New instructions for CR/INT predication

DRAFT STATUS

See:

Rationale:

Condition Registers are conceptually perfect for use as predicate masks, the only problem being that typical Vector ISAs have quite comprehensive mask-based instructions: set-before-first, popcount and much more. In fact many Vector ISAs can use Vectors as masks, consequently the entire Vector ISA is usually available for use in creating masks (one exception being AVX512 which has a dedicated Mask regfile and opcodes). Duplication of such operations (popcount etc) is not practical for SV given the strategy of leveraging pre-existing Scalar instructions in a minimalist way.

With the scalar OpenPOWER v3.0B ISA having already popcnt, cntlz and others normally seen in Vector Mask operations it makes sense to allow both scalar integers and CR-Vectors to be predicate masks. That in turn means that much more comprehensive interaction between CRs and scalar Integers is required, because with the CR Predication Modes designating CR Fields (not CR bits) as Predicate Elements, fast transfers between CR Fields and the Integer Register File is needed.

The opportunity is therefore taken to also augment CR logical arithmetic as well, using a mask-based paradigm that takes into consideration multiple bits of each CR Field (eq/lt/gt/ov). By contrast v3.0B Scalar CR instructions (crand, crxor) only allow a single bit calculation, and both mtcr and mfcr are CR-orientated rather than CR Field orientated.

Also strangely there is no v3.0 instruction for directly moving CR Fields, only CR bits, so that is corrected here with mcrfm. The opportunity is taken to allow inversion of CR Field bits, when copied.

Basic concept:

  • CR-based instructions that perform simple AND/OR from any four bits of a CR field to create a single bit value (0/1) in an integer register
  • Inverse of the same, taking a single bit value (0/1) from an integer register to selectively target any four bits of a given CR Field
  • CR-to-CR version of the same, allowing multiple bits to be AND/OR/XORed in one hit.
  • Optional Vectorisation of the same when SVP64 is implemented

Purpose:

  • To provide a merged version of what is currently a multi-sequence of CR operations (crand, cror, crxor) with mfcr and mtcrf, reducing instruction count.
  • To provide a vectorised version of the same, suitable for advanced predication

Useful side-effects:

  • mtcrweird when RA=0 is a means to set or clear multiple arbitrary CR Field bits simultaneously, using immediates embedded within the instruction.
  • With SVP64 on the weird instructions there is bit-for-bit interaction between GPR predicate masks (r3, r10, r31) and the source or destination GPR, in ways that are not possible with other SVP64 instructions because normal SVP64 is bit-per-element. On these weird instructions the element in effect is a bit.
  • mfcrweird mitigates a need to add conflictd, part of vector ops, as well as allowing more complex comparisons.

Bit ordering.

Please see appendix regarding CR bit ordering and for the definition of CR{n}

Instruction form and pseudocode

DRAFT Instruction format (use of MAJOR 19 not approved by OPF ISA WG):

0-5 6-10 11 12-15 16-18 19-20 21-25 26-30 31 name
19 RT fmsk BFA XO[0:4] XO[5:9] /
19 1 //// 00011 rsvd
19 RT M fmsk BFA 0 0 0 fmap 00011 Rc crrweird
19 RT M fmsk BFA 0 1 0 fmap 00011 Rc mfcrweird
19 RA M fmsk BF 1 0 0 fmap 00011 0 mtcrrweird
19 RA M fmsk BF 1 0 0 fmap 00011 1 mtcrweird
19 BT M fmsk BFA 1 1 0 fmap 00011 0 crweirder
19 BF // M fmsk BFA 1 1 0 fmap 00011 1 mcrfm

crrweird

fmap is encoded in XO and is 4 bits

crrweird: RT,BFA,M,fmsk,fmap

creg = CR{BFA}
n0 = fmsk[0] & (fmap[0] == creg[0])
n1 = fmsk[1] & (fmap[1] == creg[1])
n2 = fmsk[2] & (fmap[2] == creg[2])
n3 = fmsk[3] & (fmap[3] == creg[3])
n = (n0||n1||n2||n3) & fmsk
result = (n != 0) if M else (n == fmsk)
RT[63] = result # MSB0 numbering, 63 is LSB
If Rc:
    CR0 = analyse(RT)

