fp16_utils.py 16.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
#   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
from ... import layers


J
Jie Fang 已提交
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
def _rename_arg(op, old_name, new_name):
    """
    If an op has old_name input and output, rename these input 
    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)


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'


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
60
        dest_dtype (VarType): The output variable dtype of cast op.
J
Jie Fang 已提交
61 62 63 64 65 66 67 68 69

    Returns:
        num_cast_op (int): The number of cast ops that have been inserted.
    """
    num_cast_ops = 0
    valid_types = [
        core.VarDesc.VarType.LOD_TENSOR, core.VarDesc.VarType.SELECTED_ROWS,
        core.VarDesc.VarType.LOD_TENSOR_ARRAY
    ]
70

J
Jie Fang 已提交
71
    for in_name in op.input_names:
72 73 74
        if src_dtype == core.VarDesc.VarType.FP32 and op.type == 'batch_norm':
            if in_name != 'X':
                continue
J
Jie Fang 已提交
75 76 77 78 79
        for in_var_name in op.input(in_name):
            in_var = block.var(in_var_name)
            if in_var.type not in valid_types:
                continue
            if in_var.dtype == src_dtype:
80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98
                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:
                    out_var = block.create_var(
                        name=cast_name,
                        dtype=dest_dtype,
                        persistable=False,
                        stop_gradient=False)

                    block._insert_op(
                        idx,
                        type="cast",
                        inputs={"X": in_var},
                        outputs={"Out": out_var},
                        attrs={
                            "in_dtype": in_var.dtype,
                            "out_dtype": out_var.dtype
                        })
                    num_cast_ops += 1
J
Jie Fang 已提交
99 100 101 102
                _rename_arg(op, in_var.name, out_var.name)
            else:
                if op.has_attr('in_dtype'):
                    op._set_attr('in_dtype', dest_dtype)
103
    if src_dtype == core.VarDesc.VarType.FP32:
J
Jie Fang 已提交
104
        for out_name in op.output_names:
105 106
            if op.type == 'batch_norm' and out_name != 'Y':
                continue
J
Jie Fang 已提交
107 108 109 110
            for out_var_name in op.output(out_name):
                out_var = block.var(out_var_name)
                if out_var.type not in valid_types:
                    continue
111 112
                if out_var.dtype == core.VarDesc.VarType.FP32:
                    out_var.desc.set_dtype(core.VarDesc.VarType.FP16)
J
Jie Fang 已提交
113
                    if op.has_attr('out_dtype'):
114
                        op._set_attr('out_dtype', core.VarDesc.VarType.FP16)
J
Jie Fang 已提交
115 116 117
    return num_cast_ops


118 119 120 121 122 123 124 125 126 127
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
128
    for op in ops:
129 130
        if op == cur_op:
            break
J
Jie Fang 已提交
131 132 133
        for out_name in op.output_names:
            for out_var_name in op.output(out_name):
                if out_var_name == var_name:
134 135 136 137 138 139 140 141
                    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
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
def find_true_post_op(ops, cur_op, var_name):
    """
    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.
    """
    post_op = []
    for idx, op in enumerate(ops):
        if op == cur_op:
            break

    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)
    if post_op != []:
        return post_op
    return None


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


178 179 180 181 182 183 184 185 186 187 188 189
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


J
Jie Fang 已提交
190
def rewrite_program(main_prog, amp_lists):
J
Jie Fang 已提交
191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213
    """
    Traverse all ops in current block and insert cast op according to 
    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
    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 
       white list op, add it to white set.
    4. When an op isn't in the lists, add it to black op set.
    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 
       fp16 mode.

    Args:
        main_prog (Program): The main program for training.
    """
    block = main_prog.global_block()
    ops = block.ops
    white_op_set = set()
    black_op_set = set()
214
    for op in ops:
215 216 217 218 219
        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
220
        if op.type in amp_lists.black_list:
J
Jie Fang 已提交
221
            black_op_set.add(op)
J
Jie Fang 已提交
222
        elif op.type in amp_lists.white_list:
J
Jie Fang 已提交
223
            white_op_set.add(op)
J
Jie Fang 已提交
224
        elif op.type in amp_lists.gray_list:
J
Jie Fang 已提交
225 226 227 228 229 230 231 232 233 234
            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
235 236 237 238
                        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
239 240 241 242
                        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
243
                                prev_op.type in amp_lists.black_list:
J
Jie Fang 已提交
244
                            is_black_op = True
245
                        elif prev_op in white_op_set or \
J
Jie Fang 已提交
246
                                prev_op.type in amp_lists.white_list:
J
Jie Fang 已提交
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
                            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


