File size: 18,392 Bytes
a3f2e4d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 | # Copyright (c) Microsoft Corporation.
# Licensed under the MIT License.
# pylint: disable=invalid-name
"""Reduction rule for operators including softmax, layer norm, RMS norm, etc"""
from typing import List, Union
from functools import reduce
from tvm import tir
from tvm.target import Target
from ..base import normalize_prim_func, try_inline_contiguous_spatial
from ..base.analysis import get_root_block, get_reduction_blocks, BlockInfo
from .base import GPUScheduleRule
class GeneralReduction(GPUScheduleRule):
"""General Reduction rule for operators including softmax, layer norm, RMS norm, etc"""
def apply( # pylint: disable=too-many-locals
self,
func: tir.PrimFunc,
target: Target,
_: bool,
) -> Union[None, tir.Schedule, List[tir.Schedule]]:
if not isinstance(func, tir.PrimFunc) or not self.is_target_available(target):
return None
if target.kind.name == "cuda":
len_tx = 256
unroll_depth = 256
else:
len_tx = 64
unroll_depth = 64
sch = tir.Schedule(func)
block_infos = normalize_prim_func(sch)
block_infos = try_inline_contiguous_spatial(sch, block_infos)
if block_infos is None or len(block_infos) == 0:
return None
dom_kind = block_infos[0].dom_kind()
num_leading_s = len(dom_kind) - len(dom_kind.lstrip("S"))
num_trailing_r = len(dom_kind) - len(dom_kind.rstrip("R"))
# Align the number of block iters of the last block.
num_last_block_iter = len(block_infos[-1].dom_kind())
if num_last_block_iter < len(dom_kind):
index_map = tir.IndexMap.from_func(
lambda *iters: (
[tir.const(0, iters[0].dtype)] * (len(dom_kind) - num_last_block_iter)
+ list(iters)
),
ndim=num_last_block_iter,
)
sch.transform_block_layout(block_infos[-1].block_rv, index_map)
try:
# TODO: fix num_leading_s = 0 case
assert num_trailing_r > 0
for block in block_infos[1:-1]:
assert block.dom_kind() == dom_kind
assert block_infos[-1].is_injective()
assert len(block_infos[-1].dom_kind()) <= len(dom_kind)
except AssertionError:
return None
loops = sch.get_loops(block_infos[-1].block_rv)
bx = sch.fuse(*loops[:num_leading_s])
r_loop, tx = sch.split(loops[-1], [None, len_tx])
sch.reorder(tx, r_loop)
sch.bind(bx, "blockIdx.x")
sch.bind(tx, "threadIdx.x")
sch.annotate(r_loop, ann_key="pragma_auto_unroll_max_step", ann_val=unroll_depth)
sch.annotate(r_loop, ann_key="pragma_unroll_explicit", ann_val=1)
for block in reversed(block_infos[:-1]):
block = block.block_rv
for i, _ in enumerate(sch.get(block).writes):
sch.set_scope(block, buffer_index=i, storage_scope="shared")
sch.compute_at(block, bx, preserve_unit_loops=True)
r_loop = sch.fuse(*sch.get_loops(block)[-num_trailing_r:])
r_loop, tx = sch.split(r_loop, [None, len_tx])
sch.reorder(tx, r_loop)
sch.bind(tx, "threadIdx.x")
sch.annotate(r_loop, ann_key="pragma_auto_unroll_max_step", ann_val=unroll_depth)
sch.annotate(r_loop, ann_key="pragma_unroll_explicit", ann_val=1)
