You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
314 lines
18 KiB
314 lines
18 KiB
from __future__ import annotations
|
|
import functools, operator, itertools
|
|
from dataclasses import dataclass
|
|
from typing import Optional, cast, Sequence
|
|
from tinygrad.dtype import dtypes
|
|
from tinygrad.ops import resolve, UOp, Variable, sint, sym_infer, smax, smin, sint_to_uop
|
|
from tinygrad.helpers import prod, all_int, argsort, flatten, ceildiv
|
|
|
|
@functools.lru_cache(maxsize=None)
|
|
def canonicalize_strides(shape:tuple[sint, ...], strides:tuple[sint, ...]) -> tuple[sint, ...]:
|
|
return tuple(0 if s == 1 else st for s, st in zip(shape, strides))
|
|
|
|
@functools.lru_cache(maxsize=None)
|
|
def strides_for_shape(shape:tuple[sint, ...]) -> tuple[sint, ...]:
|
|
if not shape: return ()
|
|
strides = tuple(itertools.accumulate(reversed(shape[1:]), operator.mul, initial=1))[::-1]
|
|
return canonicalize_strides(shape, strides)
|
|
|
|
@functools.lru_cache(maxsize=None)
|
|
def merge_dims(shape:tuple[int, ...], strides:tuple[int, ...], mask:Optional[tuple[tuple[int, int], ...]]=None) -> tuple[tuple[int, int, int], ...]:
|
|
# merge contiguous sub-parts or zero strided dims
|
|
# any stride 0, masked from dim=1, or contiguous part is merged into next dim.
|
|
# stride != 0 to stride == 0 starts a new merging block
|
|
# ret = tuple[(merged_size, stride, merged size w/o zero stride), ...]
|
|
if not shape: return ()
|
|
assert len(shape) == len(strides) and (mask is None or len(shape) == len(mask))
|
|
ret = [(shape[0], strides[0], shape[0] if strides[0] != 0 else 0)]
|
|
# merge this dim to next dim if size is 1
|
|
merging = (mask[0][1] - mask[0][0] == 1) if mask is not None else shape[0] == 1
|
|
for i, (s, st) in enumerate(zip(shape[1:], strides[1:]), start=1):
|
|
# always merge 1
|
|
if s == 1: continue
|
|
last_s, last_st, last_pre_expand_s = ret[-1]
|
|
# merge last dim with this dim if merging or strides matched
|
|
if merging or last_st == s * st: ret[-1] = (last_s * s, st, (s if merging else last_pre_expand_s * s))
|
|
else: ret.append((s, st, s))
|
|
# merge this dim to next dim if size is 1
|
|
merging = (mask[i][1] - mask[i][0] == 1) if mask is not None else s == 1
|
|
return tuple(ret)
|
|
|
|
@functools.lru_cache(maxsize=None)
|
|
def _reshape_mask(_mask:Optional[tuple[tuple[sint, sint], ...]], old_shape:tuple[sint, ...], new_shape:tuple[sint, ...]) \
|
|
-> Optional[tuple[tuple[sint, sint], ...]]:
|
|
"""Returns the new mask if reshape is possible, and None if not possible."""
|
|
if _mask is None: return tuple((0, s) for s in new_shape)
|
|
if not all_int(flatten(_mask)): return None
|
|
|
|
new_mask: list[tuple[int, int]] = []
|
|
# _mask is all int here
|
|
r_masks, r_shape, r_new_shape = reversed(cast(tuple[tuple[int, int], ...], _mask)), reversed(old_shape), reversed(new_shape)
|
|
curr_stride, old_dim, new_dim, mask = 1, next(r_shape, 1), next(r_new_shape, 1), next(r_masks, (0,1))
|
|
|
|
while len(new_mask) < len(new_shape):
|
|
(l, r), next_stride = mask, new_dim * curr_stride
|
|
|
|
# need to split mask
|
|
if old_dim == next_stride: # simply copy the mask and get next batch for merging
|
|
new_mask.append((l // curr_stride, (r - 1) // curr_stride + 1))
|
|
curr_stride, old_dim, new_dim, mask = 1, next(r_shape, 1), next(r_new_shape, 1), next(r_masks, (0,1))
|
|
elif old_dim > next_stride: # mask can only be splitted if reshape doesn't cut across the mask.
