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/fully_connected_op.h (9351B)

#ifndef CAFFE2_OPERATORS_FULLY_CONNECTED_OP_H_
#define CAFFE2_OPERATORS_FULLY_CONNECTED_OP_H_

#include <c10/util/Optional.h>
#include "caffe2/core/context.h"
#include "caffe2/core/operator.h"
#include "caffe2/utils/conversions.h"
#include "caffe2/utils/math.h"

namespace caffe2 {

// This is Caffe's InnerProductOp, with a name that fits its purpose better.
template <
    class Context,
    class Engine = DefaultEngine,
    bool TransposeWeight = true>
class FullyConnectedOp final : public Operator<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  template <class... Args>
  explicit FullyConnectedOp(Args&&... args)
      : Operator<Context>(std::forward<Args>(args)...),
        axis_(this->template GetSingleArgument<int32_t>("axis", 1)),
        axis_w_(this->template GetSingleArgument<int32_t>("axis_w", 1)),
        float16_compute_(
            this->template GetSingleArgument<bool>("float16_compute", false)) {}
  ~FullyConnectedOp() {}

template <
      typename T_X,
      typename T_W,
      typename T_B,
      typename T_Y,
      typename MATH>
  bool DoRunWithType() {
    const auto& X = Input(0);
    const auto& W = Input(1);
    const auto& b = Input(2);

CAFFE_ENFORCE(b.dim() == 1, b.dim());
    // batch size
    const auto canonical_axis = X.canonical_axis_index(axis_);
    const auto M = X.size_to_dim(canonical_axis);
    const auto K = X.size_from_dim(canonical_axis);
    const auto canonical_axis_w = W.canonical_axis_index(axis_w_);
    const int N = TransposeWeight ? W.size_to_dim(canonical_axis_w)
                                  : W.size_from_dim(canonical_axis_w);

auto dimErrorString = [&]() {
      return c10::str(
          "Dimension mismatch: ",
          "X: ",
          X.sizes(),
          ", W: ",
          W.sizes(),
          ", b: ",
          b.sizes(),
          ", axis: ",
          axis_,
          ", M: ",
          M,
          ", N: ",
          N,
          ", K: ",
          K);
    };

// Error checking
    CAFFE_ENFORCE(M == X.numel() / K, dimErrorString());
    CAFFE_ENFORCE(K == W.numel() / N, dimErrorString());
    CAFFE_ENFORCE(N == b.dim32(0), dimErrorString());
    CAFFE_ENFORCE(N == b.numel(), dimErrorString());

Y_shape_cache_ = X.sizes().vec();
    // This is an invariant of canonical_axis, so we can DCHECK.
    DCHECK_LE(canonical_axis + 1, Y_shape_cache_.size());
    Y_shape_cache_.resize(canonical_axis + 1);
    Y_shape_cache_[canonical_axis] = N;
    auto* Y = Output(0, Y_shape_cache_, at::dtype<T_Y>());
    CAFFE_ENFORCE(M * N == Y->numel(), dimErrorString());

if (X.numel() == 0) {
      // skip the rest of the computation if X is empty
      Y->template mutable_data<T_Y>();
      return true;
    }

// default to FLOAT as math.h does.
    TensorProto::DataType math_type = TensorProto_DataType_FLOAT;
    if (fp16_type<MATH>()) {
      math_type = TensorProto_DataType_FLOAT16;
    }

// W * x
    math::Gemm<T_X, Context, Engine>(
        CblasNoTrans,
        TransposeWeight ? CblasTrans : CblasNoTrans,
        M,
        N,
        K,
        1,
        X.template data<T_X>(),
        W.template data<T_W>(),
        0,
        Y->template mutable_data<T_Y>(),
        &context_,
        math_type);

// Add bias term
    if (!bias_multiplier_.has_value()) {
      bias_multiplier_ =
          caffe2::empty({M}, at::dtype<T_B>().device(Context::GetDeviceType()));
      math::Set<T_B, Context>(
          M,
          convert::To<float, T_B>(1),
          bias_multiplier_->template mutable_data<T_B>(),
          &context_);
    } else if (bias_multiplier_->numel() != M) {
      bias_multiplier_->Resize(M);
      math::Set<T_B, Context>(
          M,
          convert::To<float, T_B>(1),
          bias_multiplier_->template mutable_data<T_B>(),
          &context_);
    }

math::Gemm<T_B, Context, Engine>(
        CblasNoTrans,
        CblasNoTrans,
        M,
        N,
        1,
        1,
        bias_multiplier_->template data<T_B>(),
        b.template data<T_B>(),
        1,
        Y->template mutable_data<T_Y>(),
        &context_,
        math_type);

return true;
  }

bool RunOnDevice() override {
    return DoRunWithType<
        float, // X
        float, // W
        float, // B
        float, // Y
        float>(); // Math
  }

protected:
  size_t axis_{1};
  size_t axis_w_{1};
  // A local vector to cache the output shape so we don't need to recreate
  // a vector object every time we run Run().
  vector<int64_t> Y_shape_cache_;
  c10::optional<Tensor> bias_multiplier_;

bool float16_compute_;
};

template <
    class Context,
    class Engine = DefaultEngine,
    bool TransposeWeight = true>
class FullyConnectedGradientOp : public Operator<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;
  template <class... Args>
  explicit FullyConnectedGradientOp(Args&&... args)
      : Operator<Context>(std::forward<Args>(args)...),
        axis_(this->template GetSingleArgument<int32_t>("axis", 1)),
        axis_w_(this->template GetSingleArgument<int32_t>("axis_w", 1)),
        float16_compute_(
            this->template GetSingleArgument<bool>("float16_compute", false)) {}
  ~FullyConnectedGradientOp() {}

template <
      typename T_X,
      typename T_W,
      typename T_DY,
      typename T_B,
      typename T_DX,
      typename T_DW,
      typename T_DB,
      typename MATH>
  bool DoRunWithType() {
    const auto& X = Input(0);
    const auto& W = Input(1);
    const auto& dY = Input(2);
    // batch size
    const auto canonical_axis = X.canonical_axis_index(axis_);
    const int M = X.size_to_dim(canonical_axis);
    const int K = X.size_from_dim(canonical_axis);
    const auto canonical_axis_w = W.canonical_axis_index(axis_w_);
    const int N = TransposeWeight ? W.size_to_dim(canonical_axis_w)
                                  : W.size_from_dim(canonical_axis_w);

auto dimErrorString = [&]() {
      return c10::str(
          "Dimension mismatch: ",
          "X: ",
          X.sizes(),
          ", W: ",
          W.sizes(),
          ", dY: ",
          dY.sizes(),
          ", axis: ",
          axis_,
          ", M: ",
          M,
          ", N: ",
          N,
          ", K: ",
          K);
    };

CAFFE_ENFORCE(M * K == X.numel(), dimErrorString());
    CAFFE_ENFORCE(K * N == W.numel(), dimErrorString());

auto* dW = Output(0, W.sizes(), at::dtype<T_DW>());
    auto* db = Output(1, {N}, at::dtype<T_DB>());

if (X.numel() == 0) {
      // generate a zero blob for db and dW when X is empty
      math::Set<T_DB, Context>(
          db->numel(),
          convert::To<float, T_DB>(0),
          db->template mutable_data<T_DB>(),
          &context_);
      math::Set<T_DW, Context>(
          dW->numel(),
          convert::To<float, T_DW>(0),
          dW->template mutable_data<T_DW>(),
          &context_);

if (OutputSize() == 3) {
        Output(2, X.sizes(), at::dtype<T_DX>());
      }

return true;
    }

// default to FLOAT as math.h does.
    TensorProto::DataType math_type = TensorProto_DataType_FLOAT;
    if (fp16_type<MATH>()) {
      math_type = TensorProto_DataType_FLOAT16;
    }

// Compute dW
    math::Gemm<T_DY, Context, Engine>(
        CblasTrans,
        CblasNoTrans,
        TransposeWeight ? N : K,
        TransposeWeight ? K : N,
        M,
        1,
        TransposeWeight ? dY.template data<T_DY>() : X.template data<T_X>(),
        TransposeWeight ? X.template data<T_X>() : dY.template data<T_DY>(),
        0,
        dW->template mutable_data<T_DW>(),
        &context_,
        math_type);
    if (!bias_multiplier_.has_value()) {
      bias_multiplier_ =
          caffe2::empty({M}, at::dtype<T_B>().device(Context::GetDeviceType()));
      math::Set<T_B, Context>(
          M,
          convert::To<float, T_B>(1),
          bias_multiplier_->template mutable_data<T_B>(),
          &context_);
    } else if (bias_multiplier_->numel() != M) {
      bias_multiplier_->Resize(M);
      math::Set<T_B, Context>(
          M,
          convert::To<float, T_B>(1),
          bias_multiplier_->template mutable_data<T_B>(),
          &context_);
    }
    // Compute dB
    math::Gemv<T_DY, Context>(
        CblasTrans,
        M,
        N,
        1,
        dY.template data<T_DY>(),
        bias_multiplier_->template data<T_B>(),
        0,
        db->template mutable_data<T_DB>(),
        &context_);

// Compute dX
    if (OutputSize() == 3) {
      auto* dX = Output(2, X.sizes(), at::dtype<T_DX>());
      math::Gemm<T_DX, Context, Engine>(
          CblasNoTrans,
          TransposeWeight ? CblasNoTrans : CblasTrans,
          M,
          K,
          N,
          1,
          dY.template data<T_DY>(),
          W.template data<T_W>(),
          0,
          dX->template mutable_data<T_DX>(),
          &context_,
          math_type);
    }
    return true;
  }

bool RunOnDevice() override {
    return DoRunWithType<
        float, //  X
        float, //  W
        float, // dY
        float, //  B
        float, // dX
        float, // dW
        float, // dB
        float>(); // Math
  }

protected:
  size_t axis_{1};
  size_t axis_w_{1};
  c10::optional<Tensor> bias_multiplier_;
  bool float16_compute_;
};

} // namespace caffe2

#endif // CAFFE2_OPERATORS_FULLY_CONNECTED_OP_H_

cmd: