unique_op.h 15.8 KB
Newer Older
Z
zhoukunsheng 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

#pragma once
Z
Zhang Ting 已提交
16
#include <algorithm>
Z
zhoukunsheng 已提交
17
#include <cmath>
Z
Zhang Ting 已提交
18 19
#include <numeric>
#include <set>
Z
zhoukunsheng 已提交
20 21 22 23
#include <unordered_map>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
Z
Zhang Ting 已提交
24 25
#include "paddle/fluid/operators/math/concat_and_split.h"
#include "paddle/fluid/operators/transpose_op.h"
26
#include "paddle/pten/kernels/funcs/math_function.h"
Z
zhoukunsheng 已提交
27 28 29 30 31 32 33 34 35

namespace paddle {
namespace operators {

template <typename InT>
struct UniqueOpFunctor {
  framework::Tensor* out_;
  framework::Tensor* index_;
  const framework::Tensor* in_;
36
  framework::Tensor* count_;
Z
zhoukunsheng 已提交
37 38

  UniqueOpFunctor(framework::Tensor* out, framework::Tensor* index,
39 40 41
                  const framework::Tensor* in,
                  framework::Tensor* count = nullptr)
      : out_(out), index_(index), in_(in), count_(count) {}
Z
zhoukunsheng 已提交
42 43 44 45 46 47 48 49 50 51 52 53

  template <typename IndexT>
  void apply() const {
    auto* in_data = in_->data<InT>();
    auto* index_data = index_->mutable_data<IndexT>(platform::CPUPlace());

    int64_t j = 0;

    // TODO(fangzeyang): Should optimize performance here.
    std::unordered_map<InT, int64_t> dict;
    std::vector<InT> uniq;

54 55 56 57 58 59
    PADDLE_ENFORCE_LT(
        in_->numel(), pow(2, 31),
        platform::errors::InvalidArgument(
            "The num of Input(X) elements should be less then INT_MAX, "
            "but received num is %d.",
            in_->numel()));
Z
zhoukunsheng 已提交
60 61 62 63

    for (auto i = 0; i < in_->numel(); i++) {
      auto it = dict.find(in_data[i]);
      if (it == dict.end()) {
64 65
        dict.emplace(std::make_pair(in_data[i], j));
        uniq.emplace_back(in_data[i]);
Z
zhoukunsheng 已提交
66 67 68 69 70 71 72
        index_data[i] = static_cast<IndexT>(j);
        j++;
      } else {
        index_data[i] = static_cast<IndexT>(it->second);
      }
    }

73 74 75 76 77 78 79
    if (count_ != nullptr) {
      // Resize the count tensor dims to allocate the memory
      count_->Resize(framework::make_ddim({static_cast<int64_t>(uniq.size())}));
      IndexT* count_data = count_->mutable_data<IndexT>(platform::CPUPlace());
      // init count_data to 0
      memset(count_data, 0, uniq.size() * sizeof(IndexT));

80
      const auto& index_type = framework::TransToProtoVarType(index_->dtype());
81 82
      bool index_type_match = index_type == framework::proto::VarType::INT32 ||
                              index_type == framework::proto::VarType::INT64;
83 84 85 86 87 88 89 90 91
      PADDLE_ENFORCE_EQ(index_type_match, true,
                        platform::errors::InvalidArgument(
                            "Index holds the wrong type, it holds %s, "
                            "but desires to be %s or %s",
                            paddle::framework::DataTypeToString(index_type),
                            paddle::framework::DataTypeToString(
                                framework::proto::VarType::INT32),
                            paddle::framework::DataTypeToString(
                                framework::proto::VarType::INT64)));
92 93 94 95 96 97 98 99 100 101 102 103 104 105

      if (index_type == framework::proto::VarType::INT32) {
        for (auto i = 0; i < in_->numel(); ++i) {
          const IndexT& index = index_data[i];
          count_data[static_cast<int32_t>(index)] += static_cast<IndexT>(1);
        }
      } else {
        for (auto i = 0; i < in_->numel(); ++i) {
          const IndexT& index = index_data[i];
          count_data[static_cast<int64_t>(index)] += static_cast<IndexT>(1);
        }
      }
    }

