+[[!tag oldstandards]]
+
SimpleV Prefix (SVprefix) Proposal v0.3
=======================================
-This proposal is designed to be able to operate without SVcsr, but not to
-require the absence of SVcsr.
+* Copyright (c) Jacob Lifshay, 2019
+* Copyright (c) Luke Kenneth Casson Leighton, 2019
+
+This proposal is designed to be able to operate without SVorig, but not to
+require the absence of SVorig. See Specification_.
+
+Principle: SVprefix embeds (unmodified) RVC and 32-bit scalar opcodes
+into 32, 48 and 64 bit RV formats, to provide Vectorisation context
+on a per-instruction basis.
+
+.. _Specification: http://libre-riscv.org/simple_v_extension/specification/
+.. _Appendix: http://libre-riscv.org/simple_v_extension/appendix/
+
+.. contents::
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, where each sub-vector
-has a length *svlen*, and an element type *etype*. 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 number of bytes in a sub-vector, *svsz*, is the
-product of *svlen* and the element size in bytes.
+* 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.
+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
=======================
-This proposal doesn't include any prefixed RVC instructions, instead, it will
-include 32-bit instructions that are compressed forms of SVprefix 48-bit
-instructions, in the same manner that RVC instructions are compressed forms of
-RVI instructions. The compressed instructions will be defined later by
-considering which 48-bit instructions are the most common.
+
+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 [[discussion]].
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.
+
+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 VL and MVL is provided.
+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, *Reserved* entries must be zero. RV32 equivalent encodings
-included for side-by-side comparison (and listed below, separately).
+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] | *Reserved* | 011111 |
-+---------------+--------+------------+------------+-----+------------+-------------+------+------------+--------+
-| P48-ST-type |vitp7[6]| rs1[5] | rs2[5] | vs2 | vs1 | vitp7[5:0] | *Reserved* | 011111 |
-+---------------+--------+------------+------------+-----+------------+-------------+------+------------+--------+
-| P48-R-type | rd[5] | rs1[5] | rs2[5] | vs2 | vs1 | vitp6 | *Reserved* | 011111 |
-+---------------+--------+------------+------------+-----+------------+--------------------+------------+--------+
-| P48-I-type | rd[5] | rs1[5] | vitp7[6] | vd | vs1 | vitp7[5:0] | *Reserved* | 011111 |
-+---------------+--------+------------+------------+-----+------------+--------------------+------------+--------+
-| P48-U-type | rd[5] | *Reserved* | *Reserved* | vd | *Reserved* | vitp6 | *Reserved* | 011111 |
-+---------------+--------+------------+------------+-----+------------+-------------+------+------------+--------+
-| P48-FR-type | rd[5] | rs1[5] | rs2[5] | vs2 | vs1 | *Reserved* | vtp5 | *Reserved* | 011111 |
-+---------------+--------+------------+------------+-----+------------+-------------+------+------------+--------+
-| P48-FI-type | rd[5] | rs1[5] | vitp7[6] | vd | vs1 | vitp7[5:0] | *Reserved* | 011111 |
-+---------------+--------+------------+------------+-----+------------+-------------+------+------------+--------+
-| P48-FR4-type | rd[5] | rs1[5] | rs2[5] | vs2 | rs3[5] | vs3 [#fr4]_ | vtp5 | *Reserved* | 011111 |
-+---------------+--------+------------+------------+-----+------------+-------------+------+------------+--------+
-
-.. [#fr4] 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 | 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 [#fr4]_ | vtp5 | *Rsvd* | 011111 |
++---------------+--------+--------+----------+-----+--------+-------------+------+--------+--------+
+
+.. [#fr4] 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 | 47:18 |
+| Encoding | RV32 Encoding |
+---------------+---------------+
-| RV32 Encoding | 31:2 |
+| 47:32 | 31:2 |
+---------------+---------------+
-| P48-LD-type | RV32-I-type |
+| P48-LD-type | RV32-I-type |
+---------------+---------------+
| P48-ST-type | RV32-S-Type |
+---------------+---------------+
+---------------+---------------+
| P48-FI-type | RV32-I-Type |
+---------------+---------------+
-| P48-FR4-type | RV32-FR-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 + 1 + 1 |
-+---------------+-------------+-------+----------+----------+--------+----------+--------+--------+------------+
-| RV32-S-type + imm[11:5] + rs2[4:0] + rs1[4:0] + funct3 | imm[4:0] + opcode + 1 + 1 |
-+---------------+-------------+-------+----------+----------+--------+----------+--------+--------+------------+
-| RV32-I-type + imm[11:0] + rs1[4:0] + funct3 | rd[4:0] + opcode + 1 + 1 |
-+---------------+-------------+-------+----------+----------+--------+----------+--------+--------+------------+
-| RV32-U-type + imm[31:12] | rd[4:0] + opcode + 1 + 1 |
-+---------------+-------------+-------+----------+----------+--------+----------+--------+--------+------------+
-| RV32-FR4-type + rs3[4:0] + fmt + rs2[4:0] + rs1[4:0] + funct3 | rd[4:0] + opcode + 1 + 1 |
-+---------------+-------------+-------+----------+----------+--------+----------+--------+--------+------------+
-| RV32-FR-type + funct5 + fmt + rs2[4:0] + rs1[4:0] + rm | rd[4:0] + opcode + 1 + 1 |
-+---------------+-------------+-------+----------+----------+--------+----------+--------+--------+------------+
++---------------+-------------+-------+----------+----------+--------+----------+--------+--------+
+| 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".
+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 |
+--------------+-------+--------+--------+--------+--------+
| Encoding | 63 | 62 | 61 | 60 | 59:48 |
+--------------+-------+--------+--------+--------+--------+
-| P64-LD-type | rd[6] | rs1[6] | | | SVtyp |
+| P64-LD-type | rd[6] | rs1[6] | | *Rsvd* | VLtyp |
+--------------+-------+--------+--------+--------+--------+
-| P64-ST-type | | rs1[6] | rs2[6] | | SVtyp |
+| P64-ST-type | | rs1[6] | rs2[6] | *Rsvd* | VLtyp |
+--------------+-------+--------+--------+--------+--------+
-| P64-R-type | rd[6] | rs1[6] | rs2[6] | | SVtyp |
+| P64-R-type | rd[6] | rs1[6] | rs2[6] | vd | VLtyp |
+--------------+-------+--------+--------+--------+--------+
-| P64-I-type | rd[6] | rs1[6] | | | SVtyp |
+| P64-I-type | rd[6] | rs1[6] | | *Rsvd* | VLtyp |
+--------------+-------+--------+--------+--------+--------+
-| P64-U-type | rd[6] | | | | SVtyp |
+| P64-U-type | rd[6] | | | *Rsvd* | VLtyp |
+--------------+-------+--------+--------+--------+--------+
-| P64-FR-type | | rs1[6] | rs2[6] | | SVtyp |
+| P64-FR-type | | rs1[6] | rs2[6] | vd | VLtyp |
+--------------+-------+--------+--------+--------+--------+
-| P64-FI-type | rd[6] | rs1[6] | rs2[6] | | SVtyp |
+| P64-FI-type | rd[6] | rs1[6] | rs2[6] | vd | VLtyp |
+--------------+-------+--------+--------+--------+--------+
-| P64-FR4-type | rd[6] | rs1[6] | rs2[6] | rs3[6] | SVtyp |
+| 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.
+prefix as well. VLtyp encodes how (whether) to set SVPSTATE.VL and SVPSTATE.MAXVL.
-SVtyp field encoding
+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.
+
+-----------+-------------+--------------+----------+----------------------+
-| SVtyp[11] | SVtyp[10:6] | SVtyp[5:1] | SVtyp[0] | comment |
-+-----------+-------------+--------------+----------+----------------------+
-| 0 | 000000 | 00000 | 0 | no change to VL/MVL |
+| VLtyp[11] | VLtyp[10:6] | VLtyp[5:1] | VLtyp[0] | comment |
+-----------+-------------+--------------+----------+----------------------+
-| 0 | VLdest!=0 | 00000 | 0 | reserved |
+| 0 | 00000 | 00000 | 0 | no change to VL/MVL |
+-----------+-------------+--------------+----------+----------------------+
| 0 | VLdest | VLEN | vlt | VL imm/reg mode (vlt)|
+-----------+-------------+--------------+----------+----------------------+
-| 1 | VLdest | MVL+VL-immed | MVL/VL immed mode |
+| 1 | VLdest | MVL+VL-immed | 0 | MVL+VL immed mode |
++-----------+-------------+--------------+----------+----------------------+
+| 1 | VLdest | MVL-immed | 1 | MVL immed mode |
+-----------+-------------+--------------+----------+----------------------+
-Notes:
-
-* When SVtyp is all zeros, neither VL nor MVL are changed
-* SVtype[11]=0, SVtype[5:0]=0 and SVtype[10:6] non-zero is a reserved encoding.
- Its uses raises an illegal instruction exception.
+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 VLIW format, when bit 11 of SVtyp is zero:
+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.)
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 SVtype is 1:
+When bit 11 of VLtype is 1:
-* both MAXVL and VL are set to (VLenimmed+1)
-* the same value goes into the scalar register VLdest (if VLdest is not x0)
+* 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 VL in a "loop mode" (SVtype[11]=0) or
-in a "one-off" mode (SVtype[11]=1) which sets both MVL and VL to the
+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:
+
+.. parsed-literal::
+
+ # 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
=======================
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:
+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 |
| 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.
+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
====================================
Notes:
-* A register strided LD/ST would require *5* registers. srcbase, vd/vs2, predicate 1, predicate 2 and the stride register.
+* 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.
+* 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?
+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 | 4 |
+| 00 | SUBVL |
+----------------+-------+
-| 01 | 1 |
+| 01 | 2 |
+----------------+-------+
-| 10 | 2 |
+| 10 | 3 |
+----------------+-------+
-| 11 | 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
=================================
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 |
+=======+============+====================+==============================================+
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
=============================
4. Otherwise,
1. Signedness is Unsigned.
-.. [#sign_enc] Like in fcvt.d.l[u], but unlike in fmv.x.w, since there is no
- fmv.x.wu
+.. [#sign_enc] 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.
+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 |
+-------+-----------------------+
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.
| XX01XX | *Reserved* |
+--------+------------------------------------------------+
-vitp7 field: only tpred=
+vitp7 field: only tpred
+---------+------------+----------+-------------+------------+
| vitp7 | itype | tpred[2] | tpred[0:1] | svlen |
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.
+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
a. If the base instruction is an I-type instruction, then
CSR Registers
=============
-+--------+-----------------+---------------------------------------------------+
-| Name | Legal Values | Meaning |
-+========+=================+===================================================+
-| VL | 0 <= VL <= XLEN | Vector Length. The number of sub-vectors operated |
-| | | on by vector instructions. |
-+--------+-----------------+---------------------------------------------------+
-| Vstart | 0 <= VL < XLEN | The sub-vector index to start execution at. |
-| | | Successful completion of all elements in a vector |
-| | | instruction sets Vstart to 0. Set to the index of |
-| | | the failing sub-vector when a vector instruction |
-| | | traps. Used to resume execution of vector |
-| | | instructions after a trap. Is *NOT* "slow" |
-+--------+-----------------+---------------------------------------------------+
-
-SetVL
-=====
-
-setvl rd, rs1, imm
-
-imm is the amount of space allocated from the register file by the compiler.
-
-Pseudocode:
-
-1. Trap if imm > XLEN.
-2. If rs1 is x0, then
- 1. Set VL to imm.
-3. Else If regs[rs1] > 2 * imm, then
- 1. Set VL to XLEN.
-4. Else If regs[rs1] > imm, then
- 1. Set VL to regs[rs1] / 2 rounded down.
-5. Otherwise,
- 1. Set VL to regs[rs1].
-6. Set regs[rd] to VL.
-
-Additional Instructions
-=======================
-
-Add instructions to convert between integer types.
+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.
-Add instructions to `swizzle`_ elements in sub-vectors. Note that the sub-vector
-lengths of the source and destination won't necessarily match.
+* VL
+* MVL
+* SVPSTATE
+* SUBVL
-.. _swizzle: https://www.khronos.org/opengl/wiki/Data_Type_(GLSL)#Swizzling
-
-Add instructions to transpose (2-4)x(2-4) element matrices.
+Associated SET and GET on the CSRs is exactly as in the main spec as well
+(including CSRRWI and CSRRW differences).
-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 partly
-by twin-predication, register cannot: requires MV.X aka VSELECT*)
+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.
-Add a register gather instruction (aka MV.X)
+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.
-# Open questions <a name="questions"></a>
+SVPSTATE fields are exactly the same layout as STATE:
-What is SUBVL and how does it work
++----------+----------+----------+----------+----------+---------+---------+
+| (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.
-SVorig goes to a lot of effort to make VL 1<= MAXVL and MAXVL 1..64 where both CSRs may be stored internally in only 6 bits.
+Additional Instructions
+=======================
-Thus, CSRRWI can reach 1..32 for VL and MAXVL.
+* 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]])
-In addition, setting a hardware loop to zero turning instructions into NOPs, um, just branch over them, to start the first loop at the end, on the test for loop variable being zero, a la c "while do" instead of "do while".
+subelement swizzle example:
-Or, does it not matter that VL only goes up to 31 on a CSRRWI, and that it only goes to a max of 63 rather than 64?
+ velswizzle x32, x64, SRCSUBVL=3, DESTSUBVL=4, ELTYPE=u8, elements=[0, 0, 2, 1]
---
+.. _swizzle: https://www.khronos.org/opengl/wiki/Data_Type_(GLSL)#Swizzling
-Should these questions be moved to Discussion subpage
+questions
+=========
---
+Moved to the discussion page (link at top of this page)
-Is MV.X good enough a substitute for swizzle?
+TODO
+====
---
+Work out a way to do sub-element swizzling.
-Is vectorised srcbase ok as a gather scatter and ok substitute for register stride? 5 dependency registers (reg stride being the 5th) is quite scary
projects / libreriscv.git / blobdiff