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/partition_ops.h (9958B)

#ifndef CAFFE2_OPERATORS_PARTITION_OPS_H_
#define CAFFE2_OPERATORS_PARTITION_OPS_H_

#include "caffe2/core/context.h"
#include "caffe2/core/operator.h"

namespace caffe2 {

template <typename Index>
static inline int moduloPartition(Index key, int numPartitions) {
  int shard = key % numPartitions;
  // equivalent to `if (shard < 0) shard += partitions;`
  shard += numPartitions & (shard >> (sizeof(int) * 8 - 1));
  return shard;
}

class GatherByKeyOp : public Operator<CPUContext> {
 public:
  USE_DISPATCH_HELPER;
  USE_OPERATOR_FUNCTIONS(CPUContext);
  template <class... Args>
  explicit GatherByKeyOp(Args&&... args)
      : Operator<CPUContext>(std::forward<Args>(args)...) {}

private:
  bool RunOnDevice() override {
    return DispatchHelper<TensorTypes<int32_t, int64_t>>::call(this, Input(0));
  }

private:
  template <typename Index>
  bool DoRunWithType() {
    const auto numPartitions = InputSize() - 1;
    CAFFE_ENFORCE_GE(numPartitions, 1);
    const auto& keysTensor = Input(0);
    const auto* keysData = keysTensor.template data<Index>();
    const auto& keysShape = Input(0).sizes();
    CAFFE_ENFORCE_EQ(
        keysShape.size(), 1, "Only 1D keys tensor supported currently.");

// 1. Shape and type consistency checks
    const auto& in0Shape = Input(1).sizes();
    CAFFE_ENFORCE_GE(in0Shape.size(), 1);

vector<int64_t> outShape(keysShape.vec());
    outShape.insert(outShape.end(), in0Shape.begin() + 1, in0Shape.end());

CAFFE_ENFORCE_GE(outShape.size(), 1);
    auto totalSize = in0Shape[0];
    auto meta = Input(1).dtype();
    for (int i = 2; i < InputSize(); ++i) {
      const auto& input = Input(i);
      CAFFE_ENFORCE(meta == input.dtype());
      CAFFE_ENFORCE_GE(input.dim(), 1);
      CAFFE_ENFORCE(std::equal(
          outShape.begin() + keysShape.size(),
          outShape.end(),
          input.sizes().begin() + 1));
      totalSize += input.size(0);
    }
    CAFFE_ENFORCE_EQ(keysTensor.numel(), totalSize);

auto* outTensor = Output(0);
    outTensor->Resize(outShape);
    auto* outData = static_cast<char*>(outTensor->raw_mutable_data(meta));
    const auto blockSize = outTensor->size_from_dim(1);

inputDatas_.resize(numPartitions);
    for (int i = 0; i < numPartitions; ++i) {
      inputDatas_[i] = static_cast<const char*>(Input(i + 1).raw_data());
    }
    inStartOffsets_.assign(numPartitions, 0);
    Index outStartOffset = 0;
    int currentShard = -1;

// 2. copy from inputs into output based on shard for each input key
    const auto numEntries = keysTensor.numel();
    for (int64_t i = 0; i <= numEntries; ++i) {
      auto newShard =
          i < numEntries ? moduloPartition(keysData[i], numPartitions) : -1;
      if (newShard != currentShard) {
        if (currentShard != -1) {
          auto inStartOffset = inStartOffsets_[currentShard];
          auto numItems = i - outStartOffset;
          context_.CopyItemsSameDevice(
              meta,
              numItems * blockSize,
              inputDatas_[currentShard] +
                  inStartOffset * blockSize * meta.itemsize(),
              outData + outStartOffset * blockSize * meta.itemsize());
          inStartOffsets_[currentShard] += numItems;
        }
        currentShard = newShard;
        outStartOffset = i;
      }
    }

return true;
  }

std::vector<const char*> inputDatas_;
  std::vector<int64_t> inStartOffsets_;
};

class PartitionOpBase : public Operator<CPUContext> {
 public:
  USE_OPERATOR_FUNCTIONS(CPUContext);

template <class... Args>
  explicit PartitionOpBase(Args&&... args)
      : Operator<CPUContext>(std::forward<Args>(args)...),
        OP_SINGLE_ARG(int, "pack_first_input", pack_first_input_, 0) {}

protected:
  template <typename Index>
  void ApplyPartition(bool skipFirstArgument) {
    CAFFE_ENFORCE_EQ(
        OutputSize() % InputSize(),
        0,
        "Output number must be a multiple of input number");
    int partitions = OutputSize() / InputSize();
    int inputSize = InputSize();
    int mainInputIndex = skipFirstArgument;
    CAFFE_ENFORCE_GT(partitions, 0, "Invalid number of partitions");

auto& main_input = Input(mainInputIndex);
    int64_t size = main_input.numel();
    const Index* data = main_input.template data<Index>();
    counts_.assign(partitions, 0);
    for (int64_t p = 0; p < size; p++) {
      int shard = moduloPartition(data[p], partitions);
      ++counts_[shard];
    }

raw_datas_.resize(inputSize);
    block_sizes_.resize(inputSize);
    metas_.resize(inputSize);
    out_datas_.resize(OutputSize());
    for (int i = mainInputIndex; i < inputSize; ++i) {
      auto& input = Input(i);
      if (i > mainInputIndex) {
        CAFFE_ENFORCE_GE(
            input.dim(),
            main_input.dim(),
            "Prefix of extra input's shape must match main input's shape, ",
            "input: ",
            i);
        for (int j = 0; j < main_input.dim(); ++j) {
          CAFFE_ENFORCE_GE(
              input.size(j),
              main_input.size(j),
              "Prefix of extra input's shape must match main input's shape, ",
              "input: ",
              i,
              ", dim ",
              j);
        }
      }
      raw_datas_[i] = input.raw_data();
      block_sizes_[i] = input.size_from_dim(main_input.dim());
      metas_[i] = input.dtype();
      // shape = partition_size + suffix of input dims
      vector<int64_t> shape(
          input.sizes().begin() + main_input.dim() - 1, input.sizes().end());
      for (int j = 0; j < partitions; ++j) {
        int out_idx = i + j * inputSize;
        auto output = Output(out_idx);
        shape[0] = counts_[j];
        output->Resize(shape);
        out_datas_[out_idx] = output->raw_mutable_data(input.dtype());
      }
    }

counts_.assign(partitions, 0);
    for (int64_t p = 0; p < size; p++) {
      int shard = moduloPartition(data[p], partitions);
      int64_t idx = counts_[shard]++;

// special case first input
      static_cast<Index*>(out_datas_[shard * inputSize + mainInputIndex])[idx] =
          pack_first_input_ ? ((data[p] - shard) / partitions) : data[p];

int baseIndex = shard * inputSize;
      for (int i = mainInputIndex + 1; i < inputSize; ++i) {
        auto bs = block_sizes_[i];
        auto meta = metas_[i];
        // special case for small bs?
        context_.CopyItemsSameDevice(
            meta,
            bs,
            static_cast<const char*>(raw_datas_[i]) + p * bs * meta.itemsize(),
            static_cast<char*>(out_datas_[baseIndex + i]) +
                idx * bs * meta.itemsize());
      }
    }
  }

bool pack_first_input_;

// use member fields to reuse memory
  vector<int64_t> counts_;
  vector<int64_t> block_sizes_;
  vector<TypeMeta> metas_;
  vector<const void*> raw_datas_;
  vector<void*> out_datas_;
};

class PartitionOp : public PartitionOpBase {
 public:
  USE_DISPATCH_HELPER;

template <class... Args>
  explicit PartitionOp(Args&&... args)
      : PartitionOpBase(std::forward<Args>(args)...) {}

bool RunOnDevice() override {
    return DispatchHelper<TensorTypes<int32_t, int64_t>>::call(this, Input(0));
  }

private:
  template <typename Index>
  bool DoRunWithType() {
    ApplyPartition<Index>(false /* skipFirstArgument */);
    return true;
  }

C10_DISABLE_COPY_AND_ASSIGN(PartitionOp);
};

class LengthsPartitionOp : public PartitionOpBase {
 public:
  USE_DISPATCH_HELPER;

template <class... Args>
  explicit LengthsPartitionOp(Args&&... args)
      : PartitionOpBase(std::forward<Args>(args)...) {}

bool RunOnDevice() override {
    return DispatchHelper<TensorTypes<int32_t, int64_t>>::call(this, Input(1));
  }

private:
  template <typename Index>
  bool DoRunWithType() {
    CAFFE_ENFORCE(
        OutputSize() % InputSize() == 0,
        "Output number must be a multiple of input number");
    int partitions = OutputSize() / InputSize();
    CAFFE_ENFORCE_GT(partitions, 0, "Invalid number of partitions");
    CAFFE_ENFORCE_EQ(
        Input(1).dim(),
        1,
        "Only 1-D tensors supported as a partitioning tensor for sharding");

if (partitions == 1) {
      // Specialization when partitions == 1 which just becomes a copy.
      for (int i = 0; i < InputSize(); ++i) {
        auto& input = Input(i);
        auto& output = *Output(i);
        output.ResizeLike(input);
        context_.CopyItemsSameDevice(
            input.dtype(),
            input.numel(),
            input.raw_data(),
            output.raw_mutable_data(input.dtype()));
      }
      return true;
    }

// Apply sharding to all parameters except lengths
    ApplyPartition<Index>(true /* skipFirstArgument */);

// Compute lengths after sharding
    auto& main_input = Input(1);
    int64_t size = main_input.numel();
    const Index* data = main_input.template data<Index>();

auto& length_input = Input(0);
    int64_t elements = length_input.numel();
    const int32_t* lengths_data = length_input.template data<int32_t>();
    out_length_.resize(partitions);
    for (int i = 0; i < partitions; ++i) {
      auto& output = *Output(i * InputSize());
      output.Resize(elements);
      out_length_[i] = output.template mutable_data<int32_t>();
    }

int total_length = 0;
    for (int i = 0; i < elements; ++i) {
      total_length += lengths_data[i];
    }
    CAFFE_ENFORCE(
        total_length == size,
        "Total length is not matching to the number of elements");

int index = 0;
    for (int i = 0; i < elements; ++i) {
      for (int j = 0; j < partitions; ++j) {
        out_length_[j][i] = 0;
      }
      for (int j = 0; j < lengths_data[i]; ++j, ++index) {
        int shard = moduloPartition(data[index], partitions);
        ++out_length_[shard][i];
      }
    }
    return true;
  }

C10_DISABLE_COPY_AND_ASSIGN(LengthsPartitionOp);

vector<int32_t*> out_length_;
};

} // namespace caffe2

#endif // CAFFE2_OPERATORS_PARTITION_OPS_H_

cmd: