Method and system for balanced-weight sparse convolution processing

1. A computer-implemented method, comprising:

obtaining an input tensor and a plurality of filters, wherein the input tensor and the plurality of filters have a same number of channels;

segmenting the input tensor into a plurality of sub-tensors according to a number of available processors;

pruning each of the plurality of filters so that every predetermined number of channels comprise a fixed number of non-zero weights;

segmenting each of the plurality of filters into a plurality of the sub-filters, wherein each of the plurality of the sub-filters comprises the same number of non-zero weights;

respectively assigning the plurality of sub-tensors and the plurality of sub-filters to the available processors, wherein each of the plurality of sub-filters is represented in a compact memory layout storing the non-zero weights of the sub-filter as index-value pairs, each of the index-value pairs comprising a channel-dimension index, a width-dimension index, and a corresponding non-zero weight;

for each of the available processors, iterating each of the index-value pairs in the compact memory layout of the assigned plurality of sub-filters, and identifying a corresponding input value in the assigned sub-tensor at a location identified by the channel-dimension index and the width-dimension index of the index-value pair to perform a multiply-and-accumulate (MAC) operation to generate a partial sum;

reassigning the plurality of sub-filters to the available processors by rotating the plurality of sub-filters among the plurality of processors; and

accumulating a plurality of the partial sums generated from each of the plurality of processors to obtain an output tensor.

2. The method of claim 1 , wherein each of the plurality of sub-tensors comprises the same number of channels as the input tensor.

3. The method of claim 1 , further comprising:

pruning each of the plurality of filters so that every a predetermined number of channels of the filter comprise the same number of non-zero values.

4. The method of claim 1 , wherein the identifying a corresponding input value in the assigned sub-tensor to perform a MAC operation to generate a partial sum comprises:

performing a multiplication operation based on the identified input value and the corresponding non-zero weight in the index-value pair to obtain a dot product; and

accumulating the dot product with a corresponding partial sum to obtained a new partial sum.

5. The method of claim 1 , wherein each of the plurality of filters comprises a channel-dimension, a height dimension, and a width dimension, and

the segmenting each of the plurality of filters into one or more sub-filters comprises:

dividing the channel-dimension into a plurality of channel groups; and

segmenting, according to the plurality of channel groups, the filter into a plurality of sub-filters with each corresponding to one of the plurality of channel groups.

6. The method of claim 5 , wherein the segmenting the filter into a plurality of sub-filters according to the channel groups comprises:

segmenting the filter into a plurality of chunks according to the plurality of channel groups; and

segmenting each of the plurality of chunks along the height dimension to obtain the plurality of sub-filters, wherein each of plurality of sub-filters shares the width dimension with the filter.

7. The method of claim 5 , wherein a subset of the plurality of sub-tensors from a same horizontal plane of the input tensor are respectively assigned to a subset of the available processors, and

the assigning the plurality of sub-filters to the available processors comprises:

identifying a subset of the plurality of sub-filters that are from a same filter but correspond to different channel groups; and

respectively assigning the subset of the plurality of sub-filters to the subset of the available processors.

8. The method of claim 1 , wherein the rotating the plurality of sub-filters among the plurality of processors comprises:

reassigning the plurality of sub-filters to the available processors by rotating a sub-filter that was assigned to an i-th processor to an (i+1)-th processor.

9. The method of claim 1 , wherein the partial sum generated by performing the MAC operation is stored in an accumulation buffer of the processor that is accessible by neighboring processors.

10. The method of claim 1 , wherein the accumulating the plurality of the partial sums generated from each of the plurality of processors to obtain an output tensor comprises:

during a current convolution layer of a neural network, accumulating the plurality of the partial sums generated from each of the plurality of processors to obtain an activation sub-tensor;

in response to the current convolution layer not being a last convolution layer, storing the activation sub-tensor in the processor for a next convolution layer of the neural network; and

in response to the current convolution layer being the last convolution layer, aggregating a plurality of the activation sub-tensors from the available processors to obtain the output tensor.

11. The method of claim 1 , wherein the representing each of the plurality of sub-filters in a compact memory layout storing the non-zero weights of the sub-filter as index-value pairs comprises:

converting each of the index-value pairs into a same number of bits; and

aggregating a plurality of the same number of bits as a bit array to represent the sub-filter.

12. A system for balanced-weight sparse convolution processing, comprising one or more processors and one or more non-transitory computer-readable memories coupled to the one or more processors and configured with instructions executable by the one or more processors to cause the system to perform operations comprising:

obtaining an input tensor and a plurality of filters, wherein the input tensor and the plurality of filters have a same number of channels;

segmenting the input tensor into a plurality of sub-tensors according to a number of available processors;

pruning each of the plurality of filters so that every predetermined number of channels comprise a fixed number of non-zero weights;

segmenting each of the plurality of filters into a plurality of the sub-filters, wherein each of the plurality of the sub-filters comprises the same number of non-zero weights;

respectively assigning the plurality of sub-tensors and the plurality of sub-filters to the available processors, wherein each of the plurality of sub-filters is represented in a compact memory layout storing the non-zero weights of the sub-filter as index-value pairs, each of the index-value pairs comprising a channel-dimension index, a width-dimension index, and a corresponding non-zero weight;

for each of the available processors, iterating each of the index-value pairs in the compact memory layout of the assigned plurality of sub-filters, and identifying a corresponding input value in the assigned sub-tensor at a location identified by the channel-dimension index and the width-dimension index of the index-value pair to perform a multiply-and-accumulate (MAC) operation to generate a partial sum;

reassigning the plurality of sub-filters to the available processors by rotating the plurality of sub-filters among the plurality of processors; and

accumulating a plurality of the partial sums generated from each of the plurality of processors to obtain an output tensor.

13. The system of claim 12 , wherein each of the plurality of sub-tensors comprises the same number of channels as the input tensor.

14. The system of claim 12 , wherein each of the plurality of filters comprises a channel dimension, a height dimension, and a width dimension, and

the segmenting each of the plurality of filters into one or more sub-filters comprises:

dividing the channel dimension into a plurality of channel groups; and

segmenting, according to the plurality of channel groups, the filter into a plurality of sub-filters with each corresponding to one of the plurality of channel groups.

15. The system of claim 12 , wherein the identifying a corresponding input value in the assigned sub-tensor to perform a MAC operation to generate a partial sum comprises:

performing a multiplication operation based on the identified input value and the corresponding non-zero weight in the index-value pair to obtain a dot product; and

accumulating the dot product with a corresponding partial sum to obtained a new partial sum.

16. The system of claim 12 , wherein the rotating the plurality of sub-filters among the plurality of processors comprises:

reassigning the plurality of sub-filters to the available processors by rotating a sub-filter that was assigned to an i-th processor to an (i+1)-th processor.

17. The system of claim 12 , wherein the accumulating the plurality of the partial sums generated from each of the plurality of processors to obtain an output tensor comprises:

during a current convolution layer of a neural network, accumulating the plurality of the partial sums generated from each of the plurality of processors to obtain an activation sub-tensor;

in response to the current convolution layer not being a last convolution layer, storing the activation sub-tensor in the processor for a next convolution layer of the neural network; and

in response to the current convolution layer being the last convolution layer, aggregating a plurality of the activation sub-tensors from the available processors to obtain the output tensor.

18. A non-transitory computer-readable storage medium for balanced-weight sparse convolution processing, configured with instructions executable by one or more processors to cause the one or more processors to perform operations comprising:

obtaining an input tensor and a plurality of filters, wherein the input tensor and the plurality of filters have a same number of channels;

segmenting the input tensor into a plurality of sub-tensors according to a number of available processors;

pruning each of the plurality of filters so that every predetermined number of channels comprise a fixed number of non-zero weights;

segmenting each of the plurality of filters into a plurality of the sub-filters, wherein each of the plurality of the sub-filters comprises the same number of non-zero weights;

respectively assigning the plurality of sub-tensors and the plurality of sub-filters to the available processors, wherein each of the plurality of sub-filters is represented in a compact memory layout storing the non-zero weights of the sub-filter as index-value pairs, each of the index-value pairs comprising a channel-dimension index, a width-dimension index, and a corresponding non-zero weight;

for each of the available processors, iterating each of the index-value pairs in the compact memory layout of the assigned plurality of sub-filters, and identifying a corresponding input value in the assigned sub-tensor at a location identified by the channel-dimension index and the width-dimension index of the index-value pair to perform a multiply-and-accumulate (MAC) operation to generate a partial sum;

reassigning the plurality of sub-filters to the available processors by rotating the plurality of sub-filters among the plurality of processors; and

accumulating a plurality of the partial sums generated from each of the plurality of processors to obtain an output tensor.

19. The storage medium of claim 18 , wherein the identifying a corresponding input value in the assigned sub-tensor to perform a MAC operation to generate a partial sum comprises:

performing a multiplication operation based on the identified input value and the corresponding non-zero weight in the index-value pair to obtain a dot product; and

accumulating the dot product with a corresponding partial sum to obtained a new partial sum.

20. The storage medium of claim 18 , wherein the accumulating the plurality of the partial sums generated from each of the plurality of processors to obtain an output tensor comprises:

during a current convolution layer of a neural network, accumulating the plurality of the partial sums generated from each of the plurality of processors to obtain an activation sub-tensor;

in response to the current convolution layer not being a last convolution layer, storing the activation sub-tensor in the processor for a next convolution layer of the neural network; and

in response to the current convolution layer being the last convolution layer, aggregating a plurality of the activation sub-tensors from the available processors to obtain the output tensor.