FM / operators

/usr/local/lib64/python3.6/site-packages/torch/include/caffe2/operators/

/usr/local/lib64/python3.6/site-packages/torch/include/caffe2/operators

Name	Size	Mode	Actions
abs_op.h	705	0644	edit dl rm
accumulate_op.h	1073	0644	edit dl rm
accuracy_op.h	652	0644	edit dl rm
acos_op.h	711	0644	edit dl rm
activation_ops_cudnn.h	4122	0644	edit dl rm
affine_channel_op.h	3450	0644	edit dl rm
alias_with_name.h	1234	0644	edit dl rm
apmeter_op.h	1027	0644	edit dl rm
arg_ops.h	2319	0644	edit dl rm
asin_op.h	711	0644	edit dl rm
assert_op.h	1335	0644	edit dl rm
async_net_barrier_op.h	904	0644	edit dl rm
atan_op.h	711	0644	edit dl rm
batch_box_cox_op.h	2287	0644	edit dl rm
batch_bucketize_op.h	720	0644	edit dl rm
batch_gather_ops.h	5264	0644	edit dl rm
batch_matmul_op.h	9602	0644	edit dl rm
batch_moments_op.h	3364	0644	edit dl rm
batch_permutation_op.h	954	0644	edit dl rm
batch_sparse_to_dense_op.h	6147	0644	edit dl rm
bbox_transform_op.h	2668	0644	edit dl rm
bisect_percentile_op.h	4921	0644	edit dl rm
boolean_mask_ops.h	2665	0644	edit dl rm
boolean_unmask_ops.h	378	0644	edit dl rm
box_with_nms_limit_op.h	4960	0644	edit dl rm
bucketize_op.h	1361	0644	edit dl rm
byte_weight_dequant_op.h	1722	0644	edit dl rm
cast_op.h	1393	0644	edit dl rm
cbrt_op.h	723	0644	edit dl rm
cc_bmm_bg_op.h	3894	0644	edit dl rm
ceil_op.h	782	0644	edit dl rm
channel_backprop_stats_op.h	737	0644	edit dl rm
channel_shuffle_op.h	1902	0644	edit dl rm
channel_stats_op.h	1807	0644	edit dl rm
clip_op.h	1639	0644	edit dl rm
collect_and_distribute_fpn_rpn_proposals_op.h	6875	0644	edit dl rm
concat_split_op.h	11850	0644	edit dl rm
conditional_op.h	487	0644	edit dl rm
conv_op.h	3125	0644	edit dl rm
conv_op_cache_cudnn.h	1935	0644	edit dl rm
conv_op_impl.h	28729	0644	edit dl rm
conv_op_shared.h	672	0644	edit dl rm
conv_pool_op_base.h	32109	0644	edit dl rm
conv_transpose_op.h	1727	0644	edit dl rm
conv_transpose_op_impl.h	18264	0644	edit dl rm
conv_transpose_op_mobile.h	1470	0644	edit dl rm
conv_transpose_op_mobile_impl.h	19587	0644	edit dl rm
conv_transpose_unpool_op_base.h	10303	0644	edit dl rm
copy_op.h	1296	0644	edit dl rm
copy_rows_to_tensor_op.h	2599	0644	edit dl rm
cosh_op.h	711	0644	edit dl rm
cosine_embedding_criterion_op.h	1127	0644	edit dl rm
cos_op.h	705	0644	edit dl rm
counter_ops.h	4596	0644	edit dl rm
create_scope_op.h	5232	0644	edit dl rm
cross_entropy_op.h	4420	0644	edit dl rm
ctc_beam_search_decoder_op.h	1102	0644	edit dl rm
ctc_greedy_decoder_op.h	817	0644	edit dl rm
cube_op.h	723	0644	edit dl rm
dataset_ops.h	5501	0644	edit dl rm
data_couple.h	464	0644	edit dl rm
deform_conv_op.h	3543	0644	edit dl rm
deform_conv_op_impl.h	13171	0644	edit dl rm
