J3ff/IQ.Pilot
1
0
Fork 0
forked from IQ.Lvbs/IQ.Pilot
Code Pull Requests Activity

IQ.Pilot Prebuilt Release @ ab07000

Browse Source
This commit is contained in:
IQ.Lvbs history cleanup
2026-08-22 23:42:42 -05:00
commit 9f9c9a70cc
3729 changed files with 778697 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
tinygrad_repo/tinygrad/codegen/opt/__init__.py Normal file
View File

@@ -0,0 +1,20 @@
# opt opinionatedly transforms an ast into an optimized ast using either heuristics or beam search
from __future__ import annotations
from enum import Enum, auto
from dataclasses import dataclass
class OptOps(Enum):
TC = auto(); UPCAST = auto(); UNROLL = auto(); LOCAL = auto(); THREAD = auto() # noqa: E702
GROUP = auto(); GROUPTOP = auto(); NOLOCALS = auto(); PADTO = auto(); SWAP = auto() # noqa: E702
def __lt__(self, x:OptOps): return self.value < x.value
@dataclass(frozen=True, order=True)
class Opt:
op: OptOps
axis: int|None = None
arg: int|tuple|None = None
def __repr__(self): return f"Opt(op={self.op}, axis={self.axis}, arg={self.arg})"
class KernelOptError(Exception): pass
def check(cond:bool, msg:str=""):
if not cond: raise KernelOptError(msg)

191
tinygrad_repo/tinygrad/codegen/opt/heuristic.py Normal file
View File

@@ -0,0 +1,191 @@
import itertools
from tinygrad.codegen.opt import Opt, OptOps, KernelOptError
from tinygrad.helpers import getenv, DEBUG, prod, NOLOCALS, TC_OPT, TC_SELECT, USE_TC, IMAGE
from tinygrad.dtype import PtrDType, ImageDType
from tinygrad.uop.ops import Ops, resolve, AxisType
from tinygrad.codegen.opt.postrange import Scheduler
def hand_coded_optimizations(k:Scheduler) -> Scheduler:
# first try the tensor cores
""" Attempts to apply a tensor core optimization to the kernel. If one exists and applies properly, return true, otherwise return false.
Tensor cores are optimized instructions that matrix multiply-accumulate across a wave of threads: D(M, N) = A(M, K) * B(K, N) + C(M, N).
Keyword arguments:
use_tensor_cores -- controls how tensor cores are applied (default 1)
0: will disable any tensor core matching
1: enable tensor cores
2: apply tensor core shape but don't use UOp.WMMA
extra_opts -- additional Opt's to apply after the tensor core instead of the hand-coded additional Opt's (default None)
tc_select -- specifies which tensor core(s) to use for optimization (default -1)
-1: iterates through all available tensor cores in order and uses the first one that matches the requirements (dims and dtypes)
[0-N]: uses only the n'th tensor core available; useful for search
tc_opt -- controls which kinds of kernels may be eligible for tensor cores application (default 2 during BEAM, 0 otherwise)
0: applies to only kernels with a single reduce axis and direct Ops.LOAD into Ops.MUL
1: allows kernels with multiple reduce axes and also multiplication of Ops.CAST'd buffers
2: allows kernels with M, N, K axes that are not multiples of the tensor core dimensions by applying padding those axes as needed
"""
# NOTE: unless TC_OPT is > 0, we only trigger tensor cores if there's only one reduce axis
if USE_TC > 0 and (len(k.axes_of(AxisType.GROUP_REDUCE, AxisType.REDUCE)) == 1 or (TC_OPT.value >= 1)):
good_tc_opt = False
tk = k.copy()
try: # check TC first and apply hand-coded opts if successful
rngs = tk.apply_opt(Opt(OptOps.TC, 0, (TC_SELECT.value, TC_OPT.value, USE_TC.value)))
good_tc_opt = True
except KernelOptError:
pass
# skip hand-coded TC opts if AMX, upcasting will make kernel slower
if good_tc_opt and "AMX" not in k.ren.target.arch:
if rngs is not None:
for tc_dim in [1,0]: # attempt to upcast M and N
szs = [sz for sz in [5,4,3,2] if rngs[tc_dim].src[0].divides(sz) is not None]
if szs:
# set it to the replaced range
rngs[tc_dim] = tk.apply_opt(Opt(OptOps.UPCAST, tk.rngs.index(rngs[tc_dim]), szs[0]))[0]
if (szs := [sz for sz in [4,2] if rngs[0].src[0].divides(sz) is not None]): # attempt to local N
tk.apply_opt(Opt(OptOps.LOCAL, tk.rngs.index(rngs[0]), szs[0]))
return tk
# make a copy so it does not mutate the input
k = k.copy()
# upcast float4 images, this must be early so we don't accidentally add locals before the upcast
if IMAGE:
for buf_index,buf in enumerate(k.bufs):
if isinstance(buf.src[0].dtype, PtrDType) and ImageDType.valid_dims(buf.src[0].dtype, k.ren.target.arch):
# part of is_expanded
unit_stride_axes_mul_4 = [k.rngs.index(c) for c in k.bufs[buf_index].src[1].get_idx().split_uop(Ops.ADD) if
c.op is Ops.RANGE and (c.vmax+1)%4 == 0]
if len(unit_stride_axes_mul_4):
if (axis:=unit_stride_axes_mul_4[0]) in k.upcastable_dims:
k.apply_opt(Opt(OptOps.UPCAST, axis, 4))
elif axis in k.unrollable_dims:
k.apply_opt(Opt(OptOps.UNROLL, k.unrollable_dims.index(axis), 4))
# should use matvec - TODO: adjust/tune based on the wide vs tall/large vs small mat
MV_BLOCKSIZE, MV_THREADS_PER_ROW, MV_ROWS_PER_THREAD = getenv("MV_BLOCKSIZE", 4), getenv("MV_THREADS_PER_ROW", 8), getenv("MV_ROWS_PER_THREAD", 4)
if k.ren.has_local and getenv("MV",1) != 0 and (MV_BLOCKSIZE > 1 or MV_THREADS_PER_ROW > 1 or MV_ROWS_PER_THREAD > 1) and \
k.reduceop is not None and k.reduceop.arg[0] is Ops.ADD and len(k.full_shape) >= 2 and k.ren.has_shared and \
(mulop:=k.reduceop.src[0]).op is Ops.MUL and mulop.src[0].op is Ops.INDEX and mulop.src[1].op is Ops.INDEX:
idx0, idx1 = mulop.src[0].src[1].get_idx(), mulop.src[1].src[1].get_idx()
if k.ranges_of(AxisType.REDUCE):
first_reduce_rng = k.ranges_of(AxisType.REDUCE)[0]
if any(u is first_reduce_rng for u in idx0.split_uop(Ops.ADD)) and all(r in idx1.ranges for r in idx0.ranges):
for global_idx in k.axes_of(AxisType.GLOBAL):
if first_reduce_rng.src[0].divides(MV_THREADS_PER_ROW) is not None and k.full_shape[global_idx]%(MV_BLOCKSIZE*MV_ROWS_PER_THREAD) == 0:
if DEBUG >= 3:
print(f"MATVEC: {k.full_shape=} {first_reduce_rng.render()} {MV_BLOCKSIZE=} {MV_THREADS_PER_ROW=} {MV_ROWS_PER_THREAD=}")
try:
if MV_THREADS_PER_ROW > 1: k.apply_opt(Opt(OptOps.GROUP, 0, MV_THREADS_PER_ROW))
except KernelOptError: pass
if MV_BLOCKSIZE > 1: k.apply_opt(Opt(OptOps.LOCAL, global_idx, MV_BLOCKSIZE))
if MV_ROWS_PER_THREAD > 1: k.apply_opt(Opt(OptOps.UPCAST, global_idx, MV_ROWS_PER_THREAD))
return k
# are we grouping? (requires local shape support)
if resolve(prod(k.output_shape[i] for i in k.upcastable_dims) <= (240 if NOLOCALS else 2048), False):
for axis, sz in itertools.product((0, 1, 2), (16,)):
try:
k.apply_opt(Opt(OptOps.GROUPTOP, axis, sz))
break
except KernelOptError: pass
# no more opt if we are grouping
if k.group_for_reduces: return k
# **** below this line need to be optional and benchmarked ****
# if there are small dims with lots of valid masks, upcast them (they might be from Tensor.stack)
to_upcast: list[int] = []
# upcast leading axes first (hack-ish for winograd; we actually want to upcast masked axes with low stride first)
for axis in k.upcastable_dims:
# for Schedule, we check if the range is used in INDEX gates or WHERE gates
is_masked = any(any(o is k.rngs[axis] for o in u.src[0].backward_slice) for u in k.ast.backward_slice if u.op is Ops.WHERE)
if k.full_shape[axis] <= 7 and is_masked and prod(k.full_shape[j] for j in to_upcast) * k.full_shape[axis] <= 7 * 7:
if DEBUG >= 4: print(f"upcasting masked axis : {axis}")
to_upcast.append(axis)
for axis in to_upcast[::-1]: k.apply_opt(Opt(OptOps.UPCAST, axis, 0))
# potentially do more upcasts of non reduce axes based on a heuristic
is_dsp = k.ren is not None and k.ren.target.device == "DSP"
upcasted_axis: set[int] = set()
while resolve(prod(k.output_shape[i] for i in k.upcastable_dims) >= 1024) and (k.upcast_size() < 32):
xb_choices = []
# consider all upcastable axes with 3 or 4 upcast (128 on the DSP)
for axis, upcast_amount in itertools.product(k.upcastable_dims, ([128] if not len(upcasted_axis) else []) if is_dsp else [3,4]):
# if we haven't upcasted it, it mods, and buffer has stride 0 on axis while having no stride 0 in the upcasted axis already
if axis in upcasted_axis or k.full_shape[axis]%upcast_amount != 0: continue
rng = k.rngs[axis]
if any(rng not in b.src[1].get_idx().backward_slice and all(r2 in b.src[1].get_idx().backward_slice
for r2 in k.ranges_of(AxisType.UPCAST, AxisType.UNROLL)) for b in k.bufs):
num_strides, sum_strides = 0, 0
for b in k.bufs:
idx = b.src[1].get_idx()
if rng in idx.backward_slice: num_strides += 1
for c in idx.split_uop(Ops.ADD):
if c is rng: sum_strides += 1
if c.op is Ops.MUL and c.src[0] is rng and c.src[1].op is Ops.CONST: sum_strides += c.src[1].arg
if c.op is Ops.MUL and c.src[1] is rng and c.src[0].op is Ops.CONST: sum_strides += c.src[0].arg
xb_choices.append((num_strides, sum_strides, axis, upcast_amount))
if xb_choices:
xb_choices = sorted(xb_choices)
if DEBUG >= 4: print(f"more upcast axis : {xb_choices}")
k.apply_opt(Opt(OptOps.UPCAST, xb_choices[0][2], xb_choices[0][3]))
upcasted_axis.add(xb_choices[0][2])
else: break
# if last reduce dim is small(ish), loop unroll the reduce
# NOTE: this can fail on multireduce with mismatching dimensions, this is okay
try:
if k.unrollable_dims and (k.upcast_size() <= 4 or not k.axes_of(AxisType.UNROLL)) and (k.upcast_size() < 64):
if (s:=k.full_shape[k.unrollable_dims[-1]]) <= 32:
k.apply_opt(Opt(OptOps.UNROLL, len(k.unrollable_dims)-1, 0))
# if it's small, upcast a second reduce dimension too
if k.unrollable_dims and s <= 3 and k.full_shape[k.unrollable_dims[-1]] <= 3:
k.apply_opt(Opt(OptOps.UNROLL, len(k.unrollable_dims)-1, 0))
else:
for splits in [4]:
if k.full_shape[axis:=k.unrollable_dims[-1]]%splits == 0:
k.apply_opt(Opt(OptOps.UNROLL, len(k.unrollable_dims)-1, splits))
break
except KernelOptError: pass
# if nothing at all is upcasted and it's easy to, do an upcast
for splits in [4]:
if not k.upcasted and k.upcastable_dims and k.full_shape[k.upcastable_dims[-1]] % splits == 0:
k.apply_opt(Opt(OptOps.UPCAST, k.upcastable_dims[-1], splits))
# **** local groups ****
if k.ren.has_local:
if NOLOCALS:
k.apply_opt(Opt(OptOps.NOLOCALS))
else:
# prioritize making expand axes local
local_axis_ranking = [(any(k.rngs[axis] not in b.src[1].get_idx().backward_slice for b in k.bufs), axis) \
for axis in k.axes_of(AxisType.GLOBAL, AxisType.LOOP) if k.rngs[axis].src[0].op is Ops.CONST]
to_local: list[tuple[int, int]] = []
for _, axis in sorted(local_axis_ranking, key=lambda x: (-x[0], -x[1])):
local_size = prod(sz for _, sz in to_local)
local_sz: int|None = next((x for x in ([32] * (axis == 0) + [16,8,4,3,2]) if k.full_shape[axis] % x == 0 and local_size * x <= 128), None)
if local_sz is not None: to_local.append((axis, local_sz))
deleted_shape = 0
for axis, local_sz in sorted(to_local[:3]):
axis = axis - deleted_shape
will_delete_shape = local_sz == k.full_shape[axis]
k.apply_opt(Opt(OptOps.LOCAL, axis, local_sz))
if will_delete_shape: deleted_shape += 1
# **** threading ****
if k.ren.has_threads and k.ren.global_max is not None:
for threads in [32,16,12,8,6,5,4,3,2]:
# Skip if too many threads. Heuristic: use about 128K ops per thread
if threads > k.ren.global_max[0] or resolve(prod(k.full_shape) // (128 << 10) < threads): continue
for axis in k.axes_of(AxisType.LOOP):
if k.full_shape[axis] % threads == 0:
try: k.apply_opt(Opt(OptOps.THREAD, axis, threads))
except KernelOptError: pass
break
if k.applied_opts and k.applied_opts[-1].op is OptOps.THREAD: break
return k

352
tinygrad_repo/tinygrad/codegen/opt/postrange.py Normal file
View File

@@ -0,0 +1,352 @@
from __future__ import annotations
import math, itertools
from collections import defaultdict
from typing import cast, Final
from tinygrad.uop.ops import Ops, UOp, KernelInfo, graph_rewrite, AxisType, ssimplify, GroupOp, remove_all_tags
from tinygrad.uop.ops import axis_letters, axis_colors, axis_to_pos
from tinygrad.device import Buffer
from tinygrad.dtype import dtypes
from tinygrad.helpers import colored, getenv, DEBUG, to_function_name, NOOPT, argsort, round_up, prod, merge_dicts, get_single_element, flatten
from tinygrad.helpers import ALLOW_TF32, count, Context
from tinygrad.codegen.opt import Opt, OptOps, KernelOptError, check
from tinygrad.codegen.simplify import pm_flatten_range
from tinygrad.renderer import Renderer
class Scheduler:
def __init__(self, ast:UOp, ren:Renderer):
self.ast, self.ren = ast, ren
self.dont_use_locals = self.ast.arg.dont_use_locals if self.ast.arg is not None else False
self.applied_opts = list(self.ast.arg.applied_opts) if self.ast.arg is not None else []
self.opt_range = count(start=max([x.arg[0] for x in self.rngs], default=0)+1)
@property
def rngs(self):
# always in order by axistype
return sorted([u for u in self.ast.backward_slice if u.op is Ops.RANGE and u.vmax > 0], key=lambda x: (axis_to_pos[x.arg[-1]],) + x.arg[0:-1])
@property
def shape_len(self) -> int: return len(self.rngs)
@property
def full_shape(self): return [ssimplify(x.src[0]) for x in self.rngs]
@property
def axis_types(self) -> list[AxisType]: return [x.arg[-1] for x in self.rngs]
# strings like ['g0', 'g1', 'l0', 'l1', 'l2', 'l3', 'l4', 'l5', 'R0', 'r0', 'r1', 'r2', 'u0', 'u1', 'u2']
def shape_str(self) -> list[str]:
ret: list[str] = []
cnt: dict[AxisType, int] = {}
for x in self.axis_types:
cnt[x] = (cnt[x] + 1) if x in cnt else 0
ret.append(f"{axis_letters[x]}{cnt[x]}")
return ret
def shape_str_to_axis(self, nms:list[str]) -> tuple[int, ...]: return tuple([self.shape_str().index(x) for x in nms])
def copy(self) -> Scheduler:
ret = Scheduler(self.ast, self.ren)
ret.dont_use_locals = self.dont_use_locals
ret.applied_opts = self.applied_opts[:]
if hasattr(self, 'tensor_core'): ret.tensor_core = self.tensor_core
return ret
kernel_cnt: Final[defaultdict[str, int]] = defaultdict(int)
def get_optimized_ast(self, name_override:str|None=None) -> UOp:
if name_override is not None: name = name_override
else:
k_type = "r" if self.reduceop is not None else "E"
special_uops = sorted([x for x in self.ast.toposort() if x.op is Ops.SPECIAL], key=lambda x: x.arg)
special_ops = [colored(str(x.vmax+1), "blue" if x.arg[0] == "g" else "cyan") for x in special_uops]
name = k_type + colored('_', 'BLACK').join(['']+special_ops+[colored(x.src[0].render(), color) for x,color in zip(self.rngs, self.colors())])
Scheduler.kernel_cnt[(function_name := to_function_name(name))] += 1
num = f"n{Scheduler.kernel_cnt[function_name]-1}" if Scheduler.kernel_cnt[function_name] > 1 else ""
name += colored(num, 'BLACK')
self.ast = graph_rewrite(self.ast, pm_flatten_range, name="flatten range")
return self.ast.replace(arg=KernelInfo(name=name, applied_opts=tuple(self.applied_opts), dont_use_locals=self.dont_use_locals), tag=1)
def _output_rngs(self) -> list[UOp]:
return flatten([[r for r in UOp.sink(*s.src[1:]).ranges if r.arg[-1] != AxisType.REDUCE] for s in self.ast.src if s.op is Ops.END])
def _globalizable_rngs(self) -> list[UOp]:
ret = [r for r in self._output_rngs() if r.arg[-1] == AxisType.LOOP]
# exclude any output ranges from global that don't appear in all BUFFERIZE
for x in self.ast.toposort():
if x.op is Ops.STAGE:
ret = [r for r in ret if r in x.ranges]
return ret
def convert_loop_to_global(self) -> None:
if not self.ren.has_local: return
globalizible_rngs = self._globalizable_rngs()
rng = [x.replace(arg=x.arg[0:-1]+(AxisType.GLOBAL,)) if x in globalizible_rngs else x for x in self.rngs]
self.ast = self.ast.substitute(dict(zip(self.rngs, rng)))
def colors(self) -> list[str]:
output_rngs = self._output_rngs()
globalizible_rngs = self._globalizable_rngs()
ret = []
for x,r in zip(self.axis_types, self.rngs):
if self.dont_use_locals and x == AxisType.GLOBAL: ret.append("BLUE")
elif r not in output_rngs and x == AxisType.LOOP: ret.append("BLACK")
elif r not in globalizible_rngs and x == AxisType.LOOP: ret.append("white")
else: ret.append(axis_colors[x])
return ret
def colored_shape(self) -> str: return ' '.join([colored(f'{x.src[0].render():>4s}', color) for x,color in zip(self.rngs, self.colors())])
def shift_to(self, rng:UOp, amount:int, new_type:AxisType, top:bool=False, input_new_rng:UOp|None=None):
if (old_sz:=rng.src[0].divides(amount)) is None:
raise KernelOptError(f"{amount} can't divide {rng.src[0]} in {self.colored_shape()}")
new_rng = UOp.range(amount, next(self.opt_range), new_type) if input_new_rng is None else input_new_rng
replaced_rng = rng.replace(src=(old_sz,))
sub_axis = (new_rng * old_sz + replaced_rng) if top else (replaced_rng * amount + new_rng)
self.ast = self.ast.substitute({rng:sub_axis}, name=f"shift {rng.arg[:-1]} {amount} {str(new_type).split('.')[1].lower()}")
return replaced_rng, new_rng
def ranges_of(self, *axis_type:AxisType) -> list[UOp]: return [r for r in self.rngs if r.arg[-1] in axis_type]
def axes_of(self, *axis_type:AxisType) -> list[int]: return [i for i,t in enumerate(self.axis_types) if t in axis_type]
def upcast_size(self): return prod(self.full_shape[a] for a in self.axes_of(AxisType.UPCAST, AxisType.UNROLL))
# copied from kernel.py
@property
def upcastable_dims(self) -> list[int]: return [i for i in self.axes_of(AxisType.GLOBAL, AxisType.LOCAL, AxisType.LOOP) \
if isinstance(s:=self.full_shape[i], int) and s > 1]
@property
def unrollable_dims(self) -> list[int]: return [i for i in self.axes_of(AxisType.GROUP_REDUCE, AxisType.REDUCE) \
if isinstance(s:=self.full_shape[i], int) and s > 1]
def real_axis(self, op:OptOps, axis:int|None) -> int:
try:
if axis is None or op is OptOps.TC: return -1
if op is OptOps.UNROLL: return self.unrollable_dims[axis]
if op in {OptOps.GROUP, OptOps.GROUPTOP}: return self.axes_of(AxisType.REDUCE)[axis]
check(axis < self.shape_len, f"invalid axis on {axis=} {op=} {self.shape_len=}")
return axis
except IndexError as e: raise KernelOptError from e
def apply_opt(self, opt:Opt, append_opt:bool=True):
if opt.op is OptOps.NOLOCALS:
check(all(x not in {AxisType.WARP, AxisType.LOCAL, AxisType.GROUP_REDUCE} for x in self.axis_types), "no locals can't have locals")
if append_opt: self.applied_opts.append(opt)
self.dont_use_locals = True
return
if opt.op in {OptOps.LOCAL, OptOps.GROUP, OptOps.GROUPTOP}:
check(self.ren.has_local, "locals needed for opt")
rng = self.rngs[real_axis] if (real_axis:=self.real_axis(opt.op, opt.axis)) >= 0 else UOp(Ops.NOOP)
opt_to_at = {
OptOps.LOCAL: AxisType.LOCAL, OptOps.UPCAST: AxisType.UPCAST,
OptOps.UNROLL: AxisType.UNROLL, OptOps.GROUP: AxisType.GROUP_REDUCE,
OptOps.GROUPTOP: AxisType.GROUP_REDUCE, OptOps.THREAD: AxisType.THREAD}
ret = None
if opt.op in opt_to_at:
amt:int = int(rng.vmax+1) if opt.arg == 0 else cast(int, opt.arg)
# copied from kernel.py. prevents METAL compiler hangs
if self.reduceop is not None and (opt.op in {OptOps.GROUP, OptOps.GROUPTOP} or \
(self.group_for_reduces and opt.op not in {OptOps.NOLOCALS, OptOps.PADTO})):
upcast_local_sz = prod([self.full_shape[a] for a in self.axes_of(AxisType.UPCAST, AxisType.WARP, AxisType.LOCAL, AxisType.GROUP_REDUCE)])
smem_sz = amt*upcast_local_sz*self.reduceop.dtype.itemsize
check(smem_sz <= self.ren.shared_max, f"exceeds maximum shared memory size: needs {smem_sz}, max {self.ren.shared_max}")
if self.reduceop is not None and (opt.op in {OptOps.GROUP, OptOps.GROUPTOP}):
# We currently dont support a group within another rudece, TODO: fix if-contexts
reduce = [u for u in self.ast.backward_slice if u.op is Ops.REDUCE and rng in merge_dicts([r.ranges for r in u.src[1:]])][0]
check(not any(u.arg[-1] in (AxisType.REDUCE, AxisType.UNROLL, AxisType.GROUP_REDUCE) for u in reduce.ranges),
"cannot have a GROUP_REDUCE inside another reduce")
if opt.op is OptOps.UNROLL:
check(amt <= 32, "don't unroll more than 32")
check(rng.arg[-1] in {AxisType.GROUP_REDUCE, AxisType.REDUCE}, "unroll is for GROUP_REDUCE/REDUCE")
if opt.op is OptOps.UPCAST:
check((self.ren is not None and self.ren.target.device == "DSP") or amt <= 16, "don't upcast more than 16")
check(rng.arg[-1] in {AxisType.GLOBAL, AxisType.LOCAL, AxisType.LOOP}, f"upcast is for GLOBAL/LOCAL/LOOP, not {rng.arg[-1]}")
if opt.op is OptOps.LOCAL:
check(not self.dont_use_locals, "can't use locals")
check(rng.arg[-1] in {AxisType.GLOBAL, AxisType.LOOP}, "local is for globals")
if opt.op is OptOps.THREAD:
check(self.ren is not None and self.ren.has_threads, "target does not support threads")
check(self.ren is not None and self.ren.global_max is not None and amt <= self.ren.global_max[0], "too many threads")
check(all(x is not AxisType.THREAD for x in self.axis_types), "already threaded")
check(rng in self._globalizable_rngs(), "can't apply range to this dim")
if opt.op in {OptOps.GROUP, OptOps.GROUPTOP}:
check(all(x.op is not OptOps.TC for x in self.applied_opts), "no grouping with tensor cores") # TODO: why is this wrong?
