Methods for processing data in an efficient convolutional engine with partitioned columns of convolver units
Contiguous columns of a convolutional engine are partitioned into two or more groups. Each group of columns may be used to process input data. Filter weights assigned to one group may be distinct from filter weights assigned to another group.
1. A method for processing input data with a first input channel and a second input channel, the method comprising:
processing, by a first plurality of convolver units belonging to a first plurality of columns of a convolutional engine, the first input channel, the processing of the first input channel comprising computing a first plurality of dot products based on the first input channel and a first plurality of filter weights, and providing the first plurality of dot products to accumulators of the first plurality of convolver units;
processing, by a second plurality of convolver units belonging to a second plurality of columns of the convolutional engine disjoint from the first plurality of columns, the first input channel, the processing of the first input channel comprising computing a second plurality of dot products based on the first input channel and a second plurality of filter weights, and providing the second plurality of dot products to accumulators of the second plurality of convolver units;
processing, by the first plurality of convolver units, the second input channel, the processing of the second input channel comprising computing a third plurality of dot products based on the second input channel and a third plurality of filter weights, and summing the first plurality of dot products and the third plurality of dot products with the accumulators of the first plurality of convolver units; and
processing, by the second plurality of convolver units, the second input channel, the processing of the second input channel comprising computing a fourth plurality of dot products based on the second input channel and a fourth plurality of filter weights, and summing the second plurality of dot products and the fourth plurality of dot products with the accumulators of the second plurality of convolver units.
2. The method of claim 1 , wherein the first plurality of convolver units computes the first plurality of dot products while the second plurality of convolver units computes the second plurality of dot products.
3. The method of claim 1 , wherein the first plurality of convolver units computes the third plurality of dot products while the second plurality of convolver units computes the fourth plurality of dot products.
4. The method of claim 1 , wherein the first plurality of columns of the convolutional engine comprise contiguous columns of the convolutional engine.
5. The method of claim 1 , wherein the second plurality of columns of the convolutional engine comprise contiguous columns of the convolutional engine.
6. The method of claim 1 , wherein a right-most column of the first plurality of columns is adjacent to a left-most column of the second plurality of columns.