Pseudo-code:
- # VL in Matrix Multiply is xd*yd*zd
+ # based on MAXVL compute other dimension
+ MVL <- SVSTATE[0:6]
+ d <- [0] * 6
+ dim <- SVd+1
+ do while d*dim < ([0]*4 || MVL)
+ d <- d + 1
# set up template, then copy once location identified
shape <- [0]*32
shape[30:31] <- 0b00 # mode
else shape[6:11] <- 0b111111 # ydim max
else
shape[18:20] <- 0b111 # indexed yd/xd
- if sk = 0 then shape[0:5] <- 0 # xdim
- else shape[0:5] <- 0b111111 # xdim max
- shape[6:11] <- (0b0 || SVd) # ydim
+ if sk = 1 then shape[6:11] <- 0 # ydim
+ else shape[6:11] <- d-1 # ydim max
+ shape[0:5] <- (0b0 || SVd) # ydim
shape[12:17] <- (0b0 || SVG) # SVGPR
shape[28:29] <- ew # element-width override
if sk = 1 then shape[28:29] <- 0b01 # skip 1st dimension
self.assertEqual(shape.svgpr, 0)
self._check_regs(sim, expected_regs)
+ def test_2_sv_index_add(self):
+ """sets VL=6 (via SVSTATE) then does 2D remapped svindex, and an add.
+
+ dim=3,yx=1
+ only RA is re-mapped via Indexing, not RB or RT
+ """
+ isa = SVP64Asm(['svindex 8, 1, 3, 0, 1, 0, 0',
+ 'sv.add *8, *0, *0',
+ ])
+ lst = list(isa)
+ print ("listing", lst)
+
+ # initial values in GPR regfile
+ initial_regs = [0] * 32
+ idxs = [1, 0, 5, 2, 4, 3] # random enough
+ for i in range(6):
+ initial_regs[16+i] = idxs[i]
+ initial_regs[i] = i
+
+ # SVSTATE vl=10
+ svstate = SVP64State()
+ svstate.vl = 6 # VL
+ svstate.maxvl = 6 # MAXVL
+ print ("SVSTATE", bin(svstate.asint()))
+
+ # copy before running
+ expected_regs = deepcopy(initial_regs)
+ for i in range(6):
+ xi = i % 3
+ yi = i // 3
+ remap = yi+xi*2
+ RA = initial_regs[0+idxs[remap]] # modulo 3 but still indexed
+ RB = initial_regs[0+i]
+ expected_regs[i+8] = RA+RB
+ print ("expected", i, expected_regs[i+8])
+
+ with Program(lst, bigendian=False) as program:
+ sim = self.run_tst_program(program, initial_regs, svstate=svstate)
+
+ print (sim.spr)
+ SVSHAPE0 = sim.spr['SVSHAPE0']
+ print ("SVSTATE after", bin(sim.svstate.asint()))
+ print (" vl", bin(sim.svstate.vl))
+ print (" mvl", bin(sim.svstate.maxvl))
+ print (" srcstep", bin(sim.svstate.srcstep))
+ print (" dststep", bin(sim.svstate.dststep))
+ print (" RMpst", bin(sim.svstate.RMpst))
+ print (" SVme", bin(sim.svstate.SVme))
+ print (" mo0", bin(sim.svstate.mo0))
+ print (" mo1", bin(sim.svstate.mo1))
+ print (" mi0", bin(sim.svstate.mi0))
+ print (" mi1", bin(sim.svstate.mi1))
+ print (" mi2", bin(sim.svstate.mi2))
+ print ("STATE0svgpr", hex(SVSHAPE0.svgpr))
+ print ("STATE0 xdim", SVSHAPE0.xdimsz)
+ print ("STATE0 ydim", SVSHAPE0.ydimsz)
+ print ("STATE0 skip", bin(SVSHAPE0.skip))
+ print ("STATE0 inv", SVSHAPE0.invxyz)
+ print ("STATE0order", SVSHAPE0.order)
+ print (sim.gpr.dump())
+ self.assertEqual(sim.svstate.RMpst, 0) # mm=0 so persist=0
+ self.assertEqual(sim.svstate.SVme, 0b00001) # same as rmm
+ # rmm is 0b00001 which means mi0=0 and all others inactive (0)
+ self.assertEqual(sim.svstate.mi0, 0)
+ self.assertEqual(sim.svstate.mi1, 0)
+ self.assertEqual(sim.svstate.mi2, 0)
+ self.assertEqual(sim.svstate.mo0, 0)
+ self.assertEqual(sim.svstate.mo1, 0)
+ self.assertEqual(SVSHAPE0.svgpr, 16) # SVG is shifted up by 1
+ for i in range(1,4):
+ shape = sim.spr['SVSHAPE%d' % i]
+ self.assertEqual(shape.svgpr, 0)
+ self._check_regs(sim, expected_regs)
+
def run_tst_program(self, prog, initial_regs=None,
svstate=None):
if initial_regs is None: