unary.cc 167.0 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* 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. */

15
#include "paddle/phi/infermeta/unary.h"
16

L
Linjie Chen 已提交
17
#include <algorithm>
18
#include <set>
W
WJJ1995 已提交
19

20
#include "gflags/gflags.h"
21
#include "paddle/phi/common/data_type.h"
22
#include "paddle/phi/common/type_traits.h"
23
#include "paddle/phi/core/enforce.h"
24
#include "paddle/phi/core/infermeta_utils.h"
25
#include "paddle/phi/core/utils/data_type.h"
26
#include "paddle/phi/kernels/funcs/parse_qr_mode.h"
F
From00 已提交
27
#include "paddle/phi/kernels/funcs/pooling.h"
H
hong 已提交
28
#include "paddle/phi/kernels/funcs/slice_utils.h"
29
#include "paddle/phi/kernels/funcs/strided_slice.h"
30
#include "paddle/phi/kernels/funcs/unfold_functor.h"
31
#include "paddle/phi/kernels/funcs/unsqueeze.h"
32
#include "paddle/phi/kernels/impl/einsum_impl.h"
33

34
namespace phi {
35

36 37 38 39 40
namespace detail {
// Used in MatrixRankInferMeta
static DDim CheckAndGetOutputDim(const DDim& dim_x) {
  auto x_vec = phi::vectorize(dim_x);
  if (x_vec.size() == 2) {
41
    return phi::make_ddim({});
42 43 44 45 46 47
  }
  x_vec.erase(x_vec.end() - 2, x_vec.end());
  return phi::make_ddim(x_vec);
}
}  // namespace detail

48 49 50 51 52
void AffineGridInferMeta(const MetaTensor& input,
                         const IntArray& outputShape,
                         bool align_corners,
                         MetaTensor* output) {
  auto theta_dims = input.dims();
53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79
  bool is_from_tensor = outputShape.FromTensor();
  if (!is_from_tensor) {
    PADDLE_ENFORCE_EQ(
        theta_dims.size(),
        3,
        phi::errors::InvalidArgument(
            "The input Theta's dimensions size should be 3. But received "
            "Theta's demensions size=[%d],  Theta's dimensions=[%s].",
            theta_dims.size(),
            theta_dims));

    PADDLE_ENFORCE_GE(
        outputShape.GetData().size(),
        4,
        phi::errors::InvalidArgument(
            "The size of attribute 'output_shape' in AffineGridOp should be >= "
            "4. But received output_shape's size=[%d].",
            outputShape.GetData().size()));

    PADDLE_ENFORCE_LE(
        outputShape.GetData().size(),
        5,
        phi::errors::InvalidArgument(
            "The size of attribute 'output_shape' in AffineGridOp should be <= "
            "5. But received output_shape's size=[%d].",
            outputShape.GetData().size()));
  }
80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98
  PADDLE_ENFORCE_GE(theta_dims[1],
                    2,
                    phi::errors::InvalidArgument(
                        "The second dimesion of input 'theta' in AffineGridOp "
                        "should be >= 2. "
                        "But received second dimesion=[%d], dimesions=[%s]",
                        theta_dims[1],
                        theta_dims));

  PADDLE_ENFORCE_LE(theta_dims[1],
                    3,
                    phi::errors::InvalidArgument(
                        "The second dimesion of input 'theta' in AffineGridOp "
                        "should be <= 3. "
                        "But received second dimesion=[%d], dimesions=[%s]",
                        theta_dims[1],
                        theta_dims));

  PADDLE_ENFORCE_GE(
99 100 101
      theta_dims[2],
      3,
      phi::errors::InvalidArgument(
102
          "The third dimesion of input 'theta' in AffineGridOp should be >= 3. "
103 104 105 106
          "But received third dimesion=[%d], dimesions=[%s]",
          theta_dims[2],
          theta_dims));

107 108 109 110 111 112 113 114
  PADDLE_ENFORCE_LE(
      theta_dims[2],
      4,
      phi::errors::InvalidArgument(
          "The third dimesion of input 'theta' in AffineGridOp should be <= 4. "
          "But received third dimesion=[%d], dimesions=[%s]",
          theta_dims[2],
          theta_dims));
115
  if (outputShape.GetData().size() == 4 && !is_from_tensor) {
116 117 118 119 120 121
    // N * H * W * 2
    output->set_dims(phi::make_ddim({theta_dims[0], -1, -1, 2}));
  } else {
    // N * D * H * W * 3
    output->set_dims(phi::make_ddim({theta_dims[0], -1, -1, -1, 3}));
  }
122 123 124 125
  output->set_dtype(input.dtype());
  output->share_lod(input);
}

TaoTao Li's avatar
TaoTao Li 已提交
126 127 128 129 130 131 132 133
void AllGatherInferMeta(const MetaTensor& x, int nranks, MetaTensor* out) {
  auto dim = x.dims();
  dim[0] = dim[0] * nranks;
  if (dim[0] < 0) dim[0] = -1;
  out->set_dtype(x.dtype());
  out->set_dims(dim);
}

134 135 136 137 138
void AllReduceInferMeta(const MetaTensor& x, MetaTensor* out) {
  out->set_dtype(x.dtype());
  out->set_dims(x.dims());
}

Z
zyfncg 已提交
139
void ArgMinMaxInferMeta(const MetaTensor& x,
140
                        const Scalar& axis,
Z
zyfncg 已提交
141 142 143 144 145 146 147 148 149 150 151
                        bool keepdims,
                        bool flatten,
                        int dtype,
                        MetaTensor* out,
                        MetaConfig config) {
  PADDLE_ENFORCE_EQ(
      (dtype < 0 || dtype == 2 || dtype == 3),
      true,
      phi::errors::InvalidArgument(
          "The attribute of dtype in argmin/argmax must be [%s] or [%s], but "
          "received [%s]",
152 153 154
          DataTypeToString(DataType::INT32),
          DataTypeToString(DataType::INT64),
          DataTypeToString(phi::TransToPhiDataType(dtype))));
Z
zyfncg 已提交
155

156 157 158
  if (!config.is_runtime && axis.FromTensor()) {
    std::vector<int64_t> vec;
    if (flatten) {
159 160 161 162 163
      if (keepdims) {
        vec = std::vector<int64_t>(x.dims().size(), -1);
      } else {
        vec = {};
      }
164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182
    } else {
      if (keepdims) {
        vec = std::vector<int64_t>(x.dims().size(), -1);
      } else {
        vec = std::vector<int64_t>(x.dims().size() - 1, -1);
      }
    }
    out->set_dims(phi::make_ddim(vec));
    if (dtype == 2) {
      out->set_dtype(DataType::INT32);
    } else if (dtype == 3) {
      out->set_dtype(DataType::INT64);
    }
    return;
  }

  auto int_axis = axis.to<int64_t>();
  const auto& x_dims = x.dims();

183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200
  auto x_rank = x.dims().size();
  if (x_rank > 0) {
    PADDLE_ENFORCE_GE(int_axis,
                      -x_rank,
                      phi::errors::InvalidArgument(
                          "'axis'(%d) must be greater than or equal to"
                          " -Rank(X)(%d).",
                          int_axis,
                          -x_rank));
    PADDLE_ENFORCE_LT(
        int_axis,
        x_rank,
        phi::errors::InvalidArgument(
            "'axis'(%d) must be less than Rank(X)(%d) of Input(X).",
            int_axis,
            x_rank));
  } else {
    // 0-dim tensor
201
    PADDLE_ENFORCE_EQ(int_axis == 0 || int_axis == -1,
202 203 204
                      true,
                      phi::errors::InvalidArgument(
                          "'axis'(%d) must be 0 or -1 if input tensor is "
205
                          "0-dim.",
206 207
                          int_axis));
  }
208 209

  if (int_axis < 0) int_axis += x_rank;
210

Z
zyfncg 已提交
211
  if (config.is_runtime) {
212
    if (dtype == phi::TransToProtoVarType(DataType::INT32)) {
Z
zyfncg 已提交
213
      int64_t all_element_num = 0;
214
      if (flatten) {
Z
zyfncg 已提交
215 216
        all_element_num = phi::product(x_dims);
      } else {
217
        all_element_num = x_dims[int_axis];
Z
zyfncg 已提交
218 219 220 221 222 223 224 225 226 227 228 229
      }
      PADDLE_ENFORCE_LE(
          all_element_num,
          INT_MAX,
          phi::errors::InvalidArgument(
              "The element num of the argmin/argmax input at axis is "
              "%d, is larger than int32 maximum value:%d, you must "
              "set the dtype of argmin/argmax to 'int64'.",
              all_element_num,
              INT_MAX));
    }
  }
230

Z
zyfncg 已提交
231
  std::vector<int64_t> vec;
232 233 234 235 236 237
  if (flatten) {
    if (keepdims) {
      vec = std::vector<int64_t>(x.dims().size(), 1);
    } else {
      vec = {};
    }
Z
zyfncg 已提交
238
  } else {
239
    for (int64_t i = 0; i < int_axis; i++) vec.emplace_back(x_dims[i]);
Z
zyfncg 已提交
240 241 242
    if (keepdims) {
      vec.emplace_back(static_cast<int64_t>(1));
    }
243
    for (int64_t i = int_axis + 1; i < x_rank; i++) vec.emplace_back(x_dims[i]);
Z
zyfncg 已提交
244
  }
245

Z
zyfncg 已提交
246 247 248 249 250 251 252 253
  out->set_dims(phi::make_ddim(vec));
  if (dtype == 2) {
    out->set_dtype(DataType::INT32);
  } else if (dtype == 3) {
    out->set_dtype(DataType::INT64);
  }
}

L
Linjie Chen 已提交
254 255 256 257 258 259 260
void ArgsortInferMeta(const MetaTensor& input,
                      int axis,
                      bool descending,
                      MetaTensor* output,
                      MetaTensor* indices) {
  auto in_dims = input.dims();
  auto num_dims = in_dims.size();
261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280
  if (num_dims > 0) {
    PADDLE_ENFORCE_GE(axis,
                      -num_dims,
                      phi::errors::InvalidArgument(
                          "'axis'(%d) must be greater than or equal to"
                          " -num_dims(%d).",
                          axis,
                          -num_dims));
    PADDLE_ENFORCE_LT(
        axis,
        num_dims,
        phi::errors::InvalidArgument(
            "'axis'(%d) must be less than num_dims(%d).", axis, num_dims));
  } else {  // 0-dim tensor
    PADDLE_ENFORCE_EQ(
        axis == 0 || axis == -1,
        1,
        phi::errors::InvalidArgument(
            "'axis'(%d) must be 0 or -1 if input tensor is 0-dim.", axis));
  }
L
Linjie Chen 已提交
281 282 283 284 285 286 287

  output->share_dims(input);
  output->set_dtype(input.dtype());
  indices->share_dims(input);
  indices->set_dtype(DataType::INT64);
  output->share_lod(input);
  indices->share_lod(input);
288 289 290 291 292 293 294
}

void AsRealInferMeta(const MetaTensor& input, MetaTensor* output) {
  auto out_dims_v = phi::vectorize(input.dims());
  out_dims_v.push_back(2);
  auto out_dims = phi::make_ddim(out_dims_v);
  output->set_dims(out_dims);
295
  output->share_lod(input);
张春乔 已提交
296
  output->set_dtype(dtype::ToReal(input.dtype()));
297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322
}

void AsComplexInferMeta(const MetaTensor& input, MetaTensor* output) {
  auto in_dims = input.dims();
  const int input_rank = in_dims.size();
  PADDLE_ENFORCE_GE(
      input_rank,
      1,
      phi::errors::InvalidArgument(
          "The rank of input(X) is less than 1. "
          "Expected the rank of input(X) to be equal to or greater than 1."
          "But received rank of input(X) = %d",
          input_rank));
  const int last_dim_size = in_dims[input_rank - 1];
  PADDLE_ENFORCE_EQ(
      last_dim_size,
      2,
      phi::errors::InvalidArgument(
          "The size of the last dimension of input(X)"
          "does not equals 2."
          "Expected the size of last dimension of input(X) to be 2."
          "But received %d",
          last_dim_size));

  const phi::DDim out_dims(in_dims.Get(), input_rank - 1);
  output->set_dims(out_dims);
323
  output->share_lod(input);
L
Linjie Chen 已提交
324 325
}

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
void BatchSizeLikeInferMeta(const MetaTensor& x,
                            const std::vector<int>& shape,
                            int x_batch_size_dim,
                            int out_batch_size_dim,
                            MetaTensor* out) {
  PADDLE_ENFORCE_GT(
      shape.size(),
      0UL,
      phi::errors::InvalidArgument(
          "Shape size must be larger than 0, but received: %s.", shape.size()));
  std::vector<int64_t> shape_int64(shape.size(), 0);
  std::transform(shape.begin(), shape.end(), shape_int64.begin(), [](int a) {
    return static_cast<int64_t>(a);
  });
  auto output_dim = phi::make_ddim(shape_int64);

  int input_dim_size = static_cast<int>(x.dims().size());
  PADDLE_ENFORCE_GE(
      x_batch_size_dim,
      0,
      phi::errors::InvalidArgument("Input dimension index must be larger "
                                   "equal than 0, but received: %s.",
                                   x_batch_size_dim));
  PADDLE_ENFORCE_GT(input_dim_size,
                    x_batch_size_dim,
                    phi::errors::InvalidArgument(
                        "Input dimension size must be larger than "
                        "input dimension index, but received input "
                        "dimension size: %s, input dimension index: %s.",
                        input_dim_size,
                        x_batch_size_dim));

  int output_dim_size = static_cast<int>(shape.size());
  PADDLE_ENFORCE_GE(
      out_batch_size_dim,
      0,
      phi::errors::InvalidArgument("Output dimension index must be larger "
                                   "equal than 0, but received: %s.",
                                   out_batch_size_dim));
  PADDLE_ENFORCE_GT(
      output_dim_size,
      out_batch_size_dim,
      phi::errors::InvalidArgument(
          "Output dimension size must be larger than output dimension index, "
          "but received output dimension size: %s, output dimension index: "
          "%s.",
          output_dim_size,
          out_batch_size_dim));

  output_dim[out_batch_size_dim] = x.dims()[x_batch_size_dim];
  out->set_dims(output_dim);
}

379 380 381 382
void CastInferMeta(const MetaTensor& x, DataType out_dtype, MetaTensor* out) {
  out->set_dims(x.dims());
  out->set_dtype(out_dtype);
  out->set_layout(x.layout());
383
  out->share_lod(x);
384 385
}

386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407
void CholeskyInferMeta(const MetaTensor& x, bool upper, MetaTensor* out) {
  auto dims = x.dims();
  auto rank = dims.size();
  PADDLE_ENFORCE_GE(rank,
                    2,
                    errors::InvalidArgument(
                        "The Input(X) should have at least 2 dimensions. But "
                        "received a %d dimension tensor.",
                        rank));
  PADDLE_ENFORCE_EQ(
      dims[rank - 2],
      dims[rank - 1],
      errors::InvalidArgument(
          "The inner-most 2 dimensions of Input(X) all should be symmetric "
          "positive-definite matrices and have the same size. But received "
          "X's shape[-2] = %d and shape[-1] = %d.",
          dims[rank - 2],
          dims[rank - 1]));
  out->set_dims(x.dims());
  out->set_dtype(x.dtype());
}

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
void ClassCenterSampleInferMeta(const MetaTensor& label,
                                int num_classes,
                                int num_samples,
                                int ring_id,
                                int rank,
                                int nranks,
                                bool fix_seed,
                                int seed,
                                MetaTensor* remapped_label,
                                MetaTensor* sampled_local_class_center) {
  PADDLE_ENFORCE_EQ(
      label.dims().size(),
      1,
      errors::InvalidArgument("Rank of Input(Label) should be equal to 1, "
                              "but the value given is %d.",
                              label.dims().size()));
  PADDLE_ENFORCE_NOT_NULL(remapped_label,
                          phi::errors::InvalidArgument(
                              "output of remapped label should not be null."));
  PADDLE_ENFORCE_NOT_NULL(
      sampled_local_class_center,
      phi::errors::InvalidArgument(
          "output of sampled local class center should not be null."));
  remapped_label->set_dims(label.dims());
  remapped_label->set_dtype(label.dtype());
  sampled_local_class_center->set_dims(phi::make_ddim({num_samples}));
  sampled_local_class_center->set_dtype(label.dtype());
}

L
lyq 已提交
437 438 439 440 441 442 443 444 445 446 447 448
void ClipByNormInferMeta(const MetaTensor& x, float max_norm, MetaTensor* out) {
  PADDLE_ENFORCE_GT(
      max_norm,
      0,
      phi::errors::InvalidArgument("max_norm should be greater than 0. "
                                   "Received max_norm is %f.",
                                   max_norm));
  out->set_dims(x.dims());
  out->set_dtype(x.dtype());
  out->share_lod(x);
}

449
void CreateLikeInferMeta(const MetaTensor& x, DataType dtype, MetaTensor* out) {
450 451
  out->set_dims(x.dims());
  out->set_dtype(dtype == DataType::UNDEFINED ? x.dtype() : dtype);
452
  out->set_layout(x.layout());
453 454
}

455 456 457 458 459 460
void CumInferMeta(const MetaTensor& x,
                  int axis,
                  bool flatten,
                  bool exclusive,
                  bool reverse,
                  MetaTensor* out) {
461 462 463 464 465
  auto x_dims = x.dims();
  if (flatten) {
    out->set_dims(phi::make_ddim({phi::product(x_dims)}));
    out->set_dtype(x.dtype());
  } else {
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
    if (x_dims.size() > 0) {
      PADDLE_ENFORCE_GE(
          axis,
          -x_dims.size(),
          phi::errors::OutOfRange(
              "axis is out of range (expected to be in range of [%ld, "
              "%ld), but got %ld).",
              -(x_dims.size()),
              x_dims.size(),
              axis));
      PADDLE_ENFORCE_LT(
          axis,
          x_dims.size(),
          phi::errors::OutOfRange(
              "axis is out of range (expected to be in range of [%ld, "
              "%ld), but got %ld).",
              -(x_dims.size()),
              x_dims.size(),
              axis));
    } else {
      PADDLE_ENFORCE_EQ(
          (axis == 0 || axis == -1),
          true,
          errors::InvalidArgument("The axis must be -1 or 0 in 0D Tensor, "
                                  "but the value given is %d.",
                                  axis));
    }
493 494 495
    out->set_dims(x_dims);
    out->set_dtype(x.dtype());
  }
496

497 498 499
  out->share_lod(x);
}

W
WangZhen 已提交
500 501 502 503 504 505 506 507 508
void CumScalarAxisInferMeta(const MetaTensor& x,
                            const Scalar& axis,
                            bool flatten,
                            bool exclusive,
                            bool reverse,
                            MetaTensor* out) {
  CumInferMeta(x, axis.to<int>(), flatten, exclusive, reverse, out);
}

509 510 511 512 513
void CropInferMeta(const MetaTensor& x,
                   const IntArray& shape,
                   const IntArray& offsets,
                   MetaTensor* out,
                   MetaConfig config) {
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
  PADDLE_ENFORCE_NE(
      out,
      nullptr,
      errors::InvalidArgument("CropTensor should have output tensor out."));

  auto x_dim = x.dims();
  auto shape_dims = shape.GetData();
  auto offsets_vec = offsets.GetData();

  PADDLE_ENFORCE_EQ(shape_dims.size(),
                    x_dim.size(),
                    errors::InvalidArgument(
                        "The number of elements (%d) of attribute 'shape' for "
                        "CropTensor must be equal to the number of "
                        "dimensions (%d) of the input.",
                        shape_dims.size(),
                        x_dim.size()));

  if (config.is_runtime) {
    out->share_lod(x);
  }

  auto out_dims = std::vector<int64_t>(shape.size(), -1);
  for (size_t i = 0; i < shape_dims.size(); ++i) {
    if (shape_dims[i] > 0) {
      out_dims[i] = static_cast<int64_t>(shape_dims[i]);
    } else {
      if (shape_dims[i] == -1 && offsets_vec[i] != -1 && x_dim[i] != -1) {
        out_dims[i] = x_dim[i] - static_cast<int64_t>(offsets_vec[i]);
      }
    }
  }
  out->set_dims(phi::make_ddim(out_dims));
  out->set_dtype(x.dtype());
}

W
wuyefeilin 已提交
550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570
void DecodeJpegInferMeta(const MetaTensor& x,
                         const std::string& mode,
                         MetaTensor* out) {
  std::vector<int> out_dims;

  if (mode == "unchanged") {
    out_dims = {-1, -1, -1};
  } else if (mode == "gray") {
    out_dims = {1, -1, -1};
  } else if (mode == "rgb") {
    out_dims = {3, -1, -1};
  } else {
    errors::Fatal("The provided mode is not supported for JPEG files on GPU: ",
                  mode);
  }
  if (out != nullptr) {
    out->set_dims(phi::make_ddim(out_dims));
    out->set_dtype(x.dtype());
  }
}

571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633
void DiagEmbedInferMeta(
    const MetaTensor& x, int offset, int dim1, int dim2, MetaTensor* out) {
  auto x_dims = x.dims();

  PADDLE_ENFORCE_GE(
      dim1,
      -(x_dims.size() + 1),
      phi::errors::OutOfRange(
          "Dim1 is out of range (expected to be in range of [%ld, "
          "%ld], but got %ld).",
          -(x_dims.size() + 1),
          x_dims.size(),
          dim1));
  PADDLE_ENFORCE_LE(
      dim1,
      x_dims.size(),
      phi::errors::OutOfRange(
          "Dim1 is out of range (expected to be in range of [%ld, "
          "%ld], but got %ld).",
          -(x_dims.size() + 1),
          x_dims.size(),
          dim1));

  PADDLE_ENFORCE_GE(
      dim2,
      -(x_dims.size() + 1),
      phi::errors::OutOfRange(
          "Dim2 is out of range (expected to be in range of [%ld, "
          "%ld], but got %ld).",
          -(x_dims.size() + 1),
          x_dims.size(),
          dim2));
  PADDLE_ENFORCE_LE(
      dim2,
      x_dims.size(),
      phi::errors::OutOfRange(
          "Dim2 is out of range (expected to be in range of [%ld, "
          "%ld], but got %ld).",
          -(x_dims.size() + 1),
          x_dims.size(),
          dim2));

  int dim1_ = dim1 < 0 ? x_dims.size() + dim1 + 1 : dim1;
  int dim2_ = dim2 < 0 ? x_dims.size() + dim2 + 1 : dim2;
  int offset_ = std::abs(offset);

  PADDLE_ENFORCE_NE(dim1_,
                    dim2_,
                    phi::errors::InvalidArgument(
                        "diagonal dimensions should not be identical "
                        "%ld vs %ld.",
                        dim1,
                        dim2));

  int new_dim_len = offset_ + x_dims[x_dims.size() - 1];
  auto sizes = vectorize(x_dims);
  sizes.pop_back();
  sizes.insert(sizes.begin() + std::min(dim1_, dim2_), new_dim_len);
  sizes.insert(sizes.begin() + std::max(dim1_, dim2_), new_dim_len);
  out->set_dims(phi::make_ddim(sizes));
  out->set_dtype(x.dtype());
}

Z
zyfncg 已提交
634 635 636 637 638
void DiagInferMeta(const MetaTensor& x,
                   int offset,
                   float padding_value,
                   MetaTensor* out) {
  auto x_dims = x.dims();
639

640 641
  if (x_dims.size() <= 1) {
    int64_t size_ = (x_dims.size() == 1UL ? x_dims[0] : 1) + std::abs(offset);
Z
zyfncg 已提交
642 643 644 645 646 647 648 649 650 651 652 653
    out->set_dims({size_, size_});
    out->set_dtype(x.dtype());
  } else if (x_dims.size() == 2UL) {
    int64_t size_ = 0;
    if (offset >= 0) {
      // Note(LutaoChu): Do not use std::min here, otherwise the calculation
      // of `size_` will have unexpected result on Windows Python3.8
      if (x_dims[0] < x_dims[1] - offset) {
        size_ = x_dims[0];
      } else {
        size_ = x_dims[1] - offset;
      }
654
    } else {
Z
zyfncg 已提交
655 656 657 658 659 660 661
      // Note(LutaoChu): Do not use std::min here, otherwise the calculation
      // of `size_` will have unexpected result on Windows Python3.8
      if (x_dims[0] + offset < x_dims[1]) {
        size_ = x_dims[0] + offset;
      } else {
        size_ = x_dims[1];
      }
662
    }
Z
zyfncg 已提交
663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696
    out->set_dims({size_});
    out->set_dtype(x.dtype());
  } else {
    PADDLE_THROW(phi::errors::InvalidArgument(
        "The input tensor X's dimensions of DiagV2Op should be either 1 or "
        "2, but received %d.",
        x_dims.size()));
  }
}

void DiagonalInferMeta(const MetaTensor& input,
                       int offset,
                       int axis1,
                       int axis2,
                       MetaTensor* out) {
  auto x_dims = input.dims();
  int offset_ = offset;
  int axis1_ = axis1 < 0 ? x_dims.size() + axis1 : axis1;
  int axis2_ = axis2 < 0 ? x_dims.size() + axis2 : axis2;
  PADDLE_ENFORCE_GE(
      x_dims.size(),
      2,
      phi::errors::OutOfRange("Input's dim is out of range (expected at "
                              "least 2 dimensions, but got %ld).",
                              x_dims.size()));
  PADDLE_ENFORCE_LT(
      axis1_,
      x_dims.size(),
      phi::errors::OutOfRange(
          "Attr(axis1) is out of range (expected to be in range of [%ld, "
          "%ld], but got %ld).",
          -(x_dims.size()),
          (x_dims.size() - 1),
          axis1));
Z
zhangbo9674 已提交
697 698 699 700 701 702 703 704 705
  PADDLE_ENFORCE_GE(
      axis1_,
      0,
      phi::errors::OutOfRange(
          "Attr(axis1) is out of range (expected to be in range of [%ld, "
          "%ld], but got %ld).",
          -(x_dims.size()),
          (x_dims.size() - 1),
          axis1));
Z
zyfncg 已提交
706 707 708 709 710 711 712 713 714
  PADDLE_ENFORCE_LT(
      axis2_,
      x_dims.size(),
      phi::errors::OutOfRange(
          "Attr(axis2) is out of range (expected to be in range of [%ld, "
          "%ld], but got %ld).",
          -(x_dims.size()),
          (x_dims.size() - 1),
          axis2));
Z
zhangbo9674 已提交
715 716 717 718 719 720 721 722 723
  PADDLE_ENFORCE_GE(
      axis2_,
      0,
      phi::errors::OutOfRange(
          "Attr(axis2) is out of range (expected to be in range of [%ld, "
          "%ld], but got %ld).",
          -(x_dims.size()),
          (x_dims.size() - 1),
          axis2));
Z
zyfncg 已提交
724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764
  PADDLE_ENFORCE_NE(
      axis1_,
      axis2_,
      phi::errors::InvalidArgument("The dimensions should not be identical "
                                   "%d vs %d.",
                                   axis1,
                                   axis2));

  auto out_dims = vectorize(x_dims);
  // from out_dims get the dim size of axis1_.
  auto axis1_size = out_dims[axis1_];
  auto axis2_size = out_dims[axis2_];
  // delete two dims by attr axis1 and axis2 from out_dims.
  /* example:
     out_dim = [2, 3, 4];
     axis1 = 0;
     axis2 = 1;
     according to the attr of axis1 and axis2, we get:
     out_dim = [4].
  */
  out_dims.erase(out_dims.begin() + std::max(axis1_, axis2_));
  out_dims.erase(out_dims.begin() + std::min(axis1_, axis2_));

  if (offset_ == 0) {
    out_dims.push_back(std::min(axis1_size, axis2_size));
  } else if (offset_ > 0) {
    if ((axis2_size - offset_) > 0) {
      out_dims.push_back(std::min(axis1_size, axis2_size - offset_));
    } else {
      out_dims.push_back(0);
    }
  } else {
    if ((axis1_size + offset_) > 0) {
      out_dims.push_back(std::min(axis1_size + offset_, axis2_size));
    } else {
      out_dims.push_back(0);
    }
  }
  out->set_dims(phi::make_ddim(out_dims));
}

765 766 767 768 769 770 771 772 773 774 775 776 777
void DirichletInferMeta(const MetaTensor& alpha, MetaTensor* out) {
  const auto alpha_dim = alpha.dims();
  PADDLE_ENFORCE_GE(alpha_dim.size(),
                    1,
                    phi::errors::InvalidArgument(
                        "ShapeError: The number of dimensions of 'Alpha' "
                        "must be greater than or euqal to 1. "
                        "But received Alpha's dimensions = %d,",
                        alpha_dim.size()));
  out->set_dims(alpha_dim);
  out->set_dtype(alpha.dtype());
}

778 779 780 781 782 783 784 785 786 787
void DistReduceInferMeta(const MetaTensor& x, MetaTensor* out) {
  out->set_dtype(x.dtype());
  out->set_dims(x.dims());
}

void DistBroadcastInferMeta(const MetaTensor& x, MetaTensor* out) {
  out->set_dtype(x.dtype());
  out->set_dims(x.dims());
}

788
void EigInferMeta(const MetaTensor& x, MetaTensor* out_w, MetaTensor* out_v) {
789
  phi::DDim x_dims = x.dims();
790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808
  int rank = x_dims.size();
  PADDLE_ENFORCE_GE(
      rank,
      2,
      phi::errors::InvalidArgument("Expects input tensor x to be not less than "
                                   "2 dimentions, but got dimention %d",
                                   rank));
  PADDLE_ENFORCE_EQ(x_dims[rank - 2],
                    x_dims[rank - 1],
                    phi::errors::InvalidArgument(
                        "The input matrix must be a square matrix, "
                        "but receive a matrix with %d rows and %d colums",
                        x_dims[rank - 2],
                        x_dims[rank - 1]));

  std::vector<int> batch_dims_vec{};
  for (int i = 0; i < rank - 1; ++i) {
    batch_dims_vec.emplace_back(x_dims[i]);
  }
809 810 811
  const DataType& x_dtype = x.dtype();
  const DataType& out_dtype =
      IsComplexType(x_dtype) ? x_dtype : ToComplexType(x_dtype);
812
  out_w->set_dims(phi::make_ddim(batch_dims_vec));
813
  out_w->set_dtype(out_dtype);
814
  out_v->set_dims(x_dims);
815
  out_v->set_dtype(out_dtype);
816 817
}

Z
zyfncg 已提交
818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846
void EighInferMeta(const MetaTensor& x,
                   const std::string& uplo,
                   MetaTensor* out_w,
                   MetaTensor* out_v) {
  auto input_dim = x.dims();
  auto rank = input_dim.size();

  PADDLE_ENFORCE_GE(rank,
                    2,
                    phi::errors::InvalidArgument(
                        "The Input(X) should have at least 2 dimensions."
                        "But received a %d dimension tensor.",
                        rank));
  PADDLE_ENFORCE_EQ(
      input_dim[rank - 2],
      input_dim[rank - 1],
      phi::errors::InvalidArgument(
          "Eigh op is designed for square matrix, consequently"
          "inner-most 2 dimensions of Input(X) should be symmetric."
          "But received X's shape[-2] = %d and shape[-1] = %d.",
          input_dim[rank - 2],
          input_dim[rank - 1]));

  std::vector<int64_t> values_dim;

  for (auto i = 0; i < rank - 1; i++) {
    values_dim.emplace_back(input_dim[i]);
  }
  out_w->set_dims(phi::make_ddim(values_dim));
847
  out_w->set_dtype(dtype::ToReal(x.dtype()));
Z
zyfncg 已提交
848
  out_v->set_dims(input_dim);
849
  out_v->set_dtype(dtype::ToReal(x.dtype()));
Z
zyfncg 已提交
850 851
}

R
Ruibiao Chen 已提交
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
void EigvalsInferMeta(const MetaTensor& x, MetaTensor* out, MetaConfig config) {
  auto x_dims = x.dims();
  PADDLE_ENFORCE_GE(x_dims.size(),
                    2,
                    errors::InvalidArgument(
                        "The dimensions of Input(X) for Eigvals operator "
                        "should be at least 2, "
                        "but received X's dimension = %d, X's shape = [%s].",
                        x_dims.size(),
                        x_dims));

  if (config.is_runtime || !phi::contain_unknown_dim(x_dims)) {
    int last_dim = x_dims.size() - 1;
    PADDLE_ENFORCE_EQ(x_dims[last_dim],
                      x_dims[last_dim - 1],
                      errors::InvalidArgument(
                          "The last two dimensions of Input(X) for Eigvals "
                          "operator should be equal, "
                          "but received X's shape = [%s].",
                          x_dims));
  }

  auto out_dims = vectorize(x_dims);
  out_dims.resize(x_dims.size() - 1);

  const DataType& x_dtype = x.dtype();
  const DataType& out_dtype =
      IsComplexType(x_dtype) ? x_dtype : ToComplexType(x_dtype);

  out->set_dims(make_ddim(out_dims));
  out->set_dtype(out_dtype);
}

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
void EigvalshInferMeta(const MetaTensor& x,
                       const std::string& uplo,
                       bool is_test,
                       MetaTensor* out_w,
                       MetaTensor* out_v) {
  auto input_dim = x.dims();
  auto rank = input_dim.size();

  PADDLE_ENFORCE_GE(
      rank,
      2,
      errors::InvalidArgument("The Input(X) should have at least 2 dimensions."
                              "But received a %d dimension tensor.",
                              rank));
  PADDLE_ENFORCE_EQ(
      input_dim[rank - 2],
      input_dim[rank - 1],
      errors::InvalidArgument(
          "Eigvalsh op is designed for square matrix, consequently"
          "inner-most 2 dimensions of Input(X) should be symmetric."
          "But received X's shape[-2] = %d and shape[-1] = %d.",
          input_dim[rank - 2],
          input_dim[rank - 1]));

  std::vector<int64_t> values_dim;

  for (auto i = 0; i < rank - 1; i++) {
    values_dim.emplace_back(input_dim[i]);
  }

  if (out_w != nullptr) {
    out_w->set_dims(phi::make_ddim(values_dim));
    out_w->set_dtype(dtype::ToReal(x.dtype()));
  }
  if (out_v != nullptr) {
    out_v->set_dims(input_dim);
    out_v->set_dtype(x.dtype());
  }
}

925 926
void EinsumInferMeta(const std::vector<const MetaTensor*>& inputs,
                     const std::string& equation,
927
                     MetaTensor* out) {
928 929 930 931 932 933 934 935 936 937 938 939 940
  // collect the following informations to prepare einsum.
  LabelMap labelshape(0);
  LabelMap labeltype(LabelType::Reduction);
  std::vector<LabelMap> label2perms(inputs.size(), LabelMap(-1));
  std::vector<char> all_labels;
  std::vector<int> broadcast_dims;
  std::vector<int> output_dims;
  std::vector<std::vector<int>> ellipsis_dims(2);

  std::vector<DDim> input_dims;
  for (auto& i : inputs) {
    input_dims.push_back(i->dims());
  }
941
  std::vector<std::string> input_strs;
942 943 944 945 946 947 948 949 950 951
  std::string right;
  ParseEinsumEquation(equation,
                      input_dims,
                      &labelshape,
                      &labeltype,
                      &all_labels,
                      &label2perms,
                      &ellipsis_dims,
                      &broadcast_dims,
                      &output_dims,
952 953
                      &right,
                      &input_strs);
954 955 956 957 958 959 960 961 962 963

  VLOG(3) << "Einsum Infershape: input dims:"
          << paddle::string::join_strings(input_dims, "\n");
  VLOG(3) << "Einsum Infershape: equation:" << equation;
  VLOG(3) << "Einsum Infershape: all_labels:"
          << paddle::string::join_strings(all_labels, ",");
  VLOG(3) << "Einsum Infershape: output dims:"
          << paddle::string::join_strings(output_dims, ",");
  VLOG(3) << "Label Type is : " << label_to_string(all_labels, labeltype);
  VLOG(3) << "Label Shape is : " << label_to_string(all_labels, labelshape);
964 965
  out->set_dims(make_ddim(output_dims));
  out->set_dtype(inputs[0]->dtype());
966 967 968 969 970 971 972 973
}

void EinsumRawInferMeta(const std::vector<const MetaTensor*>& inputs,
                        const std::string& equation,
                        MetaTensor* out,
                        std::vector<MetaTensor*> inner_cache,
                        std::vector<MetaTensor*> xshape) {
  EinsumInferMeta(inputs, equation, out);
974 975 976 977 978 979
  for (size_t i = 0; i < xshape.size(); ++i) {
    if (xshape[i] != nullptr) {
      xshape[i]->set_dims(inputs[i]->dims());
      xshape[i]->set_dtype(inputs[i]->dtype());
    }
  }
980 981
}

H
hong 已提交
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
void ExpandInferMeta(const MetaTensor& x,
                     const IntArray& shape,
                     MetaTensor* out) {
#define MAX_RANK_SUPPORTED 6
  auto x_dims = x.dims();
  auto expand_shape = shape.GetData();

  if (expand_shape.size() == 0) {
    expand_shape = std::vector<int64_t>(x_dims.size(), -1);
  }

  PADDLE_ENFORCE_GE(
      expand_shape.size(),
      static_cast<size_t>(x_dims.size()),
      phi::errors::InvalidArgument(
          "The number of elements (%d) of 'shape' for "
          "expand_v2 op must be greater than or equal to the rank "
          "(%d) of the input.",
          expand_shape.size(),
          static_cast<size_t>(x_dims.size())));
  PADDLE_ENFORCE_LE(
      expand_shape.size(),
      MAX_RANK_SUPPORTED,
      phi::errors::InvalidArgument("The number of elements (%d) of 'shape' for "
                                   "must not be greater than %d.",
                                   expand_shape.size(),
                                   MAX_RANK_SUPPORTED));
  PADDLE_ENFORCE_GE(
      expand_shape.size(),
1011
      0,
H
hong 已提交
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
      phi::errors::InvalidArgument("The number of elements (%d) of 'shape' for "
                                   "must be a positive integer.",
                                   expand_shape.size()));

  auto out_rank =
      std::max(static_cast<size_t>(x_dims.size()), expand_shape.size());
  std::vector<int64_t> out_shape(out_rank);
  for (size_t i = 0; i < expand_shape.size(); ++i) {
    if (x_dims[i] == -1) {
      out_shape[i] = -1;
    } else if (expand_shape[i] == -1) {
      if (static_cast<size_t>(x_dims.size()) > i) {
        out_shape[i] = x_dims[i];
      } else {
        out_shape[i] = -1;
      }
    } else if (expand_shape[i] == -2) {
      // We use -2 to represent the element in expand_shape is a var.
      out_shape[i] = -1;
    } else {
      PADDLE_ENFORCE_GT(
          expand_shape[i],
          0,
          phi::errors::InvalidArgument(
              "The %uth element of 'shape' for expand_v2 op must be "
              "greater than 0, but the value given is %d.",
              i,
              expand_shape[i]));
      out_shape[i] = expand_shape[i];
    }
  }

  out->set_dims(make_ddim(out_shape));
  out->set_dtype(x.dtype());
1046
  if (out_rank > 0 && out_shape[0] == x_dims[0]) {
H
hong 已提交
1047 1048 1049 1050
    out->share_lod(x);
  }
}

Z
zhiboniu 已提交
1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061
void FillDiagonalInferMeta(
    const MetaTensor& x, float value, int offset, bool wrap, MetaTensor* out) {
  PADDLE_ENFORCE_NE(
      out,
      nullptr,
      phi::errors::InvalidArgument("Tensor out should not be null if "));
  auto x_dims = x.dims();
  out->set_dims(x_dims);
  out->set_dtype(x.dtype());
}

F
Feiyu Chan 已提交
1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 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 1165 1166 1167
void FFTC2CInferMeta(const MetaTensor& x,
                     const std::vector<int64_t>& axes,
                     const std::string& normalization,
                     bool forward,
                     MetaTensor* out,
                     MetaConfig config) {
  PADDLE_ENFORCE_NOT_NULL(
      out,
      phi::errors::InvalidArgument("Output of fft_c2c should not be null."));
  // only ensure that fft axes' size greater than zero at runtime
  // they might be -1 to indicate unknown size ar compile time
  if (config.is_runtime) {
    const phi::DDim x_dim = x.dims();
    for (size_t i = 0; i < axes.size(); i++) {
      PADDLE_ENFORCE_GT(x_dim[axes[i]],
                        0,
                        phi::errors::InvalidArgument(
                            "Invalid fft n-point (%d).", x_dim[axes[i]]));
    }
  }
  out->share_meta(x);
}

void FFTC2RInferMeta(const MetaTensor& x,
                     const std::vector<int64_t>& axes,
                     const std::string& normalization,
                     bool forward,
                     int64_t last_dim_size,
                     MetaTensor* out,
                     MetaConfig config) {
  PADDLE_ENFORCE_NOT_NULL(
      out,
      phi::errors::InvalidArgument("Output of fft_c2r should not be null."));
  const phi::DDim x_dim = x.dims();
  const int64_t last_fft_axis = axes.back();

  // only ensure that fft axes' size greater than zero at runtime
  // they might be -1 to indicate unknown size ar compile time
  if (config.is_runtime) {
    size_t signal_dims = axes.size();
    for (size_t i = 0; i < signal_dims - 1; i++) {
      PADDLE_ENFORCE_GT(x_dim[axes[i]],
                        0,
                        phi::errors::InvalidArgument(
                            "Invalid fft n-point (%d).", x_dim[axes[i]]));
    }
  }

  out->set_layout(x.layout());
  out->set_dtype(ToRealType(x.dtype()));
  phi::DDim out_dim = x_dim;

  if (last_dim_size > 0) {
    out_dim.at(last_fft_axis) = last_dim_size;
  } else if (config.is_runtime) {
    const int64_t input_last_dim_size = x_dim[last_fft_axis];
    const int64_t fft_n_point = (input_last_dim_size - 1) * 2;
    PADDLE_ENFORCE_GT(
        fft_n_point,
        0,
        phi::errors::InvalidArgument("Invalid fft n-point (%d).", fft_n_point));
    out_dim.at(last_fft_axis) = fft_n_point;
  } else {
    const int64_t input_last_dim_size = x_dim[last_fft_axis];
    out_dim.at(last_fft_axis) =
        input_last_dim_size == -1 ? -1 : (input_last_dim_size - 1) * 2;
  }
  out->set_dims(out_dim);
}

void FFTR2CInferMeta(const MetaTensor& x,
                     const std::vector<int64_t>& axes,
                     const std::string& normalization,
                     bool forward,
                     bool onesided,
                     MetaTensor* out,
                     MetaConfig config) {
  PADDLE_ENFORCE_NOT_NULL(
      out,
      phi::errors::InvalidArgument("Output of fft_r2c should not be null."));
  const phi::DDim x_dim = x.dims();

  // only ensure that fft axes' size greater than zero at runtime
  // they might be -1 to indicate unknown size ar compile time
  if (config.is_runtime) {
    for (size_t i = 0; i < axes.size(); i++) {
      PADDLE_ENFORCE_GT(x_dim[axes[i]],
                        0,
                        phi::errors::InvalidArgument(
                            "Invalid fft n-point (%d).", x_dim[axes[i]]));
    }
  }

  out->set_layout(x.layout());
  out->set_dtype(ToComplexType(x.dtype()));
  if (!onesided) {
    out->share_dims(x);
  } else {
    phi::DDim out_dim = x.dims();
    const int64_t last_fft_axis = axes.back();
    const int64_t last_fft_dim_size = x_dim[last_fft_axis];
    out_dim.at(last_fft_axis) = last_fft_dim_size / 2 + 1;
    out->set_dims(out_dim);
  }
}

Z
zyfncg 已提交
1168 1169 1170 1171
void FlattenInferMeta(const MetaTensor& x,
                      int start_axis,
                      int stop_axis,
                      MetaTensor* out) {
1172 1173 1174 1175 1176 1177 1178 1179
  FlattenWithXShapeInferMeta(x, start_axis, stop_axis, out, nullptr);
}

void FlattenWithXShapeInferMeta(const MetaTensor& x,
                                int start_axis,
                                int stop_axis,
                                MetaTensor* out,
                                MetaTensor* xshape) {
Z
zyfncg 已提交
1180 1181
  auto x_dims = x.dims();
  int in_dims_size = x_dims.size();
1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198

  if (in_dims_size == 0) {
    PADDLE_ENFORCE_EQ(
        start_axis == 0 || start_axis == -1,
        true,
        phi::errors::InvalidArgument("The start_axis should be 0 or -1 when "
                                     "the input tensor is a 0D-Tensor"));
    PADDLE_ENFORCE_EQ(
        stop_axis == 0 || stop_axis == -1,
        true,
        phi::errors::InvalidArgument("The stop_axis should be 0 or -1 when the "
                                     "input tensor is a 0D-Tensor"));
    // this can ensure out shape {1}
    start_axis = 0;
    stop_axis = -1;
  }

Z
zyfncg 已提交
1199 1200 1201 1202 1203 1204
  if (start_axis < 0) {
    start_axis = start_axis + in_dims_size;
  }
  if (stop_axis < 0) {
    stop_axis = stop_axis + in_dims_size;
  }
1205 1206 1207 1208 1209 1210 1211
  if (in_dims_size > 0) {
    PADDLE_ENFORCE_GE(
        stop_axis,
        start_axis,
        phi::errors::InvalidArgument("The stop_axis should be greater"
                                     "than or equal to start_axis."));
  }
Z
zyfncg 已提交
1212 1213 1214

  int64_t outer = 1;
  std::vector<int32_t> out_shape;
1215
  out_shape.reserve(in_dims_size - stop_axis + start_axis + 1);
Z
zyfncg 已提交
1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240

  for (int i = 0; i < start_axis; ++i) {
    out_shape.push_back(x_dims[i]);
  }
  for (int i = start_axis; i <= stop_axis; i++) {
    if (x_dims[i] == -1 || outer == -1) {
      outer = -1;
    } else {
      outer *= x_dims[i];
    }
  }
  out_shape.push_back(outer);
  for (int i = stop_axis + 1; i < in_dims_size; i++) {
    out_shape.push_back(x_dims[i]);
  }
  const auto& out_dims = phi::make_ddim(out_shape);
  out->set_dims(out_dims);
  out->set_dtype(x.dtype());
  out->set_layout(x.layout());

  if (x_dims[0] == out_dims[0]) {
    // Only pass LoD when the first dimension of output and Input(X)
    // are the same.
    out->share_lod(x);
  }
1241 1242 1243 1244 1245 1246 1247 1248
  if (xshape == nullptr) return;
  std::vector<int64_t> xshape_dims(x_dims.size() + 1);
  xshape_dims[0] = 0;
  for (int i = 0; i < x_dims.size(); ++i) {
    xshape_dims[i + 1] = x_dims[i];
  }
  xshape->set_dims(phi::make_ddim(xshape_dims));
  xshape->share_lod(x);
Z
zyfncg 已提交
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 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325
void FlipInferMeta(const MetaTensor& x,
                   const std::vector<int>& axis,
                   MetaTensor* out) {
  auto x_dims = x.dims();
  auto flip_dims = axis;
  size_t flip_dims_size = axis.size();

  if (flip_dims_size > 0) {
    // check if dims axis within range
    auto min_max_d = std::minmax_element(flip_dims.begin(), flip_dims.end());
    PADDLE_ENFORCE_LT(*min_max_d.first,
                      x_dims.size(),
                      phi::errors::InvalidArgument(
                          "min(axes) should be less than the input tensor X's "
                          "axes of FlipOp. But received min(axes) = %d,  "
                          "X's axes = %d, X's shape = [%s]",
                          *min_max_d.first,
                          x_dims.size(),
                          x_dims));
    PADDLE_ENFORCE_GE(*min_max_d.first,
                      x_dims.size() * -1,
                      phi::errors::InvalidArgument(
                          "min(axes) should be greater than or equal to the "
                          "input tensor X's "
                          "axes of FlipOp times -1. But received "
                          "min(axes) = %d,  X's "
                          "axes = %d, X's shape = [%s]",
                          *min_max_d.first,
                          x_dims.size() * -1,
                          x_dims));
    PADDLE_ENFORCE_LT(*min_max_d.second,
                      x_dims.size(),
                      phi::errors::InvalidArgument(
                          "max(axes) should be less than the input tensor X's "
                          "axes of FlipOp. But received max(axes) = %d,  "
                          "X's axes = %d, X's shape = [%s]",
                          *min_max_d.second,
                          x_dims.size(),
                          x_dims));
    PADDLE_ENFORCE_GE(*min_max_d.second,
                      x_dims.size() * -1,
                      phi::errors::InvalidArgument(
                          "max(axes) should be greater than or equal to the "
                          "input tensor X's "
                          "axes of FlipOp times -1. But received "
                          "max(axes) = %d,  X's "
                          "axes = %d, X's shape = [%s]",
                          *min_max_d.second,
                          x_dims.size() * -1,
                          x_dims));

