batch_norm.cpp 13.7 KB
Newer Older
1 2 3 4
/**
 * \file src/opr/impl/dnn/batch_norm.cpp
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
5
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
6 7 8 9 10 11 12 13 14
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 */

#include "megbrain/opr/dnn/batch_norm.h"
#include "megbrain/opr/io.h"
#include "megbrain/graph/grad_impl.h"
15 16
#include "megbrain/opr/basic_arith.h"
#include "megbrain/opr/tensor_manip.h"
17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44

#include "../internal/megdnn_opr_wrapper.inl"

using namespace mgb;
using namespace opr;

namespace mgb { namespace opr { namespace intl {
template<>
struct AutoAddWorkspaceNeedLimitGetter<megdnn::BNForward> {
    static constexpr bool val = true;
};

template<>
struct AutoAddWorkspaceNeedLimitGetter<megdnn::BNBackward> {
    static constexpr bool val = true;
};
} } } // mgb::opr::intl

MGB_DYN_TYPE_OBJ_FINAL_IMPL(BatchNormForward);

BatchNormForward::BatchNormForward(VarNode *x,
        VarNode *scale, VarNode *bias,
        VarNode *mean, VarNode *variance,
        const Param &param,
        const OperatorNodeConfig &config):
    Super{x->owner_graph(), config, "batch_norm",
          {x, scale, bias, mean, variance}}
{
45
    if(owner_graph()->options().no_force_inplace) {
46 47 48
        m_force_inplace = false;
    }

M
Megvii Engine Team 已提交
49
    if (m_force_inplace && param.fwd_mode == Param::FwdMode::TRAINING) {
50 51 52 53 54
        auto check_dest = [&](VarNode* dest) {
            auto dest_opr = dest->owner_opr();
            mgb_throw_if(!(dest_opr->same_type<SharedDeviceTensor>() ||
                    dest_opr->same_type<VolatileSharedDeviceTensor>()),
                    GraphError,
55 56
                    "mean and variance in training mode BatchNorm must be"
                    "SharedDeviceTensor or VolatileSharedDeviceTensor;"
57
                    "got %s{%s} actually",
58 59 60 61 62
                    dest_opr->cname(), dest_opr->dyn_typeinfo()->name);
        };
        check_dest(mean);
        check_dest(variance);
    }
63 64 65 66 67

    init_megdnn_opr(*this, param);

    add_input({x, scale, bias, mean, variance});

68 69 70
    output(4)->add_flag(VarNode::Flag::ALLOW_EMPTY_SHAPE); // reserve
    output(5)->add_flag(VarNode::Flag::ALLOW_EMPTY_SHAPE);
    // running mean/var
M
Megvii Engine Team 已提交
71 72 73 74 75 76 77 78
    if (param.fwd_mode == Param::FwdMode::INFERENCE) {
        auto mark_empty_var = [&](VarNode *var) {
            var->add_flag(VarNode::Flag::ALLOW_EMPTY_SHAPE)
                .add_flag(VarNode::Flag::VOLATILE_CONTENT);
        };
        mark_empty_var(output(0));
        mark_empty_var(output(1));
    } else if (m_force_inplace) {
79 80 81
        output(0)->
            set_fwd_in2out_writable_force(input(3)).
            add_flag(VarNode::Flag::NO_MEM_RECLAIM);
82

83 84 85 86
        output(1)->
            set_fwd_in2out_writable_force(input(4)).
            add_flag(VarNode::Flag::NO_MEM_RECLAIM);
    }
87 88 89 90 91 92 93 94 95 96 97 98
}

BatchNormForward::BatchNormForward(VarNode *x,
        VarNode *scale, VarNode *bias,
        const Param &param,
        const OperatorNodeConfig &config):
    Super{x->owner_graph(), config, "batch_norm",
          {x, scale, bias}}
{
    init_megdnn_opr(*this, param);

    add_input({x, scale, bias});
99 100
    output(4)->add_flag(VarNode::Flag::ALLOW_EMPTY_SHAPE); // reserve
    output(5)->add_flag(VarNode::Flag::ALLOW_EMPTY_SHAPE);
101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146
    auto mark_empty_var = [&](VarNode *var) {
        var->add_flag(VarNode::Flag::ALLOW_EMPTY_SHAPE)
            .add_flag(VarNode::Flag::VOLATILE_CONTENT);
    };
    mark_empty_var(output(0));
    mark_empty_var(output(1));
}

