primrules.py 35.9 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13
# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
14
import functools
15
import math
16 17
import operator
import typing
18 19 20

import paddle

21
from . import primops
22 23 24
from .primops import (add, broadcast, concat, cos, div, eq, erf, exp,
                      fill_const, gather, ge, gt, log, matmul, max, mul, ne,
                      neg, reduce_sum, reshape, scatter_add, select, set_value,
25
                      sin, slice_assign, slice_select, split, sqrt, sub, tanh,
26
                      transpose, bernoulli, rsqrt)
27 28 29 30 31
from .primreg import (REGISTER_JVP, REGISTER_ORIG2PRIM, REGISTER_PRIM2ORIG,
                      REGISTER_TRANSPOSE, lookup_fn, lookup_jvp,
                      lookup_orig2prim, lookup_prim2orig, lookup_transpose,
                      op_position_inputs, op_position_output)
from .utils import INT_DTYPE_2_STRING, get_input_var_list, get_output_var_list
32 33
from paddle.fluid.data_feeder import convert_dtype
from paddle.fluid.framework import convert_np_dtype_to_dtype_
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


def _orig2prim(op, *args):
    _lowerrule = lookup_orig2prim(op.type)
    return _lowerrule(op, *args)


def _prim2orig(op, *args):
    _lowerrule = lookup_prim2orig(op.type)
    return _lowerrule(op, *args)


def _jvp(op, *args):
    _jvprule = lookup_jvp(op.type)
    return _jvprule(op, *args)


def _transpose(op, dot_checker, *args):
    _transposerule = lookup_transpose(op.type)
    return _transposerule(op, dot_checker, *args)


def linear_jvp(op, *args, **kwargs):
    fn = lookup_fn(op.type)
    out_dot = fn(*args, **kwargs)
    return out_dot


## Register orig2prim lower rules
"""
These original ops are fully supported:

elementwise_add
elementwise_sub
elementwise_mul
tanh
fill_zeros_like
71
fill_any_like
72 73 74 75 76
sum
index_select
scale
assign
sqrt
77 78 79 80
log
select
equal
elementwise_pow
81
dropout
82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97

These original ops are partially supported:

matmul_v2
reshape2
concat
slice
p_norm
"""


@REGISTER_ORIG2PRIM('elementwise_add')
def elementwise_add_orig2prim(op, x, y):
    if x.shape != y.shape:
        y = broadcast(y, shape=x.shape)
    if op.attr('Scale_x') - 1.0 > 1e-5:
98 99 100
        scale_x = fill_const(shape=x.shape,
                             dtype=x.dtype,
                             value=op.attr('Scale_x'))
101 102
        x = mul(x, scale_x)
    if op.attr('Scale_y') - 1.0 > 1e-5:
103 104 105
        scale_y = fill_const(shape=y.shape,
                             dtype=y.dtype,
                             value=op.attr('Scale_y'))
106 107 108
        y = mul(y, scale_y)
    z = add(x, y)
    if op.attr('Scale_out') - 1.0 > 1e-5:
109 110 111
        scale_out = fill_const(shape=z.shape,
                               dtype=z.dtype,
                               value=op.attr('Scale_out'))
112 113 114 115 116 117 118 119 120
        z = mul(z, scale_out)
    return z


@REGISTER_ORIG2PRIM('elementwise_sub')
def elementwise_sub_orig2prim(op, x, y):
    if x.shape != y.shape:
        y = broadcast(y, shape=x.shape)
    if op.attr('Scale_x') - 1.0 > 1e-5:
121 122 123
        scale_x = fill_const(shape=x.shape,
                             dtype=x.dtype,
                             value=op.attr('Scale_x'))
124 125
        x = mul(x, scale_x)
    if op.attr('Scale_y') - 1.0 > 1e-5:
126 127 128
        scale_y = fill_const(shape=y.shape,
                             dtype=y.dtype,
                             value=op.attr('Scale_y'))
129 130 131
        y = mul(y, scale_y)
    z = sub(x, y)
    if op.attr('Scale_out') - 1.0 > 1e-5:
132 133 134
        scale_out = fill_const(shape=z.shape,
                               dtype=z.dtype,
                               value=op.attr('Scale_out'))
135 136 137 138 139 140 141 142 143
        z = mul(z, scale_out)
    return z


@REGISTER_ORIG2PRIM('elementwise_mul')
def elementwise_mul_orig2prim(op, x, y):
    if x.shape != y.shape:
        y = broadcast(y, shape=x.shape)
    if op.attr('Scale_x') - 1.0 > 1e-5:
144 145 146
        scale_x = fill_const(shape=x.shape,
                             dtype=x.dtype,
                             value=op.attr('Scale_x'))
147 148
        x = mul(x, scale_x)
    if op.attr('Scale_y') - 1.0 > 1e-5:
149 150 151
        scale_y = fill_const(shape=y.shape,
                             dtype=y.dtype,
                             value=op.attr('Scale_y'))
152 153 154
        y = mul(y, scale_y)
    z = mul(x, y)
    if op.attr('Scale_out') - 1.0 > 1e-5:
155 156 157
        scale_out = fill_const(shape=z.shape,
                               dtype=z.dtype,
                               value=op.attr('Scale_out'))
158 159 160 161
        z = mul(z, scale_out)
    return z


162 163 164 165 166 167 168
@REGISTER_ORIG2PRIM('elementwise_div')
def elementwise_div_orig2prim(op, x, y):
    if x.shape != y.shape:
        y = broadcast(y, shape=x.shape)
    return primops.div(x, y)


169 170 171 172 173
@REGISTER_ORIG2PRIM('tanh')
def tanh_orig2prim(op, x):
    return tanh(x)


