Processor reduction unit for accumulation of multiple operands with or without saturation

1. A multithreaded processor comprising:

a plurality of arithmetic units;

an accumulator unit; and

a reduction unit coupled between the plurality of arithmetic units and the accumulator unit, the reduction unit being configured to receive input operands from the arithmetic units and a first accumulator value from the accumulator unit;

wherein the reduction unit is pipelined and operative to sum the input operands and the first accumulator value, and to generate a second accumulator value for delivery to the accumulator unit, the reduction unit further comprising m inputs, m adders, and an m stage pipeline, where m is greater than or equal to two, each of the in inputs being coupled to a corresponding adder by means of N−1 pipeline registers where N is a stage number greater than or equal to 1, the m stage pipeline being configured to reduce the worst case delay of the reduction unit;

wherein the reduction unit is controllable to support saturation and wrap-around arithmetic; and

wherein operations for a dot product computed for a given thread are executed concurrently with operations from other threads, the number of cycles between execution of instructions from the given thread being greater than or equal to a number of pipeline stages in the reduction unit plus any additional cycles needed to write to and read from the accumulator unit.

2. The processor of claim 1 wherein the plurality of arithmetic units comprises a plurality of multipliers arranged in parallel with one another.

3. The processor of claim 2 wherein the plurality of multipliers comprises a plurality of multiply-accumulate units arranged in parallel with one another.

4. The processor of claim 1 wherein the accumulator unit comprises an accumulator register file.

5. The processor of claim 1 wherein the reduction unit is configured to provide controllable selection between at least a first type of computation with saturation after each of a plurality of addition operations and a second type of computation with wrapping around of results of the addition operations, responsive to an applied control signal.

6. The processor of claim 1 wherein the reduction unit is further operative to permit the input operands to be subtracted from the first accumulator value.

7. The processor of claim 6 wherein the input operands are subtracted from the first accumulator value by inverting the bits of the input operands and setting a carry into each of a plurality of reduction adders of the reduction unit to a designated logic value.

8. The processor of claim 1 wherein the reduction unit is utilizable in conjunction with m arithmetic units and the accumulator unit to compute m elements of a dot product.

9. The processor of claim 8 wherein the m arithmetic units comprise m multipliers, with each of the multipliers being operative to compute P[i]=X[i]*Y[i], 2≦i≦m, the P[i]outputs of the m multipliers being fed as input operands to the reduction unit along with an accumulator value P[ 0 ] from the accumulator unit, the reduction unit being configured to compute Acc=P[ 0 ]+P[ 1 ]+P[ 2 ]+ . . . +P[m], where P[ 0 ] is set to zero for an initial iteration, and in a next iteration, m new elements of X and Y are multiplied, and P[ 0 ] is set to accumulator value Acc from the previous iteration, the iterations continuing until the m-element dot product is computed.

10. The processor of claim 9 wherein the reduction unit is operable in a saturation mode in which the reduction unit performs saturation after each addition operation, and each of the m multipliers saturates its result when overflow occurs.

11. The processor of claim 9 wherein the reduction unit is operable in an invert mode in which the reduction unit is configured to compute Acc=P[ 0 ]−P[ 1 ]−P[ 2 ]− . . . −P[m].

12. The processor of claim 8 wherein the reduction unit is configured so as to allow m elements of the dot product to be accumulated on each of a plurality of processor clock cycles.

13. The processor of claim 1 wherein the reduction unit comprises m reduction adders connected in series, each of the reduction adders being operative to add its respective input operands with or without saturation.

14. The processor of claim 1 wherein the multithreaded processor utilizes token triggered threading.

15. The processor of claim 1 wherein the reduction unit is utilizable in execution of one or more vector-reduce instructions, including at least one of a vector-multiply-and-reduce-add instruction, a vector-multiply-and-reduce-subtract instruction, and a vector-add-reduce-add instruction.

16. An integrated circuit comprising at least one multithreaded processor, the multithreaded processor comprising:

a plurality of arithmetic units;

an accumulator unit; and

a reduction unit coupled between the plurality of arithmetic units and the accumulator unit, the reduction unit being pipelined and configured to receive input operands from the arithmetic units and a first accumulator value from the accumulator unit, the reduction unit further comprising m inputs, m adders, and an m stage pipeline, where m is greater than or equal to two, each of the m inputs being coupled to a corresponding adder by means of N−1 pipeline registers where N is a stage number greater than or equal to 1, the m stage pipeline being configured to reduce the worst case delay of the reduction unit;

wherein the reduction unit is operative to sum the input operands and the first accumulator value, and to generate a second accumulator value for delivery to the accumulator unit;

wherein the reduction unit is controllable to support saturation and wrap-around arithmetic; and

wherein operations for a dot product computed for a given thread are executed concurrently with operations from other threads, the number of cycles between execution of instructions from the given thread being greater than or equal to a number of pipeline stages in the reduction unit plus any additional cycles needed to write to and read from the accumulator unit.

17. An apparatus for use in a multithreaded processor comprising a plurality of arithmetic units and an accumulator unit, the apparatus comprising:

a reduction unit coupled between the plurality of arithmetic units and the accumulator unit, the reduction unit being pipelined and configured to receive input operands from the arithmetic units and a first accumulator value from the accumulator unit, the reduction unit further comprising m inputs, m adders, and an m stage pipeline, where m is greater than or equal to two, each of the m inputs being coupled to a corresponding adder by means of N−1 pipeline registers where N is a stage number greater than or equal to 1, the m stage pipeline being configured to reduce the worst case delay of the reduction unit;

wherein the reduction unit is operative to sum the input operands and the first accumulator value, and to generate a second accumulator value for delivery to the accumulator unit;

wherein the reduction unit is controllable to support saturation and wrap-around arithmetic; and

wherein operations for a dot product computed for a given thread are executed concurrently with operations from other threads, the number of cycles between execution of instructions from the given thread being greater than or equal to a number of pipeline stages in the reduction unit plus any additional cycles needed to write to and read from the accumulator unit.