Conventions

Conventions used in this document:

  • Bits are numbered starting from 0 at the LSB, so bit 3 is 1 in the integer 8.
  • Bit ranges are inclusive on both ends, so 5:3 means bits 5, 4, and 3.
  • Operations work on variable-length vectors of sub-vectors up to VL in length, where each sub-vector has a length svlen, and svlen elements of type etype.
  • The actual total number of elements is therefore svlen times VL.
  • When the vectors are stored in registers, all elements are packed so that there is no padding in-between elements of the same vector.
  • The register file itself is thus best viewed as a byte-level SRAM that is typecast to an array of etypes
  • The number of bytes in a sub-vector, svsz, is the product of svlen and the element size in bytes.

Options

The following partial / full implementation options are possible:

  • SVPrefix augments the main Specification
  • SVPrefix operates independently, without the main spec VL (and MVL) CSRs (in any priv level)
  • SVPrefix operates independently, without the main spec SUBVL CSRs (in any priv level)
  • SVPrefix has no support for VL (or MVL) overrides in the 64 bit instruction format (VLtyp=0 as the only legal permitted value)
  • SVPrefix has no support for svlen overrides in either the 48 or 64 bit instruction format either (svlen=0 as the only legal permitted value).

All permutations of the above options are permitted, and the UNIX platform must raise illegal instruction exceptions on implementations that do not support each option. For example, an implementation that has no support for VLtyp that sees an opcode with a nonzero VLtyp must raise an illegal instruction exception.

Note that SVPrefix (VLtyp and svlen) has its own STATE CSR, SVPSTATE. This allows Prefixed operations to be re-entrant on traps, and to not affect VBLOCK use of VL or SUBVL.

If the main Specification CSRs and features are to be supported (VBLOCK), then when VLtyp or svlen are "default" they utilise the main Specification VBLOCK VL and/or SUBVL, and, correspondingly, the main VBLOCK STATE CSR will be updated and used to track hardware loops.

If however VLtyp is set to nondefault, then the SVPSTATE src and destoffs fields are used instead to create the hardware loops, and likewise if svlen is set to nondefault, SVPSTATE's svoffs field is used.

Half-Precision Floating Point (FP16)

If the F extension is supported, SVprefix adds support for FP16 in the base FP instructions by using 10 (H) in the floating-point format field fmt and using 001 (H) in the floating-point load/store width field.

Compressed Instructions

Compressed instructions are under evaluation by taking the same prefix as used in P48, embedding that and standard RVC opcodes (minus their RVC prefix) into a 32-bit space. This by taking the three remaining Major "custom" opcodes (0-2), one for each of the three RVC Quadrants. see <a href="./discussion/">discussion</a>.

48-bit Prefixed Instructions

All 48-bit prefixed instructions contain a 32-bit "base" instruction as the last 4 bytes. Since all 32-bit instructions have bits 1:0 set to 11, those bits are reused for additional encoding space in the 48-bit instructions.

64-bit Prefixed Instructions

The 48 bit format is further extended with the full 128-bit range on all source and destination registers, and the option to set both SVSTATE.VL and SVSTATE.MVL is provided.

48-bit Instruction Encodings

In the following table, Rsvd (reserved) entries must be zero. RV32 equivalent encodings included for side-by-side comparison (and listed below, separately).

First, bits 17:0:

Encoding 17 16 15 14 13 12 11:7 6 5:0
P48-LD-type rd[5] rs1[5] vitp7[6] vd vs1 vitp7[5:0] Rsvd 011111
P48-ST-type vitp7[6] rs1[5] rs2[5] vs2 vs1 vitp7[5:0] Rsvd 011111
P48-R-type rd[5] rs1[5] rs2[5] vs2 vs1 vitp6 Rsvd 011111
P48-I-type rd[5] rs1[5] vitp7[6] vd vs1 vitp7[5:0] Rsvd 011111
P48-U-type rd[5] Rsvd Rsvd vd Rsvd vitp6 Rsvd 011111
P48-FR-type rd[5] rs1[5] rs2[5] vs2 vs1 Rsvd vtp5 Rsvd 011111
P48-FI-type rd[5] rs1[5] vitp7[6] vd vs1 vitp7[5:0] Rsvd 011111
P48-FR4-type rd[5] rs1[5] rs2[5] vs2 rs3[5] vs3 [1] vtp5 Rsvd 011111
[1]Only vs2 and vs3 are included in the P48-FR4-type encoding because there is not enough space for vs1 as well, and because it is more useful to have a scalar argument for each of the multiplication and addition portions of fmadd than to have two scalars on the multiplication portion.

Table showing correspondance between P48--type and RV32--type. These are bits 47:18 (RV32 shifted up by 16 bits):

Encoding RV32 Encoding
47:32 31:2
P48-LD-type RV32-I-type
P48-ST-type RV32-S-Type
P48-R-type RV32-R-Type
P48-I-type RV32-I-Type
P48-U-type RV32-U-Type
P48-FR-type RV32-FR-Type
P48-FI-type RV32-I-Type
P48-FR4-type RV32-FR4-type

Table showing Standard RV32 encodings:

Encoding 31:27 26:25 24:20 19:15 14:12 11:7 6:2 1:0
RV32-R-type funct7 rs2[4:0] rs1[4:0] funct3 rd[4:0] opcode 0b11
RV32-S-type imm[11:5] rs2[4:0] rs1[4:0] funct3 imm[4:0] opcode 0b11
RV32-I-type imm[11:0] rs1[4:0] funct3 rd[4:0] opcode 0b11
RV32-U-type imm[31:12] rd[4:0] opcode 0b11
RV32-FR4-type rs3[4:0] fmt rs2[4:0] rs1[4:0] funct3 rd[4:0] opcode 0b11
RV32-FR-type funct5 fmt rs2[4:0] rs1[4:0] rm rd[4:0] opcode 0b11

64-bit Instruction Encodings

Where in the 48 bit format the prefix is "0b0011111" in bits 0 to 6, this is now set to "0b0111111".

63:48 47:18 17:7 6:0
64 bit prefix RV32[31:3] P48[17:7] 0b0111111
  • The 64 bit prefix format is below
  • Bits 18 to 47 contain bits 3 to 31 of a standard RV32 format
  • Bits 7 to 17 contain bits 7 through 17 of the P48 format
  • Bits 0 to 6 contain the standard RV 64-bit prefix 0b0111111

64 bit prefix format:

Encoding 63 62 61 60 59:48
P64-LD-type rd[6] rs1[6]   Rsvd VLtyp
P64-ST-type   rs1[6] rs2[6] Rsvd VLtyp
P64-R-type rd[6] rs1[6] rs2[6] vd VLtyp
P64-I-type rd[6] rs1[6]   Rsvd VLtyp
P64-U-type rd[6]     Rsvd VLtyp
P64-FR-type   rs1[6] rs2[6] vd VLtyp
P64-FI-type rd[6] rs1[6] rs2[6] vd VLtyp
P64-FR4-type rd[6] rs1[6] rs2[6] rs3[6] VLtyp

The extra bit for src and dest registers provides the full range of up to 128 registers, when combined with the extra bit from the 48 bit prefix as well. VLtyp encodes how (whether) to set SVPSTATE.VL and SVPSTATE.MAXVL.

VLtyp field encoding

NOTE: VL and MVL below are local to SVPrefix and, if non-default, will update the src and dest element offsets in SVPSTATE, not the main Specification STATE. If default (all zeros) then STATE VL and MVL apply to this instruction, and STATE.srcoffs (etc) will be used.

VLtyp[11] VLtyp[10:6] VLtyp[5:1] VLtyp[0] comment
0 00000 00000 0 no change to VL/MVL
0 VLdest VLEN vlt VL imm/reg mode (vlt)
1 VLdest MVL+VL-immed 0 MVL+VL immed mode
1 VLdest MVL-immed 1 MVL immed mode

Note: when VLtyp is all zeros, the main Specification VL and MVL apply to this instruction. If called outside of a VBLOCK or if sv.setvl has not set VL, the operation is "scalar".

Just as in the VBLOCK format, when bit 11 of VLtyp is zero:

  • if vlt is zero, bits 1 to 5 specify the VLEN as a 5 bit immediate (offset by 1: 0b00000 represents VL=1, 0b00001 represents VL=2 etc.)
  • if vlt is 1, bits 1 to 5 specify the scalar (RV standard) register from which VL is set. x0 is not permitted
  • VL goes into the scalar register VLdest (if VLdest is not x0)

When bit 11 of VLtype is 1:

  • if VLtyp[0] is zero, both SVPSTATE.MAXVL and SVPSTATE.VL are set to (imm+1). The same value goes into the scalar register VLdest (if VLdest is not x0)
  • if VLtyp[0] is 1, SVPSTATE.MAXVL is set to (imm+1). SVPSTATE.VL will be truncated to within the new range (if VL was greater than the new MAXVL). The new VL goes into the scalar register VLdest (if VLdest is not x0).