174 175 176 177 178 179 180 181 182 183 184 185 186 187 188
@REGISTER_ORIG2PRIM('sin')
def sin_orig2prim(op, x):
    return sin(x)


@REGISTER_ORIG2PRIM('cos')
def cos_orig2prim(op, x):
    return cos(x)


@REGISTER_ORIG2PRIM('exp')
def exp_orig2prim(op, x):
    return exp(x)


189 190 191 192 193
@REGISTER_ORIG2PRIM('erf')
def erf_orig2prim(op, x):
    return erf(x)


194 195 196 197 198
@REGISTER_ORIG2PRIM('abs')
def abs_orig2prim(op, x):
    return primops.abs(x)


199 200 201 202 203
@REGISTER_ORIG2PRIM('log')
def log_orig2prim(op, x):
    return log(x)


204 205 206 207 208
@REGISTER_ORIG2PRIM('fill_zeros_like')
def fill_zeros_like_orig2prim(op, x):
    return fill_const(value=0.0, shape=x.shape, dtype=x.dtype)


209 210 211 212 213 214 215 216 217 218
@REGISTER_ORIG2PRIM('fill_any_like')
def fill_any_like_orig2prim(op, x):
    if op.attr('dtype') == -1:
        return fill_const(value=op.attr('value'), shape=x.shape, dtype=x.dtype)
    return fill_const(value=op.attr('value'),
                      shape=x.shape,
                      dtype=convert_np_dtype_to_dtype_(
                          convert_dtype(INT_DTYPE_2_STRING[op.attr('dtype')])))


219 220 221 222 223 224 225 226 227 228 229 230 231 232
@REGISTER_ORIG2PRIM('fill_constant')
def fill_const_orig2prim(op,
                         shape_tensor=None,
                         shape_tensor_list=None,
                         value_tensor=None):
    if shape_tensor or shape_tensor_list or value_tensor:
        raise TypeError(
            'fill_const_orig2prim currently not support Tensor input of shape and value.'
        )
    return fill_const(value=op.attr('value'),
                      shape=op.attr('shape'),
                      dtype=paddle.dtype(op.attr('dtype')))


233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248
@REGISTER_ORIG2PRIM('sum')
def sum_orig2prim(op, xs):
    x0 = xs[0]
    for x in xs[1:]:
        x0 = add(x0, x)
    return x0


@REGISTER_ORIG2PRIM('index_select')
def index_select_orig2prim(op, index_t, x):
    return gather(x, indextensor=index_t, axis=op.attr('dim'))


@REGISTER_ORIG2PRIM('scale')
def scale_orig2prim(op, scale_t, x):
    if scale_t is None:
249 250 251
        scale_t = fill_const(shape=x.shape,
                             dtype=x.dtype,
                             value=op.attr('scale'))
252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269
    bias_t = fill_const(shape=x.shape, dtype=x.dtype, value=op.attr('bias'))
    if op.attr('bias_after_scale'):
        return add(mul(x, scale_t), bias_t)
    else:
        return mul(add(x, bias_t), scale_t)


@REGISTER_ORIG2PRIM('assign')
def assign_orig2prim(op, x):
    zero_t = fill_const(shape=x.shape, dtype=x.dtype, value=0.0)
    return add(x, zero_t)


@REGISTER_ORIG2PRIM('sqrt')
def sqrt_orig2prim(op, x):
    return sqrt(x)


J
Jiabin Yang 已提交
270 271 272 273 274
@REGISTER_ORIG2PRIM('rsqrt')
def rsqrt_orig2prim(op, x):
    return rsqrt(x)


275 276
@REGISTER_ORIG2PRIM('matmul_v2')
def matmul_v2_orig2prim(op, x, y):
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
    def trans(shape):
        ret = [i for i in range(len(shape))]
        ret[-1], ret[-2] = ret[-2], ret[-1]
        return ret

    assert len(x.shape) < 4 and len(
        y.shape) < 4, 'Do not support multi batchsize dimensions currently.'

    if len(x.shape) == 1:
        x = broadcast(x, shape=[1, x.shape[0]])
    if len(y.shape) == 1:
        y = broadcast(y, shape=[y.shape[0], 1])
    if op.attr('trans_x'):
        x = transpose(x, axis=trans(x.shape))
    if op.attr('trans_y'):
        y = transpose(y, axis=trans(y.shape))
    return matmul(x, y)


## NOTE(lml): The second output of reshape2 Xshape, which is only used in reshape2_grad, is meanlingless in new autograd mechanism, thus we use a zero tensor instead.
@REGISTER_ORIG2PRIM('reshape2')
def reshape2_orig2prim(op, shape_t, shape_tl, x):
    assert shape_t is None, 'Can not lower reshape2 into prim ops with shapetensor.'
    assert shape_tl is None, 'Can not lower reshape2 into prim ops with shapetensorlist.'
    y, xshape = get_output_var_list(op)
303 304 305
    return reshape(x, shape=y.shape), fill_const(shape=xshape.shape,
                                                 dtype=xshape.dtype,
                                                 value=0.0)
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


@REGISTER_ORIG2PRIM('concat')
def concat_orig2prim(op, axis_t, xs):
    assert axis_t is None, 'Can not lower concat into prim ops with axistensor.'
    return concat(xs, axis=op.attr('axis'))


@REGISTER_ORIG2PRIM('slice')
def slice_orig2prim(op, ends_t, ends_tl, x, starts_t, starts_tl):
    assert starts_t is None, 'Can not lower concat into prim ops with startstensor.'
    assert ends_t is None, 'Can not lower concat into prim ops with endstensor.'
    assert starts_tl is None, 'Can not lower concat into prim ops with startstensorlist.'
    assert ends_tl is None, 'Can not lower concat into prim ops with endstensorlist.'
    starts = op.attr('starts')
    ends = op.attr('ends')
    strides = [1 for _ in starts]
    axis = op.attr('axes')
    y = slice_select(x, starts=starts, ends=ends, strides=strides, axis=axis)
    if op.attr('decrease_axis'):
        y = reshape(y, shape=get_output_var_list(op)[0].shape)
    return y