dense_vector_to_id_list_op.h	1797	0644	edit dl rm
distance_op.h	8419	0644	edit dl rm
do_op.h	6981	0644	edit dl rm
dropout_op.h	1516	0644	edit dl rm
elementwise_add_op.h	2024	0644	edit dl rm
elementwise_div_op.h	1224	0644	edit dl rm
elementwise_linear_op.h	1170	0644	edit dl rm
elementwise_logical_ops.h	5083	0644	edit dl rm
elementwise_mul_op.h	1224	0644	edit dl rm
elementwise_ops.h	19115	0644	edit dl rm
elementwise_ops_utils.h	1008	0644	edit dl rm
elementwise_op_test.h	9237	0644	edit dl rm
elementwise_sub_op.h	2025	0644	edit dl rm
elu_op.h	875	0644	edit dl rm
enforce_finite_op.h	2303	0644	edit dl rm
ensure_clipped_op.h	1608	0644	edit dl rm
ensure_cpu_output_op.h	1465	0644	edit dl rm
erf_op.h	751	0644	edit dl rm
expand_op.h	3877	0644	edit dl rm
expand_squeeze_dims_op.h	3451	0644	edit dl rm
exp_op.h	425	0644	edit dl rm
fc_inference.h	775	0644	edit dl rm
feature_maps_ops.h	32437	0644	edit dl rm
feed_blob_op.h	802	0644	edit dl rm
filler_op.h	18431	0644	edit dl rm
find_duplicate_elements_op.h	1563	0644	edit dl rm
find_op.h	2055	0644	edit dl rm
flatten_op.h	1525	0644	edit dl rm
flexible_top_k.h	936	0644	edit dl rm
floor_op.h	788	0644	edit dl rm
free_op.h	777	0644	edit dl rm
fully_connected_op.h	9351	0644	edit dl rm
fused_rowwise_8bit_conversion_ops.h	6601	0644	edit dl rm
fused_rowwise_nbitfake_conversion_ops.h	4375	0644	edit dl rm
fused_rowwise_nbit_conversion_ops.h	8723	0644	edit dl rm
fused_rowwise_random_quantization_ops.h	2607	0644	edit dl rm
gather_fused_8bit_rowwise_op.h	2179	0644	edit dl rm
gather_op.h	7505	0644	edit dl rm
gather_ranges_to_dense_op.h	8188	0644	edit dl rm
gelu_op.h	1452	0644	edit dl rm
generate_proposals_op.h	6256	0644	edit dl rm
generate_proposals_op_util_boxes.h	14309	0644	edit dl rm
generate_proposals_op_util_nms.h	26214	0644	edit dl rm
generate_proposals_op_util_nms_gpu.h	2128	0644	edit dl rm
given_tensor_byte_string_to_uint8_fill_op.h	2150	0644	edit dl rm
given_tensor_fill_op.h	3002	0644	edit dl rm
glu_op.h	1458	0644	edit dl rm
group_norm_op.h	8967	0644	edit dl rm
gru_unit_op.h	6626	0644	edit dl rm
half_float_ops.h	2732	0644	edit dl rm
hard_sigmoid_op.h	994	0644	edit dl rm
heatmap_max_keypoint_op.h	939	0644	edit dl rm
histogram_op.h	2421	0644	edit dl rm
h_softmax_op.h	4954	0644	edit dl rm
if_op.h	1764	0644	edit dl rm
im2col_op.h	8943	0644	edit dl rm
index_hash_ops.h	2232	0644	edit dl rm
index_ops.h	3155	0644	edit dl rm
inference_lstm_op.h	9881	0644	edit dl rm
instance_norm_op.h	7441	0644	edit dl rm
integral_image_op.h	923	0644	edit dl rm
is_empty_op.h	558	0644	edit dl rm
jsd_op.h	721	0644	edit dl rm
key_split_ops.h	1400	0644	edit dl rm
layer_norm_op.h	8098	0644	edit dl rm
leaky_relu_op.h	1111	0644	edit dl rm
lengths_pad_op.h	2574	0644	edit dl rm
lengths_reducer_fused_8bit_rowwise_ops.h	5532	0644	edit dl rm
lengths_reducer_fused_nbit_rowwise_ops.h	23465	0644	edit dl rm
lengths_reducer_ops.h	23315	0644	edit dl rm
lengths_reducer_rowwise_8bit_ops.h	6180	0644	edit dl rm
lengths_tile_op.h	582	0644	edit dl rm
lengths_top_k_op.h	1358	0644	edit dl rm
length_split_op.h	2259	0644	edit dl rm
listwise_l2r_op.h	1677	0644	edit dl rm
load_save_op.h	14091	0644	edit dl rm
load_save_op_util.h	1642	0644	edit dl rm
locally_connected_op.h	3872	0644	edit dl rm
locally_connected_op_impl.h	26495	0644	edit dl rm
locally_connected_op_util.h	1332	0644	edit dl rm
local_response_normalization_op.h	2804	0644	edit dl rm
log1p_op.h	717	0644	edit dl rm
logit_op.h	1129	0644	edit dl rm
log_op.h	431	0644	edit dl rm
loss_op.h	1058	0644	edit dl rm
lpnorm_op.h	1279	0644	edit dl rm
lstm_unit_op.h	6733	0644	edit dl rm
lstm_utils.h	9424	0644	edit dl rm
map_ops.h	8011	0644	edit dl rm
margin_ranking_criterion_op.h	1113	0644	edit dl rm
matmul_op.h	2843	0644	edit dl rm
max_pool_with_index_gpu.h	1155	0644	edit dl rm
mean_op.h	3252	0644	edit dl rm
merge_id_lists_op.h	2570	0644	edit dl rm
minmax_ops.h	3829	0644	edit dl rm
mish_op.h	794	0644	edit dl rm
mod_op.h	984	0644	edit dl rm
moments_op.h	4051	0644	edit dl rm
multi_class_accuracy_op.h	539	0644	edit dl rm
negate_gradient_op.h	566	0644	edit dl rm
negative_op.h	451	0644	edit dl rm
ngram_ops.h	2644	0644	edit dl rm
normalize_l1_op.h	1075	0644	edit dl rm
normalize_op.h	3013	0644	edit dl rm
no_default_engine_op.h	1063	0644	edit dl rm
numpy_tile_op.h	3643	0644	edit dl rm
one_hot_ops.h	2562	0644	edit dl rm
onnx_while_op.h	10655	0644	edit dl rm
operator_fallback_gpu.h	4155	0644	edit dl rm
op_utils_cudnn.h	2112	0644	edit dl rm
order_switch_ops.h	2149	0644	edit dl rm
pack_rnn_sequence_op.h	3074	0644	edit dl rm
pack_segments.h	2729	0644	edit dl rm
pad_op.h	2902	0644	edit dl rm
partition_ops.h	9958	0644	edit dl rm
percentile_op.h	1009	0644	edit dl rm
perplexity_op.h	447	0644	edit dl rm
piecewise_linear_transform_op.h	8281	0644	edit dl rm
pool_op.h	8525	0644	edit dl rm
pool_op_util.h	1105	0644	edit dl rm
pow_op.h	4677	0644	edit dl rm
prefetch_op.h	4661	0644	edit dl rm
prelu_op.h	1067	0644	edit dl rm
prepend_dim_op.h	2760	0644	edit dl rm
quantile_op.h	4120	0644	edit dl rm
quant_decode_op.h	5370	0644	edit dl rm
rank_loss_op.h	820	0644	edit dl rm
reciprocal_op.h	721	0644	edit dl rm
reducer_functors.h	24556	0644	edit dl rm
reduce_front_back_max_ops.h	4399	0644	edit dl rm
reduce_front_back_sum_mean_ops.h	5337	0644	edit dl rm
reduce_ops.h	9962	0644	edit dl rm
reduction_ops.h	5944	0644	edit dl rm
relu_n_op.h	990	0644	edit dl rm