    // check duplicates in dims
    flip_dims.erase(std::unique(flip_dims.begin(), flip_dims.end()),
                    flip_dims.end());
    PADDLE_ENFORCE_EQ(flip_dims.size(),
                      flip_dims_size,
                      phi::errors::InvalidArgument(
                          "axes has duplicates, original flip axes size=%d, "
                          "but unique flip axes size=%d.)",
                          flip_dims_size,
                          flip_dims.size()));
  }

  VLOG(3) << "flip operator x.shape=" << x_dims;

  std::vector<int64_t> output_dims(x_dims.size());
  for (int i = 0; i < x_dims.size(); ++i) {
    output_dims[i] = x_dims[i];
  }

  out->set_dims(phi::make_ddim(output_dims));
  out->set_dtype(x.dtype());
  out->share_lod(x);
}

1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530
void FoldInferMeta(const MetaTensor& x,
                   const std::vector<int>& output_sizes,
                   const std::vector<int>& kernel_sizes,
                   const std::vector<int>& strides,
                   const std::vector<int>& paddings,
                   const std::vector<int>& dilations,
                   MetaTensor* out) {
  auto in_dims = x.dims();

  PADDLE_ENFORCE_EQ(
      output_sizes.size(),
      2,
      phi::errors::InvalidArgument(
          "It is expected output_size equals to 2, but got size %d",
          output_sizes.size()));
  PADDLE_ENFORCE_EQ(
      kernel_sizes.size(),
      2,
      phi::errors::InvalidArgument(
          "It is expected kernel_size equals to 2, but got size %d",
          kernel_sizes.size()));
  PADDLE_ENFORCE_EQ(
      strides.size(),
      2,
      phi::errors::InvalidArgument(
          "It is expected strides_size equals to 2, but got size %d",
          strides.size()));
  PADDLE_ENFORCE_EQ(
      paddings.size(),
      4,
      phi::errors::InvalidArgument(
          "It is expected paddings_size equals to 4, but got size %d",
          paddings.size()));

  PADDLE_ENFORCE_EQ(
      dilations.size(),
      2,
      phi::errors::InvalidArgument(
          "It is expected dilations_size equals to 2, but got size %d",
          dilations.size()));

  int output_height = output_sizes[0];
  int output_width = output_sizes[1];
  int kernel_height = kernel_sizes[0];
  int kernel_width = kernel_sizes[1];
  int dilation_height = dilations[0];
  int dilation_width = dilations[1];
  int stride_height = strides[0];
  int stride_width = strides[1];

  // check kernel_sizes
  PADDLE_ENFORCE_GT(kernel_height,
                    0,
                    phi::errors::InvalidArgument(
                        "The `kernel_sizes` should be greater than zero, "
                        "but received kernel_height: %d kernel_width: %d.",
                        kernel_sizes[0],
                        kernel_sizes[1]));
  PADDLE_ENFORCE_GT(kernel_width,
                    0,
                    phi::errors::InvalidArgument(
                        "The `kernel_sizes` should be greater than zero, "
                        "but received kernel_height: %d kernel_width: %d.",
                        kernel_sizes[0],
                        kernel_sizes[1]));
  // check strides
  PADDLE_ENFORCE_GT(stride_height,
                    0,
                    phi::errors::InvalidArgument(
                        "The `strides` should be greater than zero, "
                        "but received strides_height: %d strides_width: %d.",
                        strides[0],
                        strides[1]));
  PADDLE_ENFORCE_GT(stride_width,
                    0,
                    phi::errors::InvalidArgument(
                        "The `strides` should be greater than zero, "
                        "but received strides_height: %d strides_width: %d.",
                        strides[0],
                        strides[1]));
  // check dilations
  PADDLE_ENFORCE_GT(output_height,
                    1,
                    phi::errors::InvalidArgument(
                        "The `output_height` should be greater than one, "
                        "but received output_height: %d .",
                        output_height));
  PADDLE_ENFORCE_GT(output_width,
                    1,
                    phi::errors::InvalidArgument(
                        "The `output_width` should be greater than one, "
                        "but received output_width: %d .",
                        output_width));
  // check output size
  PADDLE_ENFORCE_GT(
      dilation_height,
      0,
      phi::errors::InvalidArgument(
          "The `dilations` should be greater than zero, "
          "but received dilations_height: %d dilations_width: %d.",
          dilations[0],
          dilations[1]));
  PADDLE_ENFORCE_GT(
      dilation_width,
      0,
      phi::errors::InvalidArgument(
          "The `dilations` should be greater than zero, "
          "but received dilations_height: %d dilations_width: %d.",
          dilations[0],
          dilations[1]));

  std::vector<int> out_dims;
  // batch_size
  out_dims.push_back(in_dims[0]);
  // output_plane
  int output_channels = in_dims[1] / (kernel_width * kernel_height);
  out_dims.push_back(output_channels);

  int blocks_height = (output_sizes[0] + 2 * paddings[0] -
                       (dilations[0] * (kernel_sizes[0] - 1) + 1)) /
                          strides[0] +
                      1;
  int blocks_width = (output_sizes[1] + 2 * paddings[1] -
                      (dilations[1] * (kernel_sizes[1] - 1) + 1)) /
                         strides[1] +
                     1;

  // check output height and width
  PADDLE_ENFORCE_GT(
      blocks_height,
      0,
      phi::errors::InvalidArgument(
          "The sliding blocks calculated from input spatial size (%d, %d), "
          "kernel_sizes (%d, %d), strides (%d, %d), dilations (%d, %d), "
          "is (%d, %d), which should be a positive integer.",
          in_dims[2],
          in_dims[3],
          kernel_sizes[0],
          kernel_sizes[1],
          strides[0],
          strides[1],
          dilations[0],
          dilations[1],
          output_height,
          output_width));

  PADDLE_ENFORCE_GT(
      blocks_width,
      0,
      phi::errors::InvalidArgument(
          "The sliding blocks calculated from input spatial size (%d, %d), "
          "kernel_sizes (%d, %d), strides (%d, %d), dilations (%d, %d), "
          "is (%d, %d), which should be a positive integer.",
          in_dims[2],
          in_dims[3],
          kernel_sizes[0],
          kernel_sizes[1],
          strides[0],
          strides[1],
          dilations[0],
          dilations[1],
          output_height,
          output_width));

  PADDLE_ENFORCE_EQ(
      blocks_height * blocks_width,
      in_dims[2],
      phi::errors::InvalidArgument(
          "Given input output_size (%d, %d), "
          "kernel_sizes (%d, %d), strides (%d, %d), dilations (%d, %d), "
          "which should be expected size of input's dimension "
          "2 to match the calculated number of %d * %d = %d, but got %d",
          output_height,
          output_width,
          kernel_sizes[0],
          kernel_sizes[1],
          strides[0],
          strides[1],
          dilations[0],
          dilations[1],
          blocks_height,
          blocks_width,
          blocks_height * blocks_width,
          in_dims[2]));

  PADDLE_ENFORCE_EQ(
      in_dims[1] % (kernel_sizes[0] * kernel_sizes[1]),
      0,
      phi::errors::InvalidArgument(
          "Expected size of input's dimension 1 to be divisible by the"
          "product of kernel_size, but got input.size(1)=%d and "
          "kernel_size=( %d"
          ", %d).",
          in_dims[1],
          kernel_sizes[0],
          kernel_sizes[1]));

  out_dims.push_back(output_height);
  out_dims.push_back(output_width);
  if (out != nullptr) {
    out->set_dims(phi::make_ddim(out_dims));
    out->set_dtype(x.dtype());
  }
}

C
Charles-hit 已提交
1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614
void FrameInferMeta(const MetaTensor& x,
                    int frame_length,
                    int hop_length,
                    int axis,
                    MetaTensor* out,
                    MetaConfig config) {
  PADDLE_ENFORCE_NOT_NULL(out,
                          phi::errors::InvalidArgument(
                              "Output(Out) of FrameOp should not be null."));
  const auto x_dims = x.dims();
  const int x_rank = x_dims.size();

  PADDLE_ENFORCE_GE(x_rank,
                    1,
                    phi::errors::InvalidArgument(
                        "Input(X) of FrameOp should be a tensor which contains "
                        "at least 1 dimension, but got rank %s.",
                        x_rank));
  PADDLE_ENFORCE_GT(hop_length,
                    0,
                    phi::errors::InvalidArgument(
                        "Attribute(hop_length) of FrameOp should be greater "
                        "than 0, but got %s.",
                        hop_length));
  PADDLE_ENFORCE_EQ(
      (axis == 0 || axis == -1),
      true,
      phi::errors::InvalidArgument(
          "Attribute(axis) of FrameOp should 0 or -1, but got %s.", axis));

  std::vector<int64_t> output_shape;
  int seq_length;
  int n_frames;

  int start_axis;
  int end_axis;

  if (axis == 0) {
    seq_length = x_dims[0];
    start_axis = 1;
    end_axis = x_rank - 1;
  } else {
    seq_length = x_dims[x_rank - 1];
    start_axis = 0;
    end_axis = x_rank - 2;
  }

  bool contain_unknown_dim = phi::contain_unknown_dim(x_dims);
  bool check = config.is_runtime || !contain_unknown_dim;
  if (check) {
    PADDLE_ENFORCE_LE(frame_length,
                      seq_length,
                      phi::errors::InvalidArgument(
                          "Attribute(frame_length) of FrameOp should be less "
                          "equal than sequence length, but got (%s) > (%s).",
                          frame_length,
                          seq_length));
  }

  // It won't go into for loop when x_rank == 1U.
  for (int i = start_axis; i <= end_axis; i++) {
    output_shape.push_back(x_dims[i]);
  }

  if (seq_length == -1) {
    n_frames = -1;
  } else {
    n_frames = 1 + (seq_length - frame_length) / hop_length;
  }

  if (axis == 0) {
    // (n_frames, frame_length, ...)
    output_shape.insert(output_shape.begin(), frame_length);
    output_shape.insert(output_shape.begin(), n_frames);
  } else {
    // (..., frame_length, n_frames)
    output_shape.push_back(frame_length);
    output_shape.push_back(n_frames);
  }

  out->set_dims(phi::make_ddim(output_shape));
  out->set_dtype(x.dtype());
}

1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625
void FullBatchSizeLikeInferMeta(const MetaTensor& x,
                                const std::vector<int>& shape,
                                const Scalar& val,
                                DataType dtype,
                                int x_batch_size_dim,
                                int out_batch_size_dim,
                                MetaTensor* out) {
  BatchSizeLikeInferMeta(x, shape, x_batch_size_dim, out_batch_size_dim, out);
  out->set_dtype(dtype);
}

Z
zyfncg 已提交
1626 1627 1628 1629 1630 1631 1632 1633
void GumbelSoftmaxInferMeta(const MetaTensor& x,
                            float temperature,
                            bool hard,
                            int axis,
                            MetaTensor* out) {
  UnchangedInferMetaCheckAxis(x, axis, out);
}

H
hong 已提交
1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651
void HistogramInferMeta(
    const MetaTensor& input, int64_t bins, int min, int max, MetaTensor* out) {
  PADDLE_ENFORCE_GE(bins,
                    1,
                    phi::errors::InvalidArgument(
                        "The bins should be greater than or equal to 1."
                        "But received nbins is %d",
                        bins));
  PADDLE_ENFORCE_GE(
      max,
      min,
      phi::errors::InvalidArgument("max must be larger or equal to min."
                                   "But received max is %d, min is %d",
                                   max,
                                   min));

  out->set_dims({bins});
  out->share_lod(input);
1652
  out->set_dtype(DataType::INT64);
H
hong 已提交
1653 1654
}

1655 1656 1657 1658 1659 1660 1661
void IdentityLossInferMeta(const MetaTensor& x,
                           int reduction,
                           MetaTensor* out) {
  if (reduction == 2) {
    out->set_dtype(x.dtype());
    out->set_dims(x.dims());
  } else {
1662
    out->set_dims(phi::make_ddim({}));
1663 1664 1665 1666
    out->set_dtype(x.dtype());
  }
}

Z
zyfncg 已提交
1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685
void IncrementInferMeta(const MetaTensor& x, float value, MetaTensor* out) {
  PADDLE_ENFORCE_EQ(
      product(x.dims()),
      1UL,
      errors::InvalidArgument("The number of elements in Input(X) should be 1."
                              "Now the number is %d.",
                              product(x.dims())));
  out->set_dims(x.dims());
  out->share_lod(x);
  out->set_dtype(x.dtype());
}

static phi::DDim ValidateShape(const std::vector<int64_t> shape,
                               const phi::DDim& in_dims) {
  const int64_t in_size = phi::product(in_dims);
  auto in_dims_vec = phi::vectorize(in_dims);
  std::vector<int64_t> output_shape(shape.size(), 0);
  int64_t capacity = 1;
  int unk_dim_idx = -1;
1686

Z
zyfncg 已提交
1687
  for (size_t i = 0; i < shape.size(); ++i) {
1688 1689
    if (shape[i] == -1) {
      // only one dimension can be set to -1, whose size will be infered.
Z
zyfncg 已提交
1690 1691 1692 1693 1694 1695 1696 1697 1698
      PADDLE_ENFORCE_EQ(
          unk_dim_idx,
          -1,
          phi::errors::InvalidArgument(
              "Only one dimension value of 'shape' in ReshapeOp can "
              "be -1. But received shape = [%s], shape[%d] is also -1.",
              phi::make_ddim(shape),
              i));
      unk_dim_idx = i;
1699
      output_shape[i] = shape[i];
1700
    } else if (shape[i] == 0) {
1701
      if (static_cast<int>(i) < in_dims.size()) {
1702 1703
        output_shape[i] = in_dims[i];
      } else {
1704 1705 1706 1707 1708 1709
        PADDLE_ENFORCE_EQ(
            in_size,
            0,
            phi::errors::InvalidArgument("If The index of 0 in `shape` >= "
                                         "the input tensor X's dimensions, "
                                         "It can only be Zero-Sized Tensor"));
1710 1711
      }
      capacity *= output_shape[i];
Z
zyfncg 已提交
1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722
    } else {
      PADDLE_ENFORCE_GT(
          shape[i],
          0,
          phi::errors::InvalidArgument(
              "Each dimension value of 'shape' in ReshapeOp must not "
              "be negative except one unknown dimension. "
              "But received  shape = [%s], shape[%d] = %d.",
              phi::make_ddim(shape),
              i,
              shape[i]));
1723 1724
      output_shape[i] = shape[i];
      capacity *= output_shape[i];
Z
zyfncg 已提交
1725
    }
1726
  }
Z
zyfncg 已提交
1727

1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740
  if (capacity == 0) {
    PADDLE_ENFORCE_EQ(in_size,
                      0,
                      phi::errors::InvalidArgument(
                          "Only Zero-Size Tensor'shape can contain 0"));
    PADDLE_ENFORCE_EQ(unk_dim_idx,
                      -1,
                      phi::errors::InvalidArgument(
                          "can not rehsape %s to %s, because the unspecified "
                          "dimension %i can be any number and is ambiguous",
                          in_dims,
                          phi::make_ddim(shape),
                          unk_dim_idx));
1741 1742
  }

1743 1744 1745
  bool no_negative = std::all_of(in_dims_vec.cbegin(),
                                 in_dims_vec.cend(),
                                 [](int64_t i) { return i >= 0; });
1746
  if (unk_dim_idx != -1) {
1747 1748 1749
    // in compile time, no_negative may be False.
    if (no_negative) {
      output_shape[unk_dim_idx] = in_size / capacity;
1750 1751
      PADDLE_ENFORCE_EQ(
          output_shape[unk_dim_idx] * capacity,
1752
          in_size,
1753
          phi::errors::InvalidArgument(
1754 1755 1756 1757 1758 1759 1760
              "The 'shape' attribute in ReshapeOp is invalid. "
              "The input tensor X'size must be divisible by known "
              "capacity of 'shape'. "
              "But received X's shape = [%s], X's size = %d, "
              "'shape' is [%s], known capacity of 'shape' is %d.",
              in_dims,
              in_size,
1761
              phi::make_ddim(shape),
1762 1763
              capacity));
    } else {
1764
      // such as [-1, 8, 3]->[-1, 8], out_shape will remain [-1, 8]
1765 1766 1767
      output_shape[unk_dim_idx] = -1;
    }
  } else {
1768
    if (no_negative) {
1769 1770 1771
      PADDLE_ENFORCE_EQ(
          capacity,
          in_size,
1772
          phi::errors::InvalidArgument(
1773 1774 1775 1776 1777 1778 1779
              "The 'shape' in ReshapeOp is invalid. "
              "The input tensor X'size must be equal to the capacity of "
              "'shape'. "
              "But received X's shape = [%s], X's size = %d, 'shape' is "
              "[%s], the capacity of 'shape' is %d.",
              in_dims,
              in_size,
1780
              phi::make_ddim(shape),
1781 1782 1783 1784
              capacity));
    }
  }

1785
  return phi::make_ddim(output_shape);
1786 1787
}

1788 1789 1790 1791
void InferMetaFromVecValue(const MetaTensor& x,
                           const std::vector<int64_t>& shape,
                           MetaTensor* out) {
  auto x_dims = x.dims();
1792
  auto out_dims = ValidateShape(shape, x_dims);
1793 1794 1795
  out->set_dims(out_dims);
  out->set_dtype(x.dtype());
  out->set_layout(x.layout());
1796
  if (x_dims.size() > 0 && (x_dims[0] == out_dims[0])) {
1797 1798
    // Only pass LoD when the first dimension of output and Input(X)
    // are the same.
1799
    out->share_lod(x);
1800 1801 1802
  }
}

1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836
void InverseInferMeta(const MetaTensor& x, MetaTensor* out) {
  auto input_dims = x.dims();
  int64_t input_rank = input_dims.size();
  PADDLE_ENFORCE_GE(
      input_rank,
      2,
      errors::InvalidArgument(
          "The dimension of Input(Input) is expected to be no less than 2. "
          "But received: Input(Input)'s dimension = %d, shape = [%s].",
          input_rank,
          input_dims));
  for (int64_t i = 0; i < input_rank; ++i) {
    PADDLE_ENFORCE_EQ(
        (input_dims[i] == -1) || (input_dims[i] > 0),
        true,
        errors::InvalidArgument(
            "Each dimension of input tensor is expected to be -1 or a "
            "positive number, but received %d. Input's shape is [%s].",
            input_dims[i],
            input_dims));
  }
  if (input_dims[input_rank - 2] > 0 && input_dims[input_rank - 1] > 0) {
    PADDLE_ENFORCE_EQ(input_dims[input_rank - 2],
                      input_dims[input_rank - 1],
                      errors::InvalidArgument(
                          "The last two dimensions are expected to be equal. "
                          "But received: %d and %d; "
                          "Input(Input)'s shape = [%s].",
                          input_dims[input_rank - 2],
                          input_dims[input_rank - 1],
                          input_dims));
  }

  out->set_dims(input_dims);
1837
  out->set_dtype(x.dtype());
1838 1839 1840
  out->share_lod(x);
}

W
WJJ1995 已提交
1841
void IsEmptyInferMeta(const MetaTensor& x, MetaTensor* out) {
1842
  out->set_dims(phi::make_ddim({}));
W
WJJ1995 已提交
1843 1844 1845
  out->set_dtype(DataType::BOOL);
}

Z
zyfncg 已提交
1846 1847 1848 1849 1850
void IsfiniteInferMeta(const MetaTensor& x, MetaTensor* out) {
  out->set_dims(x.dims());
  out->set_dtype(DataType::BOOL);
}

1851 1852 1853 1854 1855 1856 1857 1858 1859
void KthvalueInferMeta(const MetaTensor& x,
                       int k,
                       int axis,
                       bool keepdim,
                       MetaTensor* out,
                       MetaTensor* indices,
                       MetaConfig config) {
  auto input_dims = x.dims();
  const int& dim_size = input_dims.size();
1860
  if (dim_size > 0) {
1861 1862 1863 1864 1865 1866 1867
    PADDLE_ENFORCE_LT(axis,
                      dim_size,
                      phi::errors::InvalidArgument(
                          "the axis must be [-%d, %d), but received %d .",
                          dim_size,
                          dim_size,
                          axis));
1868 1869 1870 1871 1872 1873 1874
    PADDLE_ENFORCE_GE(axis,
                      -dim_size,
                      phi::errors::InvalidArgument(
                          "the axis must be [-%d, %d), but received %d .",
                          dim_size,
                          dim_size,
                          axis));
1875 1876 1877 1878 1879 1880 1881 1882
  } else if (dim_size == 0) {
    // 0-dim tensor
    PADDLE_ENFORCE_EQ(axis == 0 || axis == -1,
                      true,
                      phi::errors::InvalidArgument(
                          "'axis'(%d) must be 0 or -1 if input tensor is "
                          "0-dim.",
                          axis));
1883
  }
1884 1885 1886 1887 1888 1889 1890 1891
  if (axis < 0) axis += dim_size;
  PADDLE_ENFORCE_GE(
      k,
      1,
      phi::errors::InvalidArgument(
          "the k in the kthvalue must >= 1, but received %d .", k));
  PADDLE_ENFORCE_GE(
      input_dims.size(),
1892 1893 1894
      0,
      phi::errors::InvalidArgument("input of kthvalue must have >= 0d shape"));
  if (dim_size > 0 && config.is_runtime) {
1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906
    PADDLE_ENFORCE_GE(
        input_dims[axis],
        k,
        phi::errors::InvalidArgument(
            "input of kthvalue must have >= %d columns in axis of %d",
            k,
            axis));
  }
  std::vector<int64_t> dimvec;
  for (int64_t i = 0; i < axis; i++) {
    dimvec.emplace_back(input_dims[i]);
  }
1907
  if (keepdim && dim_size > 0) {
1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921
    dimvec.emplace_back(static_cast<int64_t>(1));
  }
  for (int64_t i = axis + 1; i < dim_size; i++) {
    dimvec.emplace_back(input_dims[i]);
  }
  DDim dims = phi::make_ddim(dimvec);
  out->set_dims(dims);
  out->share_lod(x);
  out->set_dtype(x.dtype());
  indices->set_dims(dims);
  indices->share_lod(x);
  indices->set_dtype(x.dtype());
}

1922 1923 1924 1925 1926
void LogsumexpInferMeta(const MetaTensor& input,
                        const std::vector<int64_t>& axis,
                        bool keepdim,
                        bool reduce_all,
                        MetaTensor* out) {
1927 1928 1929 1930 1931 1932 1933 1934
  auto input_rank = input.dims().size();
  // only supoort 0~4D, due to eigen template compile slow
  PADDLE_ENFORCE_LE(
      input_rank,
      4,
      errors::InvalidArgument("The input tensor X's dimensions of logsumexp "
                              "should be less or equal than 4. "));
  ReduceInferMetaBase(input, axis, keepdim, reduce_all, out);
1935 1936
}

1937 1938 1939 1940 1941 1942 1943 1944 1945
void MatrixPowerInferMeta(const MetaTensor& x, int n, MetaTensor* out) {
  auto dims = x.dims();
  auto n_dim = dims.size();
  PADDLE_ENFORCE_GE(n_dim,
                    2,
                    phi::errors::InvalidArgument(
                        "The Input(X) should have at least 2 dimensions. But "
                        "received a %d dimension tensor.",
                        n_dim));
1946 1947 1948 1949 1950
  for (int i = 0; i < n_dim; ++i)
    PADDLE_ENFORCE_NE(
        dims[i],
        0,
        phi::errors::InvalidArgument("The size of Input(X) should not be 0."));
1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963
  PADDLE_ENFORCE_EQ(dims[n_dim - 2],
                    dims[n_dim - 1],
                    phi::errors::InvalidArgument(
                        "The inner-most 2 dimensions of Input(X) all should "
                        "be square matrices "
                        "But received X's shape[-2] = %d and shape[-1] = %d.",
                        dims[n_dim - 2],
                        dims[n_dim - 1]));
  out->set_dims(dims);
  out->share_lod(x);
  out->set_dtype(x.dtype());
}

L
Lin Manhui 已提交
1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007
void LUInferMeta(const MetaTensor& x,
                 bool pivot,
                 MetaTensor* out,
                 MetaTensor* pivots,
                 MetaTensor* infos) {
  auto x_dims = x.dims();
  int x_rank = x_dims.size();

  PADDLE_ENFORCE_NOT_NULL(
      out, phi::errors::InvalidArgument("Output(Out) should not be nullptr."));
  PADDLE_ENFORCE_GE(
      x_rank,
      2,
      phi::errors::InvalidArgument("The rank of input must greater than 2."));
  out->set_dims(x_dims);
  out->set_dtype(x.dtype());
  int m = x_dims[x_rank - 1];
  int n = x_dims[x_rank - 2];
  int min_mn = std::min(m, n);
  auto dims_vec = phi::vectorize(x_dims);
  PADDLE_ENFORCE_NOT_NULL(
      infos,
      phi::errors::InvalidArgument("Output(Infos) should not be nullptr."));
  if (x_rank == 2) {
    auto Infos_dim = std::vector<int>(1);
    infos->set_dims(phi::make_ddim(Infos_dim));
  } else {
    auto Infos_dim =
        std::vector<int>(dims_vec.begin(), dims_vec.begin() + x_rank - 2);
    infos->set_dims(phi::make_ddim(Infos_dim));
  }
  infos->set_dtype(DataType::INT32);
  if (pivot) {
    PADDLE_ENFORCE_NOT_NULL(
        pivots,
        phi::errors::InvalidArgument("Output(Pivots) should not be nullptr."));
    auto Pivots_dim =
        std::vector<int>(dims_vec.begin(), dims_vec.begin() + x_rank - 1);
    Pivots_dim[x_rank - 2] = min_mn;
    pivots->set_dims(phi::make_ddim(Pivots_dim));
    pivots->set_dtype(DataType::INT32);
  }
}

2008
void MatrixRankInferMeta(const MetaTensor& x,
2009
                         bool use_default_tol,
Z
zhangyuqin1998 已提交
2010
                         bool hermitian,
2011 2012
                         MetaTensor* out) {
  auto dim_x = x.dims();
L
Lin Manhui 已提交
2013 2014 2015 2016
  PADDLE_ENFORCE_GE(dim_x.size(),
                    2,
                    phi::errors::InvalidArgument(
                        "The dims of input must be greater than 2."));
2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030

  if (hermitian) {
    int rows = dim_x[dim_x.size() - 2];
    int cols = dim_x[dim_x.size() - 1];
    PADDLE_ENFORCE_EQ(rows,
                      cols,
                      phi::errors::InvalidArgument(
                          "if hermitian == true, matrix should be n*n"));
  }
  DDim dim_x_batch = detail::CheckAndGetOutputDim(dim_x);
  out->set_dims(dim_x_batch);
  out->share_lod(x);
}

2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046
void MaxOutInferMeta(const MetaTensor& x,
                     int groups,
                     int axis,
                     MetaTensor* out) {
  auto in_x_dims = x.dims();
  // check groups > 1
  PADDLE_ENFORCE_GT(
      groups,
      1,
      phi::errors::InvalidArgument("Attr(groups) of Op(maxout) should be "
                                   "larger than 1. But received %d.",
                                   groups));
  PADDLE_ENFORCE_EQ(
      axis == 1 || axis == -1 || axis == 3,
      true,
      phi::errors::InvalidArgument(
L
Lin Manhui 已提交
2047
          "axis only supported 1, -1 or 3, but recevied axis is: %d.", axis));
2048 2049 2050
  PADDLE_ENFORCE_EQ(in_x_dims.size(),
                    4,
                    phi::errors::InvalidArgument(
L
Lin Manhui 已提交
2051
                        "x's dims should be 4, but received x's dims is: %d.",
2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076
                        in_x_dims.size()));

  if (axis < 0) {
    axis += in_x_dims.size();
  }
  PADDLE_ENFORCE_EQ(
      in_x_dims[axis] % groups,
      0,
      phi::errors::InvalidArgument(
          "The number of input channels for Op(maxout) "
          "should be divisible by Attr(groups). But received: the "
          "input's channels is [%d], the shape of input is [%s], "
          "the Attr(groups) is [%d], the Attr(axis) is [%d]. The "
          "error may come from wrong Attr(groups) or Attr(axis) setting.",
          in_x_dims[axis],
          in_x_dims,
          groups,
          axis));
  std::vector<int64_t> output_shape(
      {in_x_dims[0], in_x_dims[1], in_x_dims[2], in_x_dims[3]});
  output_shape[axis] = in_x_dims[axis] / groups;
  out->set_dims(phi::make_ddim(output_shape));
  out->set_dtype(x.dtype());
}

F
From00 已提交
2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090
void MaxPoolWithIndexInferMeta(const MetaTensor& x,
                               const std::vector<int>& kernel_size,
                               const std::vector<int>& strides,
                               const std::vector<int>& paddings,
                               bool global_pooling,
                               bool adaptive,
                               MetaTensor* out,
                               MetaTensor* mask,
                               MetaConfig config) {
  std::vector<int> paddings_ = paddings;
  std::vector<int> kernel_size_ = kernel_size;

  auto x_dims = x.dims();

2091 2092 2093 2094 2095 2096
  PADDLE_ENFORCE_EQ(
      (x_dims.size() == 4 || x_dims.size() == 5),
      true,
      errors::InvalidArgument("Pooling intput should be 4-D or "
                              "5-D tensor but received %dD-Tensor",
                              x_dims.size()));
F
From00 已提交
2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146

  if (global_pooling) {
    kernel_size_.resize(static_cast<size_t>(x_dims.size()) - 2);
    for (size_t i = 0; i < kernel_size_.size(); ++i) {
      paddings_[i] = 0;
      kernel_size_[i] = static_cast<int>(x_dims[i + 2]);
    }
  }

  PADDLE_ENFORCE_EQ(
      x_dims.size() - kernel_size_.size(),
      2U,
      errors::InvalidArgument(
          "The input size %d minus the kernel size %d should equal to 2.",
          x_dims.size(),
          kernel_size_.size()));
  PADDLE_ENFORCE_EQ(
      kernel_size_.size(),
      strides.size(),
      errors::InvalidArgument(
          "Strides size %d and pooling size %d should be the same.",
          strides.size(),
          kernel_size_.size()));
  PADDLE_ENFORCE_EQ(
      kernel_size_.size(),
      paddings_.size(),
      errors::InvalidArgument(
          "Paddings size %d and pooling size %d should be the same.",
          paddings_.size(),
          kernel_size_.size()));

  std::vector<int64_t> output_shape({x_dims[0], x_dims[1]});
  if (adaptive) {
    output_shape.insert(
        output_shape.end(), kernel_size_.begin(), kernel_size_.end());
  } else {
    for (size_t i = 0; i < kernel_size_.size(); ++i) {
      if ((!config.is_runtime) && (x_dims[i + 2] < 0)) {
        output_shape.push_back(x_dims[i + 2]);
      } else {
        output_shape.push_back(funcs::MaxPoolOutputSize(
            x_dims[i + 2], kernel_size_[i], paddings_[i], strides[i]));
      }
    }
  }

  out->set_dims(make_ddim(output_shape));
  out->set_dtype(x.dtype());

  mask->set_dims(make_ddim(output_shape));
2147
  mask->set_dtype(phi::CppTypeToDataType<int>::Type());
F
From00 已提交
2148 2149
}

2150
void MeanAllInferMeta(const MetaTensor& x, MetaTensor* out) {
2151
  out->set_dims(phi::make_ddim({}));
2152 2153 2154 2155
  out->set_dtype(x.dtype());
  out->set_layout(x.layout());
}

2156 2157 2158 2159 2160 2161 2162
void ModeInferMeta(const MetaTensor& x,
                   int axis,
                   bool keepdim,
                   MetaTensor* out,
                   MetaTensor* indices) {
  auto input_dims = x.dims();
  const int& dim_size = input_dims.size();
2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178
  PADDLE_ENFORCE_LE(axis,
                    dim_size,
                    phi::errors::InvalidArgument(
                        "the axis must be [-%d, %d), but received %d .",
                        dim_size,
                        dim_size,
                        axis));
  if (dim_size > 0) {
    PADDLE_ENFORCE_GE(axis,
                      -dim_size,
                      phi::errors::InvalidArgument(
                          "the axis must be [-%d, %d), but received %d .",
                          dim_size,
                          dim_size,
                          axis));
  }
2179 2180
  PADDLE_ENFORCE_GE(
      input_dims.size(),
2181 2182
      0,
      errors::InvalidArgument("input of ModeOp must have >= 0d shape"));
2183 2184 2185 2186 2187
  if (axis < 0) axis += dim_size;
  std::vector<int64_t> dimvec;
  for (int64_t i = 0; i < axis; i++) {
    dimvec.emplace_back(input_dims[i]);
  }
2188
  if (keepdim && dim_size > 0) {
2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203
    dimvec.emplace_back(static_cast<int64_t>(1));
  }
  for (int64_t i = axis + 1; i < dim_size; i++) {
    dimvec.emplace_back(input_dims[i]);
  }
  DDim dims = phi::make_ddim(dimvec);
  out->set_dims(dims);
  out->share_lod(x);
  out->set_dtype(x.dtype());

  indices->set_dims(dims);
  indices->share_lod(x);
  indices->set_dtype(x.dtype());
}

2204
void MultinomialInferMeta(const MetaTensor& x,
2205
                          const Scalar& num_samples,
2206
                          bool replacement,
2207 2208 2209
                          MetaTensor* out,
                          MetaConfig config) {
  auto int_num_samples = num_samples.to<int>();
2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229
  auto x_dim = x.dims();
  int64_t x_rank = x_dim.size();
  PADDLE_ENFORCE_GT(x_rank,
                    0,
                    errors::InvalidArgument(
                        "The number of dimensions of the input probability "
                        "distribution should be > 0, but got %d.",
                        x_rank));
  PADDLE_ENFORCE_LE(x_rank,
                    2,
                    errors::InvalidArgument(
                        "The number of dimensions of the input probability "
                        "distribution should be <= 2, but got %d.",
                        x_rank));

  std::vector<int64_t> out_dims(x_rank);
  for (int64_t i = 0; i < x_rank - 1; i++) {
    out_dims[i] = x_dim[i];
  }

2230 2231 2232 2233 2234 2235 2236 2237 2238 2239
  if (config.is_runtime || !num_samples.FromTensor()) {
    PADDLE_ENFORCE_GT(int_num_samples,
                      0,
                      errors::InvalidArgument(
                          "The number of samples should be > 0, but got %d.",
                          int_num_samples));
    out_dims[x_rank - 1] = int_num_samples;
  } else {
    out_dims[x_rank - 1] = -1;
  }
2240 2241 2242 2243 2244

  out->set_dims(make_ddim(out_dims));
  out->set_dtype(DataType::INT64);
}

2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264
void NanmedianInferMeta(const MetaTensor& x,
                        const IntArray& axes,
                        bool keep_dim,
                        MetaTensor* out,
                        MetaTensor* median_index) {
  std::vector<int64_t> axis_list = axes.GetData();
  auto x_dim = x.dims();
  int64_t x_rank = x_dim.size();
  out->set_dtype(x.dtype());
  median_index->set_dtype(DataType::INT64);
  median_index->set_dims(make_ddim({x.numel() * 2}));

  std::vector<int32_t> out_dim;
  if (axis_list.empty()) {
    if (keep_dim) {
      for (int64_t i = 0; i < x_rank; i++) {
        out_dim.push_back(1);
      }
    }
  } else {
2265
    std::vector<int64_t> formated_axis;
2266
    for (auto& axis : axis_list) {
2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290
      if (x_rank == 0) {
        PADDLE_ENFORCE_EQ(axis == 0 || axis == -1,
                          true,
                          phi::errors::InvalidArgument(
                              "When input 0D Tensor, each element of the axis "
                              "can only be -1, 0, None"));
      } else {
        PADDLE_ENFORCE_LT(axis,
                          x_rank,
                          errors::InvalidArgument(
                              "each element of the axis should be in the "
                              "range [ -dimension(X), dimension(X) ) "
                              "which dimesion = %d. But received axis = %d.",
                              x_rank,
                              axis));
        PADDLE_ENFORCE_GE(axis,
                          -x_rank,
                          errors::InvalidArgument(
                              "each element of the axis should be in the "
                              "range [ -dimension(X), dimension(X) ) "
                              "which dimesion = %d. But received axis = %d.",
                              x_rank,
                              axis));
      }
2291 2292
      if (axis < 0) axis += x_rank;
      PADDLE_ENFORCE_EQ(
2293 2294
          std::find(formated_axis.begin(), formated_axis.end(), axis),
          formated_axis.end(),
2295 2296 2297
          errors::InvalidArgument("Attr(axes) has duplicated elements: %d.",
                                  static_cast<int>(axis)));

2298
      formated_axis.push_back(axis);
2299 2300 2301
    }

    for (int64_t i = 0; i < x_rank; i++) {
2302 2303
      if (std::find(formated_axis.begin(), formated_axis.end(), i) ==
          formated_axis.end()) {
2304 2305 2306 2307 2308 2309 2310 2311 2312 2313
        out_dim.push_back(x_dim[i]);
      } else if (keep_dim) {
        out_dim.push_back(1);
      }
    }
  }

  out->set_dims(make_ddim(out_dim));
}

2314 2315 2316 2317 2318 2319 2320 2321 2322
void NMSInferMeta(const MetaTensor& x, float threshold, MetaTensor* out) {
  auto boxes_dim = x.dims();
  PADDLE_ENFORCE_EQ(boxes_dim.size(),
                    2,
                    phi::errors::InvalidArgument(
                        "The Input Boxes must be 2-dimention "
                        "whose shape must be [N, 4] "
                        "N is the number of boxes "
                        "in last dimension in format [x1, x2, y1, y2]. "));
2323 2324
  out->set_dims(phi::make_ddim({-1}));
  out->set_dtype(DataType::INT64);
2325 2326
}

2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337
void NonZeroInferMeta(const MetaTensor& condition, MetaTensor* out) {
  auto rank = condition.dims().size();
  PADDLE_ENFORCE_GE(
      rank,
      1UL,
      phi::errors::InvalidArgument(
          "Input(Condition) should have number of dimension at least 1"));
  out->set_dims(phi::make_ddim({-1, rank}));
  out->set_dtype(DataType::INT64);
}

H
hong 已提交
2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355
void NormInferMeta(const MetaTensor& x,
                   int axis,
                   float epsilon,
                   bool is_test,
                   MetaTensor* out,
                   MetaTensor* norm) {
  auto xdim = x.dims();
  out->set_dims(x.dims());
  out->set_dtype(x.dtype());

  if (is_test == false) {
    if (axis < 0) axis = xdim.size() + axis;
    xdim[axis] = 1;
    norm->set_dims(xdim);
    norm->set_dtype(x.dtype());
  }
}

2356 2357 2358 2359 2360 2361 2362 2363
void OneHotRawInferMeta(const MetaTensor& x,
                        const Scalar& depth,
                        DataType dtype,
                        bool allow_out_of_range,
                        MetaTensor* out) {
  auto x_dims = x.dims();
  PADDLE_ENFORCE_GE(
      x_dims.size(),
2364 2365
      0,
      phi::errors::InvalidArgument("Rank of Input(X) should be at least 0."));
2366 2367 2368 2369 2370 2371 2372
  auto out_dims_vec = phi::vectorize(x_dims);
  out_dims_vec.push_back(depth.to<int>());
  auto out_dims = phi::make_ddim(out_dims_vec);
  out->set_dims(out_dims);
  out->share_lod(x);
  out->set_dtype(dtype);
}
2373

2374 2375 2376 2377
void OneHotInferMeta(const MetaTensor& x,
                     const Scalar& depth_t,
                     MetaTensor* out) {
  auto x_dims = x.dims();
2378
  PADDLE_ENFORCE_GE(
2379
      x_dims.size(),
2380 2381
      0,
      phi::errors::InvalidArgument("Rank of Input(X) should be at least 0."));
2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402

  int depth = depth_t.to<int>();
  auto out_dims_vec = phi::vectorize(x_dims);
  out_dims_vec.push_back(depth);
  auto out_dims = phi::make_ddim(out_dims_vec);
  out->set_dims(out_dims);
  out->share_lod(x);

  out->set_dtype(phi::DataType::FLOAT32);
}

void OverlapAddInferMeta(const MetaTensor& x,
                         int hop_length,
                         int axis,
                         MetaTensor* out,
                         MetaConfig config) {
  const auto x_dims = x.dims();
  const int x_rank = x_dims.size();

  PADDLE_ENFORCE_GE(
      x_rank,
2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476
      2,
      errors::InvalidArgument(
          "Input(X) of OverlapAddOp should be a tensor which contains "
          "at least 2 dimensions, but got rank %s.",
          x_rank));

  PADDLE_ENFORCE_GT(
      hop_length,
      0,
      errors::InvalidArgument(
          "Attribute(hop_length) of OverlapAddOp should be greater "
          "than 0, but got %s.",
          hop_length));

  PADDLE_ENFORCE_EQ(
      (axis == 0 || axis == -1),
      true,
      errors::InvalidArgument(
          "Attribute(axis) of OverlapAddOp should 0 or -1, but got %s.", axis));

  std::vector<int64_t> output_shape;
  int n_frames;
  int frame_length;
  int seq_length;

  int start_axis;
  int end_axis;
  if (axis == 0) {
    n_frames = x_dims[0];
    frame_length = x_dims[1];
    start_axis = 2;
    end_axis = x_rank - 1;
  } else {
    n_frames = x_dims[x_rank - 1];
    frame_length = x_dims[x_rank - 2];
    start_axis = 0;
    end_axis = x_rank - 3;
  }

  bool contain_unknown_dim = phi::contain_unknown_dim(x_dims);
  bool check = config.is_runtime || !contain_unknown_dim;
  if (check) {
    PADDLE_ENFORCE_LE(
        hop_length,
        frame_length,
        errors::InvalidArgument(
            "Attribute(hop_length) of OverlapAddOp should be less or equal "
            "than frame_length, but got hop_length(%s) > frame_length(%s).",
            hop_length,
            frame_length));
  }

  if (n_frames == -1) {
    seq_length = -1;
  } else {
    seq_length = (n_frames - 1) * hop_length + frame_length;
  }

  // It won't go into for loop when x_rank == 2U.
  for (int i = start_axis; i <= end_axis; i++) {
    output_shape.push_back(x_dims[i]);
  }

  if (axis == 0) {
    // (seq_length, ...)
    output_shape.insert(output_shape.begin(), seq_length);
  } else {
    // (..., seq_length)
    output_shape.push_back(seq_length);
  }

  out->set_dims(phi::make_ddim(output_shape));
}

Z
zyfncg 已提交
2477 2478
void PadInferMeta(const MetaTensor& input,
                  const std::vector<int>& paddings,
2479
                  const Scalar& padding_value,
Z
zyfncg 已提交
2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499
                  MetaTensor* out,
                  MetaConfig config) {
  auto x_dim = input.dims();
  PADDLE_ENFORCE_EQ(
      static_cast<int>(paddings.size()),
      x_dim.size() * 2,
      phi::errors::InvalidArgument(
          "Size of 'paddings' dimension should be equal to 2 * size of "
          "Input(X)'s dimension, but received (size of 'paddings' dimension "
          "is) %d vs (2 * size of Input(X)'s dimension is) %d.",
          static_cast<int>(paddings.size()),
          x_dim.size() * 2));
  for (size_t i = 0; i < paddings.size(); ++i) {
    PADDLE_ENFORCE_GE(paddings[i],
                      0,
                      phi::errors::InvalidArgument(
                          "The element of 'paddings' should >= 0, but "
                          "received %d for index %d.",
                          paddings[i],
                          static_cast<int>(i)));
2500
  }
Z
zyfncg 已提交
2501 2502 2503 2504
  std::vector<int64_t> out_dims(x_dim.size());
  for (int i = 0; i < x_dim.size(); ++i) {
    if ((!config.is_runtime) && (x_dim[i] == -1)) {
      out_dims[i] = -1;
2505
    } else {
Z
zyfncg 已提交
2506
      out_dims[i] = x_dim[i] + paddings[i * 2] + paddings[i * 2 + 1];
2507 2508
    }
  }
Z
zyfncg 已提交
2509 2510 2511 2512 2513
  out->set_dims(phi::make_ddim(out_dims));
  if (out_dims[0] == x_dim[0]) {
    // Only pass LoD when the first dimension is equal between
    // output and input.
    out->share_lod(input);
2514
  }
Z
zyfncg 已提交
2515
  out->set_dtype(input.dtype());
2516 2517
}

2518
void Pad3dInferMeta(const MetaTensor& x,
2519
                    const IntArray& paddings_int_array,
2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532
                    const std::string& mode,
                    float value,
                    const std::string& data_format,
                    MetaTensor* out,
                    MetaConfig config) {
  auto x_dim = x.dims();
  PADDLE_ENFORCE_EQ(x_dim.size(),
                    5,
                    errors::InvalidArgument(
                        "The size of Input(X)'s dimension should be equal to "
                        "5, but received %d. ",
                        x_dim.size()));

2533
  std::vector<int64_t> out_dims(x_dim.size(), -1);
2534
  out_dims[0] = x_dim[0];
2535 2536 2537 2538 2539 2540
  auto& paddings = paddings_int_array.GetData();
  if (data_format == "NCDHW") {
    out_dims[1] = x_dim[1];
  } else {
    out_dims[4] = x_dim[4];
  }
2541
  if (paddings_int_array.FromTensor()) {
2542 2543
    if (config.is_runtime) {
      PADDLE_ENFORCE_EQ(
2544
          paddings.size(),
2545 2546 2547
          6,
          errors::InvalidArgument("Shape of Input(Paddings) should be equal to "
                                  "[6], but received [%d].",
2548 2549 2550 2551 2552 2553 2554 2555 2556 2557
                                  paddings.size()));
      if (data_format == "NCDHW") {
        out_dims[2] = x_dim[2] + paddings[4] + paddings[5];
        out_dims[3] = x_dim[3] + paddings[2] + paddings[3];
        out_dims[4] = x_dim[4] + paddings[0] + paddings[1];
      } else {
        out_dims[1] = x_dim[1] + paddings[4] + paddings[5];
        out_dims[2] = x_dim[2] + paddings[2] + paddings[3];
        out_dims[3] = x_dim[3] + paddings[0] + paddings[1];
      }
2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595
    }
  } else {
    PADDLE_ENFORCE_EQ(
        paddings.size(),
        6,
        errors::InvalidArgument(
            "Size of paddings should be equal to 6, but received %d.",
            static_cast<int>(paddings.size())));
    if (data_format == "NCDHW") {
      out_dims[2] = ((!config.is_runtime) && (x_dim[2] < 0))
                        ? x_dim[2]
                        : (x_dim[2] + paddings[4] + paddings[5]);  // depth

      out_dims[3] = ((!config.is_runtime) && (x_dim[3] < 0))
                        ? x_dim[3]
                        : (x_dim[3] + paddings[2] + paddings[3]);  // height

      out_dims[4] = ((!config.is_runtime) && (x_dim[4] < 0))
                        ? x_dim[4]
                        : (x_dim[4] + paddings[0] + paddings[1]);  // width
    } else {                                                       // NDHWC
      out_dims[1] = ((!config.is_runtime) && (x_dim[1] < 0))
                        ? x_dim[1]
                        : (x_dim[1] + paddings[4] + paddings[5]);  // depth
      out_dims[2] = ((!config.is_runtime) && (x_dim[2] < 0))
                        ? x_dim[2]
                        : (x_dim[2] + paddings[2] + paddings[3]);  // height
      out_dims[3] = ((!config.is_runtime) && (x_dim[3] < 0))
                        ? x_dim[3]
                        : (x_dim[3] + paddings[0] + paddings[1]);  // width
    }
  }

  out->set_dims(phi::make_ddim(out_dims));
  out->set_dtype(x.dtype());
  out->share_lod(x);
}

Z
zyfncg 已提交
2596 2597 2598 2599 2600 2601 2602
void PixelShuffleInferMeta(const MetaTensor& x,
                           int upscale_factor,
                           const std::string& data_format,
                           MetaTensor* out) {
  auto input_dims = x.dims();
  PADDLE_ENFORCE_EQ(input_dims.size(),
                    4,
2603
                    phi::errors::InvalidArgument(
Z
zyfncg 已提交
2604 2605 2606
                        "Input should be a 4-D tensor of format [N, C, H, W] "
                        "or [N, H, W, C], but got %u.",
                        input_dims.size()));
2607 2608 2609 2610
  PADDLE_ENFORCE_NE(
      upscale_factor,
      0,
      phi::errors::InvalidArgument("upscale_factor should not be 0."));
2611

Z
zyfncg 已提交
2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629
  const bool channel_last = (data_format == "NHWC");

  if (!channel_last) {
    PADDLE_ENFORCE_EQ(input_dims[1] % (upscale_factor * upscale_factor),
                      0,
                      phi::errors::InvalidArgument(
                          "The square of upscale_factor[%u] should divide the "
                          "number of channel[%u]",
                          upscale_factor * upscale_factor,
                          input_dims[1]));
  } else {
    PADDLE_ENFORCE_EQ(input_dims[3] % (upscale_factor * upscale_factor),
                      0,
                      phi::errors::InvalidArgument(
                          "The square of upscale_factor[%u] should divide the "
                          "number of channel[%u]",
                          upscale_factor * upscale_factor,
                          input_dims[3]));
2630
  }
Z
zyfncg 已提交
2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643
  auto output_dims = input_dims;
  output_dims[0] = input_dims[0];
  if (!channel_last) {
    output_dims[1] = input_dims[1] / (upscale_factor * upscale_factor);
    output_dims[2] = input_dims[2] * upscale_factor;
    output_dims[3] = input_dims[3] * upscale_factor;
  } else {
    output_dims[1] = input_dims[1] * upscale_factor;
    output_dims[2] = input_dims[2] * upscale_factor;
    output_dims[3] = input_dims[3] / (upscale_factor * upscale_factor);
  }
  out->set_dtype(x.dtype());
  out->set_dims(output_dims);
2644 2645
}

H
hong 已提交
2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675
void PixelShuffleGradInferMeta(const MetaTensor& out_grad,
                               int upscale_factor,
                               const std::string& data_format,
                               MetaTensor* x_grad) {
  auto do_dims = out_grad.dims();
  PADDLE_ENFORCE_EQ(do_dims.size(),
                    4,
                    phi::errors::InvalidArgument(
                        "Input should be a 4-D tensor of format [N, C, H, W] "
                        "or [N, H, W, C], but got %u.",
                        do_dims.size()));

  const bool channel_last = (data_format == "NHWC");

  auto dx_dims = do_dims;
  dx_dims[0] = do_dims[0];

  if (!channel_last) {
    dx_dims[1] = do_dims[1] * (upscale_factor * upscale_factor);
    dx_dims[2] = do_dims[2] / upscale_factor;
    dx_dims[3] = do_dims[3] / upscale_factor;
  } else {
    dx_dims[1] = do_dims[1] / upscale_factor;
    dx_dims[2] = do_dims[2] / upscale_factor;
    dx_dims[3] = do_dims[3] * (upscale_factor * upscale_factor);
  }
  x_grad->set_dims(dx_dims);
  x_grad->set_dtype(out_grad.dtype());
}

2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735
void PixelUnshuffleInferMeta(const MetaTensor& x,
                             int downscale_factor,
                             const std::string& data_format,
                             MetaTensor* out) {
  auto input_dims = x.dims();
  PADDLE_ENFORCE_EQ(input_dims.size(),
                    4,
                    phi::errors::InvalidArgument(
                        "Input should be a 4-D tensor of format [N, C, H, W] "
                        "or [N, H, W, C], but got %u.",
                        input_dims.size()));
  PADDLE_ENFORCE_GE(downscale_factor,
                    1,
                    phi::errors::InvalidArgument(
                        "downscale_factor should be larger than 0."));
  PADDLE_ENFORCE_EQ(data_format == "NCHW" || data_format == "NHWC",
                    true,
                    phi::errors::InvalidArgument(
                        "data_format must be one of "
                        "NCHW and NHWC. But recevied data_format: %s",
                        data_format));

  const bool channel_last = (data_format == "NHWC");

  if (!channel_last) {
    PADDLE_ENFORCE_EQ(
        (input_dims[2] % downscale_factor) == 0 &&
            (input_dims[3] % downscale_factor) == 0,
        true,
        phi::errors::InvalidArgument("Downscale factor[%u] should divide both "
                                     "height[%u] and width[%u]",
                                     downscale_factor,
                                     input_dims[2],
                                     input_dims[3]));
  } else {
    PADDLE_ENFORCE_EQ(
        (input_dims[1] % downscale_factor) == 0 &&
            (input_dims[2] % downscale_factor) == 0,
        true,
        phi::errors::InvalidArgument("Downscale factor[%u] should divide both "
                                     "height[%u] and width[%u]",
                                     downscale_factor,
                                     input_dims[1],
                                     input_dims[2]));
  }
  auto output_dims = input_dims;
  output_dims[0] = input_dims[0];
  if (!channel_last) {
    output_dims[1] = input_dims[1] * (downscale_factor * downscale_factor);
    output_dims[2] = input_dims[2] / downscale_factor;
    output_dims[3] = input_dims[3] / downscale_factor;
  } else {
    output_dims[1] = input_dims[1] / downscale_factor;
    output_dims[2] = input_dims[2] / downscale_factor;
    output_dims[3] = input_dims[3] * (downscale_factor * downscale_factor);
  }
  out->set_dtype(x.dtype());
  out->set_dims(output_dims);
}

2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764
void PNormInferMeta(const MetaTensor& x,
                    float porder,
                    int axis,
                    float epsilon,
                    bool keepdim,
                    bool asvector,
                    MetaTensor* out) {
  auto x_dim = x.dims();
  auto x_rank = x_dim.size();

  PADDLE_ENFORCE_GE(axis,
                    -x_rank,
                    errors::InvalidArgument(
                        "Attr(axis) value should be in range [-R, R-1], R is "
                        "the rank of Input(X). But received axis: %d, R: %d. "
                        "Current Input(X)'s shape is=[%s].",
                        axis,
                        x_rank,
                        x_dim));
  PADDLE_ENFORCE_LT(axis,
                    x_rank,
                    errors::InvalidArgument(
                        "Attr(axis) value should be in range [-R, R-1], R is "
                        "the rank of Input(X). But received axis: %d, R: %d. "
                        "Current Input(X)'s shape is=[%s].",
                        axis,
                        x_rank,
                        x_dim));

2765
  std::vector<int> out_dim_vector;
2766 2767
  if (asvector) {
    if (keepdim) {
2768 2769
      for (int i = 0; i < x_rank; ++i) {
        out_dim_vector.emplace_back(1);
2770
      }
2771 2772
    } else {
      out_dim_vector = {};
2773 2774
    }
  } else {
2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787
    if (axis < 0) axis = axis + x_rank;
    if (keepdim) {
      for (int i = 0; i < x_dim.size(); ++i) {
        if (i != axis) {
          out_dim_vector.emplace_back(x_dim[i]);
        } else {
          out_dim_vector.emplace_back(1);
        }
      }
    } else {
      for (int i = 0; i < x_dim.size(); ++i) {
        if (i != axis) out_dim_vector.emplace_back(x_dim[i]);
      }
2788 2789 2790
    }
  }

2791
  out->set_dims(phi::make_ddim(out_dim_vector));
2792 2793 2794
  out->set_dtype(x.dtype());
}

2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841
void Pool2DInferMeta(const MetaTensor& x,
                     const IntArray& kernel_size,
                     const std::vector<int>& strides,
                     const std::vector<int>& paddings,
                     bool ceil_mode,
                     bool exclusive,
                     const std::string& data_format,
                     const std::string& pooling_type,
                     bool global_pooling,
                     bool adaptive,
                     const std::string& padding_algorithm,
                     MetaTensor* out,
                     MetaConfig config) {
  const bool channel_last = (config.is_run_mkldnn_kernel == false) &&
                            (data_format == "NHWC" || data_format == "NDHWC");
  if (!config.is_runtime && kernel_size.FromTensor()) {
    auto x_dims = x.dims();
    std::vector<int64_t> output_shape = std::move(phi::vectorize(x_dims));
    // set dims of HW -1
    output_shape[x_dims.size() - 2] = -1;
    if (channel_last) {  // for NHWC, NDHWC
      output_shape[x_dims.size() - 3] = -1;
    } else {  // for NCHW
      output_shape[x_dims.size() - 1] = -1;
    }
    out->set_dims(make_ddim(output_shape));
    out->share_lod(x);
    out->set_dtype(x.dtype());
  } else {
    std::vector<int> kernel_size_val(kernel_size.GetData().begin(),
                                     kernel_size.GetData().end());
    PoolInferMeta(x,
                  kernel_size_val,
                  strides,
                  paddings,
                  ceil_mode,
                  exclusive,
                  data_format,
                  pooling_type,
                  global_pooling,
                  adaptive,
                  padding_algorithm,
                  out,
                  config);
  }
}

2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858
void PSendInferMeta(const MetaTensor& x, int peer) {
  LOG(INFO) << "SendBaseInferMeta begin";
  PADDLE_ENFORCE_GE(
      peer,
      0,
      errors::InvalidArgument(
          "The peer (%d) for p_send op must be non-negative.", peer));
}

void PSendArrayInferMeta(const MetaTensor& x, int peer) {
  PADDLE_ENFORCE_GE(
      peer,
      0,
      errors::InvalidArgument(
          "The peer (%d) for p_send op must be non-negative.", peer));
}

F
From00 已提交
2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912 2913 2914 2915 2916 2917 2918 2919 2920 2921 2922 2923 2924 2925 2926 2927 2928 2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942 2943 2944 2945 2946 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970
void PoolInferMeta(const MetaTensor& x,
                   const std::vector<int>& kernel_size,
                   const std::vector<int>& strides,
                   const std::vector<int>& paddings,
                   bool ceil_mode,
                   bool exclusive,
                   const std::string& data_format,
                   const std::string& pooling_type,
                   bool global_pooling,
                   bool adaptive,
                   const std::string& padding_algorithm,
                   MetaTensor* out,
                   MetaConfig config) {
  std::vector<int> paddings_ = paddings;
  std::vector<int> kernel_size_ = kernel_size;

  auto x_dims = x.dims();
  PADDLE_ENFORCE_EQ(
      x_dims.size() == 4 || x_dims.size() == 5,
      true,
      errors::InvalidArgument(
          "the input of Op(pool) should be 4-D or 5-D Tensor. But "
          "received: %u-D Tensor and it's shape is [%s].",
          x_dims.size(),
          x_dims));

  PADDLE_ENFORCE_EQ(x_dims.size() - kernel_size_.size(),
                    2U,
                    errors::InvalidArgument(
                        "the dimension of input minus the size of "
                        "Attr(kernel_size_) must be euqal to 2 in Op(pool). "
                        "But received: the dimension of input minus the size "
                        "of Attr(kernel_size_) is %d, the "
                        "input's dimension is %d, the shape of input "
                        "is [%s], the Attr(kernel_size_)'s size is %d, the "
                        "Attr(kernel_size_) is [%s].",
                        x_dims.size() - kernel_size_.size(),
                        x_dims.size(),
                        x_dims,
                        kernel_size_.size(),
                        make_ddim(kernel_size_)));

  PADDLE_ENFORCE_EQ(
      kernel_size_.size(),
      strides.size(),
      errors::InvalidArgument(
          "the size of Attr(kernel_size_) and Attr(strides) in "
          "Op(pool) must be equal. "
          "But received: Attr(kernel_size_)'s size is %d, Attr(strides)'s "
          "size is %d, Attr(kernel_size_) is [%s], Attr(strides)is [%s].",
          kernel_size_.size(),
          strides.size(),
          make_ddim(kernel_size_),
          make_ddim(strides)));

  // MKL-DNN Kernels are using NCHW order of dims description
  // so we ignore data_format consideration for MKL-DNN kernel
  const bool channel_last = (config.is_run_mkldnn_kernel == false) &&
                            (data_format == "NHWC" || data_format == "NDHWC");

  // update paddings if "SAME" or global_pooling
  DDim data_dims;
  if (channel_last) {
    data_dims = slice_ddim(x_dims, 1, x_dims.size() - 1);
  } else {
    data_dims = slice_ddim(x_dims, 2, x_dims.size());
  }
  funcs::UpdatePadding(&paddings_,
                       global_pooling,
                       adaptive,
                       padding_algorithm,
                       data_dims,
                       strides,
                       kernel_size_);

  if (global_pooling) {
    funcs::UpdateKernelSize(&kernel_size_, data_dims);
  }

  std::vector<int64_t> output_shape;
  if (adaptive) {
    output_shape.insert(
        output_shape.end(), kernel_size_.begin(), kernel_size_.end());
  } else {
    for (int i = 0; i < data_dims.size(); ++i) {
      if ((!config.is_runtime) && (data_dims[i] < 0)) {
        output_shape.push_back(data_dims[i]);
      } else {
        output_shape.push_back(funcs::PoolOutputSize(data_dims[i],
                                                     kernel_size_[i],
                                                     paddings_[2 * i],
                                                     paddings_[2 * i + 1],
                                                     strides[i],
                                                     ceil_mode));
      }
    }
  }

  // output_N = input_N
  output_shape.insert(output_shape.begin(), x_dims[0]);
  // output_C = input_C
  if (channel_last) {
    output_shape.push_back(x_dims[x_dims.size() - 1]);
  } else {
    output_shape.insert(output_shape.begin() + 1, x_dims[1]);
  }

  out->set_dims(make_ddim(output_shape));
  out->share_lod(x);
  out->set_dtype(x.dtype());
}

Z
zyfncg 已提交
2971 2972 2973 2974
void RealAndImagInferMeta(const MetaTensor& x, MetaTensor* out) {
  out->set_dims(x.dims());
  out->set_dtype(dtype::ToReal(x.dtype()));
  out->set_layout(x.layout());
2975 2976
}

2977 2978 2979 2980 2981 2982 2983 2984 2985 2986 2987 2988 2989 2990 2991 2992 2993 2994 2995 2996 2997 2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008 3009 3010 3011 3012 3013 3014
void QrInferMeta(const MetaTensor& x,
                 const std::string& mode,
                 MetaTensor* q,
                 MetaTensor* r) {
  auto x_dims = x.dims();
  int x_rank = x_dims.size();
  PADDLE_ENFORCE_GE(
      x_dims.size(),
      2,
      phi::errors::InvalidArgument("the rank of input must greater than 2"));
  bool compute_q;
  bool reduced_mode;
  int m = x_dims[x_rank - 2];
  int n = x_dims[x_rank - 1];
  int min_mn = std::min(m, n);
  std::tie(compute_q, reduced_mode) = phi::funcs::ParseQrMode(mode);

  if (compute_q) {
    int k = reduced_mode ? min_mn : m;
    auto q_dims_vec = phi::vectorize(x_dims);
    q_dims_vec[q_dims_vec.size() - 1] = k;
    q->set_dims(phi::make_ddim(q_dims_vec));
  } else {
    q->set_dims(phi::make_ddim({0}));
  }

  int k = reduced_mode ? min_mn : m;
  auto r_dims_vec = phi::vectorize(x_dims);
  r_dims_vec[r_dims_vec.size() - 2] = k;
  r_dims_vec[r_dims_vec.size() - 1] = n;
  r->set_dims(phi::make_ddim(r_dims_vec));

  q->share_lod(x);
  r->share_lod(x);
  q->set_dtype(x.dtype());
  r->set_dtype(x.dtype());
}

3015 3016 3017 3018
DDim ReduceInferDim(const MetaTensor& x,
                    const std::vector<int64_t>& axis,
                    bool keep_dim,
                    bool reduce_all) {
3019 3020 3021 3022
  auto x_rank = x.dims().size();

  std::vector<int64_t> formated_axis = axis;
  for (size_t i = 0; i < axis.size(); ++i) {
zhouweiwei2014's avatar
zhouweiwei2014 已提交
3023 3024 3025 3026 3027 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048
    if (x_rank == 0) {
      PADDLE_ENFORCE_EQ(
          axis[i] == 0 || axis[i] == -1,
          true,
          phi::errors::InvalidArgument(
              "When input 0D Tensor, the axis can only be -1, 0, None or []"));
    } else {
      PADDLE_ENFORCE_LT(axis[i],
                        x_rank,
                        errors::InvalidArgument(
                            "The reduce dim index %d should be in the "
                            "range [ -dimension(X), dimension(X) ) "
                            "which dimesion = %d. But received dim index = %d.",
                            i,
                            x_rank,
                            axis[i]));
      PADDLE_ENFORCE_GE(axis[i],
                        -x_rank,
                        errors::InvalidArgument(
                            "The reduce dim index %d should be in the "
                            "range [ -dimension(X), dimension(X) )  "
                            "which dimesion = %d. But received dim index = %d.",
                            i,
                            x_rank,
                            axis[i]));
    }
3049 3050 3051 3052 3053 3054 3055 3056

    if (axis[i] < 0) {
      formated_axis[i] = axis[i] + x_rank;
    }
  }

  bool full_dim = true;
  std::set<int64_t> dims_set(formated_axis.begin(), formated_axis.end());
3057
  for (int64_t i = 0; i < x_rank; ++i) {
3058
    if (dims_set.find(i) == dims_set.end()) {
3059
      full_dim = false;
3060 3061 3062
      break;
    }
  }
3063
  reduce_all = reduce_all || full_dim;
3064 3065

  std::vector<int64_t> out_dim_vector;
3066 3067 3068
  for (int64_t i = 0; i < x_rank; ++i) {
    if (reduce_all || dims_set.find(i) != dims_set.end()) {
      if (keep_dim) {
3069 3070 3071 3072
        out_dim_vector.push_back(1);
      } else {
        continue;
      }
3073 3074
    } else {
      out_dim_vector.push_back(x.dims().at(i));
3075 3076 3077
    }
  }

3078
  DDim out_dim = phi::make_ddim(out_dim_vector);
3079 3080 3081
  return out_dim;
}

3082 3083 3084 3085 3086 3087 3088 3089 3090 3091 3092 3093 3094 3095 3096 3097 3098 3099 3100 3101 3102 3103
void ReduceInferMetaBase(const MetaTensor& x,
                         const std::vector<int64_t>& axis,
                         bool keep_dim,
                         bool reduce_all,
                         MetaTensor* out) {
  DDim out_dim = ReduceInferDim(x, axis, keep_dim, reduce_all);
  out->set_dims(out_dim);
  out->set_dtype(x.dtype());
  out->set_layout(x.layout());
}

void ReduceInferMeta(const MetaTensor& x,
                     const std::vector<int64_t>& axis,
                     bool keep_dim,
                     MetaTensor* out) {
  bool reduce_all = false;
  if (axis.size() == 0) {
    reduce_all = true;
  }
  ReduceInferMetaBase(x, axis, keep_dim, reduce_all, out);
}

3104 3105 3106 3107 3108 3109 3110 3111 3112 3113
DDim ReduceInferDimForIntArrayAxis(const MetaTensor& x,
                                   const IntArray& axis,
                                   bool keep_dim,
                                   bool reduce_all) {
  std::vector<int64_t> vec_axis = axis.GetData();
  std::vector<int64_t> vec_dim;
  if (reduce_all) {
    if (keep_dim) {
      vec_dim = std::vector<int64_t>(x.dims().size(), 1);
    } else {
3114
      vec_dim = {};
3115 3116 3117 3118 3119 3120
    }
  } else {
    if (keep_dim) {
      vec_dim = std::vector<int64_t>(x.dims().size(), -1);
    } else {
      auto x_rank = static_cast<size_t>(x.dims().size());
3121
      if (vec_axis.size() > x_rank) {
3122 3123 3124 3125 3126 3127 3128 3129 3130
        vec_dim = {-1};
      } else {
        vec_dim = std::vector<int64_t>(x.dims().size() - vec_axis.size(), -1);
      }
    }
  }
  return phi::make_ddim(vec_dim);
}

3131 3132 3133 3134 3135 3136 3137 3138 3139 3140 3141
void ReduceIntArrayAxisInferMetaBase(const MetaTensor& x,
                                     const IntArray& axis,
                                     bool keep_dim,
                                     bool reduce_all,
                                     MetaTensor* out,
                                     MetaConfig config) {
  DDim out_dim;
  if (config.is_runtime || !axis.FromTensor()) {
    out_dim = ReduceInferDim(x, axis.GetData(), keep_dim, reduce_all);
  } else {
    out_dim = ReduceInferDimForIntArrayAxis(x, axis, keep_dim, reduce_all);
3142
  }
3143 3144 3145
  out->set_dims(out_dim);
  out->set_dtype(x.dtype());
  out->set_layout(x.layout());
3146 3147
}

3148 3149 3150 3151 3152 3153 3154 3155 3156 3157 3158 3159
void ReduceIntArrayAxisInferMeta(const MetaTensor& x,
                                 const IntArray& axis,
                                 bool keep_dim,
                                 MetaTensor* out,
                                 MetaConfig config) {
  bool reduce_all = false;
  if (axis.size() == 0) {
    reduce_all = true;
  }
  ReduceIntArrayAxisInferMetaBase(x, axis, keep_dim, reduce_all, out, config);
}

3160 3161 3162 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173
void ReduceScatterInferMeta(const MetaTensor& x, int nranks, MetaTensor* out) {
  auto dim = x.dims();
  if (dim[0] > 0 || dim[0] < -1) {
    PADDLE_ENFORCE_EQ(
        dim[0] % nranks,
        0,
        errors::InvalidArgument(
            "dim[0] (%d) is not divisible by nranks(%d)", dim[0], nranks));
    dim[0] /= nranks;
  }
  out->set_dims(dim);
  out->set_dtype(x.dtype());
}

S
seemingwang 已提交
3174 3175 3176 3177 3178 3179
void RepeatInterleaveInferMeta(const MetaTensor& x,
                               int repeats,
                               int dim,
                               MetaTensor* out) {
  const auto& input_dim = x.dims();
  auto output_dim = phi::vectorize(input_dim);
3180
  auto n_dim = dim;
S
seemingwang 已提交
3181

3182 3183 3184 3185 3186
  if (n_dim < 0) n_dim += input_dim.size();

  PADDLE_ENFORCE_LT(
      dim,
      input_dim.size(),
S
seemingwang 已提交
3187 3188
      phi::errors::OutOfRange(
          "Attr(dim) is out of range, It's expected "
3189 3190
          "to be in range of [%d, %d]. But received Attr(dim) = %d.",
          -input_dim.size(),
S
seemingwang 已提交
3191 3192
          input_dim.size() - 1,
          dim));
3193 3194 3195 3196 3197 3198 3199 3200 3201 3202 3203 3204 3205
  PADDLE_ENFORCE_GE(
      dim,
      (0 - input_dim.size()),
      phi::errors::OutOfRange(
          "Attr(dim) is out of range, It's expected "
          "to be in range of [%d, %d]. But received Attr(dim) = %d.",
          -input_dim.size(),
          input_dim.size() - 1,
          dim));

  PADDLE_ENFORCE_GT(
      repeats,
      0,
S
seemingwang 已提交
3206 3207
      phi::errors::InvalidArgument("repeats should be larger than zero"));

3208 3209 3210 3211
  PADDLE_ENFORCE_NOT_NULL(
      out,
      phi::errors::InvalidArgument(
          "repeat_interleave's output tensor can't be nullptr"));
S
seemingwang 已提交
3212

3213
  output_dim[n_dim] = input_dim[n_dim] * repeats;
S
seemingwang 已提交
3214 3215 3216 3217
  out->set_dims(phi::make_ddim(output_dim));
  out->share_lod(x);
  out->set_dtype(x.dtype());
}
3218

Z
zyfncg 已提交
3219
void ReshapeInferMeta(const MetaTensor& x,
3220
                      const IntArray& shape,
Z
zyfncg 已提交
3221 3222 3223 3224 3225 3226 3227 3228 3229 3230 3231 3232
                      MetaTensor* out,
                      MetaConfig config) {
  auto& shape_data = shape.GetData();
  PADDLE_ENFORCE_NOT_NULL(out,
                          phi::errors::InvalidArgument(
                              "Output(Out) of ReshapeOp should not be null."));
  if (!config.is_runtime && shape.FromTensor()) {
    out->set_dims(phi::make_ddim(shape_data));
    out->share_lod(x);
    return;
  }
  InferMetaFromVecValue(x, shape_data, out);
3233 3234
}

Z
zyfncg 已提交
3235
void ReshapeWithXShapeInferMeta(const MetaTensor& x,
3236
                                const IntArray& shape,
Z
zyfncg 已提交
3237
                                MetaTensor* out,
3238
                                MetaTensor* xshape,
Z
zyfncg 已提交
3239 3240 3241 3242 3243 3244 3245 3246 3247 3248 3249 3250 3251 3252
                                MetaConfig config) {
  PADDLE_ENFORCE_NOT_NULL(
      xshape,
      phi::errors::InvalidArgument(
          "Output(XShape) of ReshapeOp should not be null."));
  const auto& x_dims = x.dims();
  std::vector<int64_t> xshape_dims(x_dims.size() + 1);
  xshape_dims[0] = 0;
  for (int i = 0; i < x_dims.size(); ++i) {
    xshape_dims[i + 1] = x_dims[i];
  }
  xshape->set_dims(phi::make_ddim(xshape_dims));
  xshape->share_lod(x);
  ReshapeInferMeta(x, shape, out, config);
3253 3254
}

3255
void ReverseInferMeta(const MetaTensor& x,
3256 3257 3258 3259 3260 3261 3262 3263 3264 3265 3266 3267
                      const IntArray& axis,
                      MetaTensor* out,
                      MetaConfig config) {
  // NOTE(Aurelius84): In Reverse Op, output TensorMeta is always same
  // as input, so we only verify axis when it is not from Tensor or in
  // runtime.
  if (!config.is_runtime && axis.FromTensor()) {
    out->share_meta(x);
    return;
  }
  auto& axis_data = axis.GetData();
  PADDLE_ENFORCE_NE(axis_data.empty(),
3268 3269 3270
                    true,
                    phi::errors::InvalidArgument("'axis' can not be empty."));
  const auto& x_dims = x.dims();
3271
  for (int a : axis_data) {
3272 3273 3274 3275 3276 3277 3278 3279 3280 3281 3282 3283 3284 3285 3286 3287 3288 3289 3290
    PADDLE_ENFORCE_LT(a,
                      x_dims.size(),
                      phi::errors::OutOfRange(
                          "The axis must be less than input tensor's rank. "
                          "but got %d >= %d",
                          a,
                          x_dims.size()));
    PADDLE_ENFORCE_GE(
        a,
        -x_dims.size(),
        phi::errors::OutOfRange(
            "The axis must be greater than the negative number of "
            "input tensor's rank, but got %d < %d",
            a,
            -x_dims.size()));
  }
  out->share_meta(x);
}

W
wanghuancoder 已提交
3291
void ReverseArrayInferMeta(const std::vector<const phi::MetaTensor*>& x,
3292 3293 3294 3295 3296 3297 3298
                           const IntArray& axis,
                           std::vector<phi::MetaTensor*> out,
                           MetaConfig config) {
  if (!config.is_runtime && axis.FromTensor()) {
    return;
  }
  auto& axis_data = axis.GetData();
W
wanghuancoder 已提交
3299
  PADDLE_ENFORCE_EQ(
3300
      axis_data.size(),
W
wanghuancoder 已提交
3301 3302 3303 3304
      1,
      phi::errors::InvalidArgument(
          "The size of axis must be 1 when the Input(X) is LoDTensorArray, "
          "but received %d.",
3305
          axis_data.size()));
W
wanghuancoder 已提交
3306
  PADDLE_ENFORCE_EQ(
3307
      axis_data[0],
W
wanghuancoder 已提交
3308 3309 3310 3311
      0,
      phi::errors::InvalidArgument("The value of axis should be 1 when "
                                   "the Input(X) is LoDTensorArray, "
                                   "but received %d.",
3312
                                   axis_data[0]));
W
wanghuancoder 已提交
3313 3314
}

C
chenenquan 已提交
3315
void RollInferMeta(const MetaTensor& x,
3316
                   const IntArray& shifts,
C
chenenquan 已提交
3317 3318 3319 3320 3321 3322 3323 3324 3325 3326 3327 3328 3329 3330 3331 3332 3333 3334 3335 3336 3337 3338 3339 3340 3341 3342 3343 3344 3345
                   const std::vector<int64_t>& axis,
                   MetaTensor* out) {
  auto shifts_data = shifts.GetData();

  if (axis.size() != 0) {
    PADDLE_ENFORCE_EQ(
        axis.size(),
        shifts_data.size(),
        phi::errors::InvalidArgument("When dims.size() != 0, dims.size() "
                                     "should be equal to "
                                     "shifts.size(). But received "
                                     "dims.size() = %d, shifts.size() = %d",
                                     axis.size(),
                                     shifts_data.size()));
  } else {
    PADDLE_ENFORCE_EQ(
        shifts_data.size(),
        1,
        phi::errors::InvalidArgument("When dims.size() == 0, shifts.size() "
                                     "should be equal to 1, But received "
                                     "shifts.size() = %d",
                                     shifts_data.size()));
  }

  out->set_dims(x.dims());
  out->share_lod(x);
  out->set_dtype(x.dtype());
}

3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 3356 3357 3358 3359 3360 3361 3362 3363 3364 3365 3366 3367 3368 3369 3370 3371 3372 3373 3374 3375 3376 3377 3378 3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389 3390 3391 3392 3393 3394
void RReluInferMeta(const MetaTensor& x,
                    float lower,
                    float upper,
                    bool is_test,
                    MetaTensor* out,
                    MetaTensor* noise) {
  auto x_dims = x.dims();
  PADDLE_ENFORCE_GE(lower,
                    0,
                    phi::errors::InvalidArgument(
                        "The lower value should be greater than or equal to 0. "
                        "But received lower value = %f.",
                        lower));
  PADDLE_ENFORCE_LE(upper,
                    1,
                    phi::errors::InvalidArgument(
                        "The upper value should be less than or equal to 1. "
                        "But received upper value = %f.",
                        upper));
  PADDLE_ENFORCE_GE(
      upper,
      lower,
      phi::errors::InvalidArgument(
          "The upper value should be greater than or equal to lower value "
          "But received upper value = %f, lower value = %f.",
          upper,
          lower));

  out->set_dims(x_dims);
  out->set_dtype(x.dtype());
  out->set_layout(x.layout());
  out->share_lod(x);

  if (noise != nullptr) {
    noise->set_dims(x_dims);
    noise->set_dtype(x.dtype());
    noise->set_layout(x.layout());
  }
}

void RReluGradInferMeta(const MetaTensor& out_grad,
                        const MetaTensor& noise,
                        MetaTensor* x_grad) {
  auto do_dims = out_grad.dims();
  x_grad->set_dims(do_dims);
  x_grad->set_dtype(out_grad.dtype());
  x_grad->share_lod(out_grad);
}

3395 3396 3397 3398 3399 3400 3401 3402
void SetValueInferMeta(const MetaTensor& x, MetaTensor* out) {
  auto in_dims = x.dims();
  PADDLE_ENFORCE_LT(
      in_dims.size(),
      7,
      phi::errors::InvalidArgument(
          "The rank of input should be less than 7, but received %d.",
          in_dims.size()));
3403
  out->set_dims(in_dims);
3404 3405
}

3406 3407 3408 3409 3410 3411
void ShapeInferMeta(const MetaTensor& input, MetaTensor* out) {
  auto in_dim = input.dims();
  out->set_dims(phi::make_ddim({in_dim.size()}));
  out->set_dtype(DataType::INT32);
}

Z
zyfncg 已提交
3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425 3426 3427 3428 3429 3430 3431 3432 3433 3434 3435 3436 3437 3438 3439
void ShardIndexInferMeta(const MetaTensor& in,
                         int index_num,
                         int nshards,
                         int shard_id,
                         int ignore_value,
                         MetaTensor* out,
                         MetaConfig config) {
  auto x_dims = in.dims();
  PADDLE_ENFORCE_GE(
      x_dims.size(),
      2,
      phi::errors::InvalidArgument("Rank of Input(X) should be at least 2, "
                                   "but the value given is %d.",
                                   x_dims.size()));
  if (config.is_runtime || x_dims[x_dims.size() - 1] > 0) {
    PADDLE_ENFORCE_EQ(x_dims[x_dims.size() - 1],
                      1U,
                      phi::errors::InvalidArgument(
                          "The last dimension of Input(X) should be 1, "
                          "but the value given is %d.",
                          x_dims[x_dims.size() - 1]));
  }

  out->set_dims(x_dims);
  out->share_lod(in);
  out->set_dtype(in.dtype());
}

3440
void NumelInferMeta(const MetaTensor& input, MetaTensor* out) {
Z
zyfncg 已提交
3441
  out->set_dtype(DataType::INT64);
zhouweiwei2014's avatar
zhouweiwei2014 已提交
3442
  out->set_dims(phi::make_ddim({}));
Z
zyfncg 已提交
3443 3444
}

H
hong 已提交
3445 3446 3447 3448 3449 3450 3451 3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462 3463 3464 3465
void SliceRawInferMeta(const MetaTensor& input,
                       const std::vector<int64_t>& axes,
                       const IntArray& starts_arr,
                       const IntArray& ends_arr,
                       const std::vector<int64_t>& infer_flags_t,
                       const std::vector<int64_t>& decrease_axis,
                       MetaTensor* out,
                       MetaConfig config) {
  auto in_dims = input.dims();
  PADDLE_ENFORCE_LT(
      in_dims.size(),
      7,
      phi::errors::InvalidArgument("The rank of input should be less than 7."));
  DDim out_dims(in_dims);

  std::vector<int64_t> infer_flags = infer_flags_t;
  if (infer_flags.empty()) {
    // Initialize infer_flags with 1.
    // To be compatible with other op tests in which infer_flags is not set.
    infer_flags = std::vector<int64_t>(axes.size(), 1);
  }
3466 3467 3468 3469 3470 3471
  auto new_axes = axes;
  for (auto& axis : new_axes) {
    if (axis < 0) {
      axis = std::max(int64_t(0), axis + int64_t(in_dims.size()));
    }
  }
H
hong 已提交
3472

3473 3474 3475 3476 3477 3478 3479 3480 3481 3482 3483 3484 3485 3486 3487
  PADDLE_ENFORCE_EQ(
      axes.size(),
      starts_arr.size(),
      phi::errors::InvalidArgument(
          "The length of axes (%d) and length of starts (%d) should be same.",
          axes.size(),
          starts_arr.size()));
  PADDLE_ENFORCE_EQ(
      axes.size(),
      ends_arr.size(),
      phi::errors::InvalidArgument(
          "The length of axes (%d) and length of ends (%d) should be same.",
          axes.size(),
          ends_arr.size()));

H
hong 已提交
3488 3489 3490 3491 3492
  // 2.1 Check attrs.
  std::vector<int64_t> starts = starts_arr.GetData();
  std::vector<int64_t> ends = ends_arr.GetData();

  phi::funcs::CheckAndUpdateSliceAttrs<int64_t>(
3493
      in_dims, new_axes, &starts, &ends, nullptr, &infer_flags);
H
hong 已提交
3494 3495

  auto slice_dims = phi::funcs::GetSliceDims<int64_t>(
3496
      in_dims, new_axes, starts, ends, nullptr, &infer_flags);
H
hong 已提交
3497 3498 3499 3500 3501 3502 3503 3504 3505
  if (config.is_runtime) {
    out_dims = phi::funcs::GetDecreasedDims<int64_t>(
        slice_dims, decrease_axis, &infer_flags);
  } else {
    out_dims = phi::funcs::GetDecreasedDims<int64_t>(
        slice_dims, decrease_axis, nullptr);
  }

  out->set_dims(out_dims);
3506
  if (new_axes.size() > 0 && new_axes[0] != 0) {
H
hong 已提交
3507 3508 3509 3510
    out->share_lod(input);
  }
}

Z
zyfncg 已提交
3511 3512 3513
void SoftmaxInferMeta(const MetaTensor& x, int axis, MetaTensor* out) {
  auto dim_x = x.dims();
  auto rank_x = dim_x.size();
3514 3515 3516 3517 3518 3519 3520 3521 3522 3523 3524 3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536
  if (rank_x > 0) {
    PADDLE_ENFORCE_GE(axis,
                      -rank_x,
                      phi::errors::InvalidArgument(
                          "Attr(axis) value should be in range [-R, R-1], "
                          "R is the rank of Input(X)."));
    PADDLE_ENFORCE_LT(axis,
                      rank_x,
                      phi::errors::InvalidArgument(
                          "Attr(axis) value should be in range [-R, R-1], "
                          "R is the rank of Input(X)."));
  } else {
    PADDLE_ENFORCE_GE(
        axis,
        -1,
        phi::errors::InvalidArgument("Attr(axis) value should be in range [-1, "
                                     "0] when input is 0D Tensor "));
    PADDLE_ENFORCE_LE(
        axis,
        0,
        phi::errors::InvalidArgument("Attr(axis) value should be in range [-1, "
                                     "0] when input is 0D Tensor "));
  }
Z
zyfncg 已提交
3537 3538 3539 3540 3541 3542

  out->set_dims(x.dims());
  out->set_dtype(x.dtype());
  out->share_lod(x);
}

C
Charles-hit 已提交
3543 3544 3545 3546 3547 3548 3549 3550 3551 3552 3553 3554 3555 3556 3557 3558 3559 3560 3561 3562 3563 3564 3565 3566 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 3578 3579 3580 3581 3582 3583 3584 3585 3586 3587 3588 3589 3590 3591 3592 3593 3594
int GetSplitAxisValue(const MetaTensor& x,
                      const Scalar& axis,
                      MetaConfig config) {
  // Tensor has no value in static graph compile time
  if (axis.FromTensor() && !config.is_runtime) {
    return -1;
  } else {
    if (axis.dtype() == DataType::FLOAT32 ||
        axis.dtype() == DataType::FLOAT64) {
      PADDLE_THROW(
          phi::errors::InvalidArgument("%s(): argument (position 3) must be "
                                       "int, but got %s",
                                       "split",
                                       "float"));  // NOLINT
    }
    int axis_value = axis.to<int>();
    int rank = x.dims().size();
    PADDLE_ENFORCE_EQ(
        axis_value >= -rank && axis_value < rank,
        true,
        phi::errors::InvalidArgument(
            "The axis is expected to be in range of [%d, %d), but got %d",
            -rank,
            rank,
            axis_value));
    if (axis_value < 0) {
      axis_value = axis_value + rank;
    }
    return axis_value;
  }
}

void FillSplitOutDims(const MetaTensor& x,
                      const int axis_value,
                      const std::vector<int64_t>& sections_vec,
                      std::vector<MetaTensor*>* out) {
  std::vector<phi::DDim> out_dims(sections_vec.size(), x.dims());
  if (x.dims().at(axis_value) > 0) {
    for (size_t i = 0; i < sections_vec.size(); ++i) {
      out_dims[i][axis_value] = sections_vec[i];
    }
  } else {
    for (size_t i = 0; i < sections_vec.size(); ++i) {
      out_dims[i][axis_value] = -1;
    }
  }
  for (size_t i = 0; i < sections_vec.size(); ++i) {
    if (axis_value != 0) {
      // Only pass LoD when not spliting along the first dim.
      (*out)[i]->set_dtype(x.dtype());
      (*out)[i]->set_dims(out_dims[i]);
      (*out)[i]->set_layout(x.layout());
X
xiaoxiaohehe001 已提交
3595
      (*out)[i]->share_lod(x);
C
Charles-hit 已提交
3596 3597 3598 3599 3600 3601 3602 3603
    } else {
      (*out)[i]->set_dtype(x.dtype());
      (*out)[i]->set_dims(out_dims[i]);
      (*out)[i]->set_layout(x.layout());
    }
  }
}

Z
zyfncg 已提交
3604
void SplitInferMeta(const MetaTensor& x,
C
Charles-hit 已提交
3605
                    const IntArray& sections,
Z
zyfncg 已提交
3606 3607 3608
                    const Scalar& axis,
                    std::vector<MetaTensor*> out,
                    MetaConfig config) {
C
Charles-hit 已提交
3609 3610 3611 3612 3613 3614 3615
  // get axis value
  int axis_value = GetSplitAxisValue(x, axis, config);

  auto sections_data = sections.GetData();
  // fill out dims with -1
  if ((sections.FromTensor() && !config.is_runtime) || axis_value == -1 ||
      (axis_value >= 0 && x.dims().at(axis_value) <= 0)) {
3616 3617 3618 3619 3620 3621 3622 3623
    std::vector<phi::DDim> out_dims;
    if ((sections.FromTensor() && !config.is_runtime) || axis_value == -1) {
      out_dims = std::vector<phi::DDim>(
          sections_data.size(),
          phi::make_ddim(std::vector<int>(x.dims().size(), -1)));
    } else {
      out_dims = std::vector<phi::DDim>(sections_data.size(), x.dims());
    }
C
Charles-hit 已提交
3624 3625 3626 3627 3628 3629
    for (size_t i = 0; i < sections_data.size(); ++i) {
      if (axis_value != 0) {
        // Only pass LoD when not spliting along the first dim.
        out[i]->set_dtype(x.dtype());
        out[i]->set_dims(out_dims[i]);
        out[i]->set_layout(x.layout());
X
xiaoxiaohehe001 已提交
3630
        out[i]->share_lod(x);
C
Charles-hit 已提交
3631 3632 3633 3634 3635
      } else {
        out[i]->set_dtype(x.dtype());
        out[i]->set_dims(out_dims[i]);
        out[i]->set_layout(x.layout());
      }
C
chentianyu03 已提交
3636 3637
    }
  } else {
C
Charles-hit 已提交
3638 3639
    auto input_axis_dim = x.dims().at(axis_value);
    std::vector<int64_t> sections_vec;
C
chentianyu03 已提交
3640 3641 3642 3643 3644
    const int unknow_dim_val = -1;
    int unknow_dim_idx = -1;
    int num_of_unknow = 0;
    int sum_of_section = 0;

C
Charles-hit 已提交
3645 3646
    for (size_t i = 0; i < sections_data.size(); ++i) {
      sections_vec.push_back(sections_data[i]);
3647

C
Charles-hit 已提交
3648
      if (sections_data[i] == unknow_dim_val) {
C
chentianyu03 已提交
3649 3650 3651
        num_of_unknow++;
        unknow_dim_idx = i;
      } else {
C
Charles-hit 已提交
3652
        sum_of_section += sections_data[i];
C
chentianyu03 已提交
3653 3654 3655
      }
    }

C
Charles-hit 已提交
3656 3657 3658 3659 3660 3661 3662
    PADDLE_ENFORCE_LE(num_of_unknow,
                      1,
                      phi::errors::InvalidArgument(
                          "Only one dimension value of Attr(num_or_sections) "
                          "in SplitOp can be -1. "
                          "But received Attr(num_or_sections) = [%s].",
                          phi::make_ddim(sections_data)));
C
chentianyu03 已提交
3663 3664 3665 3666 3667 3668 3669 3670

    if (unknow_dim_idx != -1) {
      // for example, input shape = [4 ,5], axis = 1, sections = [2, 3, -1].
      // input_axis_dim = 5, sum_of_sections = 5.
      // the following check will fail.
      PADDLE_ENFORCE_LT(
          sum_of_section,
          input_axis_dim,
3671
          phi::errors::InvalidArgument(
C
chentianyu03 已提交
3672 3673 3674 3675 3676
              "Sum of Attr(num_or_sections) other than unknown section "
              "must be less than the input's "
              "size "
              "along the split dimension. But received Attr(num_or_sections) "
              "= [%s], input(X)'s shape = [%s], Attr(dim) = %d.",
C
Charles-hit 已提交
3677
              phi::make_ddim(sections_data),
C
chentianyu03 已提交
3678 3679 3680
              x.dims(),
              axis_value));

C
Charles-hit 已提交
3681
      sections_vec[unknow_dim_idx] = input_axis_dim - sum_of_section;
C
chentianyu03 已提交
3682 3683 3684 3685
    } else {
      PADDLE_ENFORCE_EQ(
          sum_of_section,
          input_axis_dim,
3686
          phi::errors::InvalidArgument(
C
chentianyu03 已提交
3687 3688 3689 3690
              "Sum of Attr(num_or_sections) must be equal to the input's "
              "size "
              "along the split dimension. But received Attr(num_or_sections)"
              " = [%s], input(X)'s shape = [%s], Attr(dim) = %d.",
C
Charles-hit 已提交
3691
              phi::make_ddim(sections_data),
C
chentianyu03 已提交
3692 3693 3694
              x.dims(),
              axis_value));
    }
C
Charles-hit 已提交
3695 3696
    // fill out dims
    FillSplitOutDims(x, axis_value, sections_vec, &out);
3697
  }
C
Charles-hit 已提交
3698 3699 3700 3701 3702 3703 3704 3705 3706 3707
}

void SplitWithNumInferMeta(const MetaTensor& x,
                           int num,
                           const Scalar& axis,
                           std::vector<MetaTensor*> out,
                           MetaConfig config) {
  int axis_value = GetSplitAxisValue(x, axis, config);
  // fill out dims with -1
  if (axis_value == -1 || (axis_value >= 0 && x.dims().at(axis_value) <= 0)) {
3708 3709 3710 3711 3712 3713 3714
    std::vector<phi::DDim> out_dims;
    if (axis_value == -1) {
      out_dims = std::vector<phi::DDim>(
          num, phi::make_ddim(std::vector<int>(x.dims().size(), -1)));
    } else {
      out_dims = std::vector<phi::DDim>(num, x.dims());
    }
C
Charles-hit 已提交
3715 3716 3717 3718 3719 3720
    for (int i = 0; i < num; ++i) {
      if (axis_value != 0) {
        // Only pass LoD when not spliting along the first dim.
        out[i]->set_dtype(x.dtype());
        out[i]->set_dims(out_dims[i]);
        out[i]->set_layout(x.layout());
X
xiaoxiaohehe001 已提交
3721
        out[i]->share_lod(x);
C
Charles-hit 已提交
3722 3723 3724 3725 3726
      } else {
        out[i]->set_dtype(x.dtype());
        out[i]->set_dims(out_dims[i]);
        out[i]->set_layout(x.layout());
      }
C
chentianyu03 已提交
3727
    }
3728
  } else {
C
Charles-hit 已提交
3729 3730 3731
    auto input_axis_dim = x.dims().at(axis_value);
    // step1: get formated sections
    std::vector<int64_t> sections_vec;
张春乔 已提交
3732 3733 3734 3735
    PADDLE_ENFORCE_NE(
        num,
        0,
        phi::errors::InvalidArgument("Attr(num_or_sections) should not be 0."));
C
Charles-hit 已提交
3736 3737 3738 3739 3740 3741 3742 3743 3744 3745
    PADDLE_ENFORCE_EQ(input_axis_dim % num,
                      0,
                      phi::errors::InvalidArgument(
                          "The input's size along the split dimension "
                          "must be evenly divisible by Attr(num_or_sections). "
                          "But received Attr(num_or_sections) "
                          "= %d, input(X)'s shape = [%s], Attr(dim) = %d.",
                          num,
                          x.dims(),
                          axis_value));
C
chentianyu03 已提交
3746

C
Charles-hit 已提交
3747 3748
    for (int i = 0; i < num; ++i) {
      sections_vec.push_back(input_axis_dim / num);
C
chentianyu03 已提交
3749
    }
C
Charles-hit 已提交
3750 3751
    // setp2: fill out dims
    FillSplitOutDims(x, axis_value, sections_vec, &out);
C
chentianyu03 已提交
3752
  }
C
Chen Weihang 已提交
3753 3754
}

3755 3756 3757 3758
void SquaredL2NormInferMeta(const MetaTensor& x, MetaTensor* out) {
  out->set_dims({1});
}

3759
void SqueezeInferMeta(const MetaTensor& x,
3760 3761 3762
                      const IntArray& axes,
                      MetaTensor* out,
                      MetaConfig config) {
3763 3764 3765 3766 3767 3768 3769 3770 3771 3772 3773
  const auto& x_dims = x.dims();
  // Check input tensor dims (<6) Eigen limit.
  PADDLE_ENFORCE_LE(x_dims.size(),
                    6,
                    phi::errors::InvalidArgument(
                        "The dimensions of Input(X) "
                        "should be in the range of [1, 6] (Eigen limit)."
                        "But received X's dimensions = %d, X's shape = [%s].",
                        x_dims.size(),
                        x_dims));

3774 3775 3776
  if (!config.is_runtime && axes.FromTensor()) {
    // compile time infershape, set all elements to -1.
    int output_size = x.dims().size() - axes.GetData().size();
3777 3778 3779
    if (x.dims().size() == 0 && output_size == -1) {
      output_size = 0;
    }
3780 3781 3782 3783 3784 3785 3786 3787 3788 3789 3790 3791 3792 3793 3794
    std::vector<int64_t> vec_out_dims(output_size, -1);
    out->set_dims(phi::make_ddim(vec_out_dims));
  } else {
    std::vector<int32_t> tmp;
    tmp.reserve(axes.GetData().size());
    std::for_each(axes.GetData().begin(),
                  axes.GetData().end(),
                  [&tmp](const int64_t& t) { tmp.push_back(t); });
    auto out_dims = funcs::GetOutputSqueezeShape(tmp, x_dims, false);
    out->set_dims(out_dims);
    if (x_dims[0] == out_dims[0]) {
      // Only pass LoD when the first dimension of output and Input(X)
      // are the same.
      out->share_lod(x);
    }
3795
  }
3796 3797 3798 3799
  out->set_dtype(x.dtype());
}

void SqueezeWithXShapeInferMeta(const MetaTensor& x,
3800
                                const IntArray& axes,
3801
                                MetaTensor* out,
3802 3803 3804
                                MetaTensor* xshape,
                                MetaConfig config) {
  SqueezeInferMeta(x, axes, out, config);
3805
  const auto& x_dims = x.dims();
3806 3807 3808 3809 3810
  std::vector<int64_t> xshape_dims(x_dims.size() + 1);
  xshape_dims[0] = 0;
  for (int i = 0; i < x_dims.size(); ++i) {
    xshape_dims[i + 1] = x_dims[i];
  }
3811 3812 3813 3814 3815
  if (xshape) {
    xshape->set_dims(phi::make_ddim(xshape_dims));
    xshape->share_lod(x);
    xshape->set_dtype(x.dtype());
  }
3816 3817
}

3818 3819 3820 3821 3822 3823 3824 3825 3826
void StridedSliceRawInferMeta(const MetaTensor& x,
                              const std::vector<int>& axes,
                              const IntArray& starts,
                              const IntArray& ends,
                              const IntArray& strides,
                              const std::vector<int>& infer_flags,
                              const std::vector<int>& decrease_axis,
                              MetaTensor* out,
                              MetaConfig config) {
3827 3828 3829 3830 3831 3832 3833 3834 3835 3836 3837 3838 3839 3840 3841 3842 3843 3844 3845 3846 3847 3848 3849 3850 3851 3852 3853 3854 3855 3856 3857 3858 3859 3860 3861 3862 3863
  auto in_dims = x.dims();
  PADDLE_ENFORCE_LT(
      in_dims.size(),
      7,
      errors::InvalidArgument(
          "The dimension of StridedSlice operator's input should be less "
          "than 7, but received dimension is %d.",
          in_dims.size()));

  auto starts_ = starts.GetData();
  auto ends_ = ends.GetData();
  auto strides_ = strides.GetData();

  auto starts_size = starts_.size();
  auto ends_size = ends_.size();
  auto strides_size = strides_.size();

  for (size_t i = 0; i < axes.size(); ++i) {
    PADDLE_ENFORCE_GE(
        axes[i],
        0,
        errors::InvalidArgument("The axis should be greater than or equal to 0."
                                "But received %d of axes[%d]",
                                axes[i],
                                i));
    PADDLE_ENFORCE_LT(
        axes[i],
        in_dims.size(),
        errors::InvalidArgument(
            "The axes should be less than or equal to input tensor's rank."
            "But received %d of axes[%d], input tensor shape [%d]",
            axes[i],
            i,
            in_dims.size()));
  }

  auto tensor_input = false;
3864
  auto HasInput = [](const IntArray& arr) { return arr.FromTensor(); };
3865 3866 3867 3868 3869 3870 3871 3872 3873 3874 3875 3876 3877 3878 3879 3880 3881 3882 3883 3884 3885 3886 3887 3888 3889 3890 3891 3892 3893 3894 3895 3896 3897 3898 3899 3900 3901 3902 3903 3904 3905 3906 3907 3908 3909 3910 3911 3912 3913 3914 3915 3916 3917 3918 3919 3920 3921 3922 3923 3924 3925 3926 3927 3928 3929 3930 3931 3932 3933 3934 3935 3936 3937 3938
  if (HasInput(starts) || HasInput(ends) || HasInput(strides)) {
    tensor_input = true;
  }
  if (!HasInput(ends)) {
    PADDLE_ENFORCE_EQ(
        ends_size,
        axes.size(),
        errors::InvalidArgument(
            "The size of ends attribute in StridedSlice operator is not "
            "equal to the size of axes attribute. The ends attribute's size "
            "is %d, axes attribute's size is %d.",
            ends_size,
            axes.size()));
  }
  if (!HasInput(starts)) {
    PADDLE_ENFORCE_EQ(
        starts_size,
        axes.size(),
        errors::InvalidArgument(
            "The size of starts attribute in StridedSlice operator is not "
            "equal to the size of axes attribute. The starts attribute's "
            "size is %d, axes attribute's size is %d.",
            starts_size,
            axes.size()));
  }
  if (!HasInput(strides)) {
    PADDLE_ENFORCE_EQ(
        strides_size,
        axes.size(),
        errors::InvalidArgument(
            "The size of strides attribute in StridedSlice operator is not "
            "equal to the size of axes attribute. The strides attribute's "
            "size is %d, axes attribute's size is %d.",
            strides_size,
            axes.size()));
  }
  // we need to analysis strided slice op is valid for
  // the parameter that we get from python front
  std::vector<int64_t> out_dims_vector(in_dims.size(), -1);
  if (!tensor_input || config.is_runtime) {
    phi::funcs::StridedSliceOutDims(starts_,
                                    ends_,
                                    strides_,
                                    axes,
                                    infer_flags,
                                    in_dims,
                                    decrease_axis,
                                    out_dims_vector.data(),
                                    axes.size(),
                                    true);
  }
  DDim out_dims(phi::make_ddim(out_dims_vector));
  // generate new shape
  if (decrease_axis.size() > 0) {
    std::vector<int64_t> new_out_shape;
    for (size_t i = 0; i < decrease_axis.size(); ++i) {
      if (config.is_runtime && infer_flags[i] != -1) {
        PADDLE_ENFORCE_EQ(out_dims[decrease_axis[i]],
                          1,
                          errors::InvalidArgument(
                              "the size of decrease dimension should be 1, "
                              "but received %d.",
                              out_dims[decrease_axis[i]]));
      }
      out_dims[decrease_axis[i]] = 0;
    }

    for (int i = 0; i < out_dims.size(); ++i) {
      if (out_dims[i] != 0) {
        new_out_shape.push_back(out_dims[i]);
      }
    }
    out_dims = phi::make_ddim(new_out_shape);
  }
L
Leo Chen 已提交
3939
  VLOG(4) << "out_dims: " << out_dims;
3940 3941 3942 3943 3944
  out->set_dims(out_dims);
  out->share_lod(x);
  out->set_dtype(x.dtype());
}

3945 3946 3947 3948 3949 3950 3951 3952 3953 3954 3955 3956 3957
void StridedSliceInferMeta(const MetaTensor& x,
                           const std::vector<int>& axes,
                           const IntArray& starts,
                           const IntArray& ends,
                           const IntArray& strides,
                           MetaTensor* out,
                           MetaConfig config) {
  std::vector<int> infer_flags(axes.size(), 1);
  std::vector<int> decrease_axis;
  StridedSliceRawInferMeta(
      x, axes, starts, ends, strides, infer_flags, decrease_axis, out, config);
}

Z
zyfncg 已提交
3958
void SumRawInferMeta(const MetaTensor& x,
3959
                     const IntArray& axis,
Z
zyfncg 已提交
3960 3961 3962
                     bool keep_dim,
                     bool reduce_all,
                     DataType dtype,
3963 3964 3965 3966
                     MetaTensor* out,
                     MetaConfig config) {
  DDim out_dim;
  if (config.is_runtime || !axis.FromTensor()) {
3967
    out_dim = ReduceInferDim(x, axis.GetData(), keep_dim, reduce_all);
3968
  } else {
3969
    out_dim = ReduceInferDimForIntArrayAxis(x, axis, keep_dim, reduce_all);
3970
  }
Z
zyfncg 已提交
3971 3972 3973 3974 3975

  DataType out_dtype;
  if (dtype != DataType::UNDEFINED) {
    out_dtype = dtype;
  } else {
3976
    if (x.dtype() == DataType::BOOL || x.dtype() == DataType::INT32) {
Z
zyfncg 已提交
3977 3978 3979 3980
      out_dtype = DataType::INT64;
    } else {
      out_dtype = x.dtype();
    }
L
Leo Chen 已提交
3981 3982
  }

Z
zyfncg 已提交
3983 3984 3985 3986 3987
  out->set_dims(out_dim);
  out->set_dtype(out_dtype);
  out->set_layout(x.layout());
}

3988 3989 3990 3991 3992 3993 3994 3995 3996 3997
/*  Why not use SumRawInferMeta directly?
    Because we need make InferMetaFunction's args follow the design of
   ops.yaml
*/
void SumInferMeta(const MetaTensor& x,
                  const IntArray& axis,
                  DataType dtype,
                  bool keep_dim,
                  MetaTensor* out,
                  MetaConfig config) {
3998 3999 4000 4001
  bool reduce_all = false;
  if (axis.size() == 0) {
    reduce_all = true;
  }
4002
  SumRawInferMeta(x, axis, keep_dim, reduce_all, dtype, out, config);
4003 4004
}

4005 4006 4007 4008 4009 4010 4011 4012 4013 4014 4015 4016 4017 4018 4019 4020 4021 4022 4023 4024 4025 4026 4027 4028 4029 4030 4031 4032 4033 4034 4035 4036 4037 4038 4039 4040 4041 4042 4043 4044 4045 4046 4047 4048 4049 4050 4051
void SvdInferMeta(const MetaTensor& x,
                  bool full_matrices,
                  MetaTensor* u,
                  MetaTensor* s,
                  MetaTensor* vh) {
  auto UDDim = [](const DDim& x_dim, int k) {
    // get x_dim and return the ddim of U
    auto x_vec = vectorize(x_dim);
    x_vec[x_vec.size() - 1] = k;
    return phi::make_ddim(x_vec);
  };

  auto VHDDim = [](const DDim& x_dim, int k) {
    // get x_dim and return the ddim of U
    auto x_vec = vectorize(x_dim);
    x_vec[x_vec.size() - 2] = k;
    return phi::make_ddim(x_vec);
  };

  auto SDDim = [](const DDim& x_dim, int k) {
    // get x_dim and return the ddim of U
    auto x_vec = vectorize(x_dim);
    x_vec[x_vec.size() - 2] = k;
    x_vec.erase(x_vec.end() - 1);  // rank - 1
    return phi::make_ddim(x_vec);
  };

  auto in_dims = x.dims();
  int x_rank = in_dims.size();
  PADDLE_ENFORCE_GE(
      in_dims.size(),
      2,
      phi::errors::InvalidArgument("the rank of input must greater than 2"));
  int m = in_dims[x_rank - 2];
  int n = in_dims[x_rank - 1];
  int k = std::min(m, n);
  u->set_dims(!full_matrices ? UDDim(in_dims, k) : UDDim(in_dims, m));
  vh->set_dims(!full_matrices ? VHDDim(in_dims, k) : VHDDim(in_dims, n));
  s->set_dims(SDDim(in_dims, k));
  u->share_lod(x);
  vh->share_lod(x);
  s->share_lod(x);
  u->set_dtype(x.dtype());
  vh->set_dtype(x.dtype());
  s->set_dtype(x.dtype());
}

H
hong 已提交
4052 4053 4054 4055 4056 4057 4058 4059 4060 4061 4062 4063 4064 4065 4066 4067 4068 4069 4070 4071 4072 4073 4074 4075 4076 4077 4078 4079 4080 4081 4082 4083 4084 4085 4086 4087 4088 4089 4090 4091 4092 4093 4094 4095 4096 4097
void TemporalShiftInferMeta(const MetaTensor& x,
                            int seg_num,
                            float shift_ratio,
                            const std::string& data_format,
                            MetaTensor* out,
                            MetaConfig config) {
  auto dim_x = x.dims();
  PADDLE_ENFORCE_EQ(dim_x.size(),
                    4,
                    phi::errors::InvalidArgument(
                        "Input(X) rank should be 4 in shape of [N*T, C, H, "
                        "W], but received X rank(%d)",
                        dim_x.size()));

  PADDLE_ENFORCE_GT(
      seg_num,
      0,
      phi::errors::InvalidArgument(
          "Attr(seg_num) should be greater than 0, but received %d", seg_num));
  PADDLE_ENFORCE_GT(
      shift_ratio,
      0.,
      phi::errors::InvalidArgument(
          "Attr(shift_ratio) should be greater than 0, but received %d",
          shift_ratio));
  PADDLE_ENFORCE_LT(
      shift_ratio,
      0.5,
      phi::errors::InvalidArgument(
          "Attr(shift_ratio) should be less than 0.5, but received %d",
          shift_ratio));

  if (config.is_runtime) {
    PADDLE_ENFORCE_EQ(dim_x[0] % seg_num,
                      0,
                      phi::errors::InvalidArgument(
                          "Input(X) dimension[0] should be divided exactly "
                          "by Attr(seg_num), but received X dimension[0](%d) "
                          "mod seg_num(%d) != 0",
                          dim_x[0],
                          seg_num));
  }

  out->share_meta(x);
}

Z
zyfncg 已提交
4098
void TileInferMeta(const MetaTensor& x,
4099
                   const IntArray& repeat_times,
Z
zyfncg 已提交
4100 4101 4102 4103 4104 4105 4106
                   MetaTensor* out,
                   MetaConfig config) {
#define MAX_RANK_SUPPORTED 6

  auto repeat_times_data = repeat_times.GetData();
  auto x_dims = x.dims();
  if (repeat_times_data.size() == 0) {
4107
    repeat_times_data = std::vector<int64_t>(x_dims.size(), 1);
Z
zyfncg 已提交
4108 4109 4110 4111 4112 4113 4114 4115 4116 4117 4118 4119 4120 4121 4122 4123 4124 4125 4126 4127
  }

  PADDLE_ENFORCE_LE(
      x_dims.size(),
      MAX_RANK_SUPPORTED,
      errors::InvalidArgument(
          "The rank of the input 'x' for tile op "
          "must not be greater than %d, but the value received is %d.",
          MAX_RANK_SUPPORTED,
          x_dims.size()));
  PADDLE_ENFORCE_LE(
      repeat_times_data.size(),
      MAX_RANK_SUPPORTED,
      errors::InvalidArgument(
          "The size of the shape of input 'repeat_times' for tile op "
          "must not be greater than %d, but the value received is %d.",
          MAX_RANK_SUPPORTED,
          repeat_times_data.size()));
  PADDLE_ENFORCE_GE(
      repeat_times_data.size(),
4128
      0,
Z
zyfncg 已提交
4129 4130 4131 4132 4133 4134 4135 4136 4137 4138 4139
      errors::InvalidArgument(
          "The size of the shape of input 'repeat_times' for tile op "
          "must be positive integers, but the value received is %d.",
          repeat_times_data.size()));

  auto out_rank =
      std::max(static_cast<size_t>(x_dims.size()), repeat_times_data.size());
  std::vector<int64_t> out_shape(out_rank);
  auto x_dim_vec = phi::vectorize<int>(x_dims);
  if (x_dim_vec.size() > repeat_times_data.size()) {
    auto diff = x_dim_vec.size() - repeat_times_data.size();
4140
    repeat_times_data.insert(repeat_times_data.begin(), diff, 1);
Z
zyfncg 已提交
4141 4142
  } else {
    auto diff = repeat_times_data.size() - x_dim_vec.size();
4143
    x_dim_vec.insert(x_dim_vec.begin(), diff, 1);
Z
zyfncg 已提交
4144 4145 4146 4147 4148 4149 4150 4151 4152 4153 4154 4155 4156 4157 4158 4159 4160
  }
  for (size_t i = 0; i < repeat_times_data.size(); ++i) {
    if (x_dim_vec[i] == -1 || repeat_times_data[i] == -1) {
      out_shape[i] = -1;
    } else {
      PADDLE_ENFORCE_GT(
          repeat_times_data[i],
          0,
          errors::InvalidArgument(
              "Every element of the input 'repeat_times' for tile op must be "
              "greater than 0, but the value given is %d.",
              repeat_times_data[i]));
      out_shape[i] = x_dim_vec[i] * repeat_times_data[i];
    }
  }

  out->set_dims(phi::make_ddim(out_shape));
4161
  if (out_rank > 0 && (out_shape[0] == x_dims[0])) {
Z
zyfncg 已提交
4162
    out->share_lod(x);
L
Leo Chen 已提交
4163
  }
4164
  out->set_dtype(x.dtype());
L
Leo Chen 已提交
4165 4166
}

4167 4168 4169 4170 4171 4172 4173 4174 4175 4176
void TopKInferMeta(const MetaTensor& x,
                   const Scalar& k_scalar,
                   int axis,
                   bool largest,
                   bool sorted,
                   MetaTensor* out,
                   MetaTensor* indices,
                   MetaConfig config) {
  auto input_dims = x.dims();
  const int& dim_size = input_dims.size();
4177 4178 4179 4180 4181 4182 4183 4184 4185 4186 4187 4188 4189 4190 4191 4192 4193
  if (dim_size != 0) {
    PADDLE_ENFORCE_EQ(
        (axis < dim_size) && (axis >= (-1 * dim_size)),
        true,
        phi::errors::InvalidArgument(
            "the axis of topk must be [-%d, %d), but you set axis is %d",
            dim_size,
            dim_size,
            axis));
  } else {
    PADDLE_ENFORCE_EQ(
        (axis == dim_size) || (axis == -1),
        true,
        phi::errors::InvalidArgument("the axis of topk must be 0 or -1 when "
                                     "x.dims() = 0, but you set axis is %d",
                                     axis));
  }
4194 4195 4196 4197 4198 4199 4200 4201 4202 4203 4204 4205 4206 4207 4208 4209 4210

  if (axis < 0) axis += dim_size;

  int k = k_scalar.to<int>();
  if (k_scalar.FromTensor()) {
    k = -1;
  } else {
    PADDLE_ENFORCE_EQ(k >= 1,
                      true,
                      phi::errors::InvalidArgument(
                          "the attribute of k in the topk must >= 1 or be a "
                          "Tensor, but received %d .",
                          k));
  }

  PADDLE_ENFORCE_GE(
      input_dims.size(),
4211 4212
      0,
      phi::errors::InvalidArgument("input of topk must have >= 0d shape"));
4213 4214

  phi::DDim dims = input_dims;
4215 4216 4217
  if (input_dims.size() > 0) {
    dims[axis] = k;
  }
4218 4219 4220 4221 4222 4223 4224 4225
  out->set_dims(dims);
  out->share_lod(x);
  out->set_dtype(x.dtype());
  indices->set_dims(dims);
  indices->share_lod(x);
  indices->set_dtype(DataType::INT64);
}

C
Chen Weihang 已提交
4226 4227 4228 4229 4230 4231
void TraceInferMeta(
    const MetaTensor& x, int offset, int axis1, int axis2, MetaTensor* out) {
  int dim1 = axis1;
  int dim2 = axis2;

  auto x_dims = x.dims();
C
chentianyu03 已提交
4232

C
Chen Weihang 已提交
4233 4234 4235 4236 4237 4238
  int dim1_ = dim1 < 0 ? x_dims.size() + dim1 : dim1;
  int dim2_ = dim2 < 0 ? x_dims.size() + dim2 : dim2;

  PADDLE_ENFORCE_GE(
      x_dims.size(),
      2,
4239
      phi::errors::OutOfRange(
X
XiangGao 已提交
4240
          "Input(x)'s dim is out of range (expected at least 2, but got %ld).",
C
Chen Weihang 已提交
4241 4242 4243 4244
          x_dims.size()));
  PADDLE_ENFORCE_LT(
      dim1_,
      x_dims.size(),
4245
      phi::errors::OutOfRange(
X
XiangGao 已提交
4246 4247 4248 4249 4250 4251 4252 4253 4254 4255
          "axis1 is out of range (expected to be in range of [%ld, "
          "%ld], but got %ld).",
          -(x_dims.size()),
          (x_dims.size() - 1),
          dim1));
  PADDLE_ENFORCE_GE(
      dim1_,
      0,
      phi::errors::OutOfRange(
          "axis1 is out of range (expected to be in range of [%ld, "
C
Chen Weihang 已提交
4256 4257 4258 4259 4260 4261 4262
          "%ld], but got %ld).",
          -(x_dims.size()),
          (x_dims.size() - 1),
          dim1));
  PADDLE_ENFORCE_LT(
      dim2_,
      x_dims.size(),
4263
      phi::errors::OutOfRange(
X
XiangGao 已提交
4264 4265 4266 4267 4268 4269 4270 4271 4272 4273
          "axis2 is out of range (expected to be in range of [%ld, "
          "%ld], but got %ld).",
          -(x_dims.size()),
          (x_dims.size() - 1),
          dim2));
  PADDLE_ENFORCE_GE(
      dim2_,
      0,
      phi::errors::OutOfRange(
          "axis2 is out of range (expected to be in range of [%ld, "
C
Chen Weihang 已提交
4274 4275 4276 4277 4278 4279 4280
          "%ld], but got %ld).",
          -(x_dims.size()),
          (x_dims.size() - 1),
          dim2));
  PADDLE_ENFORCE_NE(
      dim1_,
      dim2_,
4281 4282 4283 4284
      phi::errors::InvalidArgument("The dimensions should not be identical "
                                   "%ld vs %ld.",
                                   dim1,
                                   dim2));
C
Chen Weihang 已提交
4285 4286 4287 4288 4289 4290 4291 4292

  auto sizes = vectorize(x_dims);
  if (x_dims.size() == 2) {
    sizes.clear();
  } else {
    sizes.erase(sizes.begin() + std::max(dim1_, dim2_));
    sizes.erase(sizes.begin() + std::min(dim1_, dim2_));
  }
4293
  out->set_dims(phi::make_ddim(sizes));
C
Chen Weihang 已提交
4294
  out->set_dtype(x.dtype());
C
chentianyu03 已提交
4295 4296
}

Z
zyfncg 已提交
4297
void TransferLayoutInferMeta(const MetaTensor& x,
4298 4299
                             int src_layout,
                             int dst_layout,
Z
zyfncg 已提交
4300 4301 4302
                             MetaTensor* out) {
  out->set_dims(x.dims());
  out->set_dtype(x.dtype());
4303 4304
  out->set_layout(static_cast<DataLayout>(dst_layout));
  out->share_lod(x);
Z
zyfncg 已提交
4305
}
H
hong 已提交
4306

Z
zyfncg 已提交
4307 4308 4309 4310
void TransposeInferMeta(const MetaTensor& x,
                        const std::vector<int>& axis,
                        MetaTensor* out) {
  auto x_dims = x.dims();
4311 4312
  int x_rank = x_dims.size();
  int axis_size = axis.size();
H
hong 已提交
4313

4314 4315
  // Note: x_rank > axis_size when fuse squeeze2 + transpose2, else x_rank ==
  // axis_size
4316 4317 4318 4319 4320 4321 4322 4323 4324
  PADDLE_ENFORCE_GE(x_rank,
                    axis_size,
                    errors::InvalidArgument(
                        "The input tensor's dimension "
                        "should be equal to or greater than the axis's size. "
                        "But received input tensor's dimension is %d, "
                        "axis's size is %d",
                        x_rank,
                        axis_size));
H
hong 已提交
4325

4326
  std::vector<int> formated_axis = axis;
Z
zyfncg 已提交
4327
  std::vector<int> count(axis_size, 0);
4328 4329 4330 4331 4332 4333 4334 4335 4336 4337 4338 4339 4340 4341 4342 4343 4344 4345 4346
  for (int i = 0; i < axis_size; i++) {
    PADDLE_ENFORCE_LT(axis[i],
                      x_rank,
                      errors::InvalidArgument(
                          "The reduce dim index %d should be in the "
                          "range [ -dimension(X), dimension(X) ) "
                          "which dimesion = %d. But received dim index = %d.",
                          i,
                          x_rank,
                          axis[i]));
    PADDLE_ENFORCE_GE(axis[i],
                      -x_rank,
                      errors::InvalidArgument(
                          "The reduce dim index %d should be in the "
                          "range [ -dimension(X), dimension(X) )  "
                          "which dimesion = %d. But received dim index = %d.",
                          i,
                          x_rank,
                          axis[i]));
Z
zyfncg 已提交
4347

4348 4349 4350
    if (axis[i] < 0) {
      formated_axis[i] = axis[i] + x_rank;
    }
Z
zyfncg 已提交
4351
    PADDLE_ENFORCE_EQ(
4352 4353 4354 4355 4356 4357
        ++count[formated_axis[i]],
        1,
        errors::InvalidArgument("Each element of axis should be unique. but "
                                "axis[%d] is %d appear not only once",
                                i,
                                axis[i]));
H
hong 已提交
4358
  }
Z
zyfncg 已提交
4359 4360

  phi::DDim out_dims(x_dims);
4361 4362
  for (int i = 0; i < axis_size; ++i) {
    out_dims[i] = x_dims[formated_axis[i]];
Z
zyfncg 已提交
4363 4364 4365 4366 4367 4368 4369 4370
  }

  out->set_dims(out_dims);
  out->set_dtype(x.dtype());
}

void UnbindInferMeta(const MetaTensor& x,
                     int axis,
4371
                     std::vector<MetaTensor*> outs) {
Z
zyfncg 已提交
4372 4373
  auto in_dims = x.dims();
  std::vector<int> out_dim;
4374 4375 4376 4377 4378 4379 4380 4381 4382 4383 4384 4385