When used with SVP64 Prefixing this is a normal SVP64 type operation and as such can use Rc=1 and RC1 Data-dependent Mode capability

Also as noted below, element-width override bits normally used on the source is instead used to allow multiple results to be packed sequentially into the destination. Destination elwidth overrides still apply.

mfcrrweird

fmap is encoded in XO and is 4 bits

mfcrrweird: RT,BFA,fmsk,fmap

creg = CR{BFA}
n0 = fmsk[0] & (fmap[0] == creg[0])
n1 = fmsk[1] & (fmap[1] == creg[1])
n2 = fmsk[2] & (fmap[2] == creg[2])
n3 = fmsk[3] & (fmap[3] == creg[3])
result = n0||n1||n2||n3
RT[60:63] = result # MSB0 numbering, 63 is LSB
If Rc:
    CR0 = analyse(RT)

When used with SVP64 Prefixing this is a normal SVP64 type operation and as such can use Rc=1 and RC1 Data-dependent Mode capability.

Also as noted below, element-width override bits normally used on the source is instead used to allow multiple results to be packed into the destination. Destination elwidth overrides still apply

mtcrrweird

fmap is encoded in XO and is 4 bits

mtcrrweird: BF,RA,M,fmsk,fmap

a = (RA|0)
n0 = fmsk[0] & (fmap[0] == a[63])
n1 = fmsk[1] & (fmap[1] == a[62])
n2 = fmsk[2] & (fmap[2] == a[61])
n3 = fmsk[3] & (fmap[3] == a[60])
result = n0 || n1 || n2 || n3
if M:
    result |= CR{BF} & ~fmsk
CR{BF} = result

When used with SVP64 Prefixing this is a normal SVP64 type operation and as such can use RC1 Data-dependent Mode capability

mtcrweird

mtcrweird: BF,RA,M,fmsk,fmap

reg = (RA|0)
lsb = reg[63] # MSB0 numbering
n0 = fmsk[0] & (fmap[0] == lsb)
n1 = fmsk[1] & (fmap[1] == lsb)
n2 = fmsk[2] & (fmap[2] == lsb)
n3 = fmsk[3] & (fmap[3] == lsb)
result = n0 || n1 || n2 || n3
if M:
    result |= CR{BF} & ~fmsk
CR{BF} = result

Note that when M=1 this operation is a Read-Modify-Write on the CR Field BF. Masked-out bits of the 4-bit CR Field BF will not be changed when M=1. Correspondingly when M=0 this operation is an overwrite: no read of BF is required because the masked-out bits of the BF CR Field are set to zero.

When used with SVP64 Prefixing this is a cr ops SVP64 type operation that has 3-bit Data-dependent and 3-bit Predicate-result capability (BF is 3 bits)

mcrfm - Move CR Field, masked.

This instruction copies, sets, or inverts parts of a CR Field into another CR Field. mcrf copies only one bit of the CR from any arbitrary bit to any other arbitrary bit, whereas mcrfm copies an entire 4-bit CR Field (or masked parts thereof). Unlike mcrf the bits of the CR Field may not change position: the EQ bit from the source may only go into the EQ bit of the destination (optionally inverted, set, or cleared).

mcrfm: BF,BFA,M,fmsk,fmap

result = fmsk & CR{BFA}
if M:
    result |= CR{BF} & ~fmsk
result ^= fmap
CR{BF} = result

When M=1 this operation is a Read-Modify-Write on the CR Field BF. Masked-out bits of the 4-bit CR Field BF will not be changed when M=1. Correspondingly when M=0 this operation is an overwrite: no read of BF is required because the masked-out bits of the BF CR Field are set to zero.

