“7c3ec22081835371ee1e0f7167723c976fbbf2fd”上不存在“paddle/legacy/math/Matrix.h”
lstmp_op.h 21.1 KB
Newer Older
1 2 3 4 5 6
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

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

Y
Yibing Liu 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
8 9 10 11 12 13 14 15

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. */

#pragma once
Y
Yi Wang 已提交
16 17 18 19 20
#include "paddle/fluid/operators/activation_op.h"
#include "paddle/fluid/operators/math/detail/activation_functions.h"
#include "paddle/fluid/operators/math/lstm_compute.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/operators/math/sequence2batch.h"
21

Y
Yi Wang 已提交
22 23
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
24

25 26 27 28 29 30
namespace paddle {
namespace operators {

using LoDTensor = framework::LoDTensor;
using Tensor = framework::Tensor;

31 32 33 34
template <typename T, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;

35 36
template <typename DeviceContext, typename T>
inline void ReorderInitState(const DeviceContext& ctx,
D
dzhwinter 已提交
37 38
                             const framework::Tensor& src,
                             framework::Vector<size_t> index,
39 40 41 42 43 44 45 46 47
                             framework::Tensor* dst, bool indexed_src) {
  math::CopyMatrixRowsFunctor<DeviceContext, T> row_shuffle;
  dst->mutable_data<T>(src.dims(), ctx.GetPlace());
  row_shuffle(ctx, src, index, *dst, indexed_src);
}

template <typename DeviceContext, typename T>
class LSTMPKernel : public framework::OpKernel<T> {
 public:
48 49 50 51 52 53 54 55 56 57 58 59 60 61 62
  template <typename Device, typename X, typename Y>
  void ActCompute(const math::detail::ActivationType act_type, const Device& d,
                  X x, Y y) const {
    if (act_type == math::detail::ActivationType::kIdentity)
      y.device(d) = x;
    else if (act_type == math::detail::ActivationType::kSigmoid)
      SigmoidFunctor<T>()(d, x, y);
    else if (act_type == math::detail::ActivationType::kTanh)
      TanhFunctor<T>()(d, x, y);
    else if (act_type == math::detail::ActivationType::kReLU)
      ReluFunctor<T>()(d, x, y);
    else
      PADDLE_THROW("unsupported activation type");
  }

63 64 65 66 67 68 69
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* input = ctx.Input<LoDTensor>("Input");
    auto* weight = ctx.Input<Tensor>("Weight");
    auto* proj_weight = ctx.Input<Tensor>("ProjWeight");
    auto* bias = ctx.Input<Tensor>("Bias");

    auto* hidden_t0 = ctx.Input<Tensor>("H0");
70
    auto* ordered_proj0 = ctx.Output<Tensor>("OrderedP0");
71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112
    auto* cell_t0 = ctx.Input<Tensor>("C0");

    auto* batch_gate = ctx.Output<LoDTensor>("BatchGate");
    batch_gate->mutable_data<T>(ctx.GetPlace());
    auto* proj_out = ctx.Output<LoDTensor>("Projection");
    proj_out->mutable_data<T>(ctx.GetPlace());
    auto* cell_out = ctx.Output<LoDTensor>("Cell");
    cell_out->mutable_data<T>(ctx.GetPlace());

    bool is_reverse = ctx.Attr<bool>("is_reverse");
    math::LoDTensor2BatchFunctor<DeviceContext, T> to_batch;
    auto& device_ctx = ctx.template device_context<DeviceContext>();
    to_batch(device_ctx, *input, *batch_gate, true, is_reverse);

    auto in_dims = input->dims();
    int frame_size = static_cast<int>(in_dims[1] / 4);
    framework::DDim dims({in_dims[0], frame_size});
    framework::DDim proj_dims({in_dims[0], proj_weight->dims()[1]});

    if (bias) {
      Tensor b = *bias;
      b.Resize({bias->numel(), 1});
      Tensor gate_bias = b.Slice(0, 4 * frame_size);
      math::RowwiseAdd<DeviceContext, T> add_bias;
      add_bias(device_ctx, *batch_gate, gate_bias, batch_gate);
    }

    math::LstmMetaValue<T> lstmp_value;
    if (bias && ctx.Attr<bool>("use_peepholes")) {
      T* bias_data = const_cast<T*>(bias->data<T>());
      // the code style in LstmpMetaValue will be updated later.

      lstmp_value.check_ig = bias_data + 4 * frame_size;
      lstmp_value.check_fg = lstmp_value.check_ig + frame_size;
      lstmp_value.check_og = lstmp_value.check_fg + frame_size;
    } else {
      lstmp_value.check_ig = nullptr;
      lstmp_value.check_fg = nullptr;
      lstmp_value.check_og = nullptr;
    }
    lstmp_value.prev_state_value = nullptr;
    Tensor ordered_c0;
D
dzhwinter 已提交
113 114 115

    framework::Vector<size_t> order(batch_gate->lod()[2]);

116 117 118 119 120 121 122 123 124 125
    if (cell_t0) {
      // Since the batch computing for LSTMP reorders the input sequence
      // according to their length. The initialized cell state also needs
      // to reorder.
      ReorderInitState<DeviceContext, T>(device_ctx, *cell_t0, order,
                                         &ordered_c0, true);
      lstmp_value.prev_state_value = ordered_c0.data<T>();
    }

    // Use the local variable as here.
126
    LoDTensor batch_proj, batch_cell;
127
    auto* batch_cell_pre_act = ctx.Output<LoDTensor>("BatchCellPreAct");
128 129 130
    batch_cell_pre_act->mutable_data<T>(dims, ctx.GetPlace());
    auto* batch_hidden = ctx.Output<LoDTensor>("BatchHidden");
    batch_hidden->mutable_data<T>(dims, ctx.GetPlace());    // T x D
131 132 133 134 135 136 137 138 139 140 141
    batch_proj.mutable_data<T>(proj_dims, ctx.GetPlace());  // T x P
    batch_cell.mutable_data<T>(dims, ctx.GetPlace());       // T x D

    auto batch_starts = batch_gate->lod()[0];
    size_t num_batch = batch_starts.size() - 1;
    auto gate_act = math::detail::GetActivationType(
        ctx.Attr<std::string>("gate_activation"));
    auto cell_act = math::detail::GetActivationType(
        ctx.Attr<std::string>("cell_activation"));
    auto cand_act = math::detail::GetActivationType(
        ctx.Attr<std::string>("candidate_activation"));
142 143
    auto proj_act = math::detail::GetActivationType(
        ctx.Attr<std::string>("proj_activation"));
144
    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
145 146 147 148 149 150

    for (size_t n = 0; n < num_batch; n++) {
      int bstart = static_cast<int>(batch_starts[n]);
      int bend = static_cast<int>(batch_starts[n + 1]);

      Tensor gate_t = batch_gate->Slice(bstart, bend);
151
      Tensor hidden_t = batch_hidden->Slice(bstart, bend);
152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172
      Tensor proj_t = batch_proj.Slice(bstart, bend);
      Tensor cell_t = batch_cell.Slice(bstart, bend);
      Tensor cell_pre_act_t = batch_cell_pre_act->Slice(bstart, bend);

      int cur_batch_size = bend - bstart;

      if (n > 0) {
        int pre_h_start = static_cast<int>(batch_starts[n - 1]);
        int pre_h_end = pre_h_start + cur_batch_size;
        auto pre_proj_t = batch_proj.Slice(pre_h_start, pre_h_end);
        math::matmul<DeviceContext, T>(device_ctx, pre_proj_t, false, *weight,
                                       false, static_cast<T>(1.0), &gate_t,
                                       static_cast<T>(1.0));
      } else if (hidden_t0) {
        // If n == 0 and there is no initialized hidden state, that is to say
        // the H0 is zeros, the calculation W_h * H0 will be skiped.
        // If n == 0 and there is initialized hidden state, calculate W_h * H0.

        // Since the batch computing for LSTMP reorders the input sequence
        // according to their length. The initialized hidden state also needs
        // to reorder.
173 174 175

        Tensor ordered_h0;
        ordered_proj0->mutable_data<T>(ctx.GetPlace());
176 177 178 179
        ReorderInitState<DeviceContext, T>(device_ctx, *hidden_t0, order,
                                           &ordered_h0, true);
        math::matmul<DeviceContext, T>(device_ctx, ordered_h0, false,
                                       *proj_weight, false, static_cast<T>(1.0),
180
                                       ordered_proj0, static_cast<T>(0.0));
181
        if (proj_act != math::detail::ActivationType::kIdentity) {
182 183 184 185
          auto proj0_dev = EigenMatrix<T>::From(*ordered_proj0);
          ActCompute(cell_act, place, proj0_dev, proj0_dev);
        }
        math::matmul<DeviceContext, T>(device_ctx, *ordered_proj0, false,
186 187 188 189 190 191 192 193 194 195 196 197 198 199 200
                                       *weight, false, static_cast<T>(1.0),
                                       &gate_t, static_cast<T>(1.0));
      }

      lstmp_value.gate_value = gate_t.data<T>();
      lstmp_value.output_value = hidden_t.data<T>();
      lstmp_value.state_value = cell_t.data<T>();
      lstmp_value.state_active_value = cell_pre_act_t.data<T>();
      math::LstmUnitFunctor<DeviceContext, T>::compute(
          device_ctx, lstmp_value, frame_size, cur_batch_size, gate_act,
          cell_act, cand_act);
      lstmp_value.prev_state_value = lstmp_value.state_value;
      math::matmul<DeviceContext, T>(device_ctx, hidden_t, false, *proj_weight,
                                     false, static_cast<T>(1.0), &proj_t,
                                     static_cast<T>(0.0));
201
      if (proj_act != math::detail::ActivationType::kIdentity) {
202 203 204
        auto proj_t_dev = EigenMatrix<T>::From(proj_t);
        ActCompute(cell_act, place, proj_t_dev, proj_t_dev);
      }
205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220
    }

    math::Batch2LoDTensorFunctor<DeviceContext, T> to_seq;
    batch_proj.set_lod(batch_gate->lod());
    // restore the output hidden in LoDTensor from the batch hidden
    to_seq(device_ctx, batch_proj, *proj_out);

    batch_cell.set_lod(batch_gate->lod());
    // restore the output cell state in LoDTensor from the batch cell
    to_seq(device_ctx, batch_cell, *cell_out);
  }
};

template <typename DeviceContext, typename T>
class LSTMPGradKernel : public framework::OpKernel<T> {
 public:
221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236
  template <typename Device, typename X, typename Y, typename DX, typename DY>
  void ActGradCompute(const math::detail::ActivationType act_type,
                      const Device& d, X x, Y y, DX dx, DY dy) const {
    // x is dummy and won't be used even in Relu(use y instead)
    if (act_type == math::detail::ActivationType::kIdentity)
      dx.device(d) = dy;
    else if (act_type == math::detail::ActivationType::kSigmoid)
      SigmoidGradFunctor<T>()(d, x, y, dy, dx);
    else if (act_type == math::detail::ActivationType::kTanh)
      TanhGradFunctor<T>()(d, x, y, dy, dx);
    else if (act_type == math::detail::ActivationType::kReLU)
      ReluGradFunctor<T>()(d, x, y, dy, dx);
    else
      PADDLE_THROW("unsupported activation type");
  }

237 238 239
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* input = ctx.Input<LoDTensor>("Input");
    auto* weight = ctx.Input<Tensor>("Weight");
240
    auto* proj_weight = ctx.Input<Tensor>("ProjWeight");
241 242 243 244 245 246 247
    auto* bias = ctx.Input<Tensor>("Bias");

    auto* proj_out = ctx.Input<LoDTensor>("Projection");
    auto* cell_out = ctx.Input<LoDTensor>("Cell");

    auto* batch_gate = ctx.Input<LoDTensor>("BatchGate");
    auto* batch_cell_pre_act = ctx.Input<LoDTensor>("BatchCellPreAct");
248
    auto* batch_hidden = ctx.Input<LoDTensor>("BatchHidden");
249

250 251
    auto* projection_g =
        ctx.Input<LoDTensor>(framework::GradVarName("Projection"));
252 253 254

    auto* in_g = ctx.Output<LoDTensor>(framework::GradVarName("Input"));
    auto* weight_g = ctx.Output<Tensor>(framework::GradVarName("Weight"));
255 256
    auto* proj_weight_g =
        ctx.Output<Tensor>(framework::GradVarName("ProjWeight"));
257 258 259
    auto* bias_g = ctx.Output<Tensor>(framework::GradVarName("Bias"));

    auto* h0 = ctx.Input<Tensor>("H0");
260
    auto* ordered_proj0 = ctx.Input<Tensor>("OrderedP0");
261 262 263 264 265 266 267 268 269 270 271
    auto* c0 = ctx.Input<Tensor>("C0");

    auto* h0_g = ctx.Output<Tensor>(framework::GradVarName("H0"));
    auto* c0_g = ctx.Output<Tensor>(framework::GradVarName("C0"));

    auto& device_ctx = ctx.template device_context<DeviceContext>();
    math::SetConstant<DeviceContext, T> zero;
    if (weight_g) {
      weight_g->mutable_data<T>(ctx.GetPlace());
      zero(device_ctx, weight_g, static_cast<T>(0.0));
    }
272 273 274 275
    if (proj_weight_g) {
      proj_weight_g->mutable_data<T>(ctx.GetPlace());
      zero(device_ctx, proj_weight_g, static_cast<T>(0.0));
    }
276 277 278 279 280

    // ordered_h0/c0 is the reordered hidden/cell initialization.
    // ordered_h0_g/c0_g is the reordered gradient of hidden/cell
    // initialization.
    Tensor ordered_h0, ordered_c0, ordered_h0_g, ordered_c0_g;
D
dzhwinter 已提交
281 282 283

    framework::Vector<size_t> order(batch_gate->lod()[2]);

284 285 286 287 288 289 290 291 292
    if (c0) {
      ReorderInitState<DeviceContext, T>(device_ctx, *c0, order, &ordered_c0,
                                         true);
    }
    if (c0 && c0_g) {
      ordered_c0_g.mutable_data<T>(c0_g->dims(), ctx.GetPlace());
    }

    auto in_dims = input->dims();
293 294
    auto out_dims = cell_out->dims();
    framework::DDim proj_dims({in_dims[0], proj_weight->dims()[1]});
295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336
    int frame_size = static_cast<int>(in_dims[1] / 4);
    PADDLE_ENFORCE_EQ(frame_size, out_dims[1]);

    math::LstmMetaValue<T> lstmp_value;
    if (bias && ctx.Attr<bool>("use_peepholes")) {
      T* bias_data = const_cast<T*>(bias->data<T>());
      lstmp_value.check_ig = bias_data + 4 * frame_size;
      lstmp_value.check_fg = lstmp_value.check_ig + frame_size;
      lstmp_value.check_og = lstmp_value.check_fg + frame_size;
    } else {
      lstmp_value.check_ig = nullptr;
      lstmp_value.check_fg = nullptr;
      lstmp_value.check_og = nullptr;
    }

    math::LstmMetaGrad<T> lstmp_grad;

    if (bias && bias_g) {
      bias_g->mutable_data<T>(ctx.GetPlace());
      zero(device_ctx, bias_g, static_cast<T>(0.0));
    }
    if (bias && bias_g && ctx.Attr<bool>("use_peepholes")) {
      T* bias_g_data = bias_g->data<T>();
      lstmp_grad.check_ig_grad = bias_g_data + 4 * frame_size;
      lstmp_grad.check_fg_grad = lstmp_grad.check_ig_grad + frame_size;
      lstmp_grad.check_og_grad = lstmp_grad.check_fg_grad + frame_size;
    } else {
      lstmp_grad.check_ig_grad = nullptr;
      lstmp_grad.check_fg_grad = nullptr;
      lstmp_grad.check_og_grad = nullptr;
    }

    math::LoDTensor2BatchFunctor<DeviceContext, T> to_batch;

    auto ToBatch = [&batch_gate, &to_batch](
        const DeviceContext& ctx, const framework::LoDTensor& src,
        const framework::DDim& dims, framework::LoDTensor& dst) {
      dst.mutable_data<T>(dims, ctx.GetPlace());
      dst.set_lod(batch_gate->lod());
      to_batch(ctx, src, dst, false);
    };

337 338 339 340 341
    LoDTensor batch_hidden_g, batch_proj, batch_proj_g, batch_cell;
    batch_hidden_g.mutable_data<T>(out_dims, ctx.GetPlace());
    ToBatch(device_ctx, *proj_out, proj_dims, batch_proj);        // T x P
    ToBatch(device_ctx, *projection_g, proj_dims, batch_proj_g);  // T x P
    ToBatch(device_ctx, *cell_out, out_dims, batch_cell);         // T x D
342 343 344 345 346 347 348 349 350 351 352 353 354 355 356

    LoDTensor batch_cell_g, batch_gate_g;
    batch_cell_g.mutable_data<T>(out_dims, ctx.GetPlace());
    // TODO(qingqing) support the case output cell has gradient.
    // to_batch(device_ctx, *cell_g, batch_cell_g, false);
    zero(device_ctx, &batch_cell_g, static_cast<T>(0.0));
    batch_gate_g.mutable_data<T>(batch_gate->dims(), ctx.GetPlace());
    batch_gate_g.set_lod(batch_gate->lod());

    auto gate_act = math::detail::GetActivationType(
        ctx.Attr<std::string>("gate_activation"));
    auto cell_act = math::detail::GetActivationType(
        ctx.Attr<std::string>("cell_activation"));
    auto cand_act = math::detail::GetActivationType(
        ctx.Attr<std::string>("candidate_activation"));
357 358
    auto proj_act = math::detail::GetActivationType(
        ctx.Attr<std::string>("proj_activation"));
359
    auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
360 361 362 363 364 365 366

    auto batch_starts = batch_gate->lod()[0];
    size_t num_batch = batch_starts.size() - 1;
    for (int n = static_cast<int>(num_batch) - 1; n >= 0; n--) {
      int bstart = static_cast<int>(batch_starts[n]);
      int bend = static_cast<int>(batch_starts[n + 1]);

367 368
      Tensor cur_proj = batch_proj.Slice(bstart, bend);
      Tensor proj_g = batch_proj_g.Slice(bstart, bend);
369
      if (proj_act != math::detail::ActivationType::kIdentity) {
370 371 372 373 374
        auto cur_proj_dev = EigenMatrix<T>::From(cur_proj);
        auto proj_g_dev = EigenMatrix<T>::From(proj_g);
        ActGradCompute(cell_act, place, cur_proj_dev, cur_proj_dev, proj_g_dev,
                       proj_g_dev);
      }
375
      /* hidden state backwarad */
376 377 378 379
      Tensor out_g = batch_hidden_g.Slice(bstart, bend);
      math::matmul<DeviceContext, T>(device_ctx, proj_g, false, *proj_weight,
                                     true, static_cast<T>(1.0), &out_g,
                                     static_cast<T>(0.0));
380 381 382 383 384 385 386
      /* projection weight backward*/
      if (proj_weight_g) {
        Tensor hidden_t = batch_hidden->Slice(bstart, bend);
        math::matmul<DeviceContext, T>(device_ctx, hidden_t, true, proj_g,
                                       false, static_cast<T>(1.0),
                                       proj_weight_g, static_cast<T>(1.0));
      }
387

388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424
      Tensor gate = batch_gate->Slice(bstart, bend);
      Tensor cell = batch_cell.Slice(bstart, bend);
      Tensor cell_pre_act = batch_cell_pre_act->Slice(bstart, bend);
      lstmp_value.gate_value = gate.data<T>();
      lstmp_value.state_value = cell.data<T>();
      lstmp_value.state_active_value = cell_pre_act.data<T>();

      Tensor gate_g = batch_gate_g.Slice(bstart, bend);
      Tensor cell_g = batch_cell_g.Slice(bstart, bend);
      lstmp_grad.state_grad = cell_g.data<T>();
      lstmp_grad.gate_grad = gate_g.data<T>();
      lstmp_grad.output_grad = out_g.data<T>();

      if (n > 0) {
        int bstart_pre = static_cast<int>(batch_starts[n - 1]);
        Tensor cell_pre = batch_cell.Slice(bstart_pre, bstart);
        Tensor cell_pre_g = batch_cell_g.Slice(bstart_pre, bstart);
        lstmp_value.prev_state_value = cell_pre.data<T>();
        lstmp_grad.prev_state_grad = cell_pre_g.data<T>();
      } else {
        lstmp_value.prev_state_value = c0 ? ordered_c0.data<T>() : nullptr;
        lstmp_grad.prev_state_grad = c0_g ? ordered_c0_g.data<T>() : nullptr;
      }

      int cur_batch_size = bend - bstart;
      math::LstmUnitGradFunctor<DeviceContext, T>::compute(
          device_ctx, lstmp_value, lstmp_grad, frame_size, cur_batch_size,
          gate_act, cell_act, cand_act);

      if (n > 0) {
        int pre_h_start = static_cast<int>(batch_starts[n - 1]);
        int pre_h_end = pre_h_start + cur_batch_size;
        auto pre_proj_g = batch_proj_g.Slice(pre_h_start, pre_h_end);
        math::matmul<DeviceContext, T>(device_ctx, gate_g, false, *weight, true,
                                       static_cast<T>(1.0), &pre_proj_g,
                                       static_cast<T>(1.0));
        if (weight_g) {
425
          /* weight backward*/
426 427 428 429 430 431 432 433 434
          auto pre_proj = batch_proj.Slice(pre_h_start, pre_h_end);
          math::matmul<DeviceContext, T>(device_ctx, pre_proj, true, gate_g,
                                         false, static_cast<T>(1.0), weight_g,
                                         static_cast<T>(1.0));
        }
      } else {
        if (h0 && weight_g) {
          ReorderInitState<DeviceContext, T>(device_ctx, *h0, order,
                                             &ordered_h0, true);
435 436 437 438 439
          if (weight_g) {
            math::matmul<DeviceContext, T>(device_ctx, *ordered_proj0, true,
                                           gate_g, false, static_cast<T>(1.0),
                                           weight_g, static_cast<T>(1.0));
          }
440
        }
441
        if (h0 && (h0_g || proj_weight_g)) {
442
          ordered_h0_g.mutable_data<T>(h0_g->dims(), ctx.GetPlace());
443 444 445
          Tensor proj0_g;
          proj0_g.Resize({in_dims[0], proj_weight->dims()[1]});
          proj0_g.mutable_data<T>(ctx.GetPlace());
446
          math::matmul<DeviceContext, T>(device_ctx, gate_g, false, *weight,
447 448
                                         true, static_cast<T>(1.0), &proj0_g,
                                         static_cast<T>(0.0));
449
          if (proj_act != math::detail::ActivationType::kIdentity) {
450 451 452 453 454 455 456 457 458 459 460 461 462 463 464
            auto proj0_dev = EigenMatrix<T>::From(*ordered_proj0);
            auto proj0_g_dev = EigenMatrix<T>::From(proj0_g);
            ActGradCompute(cell_act, place, proj0_dev, proj0_dev, proj0_g_dev,
                           proj0_g_dev);
          }
          if (h0_g) {
            math::matmul<DeviceContext, T>(
                device_ctx, proj0_g, false, *proj_weight, true,
                static_cast<T>(1.0), &ordered_h0_g, static_cast<T>(0.0));
          }
          if (proj_weight_g) {
            math::matmul<DeviceContext, T>(device_ctx, ordered_h0, true,
                                           proj0_g, false, static_cast<T>(1.0),
                                           proj_weight_g, static_cast<T>(1.0));
          }
465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496
        }
      }
    }

    math::Batch2LoDTensorFunctor<DeviceContext, T> to_seq;
    if (in_g) {
      /* backward data */
      in_g->mutable_data<T>(ctx.GetPlace());
      to_seq(device_ctx, batch_gate_g, *in_g);
    }
    if (bias && bias_g) {
      /* backward bias */
      Tensor b_g = *bias_g;
      b_g.Resize({bias_g->numel(), 1});
      Tensor gate_bias_g = b_g.Slice(0, 4 * frame_size);
      math::ColwiseSum<DeviceContext, T> col_sum;
      col_sum(device_ctx, batch_gate_g, &gate_bias_g);
    }

    if (h0 && h0_g) {
      ReorderInitState<DeviceContext, T>(device_ctx, ordered_h0_g, order, h0_g,
                                         false);
    }
    if (c0 && c0_g) {
      ReorderInitState<DeviceContext, T>(device_ctx, ordered_c0_g, order, c0_g,
                                         false);
    }
  }
};

}  // namespace operators
}  // namespace paddle