+ 505,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 506,
+ 505,
+ 520
+ ],
+ "score": 1.0,
+ "content": "When unoptimized, the matrix multiplication and the addition will be parallelized separately. Since",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 518,
+ 505,
+ 531
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 518,
+ 505,
+ 531
+ ],
+ "score": 1.0,
+ "content": "launching compute kernels separately can be expensive, DLVM performs linear algebra fusion, which",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 529,
+ 506,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 529,
+ 506,
+ 542
+ ],
+ "score": 1.0,
+ "content": "transforms subexpressions involving both matrix multiplication and element-wise operations into a",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 540,
+ 505,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 540,
+ 505,
+ 552
+ ],
+ "score": 1.0,
+ "content": "single matrix multiplication instruction on padded tensors. Besides the simple pattern like an addition",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 551,
+ 505,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 551,
+ 505,
+ 563
+ ],
+ "score": 1.0,
+ "content": "of matrix multiplication and a vector, we can apply the same approach to a polynomial of multiple",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 561,
+ 506,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 506,
+ 575
+ ],
+ "score": 1.0,
+ "content": "matrix multiplications, turning the polynomial into a single matrix multiplication. For example, in a",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 573,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 573,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "simple recurrent neural network (RNN), each cell of the recurrence is a feed forward neural network",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 104,
+ 583,
+ 506,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 583,
+ 191,
+ 597
+ ],
+ "score": 1.0,
+ "content": "that takes two inputs:",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 585,
+ 203,
+ 595
+ ],
+ "score": 0.85,
+ "content": "\\mathbf { x } _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 583,
+ 372,
+ 597
+ ],
+ "score": 1.0,
+ "content": ", the input local to the current timestep, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 373,
+ 584,
+ 384,
+ 595
+ ],
+ "score": 0.87,
+ "content": "\\mathbf { h } _ { t }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 384,
+ 583,
+ 506,
+ 597
+ ],
+ "score": 1.0,
+ "content": ", the hidden state carried along",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 594,
+ 505,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 594,
+ 388,
+ 608
+ ],
+ "score": 1.0,
+ "content": "the recurrence. The linear algebra fusion pass can simplify operations in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 388,
+ 595,
+ 505,
+ 606
+ ],
+ "score": 0.92,
+ "content": "\\mathbf { h } _ { t } = f ( \\mathbf { W } \\mathbf { x } _ { t - 1 } + \\mathbf { U } \\mathbf { h } _ { t - 1 } + \\mathbf { b } )",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 605,
+ 506,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 605,
+ 506,
+ 619
+ ],
+ "score": 1.0,
+ "content": "from two matrix multiplications and two additions into a single matrix multiplication. A more aggres-",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 104,
+ 615,
+ 505,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 615,
+ 505,
+ 630
+ ],
+ "score": 1.0,
+ "content": "sive, interprocedural version of linear algebra fusion can optimize parameter passing and memory",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 628,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 628,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "allocation, so that the entire concatenated matrix can be created and passed around in the first place",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 639,
+ 191,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 639,
+ 191,
+ 650
+ ],
+ "score": 1.0,
+ "content": "without reallocation.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 19.5,
+ "bbox_fs": [
+ 104,
+ 474,
+ 506,
+ 650
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 666,
+ 455,
+ 678
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 457,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 457,
+ 679
+ ],
+ "score": 1.0,
+ "content": "3.1.3 ALGORITHMIC DIFFERENTIATION THROUGH ADJOINT CODE GENERATION",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 28
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 687,
+ 504,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 506,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 688,
+ 506,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Algorithmic differentiation (AD), also known as automatic differentiation, encompasses a family of",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 445,
+ 711
+ ],
+ "score": 1.0,
+ "content": "a well-known techniques for algorithmically obtaining the derivatives of a function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 445,
+ 699,
+ 505,
+ 710
+ ],
+ "score": 0.91,
+ "content": "f : \\mathbf { x } \\in \\mathbb { R } ^ { n } ",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 142,
+ 721
+ ],
+ "score": 0.89,
+ "content": "\\mathbf { y } \\in \\mathbb { R } ^ { m }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 142,
+ 709,
+ 278,
+ 722
+ ],
+ "score": 1.0,
+ "content": "(Naumann, 2011). The function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 278,
+ 710,
+ 285,
+ 721
+ ],
+ "score": 0.85,
+ "content": "f",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 285,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "can be represented as a directed acyclic computation",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 506,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 506,
+ 733
+ ],
+ "score": 1.0,
+ "content": "graph representing the composition of elementary computational operations for which the respective",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 80,
+ 505,
+ 98
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 300,
+ 98
+ ],
+ "score": 1.0,
+ "content": "derivatives are well known. The partial derivative",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 301,
+ 80,
+ 316,
+ 96
+ ],
+ "score": 0.92,
+ "content": "\\frac { \\partial y _ { j } } { \\partial x _ { i } }",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 317,
+ 80,
+ 505,
+ 98
+ ],
+ "score": 1.0,
+ "content": "can be computed through recursive applications",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 95,
+ 505,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 95,
+ 505,
+ 107
+ ],
+ "score": 1.0,
+ "content": "of the chain rule, either in the forward direction (corresponding to a bottom-up traversal of the",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 106,
+ 505,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 106,
+ 505,
+ 118
+ ],
+ "score": 1.0,
+ "content": "computation graph) or in the backward direction (corresponding to a top-down traversal of the",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 117,
+ 504,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 117,
+ 504,
+ 129
+ ],
+ "score": 1.0,
+ "content": "computation graph). The former approach, called forward-mode or tangent-mode AD, is used in",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 128,
+ 506,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 318,
+ 140
+ ],
+ "score": 1.0,
+ "content": "some research communities, most commonly when",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 319,
+ 128,
+ 352,
+ 138
+ ],
+ "score": 0.88,
+ "content": "m \\gg n",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 352,
+ 128,
+ 506,
+ 140
+ ],
+ "score": 1.0,
+ "content": "(Goodfellow et al., 2016). The latter",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 138,
+ 507,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 138,
+ 507,
+ 151
+ ],
+ "score": 1.0,
+ "content": "approach, which includes the back-propagation algorithm (Rumelhart et al., 1986) as a special case,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 150,
+ 506,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 150,
+ 506,
+ 162
+ ],
+ "score": 1.0,
+ "content": "is called reverse-mode or adjoint-mode AD, and encompasses the techniques most commonly used",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 160,
+ 281,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 160,
+ 281,
+ 173
+ ],
+ "score": 1.0,
+ "content": "for training the weights in neural networks.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 30.5,
+ "bbox_fs": [
+ 105,
+ 688,
+ 506,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 80,
+ 505,
+ 172
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 505,
+ 98
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 80,
+ 300,
+ 98
+ ],
+ "score": 1.0,
+ "content": "derivatives are well known. The partial derivative",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 301,
+ 80,
+ 316,
+ 96
+ ],
+ "score": 0.92,
+ "content": "\\frac { \\partial y _ { j } } { \\partial x _ { i } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 317,
+ 80,
+ 505,
+ 98
+ ],
+ "score": 1.0,
+ "content": "can be computed through recursive applications",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 95,
+ 505,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 95,
+ 505,
+ 107
+ ],
+ "score": 1.0,
+ "content": "of the chain rule, either in the forward direction (corresponding to a bottom-up traversal of the",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 106,
+ 505,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 106,
+ 505,
+ 118
+ ],
+ "score": 1.0,
+ "content": "computation graph) or in the backward direction (corresponding to a top-down traversal of the",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 117,
+ 504,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 117,
+ 504,
+ 129
+ ],
+ "score": 1.0,
+ "content": "computation graph). The former approach, called forward-mode or tangent-mode AD, is used in",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 128,
+ 506,
+ 140
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 318,
+ 140
+ ],
+ "score": 1.0,
+ "content": "some research communities, most commonly when",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 319,
+ 128,
+ 352,
+ 138
+ ],
+ "score": 0.88,
+ "content": "m \\gg n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 352,
+ 128,
+ 506,
+ 140
+ ],
+ "score": 1.0,
+ "content": "(Goodfellow et al., 2016). The latter",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 138,
+ 507,
+ 151
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 138,
+ 507,
+ 151
+ ],
+ "score": 1.0,
+ "content": "approach, which includes the back-propagation algorithm (Rumelhart et al., 1986) as a special case,",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 150,
+ 506,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 150,
+ 506,
+ 162
+ ],
+ "score": 1.0,
+ "content": "is called reverse-mode or adjoint-mode AD, and encompasses the techniques most commonly used",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 160,
+ 281,
+ 173
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 160,
+ 281,
+ 173
+ ],
+ "score": 1.0,
+ "content": "for training the weights in neural networks.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 3.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 177,
+ 505,
+ 331
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 177,
+ 506,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 177,
+ 506,
+ 190
+ ],
+ "score": 1.0,
+ "content": "In DLVM, the differentiation pass is responsible for performing reverse-mode AD. This pass is",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 188,
+ 506,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 188,
+ 506,
+ 200
+ ],
+ "score": 1.0,
+ "content": "responsible for generating DLVM IR code that calculates the derivative of a differentiable function.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 199,
+ 506,
+ 211
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 506,
+ 211
+ ],
+ "score": 1.0,
+ "content": "A function is marked as being automatically differentiable via gradient declarations. A gradient",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 210,
+ 506,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 210,
+ 506,
+ 222
+ ],
+ "score": 1.0,
+ "content": "declaration is a function in a module that is declared with its mathematical relation with another",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 221,
+ 505,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 221,
+ 345,
+ 233
+ ],
+ "score": 1.0,
+ "content": "function in the module and no function body. The function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 345,
+ 222,
+ 400,
+ 232
+ ],
+ "score": 0.28,
+ "content": "\\mathsf { \\Omega } \\mathsf { \\Omega } \\mathsf { \\Omega } \\mathsf { \\Omega } \\mathsf { \\Omega } \\mathsf { \\Omega } \\mathsf { f o o \\mathrm { - \\mathsf { \\Omega } \\mathsf { { g r a d } } \\Omega } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 400,
+ 221,
+ 505,
+ 233
+ ],
+ "score": 1.0,
+ "content": "in Figure 3 is an example",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 232,
+ 506,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 232,
+ 506,
+ 245
+ ],
+ "score": 1.0,
+ "content": "of such a function. Gradient declarations are configurable, e.g. specifying arguments to differentiate",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 243,
+ 506,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 243,
+ 506,
+ 256
+ ],
+ "score": 1.0,
+ "content": "with respect to, keeping original outputs, and toggling seedability to accept back-propagated gradients.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 253,
+ 505,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 505,
+ 267
+ ],
+ "score": 1.0,
+ "content": "The differentiation pass, when applied, canonicalizes every gradient declaration in the module to",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 266,
+ 506,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 266,
+ 506,
+ 277
+ ],
+ "score": 1.0,
+ "content": "a normal function definition with basic blocks and instructions. The canonicalization process first",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 276,
+ 505,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 276,
+ 505,
+ 288
+ ],
+ "score": 1.0,
+ "content": "copies basic blocks and instructions from the original function to the new function body, and then",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 507,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 507,
+ 300
+ ],
+ "score": 1.0,
+ "content": "applies adjoint code generation to the function. Unlike many of the existing deep learning frameworks,",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 297,
+ 505,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 297,
+ 505,
+ 310
+ ],
+ "score": 1.0,
+ "content": "AD in DLVM is a source code transformation, not interpretation (operator overloading) over the",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 309,
+ 505,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 309,
+ 505,
+ 322
+ ],
+ "score": 1.0,
+ "content": "same program. This makes the compiler able to perform optimizations on the gradient computation",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 320,
+ 312,
+ 332
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 320,
+ 312,
+ 332
+ ],
+ "score": 1.0,
+ "content": "separately and enables higher order differentiation.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 14.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 336,
+ 505,
+ 447
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 335,
+ 506,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 335,
+ 236,
+ 350
+ ],
+ "score": 1.0,
+ "content": "Given a differentiable function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 336,
+ 307,
+ 349
+ ],
+ "score": 0.92,
+ "content": "f ( \\mathbf { x } _ { 1 } , \\mathbf { x } _ { 2 } , \\ldots , \\mathbf { x } _ { n } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 307,
+ 335,
+ 506,
+ 350
+ ],
+ "score": 1.0,
