mlu_baseop.h 64.7 KB
Newer Older
F
fwenguang 已提交
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 30 31 32
/* Copyright (c) 2021 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. */

#pragma once
#include <cn_api.h>
#include <cnnl.h>
#include <concurrentqueue.h>

#include <string>
#include <vector>

#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/framework/tensor_util.h"
#include "paddle/fluid/framework/type_defs.h"
#include "paddle/fluid/platform/device/mlu/enforce.h"

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;
using DataLayout = framework::DataLayout;
33
using ExecutionContext = framework::ExecutionContext;
F
fwenguang 已提交
34
using DeviceContextPool = platform::DeviceContextPool;
35 36
using MLUDeviceContext = platform::MLUDeviceContext;

37 38 39 40 41 42 43
const std::map<std::string, cnnlReduceOp_t> MLUReduceOpMap = {
    {"reduce_all", CNNL_REDUCE_AND},  {"reduce_any", CNNL_REDUCE_OR},
    {"reduce_max", CNNL_REDUCE_MAX},  {"reduce_mean", CNNL_REDUCE_AVG},
    {"reduce_min", CNNL_REDUCE_MIN},  {"reduce_sum", CNNL_REDUCE_ADD},
    {"reduce_prod", CNNL_REDUCE_MUL},
};

44 45 46 47 48 49 50 51 52 53 54 55 56 57
const std::map<std::string, cnnlInterpMode_t> MLUInterpModeMap = {
    {"bilinear", CNNL_INTERP_BILINEAR},
    {"nearest", CNNL_INTERP_NEAREST},
    {"linear", CNNL_INTERP_LINEAR},
    {"trilinear", CNNL_INTERP_TRILINEAR},
    {"bicubic", CNNL_INTERP_BICUBIC}};

const std::map<std::string, cnnlInterpBackwardMode_t> MLUInterpBackwardModeMap =
    {{"bilinear", CNNL_INTERP_BACKWARD_BILINEAR},
     {"nearest", CNNL_INTERP_BACKWARD_NEAREST},
     {"linear", CNNL_INTERP_BACKWARD_LINEAR},
     {"trilinear", CNNL_INTERP_BACKWARD_TRILINEAR},
     {"bicubic", CNNL_INTERP_BACKWARD_BICUBIC}};

58 59 60 61 62 63 64 65 66
inline cnnlReduceOp_t GetMLUCnnlReduceOp(const std::string reduce_name) {
  auto iter = MLUReduceOpMap.find(reduce_name);
  if (iter != MLUReduceOpMap.end()) {
    return iter->second;
  }
  PADDLE_THROW(platform::errors::InvalidArgument(
      "Not support reduce op type of MLU Device: %s", reduce_name));
}

67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85
inline cnnlInterpMode_t GetMLUCnnlInterpMode(const std::string interp_mode) {
  auto iter = MLUInterpModeMap.find(interp_mode);
  if (iter != MLUInterpModeMap.end()) {
    return iter->second;
  }
  PADDLE_THROW(platform::errors::InvalidArgument(
      "Not support interp mode of MLU Device: %s", interp_mode));
}

inline cnnlInterpBackwardMode_t GetMLUCnnlInterpBackwardMode(
    const std::string interp_mode) {
  auto iter = MLUInterpBackwardModeMap.find(interp_mode);
  if (iter != MLUInterpBackwardModeMap.end()) {
    return iter->second;
  }
  PADDLE_THROW(platform::errors::InvalidArgument(
      "Not support interp mode of MLU Device: %s", interp_mode));
}

86 87 88 89
inline const void* GetBasePtr(const Tensor* t) { return t->data(); }

inline void* GetBasePtr(Tensor* t) { return t->data(); }

90 91
inline cnnlDataType_t ToCnnlDataType(
    const paddle::experimental::DataType& dtype) {
F
fwenguang 已提交
92
  cnnlDataType_t type = CNNL_DTYPE_FLOAT;
93 94
  switch (dtype) {
    case DataType::FLOAT16:
F
fwenguang 已提交
95 96
      type = CNNL_DTYPE_HALF;
      break;
97
    case DataType::FLOAT32:
F
fwenguang 已提交
98 99
      type = CNNL_DTYPE_FLOAT;
      break;
Q
qipengh 已提交
100 101 102
    case DataType::FLOAT64:
      type = CNNL_DTYPE_DOUBLE;
      break;
103
    case DataType::INT8:
F
fwenguang 已提交
104 105
      type = CNNL_DTYPE_INT8;
      break;
106
    case DataType::INT16:
107 108
      type = CNNL_DTYPE_INT16;
      break;
109
    case DataType::INT32:
F
fwenguang 已提交
110 111
      type = CNNL_DTYPE_INT32;
      break;
112
    case DataType::INT64:
F
fwenguang 已提交
113 114
      type = CNNL_DTYPE_INT64;
      break;
115
    case DataType::BOOL:
F
fwenguang 已提交
116 117
      type = CNNL_DTYPE_BOOL;
      break;
118
    case DataType::UINT8:
119 120
      type = CNNL_DTYPE_UINT8;
      break;
F
fwenguang 已提交
121 122 123 124 125 126
    default:
      break;
  }
  return type;
}

127 128
inline cnnlDataType_t ToCnnlDataType(
    const paddle::framework::proto::VarType::Type& type) {
129
  return ToCnnlDataType(framework::TransToPhiDataType(type));
130 131 132 133 134 135 136 137
}

template <typename T>
inline cnnlDataType_t ToCnnlDataType() {
  auto type = framework::ToDataType(std::type_index(typeid(T)));
  return ToCnnlDataType(type);
}

F
fwenguang 已提交
138 139 140 141 142 143 144 145 146 147
// Converts (via narrowing) a type T value to a type U, and checks that the
// value has no value change due to the conversion.
template <typename WideT, typename NarrowT>
NarrowT CheckedNarrowing(const WideT& wide) {
  NarrowT narrow = wide;
  CHECK_EQ(narrow, wide)
      << "checked narrowing failed; values not equal post-conversion";
  return narrow;
}

148
inline static cnnlHandle_t GetHandleFromCTX(const ExecutionContext& ctx) {
149 150 151
  return ctx.template device_context<MLUDeviceContext>().cnnl_handle();
}

152 153
inline static const MLUDeviceContext& GetDevCtxFromCTX(
    const ExecutionContext& ctx) {
154 155 156
  return ctx.template device_context<MLUDeviceContext>();
}