relu_op.h	624	0644	edit dl rm
remove_data_blocks_op.h	2651	0644	edit dl rm
replace_nan_op.h	1170	0644	edit dl rm
reshape_op.h	5723	0644	edit dl rm
resize_3d_op.h	2677	0644	edit dl rm
resize_op.h	2307	0644	edit dl rm
reverse_packed_segs_op.h	2772	0644	edit dl rm
rmac_regions_op.h	708	0644	edit dl rm
rms_norm_op.h	2968	0644	edit dl rm
roi_align_gradient_op.h	1486	0644	edit dl rm
roi_align_op.h	2857	0644	edit dl rm
roi_align_rotated_gradient_op.h	1369	0644	edit dl rm
roi_align_rotated_op.h	1636	0644	edit dl rm
roi_pool_op.h	2503	0644	edit dl rm
rowmul_op.h	1947	0644	edit dl rm
rsqrt_op.h	729	0644	edit dl rm
scale_blobs_op.h	1458	0644	edit dl rm
scale_op.h	1019	0644	edit dl rm
segment_reduction_op.h	71022	0644	edit dl rm
self_binning_histogram_op.h	6258	0644	edit dl rm
selu_op.h	1545	0644	edit dl rm
sequence_ops.h	8264	0644	edit dl rm
shape_op.h	1638	0644	edit dl rm
sigmoid_op.h	639	0644	edit dl rm
sinh_op.h	711	0644	edit dl rm
sinusoid_position_encoding_op.h	2834	0644	edit dl rm
sin_op.h	705	0644	edit dl rm
slice_op.h	10071	0644	edit dl rm
softmax_op.h	1174	0644	edit dl rm
softmax_utils.h	447	0644	edit dl rm
softmax_with_loss_op.h	2883	0644	edit dl rm
softplus_op.h	781	0644	edit dl rm
softsign_op.h	675	0644	edit dl rm
space_batch_op.h	6848	0644	edit dl rm
sparse_dropout_with_replacement_op.h	1122	0644	edit dl rm
sparse_itemwise_dropout_with_replacement_op.h	1163	0644	edit dl rm
sparse_lp_regularizer_op.h	1130	0644	edit dl rm
sparse_normalize_op.h	834	0644	edit dl rm
sparse_to_dense_mask_op.h	10051	0644	edit dl rm
sparse_to_dense_op.h	3977	0644	edit dl rm
spatial_batch_norm_op.h	15175	0644	edit dl rm
spatial_softmax_with_loss_op.h	2182	0644	edit dl rm
sqrt_op.h	448	0644	edit dl rm
sqr_op.h	431	0644	edit dl rm
square_root_divide_op.h	1857	0644	edit dl rm
stats_put_ops.h	2813	0644	edit dl rm
stop_gradient.h	548	0644	edit dl rm
string_ops.h	2067	0644	edit dl rm
stump_func_op.h	2112	0644	edit dl rm
summarize_op.h	1875	0644	edit dl rm
swish_op.h	772	0644	edit dl rm
tanh_op.h	723	0644	edit dl rm
tan_op.h	705	0644	edit dl rm
tensor_protos_db_input.h	3633	0644	edit dl rm
text_file_reader_utils.h	2900	0644	edit dl rm
thresholded_relu_op.h	1137	0644	edit dl rm
tile_op.h	8741	0644	edit dl rm
top_k.h	1061	0644	edit dl rm
transpose_op.h	2082	0644	edit dl rm
tt_linear_op.h	6501	0644	edit dl rm
unique_ops.h	1666	0644	edit dl rm
unsafe_coalesce.h	2481	0644	edit dl rm
upsample_op.h	2246	0644	edit dl rm
utility_ops.h	49994	0644	edit dl rm
variable_length_sequence_padding.h	1378	0644	edit dl rm
weighted_multi_sampling_op.h	602	0644	edit dl rm
weighted_sample_op.h	739	0644	edit dl rm
while_op.h	1961	0644	edit dl rm
zero_gradient_op.h	347	0644	edit dl rm

