openpower/isa/simplev.mdwn

   1 <!-- This defines Draft SVP64 instructions to augment PowerISA Version 3.0 -->
   2 <!-- These are not described in book 1 -->
   3
   4 # svstep
   5
   6 SVL-Form
   7
   8 * svstep RT,SVi,vf (Rc=0)
   9 * svstep. RT,SVi,vf (Rc=1)
  10
  11 Pseudo-code:
  12
  13     if SVi[3:4] = 0b11 then
  14         # store pack and unpack in SVSTATE
  15         SVSTATE[53] <- SVi[5]
  16         SVSTATE[54] <- SVi[6]
  17         RT <- [0]*62 || SVSTATE[53:54]
  18     else
  19         step <- SVSTATE_NEXT(SVi, vf)
  20         RT <- [0]*57 || step
  21
  22 Special Registers Altered:
  23
  24     CR0                     (if Rc=1)
  25
  26 # setvl
  27
  28 SVL-Form
  29
  30 * setvl RT,RA,SVi,vf,vs,ms (Rc=0)
  31 * setvl. RT,RA,SVi,vf,vs,ms (Rc=1)
  32
  33 Pseudo-code:
  34
  35     overflow <- 0b0
  36     VLimm <- SVi + 1
  37     # set or get MVL
  38     if ms = 1 then MVL <- VLimm[0:6]
  39     else           MVL <- SVSTATE[0:6]
  40     # set or get VL
  41     if vs = 0                then VL <- SVSTATE[7:13]
  42     else if _RA != 0         then
  43         if (RA) >u 0b1111111 then
  44             VL <- 0b1111111
  45             overflow <- 0b1
  46         else                      VL <- (RA)[57:63]
  47     else if _RT = 0          then VL <- VLimm[0:6]
  48     else if CTR >u 0b1111111 then
  49         VL <- 0b1111111
  50         overflow <- 0b1
  51     else                          VL <- CTR[57:63]
  52     # limit VL to within MVL
  53     if VL >u MVL then
  54         overflow <- 0b1
  55         VL <- MVL
  56     SVSTATE[0:6] <- MVL
  57     SVSTATE[7:13] <- VL
  58     if _RT != 0 then
  59        GPR(_RT) <- [0]*57 || VL
  60     # MAXVL is a static "state-reset".
  61     if ms = 1 then
  62         SVSTATE[63] <- vf   # set Vertical-First mode
  63         SVSTATE[62] <- 0b0  # clear persist bit
  64
  65 Special Registers Altered:
  66
  67     CR0                     (if Rc=1)
  68
  69 # svremap
  70
  71 SVRM-Form
  72
  73 * svremap SVme,mi0,mi1,mi2,mo0,mo1,pst
  74
  75 Pseudo-code:
  76
  77     # registers RA RB RC RT EA/FRS SVSHAPE0-3 indices
  78     SVSTATE[32:33] <- mi0
  79     SVSTATE[34:35] <- mi1
  80     SVSTATE[36:37] <- mi2
  81     SVSTATE[38:39] <- mo0
  82     SVSTATE[40:41] <- mo1
  83     # enable bit for RA RB RC RT EA/FRS
  84     SVSTATE[42:46] <- SVme
  85     # persistence bit (applies to more than one instruction)
  86     SVSTATE[62] <- pst
  87
  88 Special Registers Altered:
  89
  90     None
  91
  92 # svshape
  93
  94 SVM-Form
  95
  96 * svshape SVxd,SVyd,SVzd,SVrm,vf
  97
  98 Pseudo-code:
  99
 100     # for convenience, VL to be calculated and stored in SVSTATE
 101     vlen <- [0] * 7
 102     mscale[0:5] <- 0b000001 # for scaling MAXVL
 103     itercount[0:6] <- [0] * 7
 104     SVSTATE[0:31] <- [0] * 32
 105     # only overwrite REMAP if "persistence" is zero
 106     if (SVSTATE[62] = 0b0) then
 107         SVSTATE[32:33] <- 0b00
 108         SVSTATE[34:35] <- 0b00
 109         SVSTATE[36:37] <- 0b00
 110         SVSTATE[38:39] <- 0b00
 111         SVSTATE[40:41] <- 0b00
 112         SVSTATE[42:46] <- 0b00000
 113         SVSTATE[62] <- 0b0
 114         SVSTATE[63] <- 0b0
 115     # clear out all SVSHAPEs
 116     SVSHAPE0[0:31] <- [0] * 32
 117     SVSHAPE1[0:31] <- [0] * 32
 118     SVSHAPE2[0:31] <- [0] * 32
 119     SVSHAPE3[0:31] <- [0] * 32
 120     # set schedule up for multiply
 121     if (SVrm = 0b0000) then
 122         # VL in Matrix Multiply is xd*yd*zd
 123         xd <- (0b00 || SVxd) + 1
