from dataclasses import dataclass, field, replace import itertools from tinygrad.dtype import dtypes, PtrDType, AddrSpace, Invalid from tinygrad.uop.ops import PatternMatcher, UPat, Ops, UOp, resolve, GroupOp, _substitute, KernelInfo, ParamArg from tinygrad.uop.ops import graph_rewrite, sint, AxisType, BottomUpGate, profile_matches, identity_element from tinygrad.uop.symbolic import symbolic from tinygrad.helpers import prod, all_same, getenv, dedup, all_int, DEBUG, SPLIT_REDUCEOP, DEBUG_RANGEIFY, VIZ, MAX_KERNEL_BUFFERS from tinygrad.helpers import PCONTIG, FLOAT16, OPENPILOT_HACKS, argsort, partition, get_single_element from tinygrad.codegen.simplify import pm_flatten_range, pm_reduce_simplify from tinygrad.codegen.opt import Opt from tinygrad.schedule.indexing import run_rangeify, BufferizeOpts, IndexingContext, apply_movement_op from tinygrad.schedule.multi import multi_pm from tinygrad.schedule.allreduce import create_allreduce_function # creation can recurse a lot import sys sys.setrecursionlimit(10000) def add_ranges_to_store(ctx, x): if x.src[0]._shape is None or x.src[1]._shape is None or x.src[0].shape == (): return None assert x.src[0].shape == x.src[1].shape, "bad store shape" idxs = [UOp.range(r, next(ctx), AxisType.LOOP) for r in x.src[0].shape] return UOp.store(x.src[0].index(*idxs), x.src[1].index(*idxs)).end(*idxs) def lower_shaped_wmma(ctx, x): dims, device, threads = x.arg dtype_in, dtype_out = x.src[0].dtype.base, x.dtype upcasts = [(s, UOp.range(s.shape[-1], next(ctx), axis_type=AxisType.UPCAST)) for s in x.src] tc_upcast_axes = tuple(((u.arg[0], s.shape[-1]),) for s, u in upcasts) name = f"WMMA_{'_'.join(map(str, dims))}_{dtype_in.name}_{dtype_out.name}" wmma_arg = (name, dims, dtype_in, dtype_out, device, threads, tc_upcast_axes, ()) wmma = UOp(Ops.WMMA, dtype_out.vec(x.src[2].shape[-1]), tuple(s[u].contract(u) for s, u in upcasts), arg=wmma_arg) tmp = UOp.placeholder((x.src[2].shape[-1],), dtype_out, slot=next(ctx), addrspace=AddrSpace.REG) return tmp.after(UOp.group(*[tmp[e].store(wmma.gep(e)) for e in range(x.src[2].shape[-1])])) pm_store_ranges = PatternMatcher([ (UPat(Ops.STORE, name="x"), add_ranges_to_store), ]) pm_syntactic_sugar = PatternMatcher([ # INDEX on ptr INDEX concats them (UPat(Ops.INDEX, name="i1").f(Ops.INDEX, name="i2", allow_any_len=True), lambda i1,i2: i2.replace(src=i1.src+i2.src[1:]) if isinstance(i1.dtype, PtrDType) and not isinstance(i2.dtype, PtrDType) else None), # early rangeify (UPat(Ops.INDEX, src=(UPat(GroupOp.Elementwise | {Ops.CONST}, name="x"),), allow_any_len=True, name="idx"), lambda idx,x: x.replace(src=tuple([s.index(*idx.src[1:]) for s in x.src]))), ]) def found_after(ctx:dict[UOp, UOp], after:UOp, src:UOp): if (x:=src).op is Ops.CAST and x.dtype == dtypes.half and FLOAT16: x, after = x.src[0], after.cast(dtypes.float) while True: if x.op is Ops.PERMUTE: x, after = x.src[0], after.permute(argsort(x.marg)) elif x.op is Ops.RESHAPE: x, after = x.src[0], after.reshape(x.src[0].shape) elif x.op is Ops.WHERE and x.src[2].base.arg == Invalid and x.src[1].op is Ops.PAD: x, after = x.src[1].src[0], after.shrink(tuple((o, s+o) for (o,_),s in zip(x.src[1].marg, x.src[1].src[0].shape))) else: break ctx[x] = after # *** fold moved AFTERs (hack for openpilot) *** pm_fold_moved_after = PatternMatcher([ (UPat(Ops.AFTER, src=(UPat(), UPat(Ops.STORE, src=(UPat(), UPat((*GroupOp.Movement,Ops.CAST,Ops.WHERE), name="src")))), name="after"), found_after), # replace ALU sources with AFTER versions found above (UPat(GroupOp.ALU, name="alu"), lambda ctx,alu: alu.replace(src=new_src) if (new_src:=tuple(ctx.get(s, s) for s in alu.src)) != alu.src else None), ]) # movement op on INDEX as a PatternMatcher def _mop_index(r:UOp, idx:UOp): idxs = idx.src[1:] if len(idxs) == len(r.shape): return r.src[0].index(*apply_movement_op(r.op, r.src[0].shape, r.marg, idxs), dtype=idx.dtype, arg=idx.arg) if r.op is Ops.RESHAPE: src_prefix = len(r.src[0].shape) - len(r.shape[len(idxs):]) if src_prefix >= 0 and r.src[0].shape[src_prefix:] == r.shape[len(idxs):]: if src_prefix == 0: return r.src[0] if r.src[0].dtype == idx.dtype else None ret = r.src[0].index(*apply_movement_op(r.op, r.src[0].shape[:src_prefix], r.shape[:len(idxs)], idxs), dtype=idx.dtype, arg=idx.arg) return ret if ret.shape == idx.shape else None pm_mops = PatternMatcher([ (UPat(GroupOp.Movement, name="r").f(Ops.INDEX, allow_any_len=True, name="idx"), _mop_index), # move movement ops and INDEX after AFTER (but not when AFTER has a raw STORE with shaped children — from replace_contig_with_store_after) (UPat(GroupOp.Movement|{Ops.INDEX}, name="r").after(name="a", allow_any_len=True), lambda r,a: UOp(r.op, r.dtype, (a.replace(src=(r.src[0],)+a.src[1:]),)+r.src[1:], r.arg)), (UPat(GroupOp.Movement, name="r").end(name="a", allow_any_len=True), lambda r,a: a.replace(src=(r.src[0],)+a.src[1:])), # lower SHAPED_WMMA to WMMA with CONTRACT/UNROLL (UPat(Ops.SHAPED_WMMA, name="x"), lower_shaped_wmma), ]) # ***************** # 0. do some cleanup rewrites, mostly copied from the old stuff def fix_store_hazard(target:UOp, src:UOp): # PERMUTE and FLIP reorder indices, SHRINK can have overlapping regions when dest is also shrunk unsafe = {Ops.PERMUTE, Ops.FLIP} | ({Ops.SHRINK} if target.op_in_backward_slice_with_self(Ops.SHRINK) else set()) base = target.base reaches_base: dict[UOp, bool] = {} for s in src.toposort(gate=lambda s: s.op is not Ops.CONTIGUOUS): reaches_base[s] = s is base or any(reaches_base.get(c) for c in s.src) if reaches_base[s] and s.op in unsafe and not (s is target and s.op is Ops.SHRINK): return target.store(src.contiguous()) def split_reduceop(reduce:UOp, x:UOp): if prod(reduce.shape) == 0: return None if not SPLIT_REDUCEOP or not all_int(x.shape) or (prod(x.shape)//prod(reduce.shape))1) else UOp.const(dtypes.weakint, 0) for i,s in enumerate(x.shape)]) range_nums = [y.arg[0] for y in indexed.substitute({x.base:UOp(Ops.NOOP)}, extra_pm=pm_mops).ranges] is_expanded = [i not in range_nums for i in range(len(x.shape))] if not (split_candidates:=[(i,d) for i in reduce.arg[1] for d in range(min(256,2**getenv("REDUCEOP_SPLIT_SIZE",22)//prod(reduce.shape)),8-1,-1) if x.shape[i]%d==0 and not is_expanded[i]]): return None dim_to_split, divisor = split_candidates[0] splitted_shape = x.shape[:dim_to_split]+(divisor,)+(x.shape[dim_to_split]//divisor,)+x.shape[dim_to_split+1:] splitted = x.reshape(splitted_shape).permute(tuple([d for d in range(len(splitted_shape)) if d!=dim_to_split]+[dim_to_split])) if DEBUG >= 3: print(f"split {divisor}: {x.shape} -> {splitted.shape} -> {reduce.shape}") # reduce original axes, then split return splitted._rop(*reduce.arg).contiguous()._rop(reduce.arg[0], (len(reduce.shape),)).reshape(reduce.shape) mop_cleanup = PatternMatcher([ # merge adjacent RESHAPES (UPat(Ops.RESHAPE, src=(UPat(Ops.RESHAPE, name="x2"), UPat()), name="x"), lambda x,x2: x.replace(src=(x2.src[0], x.src[1]))), ]) pm_gather_params = PatternMatcher([ (UPat(Ops.PARAM, name="p"), lambda ctx, p: ctx.append(p)), ]) def resolve_function(c:UOp, allow_param_mismatch=True) -> UOp|None: if c.arg.precompile: return None params: list[UOp] = [] graph_rewrite(c.src[0], pm_gather_params, bottom_up=True, ctx=params, name="gather params") params = sorted(params, key=lambda x: x.arg.slot) args = c.src[1:] # NOTE: this isn't really needed. it's okay if there's unused args in the function if not allow_param_mismatch: if [x.arg.slot for x in params] != list(range(len(params))): raise RuntimeError(f"params not in order: {[x.arg.slot for x in params]}") if len(params) != len(args): raise TypeError(f"expected {len(params)} args, got {len(args)}") dict_map = {x:args[x.arg.slot] for x in params} for i, (p, a) in enumerate(dict_map.items()): if p.axis != a.axis: raise TypeError(f"arg {i} axis mismatch: expected {p.axis}, got {a.axis}") if p.max_shape != a.max_shape: raise TypeError(f"arg {i} shape mismatch: expected {p.shape}, got {a.shape}") if p.dtype != a.dtype: raise TypeError(f"arg {i} dtype mismatch: expected {p.dtype}, got {a.dtype}") return c.src[0].substitute(dict_map, walk=True) earliest_rewrites = mop_cleanup+PatternMatcher([ # resolve FUNCTION calls (inline the body) (UPat(Ops.FUNCTION, name="c"), resolve_function), # resolve TUPLE+GETTUPLE (UPat(Ops.GETTUPLE, src=(UPat(Ops.TUPLE, name="t"),), name="g"), lambda g,t: t.src[g.arg]), # resolve allreduce (must be bottom up) (UPat(Ops.ALLREDUCE, src=(UPat.var("buf"), UPat()), name="red"), create_allreduce_function), # split_reduceop (UPat(Ops.REDUCE, name="reduce", src=(UPat.var("x"),)), split_reduceop), # remove DETACH/CONTIGUOUS_BACKWARD (TODO: this is copied in allocations) (UPat((Ops.DETACH, Ops.CONTIGUOUS_BACKWARD), name="x"), lambda x: x.src[0]), # remove contiguous on movement ops before a copy on disk (UPat(GroupOp.Movement-{Ops.SHRINK, Ops.RESHAPE}, name="x").f(Ops.CONTIGUOUS).f(Ops.COPY, allow_any_len=True, name="copy"), lambda x,copy: copy.replace(src=(x,)+copy.src[1:]) if isinstance(x.device, str) and x.device.startswith("DISK") else None), # push copy past movement ops to disk (UPat(GroupOp.Movement-{Ops.SHRINK, Ops.RESHAPE}, name="x").f(Ops.COPY, allow_any_len=True, name="copy"), lambda x,copy: x.replace(src=(copy.replace(src=(x.src[0],)+copy.src[1:]),)+x.src[1:]) \ if isinstance(x.device, str) and x.device.startswith("DISK") else None), # SINK only ever references the base (UPat(Ops.SINK, name="x"), lambda x: x.replace(src=tuple(y.base for y in x.src))), # ** copy rules ** # COPY and source size need to match (UPat(Ops.COPY, src=(UPat(GroupOp.Movement, name="r"), UPat(name="d")), name="c"), lambda c,r,d: c.replace(src=(r.contiguous(), d)) if resolve(r.numel() != r.base.numel(), False) else None), # copying mselect to same device is just mselect (no NOOP kernel) (UPat(Ops.COPY, src=(UPat(Ops.MSELECT, name="ms"), UPat()), name="copy"), lambda ms,copy: ms if ms.device == copy.device else None), # copy only to different device (UPat(Ops.COPY, src=(UPat.var("x"), UPat()), name="copy"), lambda x,copy: x.f(Ops.NOOP) if x.device == copy.device else None), # ** store rules ** # fix store hazard (dest is in used in src) by adding contiguous: TestAssign.test_post_flipped_assignment (UPat(Ops.STORE, src=(UPat(name="target"), UPat(name="src"))), fix_store_hazard), # remove two STOREs that store the same thing to the same place: TestSchedule.test_dedup_assign (UPat.var("buf").after(UPat.var("buf").store(UPat.var("src")), name="a1").after(UPat.var("a1").store(UPat.var("src"))), lambda buf,src,a1:a1), # store a buffer's own current contents back into itself: TestAssign.test_nested_after_contiguous_store_no_init (UPat.var("buf").after(UPat.var("buf").store(UPat.var("buf").after(UPat.var("buf").store(UPat.var("src")), name="a1"))), lambda buf,src,a1:a1), # move bitcast from store dest to source: TestAssign.test_assign_bitcast (UPat(Ops.STORE, src=(UPat(Ops.BITCAST, src=(UPat(name="target"),)), UPat(name="src"))), lambda target, src: target.store(src.bitcast(target.dtype))), # ** size 0 ** # reduce of size 0 is the identity element (UPat(Ops.REDUCE, name="reduce", src=(UPat.var("x"),)), lambda reduce,x: reduce.const_like(identity_element(reduce.arg[0], reduce.dtype)) if 0 in x.shape and 0 not in reduce.shape else None), # handle size 0 (UPat(GroupOp.All-{Ops.SINK}, name="x"), lambda x: x.const_like(0).rtag(x.tag) if x._shape is not None and 0 in x.shape else None), ]) # ***************** # 3.5 cleanups ALWAYS_RUN_OPS = {Ops.CONTIGUOUS, Ops.COPY, Ops.NOOP} # you don't know in the first pass if axes are going to die, this happens if there's an EXPAND to the left def cleanup_dead_axes(b:UOp): # don't optimize ALWAYS_RUN_OPS or AFTER (AFTER is a buffer identity — ranges define consumer access, not computation) if b.src[0].op in ALWAYS_RUN_OPS or b.src[0].op is Ops.AFTER: return None new_rng = [] hit = False reshape: list[sint] = [] for s,rng in zip(b.shape, b.src[1:]): # skip for symbolic. TODO: fix this if rng.op is Ops.RANGE and rng.src[0].op is not Ops.CONST: return None # CONSTs are already dead axes if rng.op is Ops.CONST or (rng.op is Ops.RANGE and rng not in b.src[0].ranges): reshape.append(1) hit = True else: reshape.append(s) new_rng.append(rng) if hit: return b.replace(src=b.src[0:1]+tuple(new_rng)).reshape(tuple(reshape)).expand(b.shape) def gate_substitute(ctx, b:UOp) -> None: if not any(r in b.ranges for r in ctx.keys()): raise BottomUpGate() pm_gate_substitute = PatternMatcher([(UPat(GroupOp.All, name="b"), gate_substitute)], compiled=False) # if a buffer is being stored just for permutes or something, remove it # we want to reexpress the indexes of idx2 in terms of the implied b1 def remove_bufferize(src:UOp, buf:UOp, idx:UOp): # see if we can't do it, should this ever hit? assert len(buf.src) == len(idx.src), f"index on wrong bufferize, {len(buf.src)} != {len(idx.src)}" assert all(x.op in {Ops.RANGE, Ops.CONST} for x in buf.src[1:]) # if it's user contiguous, we never remove it if src.op in ALWAYS_RUN_OPS or not buf.arg.removable: return None # *** here is where we compute the cost *** # if we return None, the bufferize is kept accessed_buffers: list[UOp] = [] indexes: list[UOp] = [] reduces: list[UOp] = [] def red_gate(x:UOp): if x.op is Ops.AFTER: accessed_buffers.append(x.buf_uop) return False if (x.op is Ops.STAGE and x.arg.addrspace == AddrSpace.GLOBAL) or x.op is Ops.MSTACK: accessed_buffers.append(x) return False if x.op is Ops.STORE: # don't look inside stores, this doesn't count toward buffer accesses return False if x.op is Ops.PARAM: accessed_buffers.append(x) if x.op is Ops.INDEX: indexes.append(x) if x.op is Ops.REDUCE: reduces.append(x) return True src.toposort(gate=red_gate) del red_gate accessed_buffers = dedup(accessed_buffers) # if this is generated from multiple buffers, don't remove this buffer if len(accessed_buffers) > 3 and not (PCONTIG > 2): return None # if any reduces access a buffer, don't remove this buffer buffer_in_reduce = False def buf_gate(x:UOp): nonlocal buffer_in_reduce if x.op in {Ops.PARAM, Ops.STAGE, Ops.AFTER}: buffer_in_reduce = True return not buffer_in_reduce UOp.sink(*[x.src[0] for x in reduces]).toposort(gate=buf_gate) del buf_gate if buffer_in_reduce: if PCONTIG > 2: out_in_ratio = (prod(buf.shape)+1) / (sum([x.numel() for x in accessed_buffers])+1) if out_in_ratio < 10: return None # here we have to check the indexes, we might do a partial contig here local_indexes = [x for x in indexes if x.src[0].op is Ops.STAGE and x.src[0].arg.addrspace == AddrSpace.LOCAL] exclude_ranges = UOp.group(*[UOp.group(*x.src[1:]) for x in local_indexes]).ranges subs = [(k,v) for k,v in zip(buf.src[1:], idx.src[1:]) if k.op is not Ops.CONST] # if it's bufferized or a reduce, it's pcontig is_pcontig, is_subs = partition(subs, lambda x: x[0] in exclude_ranges or any([r.arg[-1] == AxisType.REDUCE for r in x[1].ranges])) if not len(is_subs): return None if len(is_pcontig): ret = src.substitute(dict(is_subs), extra_pm=pm_gate_substitute) return ret.bufferize(*[x[0] for x in is_pcontig], arg=BufferizeOpts(None, AddrSpace.LOCAL)).index(*[x[1] for x in is_pcontig]) else: return None # if it makes it here, the bufferize is removed # this is the ranges replaced # NOTE: if buf src is a const, we don't replace it. if idx is Invalid (dead load), don't replace it either replaced = {k:v for k,v in zip(buf.src[1:], idx.src[1:]) if k.op is not Ops.CONST and not (v.op is Ops.CONST and v.arg is Invalid)} return src.substitute(replaced, extra_pm=pm_gate_substitute) def remove_noop_bufferize(idx,b2): if idx.src[1:] != b2.src[1:] or idx.src[0].op is Ops.SLICE: return None return idx.src[0].shrink(tuple((0, s) for s in b2.shape)) if b2.shape else idx.src[0] pm_const_buffer_folding = pm_mops+PatternMatcher([ (UPat(Ops.STAGE, name="b"), cleanup_dead_axes), # remove noop buffers. if we look at the next index we can remove even more of these (UPat(Ops.INDEX, name="idx").f(Ops.STAGE, allow_any_len=True, name="b2"), remove_noop_bufferize), (UPat(Ops.INDEX, src=(UPat(Ops.STAGE),), allow_any_len=True, name="idx").f(Ops.NOOP).f(Ops.STAGE, allow_any_len=True, name="b2"), remove_noop_bufferize), # no buffers for const (ranges don't matter for const - it's the same value everywhere) (UPat(Ops.CONST, name='c').f(Ops.STAGE, allow_any_len=True, name="b"), lambda c,b: b.const_like(c.arg)), # indexing a const is a const (UPat(Ops.INDEX, src=(UPat(Ops.CONST, name="c"),),), lambda c: c), # copy on CONST is CONST (UPat(Ops.COPY, src=(UPat.cvar("x"), UPat()), name="copy"), lambda copy,x: copy.const_like(x.arg)), # hack if a noop turned to a const (UPat(Ops.NOOP, src=(UPat.cvar("c"),)), lambda c: c), # mstack on CONST is CONST (UPat(Ops.MSTACK, src=(UPat.var("s"),), allow_any_len=True).f(Ops.INDEX, allow_any_len=True), lambda s: UOp.const(c.dtype, c.arg) if (c:=s.base).op is Ops.CONST else None), ]) pm_remove_bufferize = PatternMatcher([ # remove reindexing with cost function (UPat.var("src").f(Ops.STAGE, allow_any_len=True, name="buf").f(Ops.INDEX, allow_any_len=True, name="idx"), remove_bufferize), # STORE to self is NOOP (UPat.var("x").store(UPat.var("x")), lambda x: UOp(Ops.NOOP)), # END on NOOP is NOOP (UPat(Ops.END, src=(UPat(Ops.NOOP, name="x"),), allow_any_len=True), lambda x: x), ]) def late_buffer_view(t:UOp, b:UOp): if not (isinstance(b.device, str) and b.device.startswith(("DISK", "TINYFS"))): return b shape = b.shape size = prod(shape) # walk up for the INDEX x = t while not any(u.op is Ops.INDEX for u in x.src): assert x.op not in GroupOp.Elementwise, "can't buffer view elementwise" x = x.src[0] x = next(u for u in x.src if u.op is Ops.INDEX) assert x.op is Ops.INDEX, "must be INDEX" if len(shape) == 0: offset = x.src[1].arg else: offset = max(sum(idx.vmin for idx in x.src[1:]), 0) return b.replace(src=(UOp(Ops.SLICE, t.dtype, (x.src[0], UOp.const(dtypes.weakint, offset)), size),)) to_bufferview = PatternMatcher([ (UPat(Ops.STAGE, src=(UPat((Ops.BITCAST, Ops.CONTIGUOUS), name="t"), UPat()), name="b"), late_buffer_view), ]) DEVICE_MAX_BUFS = {"METAL": 31, "WEBGPU": 8} # TODO: get from device? def limit_bufs(ctx:IndexingContext, root:UOp): if (device:=root.device) is None: return None # no device, index related calculations device = device if isinstance(device, str) else device[0].split(":")[0] if not (MAX_BUFS:=MAX_KERNEL_BUFFERS.value or DEVICE_MAX_BUFS.get(device, 0)): return None bufs: set[UOp] = set() def gate_input(u:UOp): # TODO: add cache to fix n^2 if is_load:=(u.op in {Ops.STAGE, Ops.AFTER, Ops.PARAM, Ops.MSELECT, Ops.MSTACK, Ops.DEFINE_VAR}): bufs.add(u) return not is_load root.toposort(gate=gate_input) if len(bufs) > MAX_BUFS - 1: # NOTE: this -1 is for the output buffer srcs = [] for s in root.src: if s.op in GroupOp.Elementwise and s.device is not None: # Insert bufferize: all AxisType.REDUCE before bufferize are AxisType.LOOP orig_ranges, end_ranges = s.ranges, [x.replace(arg=(next(ctx.range_idx), AxisType.LOOP)) if x.op is Ops.RANGE else x for x in s.ranges] s = s.substitute(dict(zip(orig_ranges, end_ranges))).bufferize(*end_ranges, arg=BufferizeOpts(device=s.device)).index(*orig_ranges) srcs.append(s) return root.replace(src=tuple(srcs)) pm_limit_bufs = PatternMatcher([(UPat(set.union(GroupOp.Binary, GroupOp.Ternary), name="root"), limit_bufs)]) # ***************** # 4. put in buffers for bufferize # TODO: should BUFFERIZE look a lot more like STORE # BUFFERIZE has device in arg # BUFFERIZE doesn't have indexing, that's implied by the ranges it closes # BUFFERIZE returns the BUFFER ready for INDEXing (doing this will make splitting a lot easier) # NOTE: this has been fixed up a bit def bufferize_to_store(ctx:itertools.count, x:UOp, idx:UOp, allow_locals=True): size = prod(x.shape) // x.dtype.count rngs = sorted(idx.ranges, key=lambda x: x.arg) assert size > 0 and isinstance(size, int), f"no zero sized or symbolic sized buffers {size}" sdtype = x.dtype.ptr(size=size, addrspace=x.arg.addrspace) # AFTER: add END to the existing STORE, return buffer with kernel dependency if (after:=x.src[0]).op is Ops.AFTER: buf = after.src[0].buf_uop.base if not (stores := [s for s in after.src[1:] if s.op is Ops.STORE and s.src[0].op is Ops.INDEX]): return buf # BUFFERIZE(INDEX(...)); store through the underlying global index instead. ended_stores = [] for store in stores: store_target = store.src[0] if store_target.src[0].op is Ops.STAGE and store_target.src[0].src[0].op is Ops.INDEX: store_target = store_target.src[0].src[0] if store.src[1] is store_target: continue # skip self-assign end_rngs = sorted(dedup(tuple(store_target.ranges) + tuple(rngs)), key=lambda x: x.arg) ended_stores.append(store_target.replace(dtype=sdtype).store(store.src[1]).end(*end_rngs)) return buf.after(*ended_stores) # NOTE: the DEFINE_LOCAL needs to be disambiguated here if sdtype.addrspace == AddrSpace.GLOBAL: buf = UOp(Ops.BUFFER, x.dtype, (UOp(Ops.LUNIQUE, arg=next(ctx)), UOp(Ops.DEVICE, arg=x.arg.device)), size) if x.src[0].op is Ops.SLICE: # no INDEX on SLICE, this could be cleaner do_store = buf.store(x.src[0]).end(*rngs) else: do_store = buf.index(idx, dtype=sdtype).store(x.src[0]).end(*rngs) return buf.after(do_store) if allow_locals: # handle locals buf = UOp.placeholder((size,), x.dtype, next(ctx), AddrSpace.LOCAL) do_store = buf.broadcast(x.src[1].dtype.count).index(idx, dtype=sdtype).store(x.src[0]).end(*rngs) return buf.after(do_store.barrier()) # collapse any BUFFERIZE to single input BUFFERIZE def flatten_bufferize(x:UOp): if len(x.src) == 2: return None ret = x.replace(src=(x.src[0], get_single_element(apply_movement_op(Ops.RESHAPE, (prod(x.shape),), x.shape, x.src[1:])))) rngs = x.src[1:] ret = ret.reshape(x.shape) if any(r.op is Ops.RANGE and r.src[0].op is not Ops.CONST for r in rngs): sym_shape = tuple([r.src[0] if r.op is not Ops.CONST else 1 for r in rngs]) ret = ret.shrink(tuple([(0,x) for x in sym_shape])) return ret pm_flatten_bufferize = PatternMatcher([(UPat(Ops.STAGE, name="x"), flatten_bufferize)]) pm_add_buffers = pm_mops+pm_flatten_bufferize+to_bufferview+PatternMatcher([ (UPat(Ops.STAGE, src=(UPat(), UPat(name="idx")), name="x"), lambda ctx,x,idx: bufferize_to_store(ctx, x, idx, allow_locals=False)), # move RESHAPEs through MSELECT/MSTACK (UPat((Ops.MSELECT, Ops.MSTACK), src=UPat(Ops.RESHAPE), name="m"), lambda m: m.replace(src=tuple([x.src[0].base for x in m.src])).reshape(m.shape)), # remove any RESHAPEs on KERNEL (UPat(Ops.CALL, name="k"), lambda k: k.replace(src=tuple(x.src[0] if x.op is Ops.RESHAPE else x for x in k.src))), # remove invalid writes (UPat(Ops.STORE, src=(UPat(), UPat(Ops.CONTIGUOUS, src=(UPat(Ops.CONST, arg=Invalid),)))), lambda: UOp(Ops.NOOP)), (UPat(Ops.STORE, src=(UPat(), UPat(Ops.CONST, arg=Invalid))), lambda: UOp(Ops.NOOP)), (UPat(Ops.AFTER, src=(UPat.var("x"), UPat(Ops.NOOP, src=()))), lambda x: x), (UPat(Ops.AFTER, src=(UPat.var("x"), UPat(Ops.END, src=(UPat(Ops.NOOP, src=()),), allow_any_len=True))), lambda x: x), ]) pm_add_buffers_local = pm_mops+pm_flatten_bufferize+to_bufferview+PatternMatcher([ (UPat(Ops.STAGE, src=(UPat(), UPat(name="idx")), name="x"), bufferize_to_store), ]) # ***************** # 5. split into kernels @dataclass class LocalAddBufferContext: dg:int = 0 map:dict = field(default_factory=dict) vars:dict = field(default_factory=dict) range:int = 0 opts:tuple|None = None def debuf(ctx:LocalAddBufferContext, buf:UOp): ret = UOp(Ops.PARAM, buf.dtype.ptr(prod(buf.max_shape), buf.addrspace), arg=ParamArg(ctx.dg, addrspace=buf.addrspace)).reshape(buf.max_shape) # if the buffer has symbolic shape, shrink the max-sized view to the actual shape if buf.max_shape != buf.shape: ret = ret.shrink(tuple((0, s) for s in buf.shape)) if buf not in ctx.map: ctx.map[buf] = buf ctx.dg += 1 return ret def unbind_kernel(ctx:LocalAddBufferContext, b:UOp): ctx.vars[b] = None return b.src[0] def handle_after(ctx:LocalAddBufferContext, after:UOp): if isinstance(after.dtype, PtrDType) and after.addrspace == AddrSpace.LOCAL: return None buf = after.buf_uop # NOTE: this is bottom up, so we only add it once if buf not in ctx.map: ctx.map[buf] = after return buf def renumber_range(ctx:LocalAddBufferContext, r:UOp): if r.tag != (): return None ret = r.replace(arg=(ctx.range,)+r.arg[1:], tag=None) ctx.range += 1 return ret def find_bufs(x:UOp): idxs = [s for s in x.toposort(gate=lambda x: x.op is not Ops.AFTER) if s.op is Ops.INDEX] read_from: dict[UOp, Ops] = {} if any((buf:=idx.buf_uop).op in {Ops.BUFFER, Ops.PARAM} and read_from.setdefault(buf, op:=idx.src[0].op) is not op for idx in idxs): raise RuntimeError(f"cycle detected while indexing {buf}") to_define_global = PatternMatcher([ (UPat(Ops.STORE, name="x"), find_bufs), (UPat((Ops.BUFFER, Ops.MSTACK, Ops.MSELECT), name="buf"), debuf), (UPat(Ops.PARAM, name="v"), lambda v: UOp.variable(v.arg.name, v.arg.vmin_vmax[0], v.arg.vmin_vmax[1], v.dtype) if v.arg.name is not None and v.arg.vmin_vmax is not None else None), (UPat(Ops.PARAM, name="buf"), lambda ctx, buf: None if isinstance(buf.dtype, PtrDType) or buf.arg.name is not None or buf._shape is None else debuf(ctx, buf)), (UPat(Ops.INDEX, src=(UPat(Ops.DEFINE_VAR, name="v"),)), lambda v: v), (UPat(Ops.BIND, name="b"), unbind_kernel), (UPat(Ops.AFTER, name="after"), handle_after), # remove device from local BUFFERIZE (UPat(Ops.STAGE, name="b"), lambda b: b.replace(arg=replace(b.arg, device=None))), # remove UNIQUE/DEVICE to dedup CONST (UPat(Ops.CONST, name="c"), lambda c: c.replace(src=()) if len(c.src) else None), # renumber the ranges starting with 0 so that kernel deduping works (UPat(Ops.RANGE, name="r"), renumber_range), ]) def get_contiguous(ctx:LocalAddBufferContext, x:UOp): if isinstance(x.arg, tuple) and all(isinstance(y, Opt) for y in x.arg): ctx.opts = x.arg return x.src[0] rangeify_codegen = PatternMatcher([ (UPat(Ops.CONTIGUOUS, name="x"), get_contiguous), # no NOOP in the kernel graph # TODO: this can be moved into codegen? (UPat(Ops.NOOP, name="x"), lambda x: x.src[0] if len(x.src) else None), # fix broadcast dtype (UPat(Ops.AFTER, name="a").broadcast(name="b"), lambda a,b: a.broadcast(len(b.src))), (UPat(Ops.DEFINE_LOCAL).f(Ops.AFTER, allow_any_len=True).broadcast(name="dg").f(Ops.INDEX, name="idx", allow_any_len=True), lambda dg,idx: None if isinstance(idx.dtype, PtrDType) else idx.replace(dtype=dg.dtype, arg=None).load(dtype=dg.dtype.base.scalar().vec(dg.dtype.vcount))), (UPat(Ops.AFTER, name="a").gep(name="b"), lambda a,b: a.gep(b.arg)), (UPat(Ops.DEFINE_LOCAL).f(Ops.AFTER, allow_any_len=True).gep(name="dg").f(Ops.INDEX, name="idx", allow_any_len=True), lambda dg,idx: None if isinstance(idx.dtype, PtrDType) else idx.replace(dtype=dg.dtype, arg=None).load(dtype=dg.dtype.base.scalar().vec(dg.dtype.vcount))), ]) pm_add_range_tags = PatternMatcher([ (UPat(Ops.RANGE, name="x"), lambda x: x.rtag(())), ]) def split_store(x:UOp) -> UOp|None: # if we have any open ranges here, we don't split if x.ranges: return None # local kernel rewrite lctx = LocalAddBufferContext() ret = graph_rewrite(x, to_define_global+pm_flatten_range+rangeify_codegen, ctx=lctx, name="kernel split", bottom_up=True) # SINK requires all buffers on the same device, but COPY/SLICE are cross-device or special hardware ops if ret.op is Ops.STORE: stored = ret.src[1] elif ret.op is Ops.END and ret.src[0].op is Ops.STORE: stored = ret.src[0].src[1] else: raise RuntimeError(f"unknown kernel type {ret.op}") if stored.op in {Ops.COPY, Ops.SLICE}: ret = stored.replace(src=stored.src + ret.ended_ranges) else: ret = ret.sink(arg=KernelInfo(opts_to_apply=lctx.opts)) kernel = ret.call(*lctx.map.values(), *lctx.vars.keys()) if ret.op is Ops.SINK and not all_same([x.device for x in kernel.src[1:] if x.op is not Ops.BIND]): raise RuntimeError(f"all buffers must be on the same device: {tuple(b.buf_uop for b in kernel.src[1:])}") return kernel split_kernels = PatternMatcher([ (UPat((Ops.STORE, Ops.END), name="x"), split_store), ]) @profile_matches def get_kernel_graph(sink:UOp) -> UOp: tsink = graph_rewrite(sink, multi_pm, name="multi_pm") if OPENPILOT_HACKS: tsink = graph_rewrite(tsink, pm_fold_moved_after, ctx={}, name="fold moved afters") tsink = graph_rewrite(tsink, pm_syntactic_sugar+pm_mops+earliest_rewrites, bottom_up=True, name="earliest rewrites") # convert movement ops to ranges tsink, rctx = run_rangeify(tsink, bool(DEBUG_RANGEIFY)) tsink = graph_rewrite(tsink, symbolic+pm_reduce_simplify+pm_const_buffer_folding+pm_remove_bufferize, name="symbolic+reduce_collapse+debuf") tsink = graph_rewrite(tsink, pm_limit_bufs, ctx=rctx, name="limit buffers") if VIZ: graph_rewrite(tsink, PatternMatcher([]), name="View Rangeify") # bufferize -> store lunique_start: int = max([-1]+[x.arg for x in tsink.toposort() if x.op is Ops.LUNIQUE]) + 1 tsink = graph_rewrite(tsink, pm_add_buffers+pm_add_range_tags, ctx=itertools.count(lunique_start), bottom_up=True, name="stage to store") tsink = graph_rewrite(tsink, split_kernels, bottom_up=True, name="split kernels") # WAR deps: if kernel U reads buffer S, and S is also written by another kernel, S's write must wait for U to finish afters = [u for u in tsink.toposort() if u.op is Ops.AFTER] kernel_assign: dict[UOp, UOp] = {u.buf_uop:u for u in afters} assign_rep: dict[UOp, UOp] = {} for u in afters: for s in u.src[1].src: # TODO: this is probably broken for MSELECT/MSTACK if s.op not in {Ops.BUFFER, Ops.PARAM} or s is u.buf_uop or (a:=kernel_assign.get(s)) is None: continue if a.src[1] is u.src[1]: continue # same kernel (multi-output custom kernels) if any(x.op is Ops.AFTER and x.buf_uop is s for x in kernel_assign[u.buf_uop].backward_slice): raise RuntimeError(f"cycle detected in assign graph, buffers {s} and {u.buf_uop} have circular dependency") assign_rep[a] = kernel_assign[s] = a.replace(src=a.src+(u,)) if assign_rep: tsink = graph_rewrite(tsink, _substitute, ctx=assign_rep, bottom_up=True, name="fix_assign") if VIZ: graph_rewrite(tsink, PatternMatcher([]), name="View Kernel Graph") return tsink