1. EE 382 Processor DesignWinter 98/99Michael Flynn 1 AT Arithmetic Most concern has gone into creating fast implementation of (especially) FP Arith. Under the AT (area-time) rule, area is (almost) as important. So it’s important to know the latency, bandwidth and area that any particular algorithm requires.

2. EE 382 Processor DesignWinter 98/99Michael Flynn 2 Integer addition Adders are the fundamental building block of the processor, defining  t. Adder types include –carry chain, carry select (conditional sum), carry lookahead (Brent-Kung), canonic (prefix) carry skip, Ling Most high speed 32b adders take about the same area (f normalized)…1 A to 1.5A

3. EE 382 Processor DesignWinter 98/99Michael Flynn 3 Integer addition Both area and time scale as n, the adder precision. The delay, t, scales slowly (log n) Area scale about linearly with n; so a 64b adder takes 2-3 A, but still fits into  t …maybe by definition of a “cycle”.

4. EE 382 Processor DesignWinter 98/99Michael Flynn 4 Carry skip adder

5. EE 382 Processor DesignWinter 98/99Michael Flynn 5 Manchester carry chain

6. EE 382 Processor DesignWinter 98/99Michael Flynn 6 Carry skip logic

7. EE 382 Processor DesignWinter 98/99Michael Flynn 7 Carry select addition

8. EE 382 Processor DesignWinter 98/99Michael Flynn 8 FP addition A basic FP adder has 5 steps –exponent difference, pre align, significand add, post align, and round. Assuming that a full shifter has about the same complexity (delay and area) as an add, then 64b FP addition takes 7 - 10 A, and has about 5  t execution

9. EE 382 Processor DesignWinter 98/99Michael Flynn 9 FP addition Advanced FP adders are faster and use more area: 1) Two path FADD creates separate paths for operands; a path for operands whose exponents close in value (subtract)  this is the only case when we need a full shift to renormalize the result a path for other cases where the exponent difference is > 2 (this is the only case that uses a full shift to prealign significands) 2) A FADD with integrated rounding. Here the rounding step is eliminated by computing both the sum/difference and the result plus 1… this is done by using 2 adders (or a compound adder) and then MUXing out the final result.

10. EE 382 Processor DesignWinter 98/99Michael Flynn 10 FP adders The two path FP adder uses an additional significand adder and exponent adder… about 3-4 A. It reduces FADD delay by one  t Integrated rounding adds another rounding adder plus MUX…another 3-4 A while reducing delay by another  t

11. EE 382 Processor DesignWinter 98/99Michael Flynn 11 FP adders Net area time tradeoff Basic… Area 10 A and delay 4-5  t Two path… Area 13.5 A and delay 3-4  t Integrated round (with two paths)… area 17 A and delay 2-3  t For pipelining add 1 A per pipe stage and use upper range on  t

12. EE 382 Processor DesignWinter 98/99Michael Flynn 12 Multipliers After add, the most important arithmetic op Approaches –encode the multiplier bits (Booth 2, Booth 3...) –assimilate the partial products one, two or n pass (iterated arrays or trees) arrays (simple, double, higher level) trees (Wallace, binary[4:2], ZD,….) –CPA to produce product

13. EE 382 Processor DesignWinter 98/99Michael Flynn 13 Multipliers Integer and FP multipliers usually have about the same execution time (with same precision, n) Booth reduces number of pp’s but adds MUXs to generate the pp’s. Most of the area, and probably delay too, is in the pp reduction tree.

14. EE 382 Processor DesignWinter 98/99Michael Flynn 14 16 bit Booth 2 multiply

15. EE 382 Processor DesignWinter 98/99Michael Flynn 15 16 bit Booth 2 example

16. EE 382 Processor DesignWinter 98/99Michael Flynn 16 16 bit Booth 2 pp selector logic

17. EE 382 Processor DesignWinter 98/99Michael Flynn 17 16 bit Booth 3 multiply

18. EE 382 Processor DesignWinter 98/99Michael Flynn 18 5 x 5 unsigned multiplication

19. EE 382 Processor DesignWinter 98/99Michael Flynn 19 1-bit adder

20. EE 382 Processor DesignWinter 98/99Michael Flynn 20 Wallace tree

21. EE 382 Processor DesignWinter 98/99Michael Flynn 21 Wallace tree reduction

22. EE 382 Processor DesignWinter 98/99Michael Flynn 22 Multipliers A full tree implementation of a 54b (FP type) with Booth 2 has tree height 28 and uses about 2500 CSAs (or about 50 A in the tree). Maybe a total of 10 A in MUXs plus 50 A in tree and 3A in the CPA, 62A total.The fastest multiplier is, maybe, 2  t Using a 2 pass tree reduces the hardware considerably; height is 14 using about 700 CSAs or 14 A…total area 5 + 14 + 3 = 22A; 3-4  t

23. EE 382 Processor DesignWinter 98/99Michael Flynn 23 Multipliers To pipeline the Multiplier we need a full tree implementation; probably 3-4  t. Perhaps Booth3, followed by a full tree (h = 17) and CPA stage. Probably area = 50 - 60A

24. EE 382 Processor DesignWinter 98/99Michael Flynn 24 Divide Infrequent op, but long latency can affect IPC achieved. Algorithms: –SRT 2 or 3 bit (32 - 36  t) maybe 6-10 A –NR or Binomial expansion (10- 14  t); needs at least 6 A for table and control plus use of MPY –Bipartite tables for small n (less than 24b)

25. EE 382 Processor DesignWinter 98/99Michael Flynn 25 Divide SRT creates quotient 2 or 3 bits/iteration –uses divisor - partial remainder lookup table for trial quotient then subtracts –result (partial rem.) is in redundant form so no restoration is needed; also result is left as a sum and carry pair (no cpa needed) –fast iteration is possible, sometimes 2 x per  t

26. EE 382 Processor DesignWinter 98/99Michael Flynn 26 Divide Multiply based use either Newton Raphson or Binomial series –if f(x) = b - 1/x; root is at x = 1/b then NR iteration is x i+1 = x i (2  b x i ) –converges is quadratic, doubles precision of result each iteration –so start with table lookup of 1/b to 8b, then 3 iterations gives 64b result then a x (1/b) is quotient

27. EE 382 Processor DesignWinter 98/99Michael Flynn 27 Divide Divide is not usually pipelined, except for small n implementations. Frequently combined with square root in the same implementation.

28. EE 382 Processor DesignWinter 98/99Michael Flynn 28 Sub word concurrency Provides 8, 16, 32b concurrent ops within “existing” integer or FP hardware In 64b integer unit can do 8 x 8, or 4 x 16, or 2 x 32 ops concurrently Since FP units are designed to be faster, may be use it: 8 x 4, or 2 x 16, or 2 x 24.

29. EE 382 Processor DesignWinter 98/99Michael Flynn 29 Sub word concurrency Usually only for add and multiply Implementations straightforward for add; more complicated for multiply –requires reorganizing partitions of the pp tree –affects multiply area and delay marginally (maybe 10% delay and 20% area) isa must define “saturating” arithmetic.