Method and apparatus with neural network operation
A neural network operation apparatus includes: a receiver configured to receive a first input feature map; a controller configured to control multiplier-accumulators (MACs) included in a first MAC array; and a first operation engine comprising the first MAC array and configured to process the first input feature map based on the MACs of which operation states are controlled.
1. A neural network operation apparatus, comprising:
a receiver configured to receive a first input feature map;
a controller configured to control multiplier-accumulators (MACs) included in a first MAC array by setting an operation state of one MAC in the first MAC array to an enable state and setting an operation state of remaining MACs in the first MAC array to a disable state; and
a first operation engine comprising the first MAC array and configured to process the first input feature map based on the MACs of which operation states are controlled.
2. The apparatus of claim 1 , wherein
for the processing, the first operation engine is configured to perform a depthwise convolution operation on the first input feature map.
3. The apparatus of claim 2 , wherein, for the receiving, the receiver is configured to receive a second input feature map.
4. The apparatus of claim 1 , further comprising:
a second operation engine comprising a second MAC array and configured to process a second input feature map based on the second MAC array.
5. The apparatus of claim 4 , wherein
for the controlling, the controller is configured to control the first operation engine and the second operation engine, and
the first operation engine and the second operation engine are configured to perform an elementwise sum operation of the first input feature map and the second input feature map by the first operation engine performing a multiplication and addition operation on the first input feature map and the second operation engine performing a multiplication and addition operation on the second input feature map, based on the controlling.
6. The apparatus of claim 5 , wherein
for the controlling, the controller is configured to:
set an operation state of one MAC in the first MAC array to an enable state and set an operation state of remaining MACs in the first MAC array to a disable state; and
set an operation state of one MAC in the second MAC array to an enable state and set an operation state of remaining MACs in the second MAC array to a disable state, and
the first operation engine and the second operation engine are configured to perform an elementwise sum operation of the first input feature map and the second input feature map, based on the MACs included in the first MAC array and the MACs included in the second MAC array of which operation states are controlled.
7. The apparatus of claim 1 , further comprising:
a memory including a weight buffer having a height of “1”, a width of “1”, and a predetermined length.
8. The apparatus of claim 7 , wherein a value of a weight of the weight buffer is “1”.
9. The apparatus of claim 1 , wherein operation results of the MACs included in the first MAC array are concatenated in an adder tree operation.
10. A neural network operation method, comprising:
receiving a first input feature map;
controlling operation states of multiplier-accumulators (MACs) included in a first MAC array by setting an operation state of one MAC in the first MAC array to an enable state and setting an operation state of remaining MACs in the first MAC array to a disable state; and
processing the first input feature map based on the MACs of which operation states are controlled.
11. The method of claim 10 , wherein
the processing comprises performing a depthwise convolution operation on the first input feature map.
12. The method of claim 10 , wherein the receiving comprises receiving a second input feature map.
13. The method of claim 10 , wherein
the controlling comprises controlling operation states of MACs included in a second MAC array, and
the processing comprises processing the second input feature map based on the MACs included in the second MAC array of which operation states are controlled.
14. The method of claim 13 , wherein the processing comprises performing an elementwise sum operation of the first input feature map and the second input feature map by the first MAC array performing a multiplication and addition operation on the first input feature map and the second MAC array performing a multiplication and addition operation on the second input feature map, based on the MACs included in the first MAC array and the MACs included in the second MAC array of which operation states are controlled.
15. The method of claim 14 , wherein the controlling of the MACs included in the first MAC array and the second MAC array comprises:
setting an operation state of one MAC in the first MAC array to an enable state;
setting an operation state of remaining MACs except for the one MAC in the first MAC array to a disable state;
setting an operation state of one MAC in the second MAC array to an enable state; and
setting an operation state of remaining MACs except for the one MAC in the second MAC array to a disable state.
16. The method of claim 10 , wherein a weight used for the processing of the first input feature map has a height of “1”, a width of “1”, and a predetermined length.
17. The method of claim 16 , wherein a value of the weight is “1”.
18. The method of claim 10 , wherein operation results of the MACs included in the first MAC array are concatenated in an adder tree operation.
19. A non-transitory computer-readable storage medium storing instructions that, when executed by a processor, configure the processor to perform the method of claim 10 .
20. A neural network operation apparatus, comprising:
a first multiplier-accumulator (MAC) array configured to perform a depthwise convolution operation on a first input feature map;
a second MAC array configured to perform the depthwise convolution operation on a second input feature map; and
a controller configured to, in each of the first MAC array and the second MAC array, set an operation state of one MAC to an enable state and set an operation state of remaining MACs except for the one MAC to a disable state,
wherein, based on the set operational states, the first MAC array and the second MAC array are configured to perform an elementwise sum operation of the first input feature map and the second input feature map by respectively performing the depthwise convolution operation.
21. The apparatus of claim 20 , wherein the depthwise convolution operation is performed based on weights, and a total number of the weights is equal to a total number of input channels of either of the first input feature map and the second input feature map.
22. The apparatus of claim 20 , wherein
the first input feature map and the second input feature map correspond to images, and
the first MAC array and the second MAC array are configured to generate an output feature map by the performing of the elementwise sum operation.