Processing core with data associative adaptive rounding
Processing cores with data associative adaptive rounding and associated methods are disclosed herein. One disclosed processing core comprises an arithmetic logic unit cluster configured to generate a value for a unit of directed graph data using input directed graph data, a comparator coupled to a threshold register and a data register, a core controller configured to load a threshold value into the threshold register when the value for the unit of directed graph data is loaded into the data register, and a rounding circuit. The rounding circuit is configured to receive the value for the unit of directed graph data from the arithmetic logic unit cluster and conditionally round the value for the unit of directed graph data based on a comparator output from the comparator.
1. A method, in which each step is conducted by a processing core, comprising:
executing, using a core controller on the processing core, a directed graph;
associating a value with a unit of directed graph data on the processing core;
generating, while executing the directed graph and using a computational unit in the processing core, a value for the unit of directed graph data;
calculating, while executing the directed graph and using the core controller, a threshold value, wherein the calculating of the threshold value uses the value associated with the unit of directed graph data;
conducting, using a comparator on the processing core, a comparison of the value for the unit of directed graph data and the threshold value; and
conditionally rounding, using a rounding circuit on the processing core and based on the comparison, the value for the unit of directed graph data.
2. The method of claim 1 ,
wherein: the calculating of the threshold value uses the value for the unit of directed graph data.
3. The method of claim 2 , wherein:
the directed graph is of an artificial neural network; and
the value for the unit of directed graph data is an activation value.
4. The method of claim 1 ,
wherein: (i) the generating of the value for the unit of directed graph data uses a value for a unit of input directed graph data; and (ii) the calculating of the threshold value uses the value associated with the unit of directed graph data and the value for the unit of input directed graph data.
5. The method of claim 4 , wherein:
the directed graph is of an artificial neural network; and
the value for the unit of directed graph data is an activation value.
6. The method of claim 1 , further comprising:
generating, while executing the directed graph, a value for a second unit of directed graph data;
calculating, while executing the directed graph, a second threshold value;
conducting a second comparison of the value for the second unit of directed graph data and the second threshold value; and
conditionally rounding, based on the second comparison, the value for the second unit of directed graph data.
7. The method of claim 1 , wherein:
the directed graph is of an artificial neural network; and
the value for the unit of directed graph data is an activation value.
8. The method of claim 7 , further comprising:
determining the value associated with the unit of directed graph data using the artificial neural network and test inputs.
9. A processing core, comprising:
a core controller configured to control an execution of a directed graph and a calculation of a threshold value while the directed graph is being executed;
a computational unit configured to generate a value for a unit of directed graph data using input directed graph data during the execution of the directed graph;
a comparator configured to generate a comparator output based on: (i) the threshold value; and (ii) the value for the unit of directed graph data; and
a rounding circuit configured to conditionally round the value for the unit of directed graph data based on the comparator output from the comparator;
wherein: (i) the directed graph is of an artificial neural network; and (ii) the value for the unit of directed graph data is an activation value.
10. The processing core of claim 9 , further comprising:
an association between the threshold value and the value for the unit of directed graph data;
wherein the core controller uses the association to provide the threshold value, along with the value for the unit of directed graph data, to the comparator.
11. The processing core of claim 9 , further comprising:
an association between a value and the value for the unit of directed graph data;
wherein the calculation of the threshold value uses the value.
12. The processing core of claim 9 , further comprising:
an association between a value and the value for the unit of directed graph data;
wherein the calculation of the threshold value uses the value and the value for the unit of directed graph data.
13. The processing core of claim 9 , further comprising:
an association between a value and the value for the unit of directed graph data;
wherein: (i) generating the value for the unit of directed graph data uses a value for a unit of input directed graph data; and (ii) calculating the threshold value uses the value and the value for the unit of input directed graph data.
14. The processing core of claim 9 , wherein:
the computational unit is configured to generate, while executing the directed graph, a value for a second unit of directed graph data;
the core controller is configured to control the calculation of a second threshold value, while executing the directed graph;
the comparator is configured to conduct a second comparison of the value for the second unit of directed graph data and the second threshold value; and
the rounding circuit is configured to conditionally round, based on the second comparison, the value for the second unit of directed graph data.
15. The processing core of claim 9 , wherein:
the core controller is configured to determine the activation value using the artificial neural network and test inputs.
16. A processing core, comprising:
a core controller configured to control an execution of a directed graph and a calculation of a threshold value while the directed graph is being executed;
a computational unit configured to generate a value for a unit of directed graph data using a value for a unit of input directed graph data;
an association between a value and the value for the unit of directed graph data, wherein the processing core is configured to calculate the threshold value using the value;
a comparator configured to generate a comparator output based on the threshold value and the value for the unit of directed graph data; and
a rounding circuit configured to conditionally round the value for the unit of directed graph data based on the comparator output from the comparator.
17. A method, in which each step is conducted by a processing core, comprising:
executing, using a core controller on the processing core, a directed graph;
associating a value with a unit of directed graph data on the processing core;
generating, while executing the directed graph and using a computational unit in the processing core, a value for the unit of directed graph data, wherein the generating of the value for the unit of directed graph data uses a value for a unit of input directed graph data;
calculating, while executing the directed graph and using the core controller, a threshold value, wherein the calculating of the threshold value uses the value associated with the unit of directed graph data and the value for the unit of input directed graph data;
conducting, using a comparator on the processing core, a comparison of the value for the unit of directed graph data and the threshold value; and
conditionally rounding, using a rounding circuit on the processing core and based on the comparison, the value for the unit of directed graph data.
18. A method, in which each step is conducted by a processing core, comprising:
executing, using a core controller on the processing core, a directed graph;
associating a value with a unit of directed graph data on the processing core;
generating, while executing the directed graph and using a computational unit in the processing core, a value for the unit of directed graph data;
calculating, while executing the directed graph and using the core controller, a threshold value, wherein the calculating of the threshold value uses the value associated with the unit of directed graph data;
conducting, using a comparator on the processing core, a comparison of the value for the unit of directed graph data and the threshold value;
conditionally rounding, using a rounding circuit on the processing core and based on the comparison, the value for the unit of directed graph data;
generating, while executing the directed graph a value for a second unit of directed graph data;
calculating, while executing the directed graph, a second threshold value;
conducting a second comparison of the value for the second unit of directed graph data and the second threshold value; and
conditionally rounding, based on the second comparison, the value for the second unit of directed graph data.
19. A method, in which each step is conducted by a processing core, comprising:
executing, using a core controller on the processing core, a directed graph;
associating a value with a unit of directed graph data on the processing core;
generating, while executing the directed graph and using a computational unit in the processing core, a value for the unit of directed graph data;
calculating, while executing the directed graph and using the core controller, a threshold value, wherein the calculating of the threshold value uses the value associated with the unit of directed graph data;
conducting, using a comparator on the processing core, a comparison of the value for the unit of directed graph data and the threshold value; and
conditionally rounding, using a rounding circuit on the processing core and based on the comparison, the value for the unit of directed graph data;
wherein: (i) the directed graph is of an artificial neural network; and (ii) the value for the unit of directed graph data is an activation value.
20. The method of claim 19 , wherein the calculating of the threshold value uses the value for the unit of directed graph data.
21. A processing core, comprising:
a core controller configured to control an execution of a directed graph and a calculation of a threshold value while the directed graph is being executed;
an association between the threshold value and a value for a unit of directed graph data;
a computational unit configured to generate the value for the unit of directed graph data using input directed graph data during the execution of the directed graph;
a comparator configured to generate a comparator output based on: (i) the threshold value; and (ii) the value for the unit of directed graph data; and
a rounding circuit configured to conditionally round the value for the unit of directed graph data based on the comparator output from the comparator;
wherein the core controller uses the association to provide the threshold value, along with the value for the unit of directed graph data, to the comparator.
22. A processing core, comprising:
a core controller configured to control an execution of a directed graph and a calculation of a threshold value while the directed graph is being executed;
a computational unit configured to generate a value for a unit of directed graph data using input directed graph data during the execution of the directed graph;
an association between a value and the value for the unit of directed graph data, wherein the calculation of the threshold value uses the value and the value for the unit of directed graph data;
a comparator configured to generate a comparator output based on: (i) the threshold value; and (ii) the value for the unit of directed graph data; and
a rounding circuit configured to conditionally round the value for the unit of directed graph data based on the comparator output from the comparator.
23. The processing core of claim 22 , wherein:
the directed graph is of an artificial neural network; and
the value for the unit of directed graph data is an activation value.
24. The processing core of claim 22 , wherein:
generating the value for the unit of directed graph data uses a value for a unit of input directed graph data; and
calculating the threshold value uses the value and the value for the unit of input directed graph data.