Systems and methods for image processing in a deep convolution network

1. A method performed by an electronic device, comprising:

interleaving a plurality of input image channels to produce an interleaved multi-channel input;

loading the interleaved multi-channel input to a single-instruction multiple data (SIMD) processor;

interleaving multiple filter channels to produce a multi-channel filter; and

convolving the interleaved multi-channel input with the multi-channel filter.

2. The method of claim 1 , further comprising performing object recognition based on the convolution.

3. The method of claim 1 , further comprising:

storing a maximum number of filter coefficients of a filter kernel in registers of the SIMD processor; and

applying the maximum number of filter coefficients in convolution operations without reloading any of the maximum number of filter coefficients from a cache.

4. The method of claim 1 , further comprising, for each input element, loading the element only once for multiple output elements.

5. The method of claim 1 , wherein convolving the interleaved multi-channel input with the multi-channel filter comprises multiplying an image input in parallel with multiple filter coefficients and accumulating sums in parallel.

6. The method of claim 1 , wherein the convolution is Deep Convolution Network (DCN) convolution.

7. The method of claim 1 , further comprising storing a kernel of the multi-channel filter in registers.

8. The method of claim 1 , further comprising storing the interleaved multi-channel input in at least one of a level 1 (L1) cache and a level 2 (L2) cache.

9. The method of claim 1 , further comprising storing an output in a level 1 (L1) cache.

10. An electronic device, comprising:

a single-instruction multiple data (SIMD) processor configured

to interleave a plurality of input image channels to produce an interleaved multi-channel input,

to load the interleaved multi-channel input to the SIMD processor,

to interleave multiple filter channels to produce a multi-channel filter, and

to convolve the interleaved multi-channel input with the multi-channel filter.

11. The electronic device of claim 10 , wherein the SIMD processor is configured to perform object recognition based on the convolution.

12. The electronic device of claim 10 , wherein the SIMD processor is configured to:

store a maximum number of filter coefficients of a filter kernel in registers of the SIMD processor; and

apply the maximum number of filter coefficients in convolution operations without reloading any of the maximum number of filter coefficients from a cache.

13. The electronic device of claim 10 , wherein the SIMD processor is configured to, for each input element, load the element only once for multiple output elements.

14. The electronic device of claim 10 , wherein the SIMD processor is configured to multiply an image input in parallel with multiple filter coefficients and to accumulate sums in parallel.

15. The electronic device of claim 10 , wherein the SIMD processor is configured to perform Deep Convolution Network (DCN) convolution.

16. The electronic device of claim 10 , wherein the SIMD processor is configured to store a kernel of the multi-channel filter in registers.

17. The electronic device of claim 10 , further comprising a memory, wherein the memory is configured to store the interleaved multi-channel input in at least one of a level 1 (L1) cache and a level 2 (L2) cache.

18. The electronic device of claim 10 , further comprising a memory, wherein the memory is configured to store an output in a level 1 (L1) cache.

19. An apparatus, comprising:

means for interleaving a plurality of input image channels to produce an interleaved multi-channel input;

means for loading the interleaved multi-channel input to a parallel processing means;

means for interleaving multiple filter channels to produce a multi-channel filter; and

means for convolving the interleaved multi-channel input with the multi-channel filter.

20. The apparatus of claim 19 , further comprising:

means for storing a maximum number of filter coefficients of a filter kernel in registers of the parallel processing means; and

means for applying the maximum number of filter coefficients in convolution operations without reloading any of the maximum number of filter coefficients from a cache.

21. The apparatus of claim 19 , further comprising means for, for each input element, loading the element only once for multiple output elements.

22. The apparatus of claim 19 , wherein the means for convolving comprises means for performing Deep Convolution Network (DCN) convolution.

23. A computer-program product, comprising a non-transitory computer-readable medium having instructions thereon, the instructions comprising:

code for causing an electronic device to interleave a plurality of input image channels to produce an interleaved multi-channel input;

code for causing the electronic device to load the interleaved multi-channel input to a single-instruction multiple data (SIMD) processor;

code for causing the electronic device to interleave multiple filter channels to produce a multi-channel filter; and

code for causing the electronic device to convolve the interleaved multi-channel input with a multi-channel filter.

24. The computer-program product of claim 23 , further comprising:

code for causing the electronic device to store a maximum number of filter coefficients of a filter kernel in registers of the SIMD processor; and

code for causing the electronic device to apply the maximum number of filter coefficients in convolution operations without reloading any of the maximum number of filter coefficients from a cache.

25. The computer-program product of claim 23 , further comprising code for causing the electronic device to, for each input element, load the element only once for multiple output elements.

26. The computer-program product of claim 23 , wherein the code for causing the electronic device to convolve comprises code for causing the electronic device to perform Deep Convolution Network (DCN) convolution.