@REGISTER_ORIG2PRIM('p_norm')
def p_norm_orig2prim(op, x):
332

333 334 335 336 337 338 339 340 341 342 343 344
    def num_el(shape):
        n = 1
        for s in shape:
            n = n * s
        return n

    assert op.attr(
        'asvector'), 'Only support lower pnorm when asvector=True currently'
    if len(x.shape) > 1:
        x = reshape(x, shape=[num_el(x.shape)])

    if abs(op.attr('porder') - 2.0) < 1e-5:
345
        return sqrt(reduce_sum(mul(x, x), axis=[0]))
346
    elif abs(op.attr('porder') - 1.0) < 1e-5:
347
        return reduce_sum(sqrt(mul(x, x)), axis=[0])
348 349 350 351
    else:
        raise RuntimeError('Only support lower l2/l1 norm currently')


352 353 354 355 356
@REGISTER_ORIG2PRIM('cast')
def cast_orig2prim(op, x):
    return primops.cast(x, paddle.dtype(op.attr('out_dtype')))


357 358 359 360 361 362 363 364 365 366 367 368 369
# TODO: support broadcast
@REGISTER_ORIG2PRIM('where')
def select_orig2prim(op, condition, x, y):
    return select(condition, x, y)


@REGISTER_ORIG2PRIM('equal')
def equal_orig2prim(op, x, y):
    if x.shape != y.shape:
        y = broadcast(y, shape=x.shape)
    return eq(x, y)


370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390
@REGISTER_ORIG2PRIM('not_equal')
def ne_orig2prim(op, x, y):
    if x.shape != y.shape:
        y = broadcast(y, shape=x.shape)
    return ne(x, y)


@REGISTER_ORIG2PRIM('greater_than')
def gt_orig2prim(op, x, y):
    if x.shape != y.shape:
        y = broadcast(y, shape=x.shape)
    return gt(x, y)


@REGISTER_ORIG2PRIM('greater_equal')
def ge_orig2prim(op, x, y):
    if x.shape != y.shape:
        y = broadcast(y, shape=x.shape)
    return ge(x, y)


391
# paddle.pow API use "elementwise_pow" operator when y is a Tensor.
392 393 394 395 396 397 398 399
@REGISTER_ORIG2PRIM('elementwise_pow')
def elementwise_pow_orig2prim(op, x, y):
    if x.shape != y.shape:
        y = broadcast(y, shape=x.shape)
    z = primops.pow(x, y)
    return z


400 401 402 403 404 405 406 407 408
# paddle.pow API use "pow" operator when y is a scalar.
@REGISTER_ORIG2PRIM('pow')
def pow_orig2prim(op, x, y):
    # x is factorTensor defined in paddle phi op. Currently it is None.
    return primops.pow(y, fill_const(op.attr('factor'), y.shape, y.dtype))


@REGISTER_ORIG2PRIM('square')
def square_orig2prim(op, x):
409
    return primops.square(x)
410 411


412 413 414 415 416 417 418
@REGISTER_ORIG2PRIM('elementwise_max')
def elementwise_max_orig2prim(op, x, y):
    if x.shape != y.shape:
        y = broadcast(y, shape=x.shape)
    return primops.max(x, y)


419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440
@REGISTER_ORIG2PRIM('gelu')
def gelu_orig2prim(op, x):
    if op.attr('approximate'):
        cdf = mul(
            fill_const(0.5, x.shape, x.dtype),
            add(
                fill_const(1.0, x.shape, x.dtype),
                tanh(
                    mul(
                        fill_const(math.sqrt(2 / math.pi), x.shape, x.dtype),
                        add(
                            x,
                            mul(
                                fill_const(0.044715, x.shape, x.dtype),
                                primops.pow(x, fill_const(3., x.shape,
                                                          x.dtype))))))))
        return mul(x, cdf)
    else:
        return mul(
            mul(fill_const(0.5, x.shape, x.dtype), x),
            add(fill_const(1.0, x.shape, x.dtype),
                erf(mul(x, fill_const(1 / math.sqrt(2.), x.shape, x.dtype)))))
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
@REGISTER_ORIG2PRIM('dropout')
def dropout_orig2prim(op, seed_t, x):
    assert seed_t is None, 'Can not lower dropout into prim ops with seedtensor.'
    mask = bernoulli(shape=x.shape, dtype=x.dtype, p=op.attr('dropout_prob'))
    if op.attr('dropout_implementation') == 'upscale_in_train':
        if op.attr('is_test') == False:
            out = div(
                mul(x, mask),
                fill_const(1.0 - op.attr('dropout_prob'), x.shape, x.dtype))
            return primops.cast(mask, dtype=paddle.uint8), out
        else:
            return primops.cast(mask, dtype=paddle.uint8), x
    elif op.attr('dropout_implementation') == 'downgrade_in_infer':
        if op.attr('is_test') == False:
            return primops.cast(mask, dtype=paddle.uint8), mul(x, mask)
        else:
            return primops.cast(mask, dtype=paddle.uint8), mul(
                x, fill_const(1.0 - op.attr('dropout_prob'), x.shape, x.dtype))
    else:
        raise RuntimeError(
            'Unsupported dropout_implementation, only support upscale_in_train and downgrade_in_infer'
        )


467 468 469 470 471 472 473 474 475 476 477
@REGISTER_ORIG2PRIM('reduce_sum')
def reduce_sum_orig2prim(op, x):
    axes = tuple(range(0, len(
        x.shape))) if op.attr('reduce_all') else op.attr('dim')
    return reduce_sum(x, axis=axes, keepdim=op.attr('keep_dim'))


