Supporting floating point 16 (FP16) in dot product architecture

A multiplier unit of a processor, wherein:

• the processor multiplies integer values of elements of a first vector and corresponding elements of a second vector to form a product-integer value;
• the processor adds exponent values corresponding to the integer values of the corresponding elements of the two vectors to form an unbiased exponent value corresponding to the product-integer value;
• the multiplier unit includes a local shifter that forms a first shifted value by shifting the product-integer value by a number of bits in a predetermined direction based on a difference value between the unbiased exponent value corresponding to the product-integer value and a predetermined value being less than or equal to a predetermined maximum bit-shift capacity of the local shifter;
• and
• the processor further comprising an auxiliary shifter coupled to an adder tree and that forms a second shifted value by shifting a second output from the adder tree by the predetermined maximum bit-shift capacity of the local shifter, wherein the processor provides an inference of a deep neural network model based at least in part on the auxiliary shifter forming the second shifted value.

A method of a processing unit to compute dot-products for floating point values, the method comprising:

• element-wise multiplying, by the processing unit associated with an array of multiplier units, integer values of elements of a first vector by integer values of corresponding elements of a second vector to form integer-product values, the first vector comprising n elements of floating point values and the second vector comprising n elements of floating point values in which n is an integer greater than 1;
• element-wise adding, by the processing unit, exponent values of the elements of the first vector and exponent values of the corresponding elements of the second vector to form exponent sum values that respectively correspond to the integer-product values;
• determining a maximum exponent sum value of the exponent sum values;
• subtracting, by the processing unit, the maximum exponent sum value from each of the exponent sum values to form relative exponent values that respectively correspond to the integer-product values;
• right-bit shifting, by first local shifters of the array of multiplier units, first integer-product values by a corresponding relative exponent value to form integer-product values that are aligned with an integer-product value corresponding to the maximum exponent sum value, each first local shifter comprising a first predetermined maximum number of bit shifts that is less than a full bit range of exponent sum values of the first vector and of the second vector, and the first integer-product values corresponding to relative exponent values that are less than or equal to the first predetermined maximum number of bit shifts;
• right-bit shifting, by an auxiliary shifter, second integer-product values by a second predetermined number of bit shifts to form second integer-product values that are aligned with the integer-product value corresponding to the maximum exponent sum value, wherein the second predetermined number of bit shifts includes the first predetermined maximum number of bits shifts;
• and
• adding the integer-product values that are aligned with the integer-product value corresponding to the maximum exponent sum value to form a dot product of the first vector and the second vector, wherein the processing unit provides an inference of a deep neural network model based at least in part on the processing unit adding the integer-product values that are aligned with the integer-product value corresponding to the maximum exponent sum value.