157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174
using VT = framework::proto::VarType;
const std::map<std::pair<VT::Type, VT::Type>, cnnlCastDataType_t>
    MLU_SUPPORTED_CAST_TYPE = {
        {{VT::FP32, /*cast to*/ VT::FP16}, CNNL_CAST_FLOAT_TO_HALF},
        {{VT::FP32, /*cast to*/ VT::INT32}, CNNL_CAST_FLOAT_TO_INT32},
        {{VT::FP32, /*cast to*/ VT::INT16}, CNNL_CAST_FLOAT_TO_INT16},
        {{VT::FP32, /*cast to*/ VT::INT8}, CNNL_CAST_FLOAT_TO_INT8},
        {{VT::FP32, /*cast to*/ VT::UINT8}, CNNL_CAST_FLOAT_TO_UINT8},
        {{VT::FP32, /*cast to*/ VT::BOOL}, CNNL_CAST_FLOAT_TO_BOOL},
        {{VT::FP16, /*cast to*/ VT::FP32}, CNNL_CAST_HALF_TO_FLOAT},
        {{VT::FP16, /*cast to*/ VT::INT32}, CNNL_CAST_HALF_TO_INT32},
        {{VT::FP16, /*cast to*/ VT::INT16}, CNNL_CAST_HALF_TO_INT16},
        {{VT::FP16, /*cast to*/ VT::INT8}, CNNL_CAST_HALF_TO_INT8},
        {{VT::FP16, /*cast to*/ VT::UINT8}, CNNL_CAST_HALF_TO_UINT8},
        {{VT::FP16, /*cast to*/ VT::BOOL}, CNNL_CAST_HALF_TO_BOOL},
        {{VT::INT32, /*cast to*/ VT::FP32}, CNNL_CAST_INT32_TO_FLOAT},
        {{VT::INT32, /*cast to*/ VT::FP16}, CNNL_CAST_INT32_TO_HALF},
        {{VT::INT32, /*cast to*/ VT::INT8}, CNNL_CAST_INT32_TO_INT8},
175
        {{VT::INT32, /*cast to*/ VT::INT16}, CNNL_CAST_INT32_TO_INT16},
176 177 178 179 180 181 182 183 184 185 186
        {{VT::INT16, /*cast to*/ VT::FP32}, CNNL_CAST_INT16_TO_FLOAT},
        {{VT::INT16, /*cast to*/ VT::FP16}, CNNL_CAST_INT16_TO_HALF},
        {{VT::INT16, /*cast to*/ VT::INT32}, CNNL_CAST_INT16_TO_INT32},
        {{VT::INT8, /*cast to*/ VT::FP32}, CNNL_CAST_INT8_TO_FLOAT},
        {{VT::INT8, /*cast to*/ VT::FP16}, CNNL_CAST_INT8_TO_HALF},
        {{VT::INT8, /*cast to*/ VT::INT32}, CNNL_CAST_INT8_TO_INT32},
        {{VT::UINT8, /*cast to*/ VT::FP32}, CNNL_CAST_UINT8_TO_FLOAT},
        {{VT::UINT8, /*cast to*/ VT::FP16}, CNNL_CAST_UINT8_TO_HALF},
        {{VT::BOOL, /*cast to*/ VT::FP32}, CNNL_CAST_BOOL_TO_FLOAT},
        {{VT::BOOL, /*cast to*/ VT::FP16}, CNNL_CAST_BOOL_TO_HALF},
        {{VT::BOOL, /*cast to*/ VT::INT32}, CNNL_CAST_BOOL_TO_INT32},
187 188
        {{VT::UINT8, /*cast to*/ VT::INT32}, CNNL_CAST_UINT8_TO_INT32},
        {{VT::INT32, /*cast to*/ VT::INT64}, CNNL_CAST_INT32_TO_INT64},
189
        {{VT::INT64, /*cast to*/ VT::INT32}, CNNL_CAST_INT64_TO_INT32},
190 191 192 193 194 195 196 197 198
        {{VT::INT32, /*cast to*/ VT::BOOL}, CNNL_CAST_INT32_TO_BOOL},
        {{VT::UINT8, /*cast to*/ VT::INT64}, CNNL_CAST_UINT8_TO_INT64},
        {{VT::INT8, /*cast to*/ VT::INT16}, CNNL_CAST_INT8_TO_INT16},
        {{VT::FP32, /*cast to*/ VT::FP64}, CNNL_CAST_FLOAT_TO_DOUBLE},
        {{VT::FP64, /*cast to*/ VT::FP32}, CNNL_CAST_DOUBLE_TO_FLOAT},
        {{VT::INT64, /*cast to*/ VT::FP32}, CNNL_CAST_INT64_TO_FLOAT},
        {{VT::INT64, /*cast to*/ VT::FP16}, CNNL_CAST_INT64_TO_HALF},
        {{VT::FP32, /*cast to*/ VT::INT64}, CNNL_CAST_FLOAT_TO_INT64},
        {{VT::FP16, /*cast to*/ VT::INT64}, CNNL_CAST_HALF_TO_INT64},
199 200 201 202
};

cnnlCastDataType_t GetCastDataType(const VT::Type& src_type,
                                   const VT::Type& dst_type);
203 204 205 206

cnnlCastDataType_t GetCastDataType(const DataType& src_type,
                                   const DataType& dst_type);

207 208
bool MLUSupportsCast(const VT::Type& src_type, const VT::Type& dst_type);

F
fwenguang 已提交
209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250
cnnlDeviceType_t GetCnnlDev(int dev_ordinal);

using CnnlTensorDesc = cnnlTensorDescriptor_t;

class MLUCnnlTensorDesc {
 public:
  MLUCnnlTensorDesc() {}

  // SE_DISALLOW_COPY_AND_ASSIGN
  MLUCnnlTensorDesc(const MLUCnnlTensorDesc& desc) = delete;
  MLUCnnlTensorDesc& operator=(const MLUCnnlTensorDesc&) = delete;

  MLUCnnlTensorDesc(MLUCnnlTensorDesc&& rhs)
      : raw_tensor_desc(rhs.raw_tensor_desc) {
    rhs.raw_tensor_desc = nullptr;
  }

  MLUCnnlTensorDesc& operator=(MLUCnnlTensorDesc&& rhs);

  MLUCnnlTensorDesc(const int tensor_dim, const int dim_sizes[],
                    const cnnlDataType_t tensor_dtype);

  MLUCnnlTensorDesc(const int tensor_dim, const int dim_sizes[],
                    const cnnlDataType_t tensor_dtype,
                    const cnnlTensorLayout_t layout);

  MLUCnnlTensorDesc(const int tensor_dim, const int dim_sizes[],
                    const cnnlDataType_t tensor_dtype, int position);

  MLUCnnlTensorDesc(const int tensor_dim, const int64_t dim_sizes[],
                    const cnnlDataType_t tensor_dtype);

  MLUCnnlTensorDesc(const int tensor_dim, const int64_t dim_sizes[],
                    const cnnlDataType_t tensor_dtype,
                    const cnnlTensorLayout_t layout);

  MLUCnnlTensorDesc(const int tensor_dim, const int64_t dim_sizes[],
                    const cnnlDataType_t tensor_dtype, int position);

  MLUCnnlTensorDesc(const Tensor& tensor, const cnnlTensorLayout_t layout,
                    const cnnlDataType_t tensor_dtype);

251 252
  explicit MLUCnnlTensorDesc(const Tensor& tensor);

F
fwenguang 已提交
253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272
  MLUCnnlTensorDesc(const Tensor& tensor, cnnlTensorLayout_t layout,
                    const cnnlDataType_t tensor_dtype, int position);

  MLUCnnlTensorDesc(const Tensor& tensor, cnnlTensorLayout_t layout,
                    const cnnlDataType_t tensor_dtype, int position,
                    float scale);

  ~MLUCnnlTensorDesc();

  const cnnlTensorDescriptor_t get() const { return raw_tensor_desc; }

 private:
  cnnlTensorDescriptor_t raw_tensor_desc = nullptr;
};

class MLUCnnlActivationDesc {
 public:
  MLUCnnlActivationDesc(const MLUCnnlActivationDesc& desc) = delete;
  MLUCnnlActivationDesc& operator=(const MLUCnnlActivationDesc& desc) = delete;
  MLUCnnlActivationDesc(const cnnlActivationMode_t act_mode, const float ceof);
273 274 275
  MLUCnnlActivationDesc(const cnnlActivationMode_t act_mode, const float ceof,
                        const float sliced_dim, const float selu_alpha,
                        const float selu_lambda);
F
fwenguang 已提交
276 277 278 279 280 281 282 283

  const cnnlActivationDescriptor_t get() const;
  ~MLUCnnlActivationDesc();

 private:
  cnnlActivationDescriptor_t active_desc_ = nullptr;
};

284 285 286 287 288 289 290 291 292
class MLUCnnlPoolingDesc {
 public:
  MLUCnnlPoolingDesc(const MLUCnnlPoolingDesc& desc) = delete;
  MLUCnnlPoolingDesc& operator=(const MLUCnnlPoolingDesc& desc) = delete;

  MLUCnnlPoolingDesc(const cnnlPoolingMode_t mode,
                     const cnnlNanPropagation_t maxpooling_nan_opt,
                     int window_rows, int window_cols, int64_t pad_up,
                     int64_t pad_down, int64_t pad_left, int64_t pad_right,
293 294
                     int row_stride, int col_stride, int row_dilation,
                     int col_dilation, bool ceil_mode);
295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311

  MLUCnnlPoolingDesc(const cnnlPoolingMode_t mode,
                     const cnnlNanPropagation_t maxpooling_nan_opt,
                     const int tensor_rank, const std::vector<int>& window,
                     const std::vector<int>& padding,
                     const std::vector<int>& stride);

  const cnnlPoolingDescriptor_t get() const;

  ~MLUCnnlPoolingDesc();

 private:
  cnnlPoolingDescriptor_t pooling_desc_ = nullptr;
};

class MLUCnnlRandomGeneratorDesc {
 public:
Q
qipengh 已提交
312
  MLUCnnlRandomGeneratorDesc(const ExecutionContext& ctx, const int seed);
313
  const cnnlRandGenerator_t get() const;
Q
qipengh 已提交
314
  Tensor& get_state();
315 316 317
  ~MLUCnnlRandomGeneratorDesc();

 private:
Q
qipengh 已提交
318
  Tensor mlu_state;
319 320 321
  cnnlRandGenerator_t mlu_generator = nullptr;
};

Q
qipengh 已提交
322 323 324
const std::shared_ptr<MLUCnnlRandomGeneratorDesc>& GetMLURandomGenerator(
    const ExecutionContext& ctx, const int64_t device_id, const int seed);

325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446
class MLUCnnlReduceDesc {
 public:
  MLUCnnlReduceDesc(const MLUCnnlReduceDesc& desc) = delete;
  MLUCnnlReduceDesc& operator=(const MLUCnnlReduceDesc& desc) = delete;

  MLUCnnlReduceDesc(const std::vector<int>& axis_vec,
                    const cnnlReduceOp_t reduce_op,
                    const cnnlDataType_t data_type,
                    const cnnlNanPropagation_t nan_propagation,
                    const cnnlReduceIndices_t reduce_indices,
                    const cnnlIndicesType_t indices_type);

  const cnnlReduceDescriptor_t get() const;

  ~MLUCnnlReduceDesc();

 private:
  cnnlReduceDescriptor_t reduction_desc_ = nullptr;
};

class MLUCnnlOpTensorDesc {
 public:
  MLUCnnlOpTensorDesc(const MLUCnnlOpTensorDesc& desc) = delete;
  void operator=(const MLUCnnlOpTensorDesc&) = delete;

  MLUCnnlOpTensorDesc(cnnlOpTensorDesc_t op_tensor_op,
                      cnnlDataType_t op_tensor_comp_type,
                      cnnlNanPropagation_t op_tensor_nan_opt);

  const cnnlOpTensorDescriptor_t get() const;

  ~MLUCnnlOpTensorDesc();

 private:
  cnnlOpTensorDescriptor_t op_tensor_desc_ = nullptr;
};

class MLUCnnlNMSDesc {
 public:
  MLUCnnlNMSDesc(const MLUCnnlNMSDesc& desc) = delete;
  MLUCnnlNMSDesc& operator=(const MLUCnnlNMSDesc& desc) = delete;

  MLUCnnlNMSDesc(const cnnlNmsOutputMode_t mode, const float iou_threshold,
                 const int max_output_size, const float confidence_threshold,
                 const int input_layout);

  const cnnlNmsDescriptor_t get() const;

  ~MLUCnnlNMSDesc();

 private:
  cnnlNmsDescriptor_t nms_desc_ = nullptr;
};

class MLUCnnlConvolutionDesc {
 public:
  MLUCnnlConvolutionDesc(const int dims, const int pad[], const int stride[],
                         const int dilation[], const int group_count,
                         const cnnlDataType_t tensor_dtype);

  MLUCnnlConvolutionDesc(const int dims, const int64_t pad[],
                         const int64_t stride[], const int64_t dilation[],
                         const int group_count,
                         const cnnlDataType_t tensor_dtype);

  MLUCnnlConvolutionDesc(const MLUCnnlConvolutionDesc& desc) = delete;

  MLUCnnlConvolutionDesc& operator=(const MLUCnnlConvolutionDesc& desc) =
      delete;

  const cnnlConvolutionDescriptor_t get() const;

  ~MLUCnnlConvolutionDesc();

 private:
  cnnlConvolutionDescriptor_t conv_desc_ = nullptr;
};

class MLUCnnlBatchSpaceDesc {
 public:
  MLUCnnlBatchSpaceDesc(uint32_t block_shape[], uint32_t paddings[],
                        const uint32_t block_shape_size,
                        const uint32_t paddings_size);

  void getBatch2spaceNdextraInputSize(const ExecutionContext& ctx,
                                      const cnnlTensorDescriptor_t input_desc);

  void getSpace2batchNdextraInputSize(const ExecutionContext& ctx,
                                      const cnnlTensorDescriptor_t input_desc);

  void initSpace2batchNdExtraInput(const ExecutionContext& ctx,
                                   const cnnlTensorDescriptor_t input_desc,
                                   void* extra_host_input);

  void initBatch2spaceNdExtraInput(const ExecutionContext& ctx,
                                   const cnnlTensorDescriptor_t input_desc,
                                   void* extra_host_input);

  const cnnlSpaceBatchNdDescriptor_t get() const;

  size_t getExtraInputSize() const;

  ~MLUCnnlBatchSpaceDesc();

 private:
  cnnlSpaceBatchNdDescriptor_t op_desc_ = nullptr;
  size_t extra_input_size_;
};

class MLUCnnlTrigonDesc {
 public:
  explicit MLUCnnlTrigonDesc(
      const cnnlTrigonFunctionMode_t trigon_function_mode);

  const cnnlTrigonDescriptor_t get() const;

  ~MLUCnnlTrigonDesc();

 private:
  cnnlTrigonDescriptor_t trigon_desc_ = nullptr;
};

447 448 449 450 451 452 453 454 455 456 457 458 459
class MLUCnnlDCNDesc {
 public:
  MLUCnnlDCNDesc(int dimNb, const int* pad, const int* stride,
                 const int* dilation, int deformable_group, int conv_group,
                 int im2col_step);
  const cnnlDCNDescriptor_t get() const;

  ~MLUCnnlDCNDesc();

 private:
  cnnlDCNDescriptor_t dcn_desc_ = nullptr;
};

F
fwenguang 已提交
460 461
class MLUCnnl {
 public:
462
  static void Active(const ExecutionContext& ctx,
F
fwenguang 已提交
463 464 465 466
                     cnnlActivationDescriptor_t active_desc,
                     const cnnlTensorDescriptor_t input_desc, const void* input,
                     const cnnlTensorDescriptor_t output_desc, void* output);

467 468 469 470 471 472 473 474 475 476 477 478
  static void ActiveGrad(
      const ExecutionContext& ctx, cnnlActivationDescriptor_t active_desc,
      const void* alpha, const void* beta, const cnnlTensorDescriptor_t y_desc,
      const void* y, const cnnlTensorDescriptor_t diff_y_desc,
      const void* diff_y, const cnnlTensorDescriptor_t x_desc, const void* x,
      const cnnlTensorDescriptor_t diff_x_desc, void* diff_x);

  static void Concat(const ExecutionContext& ctx, const int pack_num,
                     const int axis, const cnnlTensorDescriptor_t inputs_desc[],
                     const void* const inputs[],
                     const cnnlTensorDescriptor_t output_desc, void* output);

Z
zn 已提交
479 480 481 482 483
  static void Concat(const MLUDeviceContext& dev_ctx, const int pack_num,
                     const int axis, const cnnlTensorDescriptor_t inputs_desc[],
                     const void* const inputs[],
                     const cnnlTensorDescriptor_t output_desc, void* output);

484 485 486 487
  static void Cast(const ExecutionContext& ctx, cnnlCastDataType_t cast_type,
                   const cnnlTensorDescriptor_t input_desc, const void* input,
                   const cnnlTensorDescriptor_t output_desc, void* output);

488 489 490 491 492 493 494 495 496 497
  static void Clip(const ExecutionContext& ctx,
                   const cnnlTensorDescriptor_t input_desc, const void* input,
                   const void* min, const void* max, void* y);

  static void HardtanhBackward(
      const ExecutionContext& ctx, const cnnlTensorDescriptor_t x_desc,
      const void* x, const cnnlTensorDescriptor_t diff_y_desc,
      const void* diff_y, const float max_val, const float min_val,
      const cnnlTensorDescriptor_t diff_x_desc, void* diff_x);

498 499 500 501 502 503
  static void Div(const ExecutionContext& ctx,
                  cnnlComputationPreference_t prefer,
                  const cnnlTensorDescriptor_t in0_desc, const void* in0,
                  const cnnlTensorDescriptor_t in1_desc, const void* in1,
                  const cnnlTensorDescriptor_t output_desc, void* output);

504 505
  static void Fill(const ExecutionContext& ctx,
                   const cnnlPointerMode_t pointer_mode, const void* value_ptr,
506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557
                   const cnnlTensorDescriptor_t output_desc, void* output);

  static void LRN(const ExecutionContext& ctx, const int local_size,
                  const double alpha, const double beta, const double k,
                  const cnnlTensorDescriptor_t input_quant_desc,
                  const void* input_quant,
                  const cnnlTensorDescriptor_t output_desc, void* output);

  static void QuantifyOffline(const ExecutionContext& context,
                              cnnlQuantizeMode_t mode,
                              const cnnlTensorDescriptor_t input_desc,
                              const void* input,
                              const cnnlTensorDescriptor_t ouput_desc,
                              void* output);

  static void QuantifyOnline(const ExecutionContext& context,
                             const int bitwidth,
                             const cnnlTensorDescriptor_t input_desc,
                             const void* input, const bool compute_scale,
                             void* position, void* scale,
                             const cnnlTensorDescriptor_t ouput_desc,
                             void* output);

  static void SGD(const ExecutionContext& context,
                  const cnnlTensorDescriptor_t grad_desc, const void* grad,
                  const void* lr, const cnnlTensorDescriptor_t var_desc,
                  void* var);

  static void ApplyAdaGrad(const ExecutionContext& ctx,
                           const cnnlTensorDescriptor_t grad_desc,
                           const void* grad,
                           const cnnlTensorDescriptor_t accum_desc, void* accum,
                           const cnnlTensorDescriptor_t var_desc, void* var,
                           const void* lr, const bool update_slots);

  static void ApplyRMSProp(const ExecutionContext& context,
                           const cnnlTensorDescriptor_t grad_desc,
                           const void* grad, const void* lr, const void* rho,
                           const void* momentum, const void* epsilon,
                           const cnnlTensorDescriptor_t var_desc, void* var,
                           const cnnlTensorDescriptor_t ms_desc, void* ms,
                           const cnnlTensorDescriptor_t mom_desc, void* mom);

  static void ApplyCenterRMSProp(
      const ExecutionContext& ctx, const cnnlTensorDescriptor_t grad_desc,
      const void* grad, const void* lr, const void* rho, const void* momentum,
      const void* epsilon, const cnnlTensorDescriptor_t var_desc, void* var,
      const cnnlTensorDescriptor_t mg_desc, void* mg,
      const cnnlTensorDescriptor_t ms_desc, void* ms,
      const cnnlTensorDescriptor_t mom_desc, void* mom);

  static void ApplyAdam(const ExecutionContext& ctx,
558 559 560
                        const cnnlTensorDescriptor_t var_desc, void* var,
                        const cnnlTensorDescriptor_t m_desc, void* m,
                        const cnnlTensorDescriptor_t v_desc, void* v,
561 562 563 564
                        const cnnlTensorDescriptor_t grad_desc,
                        const void* grad, const void* lr, const void* beta1,
                        const void* beta2, const void* beta1_power,
                        const void* beta2_power, const void* epsilon,
565
                        const bool use_nesterov);
566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603

  static void ApplyAdaMax(const ExecutionContext& ctx,
                          const cnnlTensorDescriptor_t grad_desc,
                          const cnnlTensorDescriptor_t var_desc, void* var,
                          const cnnlTensorDescriptor_t m_desc, void* m,
                          const cnnlTensorDescriptor_t v_desc, void* v,
                          const void* diff, const void* lr, const void* beta1,
                          const void* beta2, const void* beta1_power,
                          const void* epsilon);

  static void ApplyMomentum(const ExecutionContext& ctx,
                            const cnnlTensorDescriptor_t grad_desc,
                            const void* grad, const bool use_nesterov,
                            const void* lr, const void* momentum, void* var,
                            void* accum);

  static void ApplyKerasMomentum(const ExecutionContext& ctx,
                                 const cnnlTensorDescriptor_t grad_desc,
                                 const void* grad, const bool use_nesterov,
                                 const void* lr, const void* momentum,
                                 void* var, void* accum);

  static void ApplyAdadelta(const ExecutionContext& ctx,
                            const cnnlTensorDescriptor_t grad_desc,
                            const void* diff, const void* lr, const void* rho,
                            const void* epsilon, void* var, void* accum,
                            void* accum_update);

  static void SparseSoftmaxXentWithLogits(
      const ExecutionContext& ctx, cnnlSoftmaxMode_t mode,
      const cnnlTensorDescriptor_t x_desc, const void* input,
      const cnnlTensorDescriptor_t label_desc, const void* label,
      const cnnlTensorDescriptor_t y_desc, void* output,
      const cnnlTensorDescriptor_t diff_y_desc, void* back_out);

  static void RandomUniform(const ExecutionContext& ctx, const int num,
                            const cnnlDataType_t data_type,
                            const cnnlRandGenerator_t mlu_generator,
Q
qipengh 已提交
604 605 606 607 608 609 610
                            void* mlu_state, void* output);

  static void FusedDropout(
      const ExecutionContext& ctx, const cnnlRandGenerator_t generator,
      const cnnlTensorDescriptor_t input_desc, const void* input, const float p,
      void* state, const cnnlTensorDescriptor_t mask_desc, const void* mask,
      const cnnlTensorDescriptor_t output_desc, void* output);
611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663

  static void Cumsum(const ExecutionContext& ctx, const int axis,
                     const bool exclusive, const bool reverse,
                     const cnnlTensorDescriptor_t input_desc, const void* input,
                     const cnnlTensorDescriptor_t ouput_desc, void* output);

  static void BroadcastTo(const ExecutionContext& ctx,
                          const cnnlTensorDescriptor_t input_desc,
                          const void* input,
                          const cnnlTensorDescriptor_t output_desc,
                          void* output);

  static void GatherFunctor(
      const ExecutionContext& ctx, const int axis, const int batch_dims,
      const cnnlTensorDescriptor_t params_desc, const void* params,
      const cnnlTensorDescriptor_t indices_desc, const void* indices,
      const cnnlTensorDescriptor_t output_desc, void* output);

  static void ScatterFunctor(
      const ExecutionContext& ctx, const cnnlTensorDescriptor_t params_desc,
      const void* params, const cnnlTensorDescriptor_t updates_desc,
      const void* updates, const cnnlTensorDescriptor_t indices_desc,
      const void* indices, const cnnlScatterRefMode_t mode);

  static void Range(const ExecutionContext& ctx, const void* start,
                    const void* end, const void* step,
                    const cnnlDataType_t output_dtype, void* output);

  static void Round(const ExecutionContext& ctx,
                    const cnnlTensorDescriptor_t input_desc, const void* input,
                    const cnnlTensorDescriptor_t output_desc, void* output);

  static void TopK(const ExecutionContext& ctx, const int k, const int dim,
                   const bool largest, const bool sorted,
                   const cnnlTensorDescriptor_t input_desc, const void* input,
                   const cnnlTensorDescriptor_t values_output_desc,
                   void* values_out,
                   const cnnlTensorDescriptor_t indices_output_desc,
                   void* indices_out);

  static void StridedSlice(const ExecutionContext& ctx, const int begin[],
                           const int end[], const int strides[],
                           const cnnlTensorDescriptor_t input_desc,
                           const void* input,
                           const cnnlTensorDescriptor_t output_desc,
                           void* output);

  static void Split(const ExecutionContext& ctx, int split_num, int axis,
                    const cnnlTensorDescriptor_t input_desc,
                    const void* input_ptr,
                    const cnnlTensorDescriptor_t output_descs[],
                    void* output_ptrs[]);

Z
zn 已提交
664 665 666 667 668 669
  static void Split(const MLUDeviceContext& dev_ctx, int split_num, int axis,
                    const cnnlTensorDescriptor_t input_desc,
                    const void* input_ptr,
                    const cnnlTensorDescriptor_t output_descs[],
                    void* output_ptrs[]);

670 671 672 673 674 675 676 677 678 679 680 681
  static void Scale(const ExecutionContext& ctx, const int axis,
                    const cnnlTensorDescriptor_t input_desc, const void* input,
                    const cnnlTensorDescriptor_t alpha_desc, const void* alpha,
                    const cnnlTensorDescriptor_t beta_desc, const void* beta,
                    const cnnlTensorDescriptor_t output_desc, void* output);

  static void AddN(const ExecutionContext& ctx, uint32_t input_num,
                   const cnnlTensorDescriptor_t inputs_desc[],
                   const void* inputs[],
                   const cnnlTensorDescriptor_t output_desc, void* output);

  static void Log(const ExecutionContext& ctx,
682
                  cnnlComputationPreference_t prefer, cnnlLogBase_t log_base,
683 684 685 686 687 688 689 690 691 692
                  const cnnlTensorDescriptor_t input_desc, const void* input,
                  const cnnlTensorDescriptor_t output_desc, void* output);

  static void StridedSliceGrad(const ExecutionContext& ctx, const int begin[],
                               const int end[], const int strides[],
                               const cnnlTensorDescriptor_t input_desc,
                               const void* input,
                               const cnnlTensorDescriptor_t output_desc,
                               void* output);

F
fwenguang 已提交
693
  static void Logic(const ExecutionContext& ctx, const cnnlLogicOp_t log_method,
694 695 696 697 698 699
                    const cnnlTensorDescriptor_t input1_desc,
                    const void* input1,
                    const cnnlTensorDescriptor_t input2_desc,
                    const void* input2, const cnnlTensorDescriptor_t ouput_desc,
                    void* output);

700 701 702 703 704 705
  static void Select(
      const ExecutionContext& ctx, const cnnlTensorDescriptor_t condition_desc,
      const void* condition_ptr, const cnnlTensorDescriptor_t then_desc,
      const void* then_ptr, const cnnlTensorDescriptor_t else_desc,
      const void* else_ptr, const cnnlTensorDescriptor_t output_desc,
      void* output_ptr);
706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746

  static void AssignAdd(const ExecutionContext& ctx, const void* alpha,
                        const void* beta,
                        const cnnlTensorDescriptor_t update_desc,
                        const void* update,
                        const cnnlTensorDescriptor_t param_desc, void* param);

  static void AssignSub(const ExecutionContext& ctx, const void* alpha,
                        const void* beta,
                        const cnnlTensorDescriptor_t update_desc,
                        const void* update,
                        const cnnlTensorDescriptor_t param_desc, void* param);

  static void Assign(const ExecutionContext& ctx,
                     const cnnlTensorDescriptor_t update_desc,
                     const void* update,
                     const cnnlTensorDescriptor_t param_desc, void* param);

  static void GatherNd(const ExecutionContext& ctx,
                       const cnnlTensorDescriptor_t params_desc,
                       const void* params,
                       const cnnlTensorDescriptor_t indices_desc,
                       const void* indices,
                       const cnnlTensorDescriptor_t output_desc, void* output);

  static void BatchToSpace(const ExecutionContext& ctx,
                           const cnnlTensorDescriptor_t input_desc,
                           const void* input,
                           const cnnlTensorDescriptor_t output_desc,
                           void* output, const cnnlSpaceBatchParam_t param);

  static void BatchToSpaceNd(const ExecutionContext& ctx,
                             const cnnlTensorDescriptor_t input_desc,
                             const void* input,
                             cnnlSpaceBatchNdDescriptor_t param,
                             void* extra_device_input, size_t extra_input_size,
                             const cnnlTensorDescriptor_t output_desc,
                             void* output);

  static void PoolingForward(
      const ExecutionContext& ctx, cnnlPoolingMode_t pool_mode,
747 748 749
      int64_t output_h, int64_t output_w, cnnlPoolingDescriptor_t pooling_desc,
      const void* alpha, const cnnlTensorDescriptor_t input_desc,
      const void* input, const void* beta, const void* extra_input_ptr,
750 751
      const cnnlTensorDescriptor_t output_desc, void* output);

752 753 754 755 756 757
  static void AdaptivePoolingForward(
      const ExecutionContext& ctx, cnnlPoolingMode_t pool_mode,
      const cnnlTensorDescriptor_t input_desc, const void* input,
      const cnnlTensorDescriptor_t output_desc, void* output,
      const cnnlTensorDescriptor_t index_desc, void* index);

758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781
  static void Pool3D(const ExecutionContext& ctx, cnnlPoolingMode_t pool_mode,
                     const std::vector<int64_t>& output_shape,
                     cnnlPoolingDescriptor_t pooling_desc, const void* alpha,
                     const cnnlTensorDescriptor_t input_desc, const void* input,
                     const void* beta, const cnnlTensorDescriptor_t output_desc,
                     void* output);

  static void Pad(const ExecutionContext& ctx,
                  const cnnlTensorDescriptor_t input_desc, const void* input,
                  const void* paddings, const void* padding_value,
                  const cnnlTensorDescriptor_t output_desc, void* output);

  static void Matmul(const ExecutionContext& ctx, const bool transpose_a,
                     const bool transpose_b,
                     const cnnlTensorDescriptor_t in0_desc, const void* in0,
                     const cnnlTensorDescriptor_t in1_desc, const void* in1,
                     const cnnlTensorDescriptor_t output_desc, void* output);

  static void BatchMatmul(
      const ExecutionContext& ctx, const bool transpose_a,
      const bool transpose_b, const cnnlTensorDescriptor_t in0_desc,
      const void* in0, const cnnlTensorDescriptor_t in1_desc, const void* in1,
      const cnnlTensorDescriptor_t output_desc, void* output);

Q
qipengh 已提交
782 783 784 785
  static void MulAx(const ExecutionContext& ctx,
                    const cnnlTensorDescriptor_t alpha_desc, const void* alpha,
                    const cnnlTensorDescriptor_t output_desc, void* output);

786 787 788 789 790
  static void OpTensor(const ExecutionContext& ctx,
                       const cnnlOpTensorDescriptor_t op_tensor_desc,
                       const cnnlTensorDescriptor_t a_desc, const void* a,
                       const cnnlTensorDescriptor_t b_desc, const void* b,
                       const cnnlTensorDescriptor_t output_desc, void* output,
791 792 793 794
                       const cnnlDataType_t dtype,
                       const float alpha1_float = 1.f,
                       const float alpha2_float = 1.f,
                       const float beta_float = 0.f);
795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832

  static void BiasAddGrad(const ExecutionContext& ctx, const int axis,
                          const cnnlTensorDescriptor_t out_backprop_desc,
                          const void* out_backprop,
                          const cnnlTensorDescriptor_t output_desc,
                          void* output);

  static void OneHot(const ExecutionContext& ctx,
                     const cnnlTensorDescriptor_t desc_indices,
                     const void* indices, const int depth, const void* on_value,
                     const void* off_value, const int axis,
                     cnnlDataType_t output_data_type, void* output);

  static void NonMaxSuppression(const ExecutionContext& ctx,
                                const cnnlNmsDescriptor_t nms_desc,
                                const cnnlTensorDescriptor_t boxes_desc,
                                const void* boxes,
                                const cnnlTensorDescriptor_t confidence_desc,
                                const void* confidence,
                                const cnnlTensorDescriptor_t output_desc,
                                void* output, void* output_size);

  static void SoftmaxCrossEntropyWithLogits(
      const ExecutionContext& ctx, cnnlSoftmaxMode_t mode,
      cnnlComputationPreference_t prefer,
      const cnnlTensorDescriptor_t input_desc, const void* logits_in,
      const cnnlTensorDescriptor_t label_desc, const void* labels_in,
      const cnnlTensorDescriptor_t loss_out_desc, void* loss_out,
      const cnnlTensorDescriptor_t back_out_desc, void* back_out);

  static void SoftmaxForward(const ExecutionContext& ctx,
                             cnnlSoftmaxAlgorithm_t algorithm,
                             cnnlSoftmaxMode_t mode, const void* alpha,
                             const cnnlTensorDescriptor_t input_desc,
                             const void* input, const void* beta,
                             const cnnlTensorDescriptor_t output_desc,
                             void* output);

833 834 835 836 837 838 839
  static void SoftmaxBackward(
      const ExecutionContext& ctx, cnnlSoftmaxAlgorithm_t algorithm,
      cnnlSoftmaxMode_t mode, const cnnlTensorDescriptor_t y_desc,
      const void* y, const cnnlTensorDescriptor_t diff_y_desc,
      const void* diff_y, const cnnlTensorDescriptor_t diff_x_desc,
      void* diff_x);

840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077
  static void Softplus(const ExecutionContext& ctx,
                       const cnnlTensorDescriptor_t features_desc,
                       const void* features,
                       const cnnlTensorDescriptor_t output_desc, void* output);

  static void SoftplusGrad(const ExecutionContext& ctx,
                           const cnnlTensorDescriptor_t gradients_desc,
                           const void* gradients,
                           const cnnlTensorDescriptor_t features_desc,
                           const void* features,
                           const cnnlTensorDescriptor_t output_desc,
                           void* output);

  static void RsqrtGrad(const ExecutionContext& ctx,
                        const cnnlTensorDescriptor_t data_desc, const void* y,
                        const void* diff_y, void* output);

  static void SqrtGrad(const ExecutionContext& ctx,
                       const cnnlTensorDescriptor_t data_desc, const void* y,
                       const void* diff_y, void* output);

  static void ConvolutionForward(
      const ExecutionContext& ctx, cnnlConvolutionDescriptor_t conv_desc_,
      const void* alpha, const void* beta,
      const cnnlTensorDescriptor_t bias_desc, const void* bias_ptr,
      const cnnlTensorDescriptor_t input_desc, const void* input,
      const cnnlTensorDescriptor_t filtet_desc, const void* filter,
      const cnnlTensorDescriptor_t output_desc, void* output);

  static void FusedConvBNQuantify(
      const ExecutionContext& ctx, cnnlConvolutionDescriptor_t conv_desc,
      const void* epsilon_ptr, const int fused_ops_number,
      const cnnlDataType_t tensor_dtype, const int input_position,
      const float input_scale, const int filter_position,
      const float filter_scale, const cnnlTensorDescriptor_t scale_desc,
      const void* scale_ptr, const cnnlTensorDescriptor_t offset_desc,
      const void* offset_ptr, const cnnlTensorDescriptor_t mean_desc,
      const void* mean_ptr, const cnnlTensorDescriptor_t variance_desc,
      const void* variance_ptr, const cnnlTensorDescriptor_t input_desc,
      const void* input, const cnnlTensorDescriptor_t filtet_desc,
      const void* filter, const cnnlTensorDescriptor_t output_desc,
      void* output);

  static void Tile(const ExecutionContext& ctx,
                   const cnnlTensorDescriptor_t input_desc, const void* input,
                   const cnnlTensorDescriptor_t output_desc, void* output);

  static void UnsortedSegmentSum(const ExecutionContext& ctx,
                                 const cnnlTensorDescriptor_t data_desc,
                                 const void* data,
                                 const cnnlTensorDescriptor_t ids_desc,
                                 const int* segment_ids,
                                 const cnnlTensorDescriptor_t output_desc,
                                 void* output);

  static void Reduce(const ExecutionContext& ctx, const bool need_workspace,
                     const cnnlReduceDescriptor_t reduction_desc,
                     const void* alpha, const cnnlTensorDescriptor_t input_desc,
                     const void* input, const size_t indices_size,
                     void* indices, const void* beta,
                     const cnnlTensorDescriptor_t output_desc, void* output);

  static void FloorDiv(const ExecutionContext& ctx,
                       cnnlComputationPreference_t prefer,
                       const cnnlTensorDescriptor_t input1_desc,
                       const void* input1,
                       const cnnlTensorDescriptor_t input2_desc,
                       const void* input2,
                       const cnnlTensorDescriptor_t output_desc, void* output);

  static void FloorMod(const ExecutionContext& ctx,
                       const cnnlTensorDescriptor_t input1_desc,
                       const void* input1,
                       const cnnlTensorDescriptor_t input2_desc,
                       const void* input2,
                       const cnnlTensorDescriptor_t output_desc, void* output);

  static void Maximum(const ExecutionContext& ctx,
                      const cnnlTensorDescriptor_t input1_desc,
                      const void* input1,
                      const cnnlTensorDescriptor_t input2_desc,
                      const void* input2,
                      const cnnlTensorDescriptor_t output_desc, void* output);

  static void Minimum(const ExecutionContext& ctx,
                      const cnnlTensorDescriptor_t input1_desc,
                      const void* input1,
                      const cnnlTensorDescriptor_t input2_desc,
                      const void* input2,
                      const cnnlTensorDescriptor_t output_desc, void* output);

  static void PowR(const ExecutionContext& ctx,
                   cnnlComputationPreference_t prefer,
                   const cnnlTensorDescriptor_t input1_desc, const void* input1,
                   const cnnlTensorDescriptor_t input2_desc, const void* input2,
                   const cnnlTensorDescriptor_t output_desc, void* output);

  static void DivNoNan(const ExecutionContext& ctx,
                       cnnlComputationPreference_t prefer,
                       const cnnlTensorDescriptor_t input1_desc,
                       const void* input1,
                       const cnnlTensorDescriptor_t input2_desc,
                       const void* input2,
                       const cnnlTensorDescriptor_t output_desc, void* output);

  static void SquaredDifference(const ExecutionContext& ctx,
                                const cnnlTensorDescriptor_t input1_desc,
                                const void* input1,
                                const cnnlTensorDescriptor_t input2_desc,
                                const void* input2,
                                const cnnlTensorDescriptor_t output_desc,
                                void* output);

  static void L2Loss(const ExecutionContext& ctx,
                     const cnnlTensorDescriptor_t input_desc, const void* input,
                     void* output);

  static void Abs(const ExecutionContext& ctx,
                  const cnnlTensorDescriptor_t input_desc, const void* input,
                  const cnnlTensorDescriptor_t output_desc, void* output);

  static void Neg(const ExecutionContext& ctx,
                  const cnnlTensorDescriptor_t input_desc, const void* input,
                  const cnnlTensorDescriptor_t output_desc, void* output);

  static void Floor(const ExecutionContext& ctx,
                    const cnnlTensorDescriptor_t input_desc, const void* input,
                    const cnnlTensorDescriptor_t output_desc, void* output);

  static void Ceil(const ExecutionContext& ctx,
                   const cnnlTensorDescriptor_t input_desc, const void* input,
                   const cnnlTensorDescriptor_t output_desc, void* output);

  static void IsNan(const ExecutionContext& ctx,
                    const cnnlTensorDescriptor_t input_desc, const void* input,
                    const cnnlTensorDescriptor_t output_desc, void* output);

  static void Square(const ExecutionContext& ctx,
                     const cnnlTensorDescriptor_t input_desc, const void* input,
                     const cnnlTensorDescriptor_t output_desc, void* output);

  static void Sqrt(const ExecutionContext& ctx,
                   cnnlComputationPreference_t prefer,
                   const cnnlTensorDescriptor_t input_desc, const void* input,
                   const cnnlTensorDescriptor_t output_desc, void* output);

  static void Rsqrt(const ExecutionContext& ctx,
                    cnnlComputationPreference_t prefer,
                    const cnnlTensorDescriptor_t input_desc, const void* input,
                    const cnnlTensorDescriptor_t output_desc, void* output);

  static void Cos(const ExecutionContext& ctx,
                  cnnlComputationPreference_t prefer,
                  const cnnlTensorDescriptor_t input_desc, const void* input,
                  const cnnlTensorDescriptor_t output_desc, void* output);

  static void Sin(const ExecutionContext& ctx,
                  cnnlComputationPreference_t prefer,
                  const cnnlTensorDescriptor_t input_desc, const void* input,
                  const cnnlTensorDescriptor_t output_desc, void* output);

  static void TrigonForward(const ExecutionContext& ctx,
                            const cnnlTrigonDescriptor_t trigon_desc,
                            const cnnlTensorDescriptor_t input_desc,
                            const void* input,
                            const cnnlTensorDescriptor_t output_desc,
                            void* output);

  static void Exp(const ExecutionContext& ctx,
                  cnnlComputationPreference_t prefer,
                  const cnnlTensorDescriptor_t input_desc, const void* input,
                  const cnnlTensorDescriptor_t output_desc, void* output);

  static void Sign(const ExecutionContext& ctx,
                   const cnnlTensorDescriptor_t input_desc, const void* input,
                   const cnnlTensorDescriptor_t output_desc, void* output);

  static void IsFinite(const ExecutionContext& ctx,
                       const cnnlTensorDescriptor_t input_desc,
                       const void* input,
                       const cnnlTensorDescriptor_t output_desc, void* output);

  static void IsNanInf(const ExecutionContext& ctx,
                       const cnnlTensorDescriptor_t input_desc,
                       const void* input, void* output);

  static void Erf(const ExecutionContext& ctx,
                  cnnlComputationPreference_t prefer,
                  const cnnlTensorDescriptor_t input_desc, const void* input,
                  const cnnlTensorDescriptor_t output_desc, void* output);

  static void Log1p(const ExecutionContext& ctx,
                    cnnlComputationPreference_t prefer,
                    const cnnlTensorDescriptor_t input_desc, const void* input,
                    const cnnlTensorDescriptor_t output_desc, void* output);

  static void LogicalNot(const ExecutionContext& ctx,
                         const cnnlTensorDescriptor_t input_desc,
                         const void* input,
                         const cnnlTensorDescriptor_t output_desc,
                         void* output);

  static void DynamicStitch(
      const ExecutionContext& ctx, const cnnlTensorDescriptor_t* indices_desc,
      const int** indices, const cnnlTensorDescriptor_t* data_desc,
      const void** data, const int size, int* indices_dims,
      const cnnlTensorDescriptor_t output_desc, void* output);

  static void CropAndResize(
      const ExecutionContext& ctx, const std::string method_name,
      const float extrapolation_value, const cnnlTensorDescriptor_t image_desc,
      const void* image, const cnnlTensorDescriptor_t boxes_desc,
      const void* boxes, const cnnlTensorDescriptor_t box_index_desc,
      const void* box_index, const cnnlTensorDescriptor_t output_desc,
      void* output);

  static void CropAndResizeBackwardImage(
      const ExecutionContext& ctx, const std::string method_name,
      const cnnlTensorDescriptor_t image_desc, const void* image,
      const cnnlTensorDescriptor_t boxes_desc, const void* boxes,
      const cnnlTensorDescriptor_t box_idx_desc, const void* box_idx,
      const cnnlTensorDescriptor_t grads_image_desc, void* grads_image);

  static void CropAndResizeBackwardBoxes(
      const ExecutionContext& ctx, const cnnlTensorDescriptor_t input_desc,
      const void* input, const cnnlTensorDescriptor_t image_desc,
      const void* image, const cnnlTensorDescriptor_t boxes_desc,
      const void* boxes, const cnnlTensorDescriptor_t box_idx_desc,
      const void* box_idx, const cnnlTensorDescriptor_t output_desc,
      void* output);

  static void PoolingBackward(
      const ExecutionContext& ctx, const cnnlPoolingDescriptor_t pooling_desc,
      const void* alpha, const cnnlTensorDescriptor_t y_desc, const void* y,
      const cnnlTensorDescriptor_t diff_y_desc, const void* diff_y,
      const cnnlTensorDescriptor_t x_desc, const void* x, const void* beta,
      const cnnlTensorDescriptor_t diff_x_desc, void* diff_x);

1078 1079 1080 1081 1082 1083
  static void AdaptivePoolingBackward(
      const ExecutionContext& ctx, const cnnlPoolingMode_t pool_mode,
      const cnnlTensorDescriptor_t y_desc, const void* y,
      const cnnlTensorDescriptor_t index_desc, const void* index,
      const cnnlTensorDescriptor_t diff_x_desc, void* diff_x);

1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164
  static void PoolingIndex(const ExecutionContext& ctx,
                           const cnnlPoolingDescriptor_t pooling_desc,
                           const cnnlTensorDescriptor_t x_desc, const void* x,
                           const cnnlTensorDescriptor_t y_desc, void* y);

  static void SpaceToBatch(const ExecutionContext& ctx,
                           const cnnlTensorDescriptor_t input_desc,
                           const void* input,
                           const cnnlTensorDescriptor_t output_desc,
                           void* output, const int64_t block_shape[]);

  static void SpaceToBatchNd(const ExecutionContext& ctx,
                             const cnnlTensorDescriptor_t input_desc,
                             const void* input,
                             cnnlSpaceBatchNdDescriptor_t param,
                             void* extra_device_input, size_t extra_input_size,
                             const cnnlTensorDescriptor_t output_desc,
                             void* output);

  static void Interp(const ExecutionContext& ctx, const cnnlInterpMode_t mode,
                     const bool align_corners, const bool half_pixel_centers,
                     const cnnlTensorDescriptor_t input_desc, const void* input,
                     const cnnlTensorDescriptor_t output_desc, void* output);

  static void InterpBackward(
      const ExecutionContext& ctx, const cnnlInterpBackwardMode_t mode,
      const bool align_corners, const bool half_pixel_centers,
      const cnnlTensorDescriptor_t input_desc, const void* input,
      const cnnlTensorDescriptor_t output_desc, void* output);

  static void QuantizeParam(const ExecutionContext& ctx,
                            const cnnlQuantizeMode_t mode, const int bitwidth,
                            const cnnlTensorDescriptor_t input_desc,
                            const void* input, void* position, void* scale,
                            void* offset);

  static void QuantizeMatMul(
      const ExecutionContext& ctx, const bool transpose_a,
      const bool transpose_b, const cnnlTensorDescriptor_t a_desc,
      const void* a, const void* a_position, const void* a_scale,
      const void* a_offset, const cnnlTensorDescriptor_t b_desc, const void* b,
      const void* b_position, const void* b_scale, const void* b_offset,
      const cnnlDataType_t quant_type, const cnnlDataType_t data_type,
      const cnnlTensorDescriptor_t output_desc, void* output);

  static void QuantizeBatchMatMul(
      const ExecutionContext& ctx, const bool adj_x, const bool adj_y,
      const cnnlTensorDescriptor_t a_desc, const void* a,
      const void* a_position, const void* a_scale, const void* a_offset,
      const cnnlTensorDescriptor_t b_desc, const void* b,
      const void* b_position, const void* b_scale, const void* b_offset,
      const cnnlDataType_t quant_type, const cnnlDataType_t data_type,
      const cnnlTensorDescriptor_t output_desc, void* output);

  static void QuantizeBatchMatMulBCast(
      const ExecutionContext& ctx, const bool adj_x, const bool adj_y,
      const cnnlTensorDescriptor_t a_desc, const void* a,
      const void* a_position, const void* a_scale, const void* a_offset,
      const cnnlTensorDescriptor_t b_desc, const void* b,
      const void* b_position, const void* b_scale, const void* b_offset,
      const cnnlDataType_t quant_type, const cnnlDataType_t data_type,
      const cnnlTensorDescriptor_t output_desc, void* output);

  static void FusedBatchNorm(
      const ExecutionContext& ctx, const bool is_training,
      const cnnlTensorDescriptor_t x_desc, const void* x,
      const cnnlTensorDescriptor_t scale_desc, const void* scale,
      const void* offset, const void* estimated_mean,
      const void* estimated_variance, float epsilon, float momentum,
      const cnnlTensorDescriptor_t output_desc, void* output, void* batch_mean,
      void* batch_var, void* saved_mean, void* saved_var);

  static void FusedBatchNormGrad(
      const ExecutionContext& ctx, const bool is_training,
      const cnnlTensorDescriptor_t y_backprop_desc, const void* y_backprop,
      const cnnlTensorDescriptor_t x_desc, const void* x,
      const cnnlTensorDescriptor_t scale_desc, const void* scale,
      const void* saved_mean, const void* saved_var, float epsilon,
      const cnnlTensorDescriptor_t x_backprop_desc, void* x_backprop,
      void* scale_backprop, void* offset_backprop);

1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182
  static void LayerNormForward(const ExecutionContext& ctx, int axis,
                               const cnnlTensorDescriptor_t x_desc,
                               const void* x,
                               const cnnlTensorDescriptor_t weight_bias_desc,
                               const void* weight, const void* bias, float eps,
                               const cnnlTensorDescriptor_t y_desc, void* y,
                               const cnnlTensorDescriptor_t mean_rstd_desc,
                               void* saved_mean, void* saved_rstd);

  static void LayerNormBackward(
      const ExecutionContext& ctx, int axis,
      const cnnlTensorDescriptor_t x_desc, const void* x,
      const cnnlTensorDescriptor_t diff_z_desc, const void* diff_z,
      const cnnlTensorDescriptor_t weight_bias_desc, const void* weight,
      const cnnlTensorDescriptor_t mean_rstd_desc, const void* saved_mean,
      const void* saved_rstd, const cnnlTensorDescriptor_t diff_x_desc,
      void* diff_x, void* diff_weight, void* diff_bias);

1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218
  static void Transpose(const ExecutionContext& ctx,
                        const std::vector<int> perm, const int input_dim,
                        const cnnlTensorDescriptor_t input_desc,
                        const void* input,
                        const cnnlTensorDescriptor_t output_desc, void* output);

  static void MatrixBandPart(const ExecutionContext& ctx,
                             const cnnlTensorDescriptor_t data_desc,
                             const void* input, const int num_lower,
                             const int num_upper, void* output);

  static void NumTrue(const ExecutionContext& ctx,
                      const cnnlTensorDescriptor_t x_desc, const void* x,
                      Tensor index, uint32_t* num_true);

  static void Where(const ExecutionContext& ctx,
                    const cnnlTensorDescriptor_t x_desc, const void* x,
                    const uint32_t* strides, const uint32_t* index,
                    const cnnlTensorDescriptor_t y_desc, int* y,
                    const bool as_tuple);

  static void Conv2D(const ExecutionContext& ctx,
                     const cnnlConvolutionDescriptor_t conv_desc,
                     const cnnlDataType_t tensor_dtype,
                     const cnnlDataType_t dt_onchip, const void* input_position,
                     const void* input_scale, const void* input_offset,
                     const void* filter_position, const void* filter_scale,
                     const void* filter_offset,
                     const cnnlTensorDescriptor_t input_desc, const void* input,
                     const cnnlTensorDescriptor_t filter_desc,
                     const void* filter, const cnnlTensorDescriptor_t bias_desc,
                     const void* bias, const cnnlTensorDescriptor_t output_desc,
                     void* output);

  static void ConvBackpropInput(
      const ExecutionContext& ctx, const cnnlConvolutionDescriptor_t conv_desc,
1219
      const cnnlTensorDescriptor_t filter_desc, const void* filter,
1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248
      const cnnlTensorDescriptor_t out_backprop_desc, const void* out_backprop,
      const cnnlTensorDescriptor_t in_backprop_desc, void* in_backprop);

  static void QuantizeConvBackpropInput(
      const ExecutionContext& ctx, const cnnlConvolutionDescriptor_t conv_desc,
      const cnnlDataType_t tensor_dtype, const cnnlDataType_t dt_onchip,
      const void* filter_position, const void* filter_scale,
      const void* filter_offset, const void* out_backprop_position,
      const void* out_backprop_scale, const void* out_backprop_offset,
      const cnnlTensorDescriptor_t input_desc, const void* filter,
      const cnnlTensorDescriptor_t out_backprop_desc, const void* out_backprop,
      const cnnlTensorDescriptor_t in_backprop_desc, void* in_backprop);

  static void ConvBackpropFilter(
      const ExecutionContext& ctx, const cnnlConvolutionDescriptor_t conv_desc,
      const cnnlTensorDescriptor_t input_desc, const void* input,
      const cnnlTensorDescriptor_t out_backprop_desc, const void* out_backprop,
      const cnnlTensorDescriptor_t filter_backprop_desc, void* filter_backprop);

  static void QuantizeConvBackpropFilter(
      const ExecutionContext& ctx, const cnnlConvolutionDescriptor_t conv_desc,
      const cnnlDataType_t tensor_dtype, const cnnlDataType_t dt_onchip,
      const void* input_position, const void* input_scale,
      const void* input_offset, const void* out_backprop_position,
      const void* out_backprop_scale, const void* out_backprop_offset,
      const cnnlTensorDescriptor_t input_desc, const void* input,
      const cnnlTensorDescriptor_t out_backprop_desc, const void* out_backprop,
      const cnnlTensorDescriptor_t filter_backprop_desc, void* filter_backprop);

1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277
  static void DCNForward(
      const ExecutionContext& ctx, const cnnlDCNDescriptor_t dcn_desc,
      const cnnlTensorDescriptor_t input_desc, const void* input,
      const cnnlTensorDescriptor_t offset_desc, const void* offset,
      const cnnlTensorDescriptor_t mask_desc, const void* mask,
      const cnnlTensorDescriptor_t weight_desc, const void* weight,
      const cnnlTensorDescriptor_t bias_desc, const void* bias,
      const cnnlTensorDescriptor_t output_desc, void* output);

  static void DCNBackwardData(
      const ExecutionContext& ctx, const cnnlDCNDescriptor_t dcn_desc,
      const cnnlTensorDescriptor_t input_desc, const void* input,
      const cnnlTensorDescriptor_t offset_desc, const void* offset,
      const cnnlTensorDescriptor_t mask_desc, const void* mask,
      const cnnlTensorDescriptor_t weight_desc, const void* weight,
      const cnnlTensorDescriptor_t grad_output_desc, const void* grad_output,
      const cnnlTensorDescriptor_t grad_input_desc, void* grad_input,
      const cnnlTensorDescriptor_t grad_offset_desc, void* grad_offset,
      const cnnlTensorDescriptor_t grad_mask_desc, void* grad_mask);

  static void DCNBackwardWeight(
      const ExecutionContext& ctx, const cnnlDCNDescriptor_t dcn_desc,
      const cnnlTensorDescriptor_t input_desc, const void* input,
      const cnnlTensorDescriptor_t offset_desc, const void* offset,
      const cnnlTensorDescriptor_t mask_desc, const void* mask,
      const cnnlTensorDescriptor_t grad_output_desc, const void* grad_output,
      const cnnlTensorDescriptor_t grad_weight_desc, void* grad_weight,
      const cnnlTensorDescriptor_t grad_bias_desc, void* grad_bias);

1278 1279 1280 1281 1282 1283 1284 1285
  static void InTopK(const ExecutionContext& ctx,
                     const cnnlTensorDescriptor_t predictions_desc,
                     const void* predictions,
                     const cnnlTensorDescriptor_t targets_desc,
                     const void* targets, const cnnlTensorDescriptor_t k_desc,
                     const void* k, const int k_int,
                     const cnnlTensorDescriptor_t output_desc, void* output);

1286
  static void ScatterNd(const ExecutionContext& ctx, cnnlScatterNdMode_t mode,
1287 1288 1289 1290
                        const cnnlTensorDescriptor_t indices_desc,
                        const void* indices,
                        const cnnlTensorDescriptor_t updates_desc,
                        const void* updates,
1291 1292
                        const cnnlTensorDescriptor_t input_desc,
                        const void* input,
1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312
                        const cnnlTensorDescriptor_t output_desc, void* output);

  static void BitWise(const ExecutionContext& ctx,
                      const cnnlBitComputeOp_t optype,
                      const cnnlTensorDescriptor_t input1_desc,
                      const void* input1,
                      const cnnlTensorDescriptor_t input2_desc,
                      const void* input2,
                      const cnnlTensorDescriptor_t output_desc, void* output);

  static void QR(const ExecutionContext& ctx,
                 const cnnlTensorDescriptor_t a_desc, const void* a,
                 const cnnlTensorDescriptor_t q_desc, void* q,
                 const cnnlTensorDescriptor_t r_desc, void* r, const bool some);

  static void Reciprocal(const ExecutionContext& ctx,
                         const cnnlTensorDescriptor_t input_desc,
                         const void* input,
                         const cnnlTensorDescriptor_t output_desc,
                         void* output);
1313

1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328
  static void BceLoss(
      const ExecutionContext& ctx, const cnnlBceLossReduction_t reduction,
      const cnnlTensorDescriptor_t input_desc, const void* input,
      const cnnlTensorDescriptor_t target_desc, const void* target,
      const cnnlTensorDescriptor_t weight_desc, const void* weight,
      const cnnlTensorDescriptor_t output_desc, void* output);

  static void BceLossBackward(
      const ExecutionContext& ctx, const cnnlBceLossReduction_t reduction,
      const cnnlTensorDescriptor_t grad_desc, const void* grad,
      const cnnlTensorDescriptor_t input_desc, const void* input,
      const cnnlTensorDescriptor_t target_desc, const void* target,
      const cnnlTensorDescriptor_t weight_desc, const void* weight,
      const cnnlTensorDescriptor_t output_desc, void* output);

1329 1330 1331 1332 1333 1334
  static void EmbeddingForward(
      const ExecutionContext& ctx, const int padding_idx,
      const cnnlTensorDescriptor_t weight_desc, const void* weight,
      const cnnlTensorDescriptor_t indices_desc, const int* indices,
      const cnnlTensorDescriptor_t output_desc, void* output);

1335 1336 1337 1338 1339 1340
  static void Transform(const ExecutionContext& ctx, const void* alpha,
                        const void* beta,
                        const cnnlTensorDescriptor_t input_desc,
                        const void* input,
                        const cnnlTensorDescriptor_t output_desc, void* output);

1341 1342 1343 1344 1345
  static void EmbeddingBackward(
      const ExecutionContext& ctx, int padding_idx, bool scale_grad_by_freq,
      const cnnlTensorDescriptor_t indices_desc, const void* indices,
      const cnnlTensorDescriptor_t diff_desc, const void* diff,
      const cnnlTensorDescriptor_t output_desc, void* output);
F
fwenguang 已提交
1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362

  static void BceWithLogits(
      const ExecutionContext& ctx, cnnlBceWithLogitsReduction_t reduction,
      const cnnlTensorDescriptor_t input_desc, const void* input,
      const cnnlTensorDescriptor_t target_desc, const void* target,
      const cnnlTensorDescriptor_t weight_desc, const void* weight,
      const cnnlTensorDescriptor_t pos_weight_desc, const void* pos_weight,
      const cnnlTensorDescriptor_t output_desc, void* output);

  static void BceWithLogitsBackward(
      const ExecutionContext& ctx, cnnlBceWithLogitsReduction_t reduction,
      const cnnlTensorDescriptor_t grad_desc, const void* grad,
      const cnnlTensorDescriptor_t input_desc, const void* input,
      const cnnlTensorDescriptor_t target_desc, const void* target,
      const cnnlTensorDescriptor_t weight_desc, const void* weight,
      const cnnlTensorDescriptor_t pos_weight_desc, const void* pos_weight,
      const cnnlTensorDescriptor_t diff_input_desc, void* diff_input);
F
fwenguang 已提交
1363 1364
};

Q
qipengh 已提交
1365 1366 1367 1368 1369 1370
template <typename T>
inline void TransposeFromMLUTensor(const ExecutionContext& ctx,
                                   const std::vector<int> perm,
                                   const Tensor* transformed_input,
                                   Tensor* transformed_output,
                                   bool need_reshape_or_alloc) {
1371
  const int dim_size = perm.size();
Q
qipengh 已提交
1372
  if (need_reshape_or_alloc) {
1373 1374 1375 1376 1377
    std::vector<int> output_shape;
    auto input_dims = transformed_input->dims();
    for (int i = 0; i < dim_size; ++i) {
      output_shape.push_back(input_dims[perm[i]]);
    }
Q
qipengh 已提交
1378
    transformed_output->mutable_data<T>(
1379
        framework::DDim(output_shape.data(), dim_size), ctx.GetPlace());
Q
qipengh 已提交
1380 1381 1382 1383 1384 1385
  }
  MLUCnnlTensorDesc trans_in_desc(*transformed_input, CNNL_LAYOUT_ARRAY,
                                  ToCnnlDataType<T>());
  MLUCnnlTensorDesc trans_out_desc(*transformed_output, CNNL_LAYOUT_ARRAY,
                                   ToCnnlDataType<T>());

1386
  MLUCnnl::Transpose(ctx, perm, dim_size, trans_in_desc.get(),
Q
qipengh 已提交
1387 1388 1389 1390
                     GetBasePtr(transformed_input), trans_out_desc.get(),
                     GetBasePtr(transformed_output));
}

1391 1392 1393 1394 1395 1396 1397 1398
template <typename T>
inline void FillMLUTensorWithHostValue(const ExecutionContext& ctx, T value,
                                       Tensor* out) {
  MLUCnnlTensorDesc out_desc(*out);
  MLUCnnl::Fill(ctx, CNNL_POINTER_MODE_HOST, &value, out_desc.get(),
                GetBasePtr(out));
}

F
fwenguang 已提交
1399 1400
}  // namespace operators
}  // namespace paddle