Comparative_analysis_Harmonised_RVP_vs_Andes_Packed_SIMD_ISA_proposal.mdwn

   1 # Comparative analysis of Andes Packed ISA proposal vs Harmonised RVP
   2
   3 Harmonised RVP is a proposal to provide SIMD functionality comparable to the Andes Packed SIMD ISA, but in a manner that is forwards compatible ("harmonised") with the RV Vector specification.
   4
   5 An example use case is a string copy operation - using Harmonised RVP, code can use integer register SIMD instructions to copy a string.  This code  can then also execute (unchanged) on a full RV Vector processor and use the dedicated vector unit to copy the string.  Harmonised RVP also upwards compatibility between RV32 and RV64 SIMD using this same approach.
   6
   7 ## Register file comparison
   8
   9 The Andes Packed SIMD ISA permits any GPR to be used for either INT8 or INT16 vector operations.
  10 In contrast, the default Harmonised RVP GPR register file is divided into a lower bank of Vector[INT8] and an upper banxk of Vector[INT16].
  11 (Effectively, the vector element size is encoded by the most significant bit of the 5 bit register specifiers.
  12 However programmers can reconfigure the register file data types, if the default configuration is unsuitable.)
  13
  14 | Register           | Andes ISA                              | Harmonised RVP ISA            |
  15 | ------------------ | -------------------------              | -------------------           |
  16 | v0                 | Hardwired zero                         | Hardwired zero                |
  17 | v1                 | 32bit GPR or Vector[4xINT8 or 2xINT16] | Predicate mask                |
  18 | |  |  |
  19 | v2                 | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[4xSINT8]  |
  20 | v3                 | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[4xSINT8]  |
  21 | v4                 | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[4xSINT8]  |
  22 | v5                 | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[4xSINT8]  |
  23 | v6                 | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[4xSINT8]  |
  24 | v7                 | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[4xSINT8]  |
  25 | v8                 | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[4xUINT8]  |
  26 | v9                 | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[4xUINT8]  |
  27 | v10                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[4xUINT8]  |
  28 | v11                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[4xUINT8]  |
  29 | v12                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[4xUINT8]  |
  30 | v13                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[4xUINT8]  |
  31 | v14                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[4xUINT8]  |
  32 | v15                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[4xUINT8]  |
  33 | |  |  |
  34 | v16                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[2xSINT16] |
  35 | v17                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[2xSINT16] |
  36 | v18                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[2xSINT16] |
  37 | v19                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[2xSINT16] |
  38 | v20                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[2xSINT16] |
  39 | v21                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[2xSINT16] |
  40 | v22                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[2xSINT16] |
  41 | v23                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[2xSINT16] |
  42 | v24                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[2xUINT16] |
  43 | v25                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[2xUINT16] |
  44 | v26                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[2xUINT16] |
  45 | v27                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[2xUINT16] |
  46 | v28                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[2xUINT16] |
  47 | v29                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[2xUINT16] |
  48 | |  |  |
  49 | v30                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[1xSINT32] |
  50 | v31                | 32bit GPR or Vector[4xINT8 or 2xINT16] | 32bit GPR or Vector[1xSINT32] |
  51
  52 Both Andes Packed SIMD and Harmonised RVP are intended to be "low end" SIMD implementations (for processors without dedicated vector registers).
  53 Instead, the integer register file is used for SIMD operations.  To maintain forwards compatibility with "high end" RV Vector implementations, programmer should use VLD and VST to load/store vectors.  The implementation will then load/store a vector to/from the register file supported by the implementation.
  54
  55 To keep implementations simple and focused on within-register SIMD only, there is a strict 1:1 mapping between vectors (v0-v31) and integer registers (r0-r31).  Standard calling conventions apply and so callee saved integer registers should be saved before being used as vector registers.
  56  Strided (VLDS/VSTS) and indexed (VLDX/VSTX) load/stores are complex, and simple implementations will trap on these instructions, permitting emulation in software.
  57
  58 ## Proposed Harmonised RVP vector op instruction encoding
  59
  60 Register x 2 -> register operations:
  61
  62 | 31 30 29 28 27 26 | 25  | 24 23 22 21 20 | 19 18 17 16 15 | 14 | 13 12 | 11 10 9 8 7 | 6 5 4 3 2 1 0 |
  63 | ----------------- | --  | -------------- | -------------- | -- | ----- | ----------- | ------------- |
  64 | func_6            | 0   | rs2            | rs1            | 0  | mm    | rd1         | VOP opcode    |
  65
  66 Immediate + register -> register operations:
  67
  68 | 31 30 29 | 28 27 26 | 25  | 24 23 22 21 20 | 19 18 17 16 15 | 14 | 13 12 | 11 10 9 8 7 | 6 5 4 3 2 1 0 |
  69 | -------- | -------- | --  | -------------- | -------------- | -- | ----- | ----------- | ------------- |
  70 | func_3   | imm[7:5] | 1   | imm[4:0]       | rs1            | 0  | mm    | rd1         | VOP opcode    |
  71
  72 Register x 3 -> register operations:
  73
  74 | 31 30 29 28 27 | 26 25  | 24 23 22 21 20 | 19 18 17 16 15 | 14 | 13 12 | 11 10 9 8 7 | 6 5 4 3 2 1 0 |
  75 | ----------------------- | -------------- | -------------- | -- | ----- | ----------- | ------------- |
  76 | rs3            | func_2 | rs2            | rs1            | 1  | mm    | rd1         | VOP opcode    |
  77
  78 Values for mm field (bits 12:13 above):
  79
  80 * mm = 00 -> no predicate mask, and use current global saturation / rounding settings
  81 * mm = 00 -> no predicate mask, and force saturation or rounding for this instruction only
  82 * mm = 10 -> use v1 as predicate mask, and use global saturation / rounding settings
  83 * mm = 11 -> use ~v1 as predicate mask, and use global saturation / rounding settings
  84
  85 ## 16-bit Arithmetic
  86
  87 | Andes Mnemonic           | 16-bit Instruction        | Harmonised RVP Equivalent |
  88 | ------------------ | ------------------------- | ------------------- |
  89 | ADD16 rt, ra, rb   | Add                       | VADD (v16 <= rt,ra,rb <= v29), mm=00|
  90 | RADD16 rt, ra, rb  | Signed Halving add        | RADD (v16 <= rt,ra,rb <= v23), mm=00|
  91 | URADD16 rt, ra, rb | Unsigned Halving add      | RADD (v24 <= rt,ra,rb <= v29), mm=00|
  92 | KADD16 rt, ra, rb  | Signed Saturating add     | VADD (v16 <= rt,ra,rb <= v23), mm=01|
  93 | UKADD16 rt, ra, rb | Unsigned Saturating add   | VADD (v24 <= rt,ra,rb <= v29), mm=01|
  94 | SUB16 rt, ra, rb   | Subtract                  | VSUB (v16 <= rt,ra,rb <= v29), mm=00|
  95 | RSUB16 rt, ra, rb  | Signed Halving sub        | RSUB (v16 <= rt,ra,rb <= v23), mm=00|
  96 | URSUB16 rt, ra, rb | Unsigned Halving sub                | RSUB (v24 <= rt,ra,rb <= v29), mm=00|
  97 | KSUB16 rt, ra, rb  | Signed Saturating sub               | VSUB (v16 <= rt,ra,rb <= v23), mm=01|
  98 | UKSUB16 rt, ra, rb | Unsigned Saturating sub             | VSUB (v24 <= rt,ra,rb <= v29), mm=01|
  99 | CRAS16 rt, ra, rb  | Cross Add & Sub                     | |
 100 | RCRAS16 rt, ra, rb | Signed Halving Cross Add & Sub      | |
 101 | URCRAS16 rt, ra, rb| Unsigned Halving Cross Add & Sub    | |
 102 | KCRAS16 rt, ra, rb | Signed Saturating Cross Add & Sub   | |
 103 | UKCRAS16 rt, ra, rb| Unsigned Saturating Cross Add & Sub | |
 104 | CRSA16 rt, ra, rb  | Cross Sub & Add                     | |
 105 | RCRSA16 rt, ra, rb | Signed Halving Cross Sub & Add      | |
 106 | URCRSA16 rt, ra, rb| Unsigned Halving Cross Sub & Add    | |
 107 | KCRSA16 rt, ra, rb | Signed Saturating Cross Sub & Add   | |
 108 | UKCRSA16 rt, ra, rb| Unsigned Saturating Cross Sub & Add | |
 109
 110 ## 8-bit Arithmetic
 111
 112 | Andes Mnemonic           | 8-bit Instruction        | Harmonised RVP Equivalent |
 113 | ------------------ | ------------------------- | ------------------- |
 114 | ADD8 rt, ra, rb    | Add                       | VADD (v2 <= rt,ra,rb <= v15), mm=00 |
 115 | RADD8 rt, ra, rb   | Signed Halving add        | RADD (v2 <= rt,ra,rb <= v7), mm=00 |
 116 | URADD8 rt, ra, rb  | Unsigned Halving add      | RADD (v8 <= rt,ra,rb <= v15), mm=00 |
 117 | KADD8 rt, ra, rb   | Signed Saturating add     | VADD (v2 <= rt,ra,rb <= v7), mm=01 |
 118 | UKADD8 rt, ra, rb  | Unsigned Saturating add   | VADD (v8 <= rt,ra,rb <= v15), mm=01 |
 119 | SUB8 rt, ra, rb    | Subtract                  | VSUB (v2 <= rt,ra,rb <= v15), mm=00 |
 120 | RSUB8 rt, ra, rb   | Signed Halving sub        | RSUB (v2 <= rt,ra,rb <= v7), mm=00 |
 121 | URSUB8 rt, ra, rb  | Unsigned Halving sub      | RSUB (v8 <= rt,ra,rb <= v15), mm=00 |
 122 | KSUB8 rt, ra, rb   | Signed Saturating sub        | VSUB (v2 <= rt,ra,rb <= v7), mm=01 |
 123 | UKSUB8 rt, ra, rb  | Unsigned Saturating sub      | VSUB (v8 <= rt,ra,rb <= v15), mm=01 |
 124
 125 ## 16-bit Shifts
 126
 127 SRA[I]16/SRL[I]16/SLL[I]16 to be mapped to VOP shift instructions in same manner as ADD16/SUB16
 128
 129 The “K” (Saturation) and “u” (Rounding) variants could be encoded using VOP’s mm field (mm=01 is saturated or rounded shift, mm=00 is standard VOP shift)
 130
 131 | Andes Mnemonic             | 16-bit Instruction         | Harmonised RVP Equivalent |
 132 | ------------------   | -------------------------  | ------------------- |
 133 | SRA16 rt, ra, rb     | Shift right arithmetic     | VSRA (v16 <= rt,ra,rb <= v29), mm=00|
 134 | SRAI16 rt, ra, im    | Shift right arithmetic imm | VSRAI (v16 <= rt,ra <= v29), mm=00|
 135 | SRA16.u rt, ra, rb   | Rounding Shift right arithmetic     | VSRA (v16 <= rt,ra,rb <= v29), mm=01|
 136 | SRAI16.u rt, ra, im  | Rounding Shift right arithmetic imm | VSRAI (v16 <= rt,ra <= v29), mm=01|
 137 | SRL16 rt, ra, rb     | Shift right logical        | VSRL (v16 <= rt,ra,rb <= v29), mm=00|
 138 | SRLI16 rt, ra, im    | Shift right logical imm    | VSRLI (v16 <= rt,ra <= v29), mm=00|
 139 | SRL16.u rt, ra, rb   | Rounding Shift right logical     | VSRL (v16 <= rt,ra,rb <= v29), mm=01|
 140 | SRLI16.u rt, ra, im  | Rounding Shift right logical imm | VSLRI (v16 <= rt,ra <= v29), mm=01|
 141 | SLL16 rt, ra, rb     | Shift left logical         | VSLL (v16 <= rt,ra,rb <= v29), mm=00|
 142 | SLLI16 rt, ra, im    | Shift left logical imm     | VSLLI (v16 <= rt,ra <= v29), mm=00|
