Optimized neural network input stride method and apparatus

1. A system to form at least one output feature map (OFM) from at least one input feature map (IFM) at a convolutional layer in a neural network, the system comprising:

a weight-kernel subdivider to subdivide each original weight kernel of a plurality of sets of original weight kernels based on a predetermined horizontal input stride and a predetermined vertical input stride to form a set of a plurality of sub-kernels for each set of original weight kernels, the plurality of sets of original weight kernels corresponding to a current convolutional layer of the neural network, the predetermined horizontal input stride being greater than 1, and the predetermined vertical input stride being greater than 1;

an IFM subdivider to subdivide each of a plurality of IFMs based on the predetermined horizontal input stride and the predetermined vertical input stride to form a plurality of sub-maps for each IFM of the plurality of IFMs, the plurality of IFMs corresponding to the current convolutional layer, and each sub-map corresponding to a sub-kernel in each set of the plurality of sub-kernels; and

a convolver coupled to the weight-kernel subdivider and the IFM subdivider to convolve each sub-map by the corresponding sub-kernel for a set of original weight kernels using an input stride of 1 for each set of original weight kernels, the convolver to further sum a convolved result of each sub-map and the corresponding sub-kernel to form an OFM corresponding to the set of original weight kernels for each set of original weight kernels.

2. The system of claim 1 , wherein a weight in a sub-kernel comprises a weight in the original weight kernel from which the sub-kernel was subdivided based on a modulo of the predetermined horizontal input stride and based on a modulo of the predetermined vertical input stride of a position of the weight in the original weight kernel.

3. The system of claim 2 , wherein each weight that is less than a predetermined first value has been removed from each original weight kernel.

4. The system of claim 1 , wherein the predetermined horizontal input stride is 2, and the predetermined vertical input stride is 2.

5. The system of claim 1 , wherein the predetermined horizontal input stride is 3, and the predetermined vertical input stride is 3.

6. The system of claim 1 , wherein the predetermined horizontal input stride and the predetermined vertical input stride are equal to a predetermined second value,

wherein the weight-kernel subdivider to further subdivide each original weight kernel of a plurality of sets of original weight kernels by a square of the predetermined second value to form the set of a plurality of sub-kernels for each set of original weight kernels, and

wherein the IFM subdivider to further subdivide each of the plurality of IFMs based on the square of the predetermined second value to form the plurality of sub-maps for each IFM of the plurality of IFMs.

7. The system of claim 1 , further comprising:

a domain converter coupled to the weight-kernel subdivider and the IFM subdivider to convert elements of each sub-map into a Winograd domain,

wherein the convolver is coupled to an output of the domain converter and convolves each sub-map by the corresponding sub-kernel in the Winograd domain for a set of original weight kernels using a horizontal input stride of 1 and a vertical input stride of 1 for each set of original weight kernels.

8. A system to form at least one output feature map (OFM) from at least one input feature map (IFM) at a convolutional layer in a neural network, the system comprising:

a weight-kernel subdivider to increase a dimensionality of each set of a plurality of sets of original weight kernels by subdividing each original weight kernel based on a predetermined horizontal input stride and a predetermined vertical input stride to form a plurality of sub-kernels for each set of original weight kernels, the plurality of sets of original weight kernels corresponding to a current convolutional layer of the neural network, a dimension of each sub-kernel being less than a dimension of the original weight kernel from which the sub-kernel was subdivided, the predetermined horizontal input stride being greater than 1, and the predetermined vertical input stride being greater than 1;

an IFM subdivider to increase a dimensionality of a plurality of IFMs by subdividing each IFM of the plurality of IFMs based on the predetermined horizontal input stride and the predetermined vertical input stride to form a plurality of sub-maps for each IFM, the plurality of IFMs corresponding to the current convolutional layer, each sub-map corresponding to a sub-kernel in each set of the plurality of sub-kernels, and a dimension of each sub-map being less than a dimension of the IFM was subdivided; and

a convolver coupled to the weight-kernel subdivider and the IFM subdivider to convolve each sub-map by the corresponding sub-kernel for a set of original weight kernels using an input stride of 1 for each set of original weight kernels, the convolver to further sum a convolved result of each sub-map and the corresponding sub-kernel to form an OFM corresponding to the set of original weight kernels for each set of original weight kernels.

9. The system of claim 8 , wherein a weight in a sub-kernel comprises a weight in the original weight kernel from which the sub-kernel was subdivided based on a modulo of the predetermined input horizontal stride and the predetermined vertical input stride of a position of the weight in the original weight kernel.

10. The system of claim 9 , wherein each weight that is less than a predetermined first value has been removed from each original weight kernel.

11. The system of claim 8 , wherein the predetermined horizontal input stride is 2, and

wherein the predetermined vertical input stride is 2.

12. The system of claim 8 , wherein the predetermined horizontal input stride is 3, and

wherein the predetermined vertical input stride is 3.

13. The system of claim 8 , wherein the predetermined horizontal input stride and the predetermined vertical input stride are equal to a predetermined second value,

wherein the weight-kernel subdivider to further subdivide each original weight kernel of a plurality of sets of original weight kernels by a square of the predetermined second value to form the set of a plurality of sub-kernels for each set of original weight kernels, and

wherein the IFM subdivider to further subdivide each of the plurality of IFMs based on the square of the predetermined second value to form the plurality of sub-maps for each IFM of the plurality of IFMs.

14. The system of claim 8 , further comprising:

a domain converter coupled to the weight-kernel subdivider and the IFM subdivider to convert elements of each sub-map into a Winograd domain,

wherein the convolver is coupled to an output of the domain converter and convolves each sub-map by the corresponding sub-kernel in the Winograd domain for a set of original weight kernels using a horizontal input stride of 1 and a vertical input stride of 1 for each set of original weight kernels.

15. A method to form at least one output feature map (OFM) from at least one input feature map (IFM) at a convolutional layer in a neural network, the method comprising:

subdividing each original weight kernel of a plurality of sets of original weight kernels based on a predetermined horizontal input stride and a predetermined vertical input stride to form a set of a plurality of sub-kernels for each original weight kernel in a set of original weight kernels, the plurality of sets of original weight kernels corresponding to a current convolutional layer of the neural network, the predetermined horizontal input stride being greater than 1, and the predetermined vertical input stride being greater than 1;

subdividing each of a plurality of IFMs based on the predetermined horizontal input stride and the predetermined vertical input stride to form a plurality of sub-maps for each IFM of the plurality of IFMs, the plurality of IFMs corresponding to the current convolutional layer, and each sub-map corresponding to a sub-kernel in each set of the plurality of sub-kernels;

for each set of original weight kernels, convolving each sub-map by the corresponding sub-kernel using a horizontal input stride of 1 and a vertical input stride of 1; and

for each set of original weight kernels, summing a convolved result of each sub-map and the corresponding sub-kernel to form an OFM corresponding to the set of original weight kernels.

16. The method of claim 15 , wherein a weight in a sub-kernel comprises a weight in the original weight kernel from which the sub-kernel was subdivided based on a modulo of the predetermined horizontal input stride and on a modulo of the predetermined vertical input stride of a position of the weight in the original weight kernel.

17. The method of claim 15 , wherein the predetermined horizontal input stride is 2, and

wherein the predetermined vertical input stride is 2.

18. The method of claim 15 , wherein the predetermined horizontal input stride is 3, and

wherein the predetermined vertical input stride is 3.

19. The method of claim 15 , wherein the predetermined horizontal input stride and the predetermined vertical input stride are equal to a predetermined first value,

wherein subdividing each original weight kernel of a plurality of sets of original weight kernels comprises subdividing each original weight kernel of a plurality of sets of original weight kernels by a square of the predetermined first value to form the set of a plurality of sub-kernels for each set of original weight kernels, and

wherein subdividing each of a plurality of IFMs comprises subdividing each of the plurality of IFMs based on the square of the predetermined first value to form the plurality of sub-maps for each IFM of the plurality of IFMs.

20. The method of claim 15 , further comprising converting elements of each sub-map into a Winograd domain before convolving each sub-map.