|
|
if old_dim % next_stride != 0: return None
|
|
if (l % next_stride != 0 or r % next_stride != 0) and l // next_stride != (r - 1) // next_stride: return None
|
|
new_mask.append((l % next_stride // curr_stride, (r - 1) % next_stride // curr_stride + 1))
|
|
curr_stride, new_dim = next_stride, next(r_new_shape, 1) # need to get mask for next dimension
|
|
else:
|
|
next_mask = next(r_masks, (0, 1))
|
|
# combine if the mask can unfold continuously
|
|
if mask != (0, old_dim) and l != r and next_mask[1] - next_mask[0] != 1: return None
|
|
mask, old_dim = (next_mask[0] * old_dim + l, (next_mask[1] - 1) * old_dim + r), old_dim * next(r_shape, 1)
|
|
|
|
return tuple(reversed(new_mask))
|
|
|
|
def unravel(shape:tuple[sint, ...], offset:sint) -> list[sint]:
|
|
# find the position of offset on each dimension based on shape
|
|
# similar to unravel_index in numpy/torch
|
|
acc, idxs = 1, []
|
|
for d in reversed(shape):
|
|
idxs.append((offset//acc)%d)
|
|
acc *= d
|
|
return idxs[::-1]
|
|
|
|
@dataclass(frozen=True)
|
|
class View:
|
|
shape:tuple[sint, ...]
|
|
strides:tuple[sint, ...]
|
|
offset:sint
|
|
mask:Optional[tuple[tuple[sint, sint], ...]]
|
|
contiguous:bool
|
|
|
|
def to_indexed_uops(self:View, idxs:Optional[Sequence[UOp]]=None, vexpr:UOp=UOp.const(dtypes.bool, True)) -> tuple[UOp, UOp]:
|
|
"""(idx, valid)"""
|
|
if idxs is None: idxs = [UOp.range(dtypes.int, 0, s, i) for i,s in enumerate(self.shape)]
|
|
iexpr = sint_to_uop(self.offset)
|
|
for idx,sh,st,m in zip(idxs, self.shape, self.strides, self.mask if self.mask is not None else itertools.repeat(None)):
|
|
if resolve(sh != 1) and resolve(st != 0): iexpr = iexpr + idx*st
|
|
if m is not None:
|
|
if resolve(m[0] != 0): vexpr = vexpr * (idx >= m[0])
|
|
if resolve(m[1] != sh): vexpr = vexpr * (idx < m[1])
|
|
return iexpr, vexpr
|
|
|
|
@functools.lru_cache(maxsize=None) # pylint: disable=method-cache-max-size-none
|
|
def size(self) -> int:
|
|
ret = prod([x.vmax if isinstance(x, UOp) else x for x in self.shape])
|
|
assert isinstance(ret, int), f"{ret=} is not int"
|
|
return ret
|
|
|
|
@staticmethod
|
|
@functools.lru_cache(maxsize=None)
|
|
def create(shape:tuple[sint, ...], strides:Optional[tuple[sint, ...]]=None, offset:sint=0, mask:Optional[tuple[tuple[sint, sint], ...]]=None):
|
|
if not all(s >= 0 for s in shape): raise ValueError(f"Trying to create View with negative dimension: {shape=}")
|
|
strides = canonicalize_strides(shape, strides) if strides else strides_for_shape(shape)
|
|
# canonicalize 0 in shape
|
|
if 0 in shape: return View(shape, (0,) * len(shape), offset=0, mask=None, contiguous=True)
|
|
# canonicalize no-op mask
|
|
if mask is not None and all(m == (0,s) for m,s in zip(mask, shape)): mask = None
|
|
# if any dimension has size >1, but is masked such that only one index in the dimension is unmasked
|
|
# then its stride can also be set to 0, albeit with a corresponding adjustment required to the offset
|
|
if mask and any(elim := [not resolve(b+1 < e) for b,e in mask]):
|
|
if any(not resolve(b < e) for b,e in mask):
|
|
strides, offset, mask = (0,) * len(shape), 0, ((0,0),) * len(shape)
|
|
offset += sum((strides[i] * mask[i][0]) if e else 0 for i, e in enumerate(elim))
|
|
strides = tuple(0 if e else st for st,e in zip(strides, elim))
|
|
# simplify as we go
|
|
if isinstance(offset, UOp): offset = cast(sint, offset.ssimplify())
|
|
shape = tuple(cast(sint, x.ssimplify()) if isinstance(x, UOp) else x for x in shape)
|
|
# TODO: enabling stride simplification breaks symbolic jit
|
|
"""
|
|
strides = tuple(x.ssimplify() if isinstance(x, UOp) else x for x in strides)
|
|
if mask: mask = tuple((s.ssimplify() if isinstance(s, UOp) else s, e.ssimplify() if isinstance(e, UOp) else e) for s,e in mask)
|
|
"""
|
|
contiguous = offset == 0 and mask is None and strides == strides_for_shape(shape)
|
|
return View(shape, strides, offset, mask, contiguous)
|
|
|
|
@functools.lru_cache(None) # pylint: disable=method-cache-max-size-none
|
|
def vars(self) -> set[Variable]:
|
|
flatten_mask = tuple(x for m in self.mask for x in m) if self.mask is not None else tuple()
|
|
return functools.reduce(operator.or_, [x.vars() for x in self.shape+self.strides+(self.offset,)+flatten_mask if isinstance(x, UOp)], set())
|
|
|
|
@functools.lru_cache(None) # pylint: disable=method-cache-max-size-none
|
|
def unbind(self) -> tuple[View, dict[Variable, int]]:
|
|
var_unboundvar_val = [(v, v.unbind()) for v in self.vars()]
|
|
unbound_vars = {v:uv for v,(uv,_) in var_unboundvar_val}
|
|
def substitute(x:sint): return x if isinstance(x, int) else x.substitute(unbound_vars)
|
|
new_shape = tuple(map(substitute, self.shape))
|
|
new_strides = tuple(map(substitute, self.strides))
|
|
new_offset = substitute(self.offset)
|
|
new_mask = tuple((substitute(x[0]), substitute(x[1])) for x in self.mask) if self.mask is not None else None
|
|
return View.create(new_shape, new_strides, new_offset, new_mask), dict(x[1] for x in var_unboundvar_val)
|
|
|
|
@functools.lru_cache(maxsize=None) # pylint: disable=method-cache-max-size-none
|
|
def __add__(self, vm1:View) -> Optional[View]:
|
|
vm2 = self
|
|
if vm2.contiguous: return vm1
|
|
if vm1.contiguous and vm1.shape == vm2.shape: return vm2
|
|
if vm1.contiguous and vm1.size() == vm2.size() and (ret := vm2.reshape(vm1.shape)) is not None: return ret
|
|
if vm1.mask:
|
|
if (new_vm1 := vm1.shrink(vm1.mask)) == vm1 or (merged := vm2 + new_vm1) is None: return None
|
|
return merged.pad(tuple((b,s-e) for (b,e),s in zip(vm1.mask, vm1.shape)))
|
|
if not all_int(vm1.shape): return None
|
|
|
|
# Project vm1's offset and strides on to vm2.
|
|
origin = unravel(vm2.shape, vm1.offset)
|
|
terms: list[list[tuple[int, sint]]] = [[] for _ in vm2.shape]
|
|
strides: list[sint] = [0] * len(vm1.shape)
|
|
for d1, st in enumerate(vm1.strides):
|
|
if st == 0: continue
|
|
for d2, (o, s1) in enumerate(zip(origin, unravel(vm2.shape, vm1.offset + st))):
|
|
if (s1 := s1 - o) == 0: continue
|
|
terms[d2].append((d1, s1))
|
|
strides[d1] += s1 * vm2.strides[d2]
|
|
|
|
# Merge dimensions in vm2 if required.
|
|
# NB: Merging too many dimensions can make it difficult to project vm2's mask, hence only combining when required.
|
|
idxs: list[UOp] = [UOp.variable(f"idx{i}", 0, s-1) for i,s in enumerate(vm1.shape)]
|
|
merged_size, merged_term = 1, UOp.const(dtypes.int, 0)
|
|
extents: list[tuple[sint, UOp]] = []
|
|
for term, s, o in zip(reversed(terms), reversed(vm2.shape), reversed(origin)):
|
|
merged_term += (sum([idxs[d1] * s1 for d1, s1 in term]) + o) * merged_size
|
|
merged_size *= s
|
|
if resolve(merged_term < merged_size, False) and resolve(0 <= merged_term, False):
|
|
extents.append((merged_size, merged_term))
|
|
merged_size, merged_term = 1, UOp.const(dtypes.int, 0)
|
|
if resolve(merged_term != 0): return None
|
|
if (vm2_shape := tuple(s for s,_ in reversed(extents))) != vm2.shape:
|
|
if (reshaped_vm2 := vm2.reshape(vm2_shape)) is None: return None
|
|
# NOTE: this != to prevent infinite loop
|
|
if reshaped_vm2.shape != vm2.shape: return reshaped_vm2 + vm1
|
|
|
|
if vm2.mask:
|
|
# Try to project vm2's mask on to vm1.
|
|
newb, newe, bad = [0] * len(vm1.shape), list(vm1.shape), False
|
|
for (b, e), o, term, (_, t) in zip(vm2.mask, origin, terms, reversed(extents)):
|
|
if resolve(b <= t.vmin and t.vmax < e, False): continue
|
|
if len(term) != 1:
|
|
if not term and newe: newe[0] = 0
|
|
else: bad = True
|
|
continue
|
|
d1, s1 = term[0]
|
|
newb[d1] = max(newb[d1], ceildiv(b - o if s1 > 0 else e - o - 1, s1))
|
|
newe[d1] = min(newe[d1], (b - o if s1 < 0 else e - o - 1) // s1 + 1)
|
|
|
|
# If any of vm1 was masked off, try again with that mask in place.
|
|
if any((b, e) != (0, s) for b, e, s in zip(newb, newe, vm1.shape)):
|
|
return vm2 + View.create(vm1.shape, vm1.strides, vm1.offset, tuple(zip(newb, newe)))
|
|
# Otherwise if vm2's mask was violated, then cannot merge.
|
|
if bad: return None
|
|
|
|
return View.create(vm1.shape, tuple(strides), sum(o * s for o, s in zip(origin, vm2.strides)) + vm2.offset)
|
|
|
|
@functools.lru_cache(maxsize=None) # pylint: disable=method-cache-max-size-none
|
|
def invert(self, out_shape:tuple[sint, ...]) -> Optional[View]:
|
|
ret = View.create(self.shape)
|
|
if self.mask: ret = ret.shrink(self.mask)
|
|
ret = ret.flip(tuple(x < 0 for x in self.strides)).permute(argsort(tuple(-x if x > 0 else x for x in self.strides)))
|
|
return ret if prod(ret.shape) == prod(out_shape) else None # don't support shrink, expand, or stride != (-1, 1)
|
|
|
|
@functools.lru_cache(maxsize=None) # pylint: disable=method-cache-max-size-none
|
|
def minify(self):
|
|
min_shape = tuple(x[0] for x in merge_dims(self.shape, self.strides, self.mask))
|
|
return nv if (nv := self.reshape(min_shape)) else self
|
|
|
|
def __unsafe_resize(self, arg: tuple[tuple[sint, sint], ...], mask=None) -> View:
|
|
offset = sum([s * x[0] for s, x in zip(self.strides,arg)])
|
|
if self.mask:
|
|
# move the old mask
|
|
nmask = tuple([(smax(0, smin(mx-ax,ay-ax)), smax(0, smin(my-ax,ay-ax))) for (mx,my),(ax,ay) in zip(self.mask, arg)])
|
|
# merge the masks if we have two
|
|
mask = tuple([(smax(mx1, mx2), smin(my1, my2)) for (mx1, my1), (mx2, my2) in zip(nmask, mask)]) if mask is not None else nmask
|
|
return View.create(tuple([y-x for x,y in arg]), self.strides, self.offset+offset, mask)
|
|
|
|
@functools.lru_cache(maxsize=None) # pylint: disable=method-cache-max-size-none
|
|
def pad(self, arg: tuple[tuple[sint, sint], ...]) -> View:
|
|
assert len(arg) == len(self.shape), f"invalid pad {arg} for {self.shape}"
|
|
# NOTE: not checking for symbolic arg
|
|
for b,e in arg: assert not all_int([b,e]) or b>=0 and e>=0, f"invalid pad {arg} for {self.shape}"
|
|
if any(resolve(b!=0) or resolve(e!=0) for b, e in arg):
|
|
zvarg = tuple([(-b,s+e) for s,(b,e) in zip(self.shape, arg)])
|
|
mask = tuple([(b,s+b) for s,(b,_) in zip(self.shape, arg)])
|
|
return self.__unsafe_resize(zvarg, mask=mask)
|
|
return self
|
|
|
|
@functools.lru_cache(maxsize=None) # pylint: disable=method-cache-max-size-none
|
|
def shrink(self, arg: tuple[tuple[sint, sint], ...]) -> View:
|
|
assert len(arg) == len(self.shape), f"invalid shrink {arg} for {self.shape}"
|
|
# NOTE: not checking for symbolic arg
|
|
for s,(b,e) in zip(self.shape,arg): assert not all_int([s,b,e]) or (0<=b<=e<=s), f"invalid shrink {arg} for {self.shape}"
|
|
return self.__unsafe_resize(arg)
|
|
|
|
@functools.lru_cache(maxsize=None) # pylint: disable=method-cache-max-size-none
|
|
def expand(self, new_shape: tuple[sint, ...]) -> View:
|
|
if len(new_shape) != len(self.shape): raise ValueError(f"expand arg {new_shape=} must have same number of dimensions as shape {self.shape=}")
|
|
# NOTE: does not check multiple of symbolic shape
|
|
assert all(resolve(s == ns) or s == 1 for s,ns in zip(self.shape, new_shape)), f"can't expand {self.shape} into {new_shape}"
|
|
if 0 in self.shape: return View.create(new_shape)
|
|
# TODO: this resolve may not be needed, but it's hard because vars need to be sorted
|
|
mask = tuple([(((0,0) if m != (0,1) else (0,ns)) if resolve(s != ns, False) else m) \
|
|
for m,s,ns in zip(self.mask, self.shape, new_shape)]) if self.mask else None
|
|
return View.create(new_shape, self.strides, self.offset, mask)
|
|
|
|
@functools.lru_cache(maxsize=None) # pylint: disable=method-cache-max-size-none
|
|
def permute(self, axis: tuple[int, ...]) -> View:
|
|
assert sorted(axis) == list(range(len(self.shape))), f"invalid permutation {axis} of len {len(self.shape)}"
|
|
return View.create(tuple(self.shape[a] for a in axis), tuple(self.strides[a] for a in axis), self.offset,
|
|
tuple(self.mask[a] for a in axis) if self.mask is not None else None)
|
|
|
|
@functools.lru_cache(maxsize=None) # pylint: disable=method-cache-max-size-none
|
|
def flip(self, arg: tuple[bool, ...]) -> View:
|
|
offset = sum((s-1)*z for s,z,f in zip(self.shape, self.strides, arg) if f)
|
|
mask = tuple((s-my,s-mx) if f else (mx,my) for (mx,my),s,f in zip(self.mask, self.shape, arg)) if self.mask is not None else None
|
|
return View.create(self.shape, tuple(-z if f else z for z,f in zip(self.strides, arg)), self.offset+offset, mask)
|
|
|
|
@functools.lru_cache(maxsize=None) # pylint: disable=method-cache-max-size-none
|
|
def reshape(self, new_shape: tuple[sint, ...]) -> Optional[View]:
|
|
if self.shape == new_shape: return self
|
|
|
|
if not all(x >= 0 for x in new_shape): raise ValueError(f"shape can't contain negative numbers {new_shape}")
|
|
# check for the same size
|
|
if (self_all_int := all_int(self.shape)):
|
|
assert all(isinstance(s, (int, UOp)) for s in new_shape), f"{self.shape=} -> {new_shape=} contains non (int, Variable) dim"
|
|
if resolve(prod(self.shape) != prod(new_shape), False): raise ValueError(f"size mismatched, can't reshape {self.shape=} -> {new_shape=}")
|
|
|
|
if 0 in self.shape: return View.create(new_shape)
|
|
if new_shape == () and self.mask and any(mx==my for (mx,my) in self.mask): return None
|
|
|
|
# after the asserts, it's okay to check contiguous
|
|
if self.contiguous: return View.create(new_shape)
|
|
|
|
# if it's not contiguous and new shape is symbolic, check if it's directly replaceable
|
|
if self_all_int and not all_int(new_shape):
|
|
if len(self.shape) != len(new_shape): raise ValueError(f"cannot symbolic reshape non-contiguous {self} -> {new_shape}")
|
|
for si, so in zip(self.shape, new_shape):
|
|
if not isinstance(so, int): so = sym_infer(so, dict([v.unbind() for v in so.vars()]))
|
|
if si != so: raise ValueError(f"cannot symbolic reshape non-contiguous {self} -> {new_shape}")
|
|
# all dimensions matched, return the new view directly
|
|
return View(new_shape, self.strides, self.offset, self.mask, self.contiguous)
|
|
|
|
r_strides, r_new_shape = [], reversed(new_shape)
|
|
for merged_size, new_stride, real_size in reversed(merge_dims(self.shape, self.strides, self.mask)):
|
|
# TODO: write with get_contraction
|
|
acc = 1
|
|
# TODO: third resolve shouldn't be needed
|
|
while resolve(acc <= merged_size) and resolve(acc != merged_size) and resolve((new_dim := next(r_new_shape, 0)) > 0):
|
|
r_strides.append(new_stride * acc)
|
|
acc = acc * new_dim
|
|
if not resolve(acc < real_size): new_stride = 0
|
|
if resolve(acc != merged_size): return None
|
|
new_strides = (0,) * (len(new_shape) - len(r_strides)) + tuple(r_strides[::-1])
|
|
|
|
if (new_mask:=_reshape_mask(self.mask, self.shape, new_shape)) is not None:
|
|
extra_offset = (sum(m[0] * s for m,s in zip(self.mask, self.strides)) if self.mask else 0) - \
|
|
(sum(m[0] * s for m,s in zip(new_mask, new_strides)))
|
|
return View.create(new_shape, new_strides, self.offset + extra_offset, new_mask)
|
|
|
|
return None
|
|
|