@REGISTER_ORIG2PRIM('reduce_mean')
def reduce_mean_orig2prim(op, x):
    axes = tuple(range(0, len(
        x.shape))) if op.attr('reduce_all') else op.attr('dim')
478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506
    return primops.mean(x, axes, op.attr('keep_dim'))


@REGISTER_ORIG2PRIM('batch_norm')
def batch_norm_orig2prim(op, bias, run_mean, momentum_tensor, scale, run_var,
                         x):
    momentum = op.attr('momentum')
    eps = op.attr('epsilon')
    is_test = op.attr('is_test')
    data_layout = op.attr('data_layout')
    use_global_stats = op.attr('use_global_stats')
    trainable_statistics = op.attr('trainable_statistics')
    reserve_space = None if len(
        op.output_names) == 5 else get_output_var_list(op)[1]

    feature_axis = 1 if data_layout in ('NC', 'NCL', 'NCHW',
                                        'NCHWD') else len(x.shape) - 1
    use_run_stat = (is_test and (not trainable_statistics)) or use_global_stats

    return primops.batch_norm(x,
                              feature_axis,
                              scale,
                              bias,
                              run_mean,
                              run_var,
                              eps=eps,
                              momentum=momentum,
                              use_run_stat=use_run_stat,
                              reserve_space=reserve_space)
507 508


509 510 511 512 513 514
@REGISTER_ORIG2PRIM('size')
def size_orig2prim(op, x):
    return fill_const(functools.reduce(operator.mul, x.shape), (1, ),
                      paddle.int64)


515
## Register prim2orig lower rules
516 517 518 519 520 521 522 523 524 525
@REGISTER_PRIM2ORIG('add_p')
def add_prim2orig(op, x, y):
    return paddle.add(x, y)


@REGISTER_PRIM2ORIG('sub_p')
def sub_prim2orig(op, x, y):
    return paddle.subtract(x, y)


J
Jiabin Yang 已提交
526 527 528 529 530
@REGISTER_PRIM2ORIG('rsqrt_p')
def rsqrt_prim2orig(op, x):
    return paddle.rsqrt(x)


531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550
@REGISTER_PRIM2ORIG('mul_p')
def mul_prim2orig(op, x, y):
    return paddle.multiply(x, y)


@REGISTER_PRIM2ORIG('div_p')
def div_prim2orig(op, x, y):
    return paddle.divide(x, y)


@REGISTER_PRIM2ORIG('sqrt_p')
def sqrt_prim2orig(op, x):
    return paddle.sqrt(x)


@REGISTER_PRIM2ORIG('tanh_p')
def tanh_prim2orig(op, x):
    return paddle.tanh(x)


551 552 553 554 555 556 557 558 559 560 561 562 563 564 565
@REGISTER_PRIM2ORIG('sin_p')
def sin_prim2orig(op, x):
    return paddle.sin(x)


@REGISTER_PRIM2ORIG('cos_p')
def cos_prim2orig(op, x):
    return paddle.cos(x)


@REGISTER_PRIM2ORIG('exp_p')
def exp_prim2orig(op, x):
    return paddle.exp(x)


566 567 568 569 570
@REGISTER_PRIM2ORIG('erf_p')
def erf_prim2orig(op, x):
    return paddle.erf(x)


571 572 573 574 575
@REGISTER_PRIM2ORIG('abs_p')
def abs_prim2orig(op, x):
    return paddle.abs(x)


576 577 578 579 580
@REGISTER_PRIM2ORIG('log_p')
def log_prim2orig(op, x):
    return paddle.log(x)


581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600
@REGISTER_PRIM2ORIG('reshape_p')
def reshape_prim2orig(op, x):
    return paddle.reshape(x, shape=op.attr('shape'))


@REGISTER_PRIM2ORIG('broadcast_p')
def broadcast_prim2orig(op, x):
    return paddle.broadcast_to(x, shape=op.attr('shape'))


@REGISTER_PRIM2ORIG('transpose_p')
def transpose_prim2orig(op, x):
    return paddle.transpose(x, perm=op.attr('axis'))


@REGISTER_PRIM2ORIG('split_p')
def split_prim2orig(op, x):
    num_or_sections = op.attr('num_or_sections')
    if len(num_or_sections) == 1:
        num_or_sections = num_or_sections[0]
601 602 603
    return paddle.split(x,
                        num_or_sections=num_or_sections,
                        axis=op.attr('axis'))
604 605 606 607 608 609 610


@REGISTER_PRIM2ORIG('concat_p')
def concat_prim2orig(op, xs):
    return paddle.concat(xs, axis=op.attr('axis'))


611
@REGISTER_PRIM2ORIG('reduce_sum_p')
612 613 614 615 616 617 618 619 620 621 622
def reduce_prim2orig(op, x):
    return paddle.sum(x, axis=op.attr('axis'), keepdim=op.attr('keepdim'))


@REGISTER_PRIM2ORIG('matmul_p')
def matmul_prim2orig(op, x, y):
    return paddle.matmul(x, y)


@REGISTER_PRIM2ORIG('slice_select_p')
def slice_select_prim2orig(op, x):
623 624 625 626 627
    return paddle.strided_slice(x,
                                axes=op.attr('axis'),
                                starts=op.attr('starts'),
                                ends=op.attr('ends'),
                                strides=op.attr('strides'))
628 629 630 631 632


@REGISTER_PRIM2ORIG('slice_assign_p')
def slice_assign_prim2orig(op, x, y):
    x_copy = paddle.assign(x)
633 634 635 636 637 638 639
    return set_value(x_copy,
                     y,
                     axis=op.attr('axis'),
                     starts=op.attr('starts'),
                     ends=op.attr('ends'),
                     strides=op.attr('strides'),
                     out=x_copy)
640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656