  PADDLE_ENFORCE_GE(
      axis,
      -in_dims.size(),
      phi::errors::InvalidArgument(
          "axis must be in range(%d, %d).", -in_dims.size(), in_dims.size()));
  PADDLE_ENFORCE_LT(
      axis,
      in_dims.size(),
      phi::errors::InvalidArgument(
          "axis must be in range(%d, %d).", -in_dims.size(), in_dims.size()));

Z
zyfncg 已提交
4386
  axis = axis < 0 ? in_dims.size() + axis : axis;
4387

Z
zyfncg 已提交
4388 4389 4390 4391 4392
  for (int i = 0; i < in_dims.size(); ++i) {
    if (i != axis) out_dim.push_back(in_dims[i]);
  }
  auto out_dims = phi::make_ddim(out_dim);

4393 4394 4395 4396 4397
  for (size_t i = 0; i < outs.size(); ++i) {
    outs[i]->set_dtype(x.dtype());
    outs[i]->set_dims(out_dims);
    outs[i]->set_layout(x.layout());
    outs[i]->share_lod(x);
Z
zyfncg 已提交
4398 4399 4400
  }
}

4401 4402 4403 4404
void TrilTriuInferMeta(const MetaTensor& x,
                       int diagonal,
                       bool lower,
                       MetaTensor* out) {
4405 4406 4407 4408 4409 4410 4411 4412 4413 4414
  const auto& x_dims = x.dims();
  PADDLE_ENFORCE_GE(x_dims.size(),
                    2,
                    phi::errors::InvalidArgument(
                        "Input(X)'s rank must be at least 2 in TrilTriuOp."));
  out->set_dims(x.dims());
  out->share_lod(x);
  out->set_dtype(x.dtype());
}

4415 4416 4417 4418 4419 4420 4421 4422
void TrilInferMeta(const MetaTensor& x, int diagonal, MetaTensor* out) {
  TrilTriuInferMeta(x, diagonal, true, out);
}

void TriuInferMeta(const MetaTensor& x, int diagonal, MetaTensor* out) {
  TrilTriuInferMeta(x, diagonal, false, out);
}

4423 4424 4425 4426 4427 4428 4429
// Some operator having oneDnn kernel will be set layout in kernel.
void UnchangedExceptLayoutInferMeta(const MetaTensor& x, MetaTensor* out) {
  out->set_dims(x.dims());
  out->set_dtype(x.dtype());
  out->share_lod(x);
}

Z
zyfncg 已提交
4430 4431 4432 4433 4434 4435 4436 4437 4438
void UnchangedInferMeta(const MetaTensor& x, MetaTensor* out) {
  out->share_meta(x);
}

// meta x -> out without change, check if axis in range [-Rank(x), Rank(x)-1]
void UnchangedInferMetaCheckAxis(const MetaTensor& x,
                                 int axis,
                                 MetaTensor* out) {
  auto rank = x.dims().size();
4439 4440 4441 4442 4443 4444 4445 4446 4447 4448 4449 4450 4451 4452 4453 4454 4455 4456 4457 4458 4459 4460 4461
  if (rank > 0) {
    PADDLE_ENFORCE_GE(axis,
                      -rank,
                      phi::errors::InvalidArgument(
                          "Attr(axis) value should be in range [-R, R-1], "
                          "R is the rank of Input(X)."));
    PADDLE_ENFORCE_LT(axis,
                      rank,
                      phi::errors::InvalidArgument(
                          "Attr(axis) value should be in range [-R, R-1], "
                          "R is the rank of Input(X)."));
  } else if (rank == 0) {
    PADDLE_ENFORCE_GE(
        axis,
        -1,
        phi::errors::InvalidArgument("Attr(axis) value should be in range [-1, "
                                     "0] when input is 0D Tensor "));
    PADDLE_ENFORCE_LE(
        axis,
        0,
        phi::errors::InvalidArgument("Attr(axis) value should be in range [-1, "
                                     "0] when input is 0D Tensor "));
  }
Z
zyfncg 已提交
4462
  out->share_meta(x);
H
hong 已提交
4463 4464
}

4465 4466 4467 4468 4469 4470 4471 4472 4473 4474 4475 4476 4477 4478 4479 4480 4481 4482 4483 4484 4485
void UnfoldInferMeta(const MetaTensor& x,
                     const std::vector<int>& kernel_sizes,
                     const std::vector<int>& strides,
                     const std::vector<int>& paddings,
                     const std::vector<int>& dilations,
                     MetaTensor* out,
                     MetaConfig config) {
  auto in_dims = x.dims();
  // Only [N, C, H, W] input supported now
  PADDLE_ENFORCE_EQ(
      in_dims.size(),
      4,
      phi::errors::InvalidArgument(
          "Input should be 4-D tensor of format [N, C, H, W], but get %u",
          in_dims.size()));
  PADDLE_ENFORCE_EQ(
      in_dims.size() - kernel_sizes.size(),
      2U,
      phi::errors::InvalidArgument(
          "The dims of X should be larger than that of kernel_sizes "
          "by a number of 2, due to the batch size and input channel dim. "
4486
          "But received dims(X:%u) - dims(kernel_sizes:%u) != 2",
4487 4488 4489 4490 4491 4492 4493
          in_dims.size(),
          kernel_sizes.size()));
  PADDLE_ENFORCE_EQ(
      strides.size(),
      kernel_sizes.size(),
      phi::errors::InvalidArgument(
          "The dims of strides should be the same with that of kernel_sizes. "
4494
          "But received dims(strides: %u) != dims(kernel_sizes: %u).",
4495 4496 4497 4498 4499 4500 4501
          strides.size(),
          kernel_sizes.size()));
  PADDLE_ENFORCE_EQ(
      paddings.size(),
      2 * strides.size(),
      phi::errors::InvalidArgument(
          "The dims of paddings should be 2 times of that of strides. "
4502
          "But received dims(paddings: %u) != 2*dims(strides: %u).",
4503 4504 4505 4506 4507 4508 4509
          paddings.size(),
          strides.size()));
  PADDLE_ENFORCE_EQ(
      strides.size(),
      dilations.size(),
      phi::errors::InvalidArgument(
          "The dims of strides should be the same with that of dilations. "
4510
          "But received dims(strides: %u) != dims(dilations: %u).",
4511 4512 4513 4514 4515 4516 4517 4518
          strides.size(),
          dilations.size()));

  // check kernel_sizes
  PADDLE_ENFORCE_GT(kernel_sizes[0],
                    0,
                    phi::errors::InvalidArgument(
                        "The `kernel_sizes` should be greater than zero, "
4519
                        "but received kernel_height: %d kernel_width: %d.",
4520 4521 4522 4523 4524 4525
                        kernel_sizes[0],
                        kernel_sizes[1]));
  PADDLE_ENFORCE_GT(kernel_sizes[1],
                    0,
                    phi::errors::InvalidArgument(
                        "The `kernel_sizes` should be greater than zero, "
4526
                        "but received kernel_height: %d kernel_width: %d.",
4527 4528 4529 4530 4531 4532 4533
                        kernel_sizes[0],
                        kernel_sizes[1]));
  // check strides
  PADDLE_ENFORCE_GT(strides[0],
                    0,
                    phi::errors::InvalidArgument(
                        "The `strides` should be greater than zero, "
4534
                        "but received strides_height: %d strides_width: %d.",
4535 4536 4537 4538 4539 4540
                        strides[0],
                        strides[1]));
  PADDLE_ENFORCE_GT(strides[1],
                    0,
                    phi::errors::InvalidArgument(
                        "The `strides` should be greater than zero, "
4541
                        "but received strides_height: %d strides_width: %d.",
4542 4543 4544 4545 4546 4547 4548 4549
                        strides[0],
                        strides[1]));
  // check dilations
  PADDLE_ENFORCE_GT(
      dilations[0],
      0,
      phi::errors::InvalidArgument(
          "The `dilations` should be greater than zero, "
4550
          "but received dilations_height: %d dilations_width: %d.",
4551 4552 4553 4554 4555 4556 4557
          dilations[0],
          dilations[1]));
  PADDLE_ENFORCE_GT(
      dilations[1],
      0,
      phi::errors::InvalidArgument(
          "The `dilations` should be greater than zero, "
4558
          "but received dilations_height: %d dilations_width: %d.",
4559 4560 4561 4562 4563 4564 4565 4566 4567 4568 4569 4570 4571 4572 4573 4574 4575 4576 4577 4578
          dilations[0],
          dilations[1]));

  std::vector<int> out_dims;
  out_dims.push_back(in_dims[0]);
  int output_channels = in_dims[1] * kernel_sizes[0] * kernel_sizes[1];
  out_dims.push_back(output_channels);

  int output_height = phi::funcs::CalcOutputSize(in_dims[2],
                                                 kernel_sizes[0],
                                                 dilations[0],
                                                 paddings[0],
                                                 paddings[2],
                                                 strides[0]);
  int output_width = phi::funcs::CalcOutputSize(in_dims[3],
                                                kernel_sizes[1],
                                                dilations[1],
                                                paddings[1],
                                                paddings[3],
                                                strides[1]);
4579
  int output_col_length = output_height * output_width;
4580 4581 4582 4583 4584 4585 4586 4587 4588 4589 4590 4591 4592 4593 4594 4595 4596 4597 4598 4599 4600 4601 4602 4603 4604 4605 4606 4607 4608 4609 4610 4611 4612 4613 4614 4615 4616 4617
  if (config.is_runtime) {
    // only check output height and width in runtime
    PADDLE_ENFORCE_GT(
        output_height,
        0,
        phi::errors::InvalidArgument(
            "The sliding blocks calculated from input spatial size "
            "(%d, %d), kernel_sizes (%d, %d), strides (%d, %d), "
            "dilations (%d, %d), is (%d, %d), which should be a "
            "positive integer.",
            in_dims[2],
            in_dims[3],
            kernel_sizes[0],
            kernel_sizes[1],
            strides[0],
            strides[1],
            dilations[0],
            dilations[1],
            output_height,
            output_width));
    PADDLE_ENFORCE_GT(
        output_width,
        0,
        phi::errors::InvalidArgument(
            "The sliding blocks calculated from input spatial size "
            "(%d, %d), kernel_sizes (%d, %d), strides (%d, %d), "
            "dilations (%d, %d), is (%d, %d), which should be a "
            "positive integer.",
            in_dims[2],
            in_dims[3],
            kernel_sizes[0],
            kernel_sizes[1],
            strides[0],
            strides[1],
            dilations[0],
            dilations[1],
            output_height,
            output_width));
4618 4619 4620
  } else {
    output_col_length =
        output_height == -1 || output_width == -1 ? -1 : output_col_length;
4621 4622 4623 4624 4625
  }
  out_dims.push_back(output_col_length);
  out->set_dims(phi::make_ddim(out_dims));
}

4626 4627 4628 4629 4630 4631 4632 4633 4634 4635 4636 4637 4638 4639 4640 4641 4642 4643 4644 4645 4646 4647 4648 4649 4650 4651 4652 4653 4654 4655 4656 4657 4658 4659 4660 4661 4662
void UniformRandomInplaceInferMeta(const MetaTensor& x,
                                   float min,
                                   float max,
                                   int seed,
                                   int diag_num,
                                   int diag_step,
                                   float diag_val,
                                   MetaTensor* out) {
  PADDLE_ENFORCE_LT(
      min,
      max,
      errors::InvalidArgument(
          "The uniform_random's min must less then max. But received min = "
          "%f great than or equal max = %f.",
          min,
          max));
  PADDLE_ENFORCE_GE(diag_num,
                    0,
                    errors::InvalidArgument(
                        "The uniform_random's diag_num must greater than or "
                        "equal 0. But recevied diag_num (%d) < 0.",
                        diag_num));
  PADDLE_ENFORCE_GE(diag_step,
                    0,
                    errors::InvalidArgument(
                        "The uniform_random's diag_step must greater than or "
                        "equal 0. But recevied diag_step (%d) < 0.",
                        diag_step));
  PADDLE_ENFORCE_NE(out,
                    nullptr,
                    phi::errors::InvalidArgument(
                        "uniform_random should have output tensor out."));
  auto xdim = x.dims();
  out->set_dims(xdim);
  out->set_dtype(x.dtype());
}

4663 4664 4665 4666 4667 4668 4669 4670 4671 4672 4673 4674 4675 4676
void UniqueConsecutiveInferMeta(const MetaTensor& x,
                                bool return_inverse,
                                bool return_counts,
                                const std::vector<int>& axis,
                                int dtype,
                                MetaTensor* out,
                                MetaTensor* index,
                                MetaTensor* counts) {
  PADDLE_ENFORCE_NE(out,
                    nullptr,
                    phi::errors::InvalidArgument(
                        "unique_consecutive should have output tensor out."));

  auto in_dims = x.dims();
D
duanboqiang 已提交
4677 4678 4679 4680 4681 4682 4683 4684
  if (x.dims().size() == 0) {
    PADDLE_ENFORCE_EQ(axis.empty(),
                      true,
                      phi::errors::InvalidArgument(
                          "The Input(X) with 0-D Tensor, axis must be None"
                          "But now the axis is %d.",
                          axis[0]));
  }
4685 4686 4687 4688 4689 4690 4691 4692 4693 4694 4695 4696 4697 4698 4699 4700 4701 4702 4703 4704 4705 4706 4707 4708 4709 4710 4711 4712 4713 4714 4715 4716 4717 4718 4719 4720 4721 4722 4723 4724 4725 4726 4727 4728 4729 4730
  if (return_inverse) {
    PADDLE_ENFORCE_NE(
        index,
        nullptr,
        phi::errors::InvalidArgument("Tensor index should not be null if "
                                     "return_inverse is set to True."));
  }
  if (return_counts) {
    PADDLE_ENFORCE_NE(
        counts,
        nullptr,
        phi::errors::InvalidArgument("Tensor counts should not be null if "
                                     "return_counts is set to True."));
  }

  if (axis.empty()) {
    out->set_dims({-1});
    out->set_dtype(x.dtype());
    if (return_inverse) {
      index->set_dims({phi::product(in_dims)});
    }
  } else {
    int axis_value = axis[0];
    if (axis_value < 0) {
      axis_value += in_dims.size();
    }
    PADDLE_ENFORCE_LT(
        axis_value,
        in_dims.size(),
        phi::errors::InvalidArgument("The axis(%d) should be less than "
                                     "the dimension size(%d) of x.",
                                     axis_value,
                                     in_dims.size()));
    auto out_dims = in_dims;
    out_dims[axis_value] = -1;
    out->set_dims(out_dims);
    out->set_dtype(x.dtype());
    if (return_inverse) {
      index->set_dims({in_dims[axis_value]});
    }
  }
  if (return_counts) {
    counts->set_dims({-1});
  }
}

C
csy0225 已提交
4731 4732 4733 4734 4735 4736 4737 4738 4739 4740 4741 4742 4743 4744 4745 4746 4747 4748 4749 4750 4751 4752 4753 4754 4755 4756 4757 4758 4759 4760 4761 4762 4763 4764 4765 4766
void UniqueInferMeta(const MetaTensor& x,
                     bool return_index,
                     bool return_inverse,
                     bool return_counts,
                     const std::vector<int>& axis,
                     DataType dtype,
                     MetaTensor* out,
                     MetaTensor* indices,
                     MetaTensor* index,
                     MetaTensor* counts) {
  bool is_sorted = true;
  UniqueRawInferMeta(x,
                     return_index,
                     return_inverse,
                     return_counts,
                     axis,
                     dtype,
                     is_sorted,
                     out,
                     indices,
                     index,
                     counts);
}

void UniqueRawInferMeta(const MetaTensor& x,
                        bool return_index,
                        bool return_inverse,
                        bool return_counts,
                        const std::vector<int>& axis,
                        DataType dtype,
                        bool is_sorted,
                        MetaTensor* out,
                        MetaTensor* indices,
                        MetaTensor* index,
                        MetaTensor* counts) {
  if (!is_sorted) {
S
sprouteer 已提交
4767 4768 4769 4770 4771 4772
    PADDLE_ENFORCE_EQ(x.dims().size() == 1 || x.dims().size() == 0,
                      true,
                      phi::errors::InvalidArgument(
                          "The Input(X) should be 0-D or 1-D Tensor, "
                          "But now the dims of Input(X) is %d.",
                          x.dims().size()));
C
csy0225 已提交
4773 4774 4775 4776 4777
    out->set_dims(phi::make_ddim({-1}));
    index->set_dims(x.dims());
    return;
  }

S
sprouteer 已提交
4778 4779 4780 4781 4782 4783 4784 4785 4786
  if (x.dims().size() == 0) {
    PADDLE_ENFORCE_EQ(axis.empty(),
                      true,
                      phi::errors::InvalidArgument(
                          "The Input(X) with 0-D Tensor, axis must be None"
                          "But now the axis is %d.",
                          axis[0]));
  }

C
csy0225 已提交
4787 4788 4789 4790 4791 4792 4793 4794 4795 4796
  if (axis.empty()) {
    out->set_dims(phi::make_ddim({-1}));
    if (return_inverse) {
      index->set_dims(phi::make_ddim({phi::product(x.dims())}));
    }
  } else {
    int axis_value = axis[0];
    if (axis_value < 0) {
      axis_value += x.dims().size();
    }
4797

C
csy0225 已提交
4798 4799 4800 4801 4802 4803 4804
    PADDLE_ENFORCE_LT(
        axis_value,
        x.dims().size(),
        phi::errors::InvalidArgument("The axis(%d) should be less than "
                                     "the dimension size(%d) of x.",
                                     axis_value,
                                     x.dims().size()));
4805 4806 4807 4808 4809 4810 4811 4812
    PADDLE_ENFORCE_GE(
        axis_value,
        0,
        phi::errors::InvalidArgument(
            "The axis(%d) + rank(x) (%d) should be greater than or equal to 0.",
            axis_value,
            -x.dims().size()));

C
csy0225 已提交
4813 4814 4815 4816 4817
    auto out_dims = x.dims();
    out_dims[axis_value] = -1;
    out->set_dims(out_dims);
    if (return_inverse) {
      index->set_dims(phi::make_ddim({x.dims()[axis_value]}));
4818
      index->set_dtype(dtype);
C
csy0225 已提交
4819 4820 4821 4822
    }
  }
  if (return_index) {
    indices->set_dims(phi::make_ddim({-1}));
4823
    indices->set_dtype(dtype);
C
csy0225 已提交
4824 4825 4826
  }
  if (return_counts) {
    counts->set_dims(phi::make_ddim({-1}));
4827
    counts->set_dtype(dtype);
C
csy0225 已提交
4828 4829 4830
  }
}

4831
void UnsqueezeInferMeta(const MetaTensor& x,
4832
                        const IntArray& axes,
4833 4834
                        MetaTensor* out,
                        MetaConfig config) {
4835 4836 4837 4838 4839 4840 4841 4842
  const auto& x_dims = x.dims();
  // Validity Check: input tensor dims (<6).
  PADDLE_ENFORCE_LE(x_dims.size(),
                    6,
                    phi::errors::InvalidArgument(
                        "Invalid "
                        "dimensions, the rank of Input(X) "
                        "should be in the range of [1, 6] (Eigen limit)"));
4843 4844 4845 4846 4847 4848
  if (!config.is_runtime && axes.FromTensor()) {
    // compile time infershape.  set all elements to -1.
    int output_size = x.dims().size() + axes.GetData().size();
    std::vector<int64_t> vec_out_dims(output_size, -1);
    out->set_dtype(x.dtype());
    out->set_dims(phi::make_ddim(vec_out_dims));
4849 4850
  } else {
    auto out_dims = funcs::GetUnsqueezeShape(axes.GetData(), x_dims);
4851
    out->set_dims(out_dims);
4852
    if (x_dims.size() > 0 && x_dims[0] == out_dims[0]) {
4853 4854
      out->share_lod(x);
    }
4855
    out->set_dtype(x.dtype());
4856
  }
4857
}
4858

4859 4860 4861 4862 4863 4864 4865 4866 4867 4868 4869 4870 4871 4872
void UnsqueezeWithXShapeInferMeta(const MetaTensor& x,
                                  const IntArray& axes,
                                  MetaTensor* out,
                                  MetaTensor* xshape,
                                  MetaConfig config) {
  const auto& x_dims = x.dims();
  UnsqueezeInferMeta(x, axes, out, config);
  // set xshape dims.
  std::vector<int64_t> xshape_dims(x_dims.size() + 1);
  xshape_dims[0] = 0;
  for (int i = 0; i < x_dims.size(); ++i) {
    xshape_dims[i + 1] = x_dims[i];
  }
  if (xshape) {
4873 4874 4875
    xshape->set_dims(phi::make_ddim(xshape_dims));
    xshape->share_lod(x);
    xshape->set_dtype(x.dtype());
4876 4877 4878
  }
}

C
csy0225 已提交
4879 4880 4881 4882 4883 4884 4885 4886 4887 4888 4889 4890 4891 4892 4893 4894 4895 4896 4897 4898 4899 4900 4901 4902 4903 4904 4905 4906 4907 4908 4909 4910 4911 4912 4913 4914 4915 4916 4917 4918 4919 4920 4921 4922 4923 4924 4925
void UnStackInferMeta(const MetaTensor& x,
                      int axis,
                      int num,
                      std::vector<MetaTensor*> outs) {
  auto x_dim = x.dims();
  int rank = x_dim.size();
  PADDLE_ENFORCE_GE(axis,
                    -rank,
                    phi::errors::InvalidArgument(
                        "The attribute axis is out of range, it must be "
                        "inside [-rank, rank), where rank = %d",
                        rank));
  PADDLE_ENFORCE_LT(axis,
                    rank,
                    phi::errors::InvalidArgument(
                        "The attribute axis is out of range, it must be "
                        "inside [-rank, rank), where rank = %d",
                        rank));
  if (axis < 0) axis += rank;

  size_t output_count = outs.size();
  PADDLE_ENFORCE_EQ(output_count,
                    static_cast<size_t>(num),
                    phi::errors::InvalidArgument(
                        "Number of Outputs(Y) is wrong. Got %d , but it must "
                        "equal to attribute num which is %d.",
                        output_count,
                        static_cast<size_t>(num)));
  if (x_dim[axis] > 0) {
    PADDLE_ENFORCE_EQ(
        num,
        x_dim[axis],
        phi::errors::InvalidArgument(
            "The number of attribute num is not equal to the length of the "
            "%d axis of Input(X). Expect %d but got %d.",
            axis,
            x_dim[axis],
            num));
  }
  auto vec = phi::vectorize<int>(x_dim);
  vec.erase(vec.begin() + axis);
  for (size_t i = 0; i < output_count; i++) {
    outs[i]->set_dims(phi::make_ddim(vec));
    outs[i]->set_dtype(x.dtype());
  }
}

4926 4927 4928 4929 4930 4931 4932 4933 4934 4935 4936 4937 4938 4939 4940 4941 4942 4943 4944 4945 4946 4947 4948 4949 4950 4951 4952 4953 4954 4955 4956 4957 4958 4959 4960 4961 4962 4963 4964 4965 4966 4967 4968 4969 4970 4971
void ChannelShuffleInferMeta(const MetaTensor& x,
                             int groups,
                             const std::string& data_format,
                             MetaTensor* out) {
  auto input_dims = x.dims();
  PADDLE_ENFORCE_EQ(input_dims.size(),
                    4,
                    phi::errors::InvalidArgument(
                        "Input should be a 4-D tensor of format [N, C, H, W] "
                        "or [N, H, W, C], but got %u.",
                        input_dims.size()));
  PADDLE_ENFORCE_GE(
      groups,
      1,
      phi::errors::InvalidArgument("groups should be larger than 0."));
  PADDLE_ENFORCE_EQ(data_format == "NCHW" || data_format == "NHWC",
                    true,
                    phi::errors::InvalidArgument(
                        "data_format must be one of "
                        "NCHW and NHWC. But recevied data_format: %s",
                        data_format));

  const bool channel_last = (data_format == "NHWC");

  if (!channel_last) {
    PADDLE_ENFORCE_EQ(input_dims[1] % groups,
                      0,
                      phi::errors::InvalidArgument(
                          "The number of groups to divide channels in [%u] "
                          "should divide the number of channel [%u]",
                          groups,
                          input_dims[1]));
  } else {
    PADDLE_ENFORCE_EQ(input_dims[3] % groups,
                      0,
                      phi::errors::InvalidArgument(
                          "The number of groups to divide channels in [%u] "
                          "should divide the number of channel [%u]",
                          groups,
                          input_dims[3]));
  }
  auto output_dims = input_dims;
  out->set_dtype(x.dtype());
  out->set_dims(output_dims);
}

4972
}  // namespace phi
4973

4974
PD_REGISTER_INFER_META_FN(flatten, phi::FlattenInferMeta);