The technical field of this invention is multiply-accumulators.
A digital signal processor (DSP) instruction set typically include multiply-accumulate instructions which are execute in dedicated hardware. This multiply-accumulate hardware implements the function:
Acc<=Acc+(X*Y).
As defined above the multiply-accumulate instruction forms the product of two operands X and Y and adds their product to the value stored in an accumulator. The sum is stored in the accumulator overwriting the previous value.
It is desirable to have load or move instructions having this accumulation register (Acc) as a destination. Implementing this function typically employs a multiplexer between data from the multiply-accumulate result and the load/move data. The prior art places this multiplexer just before the Acc register. This places the multiplexer on the most critical path. This critical path is the data flow from the X/Y source registers, through the multiply and add operations to the Acc register.
The multiplexer placement creates a problem. By being in the critical path, the multiplexer slows operation of the multiply-accumulate function. The prior art placement of the multiplexer causes all operations to slow, even ordinary multiply-accumulation. This limits the clock rate that can be employed potentially slowing all data processor operation.
This invention is multiply-accumulate circuit supporting a load of the accumulator. During multiply-accumulate operation a partial product generator forms partial produces from the product inputs. An adder tree sums the partial product and the accumulator value. The sum is stored back in the accumulator overwriting the prior value.
During load operation an input gate forces one of the product inputs to all 0's. Thus the partial product generator generates partial products corresponding to a zero product. The adder tree adds this zero product to the external load value. The sum, which corresponds to the external load value is stored back in the accumulator overwriting the prior value.
A multiplexer at the side input of the adder tree selects the accumulator value for normal operation or the external load value for load operation. This placement of the multiplexer is out of the critical path during normal operation. Thus this circuit may operate faster than the prior art that includes the multiplexer between the adder tree and the accumulator.
These and other aspects of this invention are illustrated in the drawings, in which:
In accordance with the prior art the multiply-accumulate operation is performed in hardware by the combination of partial product generator 103 and adder tree 104. The data stored in input register 101 and input register 102 are supplied to respective inputs of partial product generator 103. Partial product generator 103 generates a set of partial products from the inputs X and Y. Adder tree 104 sums these partial products with the proper shifts to form the product. Adder tree 104 also receives the current data stored in accumulator 106. Adder tree 104 adds the current data from accumulator 106 to the product data by summing the accumulator data with the partial products.
The accumulated product output of adder tree 104 supplies one input of multiplexer 105. A second input of multiplexer 105 receives external source data. This external source data is used to initialize or load accumulator 106. The control input to multiplexer 105 determines the data selected. During ordinary multiply-accumulate operations, multiplexer 105 selects data from adder tree 104 for storage in accumulator 106. During load operations, multiplexer 105 selects data from the external source for storage in accumulator 106. The selected output of multiplexer 105 is stored in accumulator 106 overwriting any prior data.
The placement of multiplexer 105 illustrated in
A load operation employs AND gate 202 and multiplexer 207. AND gate 202 includes the same structure for each bit of the second operand Y. A first input of AND gate 202 receives the operand Yi. A second inverting input of AND gate 202 receives a control signal ZERO. During normal (multiply-accumulate) operation control signal ZERO is all 0's. Due to the inverting action of the inverting input, all bits Yi are passed unchanged to be stored in input register 203. During load operation control signal ZERO is all 1's. Thus all bits of operand Y are blocked and the output of AND gate 202 is all 0's. This zero input is stored in input register 203 and then passed to partial product generator 204. The resulting partial products resolve all 0's. This all zero set of partial products is supplied to adder tree 204.
Multiplexer 207 selects the second input to adder tree 205 as specified by the control input. During multiply-accumulate operation multiplexer 207 selects the current contents of accumulator 206. During load operation multiplexer 207 selects the external data source. Because the output of partial product generator 204 is always zero due AND gate 202 zeroing one of its inputs during load operations, the sum of adder tree 205 is the external data. Adder tree 206 supplies this external data of accumulator 206 for storage.
This invention puts the multiplexer in the path from the accumulator register to the adder. This path is less critical than the path from the X/Y registers. The data from accumulator 206 can be fed into adder tree 205 at a later point than the partial products from partial product generator 204. AND gate 202 forces the partial products to zero, so that tree adder 205 passes the external data to accumulator register 206. This removes the multiplexer gate delay from the critical path in the multiplier-accumulator. Accordingly, the DSP is permitted to operate at a higher clock frequency.
Those skilled in the art would realize that the AND gate need not be in the second operand path. This circuit would operate equally well with the AND gate in either operand path. Those skilled in the art would realize the recited inverting input to the AND gate depends upon the sense of the control signal ZERO. It is feasible to employ a non-inverting input to the AND gate if control signal ZERO was all 1's for multiply-accumulate operation and all 0's for load operation.
This application claims priority under 35 U.S.C. 119(e)(1) to U.S. Provisional Application No. 62/078,166 filed Nov. 11, 2014.
Number | Date | Country | |
---|---|---|---|
62078166 | Nov 2014 | US |