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/gather_op.h (7505B)

#ifndef GATHER_OP_H_
#define GATHER_OP_H_

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

namespace caffe2 {

// This maintains index-mapping functions shared by Gather and BatchGather ops.
namespace gather_helper {

// New shape is concatenation:
//  [data dims before axis] + [indices dims] + [data dims after axis]
template <typename IndexType, typename DataDimsVec, typename IndexDimsVec>
static vector<IndexType> calc_output_shape_vector(
    const DataDimsVec& data_dims,
    const IndexDimsVec& indices_dims,
    int axis,
    bool match_outer) {
  vector<IndexType> shape;
  // If the dimension we are indexing is empty, just use data_dims as shape.
  // This replicates behavior in (https://github.com/pytorch/pytorch/pull/13781)
  // needed to allow workflows with empty batch to succeed.
  if (data_dims[axis] == 0) {
    shape.insert(shape.end(), data_dims.begin(), data_dims.end());
  } else {
    shape.insert(shape.end(), data_dims.begin(), data_dims.begin() + axis);
    if (match_outer) {
      shape.insert(
          shape.end(), indices_dims.begin() + axis, indices_dims.end());
    } else {
      shape.insert(shape.end(), indices_dims.begin(), indices_dims.end());
    }
    shape.insert(shape.end(), data_dims.begin() + axis + 1, data_dims.end());
  }
  return shape;
}

// Check that indices fall within dimension array size with CAFFE_ENFORCE.
template <typename IndexType>
static void check_indexarray_range(
    const IndexType* indices,
    int64_t n,
    IndexType indexing_axis_dim,
    bool wrap_indices) {
  //
  for (auto i = 0; i < n; ++i) {
    auto idx = indices[i];
    if (wrap_indices && idx < 0) {
      idx = idx + indexing_axis_dim;
    }
    CAFFE_ENFORCE(
        0 <= idx && idx < indexing_axis_dim,
        "INDICES element is out of DATA bounds, id=",
        idx,
        " axis_dim=",
        indexing_axis_dim);
  }
}

// Actual gather implementation - resizes output and copies indexed data.
template <typename Index, typename Context>
static bool gather_impl(
    Operator<Context>* op,
    int dataIdx,
    int indicesIdx,
    int outputIdx,
    int axis,
    bool wrap_indices,
    bool match_outer) {
  // If we endup using it on GPU doing O(N) memcpy is probably not best :)
  // TODO: implement prefetching if it starts mattering (TF does it)

const Tensor& data = op->Input(dataIdx);
  const Tensor& indices = op->Input(indicesIdx);
  const TypeMeta dataType = data.dtype();
  size_t item_bytesize = dataType.itemsize();

// ONNX allows negative axis to index from the back, valid range: [-r, r].
  if (axis < 0) {
    axis = data.dim() + axis;
  }
  CAFFE_ENFORCE_GE(data.dim(), axis + 1, "DATA should be at least [axis+1]-D");
  CAFFE_ENFORCE_GE(axis, 0, "Axis should be non-negative");
  CAFFE_ENFORCE_LT(axis, data.dim(), "Axis out of range");

// New shape:
  //  [data dims before axis] + [indices dims] + [data dims after axis]
  vector<int64_t> shape = calc_output_shape_vector<int64_t>(
      data.sizes(), indices.sizes(), axis, match_outer);
  Tensor* output = op->Output(outputIdx, shape, at::dtype(dataType));
  auto out = static_cast<char*>(output->raw_mutable_data(dataType));

