adam_op_mlu.cc 22.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29
/* 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. */

#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/tensor_util.h"
#include "paddle/fluid/operators/reduce_ops/reduce_op_mlu.h"

namespace paddle {
namespace operators {

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

template <typename T>
class AdamMLUKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    const auto* param_var = ctx.InputVar("Param");
30 31
    PADDLE_ENFORCE_EQ(param_var->IsType<framework::LoDTensor>(),
                      true,
32 33 34 35 36 37 38
                      platform::errors::InvalidArgument(
                          "The Var(%s)'s type should be LoDTensor, "
                          "but the received is %s",
                          ctx.InputNames("Param").front(),
                          framework::ToTypeName(param_var->Type())));
    auto* param = ctx.Input<LoDTensor>("Param");
    auto* grad_var = ctx.InputVar("Grad");
39 40
    PADDLE_ENFORCE_EQ(grad_var->IsType<framework::LoDTensor>(),
                      true,
41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62
                      platform::errors::InvalidArgument(
                          "The Grad(%s)'s type should be LoDTensor, "
                          "but the received is %s",
                          ctx.InputNames("Grad").front(),
                          framework::ToTypeName(param_var->Type())));
    auto* grad = ctx.Input<LoDTensor>("Grad");
    auto* mom1 = ctx.Input<LoDTensor>("Moment1");
    auto* mom2 = ctx.Input<LoDTensor>("Moment2");
    auto* lr = ctx.Input<LoDTensor>("LearningRate");

    auto* beta1_pow = ctx.Input<Tensor>("Beta1Pow");
    auto* beta2_pow = ctx.Input<Tensor>("Beta2Pow");

    auto* param_out = ctx.Output<LoDTensor>("ParamOut");
    auto* mom1_out = ctx.Output<LoDTensor>("Moment1Out");
    auto* mom2_out = ctx.Output<LoDTensor>("Moment2Out");
    auto* beta1_pow_out = ctx.Output<LoDTensor>("Beta1PowOut");
    auto* beta2_pow_out = ctx.Output<LoDTensor>("Beta2PowOut");

    bool skip_update = false;
    if (ctx.HasInput("SkipUpdate")) {
      auto* skip_update_tensor = ctx.Input<framework::Tensor>("SkipUpdate");
63 64
      PADDLE_ENFORCE_EQ(skip_update_tensor->numel(),
                        1,
65 66 67 68
                        platform::errors::InvalidArgument(
                            "Input(SkipUpdate) size must be 1, but get %d",
                            skip_update_tensor->numel()));
      std::vector<bool> skip_update_vec;
69 70
      paddle::framework::TensorToVector(
          *skip_update_tensor, ctx.device_context(), &skip_update_vec);
F
fwenguang 已提交
71
      ctx.device_context().Wait();
72 73 74 75 76 77 78
      skip_update = skip_update_vec[0];
    }
    // skip_update=true, just copy input to output, and TensorCopy will call
    // mutable_data
    if (skip_update) {
      VLOG(4) << "Adam skip update";
      framework::TensorCopy(
79 80 81 82
          *param,
          ctx.GetPlace(),
          ctx.template device_context<platform::MLUDeviceContext>(),
          param_out);
83
      framework::TensorCopy(
84 85 86 87
          *mom1,
          ctx.GetPlace(),
          ctx.template device_context<platform::MLUDeviceContext>(),
          mom1_out);
88
      framework::TensorCopy(
89 90 91 92
          *mom2,
          ctx.GetPlace(),
          ctx.template device_context<platform::MLUDeviceContext>(),
          mom2_out);
93
      framework::TensorCopy(
94 95
          *beta1_pow,
          beta1_pow->place(),
96 97 98
          ctx.template device_context<platform::MLUDeviceContext>(),
          beta1_pow_out);
      framework::TensorCopy(
99 100
          *beta2_pow,
          beta2_pow->place(),
101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118
          ctx.template device_context<platform::MLUDeviceContext>(),
          beta2_pow_out);
      return;
    }

    bool use_global_beta_pow = ctx.Attr<bool>("use_global_beta_pow");
    VLOG(4) << "use_global_beta_pow:" << use_global_beta_pow;

    param_out->ShareDataWith(*param);
    mom1_out->ShareDataWith(*mom1);
    mom2_out->ShareDataWith(*mom2);

    LoDTensor beta1_pow_tmp;
    LoDTensor beta2_pow_tmp;
    if (beta1_pow->place() == platform::CPUPlace()) {
      T beta1 = *beta1_pow->data<T>();
      beta1_pow_tmp.mutable_data<T>({1}, ctx.GetPlace());
      MLUCnnlTensorDesc beta1_pow_tmp_desc(beta1_pow_tmp);
119 120 121 122 123
      MLUCnnl::Fill(ctx,
                    CNNL_POINTER_MODE_HOST,
                    &beta1,
                    beta1_pow_tmp_desc.get(),
                    GetBasePtr(&beta1_pow_tmp));
124 125 126 127 128 129
      beta1_pow = &beta1_pow_tmp;
    }
    if (beta2_pow->place() == platform::CPUPlace()) {
      T beta2 = *beta2_pow->data<T>();
      beta2_pow_tmp.mutable_data<T>({1}, ctx.GetPlace());
      MLUCnnlTensorDesc beta2_pow_tmp_desc(beta2_pow_tmp);
130 131 132 133 134
      MLUCnnl::Fill(ctx,
                    CNNL_POINTER_MODE_HOST,
                    &beta2,
                    beta2_pow_tmp_desc.get(),
                    GetBasePtr(&beta2_pow_tmp));
135 136 137 138 139 140 141
      beta2_pow = &beta2_pow_tmp;
    }

    VLOG(3) << "beta1_pow.numel() : " << beta1_pow->numel()
            << "beta2_pow.numel() : " << beta2_pow->numel();
    VLOG(3) << "param.numel(): " << param->numel();

142 143
    PADDLE_ENFORCE_EQ(beta1_pow_out->numel(),
                      1,
144 145 146 147 148
                      platform::errors::InvalidArgument(
                          "beta1 pow output size should be 1, but received "
                          "value is:%d.",
                          beta1_pow_out->numel()));

149 150
    PADDLE_ENFORCE_EQ(beta2_pow_out->numel(),
                      1,
151 152 153 154 155 156 157 158 159 160 161 162 163 164 165
                      platform::errors::InvalidArgument(
                          "beta2 pow output size should be 1, but received "
                          "value is:%d.",
                          beta2_pow_out->numel()));

    const Tensor* beta1_tensor = nullptr;
    const Tensor* beta2_tensor = nullptr;
    const Tensor* epsilon_tensor = nullptr;

    Tensor beta1_tmp(experimental::DataType::FLOAT32);
    Tensor beta2_tmp(experimental::DataType::FLOAT32);
    Tensor epsilon_tmp(experimental::DataType::FLOAT32);

    if (ctx.HasInput("Beta1Tensor")) {
      beta1_tensor = ctx.Input<framework::Tensor>("Beta1Tensor");
166 167
      PADDLE_ENFORCE_EQ(beta1_tensor->numel(),
                        1,
168 169 170 171 172 173 174
                        platform::errors::InvalidArgument(
                            "Input(Beta1Tensor) size must be 1, but get %d",
                            beta1_tensor->numel()));
    } else {
      T beta1 = static_cast<T>(ctx.Attr<float>("beta1"));
      beta1_tmp.mutable_data<T>({1}, ctx.GetPlace());
      MLUCnnlTensorDesc beta1_tmp_desc(beta1_tmp);
175 176 177 178
      MLUCnnl::Fill(ctx,
                    CNNL_POINTER_MODE_HOST,
                    &beta1,
                    beta1_tmp_desc.get(),
179 180 181 182 183 184
                    GetBasePtr(&beta1_tmp));
      beta1_tensor = &beta1_tmp;
    }

    if (ctx.HasInput("Beta2Tensor")) {
      beta2_tensor = ctx.Input<framework::Tensor>("Beta2Tensor");
185 186
      PADDLE_ENFORCE_EQ(beta2_tensor->numel(),
                        1,
187 188 189 190 191 192 193
                        platform::errors::InvalidArgument(
                            "Input(Beta2Tensor) size must be 1, but get %d",
                            beta2_tensor->numel()));
    } else {
      T beta2 = static_cast<T>(ctx.Attr<float>("beta2"));
      beta2_tmp.mutable_data<T>({1}, ctx.GetPlace());
      MLUCnnlTensorDesc beta2_tmp_desc(beta2_tmp);
194 195 196 197
      MLUCnnl::Fill(ctx,
                    CNNL_POINTER_MODE_HOST,
                    &beta2,
                    beta2_tmp_desc.get(),
198 199 200 201 202 203
                    GetBasePtr(&beta2_tmp));
      beta2_tensor = &beta2_tmp;
    }

    if (ctx.HasInput("EpsilonTensor")) {
      epsilon_tensor = ctx.Input<framework::Tensor>("EpsilonTensor");
204 205
      PADDLE_ENFORCE_EQ(epsilon_tensor->numel(),
                        1,
206 207 208 209 210 211 212
                        platform::errors::InvalidArgument(
                            "Input(EpsilonTensor) size must be 1, but get %d",
                            epsilon_tensor->numel()));
    } else {
      T epsilon = static_cast<T>(ctx.Attr<float>("epsilon"));
      epsilon_tmp.mutable_data<T>({1}, ctx.GetPlace());
      MLUCnnlTensorDesc epsilon_tmp_desc(epsilon_tmp);
213 214 215 216 217
      MLUCnnl::Fill(ctx,
                    CNNL_POINTER_MODE_HOST,
                    &epsilon,
                    epsilon_tmp_desc.get(),
                    GetBasePtr(&epsilon_tmp));
218 219 220 221 222 223 224
      epsilon_tensor = &epsilon_tmp;
    }

    MLUCnnlTensorDesc param_desc(*param);
    MLUCnnlTensorDesc mom1_desc(*mom1);
    MLUCnnlTensorDesc mom2_desc(*mom2);
    MLUCnnlTensorDesc grad_desc(*grad);
225 226 227 228 229 230 231 232 233 234 235 236 237 238 239
    MLUCnnl::ApplyAdam(ctx,
                       param_desc.get(),
                       GetBasePtr(param_out),
                       mom1_desc.get(),
                       GetBasePtr(mom1_out),
                       mom2_desc.get(),
                       GetBasePtr(mom2_out),
                       grad_desc.get(),
                       GetBasePtr(grad),
                       GetBasePtr(lr),
                       GetBasePtr(beta1_tensor),
                       GetBasePtr(beta2_tensor),
                       GetBasePtr(beta1_pow),
                       GetBasePtr(beta2_pow),
                       GetBasePtr(epsilon_tensor),
240 241 242 243 244 245 246
                       /*use_nesterov*/ false);

    if (!use_global_beta_pow) {
      beta1_pow_out->mutable_data<T>(ctx.GetPlace());
      beta2_pow_out->mutable_data<T>(ctx.GetPlace());

      MLUCnnlTensorDesc beta1_desc(*beta1_tensor);
247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268
      MLUCnnlOpTensorDesc mul_op_desc(
          CNNL_OP_TENSOR_MUL, ToCnnlDataType<T>(), CNNL_NOT_PROPAGATE_NAN);

      MLUCnnl::OpTensor(ctx,
                        mul_op_desc.get(),
                        beta1_desc.get(),
                        GetBasePtr(beta1_pow),
                        beta1_desc.get(),
                        GetBasePtr(beta1_tensor),
                        beta1_desc.get(),
                        GetBasePtr(beta1_pow_out),
                        ToCnnlDataType<T>());

      MLUCnnl::OpTensor(ctx,
                        mul_op_desc.get(),
                        beta1_desc.get(),
                        GetBasePtr(beta2_pow),
                        beta1_desc.get(),
                        GetBasePtr(beta2_tensor),
                        beta1_desc.get(),
                        GetBasePtr(beta2_pow_out),
                        ToCnnlDataType<T>());
269 270 271 272 273 274 275 276 277 278 279 280 281
    }
  }
};

template <typename T>
class AdamWMLUKernel : public AdamMLUKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    VLOG(3) << "MLU AdamW Kernel";
    bool skip_update = false;
    if (ctx.HasInput("SkipUpdate")) {
      VLOG(3) << "Has SkipUpdate";
      auto* skip_update_tensor = ctx.Input<framework::Tensor>("SkipUpdate");
282 283
      PADDLE_ENFORCE_EQ(skip_update_tensor->numel(),
                        1,
284 285 286 287
                        platform::errors::InvalidArgument(
                            "Input(SkipUpdate) size must be 1, but get %d",
                            skip_update_tensor->numel()));
      std::vector<bool> skip_update_vec;
288 289
      paddle::framework::TensorToVector(
          *skip_update_tensor, ctx.device_context(), &skip_update_vec);
F
fwenguang 已提交
290
      ctx.device_context().Wait();
291 292 293
      skip_update = skip_update_vec[0];
    }
    bool with_decay = ctx.Attr<bool>("with_decay");
294
    VLOG(3) << "Skip update: " << skip_update << ", With decay: " << with_decay;
295 296 297 298 299 300
    if (!skip_update && with_decay) {
      if (ctx.HasInput("MasterParam")) {
        PADDLE_THROW(platform::errors::Unimplemented(
            "Master Param is not supported on MLU"));
      } else {
        const auto* param_var = ctx.InputVar("Param");
301 302
        PADDLE_ENFORCE_EQ(param_var->IsType<framework::LoDTensor>(),
                          true,
303 304 305 306 307 308 309 310 311 312 313 314
                          platform::errors::InvalidArgument(
                              "The Var(%s)'s type should be LoDTensor, "
                              "but the received is %s",
                              ctx.InputNames("Param").front(),
                              framework::ToTypeName(param_var->Type())));
        auto* param = ctx.Input<LoDTensor>("Param");
        auto* lr = ctx.Input<LoDTensor>("LearningRate");
        float coeff = ctx.Attr<float>("coeff");

        // update param with decay coeff: mul(-1 * lr, coeff * param) + param
        MLUCnnlTensorDesc lr_desc(*lr);
        MLUCnnlTensorDesc param_desc(*param);
315 316 317 318 319 320 321 322 323 324
        MLUCnnlOpTensorDesc mul_op_desc(
            CNNL_OP_TENSOR_MUL, ToCnnlDataType<T>(), CNNL_NOT_PROPAGATE_NAN);

        MLUCnnl::OpTensor(ctx,
                          mul_op_desc.get(),
                          lr_desc.get(),
                          GetBasePtr(lr),
                          param_desc.get(),
                          GetBasePtr(param),
                          param_desc.get(),
325 326
                          const_cast<void*>(GetBasePtr(param)),
                          ToCnnlDataType<T>(),
327 328 329
                          /*alpha1*/ -1.f,
                          /*alpha2*/ coeff,
                          /*beta*/ 1.f);
330 331 332 333 334 335
      }
    }
    AdamMLUKernel<T>::Compute(ctx);
  }
};

336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553
template <typename T>
class MergedAdamMLUKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    // Get inputs and outputs
    auto params = ctx.MultiInput<framework::Tensor>("Param");
    auto grads = ctx.MultiInput<framework::Tensor>("Grad");
    auto lrs = ctx.MultiInput<framework::Tensor>("LearningRate");
    auto mom1s = ctx.MultiInput<framework::Tensor>("Moment1");
    auto mom2s = ctx.MultiInput<framework::Tensor>("Moment2");
    auto beta1_pows = ctx.MultiInput<framework::Tensor>("Beta1Pow");
    auto beta2_pows = ctx.MultiInput<framework::Tensor>("Beta2Pow");
    auto master_params = ctx.MultiInput<framework::Tensor>("MasterParam");
    auto param_outs = ctx.MultiOutput<framework::Tensor>("ParamOut");
    auto mom1_outs = ctx.MultiOutput<framework::Tensor>("Moment1Out");
    auto mom2_outs = ctx.MultiOutput<framework::Tensor>("Moment2Out");
    auto beta1_pow_outs = ctx.MultiOutput<framework::Tensor>("Beta1PowOut");
    auto beta2_pow_outs = ctx.MultiOutput<framework::Tensor>("Beta2PowOut");

    // Check validation of inputs and outputs
    size_t param_num = params.size();
    PADDLE_ENFORCE_EQ(param_num,
                      param_outs.size(),
                      platform::errors::InvalidArgument(
                          "The size of Output(ParamOut) must be equal to "
                          "Input(Param), but got the size of Output(ParamOut) "
                          "is %d, the size of Input(Param) is %d.",
                          param_outs.size(),
                          param_num));

    bool skip_update = false;
    if (ctx.HasInput("SkipUpdate")) {
      auto* skip_update_tensor = ctx.Input<framework::Tensor>("SkipUpdate");
      PADDLE_ENFORCE_EQ(skip_update_tensor->numel(),
                        1,
                        platform::errors::InvalidArgument(
                            "Input(SkipUpdate) size must be 1, but get %d",
                            skip_update_tensor->numel()));
      std::vector<bool> skip_update_vec;
      paddle::framework::TensorToVector(
          *skip_update_tensor, ctx.device_context(), &skip_update_vec);
      ctx.device_context().Wait();
      skip_update = skip_update_vec[0];
    }
    // skip_update=true, just copy input to output, and TensorCopy will call
    // mutable_data

    if (skip_update) {
      VLOG(4) << "MergedAdam skip update";
      for (size_t i = 0; i < param_num; ++i) {
        framework::TensorCopy(
            *params[i],
            ctx.GetPlace(),
            ctx.template device_context<platform::MLUDeviceContext>(),
            param_outs[i]);
        framework::TensorCopy(
            *mom1s[i],
            ctx.GetPlace(),
            ctx.template device_context<platform::MLUDeviceContext>(),
            mom1_outs[i]);
        framework::TensorCopy(
            *mom2s[i],
            ctx.GetPlace(),
            ctx.template device_context<platform::MLUDeviceContext>(),
            mom2_outs[i]);
        framework::TensorCopy(
            *beta1_pows[i],
            beta1_pows[i]->place(),
            ctx.template device_context<platform::MLUDeviceContext>(),
            beta1_pow_outs[i]);
        framework::TensorCopy(
            *beta2_pows[i],
            beta2_pows[i]->place(),
            ctx.template device_context<platform::MLUDeviceContext>(),
            beta2_pow_outs[i]);
      }
      return;
    }

    bool use_global_beta_pow = ctx.Attr<bool>("use_global_beta_pow");
    VLOG(4) << "use_global_beta_pow:" << use_global_beta_pow;

    // Get beta1, beta2 and epsilon from attribute.
    const Tensor* beta1_tensor = nullptr;
    const Tensor* beta2_tensor = nullptr;
    const Tensor* epsilon_tensor = nullptr;

    Tensor beta1_tmp(experimental::DataType::FLOAT32);
    Tensor beta2_tmp(experimental::DataType::FLOAT32);
    Tensor epsilon_tmp(experimental::DataType::FLOAT32);

    T beta1 = static_cast<T>(ctx.Attr<float>("beta1"));
    T beta2 = static_cast<T>(ctx.Attr<float>("beta2"));
    T epsilon = static_cast<T>(ctx.Attr<float>("epsilon"));
    beta1_tmp.mutable_data<T>({1}, ctx.GetPlace());
    beta2_tmp.mutable_data<T>({1}, ctx.GetPlace());
    epsilon_tmp.mutable_data<T>({1}, ctx.GetPlace());
    MLUCnnlTensorDesc beta1_tmp_desc(beta1_tmp);
    MLUCnnlTensorDesc beta2_tmp_desc(beta2_tmp);
    MLUCnnlTensorDesc epsilon_tmp_desc(epsilon_tmp);
    MLUCnnl::Fill(ctx,
                  CNNL_POINTER_MODE_HOST,
                  &beta1,
                  beta1_tmp_desc.get(),
                  GetBasePtr(&beta1_tmp));
    MLUCnnl::Fill(ctx,
                  CNNL_POINTER_MODE_HOST,
                  &beta2,
                  beta2_tmp_desc.get(),
                  GetBasePtr(&beta2_tmp));
    MLUCnnl::Fill(ctx,
                  CNNL_POINTER_MODE_HOST,
                  &epsilon,
                  epsilon_tmp_desc.get(),
                  GetBasePtr(&epsilon_tmp));
    beta1_tensor = &beta1_tmp;
    beta2_tensor = &beta2_tmp;
    epsilon_tensor = &epsilon_tmp;

    // Loop to compute
    for (size_t i = 0; i < param_num; ++i) {
      VLOG(4) << "[MergedAdam] loop: " << i;
      param_outs[i]->ShareDataWith(*params[i]);
      mom1_outs[i]->ShareDataWith(*mom1s[i]);
      mom2_outs[i]->ShareDataWith(*mom2s[i]);

      LoDTensor beta1_pow_tmp;
      LoDTensor beta2_pow_tmp;
      if (beta1_pows[i]->place() == platform::CPUPlace()) {
        T beta1 = *beta1_pows[i]->data<T>();
        beta1_pow_tmp.mutable_data<T>({1}, ctx.GetPlace());
        MLUCnnlTensorDesc beta1_pow_tmp_desc(beta1_pow_tmp);
        MLUCnnl::Fill(ctx,
                      CNNL_POINTER_MODE_HOST,
                      &beta1,
                      beta1_pow_tmp_desc.get(),
                      GetBasePtr(&beta1_pow_tmp));
        beta1_pows[i] = &beta1_pow_tmp;
      }
      if (beta2_pows[i]->place() == platform::CPUPlace()) {
        T beta2 = *beta2_pows[i]->data<T>();
        beta2_pow_tmp.mutable_data<T>({1}, ctx.GetPlace());
        MLUCnnlTensorDesc beta2_pow_tmp_desc(beta2_pow_tmp);
        MLUCnnl::Fill(ctx,
                      CNNL_POINTER_MODE_HOST,
                      &beta2,
                      beta2_pow_tmp_desc.get(),
                      GetBasePtr(&beta2_pow_tmp));
        beta2_pows[i] = &beta2_pow_tmp;
      }

      VLOG(3) << "beta1_pow.numel() : " << beta1_pows[i]->numel()
              << "beta2_pow.numel() : " << beta2_pows[i]->numel();
      VLOG(3) << "param.numel(): " << params[i]->numel();
      PADDLE_ENFORCE_EQ(beta1_pow_outs[i]->numel(),
                        1,
                        platform::errors::InvalidArgument(
                            "beta1 pow output size should be 1, but received "
                            "value is:%d.",
                            beta1_pow_outs[i]->numel()));

      PADDLE_ENFORCE_EQ(beta2_pow_outs[i]->numel(),
                        1,
                        platform::errors::InvalidArgument(
                            "beta2 pow output size should be 1, but received "
                            "value is:%d.",
                            beta2_pow_outs[i]->numel()));
      MLUCnnlTensorDesc param_desc(*params[i]);
      MLUCnnlTensorDesc mom1_desc(*mom1s[i]);
      MLUCnnlTensorDesc mom2_desc(*mom2s[i]);
      MLUCnnlTensorDesc grad_desc(*grads[i]);
      MLUCnnl::ApplyAdam(ctx,
                         param_desc.get(),
                         GetBasePtr(param_outs[i]),
                         mom1_desc.get(),
                         GetBasePtr(mom1_outs[i]),
                         mom2_desc.get(),
                         GetBasePtr(mom2_outs[i]),
                         grad_desc.get(),
                         GetBasePtr(grads[i]),
                         GetBasePtr(lrs[i]),
                         GetBasePtr(beta1_tensor),
                         GetBasePtr(beta2_tensor),
                         GetBasePtr(beta1_pows[i]),
                         GetBasePtr(beta2_pows[i]),
                         GetBasePtr(epsilon_tensor),
                         /*use_nesterov*/ false);
      if (!use_global_beta_pow) {
        beta1_pow_outs[i]->mutable_data<T>(ctx.GetPlace());
        beta2_pow_outs[i]->mutable_data<T>(ctx.GetPlace());

        MLUCnnlTensorDesc beta1_desc(*beta1_tensor);
        MLUCnnlOpTensorDesc mul_op_desc(
            CNNL_OP_TENSOR_MUL, ToCnnlDataType<T>(), CNNL_NOT_PROPAGATE_NAN);

        MLUCnnl::OpTensor(ctx,
                          mul_op_desc.get(),
                          beta1_desc.get(),
                          GetBasePtr(beta1_pows[i]),
                          beta1_desc.get(),
                          GetBasePtr(beta1_tensor),
                          beta1_desc.get(),
                          GetBasePtr(beta1_pow_outs[i]),
                          ToCnnlDataType<T>());

        MLUCnnl::OpTensor(ctx,
                          mul_op_desc.get(),
                          beta1_desc.get(),
                          GetBasePtr(beta2_pows[i]),
                          beta1_desc.get(),
                          GetBasePtr(beta2_tensor),
                          beta1_desc.get(),
                          GetBasePtr(beta2_pow_outs[i]),
                          ToCnnlDataType<T>());
      }
    }
  }
};
554 555 556 557 558 559
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
namespace plat = paddle::platform;

560 561
REGISTER_OP_MLU_KERNEL(adam,
                       ops::AdamMLUKernel<float>,
562 563
                       ops::AdamMLUKernel<plat::float16>);

564 565
REGISTER_OP_MLU_KERNEL(adamw,
                       ops::AdamWMLUKernel<float>,
566
                       ops::AdamWMLUKernel<plat::float16>);
567 568 569 570

REGISTER_OP_MLU_KERNEL(merged_adam,
                       ops::MergedAdamMLUKernel<float>,
                       ops::MergedAdamMLUKernel<plat::float16>);