Memory mapping of activations for convolutional neural network executions
A memory controller circuit for mapping data of a convolutional neural network to a physical memory is disclosed. The memory controller circuit comprises a receiving unit to receive a selection parameter value, and a mapping unit to map pixel values of one layer of the convolutional neural network to memory words of the physical memory according to one of a plurality of mapping schemas, wherein the mapping is dependent on the value of the received selection parameter value.
1. A memory controller circuit for mapping data of a convolutional neural network to a physical memory, said memory controller circuit comprising:
a receiving unit to receive a selection parameter value; and
a mapping unit to map pixel values of one layer of said convolutional neural network to memory words of said physical memory according to one of a plurality of mapping schemas, wherein said mapping is dependent on said value of said received selection parameter value, and wherein the one of the plurality of mapping schemas is an intra-word adjacent mapping schema, and wherein bit values of portions of different pixel values are mapped to one physical memory word in a contiguous manner.
2. The memory controller circuit according to claim 1 , wherein the one of the plurality of mapping schemas is a kernel word interleaving schema, wherein bit values of portions of different pixel values are mapped to said physical memory such that at most a number of pixels equal to a size of a kernel of a convolutional neural network operation is mapped contiguously to adjacent memory words.
3. The memory controller circuit according to claim 1 , wherein the one of the plurality of mapping schemas is an activation word interleaving schema, wherein bit values of pixel values are mapped such that a memory word comprises the bit values of at most one pixel.
4. The memory controller circuit according to claim 1 , wherein said selection parameter value is comprised in a set of parameter values of a layer of said convolutional neural network.
5. The memory controller circuit according to claim 1 , wherein said selection parameter value is independent of said parameter values of a layer of said convolutional neural network.
6. The memory controller circuit according to claim 1 , wherein said physical memory is selected out of a group comprising a static random access memory, a latched based memory, and a dynamic random access memory.
7. The memory controller circuit according to claim 1 , wherein said memory controller circuit is integrated together with physical memory.
8. The memory controller circuit according to claim 1 , wherein said physical memory stores input values of a computational memory and is integrated with said computational memory.
9. The memory controller circuit according to claim 1 , wherein physical memory stores input values of a digital memory and is integrated with a digital processing unit.
10. A method for mapping data of a convolutional neural network to a physical memory, said method comprising:
receiving a selection parameter value; and
mapping pixel values of one layer of said convolutional neural network to memory words of said physical memory according to one of a plurality of mapping schemas, wherein said mapping is dependent on said value of said received selection parameter value, and wherein the one of the plurality of mapping schemas is an intra-word adjacent mapping schema, and wherein bit values of portions of different pixel values are mapped to one physical memory word in a contiguous manner.
11. The method according to claim 10 , wherein the one of the plurality of mapping schemas is a kernel word interleaving schema, wherein bit values of portions of different pixel values are mapped to said physical memory such that at most a number of pixels equal to a size of a kernel of a convolutional neural network operation is mapped contiguously to adjacent memory words.
12. The method according to claim 10 , wherein the one of the plurality of mapping schemas is an activation word interleaving schema, wherein bit values of pixel values are mapped such that a memory word comprises the bit values of at most one pixel.
13. The method according to claim 10 , wherein said selection parameter value is comprised in a set of parameter values of a layer of said convolutional neural network.
14. The method according to claim 10 , wherein said selection parameter value is independent of said parameter values of a layer of said convolutional neural network.
15. The method according to claim 10 , wherein said physical memory is selected out of a group comprising a static random access memory, a latched based memory, and a dynamic random access memory.
16. The method according to claim 10 , wherein said memory controller circuit is integrated together with physical memory.
17. The method according to claim 10 , wherein said physical memory stores input values of a computational memory and is integrated with said computational memory or with a digital processing unit.
18. A computer program product for mapping data of a convolutional neural network to a physical memory, said computer program product comprising a computer readable storage medium having program instructions embodied therewith, said program instructions being executable by one or more computing systems or controllers to cause said one or more computing systems to:
receive a selection parameter value; and
map pixel values of one layer of said convolutional neural network to memory words of said physical memory according to one of a plurality of mapping schemas, wherein said mapping is dependent on said value of said received selection parameter value, and wherein the one of the plurality of mapping schemas is an intra-word adjacent mapping schema, and wherein bit values of portions of different pixel values are mapped to one physical memory word in a contiguous manner.