Big Integer Arithmetic

DRAFT STATUS 19apr2022, last edited 23may2022

• discussion page for notes
• https://bugs.libre-soc.org/show_bug.cgi?id=817 bugreport
• https://bugs.libre-soc.org/show_bug.cgi?id=937 128/64 shifts
• analysis
• svfixedarith pseudocode

BigNum arithmetic is extremely common especially in cryptography, where for example RSA relies on arithmetic of 2048 or 4096 bits in length. The primary operations are add, multiply and divide (and modulo) with specialisations of subtract and signed multiply.

A reminder that a particular focus of SVP64 is that it is built on top of Scalar operations, where those scalar operations are useful in their own right without SVP64. Thus the operations here are proposed first as Scalar Extensions to the Power ISA.

A secondary focus is that if Vectorized, implementors may choose to deploy macro-op fusion targetting back-end 256-bit or greater Dynamic SIMD ALUs for maximum performance and effectiveness.

Analysis

Covered in analysis the summary is that standard adde is sufficient for SVP64 Vectorization of big-integer addition (and subfe for subtraction) but that big-integer shift, multiply and divide require an extra 3-in 2-out instructions, similar to Intel's shld and shrd, mulx and divq, to be efficient. The same instruction (maddedu) is used in both big-divide and big-multiply because 'maddedu''s primary purpose is to perform a fused 64-bit scalar multiply with a large vector, where that result is Big-Added for Big-Multiply, but Big-Subtracted for Big-Divide.

Chaining the operations together gives Scalar-by-Vector operations, except for sv.adde and sv.subfe which are Vector-by-Vector Chainable (through the CA flag). Macro-op Fusion and back-end massively-wide SIMD ALUs may be deployed in a fashion that is hidden from the user, behind a consistent, stable ISA API. The same macro-op fusion may theoretically be deployed even on Scalar operations.

DRAFT dsld

VA2-Form

• dsld RT,RA,RB,RC (Rc=0)
• dsld. RT,RA,RB,RC (Rc=1)

Pseudo-code:

n <- (RB)[58:63]
v <- ROTL64((RA), n)
mask <- MASK(0, 63-n)
RT <- (v[0:63] & mask) | ((RC) & ¬mask)
RS <- v[0:63] & ¬mask
overflow = 0
if RS != [0]*64:
    overflow = 1

Special Registers Altered:

CR0                    (if Rc=1)

DRAFT dsrd

VA2-Form

• dsrd RT,RA,RB,RC (Rc=0)
• dsrd. RT,RA,RB,RC (Rc=1)

Pseudo-code:

n <- (RB)[58:63]
v <- ROTL64((RA), 64-n)
mask <- MASK(n, 63)
RT <- (v[0:63] & mask) | ((RC) & ¬mask)
RS <- v[0:63] & ¬mask
overflow = 0
if RS != [0]*64:
    overflow = 1

Special Registers Altered:

CR0                    (if Rc=1)

maddedu

DRAFT

maddedu is similar to v3.0 madd, and is VA-Form despite having 2 outputs: the second destination register is implicit.

The pseudocode for maddedu RT, RA, RB, RC is:

prod[0:127] = (RA) * (RB)
sum[0:127] = EXTZ(RC) + prod
RT <- sum[64:127]
RS <- sum[0:63] # RS implicit register, see below

RC is zero-extended (not shifted, not sign-extended), the 128-bit product added to it; the lower half of that result stored in RT and the upper half in RS.

The differences here to maddhdu are that maddhdu stores the upper half in RT, where maddedu stores the upper half in RS.

The value stored in RT is exactly equivalent to maddld despite maddld performing sign-extension on RC, because RT is the full mathematical result modulo 2⁶⁴ and sign/zero extension from 64 to 128 bits produces identical results modulo 2⁶⁴. This is why there is no maddldu instruction.

Programmer's Note: As a Scalar Power ISA operation, like lq and stq, RS=RT+1. To achieve the same big-integer rolling-accumulation effect as SVP64: assuming the scalar to multiply is in r0, the vector to multiply by starts at r4 and the result vector in r20, instructions may be issued maddedu r20,r4,r0,r20 maddedu r21,r5,r0,r21 etc. where the first maddedu will have stored the upper half of the 128-bit multiply into r21, such that it may be picked up by the second maddedu. Repeat inline to construct a larger bigint scalar-vector multiply, as Scalar GPR register file space permits.

SVP64 overrides the Scalar behaviour of what defines RS. For SVP64 EXTRA register extension, the RM-1P-3S-1D format is used with the additional bit set for determining RS.

When EXTRA2_MODE is set to zero, the implicit RS register takes its Vector/Scalar setting from Rdest_EXTRA2, and takes the register number from RT, but all numbering is offset by MAXVL. Note that element-width overrides influence this offset (see SVP64 appendix for full details).

When EXTRA2_MODE is set to one, the implicit RS register is identical to RC extended with SVP64 using Rsrc3_EXTRA2 in every respect, including whether RC is set Scalar or Vector.

divmod2du RT,RA,RB,RC

DRAFT

Divide/Modulu Quad-Double Unsigned is another VA-Form instruction that is near-identical to divdeu except that:

• the lower 64 bits of the dividend, instead of being zero, contain a register, RC.
• it performs a fused divide and modulo in a single instruction, storing the modulo in an implicit RS (similar to maddedu)

RB, the divisor, remains 64 bit. The instruction is therefore a 128/64 division, producing a (pair) of 64 bit result(s), in the same way that Intel divq works. Overflow conditions are detected in exactly the same fashion as divdeu, except that rather than have UNDEFINED behaviour, RT is set to all ones and RS set to all zeros on overflow.

Programmer's note: there are no Rc variants of any of these VA-Form instructions. cmpi will need to be used to detect overflow conditions: the saving in instruction count is that both RT and RS will have already been set to useful values (all 1s and all zeros respectively) needed as part of implementing Knuth's Algorithm D

For SVP64, given that this instruction is also 3-in 2-out 64-bit registers, the exact same EXTRA format and setting of RS is used as for sv.maddedu. For Scalar usage, just as for maddedu, RS=RT+1 (similar to lq and stq).

Pseudo-code:

if ((RA) <u (RB)) & ((RB) != [0]*XLEN) then
    dividend[0:(XLEN*2)-1] <- (RA) || (RC)
    divisor[0:(XLEN*2)-1] <- [0]*XLEN || (RB)
    result <- dividend / divisor
    modulo <- dividend % divisor
    RT <- result[XLEN:(XLEN*2)-1]
    RS <- modulo[XLEN:(XLEN*2)-1]
else
    RT <- [1]*XLEN
    RS <- [0]*XLEN

[DRAFT] EXT04 Proposed Map

For the Opcode map (XO Field) see Power ISA v3.1, Book III, Appendix D, Table 13 (sheet 7 of 8), p1357. Proposed is the addition of:

• maddedu in 110010
• divmod2du in 111010
• pcdec in 111000