# TODO: It's just a workaround to avoid unroll spatial loops, because of the bug of
# the pass lower-thread-allreduce. We should fix it in the future.
# sch.annotate(bx, ann_key="pragma_auto_unroll_max_step", ann_val=unroll_depth)
# sch.annotate(bx, ann_key="pragma_unroll_explicit", ann_val=1)
return sch
def sch_inner_reduction_with_config( # pylint: disable=too-many-locals,too-many-branches,too-many-return-statements
self,
func: tir.PrimFunc,
config,
):
block_factors = config.block
thread_factors = config.thread
reduce_therad_factors = config.reduce_thread
# For inter thread reduction case, one thread must only compute one element
assert thread_factors == block_factors
# inline all the other blocks
sch = tir.Schedule(func)
block_infos = normalize_prim_func(sch)
schedule_block: tir.schedule.BlockRV = None
reduction_blocks: List[tir.schedule.BlockRV] = []
for block in block_infos:
s_loops: List[tir.schedule.LoopRV] = []
r_loops: List[tir.schedule.LoopRV] = []
o_loops: List[tir.schedule.LoopRV] = []
dom_kind = block.dom_kind()
block_rv = block.block_rv
if (
any(
[
sch.get(loop_rv).thread_binding is not None
for loop_rv in sch.get_loops(block_rv)
]
)
or len(sch.get_loops(block.block_rv)) == 0
):
continue
for loop, iter_type in zip(sch.get_loops(block_rv), dom_kind):
{"S": s_loops, "R": r_loops, "O": o_loops}[iter_type].append(loop)
if not s_loops:
s_loops.append(sch.add_unit_loop(block_rv))
if len(r_loops) > 0:
# always use the last reduction block for scheduling
schedule_block = block
reduction_blocks.append(block_rv)
# Align the number of block iters of the last block.
dom_kind = schedule_block.dom_kind()
num_leading_s = len(dom_kind) - len(dom_kind.lstrip("S"))
num_trailing_r = len(dom_kind) - len(dom_kind.rstrip("R"))
schedule_block = schedule_block.block_rv
loops = sch.get_loops(schedule_block)
s_loops = loops[:num_leading_s]
r_loops = loops[-num_trailing_r:]
block_axis = []
thread_axis = []
for s_loop, block_factor in zip(s_loops, block_factors):
block_loop, thread_loop = sch.split(s_loop, factors=[None, block_factor])
block_axis.append(block_loop)
thread_axis.append(thread_loop)
axis_order = block_axis + thread_axis
sch.reorder(*axis_order)
blck_fused = sch.fuse(*block_axis)
thrd_fused = sch.fuse(*thread_axis)
sch.bind(blck_fused, "blockIdx.x")
sch.bind(thrd_fused, "threadIdx.y")
reduce_outer_axis, reduce_inner_axis, reduce_inter_threads = [], [], []
for i in config.raxis_order:
loop = r_loops[i]
ro, ri = sch.split(loop, factors=[None, config.rstep[i]])
ri, thd = sch.split(ri, factors=[None, config.reduce_thread[i]])
reduce_inter_threads.append(thd)
reduce_outer_axis.append(ro)
reduce_inner_axis.append(ri)
axis_order = reduce_inter_threads + reduce_outer_axis + reduce_inner_axis
sch.reorder(*axis_order)
fused_reduce_inter_threads = sch.fuse(*reduce_inter_threads)
sch.bind(fused_reduce_inter_threads, "threadIdx.x")
def prod(iterable):
return reduce(lambda x, y: x * y, iterable, 1)
reg_tile = sch.cache_write(schedule_block, 0, "local")
# todo(lei): should add the shared_inputs/stride memory pad analysis at shared memory fusion stage.
for i, input_region in enumerate(sch.get(schedule_block).reads):
if input_region.buffer.name not in config.cached_tensors:
continue
# otherwise cooperative fetch in shared memory.
cache_shared = sch.cache_read(schedule_block, i, "shared")
sch.compute_at(cache_shared, reduce_outer_axis[-1])
dim_offset = (
len(reduce_inner_axis) + len(reduce_outer_axis) + 2
) # outer loops are: blck_fused, thrd_fused, vthread_axis, reduce_outer_axis
if input_region.buffer.name in config.vectorize:
vectorize = config.vectorize[input_region.buffer.name]
else:
vectorize = 1
loops = sch.get_loops(cache_shared)
if len(loops) == dim_offset:
# handle fetching only one element
loops.append(sch.add_unit_loop(schedule_block))
assert len(loops) > dim_offset
_, ty, tx, tv = sch.split(
sch.fuse(*loops[dim_offset:]),
factors=[
None,
int(prod(thread_factors)),
int(prod(reduce_therad_factors)),
vectorize,
],
)
sch.vectorize(tv)
sch.bind(ty, "threadIdx.y")
sch.bind(tx, "threadIdx.x")
sch.reverse_compute_at(reg_tile, thrd_fused)
# resolve compute_at
block_infos = try_inline_contiguous_spatial(sch, block_infos)
if block_infos is None or len(block_infos) == 0:
return None
return sch
def sch_outer_reduction_with_config( # pylint: disable=too-many-locals,too-many-branches,too-many-return-statements
self,
func: tir.PrimFunc,
config,
):
block_factors = config.block
thread_factors = config.thread
step_factors = config.step
# inline all the other blocks
sch = tir.Schedule(func)
block_infos = normalize_prim_func(sch)
schedule_block: BlockInfo = None
for block in block_infos:
s_loops: List[tir.schedule.LoopRV] = []
r_loops: List[tir.schedule.LoopRV] = []
o_loops: List[tir.schedule.LoopRV] = []
dom_kind = block.dom_kind()
block_rv = block.block_rv
if (
any(
[
sch.get(loop_rv).thread_binding is not None
for loop_rv in sch.get_loops(block_rv)
]
)
or len(sch.get_loops(block.block_rv)) == 0
):
continue
for loop, iter_type in zip(sch.get_loops(block_rv), dom_kind):
{"S": s_loops, "R": r_loops, "O": o_loops}[iter_type].append(loop)
if not s_loops:
s_loops.append(sch.add_unit_loop(block_rv))
if len(r_loops) > 0:
# always use the last reduction block for scheduling
schedule_block = block
# Align the number of block iters of the last block.
dom_kind = schedule_block.dom_kind()
num_leading_s = len(dom_kind) - len(dom_kind.lstrip("S"))
num_trailing_r = len(dom_kind) - len(dom_kind.rstrip("R"))
num_last_block_iter = len(block_infos[-1].dom_kind())
if num_last_block_iter < len(dom_kind):
index_map = tir.IndexMap.from_func(
lambda *iters: (
[tir.const(0, iters[0].dtype)] * (len(dom_kind) - num_last_block_iter)
+ list(iters)
),
ndim=num_last_block_iter,
)
sch.transform_block_layout(block_infos[-1].block_rv, index_map)
schedule_block = schedule_block.block_rv
loops = sch.get_loops(schedule_block)
s_loops = loops[:num_leading_s]
r_loops = loops[-num_trailing_r:]
reg_tile = sch.cache_write(schedule_block, 0, "local")
block_axis = []
vthread_axis = []
thread_axis = []
inner_axis = []
for s_loop, block_factor, step_factor, thread_factor in zip(
s_loops, block_factors, step_factors, thread_factors
):
block_loop, inner_loop = sch.split(s_loop, factors=[None, block_factor])
vthread_loop, inner_loop = sch.split(
inner_loop, factors=[None, thread_factor * step_factor]
)
thread_loop, inner_loop = sch.split(inner_loop, factors=[None, step_factor])
block_axis.append(block_loop)
vthread_axis.append(vthread_loop)
thread_axis.append(thread_loop)
inner_axis.append(inner_loop)
reduce_outer_axis, reduce_inner_axis = [], []
for i in config.raxis_order:
loop = r_loops[i]
ro, ri = sch.split(loop, factors=[None, config.rstep[i]])
reduce_outer_axis.append(ro)
reduce_inner_axis.append(ri)
vthread_axis = list(reversed(vthread_axis)) # inner virtual thread first
axis_order = (
block_axis
+ vthread_axis
+ thread_axis
+ reduce_outer_axis
+ reduce_inner_axis
+ inner_axis
)
sch.reorder(*axis_order)
blck_fused = sch.fuse(*block_axis)
thrd_fused = sch.fuse(*thread_axis)
sch.bind(blck_fused, "blockIdx.x")
sch.bind(thrd_fused, "threadIdx.x")
if len(vthread_axis) > 3:
vthread_axis = vthread_axis[0:2] + [sch.fuse(*vthread_axis[2:])]
for i, ax in enumerate(vthread_axis):
sch.bind(ax, "vthread" + [".x", ".y", ".z"][i])
# todo(lei): should add the shared_inputs/stride memory pad analysis at shared memory fusion stage.
for i, input_region in enumerate(sch.get(schedule_block).reads):
if input_region.buffer.name not in config.cached_tensors:
continue
# otherwise cooperative fetch in shared memory.
cache_shared = sch.cache_read(schedule_block, i, "shared")
sch.compute_at(cache_shared, reduce_outer_axis[-1])
dim_offset = (
len(vthread_axis) + len(reduce_outer_axis) + 2
) # outer loops are: blck_fused, thrd_fused, vthread_axis, reduce_outer_axis
if input_region.buffer.name in config.vectorize:
vectorize = config.vectorize[input_region.buffer.name]
else:
vectorize = 1
loops = sch.get_loops(cache_shared)
if len(loops) == dim_offset:
# handle fetching only one element
loops.append(sch.add_unit_loop(schedule_block))
assert len(loops) > dim_offset
def prod(iterable):
return reduce(lambda x, y: x * y, iterable, 1)
_, tx, tv = sch.split(
sch.fuse(*loops[dim_offset:]), factors=[None, int(prod(thread_factors)), vectorize]
)
sch.vectorize(tv)
sch.bind(tx, "threadIdx.x")
sch.reverse_compute_at(reg_tile, thrd_fused)
sch.decompose_reduction(schedule_block, reduce_outer_axis[0])
# resolve compute_at
block_infos = try_inline_contiguous_spatial(sch, block_infos)
if block_infos is None or len(block_infos) == 0:
return None
return sch
def sch_mutiple_reductions_with_config( # pylint: disable=too-many-locals,too-many-branches,too-many-return-statements
self,
func: tir.PrimFunc,
config,
):
block_factors = config.block
thread_factors = config.thread
reduce_therad_factors = config.reduce_thread
sch = tir.Schedule(func)
block_infos = normalize_prim_func(sch)
block_infos = try_inline_contiguous_spatial(sch, block_infos)
if block_infos is None or len(block_infos) == 0:
return None
def prod(iterable):
return reduce(lambda x, y: x * y, iterable, 1)
len_tx = prod(thread_factors) * prod(reduce_therad_factors)
block_factor = prod(block_factors)
dom_kind = block_infos[0].dom_kind()
num_leading_s = len(dom_kind) - len(dom_kind.lstrip("S"))
num_trailing_r = len(dom_kind) - len(dom_kind.rstrip("R"))
# Align the number of block iters of the last block.
num_last_block_iter = len(block_infos[-1].dom_kind())
if num_last_block_iter < len(dom_kind):
index_map = tir.IndexMap.from_func(
lambda *iters: (
[tir.const(0, iters[0].dtype)] * (len(dom_kind) - num_last_block_iter)
+ list(iters)
),
ndim=num_last_block_iter,
)
sch.transform_block_layout(block_infos[-1].block_rv, index_map)
try:
# TODO: fix num_leading_s = 0 case
assert num_trailing_r > 0
for block in block_infos[1:-1]:
assert block.dom_kind() == dom_kind
assert block_infos[-1].is_injective()
assert len(block_infos[-1].dom_kind()) <= len(dom_kind)
except AssertionError:
return None
loops = sch.get_loops(block_infos[-1].block_rv)
bx, _ = sch.split(sch.fuse(*loops[:num_leading_s]), factors=[None, block_factor])
r_loop, tx = sch.split(loops[-1], [None, len_tx])
sch.reorder(tx, r_loop)
sch.bind(bx, "blockIdx.x")
sch.bind(tx, "threadIdx.x")
for block in reversed(block_infos[:-1]):
block = block.block_rv
for i, _ in enumerate(sch.get(block).writes):
sch.set_scope(block, buffer_index=i, storage_scope="shared")
sch.compute_at(block, bx, preserve_unit_loops=True)
r_loop = sch.fuse(*sch.get_loops(block)[-num_trailing_r:])
r_loop, tx = sch.split(r_loop, [None, len_tx])
sch.reorder(tx, r_loop)
sch.bind(tx, "threadIdx.x")
return sch
def apply_config( # pylint: disable=too-many-locals,missing-docstring
self,
func: tir.PrimFunc,
config,
) -> tir.Schedule:
# check the number of reduction blocks
sch = tir.Schedule(func)
root_block = get_root_block(sch)
blocks = sch.get_child_blocks(root_block)
reduction_blocks = get_reduction_blocks(sch, blocks)
if len(reduction_blocks) > 1:
# schedule for multiple reduction blocks (e.g. softmax)
return self.sch_mutiple_reductions_with_config(func, config)
if any([t > 1 for t in config.reduce_thread]):
# todo(lei) should implement block reduction schedule
return self.sch_inner_reduction_with_config(func, config)
else:
return self.sch_outer_reduction_with_config(func, config)
|