This gives the option to set up SVPSTATE.VL in a "loop mode" (VLtype[11]=0) or in a "one-off" mode (VLtype[11]=1) which sets both MVL and VL to the same immediate value. This may be most useful for one-off Vectorised operations such as LOAD-MULTI / STORE-MULTI, for saving and restoration of large batches of registers in context-switches or function calls.

Note that VLtyp's VL and MVL are not the same as the main Specification VL or MVL, and that loops will alter srcoffs and destoffs in SVPSTATE in VLtype nondefault mode, but the srcoffs and destoffs in STATE, if VLtype=0.

Furthermore, the execution order and exception handling must be exactly the same as in the main spec (Program Order must be preserved)

Pseudocode for SVPSTATE.VL:

# pseudocode

regs = [0u64; 128];
vl = 0;

// instruction fields:
rd = get_rd_field();
vlmax = get_immed_field();

// handle illegal instruction decoding
if vlmax > XLEN {
    trap()
}

// calculate VL
if rs1 == 0 { // rs1 is x0
    vl = vlmax
} else {
    vl = min(regs[rs1], vlmax)
}

// write rd
if rd != 0 {
    // rd is not x0
    regs[rd] = vl
}

vs#/vd Fields' Encoding

vs#/vd Mnemonic Meaning
0 S the rs#/rd field specifies a scalar (single sub-vector); the rs#/rd field is zero-extended to get the actual 7-bit register number
1 V the rs#/rd field specifies a vector; the rs#/rd field is decoded using the Vector Register Number Encoding to get the actual 7-bit register number

If a vs#/vd field is not present, it is as if it was present with a value that is the bitwise-or of all present vs#/vd fields.

  • scalar register numbers do NOT increment when allocated in the hardware for-loop. the same scalar register number is handed to every ALU.
  • vector register numbers DO increase when allocated in the hardware for-loop. sequentially-increasing register data is handed to sequential ALUs.

Vector Register Number Encoding

For the 48 bit format, when vs#/vd is 1, the actual 7-bit register number is derived from the corresponding 6-bit rs#/rd field:

Actual 7-bit register number
Bit 6 Bits 5:1 Bit 0
rs#/rd[0] rs#/rd[5:1] 0

For the 64 bit format, the 7 bit register is constructed from the 7 bit fields: bits 0 to 4 from the 32 bit RV Standard format, bit 5 from the 48 bit prefix and bit 6 from the 64 bit prefix. Thus in the 64 bit format the full range of up to 128 registers is directly available. This for both when either scalar or vector mode is set.

Load/Store Kind (lsk) Field Encoding

vd/vs2 vs1 Meaning
0 0 srcbase is scalar, LD/ST is pure scalar.
1 0 srcbase is scalar, LD/ST is unit strided
0 1 srcbase is a vector (gather/scatter aka array of srcbases). VSPLAT and VSELECT
1 1 srcbase is a vector, LD/ST is a full vector LD/ST.

Notes:

  • A register strided LD/ST would require 5 registers. srcbase, vd/vs2, predicate 1, predicate 2 and the stride register.
  • Complex strides may all be done with a general purpose vector of srcbases.
  • Twin predication may be used even when vd/vs1 is a scalar, to give VSPLAT and VSELECT, because the hardware loop ends on the first occurrence of a 1 in the predicate when a predicate is applied to a scalar.
  • Full vectorised gather/scatter is enabled when both registers are marked as vectorised, however unlike e.g Intel AVX512, twin predication can be applied.

Open question: RVV overloads the width field of LOAD-FP/STORE-FP using the bit 2 to indicate additional interpretation of the 11 bit immediate. Should this be considered?

Sub-Vector Length (svlen) Field Encoding

NOTE: svlen is not the same as the main spec SUBVL. When nondefault (not zero) SVPSTATE context is used for Sub vector loops. However is svlen is zero, STATE and SUBVL is used instead.

Bitwidth, from VL's perspective, is a multiple of the elwidth times svlen. So within each loop of VL there are svlen sub-elements of elwidth in size, just like in a SIMD architecture. When svlen is set to 0b00 (indicating svlen=1) no such SIMD-like behaviour exists and the subvectoring is disabled.