@REGISTER_PRIM2ORIG('gather_p')
def gather_prim2orig(op, index_t, x):
    return paddle.gather(x, index_t, axis=op.attr('axis'))


@REGISTER_PRIM2ORIG('scatter_add_p')
def scatter_add_prim2orig(op, index_t, x, y):
    assert op.attr('axis') == 0, 'Only support axis==0 currently'
    zeros = paddle.zeros_like(x=x, dtype=x.dtype)
    tmp = paddle.scatter(x=zeros, index=index_t, updates=y, overwrite=False)
    return paddle.add(x, tmp)


@REGISTER_PRIM2ORIG('fill_constant_p')
def fill_constant_prim2orig(op):
657 658 659
    return paddle.full(shape=op.attr('shape'),
                       fill_value=op.attr('value'),
                       dtype=INT_DTYPE_2_STRING[op.attr('dtype')])
660 661


662 663 664 665 666 667 668 669
@REGISTER_PRIM2ORIG('bernoulli_p')
def bernoulli_prim2orig(op):
    t = paddle.full(shape=op.attr('shape'),
                    fill_value=op.attr('p'),
                    dtype=INT_DTYPE_2_STRING[op.attr('dtype')])
    return paddle.bernoulli(t)


670 671 672 673 674 675 676 677 678 679
@REGISTER_PRIM2ORIG('select_p')
def select_prim2orig(op, condition, x, y):
    return paddle.where(condition, x, y)


@REGISTER_PRIM2ORIG('eq_p')
def eq_prim2orig(op, x, y):
    return paddle.equal(x, y)


680 681 682 683 684 685 686 687 688 689 690 691 692 693 694
@REGISTER_PRIM2ORIG('gt_p')
def gt_prim2orig(op, x, y):
    return paddle.greater_than(x, y)


@REGISTER_PRIM2ORIG('ge_p')
def ge_prim2orig(op, x, y):
    return paddle.greater_equal(x, y)


@REGISTER_PRIM2ORIG('ne_p')
def ne_prim2orig(op, x, y):
    return paddle.not_equal(x, y)


695 696 697 698 699
@REGISTER_PRIM2ORIG('pow_p')
def pow_prim2orig(op, x, y):
    return paddle.pow(x, y)


700 701 702 703 704
@REGISTER_PRIM2ORIG('max_p')
def max_prim2orig(op, x, y):
    return paddle.maximum(x, y)


705 706 707 708 709
@REGISTER_PRIM2ORIG('cast_p')
def cast_prim2orig(op, x):
    return paddle.cast(x, paddle.dtype(op.attr('dtype')))


710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780
## Register linearize rules
@REGISTER_JVP('add_p')
def add_jvp(op, x_dot, y_dot):
    if x_dot is None:
        return y_dot
    elif y_dot is None:
        return x_dot
    else:
        return linear_jvp(op, x_dot, y_dot)


@REGISTER_JVP('sub_p')
def sub_jvp(op, x_dot, y_dot):
    if x_dot is None:
        return neg(y_dot)
    elif y_dot is None:
        return x_dot
    else:
        return linear_jvp(op, x_dot, y_dot)


@REGISTER_JVP('mul_p')
def mul_jvp(op, x_dot, y_dot):
    if x_dot is None and y_dot is None:
        return None
    x, y = op_position_inputs(op)
    if x_dot is None:
        return mul(x, y_dot)
    elif y_dot is None:
        return mul(x_dot, y)
    else:
        t1, t2 = mul(x_dot, y), mul(x, y_dot)
        z_dot = add(t1, t2)
        return z_dot


@REGISTER_JVP('div_p')
def div_jvp(op, x_dot, y_dot):
    if x_dot is None and y_dot is None:
        return None
    x, y = op_position_inputs(op)
    if y_dot is None:
        return div(x_dot, y)
    elif x_dot is None:
        return neg(div(mul(x, y_dot), mul(y, y)))
    else:
        t1 = div(x_dot, y)
        t2 = div(mul(x, y_dot), mul(y, y))
        return sub(t1, t2)


@REGISTER_JVP('sqrt_p')
def sqrt_jvp(op, x_dot):
    if x_dot is None:
        return None
    y = op_position_output(op)
    c2 = fill_const(value=2.0, shape=y.shape, dtype=y.dtype)
    y_dot = div(x_dot, mul(c2, y))
    return y_dot


@REGISTER_JVP('tanh_p')
def tanh_jvp(op, x_dot):
    if x_dot is None:
        return None
    y = op_position_output(op)
    c1 = fill_const(value=1.0, shape=y.shape, dtype=y.dtype)
    y_dot = mul(x_dot, sub(c1, mul(y, y)))
    return y_dot


781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804
@REGISTER_JVP('sin_p')
def sin_jvp(op, x_dot):
    if x_dot is None:
        return None
    x, = op_position_inputs(op)
    return mul(x_dot, cos(x))


@REGISTER_JVP('cos_p')
def cos_jvp(op, x_dot):
    if x_dot is None:
        return None
    x, = op_position_inputs(op)
    return mul(x_dot, neg(sin(x)))


@REGISTER_JVP('exp_p')
def exp_jvp(op, x_dot):
    if x_dot is None:
        return None
    y = op_position_output(op)
    return mul(x_dot, y)


805 806 807 808 809 810 811 812 813 814
@REGISTER_JVP('erf_p')
def erf_jvp(op, x_dot):
    if x_dot is None:
        return None
    x, = op_position_inputs(op)
    return mul(
        fill_const(2. / math.sqrt(math.pi), x.shape, x.dtype),
        mul(x_dot, exp(neg(primops.pow(x, fill_const(2., x.shape, x.dtype))))))


815 816 817 818 819 820 821 822
@REGISTER_JVP('abs_p')
def abs_jvp(op, x_dot):
    if x_dot is None:
        return None
    x, = op_position_inputs(op)
    return select(ge(x, fill_const(0., x.shape, x.dtype)), x_dot, neg(x_dot))


823 824 825 826 827 828 829 830
@REGISTER_JVP('log_p')
def log_jvp(op, x_dot):
    if x_dot is None:
        return None
    x, = op_position_inputs(op)
    return div(x_dot, x)


831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871
@REGISTER_JVP('reshape_p')
def reshape_jvp(op, x_dot):
    if x_dot is None:
        return None
    shape = op.attr('shape')
    return linear_jvp(op, x_dot, shape=shape)


@REGISTER_JVP('broadcast_p')
def broadcast_jvp(op, x_dot):
    if x_dot is None:
        return None
    shape = op.attr('shape')
    return linear_jvp(op, x_dot, shape=shape)


@REGISTER_JVP('transpose_p')
def transpose_jvp(op, x_dot):
    if x_dot is None:
        return None
    axis = op.attr('axis')
    return linear_jvp(op, x_dot, axis=axis)


@REGISTER_JVP('split_p')
def split_jvp(op, x_dot):
    if x_dot is None:
        return None
    num_or_sections = op.attr('num_or_sections')
    axis = op.attr('axis')
    return linear_jvp(op, x_dot, num_or_sections=num_or_sections, axis=axis)


@REGISTER_JVP('concat_p')
def concat_jvp(op, xs_dot):
    if xs_dot is None:
        return None
    axis = op.attr('axis')
    return linear_jvp(op, xs_dot, axis=axis)


872 873
@REGISTER_JVP('reduce_sum_p')
def reduce_sum_jvp(op, x_dot):
874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903
    if x_dot is None:
        return None
    axis = op.attr('axis')
    keepdim = op.attr('keepdim')
    return linear_jvp(op, x_dot, axis=axis, keepdim=keepdim)


@REGISTER_JVP('matmul_p')
def matmul_jvp(op, x_dot, y_dot):
    if x_dot is None and y_dot is None:
        return None
    x, y = op_position_inputs(op)
    if x_dot is None:
        return matmul(x, y_dot)
    elif y_dot is None:
        return matmul(x_dot, y)
    else:
        t1 = matmul(x, y_dot)
        t2 = matmul(x_dot, y)
        return add(t1, t2)


@REGISTER_JVP('slice_select_p')
def slice_select_jvp(op, x_dot):
    if x_dot is None:
        return x_dot
    axis = op.attr('axis')
    starts = op.attr('starts')
    ends = op.attr('ends')
    strides = op.attr('strides')
904 905 906 907 908 909
    return linear_jvp(op,
                      x_dot,
                      axis=axis,
                      starts=starts,
                      ends=ends,
                      strides=strides)
910 911 912 913 914 915 916 917 918 919 920 921 922


@REGISTER_JVP('slice_assign_p')
def slice_assign_jvp(op, x_dot, y_dot):
    if x_dot is None:
        assert y_dot is None, 'y_dot must be None.'
        return None
    else:
        assert y_dot is not None, 'y_dot should not be None.'
    axis = op.attr('axis')
    starts = op.attr('starts')
    ends = op.attr('ends')
    strides = op.attr('strides')
923 924 925 926 927 928 929
    return linear_jvp(op,
                      x_dot,
                      y_dot,
                      axis=axis,
                      starts=starts,
                      ends=ends,
                      strides=strides)
930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949


@REGISTER_JVP('gather_p')
def gather_jvp(op, x_dot, indextensor):
    if x_dot is None:
        return None
    _, indextensor = op_position_inputs(op)
    axis = op.attr('axis')
    return linear_jvp(op, x_dot, indextensor, axis=axis)


@REGISTER_JVP('scatter_add_p')
def scatter_add_jvp(op, x_dot, y_dot):
    if x_dot is None:
        return None
    _, _, indextensor = op_position_inputs(op)
    axis = op.attr('axis')
    return linear_jvp(op, x_dot, y_dot, indextensor, axis=axis)


950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971
@REGISTER_JVP('select_p')
def select_jvp(op, cond_dot, x_dot, y_dot):
    if x_dot is None and y_dot is None:
        return None

    cond, x, y = op_position_inputs(op)
    if x_dot is None:
        x_dot = fill_const(value=0.0, shape=y.shape, dtype=y.dtype)
    if y_dot is None:
        y_dot = fill_const(value=0.0, shape=y.shape, dtype=y.dtype)
    return select(cond, x_dot, y_dot)


@REGISTER_JVP('eq_p')
def eq_jvp(op, x_dot, y_dot):
    if x_dot is None and y_dot is None:
        return None
    x, _ = op_position_inputs(op)
    z_dot = fill_const(value=0., shape=x.shape, dtype=x.dtype)
    return z_dot


972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998
@REGISTER_JVP('gt_p')
def gt_jvp(op, x_dot, y_dot):
    if x_dot is None and y_dot is None:
        return None
    x, _ = op_position_inputs(op)
    z_dot = fill_const(value=0., shape=x.shape, dtype=x.dtype)
    return z_dot


@REGISTER_JVP('ge_p')
def ge_jvp(op, x_dot, y_dot):
    if x_dot is None and y_dot is None:
        return None
    x, _ = op_position_inputs(op)
    z_dot = fill_const(value=0., shape=x.shape, dtype=x.dtype)
    return z_dot


@REGISTER_JVP('ne_p')
def ne_jvp(op, x_dot, y_dot):
    if x_dot is None and y_dot is None:
        return None
    x, _ = op_position_inputs(op)
    z_dot = fill_const(value=0., shape=x.shape, dtype=x.dtype)
    return z_dot


999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025
@REGISTER_JVP('pow_p')
def pow_jvp(op, x_dot, y_dot):

    def _compute_t1(x, y):
        zero_y = fill_const(value=0.0, shape=y.shape, dtype=y.dtype)
        one_y = fill_const(value=1.0, shape=y.shape, dtype=y.dtype)

        cond = eq(y, zero_y)
        new_y = select(cond, one_y, sub(y, one_y))
        t1 = mul(x_dot, mul(y, primops.pow(x, new_y)))
        return t1

    if x_dot is None and y_dot is None:
        return None
    x, y = op_position_inputs(op)
    z = op_position_output(op)

    if y_dot is None:
        return _compute_t1(x, y)
    elif x_dot is None:
        return mul(y_dot, mul(log(x), z))
    else:
        t1, t2 = _compute_t1(x, y), mul(y_dot, mul(log(x), z))
        z_dot = add(t1, t2)
        return z_dot


1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045
@REGISTER_JVP('max_p')
def max_jvp(op, x_dot, y_dot):
    if x_dot is None and y_dot is None:
        return None

    x, y = op_position_inputs(op)
    z = op_position_output(op)
    z_zeros = fill_const(value=0.0, shape=z.shape, dtype=z.dtype)

    # To make the grad of max_p consistent with paddle.maximum when x==y,
    # we just let z_dot = y_dot when compute z_dot to y and x==y,
    # instead of using balance_eq like Jax.
    if y_dot is None:
        return select(eq(y, z), z_zeros, x_dot)
    elif x_dot is None:
        return select(eq(y, z), y_dot, z_zeros)
    else:
        return select(eq(y, z), y_dot, x_dot)


1046 1047 1048 1049 1050 1051
@REGISTER_JVP('cast_p')
def cast_jvp(op, x_dot):
    y = op_position_output(op)
    return primops.cast(x_dot, y.dtype)


J
Jiabin Yang 已提交
1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062
@REGISTER_JVP('rsqrt_p')
def rsqrt_jvp(op, x_dot):
    if x_dot is None:
        return None
    y = op_position_output(op)
    x = op_position_inputs(op)
    c2 = fill_const(value=-2.0, shape=y.shape, dtype=y.dtype)
    y_dot = mul(x_dot, div(div(y, x), c2))
    return y_dot


1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123
## Register transpose rules


@REGISTER_TRANSPOSE('add_p')
def add_transpose(op, check_dot, z_bar):
    x, y = op_position_inputs(op)
    assert check_dot(x) or check_dot(y), (
        f'(check_dot(x) or check_dot(y)) must be True, '
        f'but check_dot(x)={check_dot(x)} and check_dot(y)={check_dot(y)}.')
    x_bar = z_bar if check_dot(x) else None
    y_bar = z_bar if check_dot(y) else None
    return x_bar, y_bar


@REGISTER_TRANSPOSE('sub_p')
def sub_transpose(op, check_dot, z_bar):
    x, y = op_position_inputs(op)
    assert check_dot(x) or check_dot(y), (
        f'(check_dot(x) or check_dot(y)) must be True, '
        f'but check_dot(x)={check_dot(x)} and check_dot(y)={check_dot(y)}.')
    x_bar = z_bar if check_dot(x) else None
    y_bar = neg(z_bar) if check_dot(y) else None
    return x_bar, y_bar


@REGISTER_TRANSPOSE('mul_p')
def mul_transpose(op, check_dot, z_bar):
    x, y = op_position_inputs(op)
    assert check_dot(x) ^ check_dot(y), (
        f'(check_dot(x) ^ check_dot(y)) must be True, '
        f'but check_dot(x)={check_dot(x)} and check_dot(y)={check_dot(y)}.')
    if check_dot(x):
        return mul(z_bar, y), None
    else:
        return None, mul(x, z_bar)


@REGISTER_TRANSPOSE('div_p')
def div_transpose(op, check_dot, z_bar):
    x, y = op_position_inputs(op)
    assert not check_dot(y), 'check_dot(y) must be False'
    x_bar = div(z_bar, y) if check_dot(x) else None
    return x_bar, None


@REGISTER_TRANSPOSE('reshape_p')
def reshape_transpose(op, check_dot, y_bar):
    x, = op_position_inputs(op)
    assert check_dot(x), 'check_dot(x) must be True'
    return reshape(y_bar, shape=x.shape)


@REGISTER_TRANSPOSE('broadcast_p')
def broadcast_transpose(op, check_dot, y_bar):
    x, = op_position_inputs(op)
    assert check_dot(x), 'check_dot(x) must be True'
    bat = len(y_bar.shape) - len(x.shape)
    axis = list(range(bat))
    keepdim = [(bat + i) for i, s in enumerate(x.shape) if s == 1]
    axis += keepdim
    # TODO: Change it. keepdim boolean
1124
    out = reduce_sum(y_bar, axis=axis, keepdim=False)
1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147
    return reshape(out, x.shape)


@REGISTER_TRANSPOSE('transpose_p')
def transpose_transpose(op, check_dot, y_bar):
    x, = op_position_inputs(op)
    assert check_dot(x), 'check_dot(x) must be True'
    axis = op.attr('axis')
    reordered = sorted((k, i) for i, k in enumerate(axis))
    axis = [i for k, i in reordered]
    return transpose(y_bar, axis=axis)


@REGISTER_TRANSPOSE('split_p')
def split_transpose(op, check_dot, ys_bar):
    x, = op_position_inputs(op)
    assert check_dot(x), 'check_dot(x) must be True'
    return concat(ys_bar, axis=op.attr('axis'))


@REGISTER_TRANSPOSE('concat_p')
def concat_transpose(op, check_dot, y_bar):
    xs, = op_position_inputs(op)
1148 1149
    if not isinstance(xs, typing.Sequence):
        xs = [xs]
1150 1151 1152 1153
    for x in xs:
        assert check_dot(x), 'check_dot(x) must be True'
    axis = op.attr('axis')
    sections = [x.shape[axis] for x in xs]
1154 1155
    if len(sections) == 1:
        return y_bar
1156 1157 1158
    return split(y_bar, num_or_sections=sections, axis=axis)


1159 1160
@REGISTER_TRANSPOSE('reduce_sum_p')
def reduce_sum_transpose(op, check_dot, y_bar):
1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191
    x, = op_position_inputs(op)
    assert check_dot(x), 'check_dot(x) must be True'
    axes = op.attr('axis')
    shape = tuple(1 if i in axes else size for i, size in enumerate(x.shape))
    t = reshape(y_bar, shape=shape)
    return broadcast(t, shape=x.shape)


@REGISTER_TRANSPOSE('matmul_p')
def matmul_transpose(op, check_dot, z_bar):
    x, y = op_position_inputs(op)
    assert check_dot(x) ^ check_dot(y), (
        f'(check_dot(x) ^ check_dot(y)) must be True, '
        f'but check_dot(x)={check_dot(x)} and check_dot(y)={check_dot(y)}.')
    # TODO: replace it. this is hacky
    axis = [1, 0] if len(x.shape) == 2 else [0, 2, 1]
    if check_dot(x):
        return matmul(z_bar, transpose(y, axis=axis)), None
    else:
        return None, matmul(transpose(x, axis=axis), z_bar)


@REGISTER_TRANSPOSE('slice_select_p')
def slice_select_transpose(op, check_dot, y_bar):
    x, = op_position_inputs(op)
    assert check_dot(x), 'check_dot(x) must be True'
    zeros = fill_const(value=0.0, shape=x.shape, dtype=x.dtype)
    axis = op.attr('axis')
    starts = op.attr('starts')
    ends = op.attr('ends')
    strides = op.attr('strides')
1192 1193 1194 1195 1196 1197
    return slice_assign(zeros,
                        y_bar,
                        axis=axis,
                        starts=starts,
                        ends=ends,
                        strides=strides)
1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210


@REGISTER_TRANSPOSE('slice_assign_p')
def slice_assign_transpose(op, check_dot, z_bar):
    x, y = op_position_inputs(op)
    assert check_dot(x) and check_dot(y), (
        f'(check_dot(x) and check_dot(y)) must be True, '
        f'but check_dot(x)={check_dot(x)} and check_dot(y)={check_dot(y)}.')
    zeros = fill_const(value=0.0, shape=y.shape, dtype=y.dtype)
    axis = op.attr('axis')
    starts = op.attr('starts')
    ends = op.attr('ends')
    strides = op.attr('strides')
1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221
    x_bar = slice_assign(z_bar,
                         zeros,
                         axis=axis,
                         starts=starts,
                         ends=ends,
                         strides=strides)
    y_bar = slice_select(z_bar,
                         axis=axis,
                         starts=starts,
                         ends=ends,
                         strides=strides)
1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247
    return x_bar, y_bar


@REGISTER_TRANSPOSE('gather_p')
def gather_transpose(op, check_dot, y_bar):
    x, indextensor = op_position_inputs(op)
    assert check_dot(x), 'check_dot(x) must be True'
    axis = op.attr('axis')
    zeros = fill_const(0.0, x.shape, x.dtype)
    x_bar = scatter_add(zeros, y_bar, indextensor, axis=axis)
    indextensor_bar = None
    return x_bar, indextensor_bar


@REGISTER_TRANSPOSE('scatter_add_p')
def scatter_add_transpose(op, check_dot, z_bar):
    x, y, indextensor = op_position_inputs(op)
    assert check_dot(x) and check_dot(y), (
        f'(check_dot(x) and check_dot(y)) must be True, '
        f'but check_dot(x)={check_dot(x)} and check_dot(y)={check_dot(y)}.')
    axis = op.attr('axis')
    zeros = fill_const(value=0.0, shape=y.shape, dtype=y.dtype)
    x_bar = scatter_add(z_bar, zeros, indextensor, axis=axis)
    y_bar = gather(z_bar, indextensor, axis=axis)
    indextensor_bar = None
    return x_bar, y_bar, indextensor_bar
1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266


@REGISTER_TRANSPOSE('select_p')
def select_transpose(op, check_dot, z_bar):
    cond, x, y = op_position_inputs(op)
    assert check_dot(cond) or check_dot(x) or check_dot(y), (
        f'check_dot(cond) ^ (check_dot(x) ^ check_dot(y)) must be True, '
        f'but check_dot(cond)={check_dot(cond)}, check_dot(x)={check_dot(x)} and check_dot(y)={check_dot(y)}.'
    )

    zeros_x = fill_const(value=0.0, shape=x.shape, dtype=x.dtype)
    zeros_y = fill_const(value=0.0, shape=y.shape, dtype=y.dtype)

    cond_bar = fill_const(value=0.0, shape=y.shape,
                          dtype=cond.dtype) if check_dot(cond) else None
    x_bar = select(cond, z_bar, zeros_x) if check_dot(x) else None
    y_bar = select(cond, zeros_y, z_bar) if check_dot(y) else None

    return cond_bar, x_bar, y_bar
1267 1268 1269 1270 1271 1272


@REGISTER_TRANSPOSE('cast_p')
def cast_transpose(op, check_dot, y_bar):
    x, = op_position_inputs(op)
    return primops.cast(y_bar, x.dtype)