fp16_utils.py 28.1 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
#   Copyright (c) 2019 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.

from __future__ import print_function

from ... import core
18
from ... import framework
19
from ... import layers
20 21
from ... import global_scope
from ...log_helper import get_logger
22 23 24
from ...wrapped_decorator import signature_safe_contextmanager
from .fp16_lists import AutoMixedPrecisionLists
import collections
25 26
import logging
import numpy as np
27

28
__all__ = ["fp16_guard", "cast_model_to_fp16", "cast_parameters_to_fp16"]
29

30 31 32
_logger = get_logger(__name__,
                     logging.INFO,
                     fmt='%(asctime)s-%(levelname)s: %(message)s')
33

34 35 36 37 38 39 40
_valid_types = [
    core.VarDesc.VarType.LOD_TENSOR, core.VarDesc.VarType.SELECTED_ROWS,
    core.VarDesc.VarType.LOD_TENSOR_ARRAY
]

_fp16_guard_pattern = "__use_fp16__"

41

J
Jie Fang 已提交
42 43
def _rename_arg(op, old_name, new_name):
    """
44
    If an op has old_name input and output, rename these input
J
Jie Fang 已提交
45 46 47 48 49 50 51 52 53 54 55 56 57 58
    args new_name.

    Args:
        op (Operator): Current operator.
        old_name (str): The old name of input args.
        new_name (str): The new name of input args.
    """
    op_desc = op.desc
    if isinstance(op_desc, tuple):
        op_desc = op_desc[0]
    op_desc._rename_input(old_name, new_name)
    op_desc._rename_output(old_name, new_name)


59 60 61 62 63 64 65 66 67 68 69 70
def _rename_op_input(program, op_var_rename_map, origin_ops, keep_fp32_ops):
    for block in program.blocks:
        ops = block.ops
        block_id = block.idx
        for op in ops:
            if op not in origin_ops or op in keep_fp32_ops:
                continue
            for name in op.input_arg_names:
                if name in op_var_rename_map[block_id]:
                    op._rename_input(name, op_var_rename_map[block_id][name])


J
Jie Fang 已提交
71 72 73 74 75 76 77 78 79 80 81 82 83
def _dtype_to_str(dtype):
    """
    Convert specific variable type to its corresponding string.

    Args:
        dtype (VarType): Variable type.
    """
    if dtype == core.VarDesc.VarType.FP16:
        return 'fp16'
    else:
        return 'fp32'


84 85 86 87 88 89 90 91 92 93 94 95 96 97
_keep_layer_norm_scale_bias_to_fp32_flag = True


def _keep_layer_norm_scale_bias_to_fp32(*args):
    global _keep_layer_norm_scale_bias_to_fp32_flag
    if len(args) == 0:
        return _keep_layer_norm_scale_bias_to_fp32_flag
    else:
        assert len(args) == 1 and isinstance(args[0], bool)
        old_value = _keep_layer_norm_scale_bias_to_fp32_flag
        _keep_layer_norm_scale_bias_to_fp32_flag = args[0]
        return old_value


98 99
def _keep_fp32_input(op, in_name):
    op_type = op.type
100
    if op_type == 'batch_norm':
101 102
        # Scale, Bias, Mean, Variance should be float32.
        return in_name != 'X'
103 104
    if op_type == 'layer_norm' and _keep_layer_norm_scale_bias_to_fp32():
        return in_name != 'X'
105 106 107 108
    if op_type == 'fused_bn_add_activation':
        return in_name not in {'X', 'Z'}
    if op_type == 'resnet_unit':
        return in_name not in {'X', 'FilterX', 'Z', 'FilterZ'}
109 110 111 112
    if op_type in ['fused_attention', 'fused_feedforward']:
        return in_name in {
            'LnScale', 'LnBias', 'Ln2Scale', 'Ln2Bias', "Ln1Scale", "Ln1Bias"
        }
113 114
    if op_type == 'fused_multi_transformer':
        return in_name in {'LnScale', 'LnBias', 'FFNLnScale', 'FFNLnBias'}
115 116 117 118 119
    return False


def _keep_fp32_output(op, out_name):
    op_type = op.type
120 121 122
    if op_type in ['batch_norm', 'fused_bn_add_activation']:
        return out_name != 'Y'
    if op_type == 'layer_norm' and _keep_layer_norm_scale_bias_to_fp32():
123 124 125
        return out_name != 'Y'
    if op_type == 'resnet_unit':
        return out_name not in {'Y', 'ConvX', 'ConvZ'}
126 127 128 129 130
    if op_type in ['fused_attention', 'fused_feedforward']:
        return out_name in {
            'LnMean', 'LnVariance', 'Ln2Mean', 'Ln2Variance', 'Ln1Mean',
            'Ln1Variance'
        }
131 132 133
    return False


J
Jie Fang 已提交
134 135 136 137 138 139 140 141 142
def _insert_cast_op(block, op, idx, src_dtype, dest_dtype):
    """
    Insert cast op and rename args of input and output.

    Args:
        block (Program): The block in which the operator is.
        op (Operator): The operator to insert cast op.
        idx (int): The index of current operator.
        src_dtype (VarType): The input variable dtype of cast op.
Z
Zhen Wang 已提交
143
        dest_dtype (VarType): The output variable dtype of cast op.
J
Jie Fang 已提交
144 145 146 147 148

    Returns:
        num_cast_op (int): The number of cast ops that have been inserted.
    """
    num_cast_ops = 0
149

J
Jie Fang 已提交
150
    for in_name in op.input_names:
151 152
        if src_dtype == core.VarDesc.VarType.FP32 and _keep_fp32_input(
                op, in_name):
153
            continue
J
Jie Fang 已提交
154
        for in_var_name in op.input(in_name):
H
huangxu96 已提交
155
            in_var = block._find_var_recursive(in_var_name)
156
            if in_var.type not in _valid_types or in_var.dtype == dest_dtype:
J
Jie Fang 已提交
157 158
                continue
            if in_var.dtype == src_dtype:
159 160 161
                cast_name = in_var.name + '.cast_' + _dtype_to_str(dest_dtype)
                out_var = block.vars.get(cast_name)
                if out_var is None or out_var.dtype != dest_dtype:
162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182
                    op_device = op.attr('op_device')
                    # NOTE(wangxi): optimize for pipeline, reduce one send.
                    # if in_var is stop_gradient and prev_op device is `all`,
                    # set cast_op device to `all`, can reduce send cast_var.
                    # TODO: need remove this after we unified the dynamic
                    # and static pipeline interface.
                    if src_dtype == core.VarDesc.VarType.FP32 and in_var.stop_gradient:
                        prev_op = None
                        if in_var.op is op:
                            prev_op = find_true_prev_op(block.ops, op,
                                                        in_var_name)
                        elif in_var.op is not None:
                            prev_op = in_var.op

                        prev_op_device = None
                        if prev_op is not None:
                            prev_op_device = prev_op.attr('op_device')

                        if prev_op_device is not None and 'all' in prev_op_device:
                            op_device = prev_op_device

183 184 185 186
                    out_var = block.create_var(
                        name=cast_name,
                        dtype=dest_dtype,
                        persistable=False,
Z
Zhen Wang 已提交
187
                        stop_gradient=in_var.stop_gradient)
188

189 190 191 192 193 194 195 196 197 198 199 200
                    block._insert_op_without_sync(idx,
                                                  type="cast",
                                                  inputs={"X": in_var},
                                                  outputs={"Out": out_var},
                                                  attrs={
                                                      "in_dtype": in_var.dtype,
                                                      "out_dtype":
                                                      out_var.dtype,
                                                      "op_device": op_device,
                                                      "op_role":
                                                      op.attr("op_role"),
                                                  })
201
                    num_cast_ops += 1
J
Jie Fang 已提交
202 203 204 205
                _rename_arg(op, in_var.name, out_var.name)
            else:
                if op.has_attr('in_dtype'):
                    op._set_attr('in_dtype', dest_dtype)
Z
Zhen Wang 已提交
206
    if src_dtype == core.VarDesc.VarType.FP32 and dest_dtype == core.VarDesc.VarType.FP16:
J
Jie Fang 已提交
207
        for out_name in op.output_names:
208
            if _keep_fp32_output(op, out_name):
209
                continue
J
Jie Fang 已提交
210 211
            for out_var_name in op.output(out_name):
                out_var = block.var(out_var_name)
212
                if out_var.type not in _valid_types:
J
Jie Fang 已提交
213
                    continue
214 215
                if out_var.dtype == core.VarDesc.VarType.FP32:
                    out_var.desc.set_dtype(core.VarDesc.VarType.FP16)
J
Jie Fang 已提交
216
                    if op.has_attr('out_dtype'):
217
                        op._set_attr('out_dtype', core.VarDesc.VarType.FP16)
J
Jie Fang 已提交
218 219 220
    return num_cast_ops


221 222 223 224 225 226 227 228 229
def _insert_cast_post_op(block, op, idx, src_dtype, dest_dtype, target_name,
                         op_var_rename_map):
    num_cast_ops = 0

    target_var = block.var(target_name)
    if target_var.type not in _valid_types or target_var.dtype == dest_dtype:
        return num_cast_ops

    assert target_var.dtype == src_dtype, \
230 231
        "The real dtype({}) is not equal to the src dtype({})".format(
            _dtype_to_str(target_var.dtype), _dtype_to_str(src_dtype))
232 233 234 235

    cast_name = target_var.name + '.cast_' + _dtype_to_str(dest_dtype)
    cast_var = block.vars.get(cast_name)
    if cast_var is None or cast_var.dtype != dest_dtype:
236 237 238 239 240 241 242 243 244 245 246 247 248 249
        cast_var = block.create_var(name=cast_name,
                                    dtype=dest_dtype,
                                    persistable=False,
                                    stop_gradient=target_var.stop_gradient)
        block._insert_op(idx,
                         type="cast",
                         inputs={"X": target_var},
                         outputs={"Out": cast_var},
                         attrs={
                             "in_dtype": target_var.dtype,
                             "out_dtype": cast_var.dtype,
                             "op_device": op.attr("op_device"),
                             "op_role": op.attr("op_role"),
                         })
250 251 252 253 254 255
        num_cast_ops += 1
        op_var_rename_map[block.idx][target_var.name] = cast_var.name

    return num_cast_ops


256 257 258 259 260 261 262 263 264 265
def find_true_prev_op(ops, cur_op, var_name):
    """
    Find the true prev op that outputs var_name variable.

    Args:
        ops (list): A list of ops.
        cur_op (Operator): Current operator which has var_name variable.
        var_name (string): Variable name.
    """
    prev_op = []
J
Jie Fang 已提交
266
    for op in ops:
267 268
        if op == cur_op:
            break
J
Jie Fang 已提交
269 270 271
        for out_name in op.output_names:
            for out_var_name in op.output(out_name):
                if out_var_name == var_name:
272 273 274 275 276 277 278 279
                    prev_op.append(op)
    if prev_op:
        if not len(prev_op) == 1:
            raise ValueError("There must be only one previous op "
                             "that outputs {0} variable".format(var_name))
        else:
            return prev_op[0]
    return None
J
Jie Fang 已提交
280 281


282
def find_true_post_op(ops, cur_op, var_name, search_all=False):
M
mapingshuo 已提交
283 284 285 286 287 288 289
    """
    if there are post ops, return them, if there is no post op,
    return None instead.
    Args:
        ops (list): A list of ops.
        cur_op (Operator): Current operator which has var_name variable.
        var_name (string): Variable name.
290
        search_all (bool): The type of operator search. Use if \"cur_op\" is not in the \"ops\" set.
M
mapingshuo 已提交
291 292
    """
    post_op = []
293 294
    if search_all:
        """
295 296 297 298 299
        \"cur_op\" do not have to be in list of \"ops\". E.g. \"cur_op\" can come
        from startup_prog block and \"ops\" list from main_prog block.
        By setting idx to -1, we'll start looking for post-ops from the top of the list.
        If search_all is False, assume that \"cur_op\" is in \"ops\" list,
        so to reduce the time of search we can start iterating from \"cur_op\" idx.
300 301 302 303 304 305
        """
        idx = -1
    else:
        for idx, op in enumerate(ops):
            if op == cur_op:
                break
M
mapingshuo 已提交
306 307 308 309 310 311 312

    for i in range(idx + 1, len(ops)):
        op = ops[i]
        for in_name in op.input_names:
            for in_var_name in op.input(in_name):
                if in_var_name == var_name:
                    post_op.append(op)
313 314

    return post_op
M
mapingshuo 已提交
315 316 317 318 319 320 321 322 323 324 325


def find_op_index(block_desc, cur_op_desc):
    """
    """
    for idx in range(block_desc.op_size()):
        if cur_op_desc == block_desc.op(idx):
            return idx
    return -1


326 327 328 329 330 331 332 333 334 335 336 337
def _is_in_black_varnames(op, amp_lists):
    for in_name in op.input_arg_names:
        if in_name in amp_lists.black_varnames:
            return True

    for out_name in op.output_arg_names:
        if out_name in amp_lists.black_varnames:
            return True

    return False


338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353
def _need_keep_fp32(op, unsupported_op_list, use_fp16_guard):
    if op.type in unsupported_op_list:
        # the highest priority condition: If ops don't have fp16 computing kernels,
        # they must be executed in fp32 calculation pattern.
        return True

    # process ops about learning rate
    in_out_arg_names = []
    in_out_arg_names.extend(list(op.input_arg_names))
    in_out_arg_names.extend(list(op.output_arg_names))
    for name in in_out_arg_names:
        if "learning_rate" in name:
            return True

    if use_fp16_guard:
        if op.has_attr("op_namescope") and \
354
                (_fp16_guard_pattern in op.attr("op_namescope")):
355 356 357 358 359 360 361 362 363 364 365 366 367 368 369
            # op in fp16 guard
            return False
        else:
            # op not in fp16 guard
            return True
    else:
        return False


@signature_safe_contextmanager
def fp16_guard():
    """
    As for the pure fp16 training, if users set `use_fp16_guard` to True,
    only those ops created in the context manager `fp16_guard` will be
    transformed as float16 type.
H
huangxu96 已提交
370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385

    Examples:
        .. code-block:: python

            import numpy as np
            import paddle
            import paddle.nn.functional as F
            paddle.enable_static()
            data = paddle.static.data(name='X', shape=[None, 1, 28, 28], dtype='float32')
            conv2d = paddle.static.nn.conv2d(input=data, num_filters=6, filter_size=3)

            with paddle.static.amp.fp16_guard():
                bn = paddle.static.nn.batch_norm(input=conv2d, act="relu")
                pool = F.max_pool2d(bn, kernel_size=2, stride=2)
                hidden = paddle.static.nn.fc(pool, size=10)
                loss = paddle.mean(hidden)
386 387 388 389 390 391
    """
    with framework.name_scope(prefix=_fp16_guard_pattern):
        yield


def cast_model_to_fp16(program, amp_lists=None, use_fp16_guard=True):
392 393 394 395 396 397
    """
    Traverse all ops in the whole model and set their inputs and outputs
    to the fp16 data type. This function will do some special process for
    the batch normalization, which keeps the computational process of
    batchnorms in FP32.
    Args:
398 399 400 401
        program (Program): The used program.
        amp_lists (AutoMixedPrecisionLists): An AutoMixedPrecisionLists object.
        use_fp16_guard(bool): Determine whether to use `fp16_guard` when
                              constructing the program. Default True.
402 403
    """

404 405 406 407 408 409 410 411 412 413
    if amp_lists is None:
        amp_lists = AutoMixedPrecisionLists()
    global_block = program.global_block()
    keep_fp32_ops = set()
    to_fp16_var_names = set()
    origin_ops = []
    for block in program.blocks:
        origin_ops.extend(block.ops)

    for block in program.blocks:
414 415 416 417
        ops = block.ops
        for op in ops:
            if op.type == 'create_py_reader' or op.type == 'read':
                continue
418 419 420
            if _need_keep_fp32(op, amp_lists.unsupported_list, use_fp16_guard):
                keep_fp32_ops.add(op)
                continue  # processed below
421
            for in_name in op.input_names:
422 423 424
                # for ipu, all inputs must be converted to fp16
                if not core.is_compiled_with_ipu() and _keep_fp32_input(
                        op, in_name):
425 426 427 428 429 430 431
                    continue
                for in_var_name in op.input(in_name):
                    in_var = None
                    try:
                        in_var = block.var(in_var_name)
                    except ValueError as e:
                        _logger.debug(
432
                            "-- {}, try to get it in the global block --".
433 434 435 436
                            format(e))
                        in_var = global_block.var(in_var_name)
                        if in_var is not None:
                            _logger.debug(
437
                                "-- var {} is got in the global block --".
438 439
                                format(in_var_name))

440
                    if in_var is None or in_var.type not in _valid_types:
441 442 443 444
                        continue

                    if in_var.dtype == core.VarDesc.VarType.FP32:
                        in_var.desc.set_dtype(core.VarDesc.VarType.FP16)
445
                        to_fp16_var_names.add(in_var_name)
446 447 448 449 450 451

                    _logger.debug(
                        "-- op type: {}, in var name: {}, in var dtype: {} --".
                        format(op.type, in_var_name, in_var.dtype))

            for out_name in op.output_names:
452 453 454
                # for ipu, all outputs must be converted to fp16
                if not core.is_compiled_with_ipu() and _keep_fp32_output(
                        op, out_name):
455 456 457 458 459 460 461
                    continue
                for out_var_name in op.output(out_name):
                    out_var = None
                    try:
                        out_var = block.var(out_var_name)
                    except ValueError as e:
                        _logger.debug(
462
                            "-- {}, try to get it in the global block --".
463 464 465 466
                            format(e))
                        out_var = global_block.var(out_var_name)
                        if out_var is not None:
                            _logger.debug(
467
                                "-- var {} is got in the global block --".
468 469
                                format(out_var_name))

470
                    if out_var is None or out_var.type not in _valid_types:
471 472 473 474 475 476
                        continue

                    if out_var.dtype == core.VarDesc.VarType.FP32:
                        out_var.desc.set_dtype(core.VarDesc.VarType.FP16)

                    _logger.debug(
477 478
                        "-- op type: {}, out var name: {}, out var dtype: {} --"
                        .format(op.type, out_var_name, out_var.dtype))
479 480 481 482 483 484 485 486 487 488
            if op.has_attr('in_dtype') and op.attr(
                    'in_dtype') == core.VarDesc.VarType.FP32:
                op._set_attr('in_dtype', core.VarDesc.VarType.FP16)
            if op.has_attr('out_dtype') and op.attr(
                    'out_dtype') == core.VarDesc.VarType.FP32:
                op._set_attr('out_dtype', core.VarDesc.VarType.FP16)
            if op.has_attr('dtype') and op.attr(
                    'dtype') == core.VarDesc.VarType.FP32:
                op._set_attr('dtype', core.VarDesc.VarType.FP16)

489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523
    # process ops in keep_fp32_ops
    op_var_rename_map = [
        collections.OrderedDict() for _ in range(len(program.blocks))
    ]
    for block in program.blocks:
        ops = block.ops
        idx = 0
        while idx < len(ops):
            op = ops[idx]
            num_cast_ops = 0
            if op in keep_fp32_ops:
                pre_cast_num = _insert_cast_op(block, op, idx,
                                               core.VarDesc.VarType.FP16,
                                               core.VarDesc.VarType.FP32)
                num_cast_ops += pre_cast_num
                for out_var_name in op.output_arg_names:
                    out_var = block.vars.get(out_var_name)
                    if out_var is None or out_var.type not in _valid_types:
                        continue
                    if out_var.dtype == core.VarDesc.VarType.FP16:
                        out_var.desc.set_dtype(core.VarDesc.VarType.FP32)
                        post_ops = find_true_post_op(ops, op, out_var_name)
                        for post_op in post_ops:
                            if post_op in keep_fp32_ops:
                                continue
                            post_cast_num = _insert_cast_post_op(
                                block, op, idx + pre_cast_num + 1,
                                core.VarDesc.VarType.FP32,
                                core.VarDesc.VarType.FP16, out_var_name,
                                op_var_rename_map)
                            num_cast_ops += post_cast_num
            idx += num_cast_ops + 1

    _rename_op_input(program, op_var_rename_map, origin_ops, keep_fp32_ops)
    return to_fp16_var_names
524

525 526

def cast_parameters_to_fp16(place, program, scope=None, to_fp16_var_names=None):
527
    """
528
    Traverse all parameters in the whole model and set them to the FP16 data type.
529 530
    Whereas, this function will keep parameters of batchnorms in FP32.
    Args:
531 532 533 534 535 536 537
        place(fluid.CPUPlace|fluid.CUDAPlace): `place` is used to restore the FP16 weight tensors.
        program (Program): The used program.
        scope(fluid.Scope, optional): `scope` is used to get the FP32 weight tensor values.
                                      Default is None.
        to_fp16_var_names(set|list, optional): The data types of vars in `to_fp16_var_names`
                                               will be set to FP16. Usually, it is the returned
                                               value of `cast_model_to_fp16` API.
538
    """
539 540 541 542 543 544
    all_parameters = []
    for block in program.blocks:
        all_parameters.extend(block.all_parameters())

    fp16_var_names = to_fp16_var_names if to_fp16_var_names else set()
    var_scope = scope if scope else global_scope()
545
    for param in all_parameters:
546 547
        if param.name in fp16_var_names:
            _logger.debug("---- cast {} to fp16 dtype ----".format(param.name))
548 549 550 551 552
            param_t = var_scope.find_var(param.name).get_tensor()
            data = np.array(param_t)
            param_t.set(np.float16(data), place)


J
Jie Fang 已提交
553
def rewrite_program(main_prog, amp_lists):
J
Jie Fang 已提交
554
    """
555
    Traverse all ops in current block and insert cast op according to
J
Jie Fang 已提交
556 557 558 559
    which set current op belongs to.

    1. When an op belongs to the black list, add it to black set
    2. When an op belongs to the white list, add it to white set
560 561 562 563
    3. When an op belongs to the gray list. If one
       of its inputs is the output of black set op or black list op,
       add it to black set. If all of its previous ops are not black
       op and one of its inputs is the output of white set op or
J
Jie Fang 已提交
564 565
       white list op, add it to white set.
    4. When an op isn't in the lists, add it to black op set.
566 567
    5. Add necessary cast ops to make sure that black set op will be
       computed in fp32 mode, while white set op will be computed in
J
Jie Fang 已提交
568 569 570 571 572 573
       fp16 mode.

    Args:
        main_prog (Program): The main program for training.
    """
    block = main_prog.global_block()
F
fangshuixun007 已提交
574
    block._sync_with_cpp()
J
Jie Fang 已提交
575 576 577
    ops = block.ops
    white_op_set = set()
    black_op_set = set()
578
    for op in ops:
579

580 581
        # NOTE(zhiqiu): 'create_py_reader' and 'read' is used in non-iterable DataLoder,
        # we don't need to handle reader op and the input of 'create_py_reader' is not
582 583 584 585 586
        # in block, which may result in errors.
        # See GeneratorLoader._init_non_iterable() for details.
        if op.type == 'create_py_reader' or op.type == 'read':
            continue

587 588 589 590 591
        if amp_lists.black_varnames is not None and _is_in_black_varnames(
                op, amp_lists):
            black_op_set.add(op)
            continue

J
Jie Fang 已提交
592
        if op.type in amp_lists.black_list:
J
Jie Fang 已提交
593
            black_op_set.add(op)
J
Jie Fang 已提交
594
        elif op.type in amp_lists.white_list:
J
Jie Fang 已提交
595
            white_op_set.add(op)
J
Jie Fang 已提交
596
        elif op.type in amp_lists.gray_list:
J
Jie Fang 已提交
597 598 599 600 601 602 603 604 605 606
            is_black_op = False
            is_white_op = False
            for in_name in op.input_names:
                # if this op has inputs
                if in_name:
                    for in_var_name in op.input(in_name):
                        in_var = block.var(in_var_name)
                        # this in_var isn't the output of other op
                        if in_var.op is None:
                            continue
607 608 609 610
                        elif in_var.op is op:
                            prev_op = find_true_prev_op(ops, op, in_var_name)
                            if prev_op is None:
                                continue
J
Jie Fang 已提交
611 612 613 614
                        else:
                            prev_op = in_var.op
                        # if it's one of inputs
                        if prev_op in black_op_set or \
J
Jie Fang 已提交
615
                                prev_op.type in amp_lists.black_list:
J
Jie Fang 已提交
616
                            is_black_op = True
617
                        elif prev_op in white_op_set or \
J
Jie Fang 已提交
618
                                prev_op.type in amp_lists.white_list:
J
Jie Fang 已提交
619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648
                            is_white_op = True
            if is_black_op:
                black_op_set.add(op)
            elif is_white_op:
                white_op_set.add(op)
            else:
                pass
        else:
            # For numerical safe, we apply fp32 computation on ops that
            # are not determined which list they should stay.
            black_op_set.add(op)

    idx = 0
    while idx < len(ops):
        op = ops[idx]
        num_cast_ops = 0
        if op in black_op_set:
            num_cast_ops = _insert_cast_op(block, op, idx,
                                           core.VarDesc.VarType.FP16,
                                           core.VarDesc.VarType.FP32)
        elif op in white_op_set:
            num_cast_ops = _insert_cast_op(block, op, idx,
                                           core.VarDesc.VarType.FP32,
                                           core.VarDesc.VarType.FP16)
        else:
            pass

        idx += num_cast_ops + 1


649 650 651
def update_role_var_grad(main_prog, params_grads):
    """
    Update op_role_var attr for some ops to make sure the gradients
Z
Zhen Wang 已提交
652
    transferred across GPUs is FP16.
653 654 655 656 657 658 659 660 661 662
    1. Check whether the op that outputs gradient is cast or not.
    2. If op is cast and gradient is FP32, remove the op_role_var
       and find the prev op which outputs FP16 gradient
    3. Update the op_role_var of the prev op.

    Args:
        main_prog (Program): The main program for training.
        params_grads (list): A list of params and grads.
    """
    block = main_prog.global_block()
F
fangshuixun007 已提交
663
    block._sync_with_cpp()
664 665 666 667 668 669 670
    BACKWARD = core.op_proto_and_checker_maker.OpRole.Backward
    OPTIMIZE = core.op_proto_and_checker_maker.OpRole.Optimize
    for p, g in params_grads:
        op = g.op
        if g.dtype == core.VarDesc.VarType.FP32 and op.type == 'cast':
            role = op.attr('op_role')
            if role & int(BACKWARD) and op.has_attr('op_role_var'):
F
fangshuixun007 已提交
671
                op._remove_attr("op_role_var")
672 673 674 675 676 677 678 679 680 681 682 683 684
            else:
                raise ValueError("The cast op {0} must be in BACKWARD role "
                                 "and have op_role_var attr.".format(op))

            fp16_grad_name = op.input(op.input_names[0])[0]
            op_for_fp16_grad = find_true_prev_op(block.ops, op, fp16_grad_name)
            op_role_var_attr_name = \
                core.op_proto_and_checker_maker.kOpRoleVarAttrName()
            attr_val = [p.name, fp16_grad_name]
            if op_for_fp16_grad.has_attr(op_role_var_attr_name):
                attr_val.extend(op_for_fp16_grad.attr(op_role_var_attr_name))
            op_for_fp16_grad._set_attr(op_role_var_attr_name, attr_val)

Z
Zhen Wang 已提交
685 686
            # Maximize the all_reduce overlap, and perform the cast
            # operation after gradients transfer.
687
            op._set_attr('op_role', OPTIMIZE)
M
mapingshuo 已提交
688 689 690 691
            # optimize op should stay behind forward and backward ops
            if op == block.ops[-1]:
                continue
            post_ops = find_true_post_op(block.ops, op, g.name)
692
            if post_ops:
M
mapingshuo 已提交
693 694 695
                raise ValueError("The cast op {0}'s output should not be"
                                 "used by a non-optimize op, however, it"
                                 "is used by {1}".format(op, post_ops[0]))
696
            # add new op in the python and cpp at the same time
M
mapingshuo 已提交
697 698
            new_op_desc = block.desc.append_op()
            new_op_desc.copy_from(op.desc)
699 700 701 702 703 704
            new_op = framework.Operator(block=block,
                                        desc=new_op_desc,
                                        type=None,
                                        inputs=None,
                                        outputs=None,
                                        attrs=None)
F
fangshuixun007 已提交
705
            block.ops.append(new_op)
M
mapingshuo 已提交
706 707 708
            op_idx = find_op_index(block.desc, op.desc)
            if op_idx == -1:
                raise ValueError("The op {0} is not in program".format(op))
F
fangshuixun007 已提交
709 710
            block._remove_op(op_idx, sync=False)
    block._sync_with_cpp()