Predicate bits do not apply to the individual sub-vector elements, they apply to the entire subvector group. This saves instructions on setup of the predicate.

svlen Encoding Value
00 SUBVL
01 2
10 3
11 4

In independent standalone implementations that do not implement the main specification, the value of SUBVL in the above table (svtyp=0b00) is set to 1, such that svlen is also 1.

Behaviour of operations that set svlen are identical to those of the main spec. See section on VLtyp, above.

Predication (pred) Field Encoding

pred Mnemonic Predicate Register Meaning
000 None None The instruction is unpredicated
001 Reserved Reserved  
010 !x9 x9 (s1) execute vector op[0..i] on x9[i] == 0
011 x9 execute vector op[0..i] on x9[i] == 1
100 !x10 x10 (a0) execute vector op[0..i] on x10[i] == 0
101 x10 execute vector op[0..i] on x10[i] == 1
110 !x11 x11 (a1) execute vector op[0..i] on x11[i] == 0
111 x11 execute vector op[0..i] on x11[i] == 1

Twin-predication (tpred) Field Encoding

Twin-predication (ability to associate two predicate registers with an instruction) applies to MV, FCLASS, LD and ST. The same format also applies to integer-branch-compare operations although it is not to be considered "twin" predication. In the case of integer-branch-compare operations, the second register (if enabled) stores the results of the element comparisons. See Appendix for details.

tpred Mnemonic Predicate Register Meaning
000 None None The instruction is unpredicated
001 x9,off

src=x9, dest=none

src=none, dest=x10

src=x9, dest=x10

src[0..i] uses x9[i], dest unpredicated
010 off,x10 dest[0..i] uses x10[i], src unpredicated
011 x9,10 src[0..i] uses x9[i], dest[0..i] uses x10[i]
100 None RESERVED Instruction is unpredicated (TBD)
101 !x9,off

src=!x9, dest=none

src=none, dest=!x10

src=!x9, dest=!x10

 
110 off,!x10  
111 !x9,!x10  

Integer Element Type (itype) Field Encoding

Signedness [2] [2] itype Element Type Mnemonic in Integer Instructions Mnemonic in FP Instructions (such as fmv.x) Meaning (INT may be un/signed, FP just re-sized
Unsigned 01 u8 BU BU Unsigned 8-bit
10 u16 HU HU Unsigned 16-bit
11 u32 WU WU Unsigned 32-bit
00 uXLEN WU/DU/QU WU/LU/TU Unsigned XLEN-bit
Signed 01 i8 BS BS Signed 8-bit
10 i16 HS HS Signed 16-bit
11 i32 W W Signed 32-bit
00 iXLEN W/D/Q W/L/T Signed XLEN-bit
[2](1, 2) Signedness is defined in Signedness Decision Procedure

Note: vector mode is effectively a type-cast of the register file as if it was a sequential array being typecast to typedef itype[] (c syntax). The starting point of the "typecast" is the vector register rs#/rd.

Example: if itype=0b10 (u16), and rd is set to "vector", and VL is set to 4, the 64-bit register at rd is subdivided into FOUR 16-bit destination elements. It is NOT four separate 64-bit destination registers (rd+0, rd+1, rd+2, rd+3) that are sign-extended from the source width size out to 64-bit, because that is itype=0b00 (uXLEN).

Note also: changing elwidth creates packed elements that, depending on VL, may create vectors that do not fit perfectly onto XLEN sized registry file bit-boundaries. This does NOT result in the destruction of the MSBs of the last register written to at the end of a VL loop. More details on how to handle this are described in the main Specification.

Signedness Decision Procedure

  1. If the opcode field is either OP or OP-IMM, then
    1. Signedness is Unsigned.
  2. If the opcode field is either OP-32 or OP-IMM-32, then
    1. Signedness is Signed.
  3. If Signedness is encoded in a field of the base instruction, [3] then
    1. Signedness uses the encoded value.
  4. Otherwise,
    1. Signedness is Unsigned.
[3]Like in fcvt.d.l[u], but unlike in fmv.x.w, since there is no fmv.x.wu

Vector Type and Predication 5-bit (vtp5) Field Encoding

In the following table, X denotes a wildcard that is 0 or 1 and can be a different value for every occurrence.

vtp5 pred svlen
1XXXX vtp5[4:2] vtp5[1:0]
01XXX
000XX
001XX Reserved

Vector Integer Type and Predication 6-bit (vitp6) Field Encoding

In the following table, X denotes a wildcard that is 0 or 1 and can be a different value for every occurrence.