277 278 279
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
280
    transferred across GPUs is FP16.
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
    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()
    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'):
                op.desc.remove_attr("op_role_var")
            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
312 313
            # Maximize the all_reduce overlap, and perform the cast
            # operation after gradients transfer.
314
            op._set_attr('op_role', OPTIMIZE)
315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330
            # 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)
            if post_ops is not None:
                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]))
            new_op_desc = block.desc.append_op()
            new_op_desc.copy_from(op.desc)

            op_idx = find_op_index(block.desc, op.desc)
            if op_idx == -1:
                raise ValueError("The op {0} is not in program".format(op))
            block.desc._remove_op(op_idx, op_idx + 1)
        block._sync_with_cpp()
331 332


J
Jie Fang 已提交
333 334 335 336 337 338
def update_loss_scaling(is_overall_finite, prev_loss_scaling, num_good_steps,
                        num_bad_steps, incr_every_n_steps,
                        decr_every_n_nan_or_inf, incr_ratio, decr_ratio):
    """
    Update loss scaling according to overall gradients. If all gradients is 
    finite after incr_every_n_steps, loss scaling will increase by incr_ratio. 
Z
Zhen Wang 已提交
339
    Otherwise, loss scaling will decrease by decr_ratio after
J
Jie Fang 已提交
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
    decr_every_n_nan_or_inf steps and each step some gradients are infinite.

    Args:
        is_overall_finite (Variable): A boolean variable indicates whether 
                                     all gradients are finite.
        prev_loss_scaling (Variable): Previous loss scaling.
        num_good_steps (Variable): A variable accumulates good steps in which 
                                   all gradients are finite.
        num_bad_steps (Variable): A variable accumulates bad steps in which 
                                  some gradients are infinite.
        incr_every_n_steps (Variable): A variable represents increasing loss 
                                       scaling every n consecutive steps with 
                                       finite gradients.
        decr_every_n_nan_or_inf (Variable): A variable represents decreasing 
                                            loss scaling every n accumulated 
                                            steps with nan or inf gradients.
        incr_ratio(float): The multiplier to use when increasing the loss 
                           scaling.
        decr_ratio(float): The less-than-one-multiplier to use when decreasing 
                           loss scaling.
    """
    zero_steps = layers.fill_constant(shape=[1], dtype='int32', value=0)
    with layers.Switch() as switch:
        with switch.case(is_overall_finite):
            should_incr_loss_scaling = layers.less_than(incr_every_n_steps,
                                                        num_good_steps + 1)
            with layers.Switch() as switch1:
                with switch1.case(should_incr_loss_scaling):
                    new_loss_scaling = prev_loss_scaling * incr_ratio
                    loss_scaling_is_finite = layers.isfinite(new_loss_scaling)
                    with layers.Switch() as switch2:
                        with switch2.case(loss_scaling_is_finite):
                            layers.assign(new_loss_scaling, prev_loss_scaling)
                        with switch2.default():
                            pass
                    layers.assign(zero_steps, num_good_steps)
                    layers.assign(zero_steps, num_bad_steps)

                with switch1.default():
                    layers.increment(num_good_steps)
                    layers.assign(zero_steps, num_bad_steps)

        with switch.default():
            should_decr_loss_scaling = layers.less_than(decr_every_n_nan_or_inf,
                                                        num_bad_steps + 1)
            with layers.Switch() as switch3:
                with switch3.case(should_decr_loss_scaling):
                    new_loss_scaling = prev_loss_scaling * decr_ratio
                    static_loss_scaling = \
                        layers.fill_constant(shape=[1],
                                             dtype='float32',
                                             value=1.0)
                    less_than_one = layers.less_than(new_loss_scaling,
                                                     static_loss_scaling)
                    with layers.Switch() as switch4:
                        with switch4.case(less_than_one):
                            layers.assign(static_loss_scaling,
                                          prev_loss_scaling)
                        with switch4.default():
                            layers.assign(new_loss_scaling, prev_loss_scaling)
                    layers.assign(zero_steps, num_good_steps)
                    layers.assign(zero_steps, num_bad_steps)
                with switch3.default():
                    layers.assign(zero_steps, num_good_steps)
                    layers.increment(num_bad_steps)