 143 | KSLL16 rt, ra, rb    | Saturating Shift left logical       | VSLL (v16 <= rt,ra,rb <= v29), mm=01|
 144 | KSLLI16 rt, ra, im   | Saturating Shift left logical imm   | VSLLI (v16 <= rt,ra <= v29), mm=01|
 145 | KSLRA16 rt, ra, rb    | Saturating Shift left logical or Shift right arithmetic              ||
 146 | KSLRA16.u rt, ra, rb  | Saturating Shift left logical or Rounding Shift right arithmetic     ||
 147
 148
 149 ## 8-bit Shifts
 150
 151 Andes SIMD Packed ISA omits 8 bit shifts, but these can be encoded in Harmonised RVP as follows:
 152
 153 | Andes Mnemonic             | 8-bit Instruction         | Harmonised RVP Equivalent |
 154 | ------------------   | -------------------------  | ------------------- |
 155 | n/a     | Shift right arithmetic     | VSRA (v2 <= rt,ra,rb <= v15), mm=00|
 156 | n/a     | Shift right arithmetic imm | VSRAI (v2 <= rt,ra <= v15), mm=00|
 157 | n/a     | Rounding Shift right arithmetic     | VSRA (v2 <= rt,ra,rb <= v15), mm=01|
 158 | n/a     | Rounding Shift right arithmetic imm | VSRAI (v2 <= rt,ra <= v15), mm=01|
 159 | n/a     | Shift right logical        | VSRL (v2 <= rt,ra,rb <= v15), mm=00|
 160 | n/a     | Shift right logical imm    | VSRLI (v2 <= rt,ra <= v15), mm=00|
 161 | n/a     | Rounding Shift right logical     | VSRL (v2 <= rt,ra,rb <= v15), mm=01|
 162 | n/a     | Rounding Shift right logical imm | VSLRI (v2 <= rt,ra <= v15), mm=01|
 163 | n/a     | Shift left logical         | VSLL (v2 <= rt,ra,rb <= v15), mm=00|
 164 | n/a     | Shift left logical imm     | VSLLI (v2 <= rt,ra <= v15), mm=00|
 165 | n/a     | Saturating Shift left logical       | VSLL (v2 <= rt,ra,rb <= v15), mm=01|
 166 | n/a     | Saturating Shift left logical imm   | VSLLI (v2 <= rt,ra <= v15), mm=01|
 167
 168 ## 16-bit Comparison instructions
 169
 170 | Andes Mnemonic           | 16-bit Instruction        | Harmonised RVP Equivalent |
 171 | ------------------   | ------------------------- | ------------------- |
 172 | CMPEQ16 rt, ra, rb   | Compare equal             | VSEQ (v16 <= rt,ra,rb <= v29), mm=00|
 173 | SCMPLT16 rt, ra, rb   | Signed Compare less than        | !VSGT (v16 <= rt,ra,rb <= v23), mm=00|
 174 | SCMPLE16 rt, ra, rb   | Signed Compare less or equal    | VSLE (v16 <= rt,ra,rb <= v23), mm=00|
 175 | UCMPLT16 rt, ra, rb   | Unsigned Compare less than      | !VSGT (v24 <= rt,ra,rb <= v29), mm=00|
 176 | UCMPLE16 rt, ra, rb   | Unsigned Compare less or equal  | VSLE (v24 <= rt,ra,rb <= v29), mm=00|
 177
 178 ## 8-bit Comparison instructions
 179
 180 | Andes Mnemonic           | 8-bit Instruction        | Harmonised RVP Equivalent |
 181 | ------------------   | ------------------------- | ------------------- |
 182 | CMPEQ8 rt, ra, rb   | Compare equal             | VSEQ (v2 <= rt,ra,rb <= v7), mm=00|
 183 | SCMPLT8 rt, ra, rb   | Signed Compare less than        | !VSGT (v2 <= rt,ra,rb <= v7), mm=00|
 184 | SCMPLE8 rt, ra, rb   | Signed Compare less or equal    | VSLE (v2 <= rt,ra,rb <= v7), mm=00|
 185 | UCMPLT8 rt, ra, rb   | Unsigned Compare less than      | !VSGT (v8 <= rt,ra,rb <= v15), mm=00|
 186 | UCMPLE8 rt, ra, rb   | Unsigned Compare less or equal  | VSLE (v8 <= rt,ra,rb <= v15), mm=00|
 187
 188 ## 16-bit Miscellaneous instructions
 189
 190 | Andes Mnemonic      | 16-bit Instruction                | Harmonised RVP Equivalent |
 191 | ------------------  | ------------------------          | ------------------- |
 192 | SMIN16 rt, ra, rb   | Signed minimum                    | VMIN (v16 <= rt,ra,rb <= v23), mm=00|
 193 | UMIN16 rt, ra, rb   | Unsigned minimum                  | VMIN (v24 <= rt,ra,rb <= v29), mm=00|
 194 | SMAX16 rt, ra, rb   | Signed maximum                    | VMAX (v16 <= rt,ra,rb <= v23), mm=00|
 195 | UMAX16 rt, ra, rb   | Unsigned maximum                  | VMAX (v24 <= rt,ra,rb <= v29), mm=00|
 196 | SCLIP16 rt, ra, im  | Signed clip                       | ?VCLIP (v16 <= rt,ra,rb <= v23), mm=01|
 197 | UCLIP16 rt, ra, im  | Unsigned clip                     | ?VCLIP (v24 <= rt,ra,rb <= v29), mm=01|
 198 | KMUL16 rt, ra, rb   | Signed multiply 16x16->16         | VMUL (v16 <= rt,ra,rb <= v23), mm=01|
 199 | KMULX16 rt, ra, rb  | Signed crossed multiply 16x16->16 | |
 200 | SMUL16 rt, ra, rb   | Signed multiply 16x16->32         | VMUL (30 <= rt <= 31, v16 <= ra,rb <= v23), mm=00|
 201 | SMULX16 rt, ra, rb  | Signed crossed multiply 16x16->32 | |
 202 | UMUL16 rt, ra, rb   | Signed multiply 16x16->32         | VMUL (30 <= rt <= 31, v24 <= ra,rb <= r31), mm=00|
 203 | UMULX16 rt, ra, rb  | Signed crossed multiply 16x16->32 | |
 204 | KABS16 rt, ra       | Saturated absolute value          | VSGNX (v16 <= rt <= v29, v16 <= ra,rb <= v23, mm=01) |
 205
 206 ## 8-bit Miscellaneous instructions
 207
 208 | Andes Mnemonic           | 8-bit Instruction        | Harmonised RVP Equivalent |
 209 | ------------------   | ------------------------- | ------------------- |
 210 | SMIN8 rt, ra, rb   | Signed minimum           | VMIN (v2 <= rt,ra,rb <= v7), mm=00|
 211 | UMIN8 rt, ra, rb   | Unsigned minimum         | VMIN (v8 <= rt,ra,rb <= v15), mm=00|
 212 | SMAX8 rt, ra, rb   | Signed maximum           | VMAX (v2 <= rt,ra,rb <= v7), mm=00|
 213 | UMAX8 rt, ra, rb   | Unsigned maximum         | VMAX (v8 <= rt,ra,rb <= v15), mm=00|
 214 | KABS8 rt, ra       | Saturated absolute value | VSGNX (v2 <= rt <= v15, v2 <= ra,rb <= v8, mm=01) |
 215
 216 ## 8-bit Unpacking instructions
 217
 218 | Andes Mnemonic           | 8-bit Instruction       | Harmonised RVP Equivalent |
 219 | ------------------   | -------------------------   | ------------------- |
 220 | SUNPKD810 rt, ra     | Signed unpack bytes 1 & 0   | VMV (v16<= rt <= 23, v2 <= ra <= v7), mm=00|
 221 | SUNPKD820 rt, ra     | Signed unpack bytes 2 & 0   | |
 222 | SUNPKD830 rt, ra     | Signed unpack bytes 3 & 0   | |
 223 | SUNPKD831 rt, ra     | Signed unpack bytes 3 & 1   | |
 224 | ZUNPKD810 rt, ra     | Unsigned unpack bytes 1 & 0 | VMV (v24<= rt <= 31, v8 <= ra <= v15), mm=00|
 225 | ZUNPKD820 rt, ra     | Unsigned unpack bytes 2 & 0 | |
 226 | ZUNPKD830 rt, ra     | Unsigned unpack bytes 3 & 0 | |
 227 | ZUNPKD831 rt, ra     | Unsigned unpack bytes 3 & 1 | |