check(not self.dont_use_locals, "can't use locals")
check(rng.arg[-1] == AxisType.REDUCE, "group is for reduce")
ret = self.shift_to(rng, amt, opt_to_at[opt.op], top=opt.op in {OptOps.GROUPTOP, OptOps.THREAD})
elif opt.op is OptOps.TC:
check(len(self.applied_opts) == 0, "tensor core opts must be first") # TODO: remove the need for this by having warps
check(opt.axis is not None, "tensor core opts must have an axis")
check(opt.arg is not None and isinstance(opt.arg, tuple) and len(opt.arg) == 3, "tensor core opts must have valid arg")
check(-1 <= (tc_select:=cast(tuple, opt.arg)[0]) < len(self.ren.tensor_cores), "tensor core opts must have valid tc_select")
check(0 <= (tc_opt:=cast(tuple, opt.arg)[1]) <= 2, "tensor core opts must have valid tc_opt")
check(0 < (use_tensor_cores:=cast(tuple, opt.arg)[2]) <= 2, "use_tensor_cores value is not valid")
try: ret = self._apply_tc_opt(use_tensor_cores, cast(int, opt.axis), tc_select, tc_opt)
except ValueError as e: raise KernelOptError(str(e))
check(ret is not None, "no tensor core available")
elif opt.op is OptOps.PADTO:
check(rng.src[0].op is Ops.CONST, "only pad const axes")
check(rng.arg[-1] not in {AxisType.UPCAST, AxisType.UNROLL}, "cannot pad upcasted") # TODO: why is this wrong?
check(rng.arg[-1] is not AxisType.THREAD, "cannot pad thread")
# ok to pad SUM if all parent ALU ops have f(0) = 0
if (r:=self.reduceop) is not None and rng.arg[-1] in (AxisType.GROUP_REDUCE, AxisType.REDUCE):
check(r.arg[0] is Ops.ADD and not r.op_in_backward_slice_with_self(*GroupOp.UnsafePad), f"cannot pad {r}")
new_sz = round_up(int(rng.vmax+1), cast(int, opt.arg))
check(rng.vmax+1 > new_sz//4, "pad adds more than quadruple the work")
replaced_rng = UOp.range(new_sz, *rng.arg)
replaces = {rng:replaced_rng}
valid = replaced_rng < rng.vmax+1
for b in self.bufs:
if rng in (i:=b.src[1].get_idx()).backward_slice_with_self:
replaces[b] = b.replace(src=(b.src[0],(valid&b.src[1].get_valid()).where(i, UOp.invalid())))
self.ast = self.ast.substitute(replaces, f"padto {rng.arg[:-1]} {opt.arg}")
elif opt.op is OptOps.SWAP:
try:
altrng:UOp = self.rngs[opt.arg]
except IndexError:
raise KernelOptError
check(rng.arg[-1] == AxisType.GLOBAL and altrng.arg[-1] == AxisType.GLOBAL, "swap only for globals")
self.ast = self.ast.substitute({rng:rng.replace(arg=(*altrng.arg[0:-1], rng.arg[-1]), tag=1),
altrng:altrng.replace(arg=(*rng.arg[0:-1], altrng.arg[-1]), tag=1)},
name=f"swap {rng.arg[:-1]} {altrng.arg[:-1]}")
self.ast = graph_rewrite(self.ast, remove_all_tags, name="swap remove tags")
else:
raise KernelOptError(f"unsupported opt {opt.op}")
if append_opt: self.applied_opts.append(opt)
return ret
def _apply_tc_opt(self, use_tensor_cores:int, axis:int, tc_select:int, opt_level:int) -> None|list[UOp]:
if not (reduceops := self.reduceops): raise KernelOptError("no reduce ops for TensorCore")
reduceop = reduceops[0]
if use_tensor_cores and reduceop.arg[0] is Ops.ADD:
mul = reduceop.src[0] if reduceop.src[0].op is not Ops.CAST else reduceop.src[0].src[0]
if mul.op is not Ops.MUL: return None
in0, in1 = mul.src
try:
tensor_cores = self.ren.tensor_cores if tc_select == -1 else [self.ren.tensor_cores[tc_select]]
except IndexError:
raise KernelOptError(f"invalid tensor core choice {tc_select}")
for tc in tensor_cores:
if self.ren.target.device in ("CUDA", "NV") and tc.dtype_in == dtypes.float and not ALLOW_TF32: continue
if tc.dtype_in == in0.dtype.scalar() and tc.dtype_in == in1.dtype.scalar() and tc.dtype_out == reduceop.dtype.scalar():
# tensor cores have three ranges. X, Y, and REDUCE
in0_ranges = sorted([u for u in in0.ranges if u not in in1.ranges], key=lambda x: x.arg[0], reverse=True)
in1_ranges = sorted([u for u in in1.ranges if u not in in0.ranges], key=lambda x: x.arg[0], reverse=True)
red_ranges = sorted(reduceop.src[1:], key=lambda x: x.arg[0], reverse=True)
if DEBUG >= 3:
print(f"TC({axis}): {[(x.arg[0],x.vmax+1) for x in in0_ranges]}",
f"{[(x.arg[0],x.vmax+1) for x in in1_ranges]} {[(x.arg[0],x.vmax+1) for x in red_ranges]}")
if not len(in0_ranges) or not len(in1_ranges) or not len(red_ranges): continue
# pick ranges
# NOTE: why are in1 and in0 switched?
axis_choices = list(itertools.product(in1_ranges, in0_ranges, red_ranges))
if not (axis < len(axis_choices)): continue
axes = list(axis_choices[axis])
# tag the reduceop
self.ast = self.ast.substitute({reduceop: reduceop.replace(tag="TC")})
# do optimizations and save the ranges
try:
for i,a in enumerate(axes):
idx = self.rngs.index(a)
if (a.vmax+1) % tc.dims[i] != 0:
if opt_level < 2: raise KernelOptError("tc padding requires opt_level >= 2")