 124         yd <- (0b00 || SVyd) + 1
 125         zd <- (0b00 || SVzd) + 1
 126         n <- xd * yd * zd
 127         vlen[0:6] <- n[14:20]
 128         # set up template in SVSHAPE0, then copy to 1-3
 129         SVSHAPE0[0:5] <- (0b0 || SVxd)   # xdim
 130         SVSHAPE0[6:11] <- (0b0 || SVyd)   # ydim
 131         SVSHAPE0[12:17] <- (0b0 || SVzd)   # zdim
 132         SVSHAPE0[28:29] <- 0b11           # skip z
 133         # copy
 134         SVSHAPE1[0:31] <- SVSHAPE0[0:31]
 135         SVSHAPE2[0:31] <- SVSHAPE0[0:31]
 136         SVSHAPE3[0:31] <- SVSHAPE0[0:31]
 137         # set up FRA
 138         SVSHAPE1[18:20] <- 0b001          # permute x,z,y
 139         SVSHAPE1[28:29] <- 0b01           # skip z
 140         # FRC
 141         SVSHAPE2[18:20] <- 0b001          # permute x,z,y
 142         SVSHAPE2[28:29] <- 0b11           # skip y
 143     # set schedule up for FFT butterfly
 144     if (SVrm = 0b0001) then
 145         # calculate O(N log2 N)
 146         n <- [0] * 3
 147         do while n < 5
 148            if SVxd[4-n] = 0 then
 149                leave
 150            n <- n + 1
 151         n <- ((0b0 || SVxd) + 1) * n
 152         vlen[0:6] <- n[1:7]
 153         # set up template in SVSHAPE0, then copy to 1-3
 154         # for FRA and FRT
 155         SVSHAPE0[0:5] <- (0b0 || SVxd)   # xdim
 156         SVSHAPE0[12:17] <- (0b0 || SVzd)   # zdim - "striding" (2D FFT)
 157         mscale <- (0b0 || SVzd) + 1
 158         SVSHAPE0[30:31] <- 0b01          # Butterfly mode
 159         # copy
 160         SVSHAPE1[0:31] <- SVSHAPE0[0:31]
 161         SVSHAPE2[0:31] <- SVSHAPE0[0:31]
 162         # set up FRB and FRS
 163         SVSHAPE1[28:29] <- 0b01           # j+halfstep schedule
 164         # FRC (coefficients)
 165         SVSHAPE2[28:29] <- 0b10           # k schedule
 166     # set schedule up for (i)DCT Inner butterfly
 167     # SVrm Mode 4 (Mode 12 for iDCT) is for on-the-fly (Vertical-First Mode)
 168     if ((SVrm = 0b0100) |
 169         (SVrm = 0b1100)) then
 170         # calculate O(N log2 N)
 171         n <- [0] * 3
 172         do while n < 5
 173            if SVxd[4-n] = 0 then
 174                leave
 175            n <- n + 1
 176         n <- ((0b0 || SVxd) + 1) * n
 177         vlen[0:6] <- n[1:7]
 178         # set up template in SVSHAPE0, then copy to 1-3
 179         # set up FRB and FRS
 180         SVSHAPE0[0:5] <- (0b0 || SVxd)   # xdim
 181         SVSHAPE0[12:17] <- (0b0 || SVzd)   # zdim - "striding" (2D DCT)
 182         mscale <- (0b0 || SVzd) + 1
 183         if (SVrm = 0b1100) then
 184             SVSHAPE0[30:31] <- 0b11          # iDCT mode
 185             SVSHAPE0[18:20] <- 0b011         # iDCT Inner Butterfly sub-mode
 186         else
 187             SVSHAPE0[30:31] <- 0b01          # DCT mode
 188             SVSHAPE0[18:20] <- 0b001         # DCT Inner Butterfly sub-mode
 189             SVSHAPE0[21:23] <- 0b001         # "inverse" on outer loop
 190         SVSHAPE0[6:11] <- 0b000011       # (i)DCT Inner Butterfly mode 4
 191         # copy
 192         SVSHAPE1[0:31] <- SVSHAPE0[0:31]
 193         SVSHAPE2[0:31] <- SVSHAPE0[0:31]
 194         if (SVrm != 0b0100) & (SVrm != 0b1100) then
 195             SVSHAPE3[0:31] <- SVSHAPE0[0:31]
 196         # for FRA and FRT
 197         SVSHAPE0[28:29] <- 0b01           # j+halfstep schedule
 198         # for cos coefficient
 199         SVSHAPE2[28:29] <- 0b10           # ci (k for mode 4) schedule
 200         SVSHAPE2[12:17] <- 0b000000       # reset costable "striding" to 1
 201         if (SVrm != 0b0100) & (SVrm != 0b1100) then
 202             SVSHAPE3[28:29] <- 0b11           # size schedule
 203     # set schedule up for (i)DCT Outer butterfly
 204     if (SVrm = 0b0011) | (SVrm = 0b1011) then
 205         # calculate O(N log2 N) number of outer butterfly overlapping adds
 206         vlen[0:6] <- [0] * 7
 207         n <- 0b000
 208         size <- 0b0000001
 209         itercount[0:6] <- (0b00 || SVxd) + 0b0000001
 210         itercount[0:6] <- (0b0 || itercount[0:5])
 211         do while n < 5
 212            if SVxd[4-n] = 0 then
 213                leave
 214            n <- n + 1
 215            count <- (itercount - 0b0000001) * size
 216            vlen[0:6] <- vlen + count[7:13]
 217            size[0:6] <- (size[1:6] || 0b0)
 218            itercount[0:6] <- (0b0 || itercount[0:5])
 219         # set up template in SVSHAPE0, then copy to 1-3
 220         # set up FRB and FRS
 221         SVSHAPE0[0:5] <- (0b0 || SVxd)   # xdim
 222         SVSHAPE0[12:17] <- (0b0 || SVzd)   # zdim - "striding" (2D DCT)
 223         mscale <- (0b0 || SVzd) + 1
 224         if (SVrm = 0b1011) then
 225             SVSHAPE0[30:31] <- 0b11      # iDCT mode
 226             SVSHAPE0[18:20] <- 0b011     # iDCT Outer Butterfly sub-mode
 227             SVSHAPE0[21:23] <- 0b101     # "inverse" on outer and inner loop
 228         else
 229             SVSHAPE0[30:31] <- 0b01      # DCT mode
 230             SVSHAPE0[18:20] <- 0b100     # DCT Outer Butterfly sub-mode
 231         SVSHAPE0[6:11] <- 0b000010       # DCT Butterfly mode
 232         # copy
 233         SVSHAPE1[0:31] <- SVSHAPE0[0:31] # j+halfstep schedule
 234         SVSHAPE2[0:31] <- SVSHAPE0[0:31] # costable coefficients
 235         # for FRA and FRT
 236         SVSHAPE1[28:29] <- 0b01           # j+halfstep schedule
 237         # reset costable "striding" to 1
 238         SVSHAPE2[12:17] <- 0b000000
 239     # set schedule up for DCT COS table generation
 240     if (SVrm = 0b0101) | (SVrm = 0b1101) then
 241         # calculate O(N log2 N)
 242         vlen[0:6] <- [0] * 7
 243         itercount[0:6] <- (0b00 || SVxd) + 0b0000001
 244         itercount[0:6] <- (0b0 || itercount[0:5])
 245         n <- [0] * 3
 246         do while n < 5
 247            if SVxd[4-n] = 0 then
 248                leave
 249            n <- n + 1
 250            vlen[0:6] <- vlen + itercount
 251            itercount[0:6] <- (0b0 || itercount[0:5])
 252         # set up template in SVSHAPE0, then copy to 1-3
 253         # set up FRB and FRS
 254         SVSHAPE0[0:5] <- (0b0 || SVxd)   # xdim
 255         SVSHAPE0[12:17] <- (0b0 || SVzd)   # zdim - "striding" (2D DCT)
 256         mscale <- (0b0 || SVzd) + 1
 257         SVSHAPE0[30:31] <- 0b01          # DCT/FFT mode
 258         SVSHAPE0[6:11] <- 0b000100       # DCT Inner Butterfly COS-gen mode
 259         if (SVrm = 0b0101) then
 260             SVSHAPE0[21:23] <- 0b001     # "inverse" on outer loop for DCT
 261         # copy
 262         SVSHAPE1[0:31] <- SVSHAPE0[0:31]
 263         SVSHAPE2[0:31] <- SVSHAPE0[0:31]
 264         # for cos coefficient
 265         SVSHAPE1[28:29] <- 0b10           # ci schedule
 266         SVSHAPE2[28:29] <- 0b11           # size schedule
 267     # set schedule up for iDCT / DCT inverse of half-swapped ordering
 268     if (SVrm = 0b0110) | (SVrm = 0b1110) | (SVrm = 0b1111) then
 269         vlen[0:6] <- (0b00 || SVxd) + 0b0000001
 270         # set up template in SVSHAPE0
 271         SVSHAPE0[0:5] <- (0b0 || SVxd)   # xdim
 272         SVSHAPE0[12:17] <- (0b0 || SVzd)   # zdim - "striding" (2D DCT)
 273         mscale <- (0b0 || SVzd) + 1
 274         if (SVrm = 0b1110) then
 275             SVSHAPE0[18:20] <- 0b001     # DCT opposite half-swap
 276         if (SVrm = 0b1111) then
 277             SVSHAPE0[30:31] <- 0b01          # FFT mode
 278         else
 279             SVSHAPE0[30:31] <- 0b11          # DCT mode
 280         SVSHAPE0[6:11] <- 0b000101       # DCT "half-swap" mode
 281     # set schedule up for parallel reduction
 282     if (SVrm = 0b0111) then
 283         # calculate the total number of operations (brute-force)
 284         vlen[0:6] <- [0] * 7
 285         itercount[0:6] <- (0b00 || SVxd) + 0b0000001
 286         step[0:6] <- 0b0000001
 287         i[0:6] <- 0b0000000
 288         do while step <u itercount
 289             newstep <- step[1:6] || 0b0
 290             j[0:6] <- 0b0000000
 291             do while (j+step <u itercount)
 292                 j <- j + newstep
 293                 i <- i + 1
 294             step <- newstep
 295         # VL in Parallel-Reduce is the number of operations
 296         vlen[0:6] <- i
 297         # set up template in SVSHAPE0, then copy to 1. only 2 needed
 298         SVSHAPE0[0:5] <- (0b0 || SVxd)   # xdim
 299         SVSHAPE0[12:17] <- (0b0 || SVzd)   # zdim - "striding" (2D DCT)
 300         mscale <- (0b0 || SVzd) + 1
 301         SVSHAPE0[30:31] <- 0b10          # parallel reduce submode
 302         # copy
 303         SVSHAPE1[0:31] <- SVSHAPE0[0:31]
 304         # set up right operand (left operand 28:29 is zero)
 305         SVSHAPE1[28:29] <- 0b01           # right operand
 306     # set VL, MVL and Vertical-First
 307     m[0:12] <- vlen * mscale
 308     maxvl[0:6] <- m[6:12]
 309     SVSTATE[0:6] <- maxvl  # MAVXL
 310     SVSTATE[7:13] <- vlen  # VL
 311     SVSTATE[63] <- vf
 312
 313 Special Registers Altered:
 314
 315     None
 316
 317 # svindex
 318
 319 SVI-Form
 320
 321 * svindex SVG,rmm,SVd,ew,SVyx,mm,sk
 322
 323 Pseudo-code:
 324
 325     # based on nearest MAXVL compute other dimension
 326     MVL <- SVSTATE[0:6]
 327     d <- [0] * 6
 328     dim <- SVd+1
 329     do while d*dim <u ([0]*4 || MVL)
 330        d <- d + 1
 331     # set up template, then copy once location identified
 332     shape <- [0]*32
 333     shape[30:31] <- 0b00            # mode
 334     if SVyx = 0 then
 335         shape[18:20] <- 0b110       # indexed xd/yd
 336         shape[0:5] <- (0b0 || SVd)  # xdim
 337         if sk = 0 then shape[6:11] <- 0 # ydim
 338         else           shape[6:11] <- 0b111111 # ydim max
 339     else
 340         shape[18:20] <- 0b111       # indexed yd/xd
 341         if sk = 1 then shape[6:11] <- 0 # ydim
 342         else           shape[6:11] <- d-1 # ydim max
 343         shape[0:5] <- (0b0 || SVd) # ydim
 344     shape[12:17] <- (0b0 || SVG)        # SVGPR
 345     shape[28:29] <- ew                  # element-width override
 346     shape[21] <- sk                     # skip 1st dimension
 347     # select the mode for updating SVSHAPEs
 348     SVSTATE[62] <- mm # set or clear persistence
 349     if mm = 0 then
 350         # clear out all SVSHAPEs first
 351         SVSHAPE0[0:31] <- [0] * 32
 352         SVSHAPE1[0:31] <- [0] * 32
 353         SVSHAPE2[0:31] <- [0] * 32
 354         SVSHAPE3[0:31] <- [0] * 32
 355         SVSTATE[32:41] <- [0] * 10 # clear REMAP.mi/o
 356         SVSTATE[42:46] <- rmm # rmm exactly REMAP.SVme
 357         idx <- 0
 358         for bit = 0 to 4
 359             if rmm[4-bit] then
 360                 # activate requested shape
 361                 if idx = 0 then SVSHAPE0 <- shape
 362                 if idx = 1 then SVSHAPE1 <- shape
 363                 if idx = 2 then SVSHAPE2 <- shape
 364                 if idx = 3 then SVSHAPE3 <- shape
 365                 SVSTATE[bit*2+32:bit*2+33] <- idx
 366                 # increment shape index, modulo 4
 367                 if idx = 3 then idx <- 0
 368                 else            idx <- idx + 1
 369     else
 370         # refined SVSHAPE/REMAP update mode
 371         bit <- rmm[0:2]
 372         idx <- rmm[3:4]
 373         if idx = 0 then SVSHAPE0 <- shape
 374         if idx = 1 then SVSHAPE1 <- shape
 375         if idx = 2 then SVSHAPE2 <- shape
 376         if idx = 3 then SVSHAPE3 <- shape
 377         SVSTATE[bit*2+32:bit*2+33] <- idx
 378         SVSTATE[46-bit] <- 1
 379
 380 Special Registers Altered:
 381
 382     None
 383
 384 # svshape2
 385
 386 SVM2-Form
 387
 388 * svshape2 SVo,SVyx,rmm,SVd,sk,mm
 389
 390 Pseudo-code:
 391
 392     # based on nearest MAXVL compute other dimension
 393     MVL <- SVSTATE[0:6]
 394     d <- [0] * 6
 395     dim <- SVd+1
 396     do while d*dim <u ([0]*4 || MVL)
 397        d <- d + 1
 398     # set up template, then copy once location identified
 399     shape <- [0]*32
 400     shape[30:31] <- 0b00            # mode
 401     shape[0:5] <- (0b0 || SVd)      # x/ydim
 402     if SVyx = 0 then
 403         shape[18:20] <- 0b000       # ordering xd/yd(/zd)
 404         if sk = 0 then shape[6:11] <- 0 # ydim
 405         else           shape[6:11] <- 0b111111 # ydim max
 406     else
 407         shape[18:20] <- 0b010       # ordering yd/xd(/zd)
 408         if sk = 1 then shape[6:11] <- 0 # ydim
 409         else           shape[6:11] <- d-1 # ydim max
 410     # offset (the prime purpose of this instruction)
 411     shape[24:27] <- SVo         # offset
 412     if sk = 1 then shape[28:29] <- 0b01 # skip 1st dimension
 413     else           shape[28:29] <- 0b00 # no skipping
 414     # select the mode for updating SVSHAPEs
 415     SVSTATE[62] <- mm # set or clear persistence
 416     if mm = 0 then
 417         # clear out all SVSHAPEs first
 418         SVSHAPE0[0:31] <- [0] * 32
 419         SVSHAPE1[0:31] <- [0] * 32
 420         SVSHAPE2[0:31] <- [0] * 32
 421         SVSHAPE3[0:31] <- [0] * 32
 422         SVSTATE[32:41] <- [0] * 10 # clear REMAP.mi/o
 423         SVSTATE[42:46] <- rmm # rmm exactly REMAP.SVme
 424         idx <- 0
 425         for bit = 0 to 4
 426             if rmm[4-bit] then
 427                 # activate requested shape
 428                 if idx = 0 then SVSHAPE0 <- shape
 429                 if idx = 1 then SVSHAPE1 <- shape
 430                 if idx = 2 then SVSHAPE2 <- shape
 431                 if idx = 3 then SVSHAPE3 <- shape
 432                 SVSTATE[bit*2+32:bit*2+33] <- idx
 433                 # increment shape index, modulo 4
 434                 if idx = 3 then idx <- 0
 435                 else            idx <- idx + 1
 436     else
 437         # refined SVSHAPE/REMAP update mode
 438         bit <- rmm[0:2]
 439         idx <- rmm[3:4]
 440         if idx = 0 then SVSHAPE0 <- shape
 441         if idx = 1 then SVSHAPE1 <- shape
 442         if idx = 2 then SVSHAPE2 <- shape
 443         if idx = 3 then SVSHAPE3 <- shape
 444         SVSTATE[bit*2+32:bit*2+33] <- idx
 445         SVSTATE[46-bit] <- 1
 446
 447 Special Registers Altered:
 448
 449     None
 450