+ "content": ", this pass creates a new function that computes",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 347,
+ 505,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 347,
+ 160,
+ 360
+ ],
+ "score": 1.0,
+ "content": "the Jacobian",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 160,
+ 348,
+ 172,
+ 360
+ ],
+ "score": 0.87,
+ "content": "\\mathbf { J } _ { f }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 172,
+ 347,
+ 505,
+ 360
+ ],
+ "score": 1.0,
+ "content": ". This approach to AD has several advantages with respect to AD performed by",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 359,
+ 505,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 359,
+ 195,
+ 371
+ ],
+ "score": 1.0,
+ "content": "operator overloading",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 195,
+ 359,
+ 201,
+ 369
+ ],
+ "score": 0.32,
+ "content": "/",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 359,
+ 505,
+ 371
+ ],
+ "score": 1.0,
+ "content": "graph interpretation. Unlike operator overloading, the gradient function",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 369,
+ 506,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 369,
+ 506,
+ 382
+ ],
+ "score": 1.0,
+ "content": "produced by AD is a standalone function that sits uniformly alongside other functions in an IR",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 381,
+ 505,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 381,
+ 505,
+ 393
+ ],
+ "score": 1.0,
+ "content": "module, representationally unrelated to the original function. The generated function takes original",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 392,
+ 506,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 392,
+ 506,
+ 405
+ ],
+ "score": 1.0,
+ "content": "inputs and produces a tuple of partial derivatives with respect to the inputs. In AD, not all values in",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 415
+ ],
+ "score": 1.0,
+ "content": "the forward evaluation will necessarily be used to compute derivatives. In DLVM, unused operations",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 412,
+ 506,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 412,
+ 506,
+ 427
+ ],
+ "score": 1.0,
+ "content": "can be easily eliminated by the aggressive dead code elimination pass in the compilation pipeline (see",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 424,
+ 506,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 506,
+ 437
+ ],
+ "score": 1.0,
+ "content": "Section 3.1.4). In addition, an AD-specific optimization technique called checkpointing can further",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 436,
+ 342,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 436,
+ 342,
+ 448
+ ],
+ "score": 1.0,
+ "content": "reduce memory consumption during gradient computation.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 452,
+ 505,
+ 551
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 452,
+ 505,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 452,
+ 505,
+ 464
+ ],
+ "score": 1.0,
+ "content": "AD in DLVM is configurable. The front-end can choose to differentiate a function with respect to",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 464,
+ 505,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 464,
+ 505,
+ 475
+ ],
+ "score": 1.0,
+ "content": "selected arguments, to keep some of the outputs of the original function, to apply differentiation",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 475,
+ 506,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 506,
+ 486
+ ],
+ "score": 1.0,
+ "content": "to a specific output when there are multiple return values, or to enable the function to accept back-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 485,
+ 506,
+ 498
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 485,
+ 506,
+ 498
+ ],
+ "score": 1.0,
+ "content": "propagated gradients (seeds) through function composition, all by gradient declarations. If the",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 496,
+ 505,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 496,
+ 505,
+ 509
+ ],
+ "score": 1.0,
+ "content": "function returns multiple values in a tuple, the gradient declaration can also specify which tuple",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 507,
+ 505,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 507,
+ 505,
+ 520
+ ],
+ "score": 1.0,
+ "content": "element to differentiate. Our approach to AD is implemented as a transformation from one function",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 518,
+ 504,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 518,
+ 504,
+ 530
+ ],
+ "score": 1.0,
+ "content": "to another function. This approach also makes higher-order differentiation possible; this can be",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 529,
+ 505,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 529,
+ 505,
+ 542
+ ],
+ "score": 1.0,
+ "content": "accomplished by declaring a higher-order gradient function that differentiates the original gradient",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 541,
+ 144,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 144,
+ 552
+ ],
+ "score": 1.0,
+ "content": "function.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 599,
+ 348,
+ 610
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 349,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 349,
+ 612
+ ],
+ "score": 1.0,
+ "content": "3.1.4 GENERAL-PURPOSE OPTIMIZATIONS FOR DLVM",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 41
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 632,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 506,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 506,
+ 646
+ ],
+ "score": 1.0,
+ "content": "General-purpose optimizations refer to traditional compiler optimizations applied to DLVM IR.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 644,
+ 506,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 506,
+ 657
+ ],
+ "score": 1.0,
+ "content": "These optimizations are important at the DLVM stage in the compilation pipeline, since linear",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "algebra computation can be highly optimized or eliminated before they get lowered to LLVM IR",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "which contain parallel execution and low-level information that prevent LLVM optimizations from",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "identifying high-level patterns. Some of these optimizations are aggressive dead code elimination,",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 701
+ ],
+ "score": 1.0,
+ "content": "common subexpression elimination, and sparse conditional constant propagation. Applying such",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "optimizations on gradient computation is not feasible in other approaches to AD that use graph",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "interpretation (operator overloading), because the forward pass and the backward pass are tied",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 485,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 485,
+ 734
+ ],
+ "score": 1.0,
+ "content": "together in a single graph; mutating either evaluation pass will alter the semantics of the other.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "page_idx": 5,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 224,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 225,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 225,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Workshop track - ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "6",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 80,
+ 505,
+ 172
+ ],
+ "lines": [],
+ "index": 3.5,
+ "bbox_fs": [
+ 105,
+ 80,
+ 507,
+ 173
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 177,
+ 505,
+ 331
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 177,
+ 506,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 177,
+ 506,
+ 190
+ ],
+ "score": 1.0,
+ "content": "In DLVM, the differentiation pass is responsible for performing reverse-mode AD. This pass is",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 188,
+ 506,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 188,
+ 506,
+ 200
+ ],
+ "score": 1.0,
+ "content": "responsible for generating DLVM IR code that calculates the derivative of a differentiable function.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 199,
+ 506,
+ 211
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 506,
+ 211
+ ],
+ "score": 1.0,
+ "content": "A function is marked as being automatically differentiable via gradient declarations. A gradient",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 210,
+ 506,
+ 222
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 210,
+ 506,
+ 222
+ ],
+ "score": 1.0,
+ "content": "declaration is a function in a module that is declared with its mathematical relation with another",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 221,
+ 505,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 221,
+ 345,
+ 233
+ ],
+ "score": 1.0,
+ "content": "function in the module and no function body. The function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 345,
+ 222,
+ 400,
+ 232
+ ],
+ "score": 0.28,
+ "content": "\\mathsf { \\Omega } \\mathsf { \\Omega } \\mathsf { \\Omega } \\mathsf { \\Omega } \\mathsf { \\Omega } \\mathsf { \\Omega } \\mathsf { f o o \\mathrm { - \\mathsf { \\Omega } \\mathsf { { g r a d } } \\Omega } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 400,
+ 221,
+ 505,
+ 233
+ ],
+ "score": 1.0,
+ "content": "in Figure 3 is an example",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 232,
+ 506,
+ 245
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 232,
+ 506,
+ 245
+ ],
+ "score": 1.0,
+ "content": "of such a function. Gradient declarations are configurable, e.g. specifying arguments to differentiate",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 243,
+ 506,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 243,
+ 506,
+ 256
+ ],
+ "score": 1.0,
+ "content": "with respect to, keeping original outputs, and toggling seedability to accept back-propagated gradients.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 253,
+ 505,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 505,
+ 267
+ ],
+ "score": 1.0,
+ "content": "The differentiation pass, when applied, canonicalizes every gradient declaration in the module to",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 266,
+ 506,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 266,
+ 506,
+ 277
+ ],
+ "score": 1.0,
+ "content": "a normal function definition with basic blocks and instructions. The canonicalization process first",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 276,
+ 505,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 276,
+ 505,
+ 288
+ ],
+ "score": 1.0,
+ "content": "copies basic blocks and instructions from the original function to the new function body, and then",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 507,
+ 300
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 507,
+ 300
+ ],
+ "score": 1.0,
+ "content": "applies adjoint code generation to the function. Unlike many of the existing deep learning frameworks,",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 297,
+ 505,
+ 310
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 297,
+ 505,
+ 310
+ ],
+ "score": 1.0,
+ "content": "AD in DLVM is a source code transformation, not interpretation (operator overloading) over the",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 309,
+ 505,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 309,
+ 505,
+ 322
+ ],
+ "score": 1.0,
+ "content": "same program. This makes the compiler able to perform optimizations on the gradient computation",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 320,
+ 312,
+ 332
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 320,
+ 312,
+ 332
+ ],
+ "score": 1.0,
+ "content": "separately and enables higher order differentiation.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 14.5,
+ "bbox_fs": [
+ 105,
+ 177,
+ 507,
+ 332
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 336,
+ 505,
+ 447
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 335,
+ 506,
+ 350
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 335,
+ 236,
+ 350
+ ],
+ "score": 1.0,
+ "content": "Given a differentiable function",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 236,
+ 336,
+ 307,
+ 349
+ ],
+ "score": 0.92,
+ "content": "f ( \\mathbf { x } _ { 1 } , \\mathbf { x } _ { 2 } , \\ldots , \\mathbf { x } _ { n } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 307,
+ 335,
+ 506,
+ 350
+ ],
+ "score": 1.0,
+ "content": ", this pass creates a new function that computes",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 347,
+ 505,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 347,
+ 160,
+ 360
+ ],
+ "score": 1.0,
+ "content": "the Jacobian",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 160,
+ 348,
+ 172,
+ 360
+ ],
+ "score": 0.87,
+ "content": "\\mathbf { J } _ { f }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 172,
+ 347,
+ 505,
+ 360
+ ],
+ "score": 1.0,
+ "content": ". This approach to AD has several advantages with respect to AD performed by",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 359,
+ 505,
+ 371
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 359,
+ 195,
+ 371
+ ],
+ "score": 1.0,
+ "content": "operator overloading",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 195,
+ 359,
+ 201,
+ 369
+ ],
+ "score": 0.32,
+ "content": "/",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 359,
+ 505,
+ 371
+ ],
+ "score": 1.0,
+ "content": "graph interpretation. Unlike operator overloading, the gradient function",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 369,
+ 506,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 369,
+ 506,
+ 382
+ ],
+ "score": 1.0,
+ "content": "produced by AD is a standalone function that sits uniformly alongside other functions in an IR",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 381,
+ 505,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 381,
+ 505,
+ 393
+ ],
+ "score": 1.0,
+ "content": "module, representationally unrelated to the original function. The generated function takes original",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 392,
+ 506,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 392,
+ 506,
+ 405
+ ],
+ "score": 1.0,
+ "content": "inputs and produces a tuple of partial derivatives with respect to the inputs. In AD, not all values in",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 402,
+ 505,
+ 415
+ ],
+ "score": 1.0,
+ "content": "the forward evaluation will necessarily be used to compute derivatives. In DLVM, unused operations",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 412,
+ 506,
+ 427
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 412,
+ 506,
+ 427
+ ],
+ "score": 1.0,
+ "content": "can be easily eliminated by the aggressive dead code elimination pass in the compilation pipeline (see",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 424,
+ 506,
+ 437
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 506,
+ 437
+ ],
+ "score": 1.0,
+ "content": "Section 3.1.4). In addition, an AD-specific optimization technique called checkpointing can further",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 436,
+ 342,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 436,
+ 342,
+ 448
+ ],
+ "score": 1.0,
+ "content": "reduce memory consumption during gradient computation.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 105,
+ 335,
+ 506,
+ 448
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 452,
+ 505,
+ 551
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 452,
+ 505,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 452,
+ 505,
+ 464
+ ],
+ "score": 1.0,
+ "content": "AD in DLVM is configurable. The front-end can choose to differentiate a function with respect to",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 464,
+ 505,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 464,
+ 505,
+ 475
+ ],
+ "score": 1.0,
+ "content": "selected arguments, to keep some of the outputs of the original function, to apply differentiation",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 475,
+ 506,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 475,
+ 506,
+ 486
+ ],
+ "score": 1.0,
+ "content": "to a specific output when there are multiple return values, or to enable the function to accept back-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 485,
+ 506,
+ 498
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 485,
+ 506,
+ 498
+ ],
+ "score": 1.0,
+ "content": "propagated gradients (seeds) through function composition, all by gradient declarations. If the",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 496,
+ 505,
+ 509
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 496,
+ 505,
+ 509
+ ],
+ "score": 1.0,
+ "content": "function returns multiple values in a tuple, the gradient declaration can also specify which tuple",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 507,
+ 505,
+ 520
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 507,
+ 505,
+ 520
+ ],
+ "score": 1.0,
+ "content": "element to differentiate. Our approach to AD is implemented as a transformation from one function",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 518,
+ 504,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 518,
+ 504,
+ 530
+ ],
+ "score": 1.0,
+ "content": "to another function. This approach also makes higher-order differentiation possible; this can be",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 529,
+ 505,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 529,
+ 505,
+ 542
+ ],
+ "score": 1.0,
+ "content": "accomplished by declaring a higher-order gradient function that differentiates the original gradient",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 541,
+ 144,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 144,
+ 552
+ ],
+ "score": 1.0,
+ "content": "function.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 36,
+ "bbox_fs": [
+ 105,
+ 452,
+ 506,
+ 552
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 599,
+ 348,
+ 610
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 349,
+ 612
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 599,
+ 349,
+ 612
+ ],
+ "score": 1.0,
+ "content": "3.1.4 GENERAL-PURPOSE OPTIMIZATIONS FOR DLVM",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 41
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 632,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 506,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 506,
+ 646
+ ],
+ "score": 1.0,
+ "content": "General-purpose optimizations refer to traditional compiler optimizations applied to DLVM IR.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 644,
+ 506,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 506,
+ 657
+ ],
+ "score": 1.0,
+ "content": "These optimizations are important at the DLVM stage in the compilation pipeline, since linear",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "algebra computation can be highly optimized or eliminated before they get lowered to LLVM IR",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "which contain parallel execution and low-level information that prevent LLVM optimizations from",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "identifying high-level patterns. Some of these optimizations are aggressive dead code elimination,",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 701
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 688,
+ 505,
+ 701
+ ],
+ "score": 1.0,
+ "content": "common subexpression elimination, and sparse conditional constant propagation. Applying such",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "optimizations on gradient computation is not feasible in other approaches to AD that use graph",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "interpretation (operator overloading), because the forward pass and the backward pass are tied",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 720,
+ 485,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 720,
+ 485,
+ 734
+ ],
+ "score": 1.0,
+ "content": "together in a single graph; mutating either evaluation pass will alter the semantics of the other.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 46,
+ "bbox_fs": [
+ 105,
+ 633,
+ 506,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 82,
+ 214,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 216,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 216,
+ 95
+ ],
+ "score": 1.0,
+ "content": "3.2 CODE GENERATION",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 102,
+ 505,
+ 191
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 102,
+ 505,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 102,
+ 505,
+ 116
+ ],
+ "score": 1.0,
+ "content": "Two major design goals of DLVM are the ability to target multiple heterogenous parallel architectures",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 114,
+ 504,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 114,
+ 504,
+ 127
+ ],
+ "score": 1.0,
+ "content": "from the same front-end DSL code (and corresponding DLVM IR), and the ability to perform",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 137
+ ],
+ "score": 1.0,
+ "content": "aggressive optimizations on lowered programs. In order to attain these goals, DLVM code generation",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 136,
+ 504,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 136,
+ 504,
+ 148
+ ],
+ "score": 1.0,
+ "content": "transforms DLVM IR into LLVM IR. LLVM is a robust and mature compiler infrastructure with",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 147,
+ 505,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 147,
+ 505,
+ 159
+ ],
+ "score": 1.0,
+ "content": "multiple back-ends, including NVIDIA GPUs. Many high-level DLVM IR linear algebra instructions",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 158,
+ 506,
+ 170
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 158,
+ 506,
+ 170
+ ],
+ "score": 1.0,
+ "content": "over tensors abstract lower-level operations. The DLVM compiler transforms the high-level DLVM",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 169,
+ 507,
+ 181
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 169,
+ 507,
+ 181
+ ],
+ "score": 1.0,
+ "content": "IR into lower-level stages and ultimately into calls to BLAS and compute kernels in LLVM IR.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 179,
+ 457,
+ 192
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 179,
+ 457,
+ 192
+ ],
+ "score": 1.0,
+ "content": "Existing LLVM utilities are used to compile the generated LLVM IR to the final binary.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 4.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 196,
+ 505,
+ 241
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 196,
+ 505,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 196,
+ 505,
+ 209
+ ],
+ "score": 1.0,
+ "content": "In order to take full advantage of a variety of emerging heterogeneous parallel architectures, we",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 208,
+ 506,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 208,
+ 506,
+ 219
+ ],
+ "score": 1.0,
+ "content": "plan for future versions of DLVM to target the IR of HPVM (Srivastava et al., 2016), a higher-",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 218,
+ 505,
+ 231
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 218,
+ 505,
+ 231
+ ],
+ "score": 1.0,
+ "content": "level heterogeneous compiler extension to LLVM IR that allows for transparent targeting of diverse",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 230,
+ 255,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 230,
+ 255,
+ 243
+ ],
+ "score": 1.0,
+ "content": "architectures from a data flow graph.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 10.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 109,
+ 254,
+ 286,
+ 264
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 253,
+ 288,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 253,
+ 288,
+ 266
+ ],
+ "score": 1.0,
+ "content": "3.3 DLVM COMMAND LINE TOOLCHAIN",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 274,
+ 505,
+ 340
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 272,
+ 506,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 272,
+ 506,
+ 288
+ ],
+ "score": 1.0,
+ "content": "The front-end software associated with each DSL (see Section 3.4) is responsible for generating a",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 285,
+ 506,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 285,
+ 506,
+ 299
+ ],
+ "score": 1.0,
+ "content": "DLVM IR for a given source language program to be compiled. The DLVM compiler infrastructure",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 296,
+ 505,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 296,
+ 505,
+ 309
+ ],
+ "score": 1.0,
+ "content": "itself is a compiler from DLVM IR to LLVM IR, therefore having a command line toolchain is",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 307,
+ 506,
+ 319
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 307,
+ 417,
+ 319
+ ],
+ "score": 1.0,
+ "content": "necessary for verifying, transforming and compiling batches of DLVM IR files",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 418,
+ 308,
+ 448,
+ 318
+ ],
+ "score": 0.34,
+ "content": "( \\star _ { } . \\mathrm { d l } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 449,
+ 307,
+ 506,
+ 319
+ ],
+ "score": 1.0,
+ "content": ". Unlike XLA",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 317,
+ 506,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 317,
+ 307,
+ 331
+ ],
+ "score": 1.0,
+ "content": "and NNVM/TVM, which only provide a Python",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 308,
+ 319,
+ 329,
+ 329
+ ],
+ "score": 0.8,
+ "content": "/ \\mathrm { C } { + + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 329,
+ 317,
+ 506,
+ 331
+ ],
+ "score": 1.0,
+ "content": "interface to their users, DLVM provides a",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 330,
+ 348,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 330,
+ 348,
+ 341
+ ],
+ "score": 1.0,
+ "content": "command line interface like any industry-standard compiler.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 16.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 346,
+ 505,
+ 423
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 505,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 290,
+ 359
+ ],
+ "score": 1.0,
+ "content": "The DLVM optimizer utility, dlopt, accepts",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 290,
+ 348,
+ 297,
+ 356
+ ],
+ "score": 0.76,
+ "content": "\\star",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 298,
+ 345,
+ 505,
+ 359
+ ],
+ "score": 1.0,
+ "content": ".dl IR files and applies user-specified optimization",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 357,
+ 505,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 357,
+ 337,
+ 369
+ ],
+ "score": 1.0,
+ "content": "passes on them. The DLVM compiler driver, dlc, accepts",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 338,
+ 359,
+ 344,
+ 367
+ ],
+ "score": 0.71,
+ "content": "\\star",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 345,
+ 357,
+ 505,
+ 369
+ ],
+ "score": 1.0,
+ "content": ".dl IR files and performs user-specified",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 368,
+ 506,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 368,
+ 506,
+ 381
+ ],
+ "score": 1.0,
+ "content": "tasks, such as verification, differentiation, optimization passes, stage lowering passes, and code",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 378,
+ 507,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 378,
+ 507,
+ 393
+ ],
+ "score": 1.0,
+ "content": "generation; the driver invokes the DLVM core library to achieve these tasks. Because of having a",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 390,
+ 506,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 390,
+ 506,
+ 402
+ ],
+ "score": 1.0,
+ "content": "textual format of the IR, the DLVM framework can easily make use of the LLVM Integrated Tester",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 401,
+ 506,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 401,
+ 506,
+ 413
+ ],
+ "score": 1.0,
+ "content": "(lit) and FileCheck to perform robust unit testing. In future development, we plan to introduce a",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 412,
+ 474,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 412,
+ 474,
+ 425
+ ],
+ "score": 1.0,
+ "content": "DLVM bitcode format for compact storage and high-throughput processing of DLVM code.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 435,
+ 241,
+ 447
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 435,
+ 243,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 435,
+ 243,
+ 448
+ ],
+ "score": 1.0,
+ "content": "3.4 NEURAL NETWORK DSLS",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 456,
+ 505,
+ 512
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 505,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 505,
+ 470
+ ],
+ "score": 1.0,
+ "content": "Most of existing deep learning DSLs are embedded in a dynamically typed scripting language",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 467,
+ 506,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 467,
+ 506,
+ 480
+ ],
+ "score": 1.0,
+ "content": "such as Python and Lua. While these scripting languages provide flexibility and a large number of",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 477,
+ 505,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 477,
+ 505,
+ 492
+ ],
+ "score": 1.0,
+ "content": "libraries for scientific computing, they can act as a barrier between lightweight prototyping code",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 104,
+ 487,
+ 507,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 487,
+ 507,
+ 504
+ ],
+ "score": 1.0,
+ "content": "and systematic production code. This barrier significantly reduces the reliability of ML software,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 495,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 495,
+ 514
+ ],
+ "score": 1.0,
+ "content": "resulting in suboptimal programming experience and unnecessarily effortful development cycles.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 517,
+ 505,
+ 627
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 517,
+ 506,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 517,
+ 506,
+ 530
+ ],
+ "score": 1.0,
+ "content": "In software engineering, a proven approach to tackle this problem is language and compiler technolo-",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 528,
+ 506,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 528,
+ 506,
+ 541
+ ],
+ "score": 1.0,
+ "content": "gies, starting from a language that is amenable to static analysis. A well-designed deep learning DSL",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 539,
+ 505,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 539,
+ 505,
+ 552
+ ],
+ "score": 1.0,
+ "content": "should support the needs of deep learning software developers by providing a safe environment for",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 550,
+ 506,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 550,
+ 506,
+ 563
+ ],
+ "score": 1.0,
+ "content": "rapid prototyping, while simultaneously generating highly efficient code for training and inference.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 561,
+ 506,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 506,
+ 575
+ ],
+ "score": 1.0,
+ "content": "DSLs in a scripting language can easily achieve rapid prototyping, but they are generally incapable",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 572,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 572,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "of providing a safe environment with optimal performance. We believe that the best solution is",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 596
+ ],
+ "score": 1.0,
+ "content": "DSLs embedded in a type-safe, type-inferring programming language that is both fast and easy to",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 593,
+ 506,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 593,
+ 506,
+ 607
+ ],
+ "score": 1.0,
+ "content": "learn. In our initial release of DLVM, we provide one such DSL, both as a proof-of-concept and as a",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 104,
+ 603,
+ 506,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 603,
+ 506,
+ 619
+ ],
+ "score": 1.0,
+ "content": "reference implementation that showcases the capabilities of DLVM as a platform for deep learning",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 184,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 184,
+ 629
+ ],
+ "score": 1.0,
+ "content": "DSL development.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 37.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 632,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 505,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 505,
+ 646
+ ],
+ "score": 1.0,
+ "content": "NNKit is a staged DSL embedded in Swift, featuring natural expression of tensor computation",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 643,
+ 506,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 643,
+ 506,
+ 657
+ ],
+ "score": 1.0,
+ "content": "alongside the host program without losing static analyses and optimizations on the tensor program.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 505,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 505,
+ 669
+ ],
+ "score": 1.0,
+ "content": "Inspired by Lightweight Modular Staging (Rompf & Odersky, 2010), NNKit leverages the static type",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 428,
+ 680
+ ],
+ "score": 1.0,
+ "content": "system of Swift to guarantee type safety of the DSL. Tensor types are wrapped in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 428,
+ 667,
+ 465,
+ 677
+ ],
+ "score": 0.83,
+ "content": "\\mathrm { R e p } { < } \\mathrm { T } >",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 465,
+ 665,
+ 506,
+ 680
+ ],
+ "score": 1.0,
+ "content": ", meaning",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 691
+ ],
+ "score": 1.0,
+ "content": "the representation of some computation that produces data of type T. Tensor operators overloaded for",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Rep are essentially AST builders for delayed evaluation. Instead of generating computation nodes at",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "runtime and performing operator-overloading AD like PyTorch or TensorFlow Eager (Google Brain",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Team, 2017), NNKit tensor computations are staged once during the lifetime of the host program. At",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 734
+ ],
+ "score": 1.0,
+ "content": "invocation time of staged functions, NNKit emits shape-specialized DLVM IR and leverages DLVM",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "page_idx": 6,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 224,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 225,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 225,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Workshop track - ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 303,
+ 751,
+ 308,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 763
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 750,
+ 309,
+ 763
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 13,
+ "width": 7
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 82,
+ 214,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 216,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 82,
+ 216,
+ 95
+ ],
+ "score": 1.0,
+ "content": "3.2 CODE GENERATION",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 102,
+ 505,
+ 191
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 102,
+ 505,
+ 116
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 102,
+ 505,
+ 116
+ ],
+ "score": 1.0,
+ "content": "Two major design goals of DLVM are the ability to target multiple heterogenous parallel architectures",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 105,
+ 114,
+ 504,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 114,
+ 504,
+ 127
+ ],
+ "score": 1.0,
+ "content": "from the same front-end DSL code (and corresponding DLVM IR), and the ability to perform",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 126,
+ 505,
+ 137
+ ],
+ "score": 1.0,
+ "content": "aggressive optimizations on lowered programs. In order to attain these goals, DLVM code generation",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 136,
+ 504,
+ 148
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 136,
+ 504,
+ 148
+ ],
+ "score": 1.0,
+ "content": "transforms DLVM IR into LLVM IR. LLVM is a robust and mature compiler infrastructure with",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 147,
+ 505,
+ 159
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 147,
+ 505,
+ 159
+ ],
+ "score": 1.0,
+ "content": "multiple back-ends, including NVIDIA GPUs. Many high-level DLVM IR linear algebra instructions",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 158,
+ 506,
+ 170
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 158,
+ 506,
+ 170
+ ],
+ "score": 1.0,
+ "content": "over tensors abstract lower-level operations. The DLVM compiler transforms the high-level DLVM",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 169,
+ 507,
+ 181
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 169,
+ 507,
+ 181
+ ],
+ "score": 1.0,
+ "content": "IR into lower-level stages and ultimately into calls to BLAS and compute kernels in LLVM IR.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 179,
+ 457,
+ 192
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 179,
+ 457,
+ 192
+ ],
+ "score": 1.0,
+ "content": "Existing LLVM utilities are used to compile the generated LLVM IR to the final binary.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 4.5,
+ "bbox_fs": [
+ 105,
+ 102,
+ 507,
+ 192
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 196,
+ 505,
+ 241
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 196,
+ 505,
+ 209
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 196,
+ 505,
+ 209
+ ],
+ "score": 1.0,
+ "content": "In order to take full advantage of a variety of emerging heterogeneous parallel architectures, we",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 208,
+ 506,
+ 219
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 208,
+ 506,
+ 219
+ ],
+ "score": 1.0,
+ "content": "plan for future versions of DLVM to target the IR of HPVM (Srivastava et al., 2016), a higher-",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 218,
+ 505,
+ 231
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 218,
+ 505,
+ 231
+ ],
+ "score": 1.0,
+ "content": "level heterogeneous compiler extension to LLVM IR that allows for transparent targeting of diverse",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 230,
+ 255,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 230,
+ 255,
+ 243
+ ],
+ "score": 1.0,
+ "content": "architectures from a data flow graph.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 10.5,
+ "bbox_fs": [
+ 105,
+ 196,
+ 506,
+ 243
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 109,
+ 254,
+ 286,
+ 264
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 253,
+ 288,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 253,
+ 288,
+ 266
+ ],
+ "score": 1.0,
+ "content": "3.3 DLVM COMMAND LINE TOOLCHAIN",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 274,
+ 505,
+ 340
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 272,
+ 506,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 272,
+ 506,
+ 288
+ ],
+ "score": 1.0,
+ "content": "The front-end software associated with each DSL (see Section 3.4) is responsible for generating a",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 285,
+ 506,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 285,
+ 506,
+ 299
+ ],
+ "score": 1.0,
+ "content": "DLVM IR for a given source language program to be compiled. The DLVM compiler infrastructure",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 296,
+ 505,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 296,
+ 505,
+ 309
+ ],
+ "score": 1.0,
+ "content": "itself is a compiler from DLVM IR to LLVM IR, therefore having a command line toolchain is",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 307,
+ 506,
+ 319
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 307,
+ 417,
+ 319
+ ],
+ "score": 1.0,
+ "content": "necessary for verifying, transforming and compiling batches of DLVM IR files",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 418,
+ 308,
+ 448,
+ 318
+ ],
+ "score": 0.34,
+ "content": "( \\star _ { } . \\mathrm { d l } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 449,
+ 307,
+ 506,
+ 319
+ ],
+ "score": 1.0,
+ "content": ". Unlike XLA",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 317,
+ 506,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 317,
+ 307,
+ 331
+ ],
+ "score": 1.0,
+ "content": "and NNVM/TVM, which only provide a Python",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 308,
+ 319,
+ 329,
+ 329
+ ],
+ "score": 0.8,
+ "content": "/ \\mathrm { C } { + + }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 329,
+ 317,
+ 506,
+ 331
+ ],
+ "score": 1.0,
+ "content": "interface to their users, DLVM provides a",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 330,
+ 348,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 330,
+ 348,
+ 341
+ ],
+ "score": 1.0,
+ "content": "command line interface like any industry-standard compiler.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 16.5,
+ "bbox_fs": [
+ 105,
+ 272,
+ 506,
+ 341
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 346,
+ 505,
+ 423
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 505,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 290,
+ 359
+ ],
+ "score": 1.0,
+ "content": "The DLVM optimizer utility, dlopt, accepts",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 290,
+ 348,
+ 297,
+ 356
+ ],
+ "score": 0.76,
+ "content": "\\star",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 298,
+ 345,
+ 505,
+ 359
+ ],
+ "score": 1.0,
+ "content": ".dl IR files and applies user-specified optimization",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 357,
+ 505,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 357,
+ 337,
+ 369
+ ],
+ "score": 1.0,
+ "content": "passes on them. The DLVM compiler driver, dlc, accepts",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 338,
+ 359,
+ 344,
+ 367
+ ],
+ "score": 0.71,
+ "content": "\\star",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 345,
+ 357,
+ 505,
+ 369
+ ],
+ "score": 1.0,
+ "content": ".dl IR files and performs user-specified",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 368,
+ 506,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 368,
+ 506,
+ 381
+ ],
+ "score": 1.0,
+ "content": "tasks, such as verification, differentiation, optimization passes, stage lowering passes, and code",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 378,
+ 507,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 378,
+ 507,
+ 393
+ ],
+ "score": 1.0,
+ "content": "generation; the driver invokes the DLVM core library to achieve these tasks. Because of having a",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 390,
+ 506,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 390,
+ 506,
+ 402
+ ],
+ "score": 1.0,
+ "content": "textual format of the IR, the DLVM framework can easily make use of the LLVM Integrated Tester",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 401,
+ 506,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 401,
+ 506,
+ 413
+ ],
+ "score": 1.0,
+ "content": "(lit) and FileCheck to perform robust unit testing. In future development, we plan to introduce a",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 412,
+ 474,
+ 425
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 412,
+ 474,
+ 425
+ ],
+ "score": 1.0,
+ "content": "DLVM bitcode format for compact storage and high-throughput processing of DLVM code.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 23,
+ "bbox_fs": [
+ 105,
+ 345,
+ 507,
+ 425
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 435,
+ 241,
+ 447
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 435,
+ 243,
+ 448
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 435,
+ 243,
+ 448
+ ],
+ "score": 1.0,
+ "content": "3.4 NEURAL NETWORK DSLS",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 456,
+ 505,
+ 512
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 505,
+ 470
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 456,
+ 505,
+ 470
+ ],
+ "score": 1.0,
+ "content": "Most of existing deep learning DSLs are embedded in a dynamically typed scripting language",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 467,
+ 506,
+ 480
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 467,
+ 506,
+ 480
+ ],
+ "score": 1.0,
+ "content": "such as Python and Lua. While these scripting languages provide flexibility and a large number of",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 477,
+ 505,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 477,
+ 505,
+ 492
+ ],
+ "score": 1.0,
+ "content": "libraries for scientific computing, they can act as a barrier between lightweight prototyping code",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 104,
+ 487,
+ 507,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 487,
+ 507,
+ 504
+ ],
+ "score": 1.0,
+ "content": "and systematic production code. This barrier significantly reduces the reliability of ML software,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 495,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 495,
+ 514
+ ],
+ "score": 1.0,
+ "content": "resulting in suboptimal programming experience and unnecessarily effortful development cycles.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 104,
+ 456,
+ 507,
+ 514
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 517,
+ 505,
+ 627
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 517,
+ 506,
+ 530
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 517,
+ 506,
+ 530
+ ],
+ "score": 1.0,
+ "content": "In software engineering, a proven approach to tackle this problem is language and compiler technolo-",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 528,
+ 506,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 528,
+ 506,
+ 541
+ ],
+ "score": 1.0,
+ "content": "gies, starting from a language that is amenable to static analysis. A well-designed deep learning DSL",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 539,
+ 505,
+ 552
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 539,
+ 505,
+ 552
+ ],
+ "score": 1.0,
+ "content": "should support the needs of deep learning software developers by providing a safe environment for",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 550,
+ 506,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 550,
+ 506,
+ 563
+ ],
+ "score": 1.0,
+ "content": "rapid prototyping, while simultaneously generating highly efficient code for training and inference.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 561,
+ 506,
+ 575
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 561,
+ 506,
+ 575
+ ],
+ "score": 1.0,
+ "content": "DSLs in a scripting language can easily achieve rapid prototyping, but they are generally incapable",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 572,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 572,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "of providing a safe environment with optimal performance. We believe that the best solution is",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 582,
+ 506,
+ 596
+ ],
+ "score": 1.0,
+ "content": "DSLs embedded in a type-safe, type-inferring programming language that is both fast and easy to",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 593,
+ 506,
+ 607
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 593,
+ 506,
+ 607
+ ],
+ "score": 1.0,
+ "content": "learn. In our initial release of DLVM, we provide one such DSL, both as a proof-of-concept and as a",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 104,
+ 603,
+ 506,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 603,
+ 506,
+ 619
+ ],
+ "score": 1.0,
+ "content": "reference implementation that showcases the capabilities of DLVM as a platform for deep learning",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 616,
+ 184,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 616,
+ 184,
+ 629
+ ],
+ "score": 1.0,
+ "content": "DSL development.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 37.5,
+ "bbox_fs": [
+ 104,
+ 517,
+ 506,
+ 629
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 632,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 505,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 505,
+ 646
+ ],
+ "score": 1.0,
+ "content": "NNKit is a staged DSL embedded in Swift, featuring natural expression of tensor computation",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 106,
+ 643,
+ 506,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 643,
+ 506,
+ 657
+ ],
+ "score": 1.0,
+ "content": "alongside the host program without losing static analyses and optimizations on the tensor program.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 505,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 654,
+ 505,
+ 669
+ ],
+ "score": 1.0,
+ "content": "Inspired by Lightweight Modular Staging (Rompf & Odersky, 2010), NNKit leverages the static type",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 665,
+ 506,
+ 680
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 665,
+ 428,
+ 680
+ ],
+ "score": 1.0,
+ "content": "system of Swift to guarantee type safety of the DSL. Tensor types are wrapped in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 428,
+ 667,
+ 465,
+ 677
+ ],
+ "score": 0.83,
+ "content": "\\mathrm { R e p } { < } \\mathrm { T } >",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 465,
+ 665,
+ 506,
+ 680
+ ],
+ "score": 1.0,
+ "content": ", meaning",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 676,
+ 506,
+ 691
+ ],
+ "score": 1.0,
+ "content": "the representation of some computation that produces data of type T. Tensor operators overloaded for",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 505,
+ 700
+ ],
+ "score": 1.0,
+ "content": "Rep are essentially AST builders for delayed evaluation. Instead of generating computation nodes at",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 699,
+ 506,
+ 711
+ ],
+ "score": 1.0,
+ "content": "runtime and performing operator-overloading AD like PyTorch or TensorFlow Eager (Google Brain",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 709,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Team, 2017), NNKit tensor computations are staged once during the lifetime of the host program. At",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 734
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 721,
+ 505,
+ 734
+ ],
+ "score": 1.0,
+ "content": "invocation time of staged functions, NNKit emits shape-specialized DLVM IR and leverages DLVM",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 106,
+ 398,
+ 505,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 505,
+ 411
+ ],
+ "score": 1.0,
+ "content": "to perform AD, optimizations, and low-level code generation. A sample of Swift code using NNKit",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 410,
+ 193,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 410,
+ 193,
+ 423
+ ],
+ "score": 1.0,
+ "content": "is shown in Figure 4.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 47,
+ "bbox_fs": [
+ 105,
+ 633,
+ 506,
+ 734
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 107,
+ 82,
+ 503,
+ 347
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 107,
+ 82,
+ 503,
+ 347
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 82,
+ 503,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 82,
+ 503,
+ 347
+ ],
+ "score": 0.954,
+ "type": "image",
+ "image_path": "0d26f4d3b2eee12db38e5fe6c1e74f9f289abe5662e934db33064114cc501489.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 107,
+ 82,
+ 503,
+ 170.33333333333331
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 107,
+ 170.33333333333331,
+ 503,
+ 258.66666666666663
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 107,
+ 258.66666666666663,
+ 503,
+ 346.99999999999994
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 145,
+ 364,
+ 461,
+ 376
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 362,
+ 463,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 362,
+ 463,
+ 379
+ ],
+ "score": 1.0,
+ "content": "Figure 4: Example code in Swift using NNKit, a staged DSL targeting DLVM.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 398,
+ 505,
+ 421
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 505,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 505,
+ 411
+ ],
+ "score": 1.0,
+ "content": "to perform AD, optimizations, and low-level code generation. A sample of Swift code using NNKit",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 410,
+ 193,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 410,
+ 193,
+ 423
+ ],
+ "score": 1.0,
+ "content": "is shown in Figure 4.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 427,
+ 505,
+ 493
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 425,
+ 507,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 425,
+ 507,
+ 441
+ ],
+ "score": 1.0,
+ "content": "The NNKit just-in-time compiler has four important phases: The first phase, expression staging,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 438,
+ 507,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 438,
+ 507,
+ 451
+ ],
+ "score": 1.0,
+ "content": "produces an unshaped graph IR of the tensor computation. The second phase, shape specialization,",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 449,
+ 506,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 449,
+ 506,
+ 462
+ ],
+ "score": 1.0,
+ "content": "prepares to generate statically shaped DLVM IR for staged functions when they are applied to shaped",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 459,
+ 506,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 459,
+ 506,
+ 473
+ ],
+ "score": 1.0,
+ "content": "tensors. The third phase, lowering, generates DLVM IR and passes it through DLVM, producing a",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 484
+ ],
+ "score": 1.0,
+ "content": "dynamic library containing a function symbol. The final phase, function reification, loads the binary",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 482,
+ 319,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 319,
+ 494
+ ],
+ "score": 1.0,
+ "content": "and wraps the low-level function to a Swift function.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 8.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 498,
+ 504,
+ 521
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 505,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 505,
+ 512
+ ],
+ "score": 1.0,
+ "content": "We anticipate other existing deep learning frameworks, such as TensorFlow, could be adapted to use",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 510,
+ 302,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 510,
+ 302,
+ 522
+ ],
+ "score": 1.0,
+ "content": "DLVM as a back-end to their tensor math DSLs.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 12.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 540,
+ 195,
+ 552
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 197,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 197,
+ 556
+ ],
+ "score": 1.0,
+ "content": "4 CONCLUSION",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 567,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 506,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 506,
+ 581
+ ],
+ "score": 1.0,
+ "content": "The deep learning research community has a rich variety of available frameworks. While two existing",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 578,
+ 506,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 578,
+ 506,
+ 592
+ ],
+ "score": 1.0,
+ "content": "projects have attempted a compilers approach to deep learning frameworks, and have respectively",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 588,
+ 506,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 588,
+ 506,
+ 603
+ ],
+ "score": 1.0,
+ "content": "achieved good integration with existing systems (TensorFlow XLA) and good performance (NNVM",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 599,
+ 506,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 138,
+ 611
+ ],
+ "score": 0.5,
+ "content": "+ \\mathrm { T V M } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 139,
+ 599,
+ 506,
+ 614
+ ],
+ "score": 1.0,
+ "content": ", their design philosophies have not entirely followed established best practices in optimizing",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 104,
+ 610,
+ 506,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 610,
+ 506,
+ 623
+ ],
+ "score": 1.0,
+ "content": "compiler design. While well intentioned, the remaining vast majority of other frameworks have failed",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 635
+ ],
+ "score": 1.0,
+ "content": "to observe that the problem of front-end DSLs, algorithmic differentiation, and converting a neural",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 646
+ ],
+ "score": 1.0,
+ "content": "network into efficient executable code is, at its core, a compilers problem. As a result, important issues",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 644,
+ 506,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 506,
+ 656
+ ],
+ "score": 1.0,
+ "content": "of extensibility and optimization have been addressed in less than optimal fashion in such frameworks.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "Nevertheless, several such frameworks have achieved wide adoption. We believe that the principled",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "application of optimizing compiler techniques will lead to substantial improvements in the tools",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "available to deep learning researchers. DLVM and its associated front-end DSLs have a major role",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 507,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 507,
+ 702
+ ],
+ "score": 1.0,
+ "content": "to play in this future. Our existing implementation supports reverse-mode AD in the core language,",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 506,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 506,
+ 712
+ ],
+ "score": 1.0,
+ "content": "and utilizes LLVM to target NVIDIA GPUs. In our ongoing work, we plan to substantially increase",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "the number of supported hardware architectures by utilizing HPVM as an additional back-end, and",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 438,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 438,
+ 732
+ ],
+ "score": 1.0,
+ "content": "explore more advanced AD techniques such as mixing forward and reverse modes.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "page_idx": 7,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 224,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 225,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 26,
+ 225,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Workshop track - ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 752,
+ 308,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 750,
+ 310,
+ 762
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 12,
+ "width": 9
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 107,
+ 82,
+ 503,
+ 347
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 107,
+ 82,
+ 503,
+ 347
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 82,
+ 503,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 82,
+ 503,
+ 347
+ ],
+ "score": 0.954,
+ "type": "image",
+ "image_path": "0d26f4d3b2eee12db38e5fe6c1e74f9f289abe5662e934db33064114cc501489.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 107,
+ 82,
+ 503,
+ 170.33333333333331
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 107,
+ 170.33333333333331,
+ 503,
+ 258.66666666666663
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 107,
+ 258.66666666666663,
+ 503,
+ 346.99999999999994
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 145,
+ 364,
+ 461,
+ 376
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 362,
+ 463,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 362,
+ 463,
+ 379
+ ],
+ "score": 1.0,
+ "content": "Figure 4: Example code in Swift using NNKit, a staged DSL targeting DLVM.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 398,
+ 505,
+ 421
+ ],
+ "lines": [],
+ "index": 4.5,
+ "bbox_fs": [
+ 105,
+ 398,
+ 505,
+ 423
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 427,
+ 505,
+ 493
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 425,
+ 507,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 425,
+ 507,
+ 441
+ ],
+ "score": 1.0,
+ "content": "The NNKit just-in-time compiler has four important phases: The first phase, expression staging,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 438,
+ 507,
+ 451
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 438,
+ 507,
+ 451
+ ],
+ "score": 1.0,
+ "content": "produces an unshaped graph IR of the tensor computation. The second phase, shape specialization,",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 449,
+ 506,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 449,
+ 506,
+ 462
+ ],
+ "score": 1.0,
+ "content": "prepares to generate statically shaped DLVM IR for staged functions when they are applied to shaped",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 459,
+ 506,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 459,
+ 506,
+ 473
+ ],
+ "score": 1.0,
+ "content": "tensors. The third phase, lowering, generates DLVM IR and passes it through DLVM, producing a",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 484
+ ],
+ "score": 1.0,
+ "content": "dynamic library containing a function symbol. The final phase, function reification, loads the binary",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 482,
+ 319,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 319,
+ 494
+ ],
+ "score": 1.0,
+ "content": "and wraps the low-level function to a Swift function.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 8.5,
+ "bbox_fs": [
+ 105,
+ 425,
+ 507,
+ 494
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 498,
+ 504,
+ 521
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 505,
+ 512
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 497,
+ 505,
+ 512
+ ],
+ "score": 1.0,
+ "content": "We anticipate other existing deep learning frameworks, such as TensorFlow, could be adapted to use",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 510,
+ 302,
+ 522
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 510,
+ 302,
+ 522
+ ],
+ "score": 1.0,
+ "content": "DLVM as a back-end to their tensor math DSLs.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 12.5,
+ "bbox_fs": [
+ 105,
+ 497,
+ 505,
+ 522
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 540,
+ 195,
+ 552
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 197,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 197,
+ 556
+ ],
+ "score": 1.0,
+ "content": "4 CONCLUSION",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 567,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 506,
+ 581
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 506,
+ 581
+ ],
+ "score": 1.0,
+ "content": "The deep learning research community has a rich variety of available frameworks. While two existing",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 578,
+ 506,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 578,
+ 506,
+ 592
+ ],
+ "score": 1.0,
+ "content": "projects have attempted a compilers approach to deep learning frameworks, and have respectively",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 588,
+ 506,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 588,
+ 506,
+ 603
+ ],
+ "score": 1.0,
+ "content": "achieved good integration with existing systems (TensorFlow XLA) and good performance (NNVM",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 599,
+ 506,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 138,
+ 611
+ ],
+ "score": 0.5,
+ "content": "+ \\mathrm { T V M } )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 139,
+ 599,
+ 506,
+ 614
+ ],
+ "score": 1.0,
+ "content": ", their design philosophies have not entirely followed established best practices in optimizing",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 104,
+ 610,
+ 506,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 610,
+ 506,
+ 623
+ ],
+ "score": 1.0,
+ "content": "compiler design. While well intentioned, the remaining vast majority of other frameworks have failed",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 635
+ ],
+ "score": 1.0,
+ "content": "to observe that the problem of front-end DSLs, algorithmic differentiation, and converting a neural",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 646
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 633,
+ 506,
+ 646
+ ],
+ "score": 1.0,
+ "content": "network into efficient executable code is, at its core, a compilers problem. As a result, important issues",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 644,
+ 506,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 644,
+ 506,
+ 656
+ ],
+ "score": 1.0,
+ "content": "of extensibility and optimization have been addressed in less than optimal fashion in such frameworks.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 505,
+ 667
+ ],
+ "score": 1.0,
+ "content": "Nevertheless, several such frameworks have achieved wide adoption. We believe that the principled",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 506,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 506,
+ 678
+ ],
+ "score": 1.0,
+ "content": "application of optimizing compiler techniques will lead to substantial improvements in the tools",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 677,
+ 505,
+ 689
+ ],
+ "score": 1.0,
+ "content": "available to deep learning researchers. DLVM and its associated front-end DSLs have a major role",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 687,
+ 507,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 687,
+ 507,
+ 702
+ ],
+ "score": 1.0,
+ "content": "to play in this future. Our existing implementation supports reverse-mode AD in the core language,",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 699,
+ 506,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 699,
+ 506,
+ 712
+ ],
+ "score": 1.0,
+ "content": "and utilizes LLVM to target NVIDIA GPUs. In our ongoing work, we plan to substantially increase",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 506,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 506,
+ 722
+ ],
+ "score": 1.0,
+ "content": "the number of supported hardware architectures by utilizing HPVM as an additional back-end, and",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 721,
+ 438,
+ 732
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 721,
+ 438,
+ 732
+ ],
+ "score": 1.0,
+ "content": "explore more advanced AD techniques such as mixing forward and reverse modes.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 22,
+ "bbox_fs": [
+ 104,
+ 565,
+ 507,
+ 732
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 82,
+ 175,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 81,
+ 176,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 81,
+ 176,
+ 95
+ ],
+ "score": 1.0,
+ "content": "REFERENCES",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 100,
+ 506,
+ 200
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 100,
+ 507,
+ 113
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 100,
+ 507,
+ 113
+ ],
+ "score": 1.0,
+ "content": "Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Gregory S.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 116,
+ 112,
+ 505,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 112,
+ 505,
+ 124
+ ],
+ "score": 1.0,
+ "content": "Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian J. Goodfellow, Andrew",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 115,
+ 122,
+ 505,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 122,
+ 505,
+ 135
+ ],
+ "score": 1.0,
+ "content": "Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Józefowicz, Lukasz Kaiser, Manjunath",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 115,
+ 133,
+ 506,
+ 146
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 133,
+ 506,
+ 146
+ ],
+ "score": 1.0,
+ "content": "Kudlur, Josh Levenberg, Dan Mané, Rajat Monga, Sherry Moore, Derek Gordon Murray, Chris",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 115,
+ 144,
+ 506,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 144,
+ 506,
+ 157
+ ],
+ "score": 1.0,
+ "content": "Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul A.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 115,
+ 154,
+ 506,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 154,
+ 506,
+ 169
+ ],
+ "score": 1.0,
+ "content": "Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda B. Viégas, Oriol Vinyals, Pete Warden,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 116,
+ 167,
+ 505,
+ 179
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 167,
+ 505,
+ 179
+ ],
+ "score": 1.0,
+ "content": "Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 115,
+ 177,
+ 506,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 177,
+ 506,
+ 190
+ ],
+ "score": 1.0,
+ "content": "machine learning on heterogeneous distributed systems. CoRR, abs/1603.04467, 2016. URL",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 115,
+ 189,
+ 307,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 189,
+ 307,
+ 200
+ ],
+ "score": 1.0,
+ "content": "http://arxiv.org/abs/1603.04467.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 208,
+ 506,
+ 252
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 208,
+ 506,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 208,
+ 506,
+ 221
+ ],
+ "score": 1.0,
+ "content": "Tianqi Chen, Mu Li, Yutian Li, Min Lin, Naiyan Wang, Minjie Wang, Tianjun Xiao, Bing Xu,",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 116,
+ 219,
+ 506,
+ 231
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 219,
+ 506,
+ 231
+ ],
+ "score": 1.0,
+ "content": "Chiyuan Zhang, and Zheng Zhang. Mxnet: A flexible and efficient machine learning library for",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 115,
+ 231,
+ 505,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 231,
+ 505,
+ 243
+ ],
+ "score": 1.0,
+ "content": "heterogeneous distributed systems. CoRR, abs/1512.01274, 2015. URL http://arxiv.org/",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 116,
+ 241,
+ 206,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 241,
+ 206,
+ 253
+ ],
+ "score": 1.0,
+ "content": "abs/1512.01274.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 261,
+ 505,
+ 295
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 261,
+ 505,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 261,
+ 505,
+ 273
+ ],
+ "score": 1.0,
+ "content": "Tianqi Chen, Thierry Moreau, Ziheng Jiang, and Haichen Shen. TVM: An end to end IR stack for",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 116,
+ 272,
+ 505,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 272,
+ 505,
+ 285
+ ],
+ "score": 1.0,
+ "content": "deploying deep learning workloads on hardware platforms. http://tvmlang.org/2017/",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 116,
+ 283,
+ 357,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 283,
+ 357,
+ 295
+ ],
+ "score": 1.0,
+ "content": "08/17/tvm-release-announcement.html, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 303,
+ 505,
+ 337
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 303,
+ 505,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 303,
+ 505,
+ 316
+ ],
+ "score": 1.0,
+ "content": "Ronan Collobert, Koray Kavukcuoglu, and Clément Farabet. Torch7: A Matlab-like environment for",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 314,
+ 505,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 314,
+ 505,
+ 328
+ ],
+ "score": 1.0,
+ "content": "machine learning. In NIPS Big Learning Workshop: Algorithms, Systems, and Tools for Learning",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 115,
+ 326,
+ 222,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 326,
+ 222,
+ 336
+ ],
+ "score": 1.0,
+ "content": "at Scale, December 2011.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 345,
+ 506,
+ 412
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 506,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 506,
+ 358
+ ],
+ "score": 1.0,
+ "content": "Amol Ghoting, Rajasekar Krishnamurthy, Edwin Pednault, Berthold Reinwald, Vikas Sindhwani,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 117,
+ 357,
+ 505,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 357,
+ 505,
+ 369
+ ],
+ "score": 1.0,
+ "content": "Shirish Tatikonda, Yuanyuan Tian, and Shivakumar Vaithyanathan. Systemml: Declarative machine",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 115,
+ 367,
+ 506,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 367,
+ 506,
+ 381
+ ],
+ "score": 1.0,
+ "content": "learning on MapReduce. In Proceedings of the 2011 IEEE 27th International Conference on",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 114,
+ 378,
+ 507,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 378,
+ 507,
+ 392
+ ],
+ "score": 1.0,
+ "content": "Data Engineering, ICDE ’11, pp. 231–242, Washington, DC, USA, 2011. IEEE Computer Society.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 115,
+ 389,
+ 506,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 389,
+ 506,
+ 403
+ ],
+ "score": 1.0,
+ "content": "ISBN 978-1-4244-8959-6. doi: 10.1109/ICDE.2011.5767930. URL http://dx.doi.org/",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 116,
+ 401,
+ 271,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 401,
+ 271,
+ 411
+ ],
+ "score": 1.0,
+ "content": "10.1109/ICDE.2011.5767930.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 420,
+ 505,
+ 443
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 507,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 507,
+ 434
+ ],
+ "score": 1.0,
+ "content": "Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016. http:",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 116,
+ 430,
+ 277,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 430,
+ 277,
+ 446
+ ],
+ "score": 1.0,
+ "content": "//www.deeplearningbook.org.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 451,
+ 506,
+ 486
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 451,
+ 506,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 451,
+ 203,
+ 464
+ ],
+ "score": 1.0,
+ "content": "Google Brain Team.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 220,
+ 451,
+ 506,
+ 465
+ ],
+ "score": 1.0,
+ "content": "Eager Execution: An imperative, define-by-run interface to",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 462,
+ 505,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 462,
+ 169,
+ 474
+ ],
+ "score": 1.0,
+ "content": "TensorFlow.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 188,
+ 462,
+ 216,
+ 474
+ ],
+ "score": 1.0,
+ "content": "2017.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 234,
+ 462,
+ 505,
+ 475
+ ],
+ "score": 1.0,
+ "content": "URL https://research.googleblog.com/2017/10/",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 116,
+ 474,
+ 369,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 474,
+ 369,
+ 486
+ ],
+ "score": 1.0,
+ "content": "eager-execution-imperative-define-by.html.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 493,
+ 506,
+ 527
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 507
+ ],
+ "score": 1.0,
+ "content": "Joe Groff and Chris Lattner. Swift’s High-Level IR: A Case Study of Complementing LLVM IR",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 115,
+ 504,
+ 507,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 504,
+ 507,
+ 518
+ ],
+ "score": 1.0,
+ "content": "with Language-Specific Optimization. 2015 LLVM Developers’ Meeting, 2015. URL http:",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 117,
+ 517,
+ 504,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 517,
+ 504,
+ 528
+ ],
+ "score": 1.0,
+ "content": "//llvm.org/devmtg/2015-10/slides/GroffLattner-SILHighLevelIR.pdf.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 536,
+ 506,
+ 689
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 535,
+ 506,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 535,
+ 506,
+ 549
+ ],
+ "score": 1.0,
+ "content": "Norman P. Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 547,
+ 506,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 547,
+ 506,
+ 560
+ ],
+ "score": 1.0,
+ "content": "Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, Rick Boyle, Pierre-luc Cantin, Clifford",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 116,
+ 558,
+ 506,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 558,
+ 506,
+ 570
+ ],
+ "score": 1.0,
+ "content": "Chao, Chris Clark, Jeremy Coriell, Mike Daley, Matt Dau, Jeffrey Dean, Ben Gelb, Tara Vazir",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 115,
+ 567,
+ 506,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 567,
+ 506,
+ 583
+ ],
+ "score": 1.0,
+ "content": "Ghaemmaghami, Rajendra Gottipati, William Gulland, Robert Hagmann, Richard C. Ho, Doug",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 115,
+ 579,
+ 506,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 579,
+ 506,
+ 592
+ ],
+ "score": 1.0,
+ "content": "Hogberg, John Hu, Robert Hundt, Dan Hurt, Julian Ibarz, Aaron Jaffey, Alek Jaworski, Alexander",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 115,
+ 590,
+ 506,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 590,
+ 506,
+ 604
+ ],
+ "score": 1.0,
+ "content": "Kaplan, Harshit Khaitan, Andy Koch, Naveen Kumar, Steve Lacy, James Laudon, James Law,",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 115,
+ 601,
+ 505,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 601,
+ 505,
+ 614
+ ],
+ "score": 1.0,
+ "content": "Diemthu Le, Chris Leary, Zhuyuan Liu, Kyle Lucke, Alan Lundin, Gordon MacKean, Adriana",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 115,
+ 612,
+ 505,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 612,
+ 505,
+ 626
+ ],
+ "score": 1.0,
+ "content": "Maggiore, Maire Mahony, Kieran Miller, Rahul Nagarajan, Ravi Narayanaswami, Ray Ni, Kathy",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 115,
+ 623,
+ 505,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 623,
+ 505,
+ 636
+ ],
+ "score": 1.0,
+ "content": "Nix, Thomas Norrie, Mark Omernick, Narayana Penukonda, Andy Phelps, Jonathan Ross, Amir",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 115,
+ 634,
+ 505,
+ 648
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 634,
+ 505,
+ 648
+ ],
+ "score": 1.0,
+ "content": "Salek, Emad Samadiani, Chris Severn, Gregory Sizikov, Matthew Snelham, Jed Souter, Dan",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 115,
+ 645,
+ 506,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 645,
+ 506,
+ 658
+ ],
+ "score": 1.0,
+ "content": "Steinberg, Andy Swing, Mercedes Tan, Gregory Thorson, Bo Tian, Horia Toma, Erick Tuttle,",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 116,
+ 657,
+ 506,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 657,
+ 506,
+ 669
+ ],
+ "score": 1.0,
+ "content": "Vijay Vasudevan, Richard Walter, Walter Wang, Eric Wilcox, and Doe Hyun Yoon. In-datacenter",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 114,
+ 667,
+ 507,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 667,
+ 507,
+ 681
+ ],
+ "score": 1.0,
+ "content": "performance analysis of a tensor processing unit. CoRR, abs/1704.04760, 2017. URL http:",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 116,
+ 679,
+ 276,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 679,
+ 276,
+ 690
+ ],
+ "score": 1.0,
+ "content": "//arxiv.org/abs/1704.04760.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 40.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 699,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 712
+ ],
+ "score": 1.0,
+ "content": "Fredrik Kjolstad, Shoaib Kamil, Stephen Chou, David Lugato, and Saman Amarasinghe. The tensor",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 115,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "algebra compiler. Proc. ACM Program. Lang., 1(OOPSLA):77:1–77:29, October 2017. ISSN",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 115,
+ 720,
+ 492,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 720,
+ 492,
+ 733
+ ],
+ "score": 1.0,
+ "content": "2475-1421. doi: 10.1145/3133901. URL http://doi.acm.org/10.1145/3133901.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "page_idx": 8,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 27,
+ 224,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 225,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 225,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Workshop track - ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 759
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 751,
+ 309,
+ 762
+ ],
+ "score": 1.0,
+ "content": "9",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 82,
+ 175,
+ 93
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 81,
+ 176,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 81,
+ 176,
+ 95
+ ],
+ "score": 1.0,
+ "content": "REFERENCES",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 100,
+ 506,
+ 200
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 100,
+ 507,
+ 113
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 100,
+ 507,
+ 113
+ ],
+ "score": 1.0,
+ "content": "Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Gregory S.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 116,
+ 112,
+ 505,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 112,
+ 505,
+ 124
+ ],
+ "score": 1.0,
+ "content": "Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian J. Goodfellow, Andrew",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 115,
+ 122,
+ 505,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 122,
+ 505,
+ 135
+ ],
+ "score": 1.0,
+ "content": "Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Józefowicz, Lukasz Kaiser, Manjunath",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 115,
+ 133,
+ 506,
+ 146
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 133,
+ 506,
+ 146
+ ],
+ "score": 1.0,
+ "content": "Kudlur, Josh Levenberg, Dan Mané, Rajat Monga, Sherry Moore, Derek Gordon Murray, Chris",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 115,
+ 144,
+ 506,
+ 157
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 144,
+ 506,
+ 157
+ ],
+ "score": 1.0,
+ "content": "Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul A.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 115,
+ 154,
+ 506,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 154,
+ 506,
+ 169
+ ],
+ "score": 1.0,
+ "content": "Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda B. Viégas, Oriol Vinyals, Pete Warden,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 116,
+ 167,
+ 505,
+ 179
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 167,
+ 505,
+ 179
+ ],
+ "score": 1.0,
+ "content": "Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 115,
+ 177,
+ 506,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 177,
+ 506,
+ 190
+ ],
+ "score": 1.0,
+ "content": "machine learning on heterogeneous distributed systems. CoRR, abs/1603.04467, 2016. URL",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 115,
+ 189,
+ 307,
+ 200
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 189,
+ 307,
+ 200
+ ],
+ "score": 1.0,
+ "content": "http://arxiv.org/abs/1603.04467.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 5,
+ "bbox_fs": [
+ 105,
+ 100,
+ 507,
+ 200
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 208,
+ 506,
+ 252
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 208,
+ 506,
+ 221
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 208,
+ 506,
+ 221
+ ],
+ "score": 1.0,
+ "content": "Tianqi Chen, Mu Li, Yutian Li, Min Lin, Naiyan Wang, Minjie Wang, Tianjun Xiao, Bing Xu,",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 116,
+ 219,
+ 506,
+ 231
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 219,
+ 506,
+ 231
+ ],
+ "score": 1.0,
+ "content": "Chiyuan Zhang, and Zheng Zhang. Mxnet: A flexible and efficient machine learning library for",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 115,
+ 231,
+ 505,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 231,
+ 505,
+ 243
+ ],
+ "score": 1.0,
+ "content": "heterogeneous distributed systems. CoRR, abs/1512.01274, 2015. URL http://arxiv.org/",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 116,
+ 241,
+ 206,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 241,
+ 206,
+ 253
+ ],
+ "score": 1.0,
+ "content": "abs/1512.01274.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 11.5,
+ "bbox_fs": [
+ 105,
+ 208,
+ 506,
+ 253
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 261,
+ 505,
+ 295
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 261,
+ 505,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 261,
+ 505,
+ 273
+ ],
+ "score": 1.0,
+ "content": "Tianqi Chen, Thierry Moreau, Ziheng Jiang, and Haichen Shen. TVM: An end to end IR stack for",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 116,
+ 272,
+ 505,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 272,
+ 505,
+ 285
+ ],
+ "score": 1.0,
+ "content": "deploying deep learning workloads on hardware platforms. http://tvmlang.org/2017/",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 116,
+ 283,
+ 357,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 283,
+ 357,
+ 295
+ ],
+ "score": 1.0,
+ "content": "08/17/tvm-release-announcement.html, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 15,
+ "bbox_fs": [
+ 106,
+ 261,
+ 505,
+ 295
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 303,
+ 505,
+ 337
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 303,
+ 505,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 303,
+ 505,
+ 316
+ ],
+ "score": 1.0,
+ "content": "Ronan Collobert, Koray Kavukcuoglu, and Clément Farabet. Torch7: A Matlab-like environment for",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 314,
+ 505,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 314,
+ 505,
+ 328
+ ],
+ "score": 1.0,
+ "content": "machine learning. In NIPS Big Learning Workshop: Algorithms, Systems, and Tools for Learning",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 115,
+ 326,
+ 222,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 326,
+ 222,
+ 336
+ ],
+ "score": 1.0,
+ "content": "at Scale, December 2011.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 18,
+ "bbox_fs": [
+ 106,
+ 303,
+ 505,
+ 336
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 345,
+ 506,
+ 412
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 506,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 345,
+ 506,
+ 358
+ ],
+ "score": 1.0,
+ "content": "Amol Ghoting, Rajasekar Krishnamurthy, Edwin Pednault, Berthold Reinwald, Vikas Sindhwani,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 117,
+ 357,
+ 505,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 357,
+ 505,
+ 369
+ ],
+ "score": 1.0,
+ "content": "Shirish Tatikonda, Yuanyuan Tian, and Shivakumar Vaithyanathan. Systemml: Declarative machine",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 115,
+ 367,
+ 506,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 367,
+ 506,
+ 381
+ ],
+ "score": 1.0,
+ "content": "learning on MapReduce. In Proceedings of the 2011 IEEE 27th International Conference on",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 114,
+ 378,
+ 507,
+ 392
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 378,
+ 507,
+ 392
+ ],
+ "score": 1.0,
+ "content": "Data Engineering, ICDE ’11, pp. 231–242, Washington, DC, USA, 2011. IEEE Computer Society.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 115,
+ 389,
+ 506,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 389,
+ 506,
+ 403
+ ],
+ "score": 1.0,
+ "content": "ISBN 978-1-4244-8959-6. doi: 10.1109/ICDE.2011.5767930. URL http://dx.doi.org/",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 116,
+ 401,
+ 271,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 401,
+ 271,
+ 411
+ ],
+ "score": 1.0,
+ "content": "10.1109/ICDE.2011.5767930.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22.5,
+ "bbox_fs": [
+ 105,
+ 345,
+ 507,
+ 411
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 420,
+ 505,
+ 443
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 507,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 507,
+ 434
+ ],
+ "score": 1.0,
+ "content": "Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep Learning. MIT Press, 2016. http:",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 116,
+ 430,
+ 277,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 430,
+ 277,
+ 446
+ ],
+ "score": 1.0,
+ "content": "//www.deeplearningbook.org.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 105,
+ 419,
+ 507,
+ 446
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 451,
+ 506,
+ 486
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 451,
+ 506,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 451,
+ 203,
+ 464
+ ],
+ "score": 1.0,
+ "content": "Google Brain Team.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 220,
+ 451,
+ 506,
+ 465
+ ],
+ "score": 1.0,
+ "content": "Eager Execution: An imperative, define-by-run interface to",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 462,
+ 505,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 462,
+ 169,
+ 474
+ ],
+ "score": 1.0,
+ "content": "TensorFlow.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 188,
+ 462,
+ 216,
+ 474
+ ],
+ "score": 1.0,
+ "content": "2017.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 234,
+ 462,
+ 505,
+ 475
+ ],
+ "score": 1.0,
+ "content": "URL https://research.googleblog.com/2017/10/",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 116,
+ 474,
+ 369,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 474,
+ 369,
+ 486
+ ],
+ "score": 1.0,
+ "content": "eager-execution-imperative-define-by.html.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 29,
+ "bbox_fs": [
+ 106,
+ 451,
+ 506,
+ 486
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 493,
+ 506,
+ 527
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 507
+ ],
+ "score": 1.0,
+ "content": "Joe Groff and Chris Lattner. Swift’s High-Level IR: A Case Study of Complementing LLVM IR",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 115,
+ 504,
+ 507,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 504,
+ 507,
+ 518
+ ],
+ "score": 1.0,
+ "content": "with Language-Specific Optimization. 2015 LLVM Developers’ Meeting, 2015. URL http:",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 117,
+ 517,
+ 504,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 517,
+ 504,
+ 528
+ ],
+ "score": 1.0,
+ "content": "//llvm.org/devmtg/2015-10/slides/GroffLattner-SILHighLevelIR.pdf.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32,
+ "bbox_fs": [
+ 105,
+ 492,
+ 507,
+ 528
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 536,
+ 506,
+ 689
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 535,
+ 506,
+ 549
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 535,
+ 506,
+ 549
+ ],
+ "score": 1.0,
+ "content": "Norman P. Jouppi, Cliff Young, Nishant Patil, David Patterson, Gaurav Agrawal, Raminder Bajwa,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 547,
+ 506,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 547,
+ 506,
+ 560
+ ],
+ "score": 1.0,
+ "content": "Sarah Bates, Suresh Bhatia, Nan Boden, Al Borchers, Rick Boyle, Pierre-luc Cantin, Clifford",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 116,
+ 558,
+ 506,
+ 570
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 558,
+ 506,
+ 570
+ ],
+ "score": 1.0,
+ "content": "Chao, Chris Clark, Jeremy Coriell, Mike Daley, Matt Dau, Jeffrey Dean, Ben Gelb, Tara Vazir",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 115,
+ 567,
+ 506,
+ 583
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 567,
+ 506,
+ 583
+ ],
+ "score": 1.0,
+ "content": "Ghaemmaghami, Rajendra Gottipati, William Gulland, Robert Hagmann, Richard C. Ho, Doug",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 115,
+ 579,
+ 506,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 579,
+ 506,
+ 592
+ ],
+ "score": 1.0,
+ "content": "Hogberg, John Hu, Robert Hundt, Dan Hurt, Julian Ibarz, Aaron Jaffey, Alek Jaworski, Alexander",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 115,
+ 590,
+ 506,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 590,
+ 506,
+ 604
+ ],
+ "score": 1.0,
+ "content": "Kaplan, Harshit Khaitan, Andy Koch, Naveen Kumar, Steve Lacy, James Laudon, James Law,",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 115,
+ 601,
+ 505,
+ 614
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 601,
+ 505,
+ 614
+ ],
+ "score": 1.0,
+ "content": "Diemthu Le, Chris Leary, Zhuyuan Liu, Kyle Lucke, Alan Lundin, Gordon MacKean, Adriana",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 115,
+ 612,
+ 505,
+ 626
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 612,
+ 505,
+ 626
+ ],
+ "score": 1.0,
+ "content": "Maggiore, Maire Mahony, Kieran Miller, Rahul Nagarajan, Ravi Narayanaswami, Ray Ni, Kathy",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 115,
+ 623,
+ 505,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 623,
+ 505,
+ 636
+ ],
+ "score": 1.0,
+ "content": "Nix, Thomas Norrie, Mark Omernick, Narayana Penukonda, Andy Phelps, Jonathan Ross, Amir",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 115,
+ 634,
+ 505,
+ 648
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 634,
+ 505,
+ 648
+ ],
+ "score": 1.0,
+ "content": "Salek, Emad Samadiani, Chris Severn, Gregory Sizikov, Matthew Snelham, Jed Souter, Dan",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 115,
+ 645,
+ 506,
+ 658
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 645,
+ 506,
+ 658
+ ],
+ "score": 1.0,
+ "content": "Steinberg, Andy Swing, Mercedes Tan, Gregory Thorson, Bo Tian, Horia Toma, Erick Tuttle,",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 116,
+ 657,
+ 506,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 657,
+ 506,
+ 669
+ ],
+ "score": 1.0,
+ "content": "Vijay Vasudevan, Richard Walter, Walter Wang, Eric Wilcox, and Doe Hyun Yoon. In-datacenter",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 114,
+ 667,
+ 507,
+ 681
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 667,
+ 507,
+ 681
+ ],
+ "score": 1.0,
+ "content": "performance analysis of a tensor processing unit. CoRR, abs/1704.04760, 2017. URL http:",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 116,
+ 679,
+ 276,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 679,
+ 276,
+ 690
+ ],
+ "score": 1.0,
+ "content": "//arxiv.org/abs/1704.04760.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 106,
+ 535,
+ 507,
+ 690
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 699,
+ 505,
+ 732
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 712
+ ],
+ "score": 1.0,
+ "content": "Fredrik Kjolstad, Shoaib Kamil, Stephen Chou, David Lugato, and Saman Amarasinghe. The tensor",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 115,
+ 709,
+ 505,
+ 722
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 709,
+ 505,
+ 722
+ ],
+ "score": 1.0,
+ "content": "algebra compiler. Proc. ACM Program. Lang., 1(OOPSLA):77:1–77:29, October 2017. ISSN",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 115,
+ 720,
+ 492,
+ 733
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 720,
+ 492,
+ 733
+ ],
+ "score": 1.0,
+ "content": "2475-1421. doi: 10.1145/3133901. URL http://doi.acm.org/10.1145/3133901.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 49,
+ "bbox_fs": [
+ 105,
+ 698,
+ 506,
+ 733
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 82,
+ 504,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 82,
+ 506,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 82,
+ 506,
+ 96
+ ],
+ "score": 1.0,
+ "content": "Chris Lattner and Vikram Adve. LLVM: A Compilation Framework for Lifelong Program Analysis",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 115,
+ 94,
+ 506,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 94,
+ 506,
+ 106
+ ],
+ "score": 1.0,
+ "content": "& Transformation. In Proceedings of the 2004 International Symposium on Code Generation and",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 116,
+ 105,
+ 349,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 105,
+ 349,
+ 117
+ ],
+ "score": 1.0,
+ "content": "Optimization (CGO’04), Palo Alto, California, Mar 2004.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 123,
+ 504,
+ 146
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 121,
+ 506,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 121,
+ 506,
+ 137
+ ],
+ "score": 1.0,
+ "content": "Chris Leary and Todd Wang. XLA: TensorFlow, compiled! TensorFlow Dev Summit 2017, February",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 114,
+ 133,
+ 144,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 133,
+ 144,
+ 147
+ ],
+ "score": 1.0,
+ "content": "2017.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 153,
+ 504,
+ 186
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 152,
+ 505,
+ 166
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 152,
+ 505,
+ 166
+ ],
+ "score": 1.0,
+ "content": "Uwe Naumann. The Art of Differentiating Computer Programs. Society for Industrial and Applied",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 115,
+ 163,
+ 505,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 163,
+ 505,
+ 177
+ ],
+ "score": 1.0,
+ "content": "Mathematics, 2011. doi: 10.1137/1.9781611972078. URL http://epubs.siam.org/doi/",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 117,
+ 176,
+ 283,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 176,
+ 283,
+ 186
+ ],
+ "score": 1.0,
+ "content": "abs/10.1137/1.9781611972078.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 194,
+ 503,
+ 217
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 194,
+ 505,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 194,
+ 505,
+ 207
+ ],
+ "score": 1.0,
+ "content": "NNVM. NNVM compiler: Open compiler for AI frameworks. http://tvmlang.org/2017/",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 116,
+ 204,
+ 369,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 204,
+ 369,
+ 217
+ ],
+ "score": 1.0,
+ "content": "10/06/nnvm-compiler-announcement.html, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 223,
+ 503,
+ 247
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 223,
+ 504,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 223,
+ 504,
+ 236
+ ],
+ "score": 1.0,
+ "content": "PyTorch Development Team. Tensors and Dynamic neural networks in Python with strong GPU",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 115,
+ 233,
+ 333,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 233,
+ 333,
+ 249
+ ],
+ "score": 1.0,
+ "content": "acceleration. 2016. URL http://pytorch.org.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 253,
+ 506,
+ 320
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 266
+ ],
+ "score": 1.0,
+ "content": "Jonathan Ragan-Kelley, Connelly Barnes, Andrew Adams, Sylvain Paris, Frédo Durand, and Saman",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 115,
+ 264,
+ 506,
+ 278
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 264,
+ 506,
+ 278
+ ],
+ "score": 1.0,
+ "content": "Amarasinghe. Halide: A language and compiler for optimizing parallelism, locality, and recom-",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 114,
+ 275,
+ 506,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 275,
+ 506,
+ 289
+ ],
+ "score": 1.0,
+ "content": "putation in image processing pipelines. In Proceedings of the 34th ACM SIGPLAN Conference",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 114,
+ 286,
+ 507,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 286,
+ 507,
+ 299
+ ],
+ "score": 1.0,
+ "content": "on Programming Language Design and Implementation, PLDI ’13, pp. 519–530, New York,",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 116,
+ 298,
+ 506,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 298,
+ 506,
+ 309
+ ],
+ "score": 1.0,
+ "content": "NY, USA, 2013. ACM. ISBN 978-1-4503-2014-6. doi: 10.1145/2491956.2462176. URL",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 116,
+ 309,
+ 373,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 309,
+ 373,
+ 320
+ ],
+ "score": 1.0,
+ "content": "http://doi.acm.org/10.1145/2491956.2462176.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 14.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 327,
+ 505,
+ 383
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 326,
+ 505,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 326,
+ 505,
+ 341
+ ],
+ "score": 1.0,
+ "content": "Tiark Rompf and Martin Odersky. Lightweight modular staging: A pragmatic approach to runtime",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 115,
+ 338,
+ 505,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 338,
+ 505,
+ 352
+ ],
+ "score": 1.0,
+ "content": "code generation and compiled dsls. In Proceedings of the Ninth International Conference on",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 115,
+ 349,
+ 506,
+ 362
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 349,
+ 506,
+ 362
+ ],
+ "score": 1.0,
+ "content": "Generative Programming and Component Engineering, GPCE ’10, pp. 127–136, New York,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 361,
+ 506,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 361,
+ 506,
+ 372
+ ],
+ "score": 1.0,
+ "content": "NY, USA, 2010. ACM. ISBN 978-1-4503-0154-1. doi: 10.1145/1868294.1868314. URL",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 115,
+ 372,
+ 374,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 372,
+ 374,
+ 384
+ ],
+ "score": 1.0,
+ "content": "http://doi.acm.org/10.1145/1868294.1868314.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 389,
+ 505,
+ 424
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 390,
+ 505,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 390,
+ 505,
+ 402
+ ],
+ "score": 1.0,
+ "content": "David E. Rumelhart, James L. McClelland, and CORPORATE PDP Research Group (eds.). Parallel",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 115,
+ 401,
+ 505,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 401,
+ 505,
+ 414
+ ],
+ "score": 1.0,
+ "content": "Distributed Processing: Explorations in the Microstructure of Cognition, Vol. 1: Foundations. MIT",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 115,
+ 412,
+ 356,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 412,
+ 356,
+ 424
+ ],
+ "score": 1.0,
+ "content": "Press, Cambridge, MA, USA, 1986. ISBN 0-262-68053-X.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 431,
+ 506,
+ 465
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 431,
+ 505,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 431,
+ 505,
+ 443
+ ],
+ "score": 1.0,
+ "content": "Prakalp Srivastava, Maria Kotsifakou, and Vikram S. Adve. HPVM: A portable virtual instruction set",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 115,
+ 441,
+ 505,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 441,
+ 505,
+ 455
+ ],
+ "score": 1.0,
+ "content": "for heterogeneous parallel systems. CoRR, abs/1611.00860, 2016. URL http://arxiv.org/",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 116,
+ 454,
+ 205,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 454,
+ 205,
+ 465
+ ],
+ "score": 1.0,
+ "content": "abs/1611.00860.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "page_idx": 9,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 107,
+ 26,
+ 224,
+ 37
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 225,
+ 38
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 25,
+ 225,
+ 38
+ ],
+ "score": 1.0,
+ "content": "Workshop track - ICLR 2018",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 751,
+ 311,
+ 760
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 298,
+ 750,
+ 312,
+ 764
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 298,
+ 750,
+ 312,
+ 764
+ ],
+ "score": 1.0,
+ "content": "10",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 82,
+ 504,
+ 116
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 82,
+ 506,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 82,
+ 506,
+ 96
+ ],
+ "score": 1.0,
+ "content": "Chris Lattner and Vikram Adve. LLVM: A Compilation Framework for Lifelong Program Analysis",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 115,
+ 94,
+ 506,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 94,
+ 506,
+ 106
+ ],
+ "score": 1.0,
+ "content": "& Transformation. In Proceedings of the 2004 International Symposium on Code Generation and",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 116,
+ 105,
+ 349,
+ 117
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 105,
+ 349,
+ 117
+ ],
+ "score": 1.0,
+ "content": "Optimization (CGO’04), Palo Alto, California, Mar 2004.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1,
+ "bbox_fs": [
+ 107,
+ 82,
+ 506,
+ 117
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 123,
+ 504,
+ 146
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 121,
+ 506,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 121,
+ 506,
+ 137
+ ],
+ "score": 1.0,
+ "content": "Chris Leary and Todd Wang. XLA: TensorFlow, compiled! TensorFlow Dev Summit 2017, February",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 114,
+ 133,
+ 144,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 133,
+ 144,
+ 147
+ ],
+ "score": 1.0,
+ "content": "2017.",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ }
+ ],
+ "index": 3.5,
+ "bbox_fs": [
+ 105,
+ 121,
+ 506,
+ 147
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 153,
+ 504,
+ 186
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 152,
+ 505,
+ 166
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 152,
+ 505,
+ 166
+ ],
+ "score": 1.0,
+ "content": "Uwe Naumann. The Art of Differentiating Computer Programs. Society for Industrial and Applied",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 115,
+ 163,
+ 505,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 163,
+ 505,
+ 177
+ ],
+ "score": 1.0,
+ "content": "Mathematics, 2011. doi: 10.1137/1.9781611972078. URL http://epubs.siam.org/doi/",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 117,
+ 176,
+ 283,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 176,
+ 283,
+ 186
+ ],
+ "score": 1.0,
+ "content": "abs/10.1137/1.9781611972078.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 106,
+ 152,
+ 505,
+ 186
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 194,
+ 503,
+ 217
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 194,
+ 505,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 194,
+ 505,
+ 207
+ ],
+ "score": 1.0,
+ "content": "NNVM. NNVM compiler: Open compiler for AI frameworks. http://tvmlang.org/2017/",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 116,
+ 204,
+ 369,
+ 217
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 204,
+ 369,
+ 217
+ ],
+ "score": 1.0,
+ "content": "10/06/nnvm-compiler-announcement.html, 2017.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8.5,
+ "bbox_fs": [
+ 106,
+ 194,
+ 505,
+ 217
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 223,
+ 503,
+ 247
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 223,
+ 504,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 223,
+ 504,
+ 236
+ ],
+ "score": 1.0,
+ "content": "PyTorch Development Team. Tensors and Dynamic neural networks in Python with strong GPU",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 115,
+ 233,
+ 333,
+ 249
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 233,
+ 333,
+ 249
+ ],
+ "score": 1.0,
+ "content": "acceleration. 2016. URL http://pytorch.org.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10.5,
+ "bbox_fs": [
+ 106,
+ 223,
+ 504,
+ 249
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 253,
+ 506,
+ 320
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 254,
+ 505,
+ 266
+ ],
+ "score": 1.0,
+ "content": "Jonathan Ragan-Kelley, Connelly Barnes, Andrew Adams, Sylvain Paris, Frédo Durand, and Saman",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 115,
+ 264,
+ 506,
+ 278
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 264,
+ 506,
+ 278
+ ],
+ "score": 1.0,
+ "content": "Amarasinghe. Halide: A language and compiler for optimizing parallelism, locality, and recom-",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 114,
+ 275,
+ 506,
+ 289
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 275,
+ 506,
+ 289
+ ],
+ "score": 1.0,
+ "content": "putation in image processing pipelines. In Proceedings of the 34th ACM SIGPLAN Conference",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 114,
+ 286,
+ 507,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 286,
+ 507,
+ 299
+ ],
+ "score": 1.0,
+ "content": "on Programming Language Design and Implementation, PLDI ’13, pp. 519–530, New York,",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 116,
+ 298,
+ 506,
+ 309
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 298,
+ 506,
+ 309
+ ],
+ "score": 1.0,
+ "content": "NY, USA, 2013. ACM. ISBN 978-1-4503-2014-6. doi: 10.1145/2491956.2462176. URL",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 116,
+ 309,
+ 373,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 309,
+ 373,
+ 320
+ ],
+ "score": 1.0,
+ "content": "http://doi.acm.org/10.1145/2491956.2462176.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 14.5,
+ "bbox_fs": [
+ 106,
+ 254,
+ 507,
+ 320
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 327,
+ 505,
+ 383
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 326,
+ 505,
+ 341
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 326,
+ 505,
+ 341
+ ],
+ "score": 1.0,
+ "content": "Tiark Rompf and Martin Odersky. Lightweight modular staging: A pragmatic approach to runtime",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 115,
+ 338,
+ 505,
+ 352
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 338,
+ 505,
+ 352
+ ],
+ "score": 1.0,
+ "content": "code generation and compiled dsls. In Proceedings of the Ninth International Conference on",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 115,
+ 349,
+ 506,
+ 362
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 349,
+ 506,
+ 362
+ ],
+ "score": 1.0,
+ "content": "Generative Programming and Component Engineering, GPCE ’10, pp. 127–136, New York,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 361,
+ 506,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 361,
+ 506,
+ 372
+ ],
+ "score": 1.0,
+ "content": "NY, USA, 2010. ACM. ISBN 978-1-4503-0154-1. doi: 10.1145/1868294.1868314. URL",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 115,
+ 372,
+ 374,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 372,
+ 374,
+ 384
+ ],
+ "score": 1.0,
+ "content": "http://doi.acm.org/10.1145/1868294.1868314.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 20,
+ "bbox_fs": [
+ 106,
+ 326,
+ 506,
+ 384
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 389,
+ 505,
+ 424
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 390,
+ 505,
+ 402
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 390,
+ 505,
+ 402
+ ],
+ "score": 1.0,
+ "content": "David E. Rumelhart, James L. McClelland, and CORPORATE PDP Research Group (eds.). Parallel",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 115,
+ 401,
+ 505,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 401,
+ 505,
+ 414
+ ],
+ "score": 1.0,
+ "content": "Distributed Processing: Explorations in the Microstructure of Cognition, Vol. 1: Foundations. MIT",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 115,
+ 412,
+ 356,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 412,
+ 356,
+ 424
+ ],
+ "score": 1.0,
+ "content": "Press, Cambridge, MA, USA, 1986. ISBN 0-262-68053-X.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24,
+ "bbox_fs": [
+ 106,
+ 390,
+ 505,
+ 424
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 431,
+ 506,
+ 465
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 431,
+ 505,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 431,
+ 505,
+ 443
+ ],
+ "score": 1.0,
+ "content": "Prakalp Srivastava, Maria Kotsifakou, and Vikram S. Adve. HPVM: A portable virtual instruction set",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 115,
+ 441,
+ 505,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 441,
+ 505,
+ 455
+ ],
+ "score": 1.0,
+ "content": "for heterogeneous parallel systems. CoRR, abs/1611.00860, 2016. URL http://arxiv.org/",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 116,
+ 454,
+ 205,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 454,
+ 205,
+ 465
+ ],
+ "score": 1.0,
+ "content": "abs/1611.00860.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27,
+ "bbox_fs": [
+ 106,
+ 431,
+ 505,
+ 465
+ ]
+ }
+ ]
+ }
+ ],
+ "_backend": "pipeline",
+ "_version_name": "2.2.2"
+}
\ No newline at end of file