SymbolVarArray BatchNormForward::make(SymbolVar x,
        SymbolVar scale, SymbolVar bias,
        SymbolVar mean, SymbolVar variance,
        const Param &param,
        const OperatorNodeConfig &config) {
    auto&& out = x.node()
                    ->owner_graph()
                    ->insert_opr(std::make_unique<BatchNormForward>(
                        x.node(), scale.node(), bias.node(),
                        mean.node(), variance.node(), param, config))
                    ->output();
    SymbolVarArray ret(out.size());
    for (size_t i = 0; i < ret.size(); i++) {
        ret[i] = out[i];
    }
    return ret;
}

SymbolVarArray BatchNormForward::make(SymbolVar x,
        SymbolVar scale, SymbolVar bias,
        const Param &param,
        const OperatorNodeConfig &config) {
    auto&& out = x.node()
                    ->owner_graph()
                    ->insert_opr(std::make_unique<BatchNormForward>(
                        x.node(), scale.node(), bias.node(),
                        param, config))
                    ->output();
    SymbolVarArray ret(out.size());
    for (size_t i = 0; i < ret.size(); i++) {
        ret[i] = out[i];
    }
    return ret;
}

cg::OperatorNodeBase::NodeProp*
BatchNormForward::do_make_node_prop() const {
    auto ret = Super::do_make_node_prop();
147 148
    ret->add_dep_type_existing_var(input(0),
                                   NodeProp::DepType::VALUE_ALLOW_EMPTY);
149
    if (need_stats() && m_force_inplace) {
150 151 152 153 154 155 156
        ret->add_flag(NodeProp::Flag::FORCE_UPDATE_INPUT_VAR);
    }
    return ret;
}

void BatchNormForward::scn_do_execute() {
    auto &&x = input(0)->dev_tensor();
157
    auto &&y = output(5)->dev_tensor();
M
Megvii Engine Team 已提交
158
    if (need_stats()) {
159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179
        auto &&o0 = output(0)->dev_tensor(),
             &&o1 = output(1)->dev_tensor(),
             &&i0 = input(3)->dev_tensor(),
             &&i1 = input(4)->dev_tensor();
        mgb_assert(o0.raw_ptr() && o1.raw_ptr()); // non-empty tensor
        mgb_assert(o0.comp_node() == i0.comp_node() &&
                   o1.comp_node() == i1.comp_node() &&
                   o0.layout().eq_layout(i0.layout()) &&
                   o1.layout().eq_layout(i1.layout()));
        if (!m_force_inplace) {
            if (o0.raw_ptr() != i0.raw_ptr()) {
                o0.copy_from_fixlayout(i0);
            }
            if (o1.raw_ptr() != i1.raw_ptr()) {
                o1.copy_from_fixlayout(i1);
            }
        } else {
            mgb_assert(o0.raw_ptr() == i0.raw_ptr()
                    && o1.raw_ptr() == i1.raw_ptr());
        }
    }
180 181 182 183 184 185
    mgb_assert(x.layout().eq_layout(y.layout()));
    if (x.layout().is_empty()) {
        return;
    }
    mgb_assert(x.layout().is_contiguous() &&
               y.layout().is_contiguous());
186 187
    auto scale = input(1)->dev_tensor().as_megdnn();
    auto bias = input(2)->dev_tensor().as_megdnn();
M
Megvii Engine Team 已提交
188 189 190 191 192 193 194 195
    megdnn::TensorND mean, variance;
    if (param().fwd_mode == Param::FwdMode::INFERENCE) {
        mean = input(3)->dev_tensor().as_megdnn();
        variance = input(4)->dev_tensor().as_megdnn();
    } else {
        mean = output(0)->dev_tensor().as_megdnn();
        variance = output(1)->dev_tensor().as_megdnn();
    }
196 197
    auto save_mean = output(2)->dev_tensor().as_megdnn();
    auto save_variance = output(3)->dev_tensor().as_megdnn();
198
    auto reserve = output(4)->dev_tensor().as_megdnn();
199
    auto workspace = intl::get_megdnn_workspace_from_var(output().back());
200
    megdnn_opr()->exec(x.as_megdnn(), scale, bias, mean, variance,
201
        save_mean, save_variance, reserve, y.as_megdnn(), workspace);
202 203 204 205 206 207 208 209 210
}