When used with SVP64 Prefixing this is a cr ops SVP64 type operation that has 3-bit Data-dependent and 3-bit Predicate-result capability (BF is 3 bits)

Programmer's note: fmap being XORed onto the result provides considerable flexibility. individual bits of BFA may be copied inverted to BF by ensuring that fmsk and fmap have the same bit set. Also, individual bits in BF may be set to 1 by ensuring that the required bit of fmsk is set to zero and the same bit in fmap is set to 1

crweirder

crweirder: BT,BFA,fmsk,fmap

creg = CR{BFA}
n0 = fmsk[0] & (fmap[0] == creg[0])
n1 = fmsk[1] & (fmap[1] == creg[1])
n2 = fmsk[2] & (fmap[2] == creg[2])
n3 = fmsk[3] & (fmap[3] == creg[3])
bf = BT[2:4] # select CR field
bit = BT[0:1] # select bit of CR field
n = (n0||n1||n2||n3) & fmsk
result = (n != 0) if M else (n == fmsk)
CR{bf}[bit] = result

When used with SVP64 Prefixing this is a cr ops SVP64 type operation that has 5-bit Data-dependent and 5-bit Predicate-result capability (BT is 5 bits)

Example Pseudo-ops:

mtcri BF, fmap    mtcrweird BF, r0, 0, 0b1111,~fmap
mtcrset BF, fmsk  mtcrweird BF, r0, 1, fmsk,0b0000
mtcrclr BF, fmsk  mtcrweird BF, r0, 1, fmsk,0b1111

Vectorised versions involving GPRs

The name "weird" refers to a minor violation of SV rules when it comes to deriving the Vectorised versions of these instructions.

Normally the progression of the SV for-loop would move on to the next register. Instead however in the scalar case these instructions remain in the same register and insert or transfer between bits of the scalar integer source or destination. The reason is that when using CR Fields as predicate masks and there is a need to transfer into a GPR, again for use as a predicate mask, the CR Field bits need to be efficiently packed into that one GPR (r3, r10 or r31).

Further useful violation of the normal SV Elwidth override rules allows for packing (or unpacking) of multiple CR test results into (or out of) an Integer Element. Note that the CR (source operand) elwidth field is utilised to determine the bit- packing size (1/2/4/8 with remaining bits within the Integer element set to zero) whilst the INT (dest operand) elwidth field still sets the Integer element size as usual (8/16/32/default)

crrweird: RT, BB, fmsk.fmap

for i in range(VL):
    if BB.isvec:
        creg = CR{BB+i}
    else:
        creg = CR{BB}
    n0 = fmsk[0] & (fmap[0] == creg[0])
    n1 = fmsk[1] & (fmap[1] == creg[1])
    n2 = fmsk[2] & (fmap[2] == creg[2])
    n3 = fmsk[3] & (fmap[3] == creg[3])
    # OR or AND to a single bit
    n = (n0||n1||n2||n3) & fmsk
    result = (n != 0) if M else (n == fmsk)
    if RT.isvec:
        # TODO: RT.elwidth override to be also added here
        # note, yes, really, the CR's elwidth field determines
        # the bit-packing into the INT!
        if BB.elwidth == 0b00:
            # pack 1 result into 64-bit registers
            iregs[RT+i][0..62] = 0
            iregs[RT+i][63] = result # sets LSB to result
        if BB.elwidth == 0b01:
            # pack 2 results sequentially into INT registers
            iregs[RT+i//2][0..61] = 0
            iregs[RT+i//2][63-(i%2)] = result
        if BB.elwidth == 0b10:
            # pack 4 results sequentially into INT registers
            iregs[RT+i//4][0..59] = 0
            iregs[RT+i//4][63-(i%4)] = result
        if BB.elwidth == 0b11:
            # pack 8 results sequentially into INT registers
            iregs[RT+i//8][0..55] = 0
            iregs[RT+i//8][63-(i%8)] = result
    else:
        iregs[RT][63-i] = result # results also in scalar INT

Note that:

  • in the scalar case the CR-Vector assessment is stored bit-wise starting at the LSB of the destination scalar INT
  • in the INT-vector case the results are packed into LSBs of the INT Elements, the packing arrangement depending on both elwidth override settings.

mfcrrweird: RT, BFA, fmsk.fmap

Unlike crrweird the results are 4-bit wide, so the packing will begin to spill over to other destination elements. 8 results per destination at 4-bits each still fits into destination elwidth at 32-bit, but for 16-bit and 8-bit obviously this does not fit, and must split across to the next element

When for example destination elwidth is 16-bit (0b10) the following packing occurs:

  • SVRM bits 6:7 equal to 0b00 - one 4-bit result element packed into the first 4-bits of the 16-bit destination element (in the first 4 LSBs)
  • SVRM bits 6:7 equal to 0b01 - two 4-bit result elements packed into the first 8-bits of the 16-bit destination element (in the first 8 LSBs)
  • SVRM bits 6:7 equal to 0b10 - four 4-bit result elements packed into each 16-bit destination element
  • SVRM bits 6:7 equal to 0b11 - eight 4-bit result elements, the first four of which are packed into the first 16-bit destination element, the second four of which are packed into the second 16-bit destination element.

Pseudocode example: note that dest elwidth overrides affect the packing of results. BB.elwidth in effect requests how many 4-bit result elements would like to be packed, but RT.elwidth determines the limit. Any parts of the destination elements not containing results are set to zero.

for i in range(VL):
    if BB.isvec:
        creg = CR{BB+i}
    else:
        creg = CR{BB}
    n0 = fmsk[0] & (fmap[0] == creg[0])
    n1 = fmsk[1] & (fmap[1] == creg[1])
    n2 = fmsk[2] & (fmap[2] == creg[2])
    n3 = fmsk[3] & (fmap[3] == creg[3])
    result = n0||n1||n2||n3 # 4-bit result
    if RT.isvec:
        # RT.elwidth override can affect the packing
        bwid = {0b00:64, 0b01:8, 0b10:16, 0b11:32}[RT.elwidth]
        t4, t8 = min(4, bwid//2), min(8, bwid//2)
        # yes, really, the CR's elwidth field determines
        # the bit-packing into the INT!
        if BB.elwidth == 0b00:
            # pack 1 result into 64-bit registers
            idx, boff = i, 0
        if BB.elwidth == 0b01:
            # pack 2 results sequentially into INT registers
            idx, boff = i//2, i%2
        if BB.elwidth == 0b10:
            # pack 4 results sequentially into INT registers
            idx, boff = i//t4, i%t4
        if BB.elwidth == 0b11:
            # pack 8 results sequentially into INT registers
            idx, boff = i//t8, i%t8
    else:
        # exceeding VL=16 is UNDEFINED
        idx, boff = 0, i
    iregs[RT+idx][60-boff*4:63-boff*4] = result

v3.1 setbc instructions

There are additional setb conditional instructions in v3.1 (p129)

RT = (CR[BI] == 1) ? 1 : 0

which also negate that, and also return -1 / 0. these are similar to crweird but not the same purpose. most notable is that crweird acts on CR fields rather than the entire 32 bit CR.

Predication Examples

Take the following example:

r10 = 0b00010
sv.mtcrweird/dm=r10/dz cr8.v, 0, 0b0011.0000

Here, RA is zero, so the source input is zero. The destination is CR Field 8, and the destination predicate mask indicates to target the first two elements. Destination predicate zeroing is enabled, and the destination predicate is only set in the 2nd bit. fmsk is 0b0011, fmap is all zeros.

Let us first consider what should go into element 0 (CR Field 8):

  • The destination predicate bit is zero, and zeroing is enabled.
  • Therefore, what is in the source is irrelevant: the result must be zero.
  • Therefore all four bits of CR Field 8 are therefore set to zero.

Now the second element, CR Field 9 (CR9):

  • Bit 2 of the destination predicate, r10, is 1. Therefore the computation of the result is relevant.
  • RA is zero therefore bit 2 is zero. fmsk is 0b0011 and fmap is 0b0000
  • When calculating n0 thru n3 we get n0=1, n1=2, n2=0, n3=0
  • Therefore, CR9 is set (using LSB0 ordering) to 0b0011, i.e. to fmsk.

It should be clear that this instruction uses bits of the integer predicate to decide whether to set CR Fields to (fmsk & ~fmap) or to zero. Thus, in effect, it is the integer predicate that has been copied into the CR Fields.

By using twin predication, zeroing, and inversion (sm=~r3, dm=r10) for example, it becomes possible to combine two Integers together in order to set bits in CR Fields. Likewise there are dozens of ways that CR Predicates can be used, on the same sv.mtcrweird instruction.