Processing apparatus and method of processing add operation therein
A method of processing an add operation in a processing apparatus includes acquiring sub-operands from input operands each having an n-bit precision, acquiring intermediate addition results by performing add operations of sub-operands in parallel by using adders, bit-shifting each of the intermediate addition results such that the intermediate addition results correspond to original bit positions in the input operands, and outputting a final addition result of the add operations of the input operands based on the bit-shifted intermediate addition results.
1. A processor-implemented method of a processor including a plurality of adder circuitries, the method comprising:
generating a plurality of sub-operands, respectively corresponding to bit values of bit sections of a plurality of operands of an input feature map of a neural network, having a first pixel configuration, by dividing the plurality of operands into the bit sections each having a predetermined bit size, wherein each of the plurality of operands has an n-bit precision, and n is a natural number;
generating respective intermediate bit results by operating the plurality of adder circuitries in parallel through each of the plurality of adder circuitries being provided respective multiple sub-operands of the plurality of sub-operands;
generating an output feature map, having a second pixel configuration different from the first pixel configuration, by combining bit-shifted intermediate bit results that result from respective bit-shiftings of the respective intermediate bit results based on respective bit positions in the plurality of operands, such that the combined bit-shifted intermediate bit results correspond to a summation result for the plurality of operands; and
dependent on a performed determination that there is a zero-bit section, among the bit sections, in which a corresponding sub-operand has a zero-value:
in the operating of the plurality of adder circuitries in parallel, controlling a corresponding one adder circuitry, of the plurality of adder circuitries, allocated to the zero-bit section to not operate with respect to the corresponding sub-operand, thereby skipping performance of an add operation with respect to the corresponding sub-operand; and
controlling the corresponding one adder circuitry to operate with respect to an other operand, other than the plurality of operands, to decrease an idling of the corresponding one adder circuitry.
2. The method of claim 1 , wherein
the predetermined bit size is k bits,
the generating of the respective intermediate bit results comprises generating the respective intermediate bit results by using multi-input adder circuitries each having a k-bit precision performing add operations of the respective multiple sub-operands each having a k-bit precision, and
k is a natural number less than n.
3. The method of claim 2 , wherein
a total number of the plurality of operands is m,
each of the respective intermediate bit results comprises a result of executing an add operation of m sub-operands, each having the k-bit precision, acquired from each corresponding bit section in the m operands,
each of the respective intermediate bit results has a precision of (k+log 2m) bits, and
m is a natural number.
4. The method of claim 3 , wherein
each of the plurality of operands is divided into (n/k) bit sections, and
the respective intermediate bit results comprise intermediate bit results from an intermediate bit result of sub-operands of a first bit section of each of the plurality of operands to an intermediate bit result of sub-operands of (n/k)-th bit section of each of the plurality of operands.
5. The method of claim 4 , further comprising performing the determination that there is the zero-bit section by determining whether there is the zero-bit section in which the corresponding sub-operand has the zero-value from among the first bit section to the (n/k)-th bit section,
wherein the controlling of the corresponding one adder circuitry to not operate with respect to the corresponding sub-operand and to operate with respect to the other operand is performed for avoiding the corresponding one adder circuitry from being in an idling state during the generating of the respective intermediate bit results.
6. The method of claim 2 , further comprising performing the respective bit-shiftings of the respective intermediate bit results, including bit-shifting each of the respective intermediate bit results by integer-multiple-of-k bits to correspond to the respective bit positions in each of the plurality of operands.
7. A processing apparatus comprising:
one or more memories storing instructions; and
one or more processors configured to execute the instructions, wherein the execution of the instructions configures the one or more processors to:
generate a plurality of sub-operands, respectively corresponding to bit values of bit sections of a plurality of operands of an input feature map of a neural network, having a first pixel configuration, by dividing the plurality of operands into the bit sections each having a certain bit size, wherein each of the plurality of operands has an n-bit precision, and n is a natural number;
generate respective intermediate bit results by operating a plurality of adder circuitries in parallel through each of the plurality of adder circuitries being provided different respective multiple sub-operands of the plurality of sub-operands;
generate an output feature map, having a second pixel configuration different from the first pixel configuration, by combining bit-shifted intermediate bit results that result from respective bit-shiftings of the respective intermediate bit results based on respective bit positions in the plurality of operands, such that the combined bit-shifted intermediate bit results correspond to a summation result for the plurality of operands; and
dependent on a performed determination that there is a zero-bit section, among the bit sections, in which a corresponding sub-operand has a zero-value:
in the operating of the plurality of adder circuitries in parallel, control a corresponding one adder circuitry, of the plurality of adder circuitries, allocated to the zero-bit section to not operate with respect to the corresponding sub-operand, thereby skipping performance of an add operation with respect to the corresponding sub-operand; and
control the corresponding one adder circuitry to operate with respect to an other operand, other than the plurality of operands, to decrease an idling of the corresponding one adder circuitry.
8. The processing apparatus of claim 7 , wherein
the certain bit size is k bits,
the generation of the respective intermediate bit results comprises generation of the respective intermediate bit results by using multi-input adder circuitries each having a k-bit precision performing add operations of the different respective multiple sub-operands each having a k-bit precision, and
k is a natural number less than n.
9. The processing apparatus of claim 8 , wherein
a total number of the plurality of operands is m,
each of the respective intermediate bit results comprises a result of performing an add operation of m sub-operands each having the k-bit precision and obtained from corresponding bit sections in each of the m operands,
each of the respective intermediate bit results has a precision of (k+log 2m) bits, and
m is a natural number.
10. The processing apparatus of claim 9 , wherein
each of the plurality of operands is divided into (n/k) sub-operands, and
the respective intermediate bit results comprise
intermediate bit results from an intermediate bit result of sub-operands of a first bit section of each of the plurality of operands to an intermediate bit result of sub-operands of (n/k)-th bit section of each of the plurality of operands.
11. The processing apparatus of claim 10 , wherein the one or more processors is further configured to
perform the determination that there is the zero-bit section by determining whether there is the zero-bit section in which the corresponding sub-operand has the zero-value from among the first bit section to the (n/k)-th bit section, and
wherein the controlling of the corresponding one adder circuitry to not operate with respect to the corresponding sub-operand and to operate with respect to the other operand is performed for avoiding the corresponding one adder circuitry from being in an idling state during the generating of the respective intermediate bit results.