# apply_opt should return the updated range?
self.apply_opt(Opt(OptOps.PADTO, idx, tc.dims[i]), append_opt=False) # PADTO might fail
axes[i] = self.rngs[idx]
except KernelOptError: continue
# we create the warp as a whole thing, in case some of these ranges are moved/removed later
warp = UOp.range(tc.threads, -1, AxisType.WARP)
ne: list[UOp] = []
for opt in tc.opts:
if opt[0] == "l":
axes[int(opt[1])], new_range = self.shift_to(axes[int(opt[1])], 2, AxisType.LOCAL, input_new_rng=warp%2)
warp //= 2
elif opt[0] == "u":
axes[int(opt[1])], new_range = self.shift_to(axes[int(opt[1])], 2, AxisType.UPCAST)
else: raise RuntimeError(f"unsupported opt {opt[0]} in tensor cores")
ne.append(new_range)
for _, amt in tc.get_reduce_axes():
axes[2], new_range = self.shift_to(axes[2], amt, AxisType.UNROLL)
ne.append(new_range)
if use_tensor_cores != 2:
# fix the srcs
reduceop = get_single_element([x for x in self.ast.toposort() if x.op is Ops.REDUCE and x.tag == "TC"])
tne = [x.replace(tag=1) for x in ne]
ret = reduceop.substitute(dict(zip(ne, tne)))
srcs = list((ret.src[0] if ret.src[0].op is not Ops.CAST else ret.src[0].src[0]).src)
srcs = [x.substitute(dict(zip(tne, [ne[i] for i in argsort(p)]))) for x,p in zip(srcs, tc.permutes_for_shape_str(tc.base_shape_str()))]
# get reduce/upcast axes for the tensor cores
tc_reduce_axes = self.shape_str_to_axis([f"r{i}" for i in range(len(tc.get_reduce_axes()))])
base_upcast_axes = tuple([(s,2) for s in self.shape_str_to_axis(tc.base_upcast_axes())])
tc_upcast_axes = tuple([base_upcast_axes[:int(math.log2(tc.elements_per_thread[i]))] for i in range(3)])
# axes to range number (was done in lowerer)
tc_upcast_axes = tuple([tuple([(self.rngs[a].arg[0], sz) for a,sz in v]) for v in tc_upcast_axes])
tc_reduce_axes = tuple([self.rngs[a].arg[0] for a in tc_reduce_axes])
# construct the op
# TODO: remove tc_upcast_axes from the arg
# do the reduce_axes always disappear? i think they don't
# they need to be moved into the WMMA srcs
wmma_arg = (str(tc), tc.dims, tc.dtype_in, tc.dtype_out, self.ren.target.device, tc.threads, tc_upcast_axes, ()) #, tc_reduce_axes)
wmma = UOp(Ops.WMMA, dtype=tc.dtype_out.vec(tc.elements_per_thread[2]), src=(
UOp(Ops.CONTRACT, dtype=srcs[0].dtype.vec(tc.elements_per_thread[0]), src=(srcs[0],), arg=tc_upcast_axes[0], tag=1),
UOp(Ops.CONTRACT, dtype=srcs[1].dtype.vec(tc.elements_per_thread[1]), src=(srcs[1],), arg=tc_upcast_axes[1], tag=1),
UOp.const(tc.dtype_out.vec(tc.elements_per_thread[2]), 0.0)), arg=wmma_arg, tag=1)
tc_uop = UOp(Ops.UNROLL, tc.dtype_out, (wmma,), arg=tc_upcast_axes[2], tag=1)
# preserve extra reduces
reduce_ranges = [x for x in UOp.sink(*reduceop.src[1:]).toposort() if x.op is Ops.RANGE and x.arg[0] not in tc_reduce_axes]
if len(reduce_ranges): tc_uop = UOp(Ops.REDUCE, tc_uop.dtype, (tc_uop,)+tuple(reduce_ranges), (Ops.ADD, ()))
self.ast = self.ast.substitute({reduceop: tc_uop})
self.tensor_core = tc
return axes
return None
# helpers for hand_coded_optimizations
@property
def reduceops(self) -> list[UOp]: return [x for x in self.ast.backward_slice if x.op is Ops.REDUCE]
@property
def reduceop(self) -> UOp|None:
if not (red := self.reduceops): return None
return UOp(Ops.REDUCE, red[0].dtype, red[0].src, red[0].arg)
@property
def bufs(self) -> list[UOp]: return [x for x in self.ast.toposort() if x.op is Ops.INDEX][::-1]
@property
def output_shape(self):
return [s if at not in {AxisType.REDUCE, AxisType.UNROLL, AxisType.GROUP_REDUCE} else 1 for s,at in zip(self.full_shape, self.axis_types)]
@property
def upcasted(self) -> int: return len(self.axes_of(AxisType.UPCAST, AxisType.UNROLL))
@property
def group_for_reduces(self) -> int: return len(self.axes_of(AxisType.GROUP_REDUCE))
def bufs_from_ast(ast:UOp, dname:str) -> list[Buffer]:
glbls = sorted([x for x in ast.backward_slice if x.op is Ops.PARAM], key=lambda x: x.arg.slot)
return [Buffer(dname, x.max_numel(), x.dtype.base) for x in glbls]
def apply_opts(ast:UOp, ren:Renderer, beam:int=0) -> UOp:
if ast.tag is not None: return ast
k = Scheduler(ast, ren)
k.convert_loop_to_global()
if ast.arg is not None and ast.arg.opts_to_apply is not None:
for opt in ast.arg.opts_to_apply: k.apply_opt(opt)
elif beam >= 1:
from tinygrad.codegen.opt.search import beam_search
rawbufs = bufs_from_ast(ast, ren.target.device)
# beam search may open devices
with Context(ALLOW_DEVICE_USAGE=1):
k = beam_search(k, rawbufs, beam, bool(getenv("BEAM_ESTIMATE", 1)))
elif not NOOPT and (ast.arg is None or ast.arg.applied_opts == ()):
from tinygrad.codegen.opt.heuristic import hand_coded_optimizations
# NOTE: hand_coded_optimizations doesn't support multiblock opts yet
if not any(u.op is Ops.STAGE for u in ast.backward_slice):
k = hand_coded_optimizations(k)
return k.get_optimized_ast(name_override=ast.arg.name if ast.arg is not None and ast.arg.name != "test" else None)

187
tinygrad_repo/tinygrad/codegen/opt/search.py Normal file
View File

@@ -0,0 +1,187 @@
import math, time, multiprocessing, traceback, signal, atexit
from dataclasses import replace
from tinygrad.uop.ops import sym_infer, AxisType, UOp
from tinygrad.uop.render import pyrender
from tinygrad.device import Device, Buffer
from tinygrad.helpers import prod, flatten, DEBUG, CACHELEVEL, diskcache_get, diskcache_put, getenv, Context, colored, time_to_str
from tinygrad.helpers import IGNORE_BEAM_CACHE
from tinygrad.codegen.opt import Opt, OptOps, KernelOptError
from tinygrad.engine.realize import time_call
from tinygrad.codegen import to_program
from tinygrad.codegen.opt.postrange import Scheduler
actions = [Opt(op=OptOps.UPCAST, axis=axis, arg=amt) for amt in [0,2,3,4,5,7] for axis in range(8)]
actions += [Opt(op=OptOps.UNROLL, axis=axis, arg=amt) for amt in [0,4,7] for axis in range(5)]
actions += [Opt(op=OptOps.LOCAL, axis=axis, arg=amt) for amt in [2,3,4,8,13,16,29] for axis in range(6)]
actions += [Opt(op=OptOps.GROUPTOP, axis=axis, arg=amt) for amt in [13,16,28,29,32,49,64,256] for axis in range(3)]
actions += [Opt(op=OptOps.GROUP, axis=axis, arg=amt) for amt in [0,4,8,16] for axis in range(3)]
if getenv("BEAM_PADTO", 0): actions += [Opt(op=OptOps.PADTO, axis=axis, arg=amt) for amt in [32] for axis in range(7)]
actions += [Opt(op=OptOps.LOCAL, axis=0, arg=32), Opt(op=OptOps.LOCAL, axis=6, arg=2)]
actions += [Opt(op=OptOps.TC, axis=0, arg=(-1, 0, getenv("TC", 1)))]
# covers resnet kernels (3 global * 3 reduce)
actions += [Opt(op=OptOps.TC, axis=axis, arg=(-1, getenv("TC_OPT", 2), getenv("TC", 1))) for axis in range(9)]
actions += [Opt(op=OptOps.SWAP, axis=axis_0, arg=axis_1) for axis_0 in range(5) for axis_1 in range(axis_0+1, 5)]
actions += [Opt(op=OptOps.THREAD, axis=axis, arg=amt) for amt in [2,3,4,5,8,12,16,24,32,64] for axis in range(3)]
if getenv("NOLOCALS"): actions += [Opt(op=OptOps.NOLOCALS)]
def get_test_global_size(global_size, max_global_size, var_vals):
test_global_size = [sym_infer(sz, var_vals) for sz in global_size]
input_size = prod(test_global_size)
while prod(test_global_size) > max_global_size:
for j in range(len(global_size)-1,-1,-1):
if test_global_size[j] > 16:
test_global_size[j] //= 2
break
return test_global_size, input_size / prod(test_global_size)
def _time_program(prg:UOp, var_vals:dict[str, int], rawbufs:list[Buffer], early_stop:float|None=None,
allow_test_size:int=True, max_global_size:int|None=65536, clear_l2=False, cnt=3, name="test", dev_timeout=False) -> list[float]:
timeout = int(early_stop * 1e3) if dev_timeout and early_stop is not None and early_stop < math.inf else None
factor = 1
if allow_test_size and max_global_size is not None:
global_size, factor = get_test_global_size(prg.arg.global_size, max_global_size, var_vals)
prg = prg.replace(arg=replace(prg.arg, global_size=tuple(global_size)))
call = prg.call(*[UOp.from_buffer(b) for b in rawbufs])
tms = []
for _ in range(cnt):
try: tms.append(time_call(call, var_vals, timeout=timeout, clear_l2=clear_l2) * factor)
except AssertionError: return [math.inf] * cnt
if early_stop is not None and early_stop < min(tms): break
return tms
class TimeoutException(Exception): pass
def timeout_handler(signum, frame):
if DEBUG >= 2: print("*** BEAM COMPILE TIMEOUT")
raise TimeoutException()
def _try_compile(x:tuple[int,Scheduler]) -> tuple[int, tuple[UOp, float]|None]:
if hasattr(signal, "alarm"):
signal.signal(getattr(signal, 'SIGALRM'), timeout_handler)
# set timeout
signal.alarm(getenv("BEAM_TIMEOUT_SEC", 10))
ret = None
try:
st = time.perf_counter()
prg = to_program(x[1].copy().get_optimized_ast(name_override="test"), x[1].ren)
et = time.perf_counter() - st
uops = prg.src[2].src
if len(uops) >= (uops_max:=getenv("BEAM_UOPS_MAX", 3000)) > 0:
if getenv("BEAM_LOG_SURPASS_MAX"): print(f"too many uops. {len(uops)=}, {uops_max=}")
raise RuntimeError("too many uops")
ret = (prg, et)
except RuntimeError:
if DEBUG >= 4: traceback.print_exc()
except Exception as e:
if getenv("BEAM_STRICT_MODE"): raise e
finally:
if hasattr(signal, "alarm"): signal.alarm(0)
return x[0], ret
# workers should not open devices and should ignore ctrl c and should not launch VIZ
def _init_worker():
Context(ALLOW_DEVICE_USAGE=0, VIZ=0, TRACK_MATCH_STATS=0).__enter__()
signal.signal(signal.SIGINT, signal.SIG_IGN)
def _ensure_buffer_alloc(bufs:list[Buffer]) -> list[Buffer]: return [buf.ensure_allocated() if buf is not None else buf for buf in bufs]
# *** external API ***
# get dictionary of all possible actions
def get_kernel_actions(s:Scheduler, include_0=True, max_up:int|None=None) -> dict[int, Scheduler]:
acted, max_up, max_lcl = {0:s} if include_0 else {}, getenv("BEAM_UPCAST_MAX", 256) if max_up is None else max_up, getenv("BEAM_LOCAL_MAX", 1024)
kernel_actions = actions.copy()
for i,a in enumerate(kernel_actions):
if a.axis is not None and a.op is not OptOps.TC:
try: ax = s.real_axis(a.op, a.axis)
except KernelOptError: continue
if (ax >= s.shape_len) or (s.full_shape[ax] == a.arg and Opt(a.op, a.axis, 0) in kernel_actions): continue
s2 = s.copy()
try:
s2.apply_opt(a)
up, lcl, tc_up = 1, 1, prod(tc.dims)//tc.threads if hasattr(s2, 'tensor_core') and (tc:=s2.tensor_core) else 1
for x,t in zip(s2.full_shape, s2.axis_types):
if t in (AxisType.UPCAST, AxisType.UNROLL): up *= x
elif t in (AxisType.WARP, AxisType.LOCAL, AxisType.GROUP_REDUCE): lcl *= x
if up//tc_up > max_up or lcl > max_lcl:
if getenv("BEAM_LOG_SURPASS_MAX"): print(f"too many upcast/local. {up//tc_up=}, {max_up=}, {lcl=}, {max_lcl=}")
continue
acted[i+1] = s2
except KernelOptError: pass
return acted
beam_pool, BEAM_DEBUG = None, getenv("BEAM_DEBUG")
def beam_search(s:Scheduler, rawbufs:list[Buffer], amt:int, allow_test_size=True, disable_cache=IGNORE_BEAM_CACHE.value):
global beam_pool
key = {"ast": s.ast.key, "amt": amt, "allow_test_size": allow_test_size, "device": s.ren.target.device, "suffix": s.ren.suffix}
if not disable_cache and CACHELEVEL >= 1 and (val:=diskcache_get("beam_search", key)) is not None:
ret = s.copy()
for o in val[len(s.applied_opts):]: ret.apply_opt(o)
return ret
beam: list[tuple[Scheduler, float]] = [(s, float("inf"))]
seen_libs = set()
default_parallel = multiprocessing.cpu_count() if s.ren.target.device in {"CUDA", "AMD", "NV", "METAL", "HIP"} else 0
if beam_pool is None and (workers := getenv("PARALLEL", default_parallel)):
beam_pool = multiprocessing.get_context("spawn").Pool(workers, _init_worker, (), getenv("BEAM_MAX_TASKS_PER_CHILD", 16))
@atexit.register
def close_pool(): beam_pool.close()
min_progress = getenv("BEAM_MIN_PROGRESS", 0.01)/1e6
if BEAM_DEBUG:
print("BEAM_SEARCH:")
print(pyrender(s.ast.replace(arg=None)))
if DEBUG >= 2: print(f" 0.00s: from 1 -> 1 actions {s.colored_shape()}")
try:
rawbufs = _ensure_buffer_alloc(rawbufs)
var_vals: dict[str, int] = {k.expr:int(k.vmax+k.vmin)//2 for k in s.ast.variables()}
exiting, st = False, time.perf_counter()
dev = Device[s.ren.target.device]
while not exiting:
candidates: list[Scheduler] = flatten([get_kernel_actions(si, include_0=False).values() for si,_ in beam])
timed: list[tuple[Scheduler, float]] = []
least_compute_ops = math.inf
for i, proc in ((map if beam_pool is None else beam_pool.imap_unordered)(_try_compile, enumerate(candidates))):
if proc is None: continue
prg, compile_et = proc
if (lib:=prg.src[4].arg) in seen_libs: continue
# filter out kernels that use 1000x more compute than the smallest
estimates = prg.src[0].arg.estimates
least_compute_ops = min(this_compute_ops:=sym_infer(estimates.ops if estimates is not None else 0, var_vals), least_compute_ops)
if least_compute_ops*1000 < this_compute_ops:
if getenv("BEAM_LOG_SURPASS_MAX"): print(f"too much compute. {this_compute_ops} when least is {least_compute_ops}")
continue
seen_libs.add(lib)
try: tms = _time_program(prg, var_vals, rawbufs, early_stop=beam[0][1]*3 if len(beam) else 1.0,
allow_test_size=allow_test_size, clear_l2=hasattr(dev, 'invalidate_caches'),
dev_timeout=getenv("BEAM_DEV_TIMEOUT", 1))
except Exception as e:
if BEAM_DEBUG: print(f"BEAM failed for opts: {candidates[i].applied_opts}\n{e}")
if isinstance(e, RuntimeError): continue
raise
timed.append((candidates[i], min(tms)))
if BEAM_DEBUG > 1:
print(f"{time.perf_counter() - st:7.2f}s: {i:5d} {len(prg.src[2].src):5d} uops",
f"{time_to_str(compile_et, w=12)} compile/{time_to_str(timed[-1][1], w=12)} run",
f" {len(timed):4d}/{len(candidates):4d} {timed[-1][0].colored_shape()}")
elif DEBUG >= 2:
print(f"\r{time.perf_counter() - st:7.2f}s: {time_to_str(timed[-1][1], w=12)}",
f" {len(timed):4d}/{len(candidates):4d} {timed[-1][0].colored_shape()}\033[K", end="")
# done
opts = sorted(timed, key=lambda x: x[1])
exiting = len(opts) == 0 or (opts[0][1] < min_progress) or (len(beam) > 0 and ((beam[0][1]-opts[0][1]) < min_progress))
if not exiting: beam = opts[:amt]
elif len(opts) > 0 and opts[0][1] < beam[0][1]: beam = opts[:1]
if DEBUG >= 2:
print(f"\r{time.perf_counter() - st:7.2f}s:", colored(time_to_str(beam[0][1], w=12), "green" if exiting else None),
f"from {len(candidates):3d} -> {len(opts):3d} actions\033[K", beam[0][0].colored_shape())
except KeyboardInterrupt as e:
if beam_pool is not None: beam_pool.terminate()
raise e
if CACHELEVEL >= 1: diskcache_put("beam_search", key, beam[0][0].applied_opts)
if BEAM_DEBUG: print(f"BEAM_SEARCH: final tm={time_to_str(beam[0][1], w=0)}, applied_opts={beam[0][0].applied_opts}")
return beam[0][0]

159
tinygrad_repo/tinygrad/codegen/opt/tc.py Normal file
View File

@@ -0,0 +1,159 @@
import math, functools
from dataclasses import dataclass
from tinygrad.dtype import DType, dtypes
@dataclass(frozen=True)
class TensorCore: # D = A * B + C, A is (M x K), B is (K x N), C and D are (M x N)
dims: tuple[int,int,int] # N, M, K
threads: int # number of threads that construct the warp
elements_per_thread: tuple[int, int, int] # elements per-thread to load/store from A/B/C
dtype_in: DType # dtype for A and B
dtype_out: DType # dtype for C and D
opts: tuple[str, ...] # ordered tuple of "ux" or "lx" specifying kernel opts to perform. "ux" upcasts dim x and "lx" localizes dim x
# (local_swizzle, upcast_swizzle, reduce_swizzle)
# l<num> is the num axis of the locals, similar for u<num> and upcasts, r<num> and reduces
swizzle: tuple[tuple[tuple[str, ...], tuple[str, ...], tuple[str, ...]], tuple[tuple[str, ...], tuple[str, ...], tuple[str, ...]]]
@functools.cache # pylint: disable=method-cache-max-size-none
def _remaps(self) -> list[dict[str, str]]:
local_axes, upcast_axes, reduce_axes = len(self.get_local_axes()), len(self.get_upcast_axes()), len(self.get_reduce_axes())
fwd_st = [f"l{i}" for i in range(local_axes)] + [f"u{i}" for i in range(upcast_axes)] + [f"r{i}" for i in range(reduce_axes)]
return [dict(zip(fwd_st, sum(s, ()))) for s in self.swizzle]
def permutes_for_shape_str(self, shape_str:list[str]) -> tuple[tuple[int, ...], tuple[int, ...]]:
ret = [[shape_str.index(remap[ss]) if ss in remap else i for i,ss in enumerate(shape_str)] for remap in self._remaps()]
return tuple(ret[0]), tuple(ret[1])
@functools.cache # pylint: disable=method-cache-max-size-none
def base_shape_str(self) -> list[str]:
ret = []
cnt = {'u': 0, 'l': 0}
for opt in self.opts:
ret.append(f"{opt[0]}{cnt[opt[0]]}")
cnt[opt[0]] += 1
# assumes you do the UNROLL after the opts
return ret + [f"r{i}" for i in range(len(self.get_reduce_axes()))]
def get_reduce_axes(self): return [(i, 2) for i in range(int(math.log2(self.dims[2])))]
def get_upcast_axes(self): return [opt for opt in self.opts if opt[0] == "u"]
def get_local_axes(self): return [opt for opt in self.opts if opt[0] == "l"]
def base_upcast_axes(self):
# this is defined in the swizzle. first we use the upcast axes, then the reduce
return ([f"r{i}" for i in range(len(self.get_reduce_axes()))] + [f"u{i}" for i in range(len(self.get_upcast_axes()))])[::-1]
def __str__(self): return "_".join(["WMMA"] + list(map(str, self.dims)) + [self.dtype_in.name, self.dtype_out.name])
def __post_init__(self):
# all axes have size 2, <local> <reduce> <upcast> is the order
local_axes, upcast_axes, reduce_axes = len(self.get_local_axes()), len(self.get_upcast_axes()), len(self.get_reduce_axes())
assert self.dims[0] * self.dims[1] == 2**(local_axes + upcast_axes), \
f"N({self.dims[0]}) x M({self.dims[1]}) != local({2**local_axes}) x upcast({2**upcast_axes}) with opts({self.opts})"
assert 2**local_axes == self.threads, f"{self.threads} threads construct the warp but found {2**local_axes} in {self.opts}"
assert 2**upcast_axes == self.elements_per_thread[2], \
f"{self.elements_per_thread[2]} elements from C are processed per thread but found {2**upcast_axes} in {self.opts}"
# check dims match opts
assert self.dims[0] == 2**len(gd:=[x for x in self.opts if x[1] == '0']), f"opts wrong on dims[0], {self.dims[0]} vs {gd}"
assert self.dims[1] == 2**len(gd:=[x for x in self.opts if x[1] == '1']), f"opts wrong on dims[1], {self.dims[1]} vs {gd}"
# NOTE: the K opts is implictly set by the dim
# check swizzle
assert len(self.swizzle[0]) == 3 and len(self.swizzle[1]) == 3, "swizzle has wrong part count"
assert len(self.swizzle[0][0]) == len(self.swizzle[1][0]) == local_axes, "local swizzle size is wrong"
assert len(self.swizzle[0][1]) == len(self.swizzle[1][1]) == upcast_axes, "upcast swizzle size is wrong"
assert len(self.swizzle[0][2]) == len(self.swizzle[1][2]) == reduce_axes, "reduce swizzle size is wrong"
assert all(len(s) == local_axes+upcast_axes+reduce_axes for s in self._remaps()), "remaps are the wrong size"
# check elements_per_thread
un, ln = 0, 0
zero_stride_0 = []
zero_stride_1 = []
for o in self.opts:
if o[1] == '0': zero_stride_0.append(o[0] + str(un if o[0] == 'u' else ln))
if o[1] == '1': zero_stride_1.append(o[0] + str(un if o[0] == 'u' else ln))
if o[0] == 'u': un += 1
if o[0] == 'l': ln += 1
# NOTE: all the zero_stride dims can be placed in any order in the swizzle
upcasted_0 = [x for x in (self.swizzle[0][1] + self.swizzle[0][2]) if x not in zero_stride_0 and x[0] != 'l']
upcasted_1 = [x for x in (self.swizzle[1][1] + self.swizzle[1][2]) if x not in zero_stride_1 and x[0] != 'l']
assert 2**len(upcasted_0) == self.elements_per_thread[0], f"mismatch in elements_per_thread[0], {upcasted_0} vs {self.elements_per_thread[0]}"
assert 2**len(upcasted_1) == self.elements_per_thread[1], f"mismatch in elements_per_thread[1], {upcasted_1} vs {self.elements_per_thread[1]}"
# ***** NVIDIA *****
cuda_tc_opts = ("u0","l0","l0","l1","l1","l1","u1") # shared by all shapes with M=16 N=8
# https://docs.nvidia.com/cuda/parallel-thread-execution/#warp-level-matrix-multiply-accumulate-instructions
cuda_81616 = [TensorCore(dims=(8,16,16), threads=32, elements_per_thread=(8,4,4), dtype_in=di, dtype_out=do, opts=cuda_tc_opts,
swizzle=((('r1', 'r2', 'l2', 'l3', 'l4'), ('u1', 'r3'), ('l0', 'l1', 'u0', 'r0')),
(('r1', 'r2', 'u0', 'l0', 'l1'), ('r0', 'r3'), ('l2', 'l3', 'l4', 'u1'))))
for di,do in [(dtypes.half,dtypes.float), (dtypes.bfloat16,dtypes.float), (dtypes.half,dtypes.half)]]
cuda_81632_f8 = [TensorCore(dims=(8,16,32), threads=32, elements_per_thread=(16,8,4), dtype_in=di, dtype_out=do, opts=cuda_tc_opts,
swizzle=((('r2', 'r3', 'l2', 'l3', 'l4'), ('u1', 'r4'), ('l0', 'l1', 'u0', 'r0', 'r1')),
(('r2', 'r3', 'u0', 'l0', 'l1'), ('r1', 'r4'), ('l2', 'l3', 'l4', 'u1', 'r0'))))
for di,do in [(dtypes.fp8e4m3,dtypes.float),(dtypes.fp8e5m2,dtypes.float)]]
cuda_8168_f16 = [TensorCore(dims=(8,16,8), threads=32, elements_per_thread=(4,2,4), dtype_in=di, dtype_out=do, opts=cuda_tc_opts,
swizzle=((('r1', 'r2', 'l2', 'l3', 'l4'), ('r0', 'u1'), ('l0', 'l1', 'u0')),
(('r1', 'r2', 'u0', 'l0', 'l1'), ('u1', 'r0'), ('l2', 'l3', 'l4'))))
for di,do in [(dtypes.half,dtypes.float), (dtypes.half,dtypes.half)]]
cuda_8168_tf32 = [TensorCore(dims=(8,16,8), threads=32, elements_per_thread=(4,2,4), dtype_in=dtypes.float, dtype_out=dtypes.float, opts=cuda_tc_opts,
swizzle=((('r0', 'r1', 'l2', 'l3', 'l4'), ('u1', 'r2'), ('l0', 'l1', 'u0')),
(('r0', 'r1', 'u0', 'l0', 'l1'), ('u1', 'r2'), ('l2', 'l3', 'l4'))))]
cuda_sm75: list[TensorCore] = cuda_8168_f16
cuda_sm80: list[TensorCore] = cuda_81616 + cuda_8168_f16 + cuda_8168_tf32
cuda_sm89: list[TensorCore] = cuda_sm80 + cuda_81632_f8
def get_cuda(arch): return cuda_sm89 if (ver:=int(arch[3:])) >= 89 else cuda_sm80 if ver >= 80 else cuda_sm75 if ver >= 75 else []
# ***** AMD *****
# https://gpuopen.com/learn/wmma_on_rdna3/
amd_rdna3 = [TensorCore(dims=(16,16,16), threads=32, elements_per_thread=(16,16,8), dtype_in=di, dtype_out=do,
opts=("l0","l0","l0","l0","l1","u1","u1","u1"),
swizzle=((('l4', 'u0', 'u1', 'u2', 'l0'), ('r1', 'r2', 'r3'), ('l1', 'l2', 'l3', 'r0')),
(('l0', 'l1', 'l2', 'l3', 'l4'), ('r1', 'r2', 'r3'), ('u0', 'u1', 'u2', 'r0'))))
for di,do in [(dtypes.half,dtypes.float),(dtypes.half,dtypes.half),(dtypes.bfloat16,dtypes.float)]]
amd_rdna4 = [TensorCore(dims=(16,16,16), threads=32, elements_per_thread=(8,8,8), dtype_in=di, dtype_out=do,
opts=("l0","l0","l0","l0","u1","u1","u1","l1"),
swizzle=((('u0', 'u1', 'u2', 'l4', 'r2'), ('r0', 'r1', 'r3'), ('l0', 'l1', 'l2', 'l3')),
(('l0', 'l1', 'l2', 'l3', 'r2'), ('r0', 'r1', 'r3'), ('l4', 'u0', 'u1', 'u2'))))
for di,do in [(dtypes.half,dtypes.float),(dtypes.half,dtypes.half),(dtypes.bfloat16,dtypes.float),(dtypes.bfloat16,dtypes.bfloat16)]]
# https://gpuopen.com/learn/amd-lab-notes/amd-lab-notes-matrix-cores-readme
amd_cdna_161616 = [TensorCore(dims=(16,16,16), threads=64, elements_per_thread=(4,4,4), dtype_in=di, dtype_out=do,
opts=("l0","l0","l0","l0","u1","u1","l1","l1"),
swizzle=((('u0', 'u1', 'l4', 'l5', 'r2', 'r3'), ('r0', 'r1'), ('l0', 'l1', 'l2', 'l3')),
(('l0', 'l1', 'l2', 'l3', 'r2', 'r3'), ('r0', 'r1'), ('l4', 'l5', 'u0', 'u1'))))
for di,do in [(dtypes.half,dtypes.float),(dtypes.bfloat16,dtypes.float)]]
amd_cdna_161632 = [TensorCore(dims=(16,16,32), threads=64, elements_per_thread=(8,8,4), dtype_in=di, dtype_out=do,
opts=("l0","l0","l0","l0","u1","u1","l1","l1"),
swizzle=((('u0', 'u1', 'l4', 'l5', 'r3', 'r4'), ('r0', 'r1'), ('l0', 'l1', 'l2', 'l3', 'r2')),
(('l0', 'l1', 'l2', 'l3', 'r3', 'r4'), ('r0', 'r1'), ('l4', 'l5', 'u0', 'u1', 'r2'))))
for di,do in [(dtypes.fp8e5m2,dtypes.float),(dtypes.fp8e4m3,dtypes.float),(dtypes.half,dtypes.float),(dtypes.bfloat16,dtypes.float)]]
amd_cdna_1616128 = [TensorCore(dims=(16,16,128), threads=64, elements_per_thread=(32,32,4), dtype_in=di, dtype_out=do,
opts=("l0","l0","l0","l0","u1","u1","l1","l1"),
swizzle=((('u0', 'u1', 'l4', 'l5', 'r5', 'r6'), ('r0', 'r1'), ('l0', 'l1', 'l2', 'l3', 'r2', 'r3', 'r4')),
(('l0', 'l1', 'l2', 'l3', 'r5', 'r6'), ('r0', 'r1'), ('l4', 'l5', 'u0', 'u1', 'r2', 'r3', 'r4'))))
for di,do in [(dtypes.fp8e5m2,dtypes.float),(dtypes.fp8e4m3,dtypes.float)]]
amd_cdna3 = amd_cdna_161632[:2] + amd_cdna_161616
amd_cdna4 = amd_cdna_1616128 + amd_cdna_161632 + amd_cdna_161616
def get_amd(arch): return {"gfx942": amd_cdna3, "gfx950": amd_cdna4, "gfx1200": amd_rdna4, "gfx1201": amd_rdna4}.get(arch, amd_rdna3)
# ***** Apple Metal *****
metal = [TensorCore(dims=(8,8,8), threads=32, elements_per_thread=(2,2,2), dtype_in=di, dtype_out=do,
opts=("u0","l0","l1","l1","l0","l1"),
swizzle=((('r1', 'l1', 'l2', 'r2', 'l4'), ('r0',), ('u0', 'l0', 'l3')),
(('l0', 'r0', 'r1', 'l3', 'r2'), ('u0',), ('l1', 'l2', 'l4'))))
for di,do in [(dtypes.float,dtypes.float),(dtypes.half,dtypes.float),
(dtypes.half,dtypes.half),(dtypes.bfloat16,dtypes.float),(dtypes.bfloat16,dtypes.bfloat16)]]
# ***** Apple AMX *****
amx = [TensorCore(dims=(sz,sz,1), threads=1, elements_per_thread=(sz,sz,sz*sz), dtype_in=dt, dtype_out=dt,
swizzle=(((), ('u0', 'u1', 'u2', 'u3', 'u4', 'u5', 'u6', 'u7'), ()),
((), ('u4', 'u5', 'u6', 'u7', 'u0', 'u1', 'u2', 'u3'), ())),
opts=("u0","u0","u0","u0","u1","u1","u1","u1")) for dt,sz in [(dt, 64 // dt.itemsize) for dt in [dtypes.float]]]
# ***** Intel ****
intel = [TensorCore(dims=(8,8,16), threads=8, elements_per_thread=(16,16,8), dtype_in=dtypes.half, dtype_out=dtypes.float,
opts=("l0","l0","l0","u1","u1","u1"),
swizzle=((('r1', 'r2', 'r3'), ('u0', 'u1', 'u2'), ('l0', 'l1', 'l2', 'r0')),
(('l0', 'l1', 'l2'), ('r1', 'r2', 'r3'), ('u0', 'u1', 'u2', 'r0'))))]