void BatchNormForward::add_input_layout_constraint() {
    mixin::megdnn_utils::add_input_layout_constraint_contig(*this);
}

void BatchNormForward::get_output_var_shape(
        const TensorShapeArray &inp_shape,
        TensorShapeArray &out_shape) const {
211 212 213 214
    mgb_assert(inp_shape[0].ndim == 4 && inp_shape[0].ndim == 4 && inp_shape[1].ndim == 4,
        "expect input, scale and bias to be 4 dim tensor, but "
        "got input dim: %zu, scale dim: %zu, bias dim: %zu",
        inp_shape[0].ndim, inp_shape[1].ndim, inp_shape[2].ndim);
215 216 217 218 219 220 221 222 223 224
    
    size_t channel_idx;
    if (param().param_dim == Param::ParamDim::DIM_111C) {
        channel_idx = 3;
    } else {
        channel_idx = 1;
    }
    size_t inp_c = inp_shape[0][channel_idx],
           scale_c = inp_shape[1][channel_idx],
           bias_c = inp_shape[2][channel_idx];
225 226 227 228
    mgb_assert(inp_c == scale_c && inp_c == bias_c,
        "inconsistent channel size, input chennel: %zu, scale channel: %zu, bias channel: %zu",
        inp_c, scale_c, bias_c);

229
    out_shape[5] = inp_shape[0];
230 231 232
    for (size_t i = 0; i < 4; ++ i) {
        out_shape[i] = inp_shape[1];
    }
M
Megvii Engine Team 已提交
233
    if (!need_stats()) {
234 235
        out_shape[0] = out_shape[1] = {0};
    }
236 237 238 239 240
    if (inp_shape[0].is_empty()) {
        out_shape[4] = {0};
    } else {
        out_shape[4] = {megdnn_opr()->get_reserve_in_bytes({inp_shape[0], input(0)->dtype()})};
    }
241 242 243 244 245
}

size_t BatchNormForward::get_workspace_size_bytes(
        const TensorShapeArray &input_shapes,
        const TensorShapeArray &output_shapes) const {
246 247
    if (input_shapes[0].is_empty())
        return 0;
248 249 250
#define in(x) {input_shapes[x], input(x)->dtype()}
#define out(x) {output_shapes[x], output(x)->dtype()}
    return megdnn_opr()->get_workspace_in_bytes(
251
            in(0), in(1), in(2), out(0), out(1), out(2), out(3), out(4), out(5));
252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268
#undef in
#undef out
}

void BatchNormForward::init_output_static_infer_desc() {
    Super::set_nr_managed_outputs(this->output().size() - 1);
    Super::init_output_static_infer_desc();
    this->init_output_static_infer_desc_workspace(
            intl::AutoAddWorkspaceNeedLimitGetter<megdnn::BNForward>::val);
}

void BatchNormForward::init_output_dtype() {
    size_t nr_inp = input().size();
    mgb_assert(input(0)->dtype().category() == input(1)->dtype().category());
    for (size_t i = 2; i < nr_inp; ++ i) {
        mgb_assert(input(1)->dtype() == input(i)->dtype());
    }
269 270
    output(4)->dtype(dtype::Byte()); // reserve
    output(5)->dtype(input(0)->dtype()); // output
271 272 273 274 275
    for (size_t i = 0; i < 4; ++ i) {
        output(i)->dtype(input(1)->dtype());
    }
}

276
void BatchNormForward::mem_plan_fwd_in2out_writable() {
M
Megvii Engine Team 已提交
277
    if (need_stats() && !m_force_inplace) {
278 279 280 281 282 283
        // TODO: testing
        output(0)->set_fwd_in2out_writable(input(3));
        output(1)->set_fwd_in2out_writable(input(4));
    }
}

284
#if MGB_ENABLE_GRAD
285
MGB_IMPL_OPR_GRAD(BatchNormForward) {
286 287
    mgb_assert(wrt_idx < 5, "wrt_idx %zu is out of range", wrt_idx);
    VarNodeArray ret(opr.input().size(), nullptr);
288 289 290 291
    SymbolVarArray grad;
    switch (opr.param().fwd_mode) {
    case BatchNorm::Param::FwdMode::TRAINING:
        grad = BatchNormBackward::make(
292
                opr.input(0), out_grad[5],
293
                opr.output(2), opr.output(3),
294 295
                opr.input(1), opr.output(4), // reserve
                opr.param());
296 297 298 299 300 301 302
        for (size_t i = 0; i < 3; ++ i) {
            ret[i] = grad[(i + 2) % 3].node();
        }
        return ret;
    case BatchNorm::Param::FwdMode::INFERENCE:
        auto sqrt_var = PowC::make((SymbolVar{opr.input(4)}
                        + static_cast<dt_float32>(opr.param().epsilon)), 0.5, opr.config());
303
        auto d_bn_scale_unreduced = SymbolVar{out_grad[5]} *
304 305 306
                            (SymbolVar{opr.input(0)} - SymbolVar{opr.input(3)}) / sqrt_var;
        auto d_bn_scale = Reduce::make(d_bn_scale_unreduced,
                            Reduce::Param::Mode::SUM, GetVarShape::make(opr.input(1)));
307
        auto d_bn_bias = Reduce::make(out_grad[5],
308
                            Reduce::Param::Mode::SUM, GetVarShape::make(opr.input(2)));
309
        auto dx = SymbolVar{out_grad[5]} * SymbolVar{opr.input(1)} / sqrt_var;
310 311 312 313 314

        ret[0] = dx.node();
        ret[1] = d_bn_scale.node();
        ret[2] = d_bn_bias.node();
        return ret;
315
    }
316
    return ret;
317
}
318
#endif
319 320 321 322 323

MGB_DYN_TYPE_OBJ_FINAL_IMPL(BatchNormBackward);

BatchNormBackward::BatchNormBackward(VarNode *x,
        VarNode *y_grad, VarNode *save_mean,
324
        VarNode* save_variance, VarNode *scale, VarNode *reserve,
325 326
        const Param &param, const OperatorNodeConfig &config):
    Super({x->owner_graph(), config, "batch_norm_bwd",
327
            {x, y_grad, save_mean, save_variance, scale, reserve}},
328 329 330
            0, true)
{
    init_megdnn_opr(*this, param);
331
    add_input({x, y_grad, save_mean, save_variance, scale, reserve});
332 333 334 335
}

SymbolVarArray BatchNormBackward::make(SymbolVar x,
        SymbolVar y_grad, SymbolVar save_mean,
336
        SymbolVar save_variance, SymbolVar scale, SymbolVar reserve,
337 338 339 340 341 342
        const Param &param,
        const OperatorNodeConfig &config) {
    auto&& out = x.node()
                    ->owner_graph()
                    ->insert_opr(std::make_unique<BatchNormBackward>(
                        x.node(), y_grad.node(), save_mean.node(),
343
                        save_variance.node(), scale.node(), reserve.node(), param, config))
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
                    ->output();
    SymbolVarArray ret(out.size());
    for (size_t i = 0; i < ret.size(); i++) {
        ret[i] = out[i];
    }
    return ret;
}

void BatchNormBackward::init_output_static_infer_desc() {

    using namespace cg::static_infer;
    auto &&mgr = owner_graph()->static_infer_manager();

    mgr.register_shape_infer(output(0),
            ShapeInferDesc::make_identity(input(4)));
    mgr.register_shape_infer(output(1),
            ShapeInferDesc::make_identity(input(4)));
    mgr.register_shape_infer(output(2),
            ShapeInferDesc::make_identity(input(0)));
    this->init_output_static_infer_desc_workspace(
            intl::AutoAddWorkspaceNeedLimitGetter<megdnn::BNBackward>::val);
}

void BatchNormBackward::init_output_dtype() {
    mgb_assert(input(0)->dtype().category() == input(2)->dtype().category());
    mgb_assert(input(0)->dtype() == input(1)->dtype());
    mgb_assert(input(2)->dtype() == input(3)->dtype());
    mgb_assert(input(2)->dtype() == input(4)->dtype());
    output(0)->dtype(input(2)->dtype());
    output(1)->dtype(input(2)->dtype());
    output(2)->dtype(input(0)->dtype());
}

377 378 379 380 381 382 383
cg::OperatorNodeBase::NodeProp*
BatchNormBackward::do_make_node_prop() const {
    auto ret = Super::do_make_node_prop();
    ret->add_dep_type_existing_var(input(5),
                                   NodeProp::DepType::VALUE_ALLOW_EMPTY);
    return ret;
}
384
// vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}