Systolic super summation device

• US 4,751,665 A
• Filed: 02/24/1986
• Issued: 06/14/1988
• Est. Priority Date: 02/24/1986
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Cappello, Peter R., Miranker, Willard L.
• Primary Examiner(s): Malzahn, David H.
• Application Number:
  US06/832,282
• Time in Patent Office: 841 Days
• Field of Search: 364/748
• US Class Current: 708/505
• CPC Class Codes: G06F 2207/388 Skewing
• CPC Class Codes: G06F 2207/3892 Systolic array
• CPC Class Codes: G06F 5/015 having at least two separately controlled shifting levels, e.g. using shifting matrices G06F5/012 takes precedence
• CPC Class Codes: G06F 7/485 Adding; Subtracting G06F7/4833, G06F7/4836 take precedence
• CPC Class Codes: G06F 7/49936 Normalisation mentioned as feature only
• CPC Class Codes: G06F 7/49942 Significance control
• CPC Class Codes: G06F 7/49947 Rounding
• CPC Class Codes: G06F 7/5095 word-serial, i.e. with an accumulator-register

Abstract

A cellular architecture for a systolic super summer for high throughput performance of repeated sums of floating-point numbers. The summer receives pipelined inputs of streams of summands, converts the floating-point summands into a fixed-point form by a sieve-like pipelined cellular network, and sums the emerging fixed-point numbers in a corresponding network of extremely long accumulators. The throughput per unit area of the hardware approaches that of a tree network, but without the long wires and signal propagation that are intrinsic to tree networks.

Claims

1. A device for calculating sums of floating point summands, each summand including a characteristic and a mantissa, said device comprising:

a sieve comprising a characteristic decode section for determining the value of the characteristic of each summand and a sieve section responsive to the determined characteristic of a summand for routing that summand toward an appropriate position at the output of said sieve corresponding to a fixed-point representation of said summand, said sieve including coalescence means for combining mantissas of different summands prior to said sieve output; and

accumulator means receiving and accumulating outputs from said sieve and converting the accumulated sum to floating point form, to thereby obtain a floating-point representation of each sum.

2. A device as defined in claim 1, wherein said sieve section passes a first mantissa toward its appropriate sieve output position along first and second path segments and passes a second mantissa toward its appropriate sieve output position along a third path segment and said second path segment, and wherein said coalescence means combines said first and second mantissas in said second path segment.

3. A device as defined in claim 2, wherein each sieve cell includes first and second inputs for receiving first and second mantissa bits and first and second outputs, said sieve cell including means for passing said first and second mantissa bits to said first and second outputs, respectively, when said first and second mantissa bits are en route toward different sieve output positions, with said coalescence means combining said first and second mantissa bits when they are en route toward the same sieve output position.

4. A device as defined in claim 3, wherein said coalescence means includes means for generating sum and carry bits corresponding to the combination of said first and second mantissa bits, said sum bit being provided to said second output, and said sieve cell includes a carry bit for providing said carry bit to an adjacent cell in said sieve.

5. A device as defined in claim 1, wherein each mantissa includes a plurality of mantissa bits passing toward respective sieve output positions along respective paths. and wherein said coalescence means combines only those mantissa bits which are destined for the same sieve output position.

6. A device as defined in claim 1, wherein said summands are received by said sieve in serial-parallel order with said characteristic decoding section comprising a plurality of levels, each level including a plurality of individual bit comparators for comparing respective bits of a characteristic to reference bit values corresponding to an order of magnitude associated with said each level, said sieve section including a plurality of rows of sieve cells and each level of said characteristic decoding section corresponding to a respective row of sieve cells in said sieve section and providing a control signal to cause the sieve cells in said respective row to change the movement direction of the mantissa bits corresponding to the characteristic decoded at said each level.

7. A device as defined in claim 1, wherein each summand includes characteristic and mantissa bits and wherein said sieve section comprises a plurality of sieve cells through which said mantissa bits pass en route toward their respective appropriate positions at the output of said sieve. said coalescence means comprising means for combining mantissa bits which simultaneously occupy a sieve cell en route toward the same position at said sieve output.

8. A device as defined in claim 7, wherein said coalescence means includes means for generating sum and carry bits corresponding to the sum of the combined mantissa bits being routed toward said same position, said sieve passing said sum bits to said same position while passing said carry bits toward a higher order position at said sieve output.

9. A device as defined in claim 8, wherein said coalescence means further comprises means for combining sum, carry or mantissa bits which occupy the same sieve cell en route toward the same sieve output position.

10. A device as defined in claim 7, wherein each said summand further includes a sign bit indicating the sign of said summand, said device including means for converting said mantissa bits to signed mantissa bits and for providing said signed mantissa bits to said sieve, and wherein said accumulator means includes a plurality of bit positions from a most significant bit position to a least significant bit position, and sign detection means for determining the sign of the sum in said accumulator means by successively examining the signs of individual bit positions commencing with said most significant bit position.

11. A device as defined in claim 10, wherein said sign detection means comprises a finite automaton having a plurality of states including at least one terminal state, said sign detection means completing detection of the sign of said sum in said accumulator means upon reaching said at least one terminal state regardless of the number of individual bit positions examined.

12. A device as defined in claim 1, wherein said sieve section comprises a matrix of sieve cells arranged in rows and columns including a sinking section and a ramp section and passes each mantissa vertically through a column in said sinking section until the corresponding characteristic is decoded, said ramp section thereafter passing said mantissa diagonally through at least a portion of said ramp section toward said appropriate position at the output of said sieve.

13. A device as defined in claim 12, wherein each sieve cell in said sinking section includes first (m^s) and second (m^r) inputs and first (m^s) and second (m^r) outputs, and a control input (f) for receiving a control signal having a value indicating whether or not the characteristic corresponding to the mantissa bit at said first input has been decoded, said sieve cell passing its second input to its second output regardless of the value of said control signal and passing its first input to its first or second outputs in accordance with the value of said control signal.

14. A device as claimed in claim 13, wherein said sieve section includes a plurality of sinking sections, each of said sinking sections having a corresponding characteristic decoding section decoding said characteristics in a different sequence.

15. A device as defined in claim 14, wherein said accumulator and sieve are arranged in a cylindrical configuration, with said plurality of sinking sections offset with respect to one another around the periphery of the cylinder.

16. A device as defined in claim 1, wherein said accumulator means comprises an accumulator pipeline formed of a plurality of accumulators including a first accumulator α

₀ and at least one further accumulator, said device including shifting means, upon entry of a last summand mantissa into said first accumulator, for shifting the contents of said first accumulator to said at least one further accumulator for conversion to floating point form while said first accumulator commences accumulation of summands corresponding to a further sum.

17. A device as defined in claim 16, further comprising means for providing an end of sum signal to said accumulator means following entry of said last summand mantissa into said accumulator means, wherein said accumulator pipeline includes a plurality of further accumulators, the contents of each accumulator being shifted to a subsequent accumulator upon entry of said end of sum signal into said each accumulator while continuing the process of conversion to floating point form, said device further including means for indicating the completion of conversion to floating point form and means, responsive to such indication, for passing the floating point sum to a device output.

18. A device for calculating sums of floating point summands, each summand including a characteristic and a mantissa, and for then calculating an iterative summation of a plurality of successive sums, said device comprising:

a sieve comprising a characteristic decode section for determining the value of the characteristic of each summand and a sieve section responsive to the determined characteristic of a summand for routing that summand toward an appropriate position at the output of said sieve corresponding to a fixed-point representation of said summand, said sieve including coalescence means for combining mantissas of different summands prior to said sieve output;

accumulator means receiving and accumulating outputs from said sieve and converting the accumulated sum to floating point form, to thereby obtain a floating-point representation of each sum, said accumulator means including a plurality of accumulators including a first accumulator α

₀, a holding accumulator β

₀ and at least one further accumulator;

shifting means, upon entry into said first accumulator of a last summand mantissa in each of said successive sums, for shifting the contents of said first accumulator to said holding accumulator for summation with the contents of said holding accumulator, and upon entry into said holding accumulator of a last sum in said iterative summation, for shifting the contents of said holding accumulator to said at least one further accumulator for conversion to floating point form while said first accumulator and holding accumulator commence calculation of a further iterative summation.

19. A device as defined in claim 18 for concurrently performing a plurality of iterative summations, said accumulator means including a plurality of holding accumulators each for accumulating sums from said first accumulator corresponding to a respective iterative summation.