shape[30:31] <- 0b11 # triangle/bigmul mode
shape[28:29] <- sm # mode
shape[24:27] <- SVo # offset
- shape[23] <- SVyx # swap x/y priority
- shape[22] <- sk # skipping
+ shape[22:23] <- SVyx # swap x/y priority
+ shape[21] <- sk # skipping
shape[19:20] <- sm2 # submode2
shape[18] <- 0 # reserved
shape[12:17] <- 1 # zdim
- shape[6:11] <- (0b0 || SVyd) # xdim
+ shape[6:11] <- (0b0 || SVyd) # ydim
shape[0:5] <- (0b0 || SVxd) # xdim
# activate requested shape
if si = 0 then SVSHAPE0 <- shape
# 1.6.29.1 SVI2-FORM
- |0 |6 |9 |11 |16 |21 |22 |24 |26 |27 31|
- |PO |SVo |si |SVyd|SVxd|sk |sm2 |sm |SVyx|XO |
+ |0 |6 |9 |11 |16 |21 |22 |24 |26 |28 |29 |31 |
+ |PO |SVo |si |SVyd|SVxd|sk |sm2 |sm |XO |SVyx|XO |SVyx|
# 1.6.30 SVL-FORM
|0 |6 |11 |16 |23 |24 |25 |26 |31 |
Immediate field used to specify a 5-bit signed inte-
ger.
Formats: VX
+ sk (21)
+ Field used to specify dimensional skipping in svindex
+ Formats: SVI2
sk (25)
Field used to specify dimensional skipping in svindex
- Formats: SVI, SVM2, SVI2
+ Formats: SVI, SVM2
sm (24:25)
submode for svindex4
Formats: SVI2
XO (26:27)
Extended opcode field.
Formats: XX4
+ XO (26:27,29:30)
+ Extended opcode field.
+ Formats: SVI2
XO (26:30)
Extended opcode field.
Formats: A, DX, VA2, SVL, CRB, DCT
Formats: MDS
XO (27:31)
Extended opcode field.
- Formats: BM2, SVI2
+ Formats: BM2
XO (29:31)
Extended opcode field.
Formats: DQ
Formats: DQE, DS, SC
SVyx (23)
Field used to specify loop dimension order in svindex
- Formats: SVI, SVI2
+ Formats: SVI
+ SVyx (28,31)
+ Field used to specify loop dimension order in svindex
+ Formats: SVI2
SVyx (10)
Field used to specify loop dimension order in svshape2
Formats: SVM2
100--011001,VL,OP_SVSHAPE,NONE,NONE,NONE,NONE,NONE,NONE,0,0,ZERO,0,NONE,0,0,0,0,0,0,NONE,0,0,svshape2,SVM2,,1,unofficial until submitted and approved/renumbered by the opf isa wg
# svshape3 and svshape4, none of the horribleness of shoe-horning svshape2 in it
------10000,VL,OP_SVSHAPE,NONE,NONE,NONE,NONE,NONE,NONE,0,0,ZERO,0,NONE,0,0,0,0,0,0,NONE,0,0,svshape3,SVM,,1,unofficial until submitted and approved/renumbered by the opf isa wg
-------11000,VL,OP_SVSHAPE,NONE,NONE,NONE,NONE,NONE,NONE,0,0,ZERO,0,NONE,0,0,0,0,0,0,NONE,0,0,svshape4,SVI2,,1,unofficial until submitted and approved/renumbered by the opf isa wg
# A/V bitmanip
1101001110-,ALU,OP_AVGADD,RA,RB,NONE,RT,NONE,CR0,0,0,ZERO,0,NONE,0,0,0,0,0,0,RC_ONLY,0,0,avgadd,X,,1,unofficial until submitted and approved/renumbered by the opf isa wg
--01101110-,ALU,OP_SHADD,RA,RB,NONE,RT,NONE,CR0,0,0,ZERO,0,NONE,0,0,0,0,0,0,RC_ONLY,0,0,sadd,Z23,,1,unofficial until submitted and approved/renumbered by the opf isa wg
opcode,unit,internal op,in1,in2,in3,out,CR in,CR out,inv A,inv out,cry in,cry out,ldst len,BR,sgn ext,upd,rsrv,32b,sgn,rc,lk,sgl pipe,comment,form,CONDITIONS,unofficial,comment2
--------00-,SHIFT_ROT,OP_TERNLOG,RA,RB,RT,RT,NONE,CR0,0,0,ZERO,0,NONE,0,0,0,0,0,0,RC_ONLY,0,0,ternlogi,TLI,,1,unofficial until submitted and approved/renumbered by the opf isa wg
+-----01-01-,VL,OP_SVSHAPE,NONE,NONE,NONE,NONE,NONE,NONE,0,0,ZERO,0,NONE,0,0,0,0,0,0,NONE,0,0,svshape4,SVI2,,1,unofficial until submitted and approved/renumbered by the opf isa wg
s[1] = 1<<51
r[1] = 1<<51
check(r, s)
- for i in range(100000):
+ for i in range(1):
r, s = [0]*5, [0]*5
for j in range(5):
r[j] = random.randint(0, 1<<50)
# the indices are generated by using natural-looking nested loops
def iterate_indices(SVSHAPE, VL=None):
# establish if this is triangular mode
- triangle = SVSHAPE.mode == 0b11 and SVSHAPE.submode & 0b10 == 0b10
- bigmul = SVSHAPE.mode == 0b11 and SVSHAPE.submode & 0b01 == 0b01
+ triangle = SVSHAPE.mode == 0b11 and SVSHAPE._skip & 0b10 == 0b10
+ bigmul = SVSHAPE.mode == 0b11 and SVSHAPE._skip & 0b01 == 0b01
in_offs = VL is not None and SVSHAPE.mode == 0b11
# get indices to iterate over, in the required order
- xd = SVSHAPE.lims[0]
- yd = SVSHAPE.lims[1]
- zd = SVSHAPE.lims[2]
+ if SVSHAPE.mode == 0b11: # bigmul/triangle has no Z-dim
+ xd = 1
+ yd = SVSHAPE.lims[0]
+ zd = SVSHAPE.lims[1]
+ else:
+ xd = SVSHAPE.lims[0]
+ yd = SVSHAPE.lims[1]
+ zd = SVSHAPE.lims[2]
# create lists of indices to iterate over in each dimension
_x_r = list(range(xd))
_y_r = list(range(yd))
if SVSHAPE.invxyz[2]: z_r.reverse()
# start an infinite (wrapping) loop
step = 0 # track src/dst step
+ print("iterate_indices", VL, triangle, bigmul, in_offs, "dims", xd, yd, zd,
+ "mode", bin(SVSHAPE.mode), "skip", SVSHAPE.skip,
+ "inv", SVSHAPE.invxyz)
while True:
for zi, z in enumerate(z_r): # loop over 1st order dimension
z_end = z == z_r[-1]
# ok work out which order to construct things in.
# start by creating a list of tuples of the dimension
# and its limit
- vals = [(SVSHAPE.lims[0], x, "x"),
- (SVSHAPE.lims[1], y, "y"),
- (SVSHAPE.lims[2], z, "z")
+ vals = [(xd, x, "x"),
+ (yd, y, "y"),
+ (zd, z, "z")
]
# now select those by order. this allows us to
# create schedules for [z][x], [x][y], or [y][z]
break
triangle = SVSHAPE.mode == 0b11 and SVSHAPE.submode & 0b10 == 0b10
if triangle:
- if SVSHAPE.lims[0] == 1: end &= ~0b001
+ if SVSHAPE.lims[0] == 1: end &= ~0b010
if triangle:
- if SVSHAPE.lims[1] == 1: end &= ~0b010
+ if SVSHAPE.lims[1] == 1: end &= ~0b100
+ if triangle:
+ if SVSHAPE.lims[2] == 1: end &= ~0b001
if end != 0:
print()
SVSHAPE0.order = [0,2,1] # experiment with different permutations, here
SVSHAPE0.mode = 0b00
SVSHAPE0.skip = 0b00
+ SVSHAPE0._skip = 0b00
SVSHAPE0.submode2 = 0b000 #
SVSHAPE0.offset = 0 # experiment with different offset, here
SVSHAPE0.invxyz = [0,0,0] # inversion if desired
SVSHAPE0.order = [0,1,2] # experiment with different permutations, here
SVSHAPE0.mode = 0b00
SVSHAPE0.skip = 0b10
+ SVSHAPE0._skip = 0b10
SVSHAPE0.submode2 = 0b000 #
SVSHAPE0.offset = 0 # experiment with different offset, here
SVSHAPE0.invxyz = [0,0,0] # inversion if desired
SVSHAPE0.order = [0,1,2] # experiment with different permutations, here
SVSHAPE0.mode = 0b00
SVSHAPE0.skip = 0b01
+ SVSHAPE0._skip = 0b01
SVSHAPE0.submode2 = 0b000 #
SVSHAPE0.offset = 0 # experiment with different offset, here
SVSHAPE0.invxyz = [0,0,0] # inversion if desired
SVSHAPE0.order = [1,0,2] # experiment with different permutations, here
SVSHAPE0.mode = 0b00
SVSHAPE0.skip = 0b00
+ SVSHAPE0._skip = 0b00
SVSHAPE0.submode2 = 0b000 #
SVSHAPE0.offset = 0 # experiment with different offset, here
SVSHAPE0.invxyz = [0,0,0] # inversion if desired
SVSHAPE0.submode = 0b01 # rhombus mode
SVSHAPE0.submode2 = 0b000 #
SVSHAPE0.skip = 0b00
+ SVSHAPE0._skip = 0b00
SVSHAPE0.offset = 0 # experiment with different offset, here
SVSHAPE0.invxyz = [0,0,0] # inversion if desired
dump_shape(SVSHAPE0)
# triangle
- xdim, ydim, zdim = 1, 5, 5 # set the dimension sizes here
+ xdim, ydim, zdim = 5, 5, 1 # set the dimension sizes here
print("\ntriangle", xdim, ydim, zdim)
SVSHAPE0 = SVSHAPE()
SVSHAPE0.lims = [xdim, ydim, zdim]
SVSHAPE0.submode = 0b10 # triangle mode
SVSHAPE0.submode2 = 0b000 #
SVSHAPE0.skip = 0b11
+ SVSHAPE0._skip = 0b11
SVSHAPE0.offset = 0 # experiment with different offset, here
SVSHAPE0.invxyz = [0,0,0] # inversion if desired
# triangle
print("\ntriangle-bigmul", xdim, ydim, zdim)
- xdim, ydim, zdim = 1, 3, 5 # set the dimension sizes here
+ xdim, ydim, zdim = 3, 5, 1 # set the dimension sizes here
SVSHAPE0 = SVSHAPE()
SVSHAPE0.lims = [xdim, ydim, zdim]
SVSHAPE0.order = [0,1,2] # experiment with different permutations, here
SVSHAPE0.submode = 0b11 # triangle mode, with sum
SVSHAPE0.submode2 = 0b000 #
SVSHAPE0.skip = 0b00
+ SVSHAPE0._skip = 0b00
SVSHAPE0.offset = 0 # experiment with different offset, here
SVSHAPE0.invxyz = [0,0,0] # inversion if desired
SVSHAPE0.mode = 0b11
SVSHAPE0.submode = 0b00 #
SVSHAPE0.skip = 0b00
+ SVSHAPE0._skip = 0b00
SVSHAPE0.submode2 = 0b000 # cols or rows modulo
SVSHAPE0.offset = 1 # experiment with different offset, here
SVSHAPE0.invxyz = [0,0,0] # inversion if desired
fs = tuple(range(offs, end))
v = FieldSelectableInt(self, fs)
self.fsi[field] = v
- log("SVSHAPE setup field", field, offs, end)
+ log("SVSHAPE setup field", field, offs, end, v.asint(msb0=True))
offs = end
def copy(self):
"REMAP Triangle Mode"
return self.mode == 0b11 and self._skip in [0b10, 0b11]
+ def is_bigmul(self):
+ "REMAP Bigmul (Modulo) Mode"
+ return self.mode == 0b11 and self._skip in [0b01, 0b11]
+
@property
def submode2(self):
return self.fsi['permute'].asint(msb0=True)
@property
def zdimsz(self):
- z = self.fsi['zdimsz'].asint(msb0=True)+1
+ if self.mode == 0b11: # no 3rd dimension in bigmul/triangle
+ return 1
if self.is_indexed():
- z = 1 # no z dimension when indexed
- return z
+ return 1 # no z dimension when indexed
+ return self.fsi['zdimsz'].asint(msb0=True)+1
@zdimsz.setter
def zdimsz(self, value):
return remap
def get_iterator(self):
- log ("SVSHAPE get_iterator", self.mode, self.ydimsz, self.is_indexed())
+ log ("SVSHAPE get_iterator",
+ self.mode, self.ydimsz, "index", self.is_indexed(),
+ "triangle", self.is_triangle())
if self.mode == 0b00:
iterate_fn = iterate_indices
elif self.mode == 0b10:
iterate_fn = iterate_preduce_indices # parallel-reduce
elif self.mode in [0b01, 0b11]:
# further sub-selection
- if self.ydimsz == 1:
+ if self.mode == 0b11:
+ iterate_fn = iterate_indices
+ elif self.ydimsz == 1:
iterate_fn = iterate_butterfly_indices
elif self.ydimsz in [2, 4]:
iterate_fn = iterate_dct_inner_butterfly_indices
iterate_fn = iterate_dct_inner_costable_indices
elif self.ydimsz in [6, 14, 15]:
iterate_fn = iterate_dct_inner_halfswap_loadstore
+ log ("SVSHAPE get_iterator", iterate_fn)
# create a **NEW** iterator each time this is called
return iterate_fn(self.copy())
projects / openpower-isa.git / commitdiff