// Succeed if size of output is zero, which can happen for empty batch which
  // would have data dimension size of 0.
  // This *must* be done AFTER output->raw_mutable_data() above as that has
  // important allocation side effect that we must see.
  if (output->numel() == 0) {
    return true;
  }

const Index* idxs = indices.template data<Index>();
  auto src_base = static_cast<const char*>(data.raw_data());

auto outer_dims_product = data.size_to_dim(axis);
  auto block_size = data.size_from_dim(axis + 1);
  auto block_bytesize = block_size * item_bytesize;

auto src_indexing_axis_dim = data.size(axis);
  auto src_batch_bytesize = data.size_from_dim(axis) * item_bytesize;
  // Treat indices as a single block even if they have multiple dimensions.
  // The "gathered batch" is a cumulative result combining indexed blocks.
  auto idx_inner_dims_product = indices.size_from_dim(axis);
  auto N = indices.numel();
  if (match_outer) {
    CAFFE_ENFORCE_GE(axis, 1, "Axis should be at least 1");
    for (auto i = 0; i < axis; i++) {
      CAFFE_ENFORCE_EQ(
          data.size(i),
          indices.size(i),
          "INDICES must have the same outer dims as DATA (before dim AXIS)");
    }
    N = idx_inner_dims_product;
  }

auto gathered_batch_bytesize = N * block_size * item_bytesize;

check_indexarray_range<Index>(idxs, N, src_indexing_axis_dim, wrap_indices);

// Special-case single-float copy for efficiency
  if (data.template IsType<float>() && block_size == 1) {
    for (auto batch = 0; batch < outer_dims_product; ++batch) {
      const float* src_floats =
          (const float*)(src_base + batch * src_batch_bytesize);
      float* dst_floats = (float*)(out + batch * gathered_batch_bytesize);

for (auto i = 0; i < N; ++i) {
        auto idx = idxs[i];
        if (match_outer) {
          idx = idxs[batch * idx_inner_dims_product + i];
        }
        if (wrap_indices && idx < 0) {
          idx = idx + src_indexing_axis_dim;
        }
        dst_floats[i] = src_floats[idx];
      }
    }
  } else {
    // outer_dims_product specifies how many times we repeat inner dimensions,
    // so we just iterate over it to cover all outer dimensions.
    for (auto batch = 0; batch < outer_dims_product; ++batch) {
      for (auto i = 0; i < N; ++i) {
        auto idx = idxs[i];
        if (match_outer) {
          idx = idxs[batch * idx_inner_dims_product + i];
        }
        if (wrap_indices && idx < 0) {
          idx = idx + src_indexing_axis_dim;
        }

auto src = src_base + batch * src_batch_bytesize + idx * block_bytesize;
        auto dst = out + batch * gathered_batch_bytesize + i * block_bytesize;
        op->getContext()->CopyItemsSameDevice(dataType, block_size, src, dst);
      }
    }
  }
  return true;
}

} // namespace gather_helper

template <class Context>
class GatherOp : public Operator<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;

template <class... Args>
  explicit GatherOp(Args&&... args)
      : Operator<Context>(std::forward<Args>(args)...),
        OP_SINGLE_ARG(int, "axis", axis_, 0),
        OP_SINGLE_ARG(bool, "match_outer", match_outer_, false) {
    // TBD: We may want to fix the old index wrap behaviour once we have
    // operator versioning, to only apply it when needed as otherwise its likely
    // an error.
    // Right now, we apply index wrapping by default only to axis == 0,
    // since we have ONNX conversion code that uses it. For other ops it
    // needs to be specified explicitly with argument or you don't get it.
    if (OperatorBase::HasArgument("wrap_indices")) {
      wrap_indices_ = Operator<Context>::template GetSingleArgument<bool>(
          "wrap_indices", (false));
    } else {
      wrap_indices_ = (axis_ == 0) ? true : false;
    }
  }

virtual ~GatherOp() noexcept {}

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

template <typename Index>
  bool DoRunWithType() {
    return gather_helper::gather_impl<Index, Context>(
        this, DATA, INDICES, 0, axis_, wrap_indices_, match_outer_);
  }

INPUT_TAGS(DATA, INDICES);

protected:
  int axis_;
  bool wrap_indices_;
  bool match_outer_;
};

} // namespace caffe2
#endif // GATHER_OP_H_

cmd: