Loop - 13 vs 16

Files changed (1) hide show

1. Loop13 → Loop16 +36 -1

Loop13 → Loop16 RENAMED Viewed

	@@ -1 +1 @@
1 1	Generic Looping construct. This loop has multiple termination conditions:
2 2	1) Trip count. Iteration count specified at runtime. Set by
3 3	specifying the input M. Optional. Set to empty string to omit.
4 4	Note that a static trip count (specified at graph construction time) can be
5 5	specified by passing in a constant node for input M.
6 6	2) Loop termination condition. This is an input to the op that determines
7 7	whether to run the first iteration and also a loop-carried dependency for
8 8	the body graph. The body graph must yield a value for the condition variable,
9 9	whether this input is provided or not.
10 10	This table summarizes the operating modes of this operator with equivalent
11 11	C-style code:
12 12	Operator inputs defined as (max_trip_count, condition_var).
13 13	input ("", ""):
14 14	for (int i=0; ; ++i) {
15 15	cond = ... // Note this value is ignored, but is required in the body
16 16	}
17 17	input ("", cond) // Note this is analogous to a while loop
18 18	bool cond = ...;
19 19	for (int i=0; cond; ++i) {
20 20	cond = ...;
21 21	}
22 22	input ("", 1) // Note this is analogous to a do-while loop
23 23	bool cond = true
24 24	for (int i=0; cond; ++i) {
25 25	cond = ...;
26 26	}
27 27	input (trip_count, "") // Note this is analogous to a for loop
28 28	int trip_count = ...
29 29	for (int i=0; i < trip_count; ++i) {
30 30	cond = ...; // ignored
31 31	}
32 32	input (trip_count, cond)
33 33	int trip_count = ...;
34 34	bool cond = ...;
35 35	for (int i=0; i < trip_count && cond; ++i) {
36 36	cond = ...;
37 37	}
38 38	*Sample usage - cond as well as trip count*
39 39	graph predict-net {
40 40	%a = Constant[value = <Scalar Tensor [3]>]()
41 41	%b = Constant[value = <Scalar Tensor [6]>]()
42 42	%keepgoing = Constant[value = <Scalar Tensor [1]>]()
43 43	%max_trip_count = Constant[value = <Scalar Tensor [10]>]()
44 44	%keepgoing_out, %b_out, %user_defined_vals = Loop[body = <graph body-net>](%max_trip_count, %keepgoing, %b)
45 45	return
46 46	}
47 47	graph body-net (
48 48	%i[INT32, scalar] // iteration number
49 49	%keepgoing_in[BOOL, scalar] // incoming loop-termination-condition; not used
50 50	%b_in[INT32, scalar] // incoming value of loop-carried-dependency b
51 51	) {
52 52	%my_local = Add(%a, %b_in)
53 53	%b_out = Sub(%a, %b_in) // outgoing value of loop-carried-dependency b
54 54	%keepgoing_out = Greater(%my_local, %b_out) // outgoing loop-termination-condition
55 55	%user_defined_val = Add(%b_in, %b_in) // scan-output value to be accumulated
56 56	return %keepgoing_out, %b_out, %user_defined_val
57 57	}
58 58	*Sample equivalent C code*
59 59	{
60 60	/* User-defined code (enclosing scope) */
61 61	int a = 3, b = 6;
62 62	bool keepgoing = true; // Analogous to input cond
63 63	/* End user-defined code */
64 64	/* Implicitly-defined code */
65 65	const int max_trip_count = 10; // Analogous to input M
66 66	int user_defined_vals[]; // Imagine this is resizable
67 67	/* End implicitly-defined code */
68 68	/* initialize loop-carried variables and scan-output variables */
69 69	bool keepgoing_out = keepgoing
70 70	int b_out = b
71 71	for (int i=0; i < max_trip_count && keepgoing_out; ++i) {
72 72	/* Implicitly-defined code: bind actual parameter values
73 73	to formal parameter variables of loop-body */
74 74	bool keepgoing_in = keepgoing_out;
75 75	bool b_in = b_out;
76 76	/* User-defined code (loop body) */
77 77	int my_local = a + b_in; // Reading value "a" from the enclosing scope is fine
78 78	b_out = a - b_in;
79 79	keepgoing_out = my_local > b_out;
80 80	user_defined_val = b_in + b_in; // b_in and b_out are different variables
81 81	/* End user-defined code */
82 82	/* Implicitly defined-code */
83 83	user_defined_vals[i] = user_defined_val // accumulate scan-output values
84 84	}
85 85	// int t = my_local; // Can't do this. my_local is not accessible here.
86 86	// The values below are bound to the output variables of the loop and therefore accessible
87 87	// b_out; user_defined_vals; keepgoing_out;
88 88	}
89 89	There are several things of note in this code snippet:
90 90	1) Values from the enclosing scope (i.e. variable "a" here) are in scope and can
91 91	be referenced in the inputs of the loop.
92 92	2) Any values computed in the loop body that needs to be used in a subsequent
93 93	iteration or after the loop are modelled using a pair of variables in the loop-body,
94 94	consisting of an input variable (eg., b_in) and an output variable (eg., b_out).
95 95	These are referred to as loop-carried dependences. The loop operation node
96 96	supplies the input value of the input variable for the first iteration, and
97 97	returns the output value of the output variable produced by the final
98 98	iteration.
99 99	3) Scan_output variables are used to implicitly concatenate values computed across
100 100	all the iterations. In the above example, the value of user_defined_val computed
101 101	over all iterations are concatenated and returned as the value of user_defined_vals
102 102	after the loop.
103 103	4) Values created in the body cannot be accessed in the enclosing scope,
104 104	except using the mechanism described above.
105 105	Note that the semantics of this op support "diagonal" or "wavefront" execution.
106 106	(See Step 3 here for an example:
107 107	https://devblogs.nvidia.com/optimizing-recurrent-neural-networks-cudnn-5/).
108 108	Frontends should emit multi-layer RNNs as a series of While operators (with
109 109	time being the inner looping dimension), with each successive layer consuming
110 110	the scan_outputs from the previous layer, possibly going through several
111 111	point-wise operators (e.g. dropout, residual connections, linear layer).
112 112	The input/output of subgraph (produced by loop node) matching is based on order instead of name. The implementation will figure out the names based on this order.
113 113	**Attributes**
114 114	* **body** (required):
115 115	The graph run each iteration. It has 2+N inputs: (iteration_num,
116 116	condition, loop carried dependencies...). It has 1+N+K outputs:
117 117	(condition, loop carried dependencies..., scan_outputs...). Each
118 118	scan_output is created by concatenating the value of the specified
119 119	output value at the end of each iteration of the loop. It is an
120 120	error if the dimensions or data type of these scan_outputs change
121 121	across loop iterations.
122 122	**Inputs**
123 123	Between 2 and 2147483647 inputs.
124 124	* **M** (optional, heterogeneous) - **I**:
125 125	A maximum trip-count for the loop specified at runtime. Optional.
126 126	Pass empty string to skip.
127 127	* **cond** (optional, heterogeneous) - **B**:
128 128	A boolean termination condition. Optional. Pass empty string to
129 129	skip.
130 130	* **v_initial** (variadic) - **V**:
131 131	The initial values of any loop-carried dependencies (values that
132 132	change across loop iterations)
133 133	**Outputs**
134 134	Between 1 and 2147483647 outputs.
135 135	* **v_final_and_scan_outputs** (variadic) - **V**:
136 136	Final N loop carried dependency values then K scan_outputs. Scan
137 137	outputs must be Tensors.
138 138	**Type Constraints**
139 139	* **V** in (
140	+ optional(seq(tensor(bfloat16))),
141	+ optional(seq(tensor(bool))),
142	+ optional(seq(tensor(complex128))),
143	+ optional(seq(tensor(complex64))),
144	+ optional(seq(tensor(double))),
145	+ optional(seq(tensor(float))),
146	+ optional(seq(tensor(float16))),
147	+ optional(seq(tensor(int16))),
148	+ optional(seq(tensor(int32))),
149	+ optional(seq(tensor(int64))),
150	+ optional(seq(tensor(int8))),
151	+ optional(seq(tensor(string))),
152	+ optional(seq(tensor(uint16))),
153	+ optional(seq(tensor(uint32))),
154	+ optional(seq(tensor(uint64))),
155	+ optional(seq(tensor(uint8))),
156	+ optional(tensor(bfloat16)),
157	+ optional(tensor(bool)),
158	+ optional(tensor(complex128)),
159	+ optional(tensor(complex64)),
160	+ optional(tensor(double)),
161	+ optional(tensor(float)),
162	+ optional(tensor(float16)),
163	+ optional(tensor(int16)),
164	+ optional(tensor(int32)),
165	+ optional(tensor(int64)),
166	+ optional(tensor(int8)),
167	+ optional(tensor(string)),
168	+ optional(tensor(uint16)),
169	+ optional(tensor(uint32)),
170	+ optional(tensor(uint64)),
171	+ optional(tensor(uint8)),
172	+ seq(tensor(bfloat16)),
140 173	seq(tensor(bool)),
141 174	seq(tensor(complex128)),
142 175	seq(tensor(complex64)),
143 176	seq(tensor(double)),
144 177	seq(tensor(float)),
145 178	seq(tensor(float16)),
146 179	seq(tensor(int16)),
147 180	seq(tensor(int32)),
148 181	seq(tensor(int64)),
149 182	seq(tensor(int8)),
150 183	seq(tensor(string)),
151 184	seq(tensor(uint16)),
152 185	seq(tensor(uint32)),
153 186	seq(tensor(uint64)),
154 187	seq(tensor(uint8)),
188	+ tensor(bfloat16),
155 189	tensor(bool),
156 190	tensor(complex128),
157 191	tensor(complex64),
158 192	tensor(double),
159 193	tensor(float),
160 194	tensor(float16),
161 195	tensor(int16),
162 196	tensor(int32),
163 197	tensor(int64),
164 198	tensor(int8),
165 199	tensor(string),
166 200	tensor(uint16),
167 201	tensor(uint32),
168 202	tensor(uint64),
169 203	tensor(uint8)
170 204	):
205	+ All Tensor, Sequence(Tensor), Optional(Tensor), and
171	- All Tensor and Sequence types
206	+ Optional(Sequence(Tensor)) types
172 207	* **I** in (
173 208	tensor(int64)
174 209	):
175 210	tensor of int64, which should be a scalar.
176 211	* **B** in (
177 212	tensor(bool)
178 213	):
179 214	tensor of bool, which should be a scalar.