Z
zhoukunsheng 已提交
106 107 108 109 110 111
    out_->Resize(framework::make_ddim({static_cast<int64_t>(uniq.size())}));
    auto out_data = out_->mutable_data<InT>(platform::CPUPlace());
    std::memcpy(out_data, uniq.data(), uniq.size() * sizeof(InT));
  }
};

Z
Zhang Ting 已提交
112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133
static std::vector<framework::Tensor> Unbind(const framework::Tensor& in) {
  int64_t size = in.dims()[0];
  std::vector<framework::Tensor> tensors(size);
  for (int64_t i = 0; i < size; ++i) {
    tensors[i] = in.Slice(i, i + 1);
  }
  return tensors;
}

template <typename T>
static bool Equal(const framework::Tensor& a, const framework::Tensor& b) {
  if (a.numel() != b.numel()) {
    return false;
  }
  for (int64_t i = 0; i < a.numel(); ++i) {
    if (a.data<T>()[i] != b.data<T>()[i]) {
      return false;
    }
  }
  return true;
}

Z
Zhang Ting 已提交
134
template <typename InT, typename IndexT>
Z
Zhang Ting 已提交
135 136 137 138
static void UniqueFlattendTensor(const framework::ExecutionContext& context,
                                 const framework::Tensor& in,
                                 framework::Tensor* out, bool return_index,
                                 bool return_inverse, bool return_counts) {
Z
Zhang Ting 已提交
139 140
  const InT* in_data = in.data<InT>();
  std::set<InT> unique(in_data, in_data + in.numel());
Z
Zhang Ting 已提交
141
  out->Resize(framework::make_ddim({static_cast<int64_t>(unique.size())}));
Z
Zhang Ting 已提交
142
  auto out_data = out->mutable_data<InT>(context.GetPlace());
Z
Zhang Ting 已提交
143 144 145 146 147
  std::copy(unique.begin(), unique.end(), out_data);

  if (return_index) {
    auto* indices = context.Output<framework::Tensor>("Indices");
    indices->Resize(framework::make_ddim({out->numel()}));
Z
Zhang Ting 已提交
148 149
    auto indices_data = indices->mutable_data<IndexT>(context.GetPlace());
    std::unordered_map<InT, IndexT> indices_map;
Z
Zhang Ting 已提交
150 151 152 153 154 155 156 157 158 159 160 161 162
    indices_map.reserve(out->numel());
    for (int64_t i = 0; i < in.numel(); ++i) {
      if (indices_map.find(in_data[i]) != indices_map.end()) continue;
      indices_map[in_data[i]] = i;
    }
    for (int64_t i = 0; i < out->numel(); ++i) {
      indices_data[i] = indices_map[out_data[i]];
    }
  }

  if (return_inverse) {
    auto* inverse = context.Output<framework::Tensor>("Index");
    inverse->Resize(framework::make_ddim({in.numel()}));
Z
Zhang Ting 已提交
163 164
    auto inverse_data = inverse->mutable_data<IndexT>(context.GetPlace());
    std::unordered_map<InT, IndexT> inverse_map;
Z
Zhang Ting 已提交
165 166 167 168 169 170 171 172 173 174 175 176
    inverse_map.reserve(out->numel());
    for (int64_t i = 0; i < out->numel(); ++i) {
      inverse_map[out_data[i]] = i;
    }
    for (int64_t i = 0; i < in.numel(); ++i) {
      inverse_data[i] = inverse_map[in_data[i]];
    }
  }

  if (return_counts) {
    auto* count = context.Output<framework::Tensor>("Counts");
    count->Resize(framework::make_ddim({out->numel()}));
Z
Zhang Ting 已提交
177 178
    auto count_data = count->mutable_data<IndexT>(context.GetPlace());
    std::unordered_map<InT, IndexT> counts_map;
Z
Zhang Ting 已提交
179 180 181 182 183 184 185 186 187 188 189 190 191
    counts_map.reserve(out->numel());
    for (int64_t i = 0; i < out->numel(); ++i) {
      counts_map[out_data[i]] = 0;
    }
    for (int64_t i = 0; i < in.numel(); i++) {
      counts_map[in_data[i]] += 1;
    }
    for (int64_t i = 0; i < out->numel(); i++) {
      count_data[i] = counts_map[out_data[i]];
    }
  }
}

Z
Zhang Ting 已提交
192
template <class ForwardIt, typename InT, typename IndexT>
Z
Zhang Ting 已提交
193 194
static ForwardIt UniqueDimImpl(const framework::ExecutionContext& context,
                               ForwardIt first, ForwardIt last,
Z
Zhang Ting 已提交
195 196 197 198
                               const std::vector<IndexT>& sorted_indices_vec,
                               std::vector<IndexT>* inverse_vec,
                               std::vector<IndexT>* counts_vec,
                               std::vector<IndexT>* indices_vec) {
Z
Zhang Ting 已提交
199 200 201 202 203 204 205 206 207 208 209 210 211 212
  if (first == last) {
    return last;
  }

  (*inverse_vec)[sorted_indices_vec[0]] = 0;
  (*counts_vec)[0] = 1;
  (*indices_vec)[0] = sorted_indices_vec[0];

  ForwardIt begin = first;
  ForwardIt result = first;

  while (++first != last) {
    int64_t idx_first = std::distance(begin, first);
    int64_t idx_result = std::distance(begin, result);
Z
Zhang Ting 已提交
213
    if (!Equal<InT>(*result, *first)) {
Z
Zhang Ting 已提交
214 215 216 217 218 219 220 221 222 223 224 225
      if (++result != first) {
        *result = std::move(*first);
      }
      idx_result += 1;
      (*indices_vec)[idx_result] = sorted_indices_vec[idx_first];
    }
    (*inverse_vec)[sorted_indices_vec[idx_first]] = idx_result;
    (*counts_vec)[idx_result] += 1;
  }
  return ++result;
}

Z
Zhang Ting 已提交
226
template <typename DeviceContext, typename InT, typename IndexT>
Z
Zhang Ting 已提交
227 228 229 230 231 232 233 234 235 236 237 238 239 240 241
static void UniqueDim(const framework::ExecutionContext& context,
                      const framework::Tensor& in, framework::Tensor* out,
                      bool return_index, bool return_inverse,
                      bool return_counts, int axis) {
  // transpose tensor: eg. axis=1, [dim0, dim1, dim2] -> [dim1, dim0, dim2]
  std::vector<int> permute(in.dims().size());
  std::iota(permute.begin(), permute.end(), 0);
  permute[axis] = 0;
  permute[0] = axis;
  std::vector<int64_t> in_trans_dims_vec(framework::vectorize(in.dims()));
  in_trans_dims_vec[axis] = in.dims()[0];
  in_trans_dims_vec[0] = in.dims()[axis];
  framework::Tensor in_trans;
  framework::DDim in_trans_dims = framework::make_ddim(in_trans_dims_vec);
  in_trans.Resize(in_trans_dims);
Z
Zhang Ting 已提交
242
  in_trans.mutable_data<InT>(context.GetPlace());
Z
Zhang Ting 已提交
243
  auto& dev_ctx = context.template device_context<DeviceContext>();
Z
Zhang Ting 已提交
244 245
  TransCompute<DeviceContext, InT>(in.dims().size(), dev_ctx, in, &in_trans,
                                   permute);
Z
Zhang Ting 已提交
246 247 248 249 250 251
  // reshape tensor: eg. [dim1, dim0, dim2] -> [dim1, dim0*dim2]
  framework::DDim in_trans_flat_dims =
      framework::flatten_to_2d(in_trans_dims, 1);
  in_trans.Resize(in_trans_flat_dims);

  // sort indices
Z
Zhang Ting 已提交
252
  std::vector<IndexT> sorted_indices_vec(in_trans.dims()[0]);
Z
Zhang Ting 已提交
253 254
  std::iota(sorted_indices_vec.begin(), sorted_indices_vec.end(), 0);
  int64_t col = in_trans.dims()[1];
Z
Zhang Ting 已提交
255
  const InT* in_trans_data = in_trans.data<InT>();
Z
Zhang Ting 已提交
256 257 258
  std::sort(sorted_indices_vec.begin(), sorted_indices_vec.end(),
            [&](int64_t a, int64_t b) -> bool {
              for (int64_t i = 0; i < col; ++i) {
Z
Zhang Ting 已提交
259 260
                InT lhs = in_trans_data[i + a * col];
                InT rhs = in_trans_data[i + b * col];
Z
Zhang Ting 已提交
261 262 263 264 265 266 267 268 269 270 271 272
                if (lhs < rhs) {
                  return true;
                } else if (lhs > rhs) {
                  return false;
                }
              }
              return false;
            });

  // sort tensor according to indices
  framework::Tensor input_sorted;
  input_sorted.Resize(in_trans_dims);
Z
Zhang Ting 已提交
273 274
  input_sorted.mutable_data<InT>(context.GetPlace());
  InT* input_sorted_data = input_sorted.data<InT>();
Z
Zhang Ting 已提交
275 276
  for (size_t i = 0; i < sorted_indices_vec.size(); ++i) {
    memcpy(input_sorted_data + i * col,
Z
Zhang Ting 已提交
277 278
           in_trans_data + static_cast<int64_t>(sorted_indices_vec[i]) * col,
           col * sizeof(InT));
Z
Zhang Ting 已提交
279 280 281
  }

  std::vector<framework::Tensor> input_unbind = Unbind(input_sorted);
Z
Zhang Ting 已提交
282 283 284 285
  std::vector<IndexT> inverse_vec(sorted_indices_vec.size(), 0);
  std::vector<IndexT> counts_vec(sorted_indices_vec.size(), 0);
  std::vector<IndexT> indices_vec(sorted_indices_vec.size(), 0);
  auto last = UniqueDimImpl<std::vector<framework::Tensor>::iterator, InT>(
Z
Zhang Ting 已提交
286 287 288 289 290 291 292
      context, input_unbind.begin(), input_unbind.end(), sorted_indices_vec,
      &inverse_vec, &counts_vec, &indices_vec);
  input_unbind.erase(last, input_unbind.end());
  counts_vec.erase(counts_vec.begin() + input_unbind.size(), counts_vec.end());
  indices_vec.erase(indices_vec.begin() + input_unbind.size(),
                    indices_vec.end());

Z
Zhang Ting 已提交
293
  math::ConcatFunctor<DeviceContext, InT> concat_functor;
Z
Zhang Ting 已提交
294 295 296 297
  framework::Tensor out_trans;
  std::vector<int64_t> out_trans_dims_vec = in_trans_dims_vec;
  out_trans_dims_vec[0] = input_unbind.size();
  out_trans.Resize(framework::make_ddim(out_trans_dims_vec));
Z
Zhang Ting 已提交
298
  out_trans.mutable_data<InT>(context.GetPlace());
Z
Zhang Ting 已提交
299 300
  std::swap(out_trans_dims_vec[0], out_trans_dims_vec[axis]);
  out->Resize(framework::make_ddim(out_trans_dims_vec));
Z
Zhang Ting 已提交
301
  out->mutable_data<InT>(context.GetPlace());
Z
Zhang Ting 已提交
302
  concat_functor(dev_ctx, input_unbind, 0, &out_trans);
Z
Zhang Ting 已提交
303 304
  TransCompute<DeviceContext, InT>(out_trans.dims().size(), dev_ctx, out_trans,
                                   out, permute);
Z
Zhang Ting 已提交
305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321

  if (return_inverse) {
    auto* inverse = context.Output<framework::Tensor>("Index");
    framework::TensorFromVector(inverse_vec, context.device_context(), inverse);
  }

  if (return_counts) {
    auto* count = context.Output<framework::Tensor>("Counts");
    framework::TensorFromVector(counts_vec, context.device_context(), count);
  }

  if (return_index) {
    auto* indices = context.Output<framework::Tensor>("Indices");
    framework::TensorFromVector(indices_vec, context.device_context(), indices);
  }
}

Z
Zhang Ting 已提交
322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377
template <typename DeviceContext, typename InT>
struct UniqueFlattendTensorFunctor {
  const framework::ExecutionContext& ctx_;
  const framework::Tensor& in_;
  framework::Tensor* out_;
  const bool return_index_;
  const bool return_inverse_;
  const bool return_counts_;

  UniqueFlattendTensorFunctor(const framework::ExecutionContext& context,
                              const framework::Tensor& in,
                              framework::Tensor* out, bool return_index,
                              bool return_inverse, bool return_counts)
      : ctx_(context),
        in_(in),
        out_(out),
        return_index_(return_index),
        return_inverse_(return_inverse),
        return_counts_(return_counts) {}

  template <typename IndexT>
  void apply() const {
    UniqueFlattendTensor<InT, IndexT>(ctx_, in_, out_, return_index_,
                                      return_inverse_, return_counts_);
  }
};

template <typename DeviceContext, typename InT>
struct UniqueDimFunctor {
  const framework::ExecutionContext& ctx_;
  const framework::Tensor& in_;
  framework::Tensor* out_;
  const int axis_;
  const bool return_index_;
  const bool return_inverse_;
  const bool return_counts_;

  UniqueDimFunctor(const framework::ExecutionContext& context,
                   const framework::Tensor& in, framework::Tensor* out,
                   const int axis, bool return_index, bool return_inverse,
                   bool return_counts)
      : ctx_(context),
        in_(in),
        out_(out),
        axis_(axis),
        return_index_(return_index),
        return_inverse_(return_inverse),
        return_counts_(return_counts) {}

  template <typename IndexT>
  void apply() const {
    UniqueDim<DeviceContext, InT, IndexT>(
        ctx_, in_, out_, return_index_, return_inverse_, return_counts_, axis_);
  }
};

Z
Zhang Ting 已提交
378
template <typename DeviceContext, typename T>
Z
zhoukunsheng 已提交
379 380 381 382 383
class UniqueKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    auto* x = context.Input<framework::Tensor>("X");
    auto* out = context.Output<framework::Tensor>("Out");
Z
Zhang Ting 已提交
384 385 386 387 388 389 390 391 392 393 394
    auto data_type = static_cast<framework::proto::VarType::Type>(
        context.Attr<int>("dtype"));
    if (data_type == framework::proto::VarType::INT32) {
      PADDLE_ENFORCE_LE(
          x->numel(), INT_MAX,
          platform::errors::InvalidArgument(
              "The number of elements in Input(X) should be less than or "
              "equal to INT_MAX, but received num is %d. Please set `dtype` to "
              "int64.",
              x->numel()));
    }
Z
Zhang Ting 已提交
395 396 397 398 399 400
    if (!context.Attr<bool>("is_sorted")) {
      auto* index = context.Output<framework::Tensor>("Index");

      framework::VisitDataType(data_type, UniqueOpFunctor<T>(out, index, x));
      return;
    }
Z
zhoukunsheng 已提交
401

Z
Zhang Ting 已提交
402 403 404 405
    std::vector<int> axis_vec = context.Attr<std::vector<int>>("axis");
    bool return_index = context.Attr<bool>("return_index");
    bool return_inverse = context.Attr<bool>("return_inverse");
    bool return_counts = context.Attr<bool>("return_counts");
406 407 408 409
    if (x->numel() == 0) {
      out->mutable_data<T>(context.GetPlace());
      return;
    }
Z
Zhang Ting 已提交
410
    if (axis_vec.empty()) {
411
      framework::VisitDataTypeTiny(
Z
Zhang Ting 已提交
412 413 414
          data_type,
          UniqueFlattendTensorFunctor<DeviceContext, T>(
              context, *x, out, return_index, return_inverse, return_counts));
Z
Zhang Ting 已提交
415 416
    } else {
      int axis = axis_vec[0];
417
      framework::VisitDataTypeTiny(
Z
Zhang Ting 已提交
418 419 420
          data_type, UniqueDimFunctor<DeviceContext, T>(
                         context, *x, out, axis, return_index, return_inverse,
                         return_counts));
Z
Zhang Ting 已提交
421
    }
Z
zhoukunsheng 已提交
422 423 424 425 426
  }
};

}  // namespace operators
}  // namespace paddle