vitp6 itype pred[2] pred[0:1] svlen
XX1XXX vitp6[5:4] 0 vitp6[3:2] vitp6[1:0]
XX00XX
XX01XX Reserved

vitp7 field: only tpred

vitp7 itype tpred[2] tpred[0:1] svlen
XXXXXXX vitp7[5:4] vitp7[6] vitp7[3:2] vitp7[1:0]

48-bit Instruction Encoding Decision Procedure

In the following decision procedure, Reserved means that there is not yet a defined 48-bit instruction encoding for the base instruction.

  1. If the base instruction is a load instruction, then
    1. If the base instruction is an I-type instruction, then
      1. The encoding is P48-LD-type.
    2. Otherwise
      1. The encoding is Reserved.
  2. If the base instruction is a store instruction, then
    1. If the base instruction is an S-type instruction, then
      1. The encoding is P48-ST-type.
    2. Otherwise
      1. The encoding is Reserved.
  3. If the base instruction is a SYSTEM instruction, then
    1. The encoding is Reserved.
  4. If the base instruction is an integer instruction, then
    1. If the base instruction is an R-type instruction, then
      1. The encoding is P48-R-type.
    2. If the base instruction is an I-type instruction, then
      1. The encoding is P48-I-type.
    3. If the base instruction is an S-type instruction, then
      1. The encoding is Reserved.
    4. If the base instruction is an B-type instruction, then
      1. The encoding is Reserved.
    5. If the base instruction is an U-type instruction, then
      1. The encoding is P48-U-type.
    6. If the base instruction is an J-type instruction, then
      1. The encoding is Reserved.
    7. Otherwise
      1. The encoding is Reserved.
  5. If the base instruction is a floating-point instruction, then
    1. If the base instruction is an R-type instruction, then
      1. The encoding is P48-FR-type.
    2. If the base instruction is an I-type instruction, then
      1. The encoding is P48-FI-type.
    3. If the base instruction is an S-type instruction, then
      1. The encoding is Reserved.
    4. If the base instruction is an B-type instruction, then
      1. The encoding is Reserved.
    5. If the base instruction is an U-type instruction, then
      1. The encoding is Reserved.
    6. If the base instruction is an J-type instruction, then
      1. The encoding is Reserved.
    7. If the base instruction is an R4-type instruction, then
      1. The encoding is P48-FR4-type.
    8. Otherwise
      1. The encoding is Reserved.
  6. Otherwise
    1. The encoding is Reserved.

CSR Registers

CSRs are the same as in the main Specification, if associated functionality is implemented. They have the exact same meaning as in the main specification.

  • VL
  • MVL
  • SVPSTATE
  • SUBVL

Associated SET and GET on the CSRs is exactly as in the main spec as well (including CSRRWI and CSRRW differences).

Note that if both VLtyp and svlen are not implemented, SVPSTATE is not required. Also if VL and SUBVL are not implemented, STATE from the main Specification is not required either.

However if partial functionality is implemented, the unimplemented bits in STATE and SVPSTATE must be zero, and, in the UNIX Platform, an illegal exception MUST be raised if unsupported bits are written to.

SVPSTATE fields are exactly the same layout as STATE:

(31..28) (27..26) (25..24) (23..18) (17..12) (11..6) (5...0)
rsvd dsvoffs subvl destoffs srcoffs vl maxvl

However note that where STATE stores the scalar register number to be used as VL, SVPSTATE.VL actually contains the actual VL value, in an identical fashion to RVV.

Additional Instructions

  • Add instructions to convert between integer types.
  • Add instructions to swizzle elements in sub-vectors. Note that the sub-vector lengths of the source and destination won't necessarily match.
  • Add instructions to transpose (2-4)x(2-4) element matrices.
  • Add instructions to insert or extract a sub-vector from a vector, with the index allowed to be both immediate and from a register (immediate can be covered by twin-predication, register might be, by virtue of predicates being registers)
  • Add a register gather instruction (aka MV.X: regfile[rd] = regfile[regfile[rs1]])

subelement swizzle example:

velswizzle x32, x64, SRCSUBVL=3, DESTSUBVL=4, ELTYPE=u8, elements=[0, 0, 2, 1]

questions

Moved to the discussion page (link at top of this page)

TODO

Work out a way to do sub-element swizzling.