Edit: /usr/local/lib64/python3.6/site-packages/torch/include/caffe2/operators/sparse_to_dense_mask_op.h (10051B)

#ifndef CAFFE2_OPERATORS_SPARSE_TO_DENSE_MASK_OP_H_
#define CAFFE2_OPERATORS_SPARSE_TO_DENSE_MASK_OP_H_

#include <algorithm>
#include <unordered_map>
#include <vector>
#include "caffe2/core/context.h"
#include "caffe2/core/export_caffe2_op_to_c10.h"
#include "caffe2/core/operator.h"
#include "caffe2/core/tensor.h"
#include "caffe2/utils/math.h"

C10_DECLARE_EXPORT_CAFFE2_OP_TO_C10(SparseToDenseMask);

namespace caffe2 {

template <class Context>
class SparseToDenseMaskBase : public Operator<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  template <class... Args>
  explicit SparseToDenseMaskBase(Args&&... args)
      : Operator<Context>(std::forward<Args>(args)...) {
    std::vector<int64_t> mask =
        this->template GetRepeatedArgument<int64_t>("mask");
    featuresCount_ = mask.size();

CAFFE_ENFORCE(!mask.empty(), "mask can't be empty");
    auto biggest = *std::max_element(mask.begin(), mask.end());
    dense_.assign(std::min(kMaxDenseSize, biggest + 1), -1);
    // NOLINTNEXTLINE(clang-diagnostic-sign-compare)
    for (int i = 0; i < mask.size(); i++) {
      int64_t id = mask[i];
      CAFFE_ENFORCE_GE(id, 0, "Only positive IDs are allowed.");
      if (id >= kMaxDenseSize) {
        CAFFE_ENFORCE(sparse_.count(id) == 0, "Duplicated id: ", id);
        sparse_[id] = i;
      } else {
        CAFFE_ENFORCE(dense_[id] == -1, "Duplicated id: ", id);
        dense_[id] = i;
      }
    }
  }

protected:
  const int64_t kMaxDenseSize = 1024 * 128;

std::unordered_map<int64_t, int> sparse_;
  std::vector<int> dense_;
  size_t featuresCount_;

inline int getFeatureIdx(int64_t id) const {
    if (id >= kMaxDenseSize) {
      const auto& iter = sparse_.find(id);
      if (iter == sparse_.end()) {
        return -1;
      } else {
        return iter->second;
      }
    } else {
      // NOLINTNEXTLINE(clang-diagnostic-sign-compare)
      return (id >= dense_.size()) ? -1 : dense_[id];
    }
  }
};

template <class Context>
class SparseToDenseMaskOp : public SparseToDenseMaskBase<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  template <class... Args>
  explicit SparseToDenseMaskOp(Args&&... args)
      : SparseToDenseMaskBase<Context>(std::forward<Args>(args)...) {
    returnPresenceMask_ =
        this->template GetSingleArgument<bool>("return_presence_mask", false);
    maxSkippedRows_ = this->template GetSingleArgument<int32_t>(
        "max_skipped_indices", kMaxSkippedSparseIndices);
  }

bool RunOnDevice() override {
    return DispatchHelper<TensorTypes<int32_t, int64_t>>::call(
        this, Input(INDICES));
  }

template <typename TInd>
  bool DoRunWithType() {
    auto& sparse_indices = Input(INDICES);
    CAFFE_ENFORCE_EQ(sparse_indices.dim(), 1);
    auto& sparse_values = Input(VALUES);
    CAFFE_ENFORCE_GE(sparse_values.dim(), 1);
    CAFFE_ENFORCE_EQ(sparse_indices.numel(), sparse_values.size(0));
    auto& default_value = Input(DEFAULT);
    CAFFE_ENFORCE_EQ(default_value.dim() + 1, sparse_values.dim());
    CAFFE_ENFORCE_EQ(default_value.numel(), sparse_values.size_from_dim(1));
    CAFFE_ENFORCE(sparse_values.dtype() == default_value.dtype());

