modeci_mdf.functions.onnx.loop
- modeci_mdf.functions.onnx.loop(*args, **kwargs)
Generic Looping construct. This loop has multiple termination conditions:
Trip count. Iteration count specified at runtime. Set by specifying the input M. Optional. Set to empty string to omit. Note that a static trip count (specified at graph construction time) can be specified by passing in a constant node for input M.
Loop termination condition. This is an input to the op that determines whether to run the first iteration and also a loop-carried dependency for the body graph. The body graph must yield a value for the condition variable, whether this input is provided or not.
This table summarizes the operating modes of this operator with equivalent C-style code:
Operator inputs defined as (max_trip_count, condition_var).
- input (“”, “”):
- for (int i=0; ; ++i) {
cond = … // Note this value is ignored, but is required in the body
}
- input (“”, cond) // Note this is analogous to a while loop
bool cond = …; for (int i=0; cond; ++i) {
cond = …;
}
- input (“”, 1) // Note this is analogous to a do-while loop
bool cond = true for (int i=0; cond; ++i) {
cond = …;
}
- input (trip_count, “”) // Note this is analogous to a for loop
int trip_count = … for (int i=0; i < trip_count; ++i) {
cond = …; // ignored
}
- input (trip_count, cond)
int trip_count = …; bool cond = …; for (int i=0; i < trip_count && cond; ++i) {
cond = …;
}
Sample usage - cond as well as trip count
- graph predict-net {
%a = Constant[value = <Scalar Tensor [3]>]() %b = Constant[value = <Scalar Tensor [6]>]() %keepgoing = Constant[value = <Scalar Tensor [1]>]() %max_trip_count = Constant[value = <Scalar Tensor [10]>]() %keepgoing_out, %b_out, %user_defined_vals = Loop[body = <graph body-net>](%max_trip_count, %keepgoing, %b) return
}
- graph body-net (
%i[INT32, scalar] // iteration number %keepgoing_in[BOOL, scalar] // incoming loop-termination-condition; not used %b_in[INT32, scalar] // incoming value of loop-carried-dependency b
- ) {
%my_local = Add(%a, %b_in) %b_out = Sub(%a, %b_in) // outgoing value of loop-carried-dependency b %keepgoing_out = Greater(%my_local, %b_out) // outgoing loop-termination-condition %user_defined_val = Add(%b_in, %b_in) // scan-output value to be accumulated return %keepgoing_out, %b_out, %user_defined_val
}
Sample equivalent C code
- {
/* User-defined code (enclosing scope) / int a = 3, b = 6; bool keepgoing = true; // Analogous to input cond / End user-defined code */
/* Implicitly-defined code / const int max_trip_count = 10; // Analogous to input M int user_defined_vals[]; // Imagine this is resizable / End implicitly-defined code / / initialize loop-carried variables and scan-output variables */ bool keepgoing_out = keepgoing int b_out = b
- for (int i=0; i < max_trip_count && keepgoing_out; ++i) {
- /* Implicitly-defined code: bind actual parameter values
to formal parameter variables of loop-body */
bool keepgoing_in = keepgoing_out; bool b_in = b_out;
/* User-defined code (loop body) / int my_local = a + b_in; // Reading value “a” from the enclosing scope is fine b_out = a - b_in; keepgoing_out = my_local > b_out; user_defined_val = b_in + b_in; // b_in and b_out are different variables / End user-defined code */
/* Implicitly defined-code */ user_defined_vals[i] = user_defined_val // accumulate scan-output values
} // int t = my_local; // Can’t do this. my_local is not accessible here.
// The values below are bound to the output variables of the loop and therefore accessible // b_out; user_defined_vals; keepgoing_out;
}
There are several things of note in this code snippet:
Values from the enclosing scope (i.e. variable “a” here) are in scope and can be referenced in the inputs of the loop.
Any values computed in the loop body that needs to be used in a subsequent iteration or after the loop are modelled using a pair of variables in the loop-body, consisting of an input variable (eg., b_in) and an output variable (eg., b_out). These are referred to as loop-carried dependences. The loop operation node supplies the input value of the input variable for the first iteration, and returns the output value of the output variable produced by the final iteration.
Scan_output variables are used to implicitly concatenate values computed across all the iterations. In the above example, the value of user_defined_val computed over all iterations are concatenated and returned as the value of user_defined_vals after the loop.
Values created in the body cannot be accessed in the enclosing scope, except using the mechanism described above.
Note that the semantics of this op support “diagonal” or “wavefront” execution. (See Step 3 here for an example: https://devblogs.nvidia.com/optimizing-recurrent-neural-networks-cudnn-5/). Frontends should emit multi-layer RNNs as a series of While operators (with time being the inner looping dimension), with each successive layer consuming the scan_outputs from the previous layer, possibly going through several point-wise operators (e.g. dropout, residual connections, linear layer).
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.