const TInd* sparse_indices_vec = sparse_indices.template data<TInd>();
    const char* sparse_values_vec =
        static_cast<const char*>(sparse_values.raw_data());
    const void* default_val = default_value.raw_data();

int64_t block_size = default_value.numel();
    size_t block_nbytes = default_value.nbytes();

const size_t cols = this->featuresCount_;
    int rows = -1;
    int32_t sparse_indices_length = sparse_indices.dim32(0);
    const int32_t* lengths_vec = nullptr;
    auto* output = Output(OUTPUTVALUE);
    Tensor* presence_mask = nullptr;
    if (returnPresenceMask_) {
      presence_mask = Output(PRESENCEMASK);
    }
    vector<int64_t> shape;
    if (InputSize() == 4) {
      auto& lengths = Input(LENGTHS);
      CAFFE_ENFORCE_EQ(lengths.dim(), 1);
      lengths_vec = lengths.template data<int32_t>();
      rows = lengths.dim32(0);
    }
    if (rows == -1) {
      // if the LENGTHS is not set, the output will be a vector
      rows = 1;
      lengths_vec = &sparse_indices_length;
    } else {
      shape.push_back(rows);
    }
    shape.push_back(cols);
    if (returnPresenceMask_) {
      presence_mask->Resize(shape);
    }
    shape.insert(
        shape.end(),
        default_value.sizes().begin(),
        default_value.sizes().end());
    output->Resize(shape);

// init
    // TODO: consider unrolling CopyItems to make elemental types copy faster
    char* output_data =
        static_cast<char*>(output->raw_mutable_data(sparse_values.dtype()));
    // NOLINTNEXTLINE(clang-diagnostic-sign-compare)
    for (int i = 0; i < cols * rows; i++) {
      context_.CopyItemsSameDevice(
          default_value.dtype(),
          block_size,
          default_val,
          output_data + i * block_nbytes);
    }
    bool* presence_mask_data = nullptr;
    if (returnPresenceMask_) {
      presence_mask_data = presence_mask->template mutable_data<bool>();
      math::Set<bool, Context>(
          rows * cols, false, presence_mask_data, &context_);
    }

int64_t offset = 0;
    for (int r = 0; r < rows; r++) {
      bool skippedSparseIndex = false;
      for (int c = 0; c < lengths_vec[r]; c++) {
        const auto sparse_index = sparse_indices_vec[offset + c];
        if (sparse_index < 0 ||
            sparse_index >= std::numeric_limits<TInd>::max()) {
          skippedSparseIndex = true;
          LOG(WARNING) << "Skipping invalid sparse index: " << sparse_index;
          continue;
        }
        int idx = this->getFeatureIdx(sparse_index);
        if (idx != -1) {
          context_.CopyItemsSameDevice(
              sparse_values.dtype(),
              block_size,
              sparse_values_vec + (offset + c) * block_nbytes,
              output_data + (r * cols + idx) * block_nbytes);
          if (returnPresenceMask_) {
            presence_mask_data[r * cols + idx] = true;
          }
        }
      }
      skippedRows_ += skippedSparseIndex;
      CAFFE_ENFORCE_LT(
          skippedRows_,
          maxSkippedRows_,
          "Too many rows with invalid sparse indices skipped");
      offset += lengths_vec[r];
    }

return true;
  }

private:
  static const uint32_t kMaxSkippedSparseIndices = 50;

bool returnPresenceMask_;
  uint32_t maxSkippedRows_ = 0;
  uint32_t skippedRows_ = 0;

INPUT_TAGS(INDICES, VALUES, DEFAULT, LENGTHS);
  OUTPUT_TAGS(OUTPUTVALUE, PRESENCEMASK);
};

template <class Context>
class SparseToDenseMaskGradientOp : public SparseToDenseMaskBase<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  template <class... Args>
  explicit SparseToDenseMaskGradientOp(Args&&... args)
      : SparseToDenseMaskBase<Context>(std::forward<Args>(args)...) {}

bool RunOnDevice() override {
    return DispatchHelper<TensorTypes<int32_t, int64_t>>::call(
        this, Input(INDICES));
  }

template <typename TInd>
  bool DoRunWithType() {
    auto& sparse_indices = Input(INDICES);
    CAFFE_ENFORCE_EQ(sparse_indices.dim(), 1);
    auto& gradient_output = Input(GOUTPUT);

int64_t block_size = gradient_output.size_from_dim(1);
    size_t block_nbytes = gradient_output.itemsize() * block_size;

const size_t cols = this->featuresCount_;
    int rows = -1;
    int iter_offset = 1;
    int32_t default_length = sparse_indices.dim32(0);
    const int32_t* lengths_vec = nullptr;
    auto* output = Output(GVALUES);
    vector<int64_t> shape;
    if (InputSize() > LENGTHS) {
      // if the LENGTHS is set, the gradient_output has dim:
      // lengths * mask.size() * feature_dim
      auto& lengths = Input(LENGTHS);
      lengths_vec = lengths.template data<int32_t>();
      rows = lengths.dim32(0);
      CAFFE_ENFORCE_EQ(lengths.dim(), 1);
      CAFFE_ENFORCE_GE(gradient_output.dim(), 2);
      CAFFE_ENFORCE_EQ(gradient_output.size(0), rows);
      CAFFE_ENFORCE_EQ(gradient_output.size(1), cols);
      block_nbytes /= gradient_output.size(1);
      block_size /= gradient_output.size(1);
      iter_offset += 1;
    }
    if (rows == -1) {
      // if the LENGTHS is not set, the gradient_output has dim:
      // mask.size() * feature_dim
      rows = 1;
      lengths_vec = &default_length;
      CAFFE_ENFORCE_GE(gradient_output.dim(), 1);
      CAFFE_ENFORCE_EQ(gradient_output.size(0), cols);
    }
    shape.push_back(default_length);
    // insert feature_dim
    shape.insert(
        shape.end(),
        gradient_output.sizes().begin() + iter_offset,
        gradient_output.sizes().end());
    output->Resize(shape);

const TInd* sparse_indices_vec = sparse_indices.template data<TInd>();
    const char* gradient_output_vec =
        static_cast<const char*>(gradient_output.raw_data());

char* output_data =
        static_cast<char*>(output->raw_mutable_data(gradient_output.dtype()));
    memset(output_data, 0, output->nbytes());
    math::Set<char, Context>(
        default_length * gradient_output.itemsize(), 0, output_data, &context_);

int32_t offset = 0;
    // SparseToDenseMask is not injective; gradient_used records
    // if the gradient is used for other input value from the same row
    vector<bool> gradient_used(cols, false);
    for (int r = 0; r < rows; r++) {
      std::fill(gradient_used.begin(), gradient_used.end(), false);
      for (int c = lengths_vec[r] - 1; c >= 0; c--) {
        int idx = this->getFeatureIdx(sparse_indices_vec[offset + c]);
        if (idx != -1 && !gradient_used[idx]) {
          gradient_used[idx] = true;
          context_.CopyItemsSameDevice(
              gradient_output.dtype(),
              block_size,
              gradient_output_vec + (r * cols + idx) * block_nbytes,
              output_data + (offset + c) * block_nbytes);
        }
      }
      offset += lengths_vec[r];
    }
    return true;
  }

private:
  INPUT_TAGS(INDICES, GOUTPUT, LENGTHS);
  OUTPUT_TAGS(GVALUES);
};

} // namespace caffe2

#endif // CAFFE2_OPERATORS_SPARSE_TO_DENSE_MASK_OP_H_

cmd: