tree: d08b7418db5a39b2a6626b78d799f686437556b4 [path history] [tgz]
  1. experimental/
  2. hip/
  3. quantized/
  4. rnn/
  5. abs_op.cc
  6. abs_op.cu
  7. abs_op.h
  8. accumulate_op.cc
  9. accumulate_op.cu
  10. accumulate_op.h
  11. accuracy_op.cc
  12. accuracy_op.cu
  13. accuracy_op.h
  14. acos_op.cc
  15. acos_op.cu
  16. acos_op.h
  17. activation_ops_cudnn.h
  18. affine_channel_op.cc
  19. affine_channel_op.cu
  20. affine_channel_op.h
  21. apmeter_op.cc
  22. apmeter_op.h
  23. arg_ops.cc
  24. arg_ops.cu
  25. arg_ops.h
  26. asin_op.cc
  27. asin_op.cu
  28. asin_op.h
  29. assert_op.cc
  30. assert_op.cu
  31. assert_op.h
  32. atan_op.cc
  33. atan_op.cu
  34. atan_op.h
  35. atomic_ops.cc
  36. batch_box_cox_op.cc
  37. batch_box_cox_op.h
  38. batch_bucketize_op.cc
  39. batch_bucketize_op.h
  40. batch_gather_ops.cc
  41. batch_gather_ops.cu
  42. batch_gather_ops.h
  43. batch_matmul_op.cc
  44. batch_matmul_op.cu
  45. batch_matmul_op.h
  46. batch_matmul_op_gpu_test.cc
  47. batch_matmul_op_test.cc
  48. batch_moments_op.cc
  49. batch_moments_op.cu
  50. batch_moments_op.h
  51. batch_sparse_to_dense_op.cc
  52. batch_sparse_to_dense_op.h
  53. bbox_transform_op.cc
  54. bbox_transform_op.h
  55. bisect_percentile_op.cc
  56. bisect_percentile_op.h
  57. boolean_mask_ops.cc
  58. boolean_mask_ops.cu
  59. boolean_mask_ops.h
  60. boolean_unmask_ops.cc
  61. boolean_unmask_ops.cu
  62. boolean_unmask_ops.h
  63. boolean_unmask_ops_test.cc
  64. box_with_nms_limit_op.cc
  65. box_with_nms_limit_op.h
  66. bucketize_op.cc
  67. bucketize_op.cu
  68. bucketize_op.h
  69. byte_weight_dequant_op.cc
  70. byte_weight_dequant_op.h
  71. cast_op.cc
  72. cast_op.cu
  73. cast_op.h
  74. cbrt_op.cc
  75. cbrt_op.cu
  76. cbrt_op.h
  77. cc_bmm_bg_op.cc
  78. cc_bmm_bg_op.h
  79. ceil_op.cc
  80. ceil_op.cu
  81. ceil_op.h
  82. channel_backprop_stats_op.cc
  83. channel_backprop_stats_op.cu
  84. channel_backprop_stats_op.h
  85. channel_shuffle_op.cc
  86. channel_shuffle_op.cu
  87. channel_shuffle_op.h
  88. channel_stats_op.cc
  89. channel_stats_op.cu
  90. channel_stats_op.h
  91. clip_op.cc
  92. clip_op.cu
  93. clip_op.h
  94. CMakeLists.txt
  95. collect_and_distribute_fpn_rpn_proposals_op.cc
  96. collect_and_distribute_fpn_rpn_proposals_op.h
  97. communicator_op.cc
  98. communicator_op_gpu.cc
  99. concat_split_op.cc
  100. concat_split_op.h
  101. concat_split_op_gpu.cc
  102. conditional_op.cc
  103. conditional_op.h
  104. conv_gradient_op.cc
  105. conv_op.cc
  106. conv_op.h
  107. conv_op_cache_cudnn.cc
  108. conv_op_cache_cudnn.h
  109. conv_op_cache_cudnn_test.cc
  110. conv_op_cudnn.cc
  111. conv_op_eigen.cc
  112. conv_op_gpu.cc
  113. conv_op_impl.h
  114. conv_op_shared.cc
  115. conv_op_shared.h
  116. conv_op_shared_gpu.cc
  117. conv_pool_op_base.h
  118. conv_transpose_gradient_op.cc
  119. conv_transpose_op.cc
  120. conv_transpose_op.h
  121. conv_transpose_op_cudnn.cc
  122. conv_transpose_op_gpu.cc
  123. conv_transpose_op_impl.h
  124. conv_transpose_op_mobile.cc
  125. conv_transpose_op_mobile.h
  126. conv_transpose_op_mobile_impl.h
  127. conv_transpose_op_mobile_test.cc
  128. conv_transpose_unpool_op_base.h
  129. copy_op.cc
  130. copy_op.cu
  131. copy_op.h
  132. copy_rows_to_tensor_op.cc
  133. copy_rows_to_tensor_op.h
  134. cos_op.cc
  135. cos_op.cu
  136. cos_op.h
  137. cosh_op.cc
  138. cosh_op.cu
  139. cosh_op.h
  140. cosine_embedding_criterion_op.cc
  141. cosine_embedding_criterion_op.cu
  142. cosine_embedding_criterion_op.h
  143. counter_ops.cc
  144. counter_ops.h
  145. counter_ops_gpu.cc
  146. crash_op.cc
  147. create_scope_op.cc
  148. create_scope_op.h
  149. crf_viterbi_op.cc
  150. cross_entropy_op.cc
  151. cross_entropy_op.cu
  152. cross_entropy_op.h
  153. ctc_beam_search_decoder_op.cc
  154. ctc_beam_search_decoder_op.h
  155. ctc_greedy_decoder_op.cc
  156. ctc_greedy_decoder_op.h
  157. cube_op.cc
  158. cube_op.cu
  159. cube_op.h
  160. data_couple.cc
  161. data_couple.h
  162. data_couple_gpu.cu
  163. dataset_ops.cc
  164. dataset_ops.h
  165. deform_conv_gradient_op.cc
  166. deform_conv_op.cc
  167. deform_conv_op.cu
  168. deform_conv_op.h
  169. deform_conv_op_impl.h
  170. dense_vector_to_id_list_op.cc
  171. dense_vector_to_id_list_op.h
  172. depthwise_3x3_conv_op_cudnn.cu
  173. distance_op.cc
  174. distance_op.cu
  175. distance_op.h
  176. do_op.cc
  177. do_op.h
  178. do_op_gpu.cc
  179. dropout_op.cc
  180. dropout_op.cu
  181. dropout_op.h
  182. dropout_op_cudnn.cc
  183. elementwise_add_gradient_op.cc
  184. elementwise_add_op.cc
  185. elementwise_add_op.h
  186. elementwise_add_op_gpu.cc
  187. elementwise_div_gradient_op.cc
  188. elementwise_div_op.cc
  189. elementwise_div_op.cu
  190. elementwise_div_op.h
  191. elementwise_linear_op.cc
  192. elementwise_linear_op.cu
  193. elementwise_linear_op.h
  194. elementwise_logical_ops.cc
  195. elementwise_logical_ops.h
  196. elementwise_mul_gradient_op.cc
  197. elementwise_mul_op.cc
  198. elementwise_mul_op.cu
  199. elementwise_mul_op.h
  200. elementwise_op_gpu_test.cc
  201. elementwise_op_test.cc
  202. elementwise_op_test.h
  203. elementwise_ops.cc
  204. elementwise_ops.cu
  205. elementwise_ops.h
  206. elementwise_ops_schema.cc
  207. elementwise_ops_utils.cc
  208. elementwise_ops_utils.h
  209. elementwise_sub_gradient_op.cc
  210. elementwise_sub_op.cc
  211. elementwise_sub_op.h
  212. elementwise_sub_op_gpu.cc
  213. elementwise_sum_op.cc
  214. elu_op.cc
  215. elu_op.cu
  216. elu_op.h
  217. elu_op_cudnn.cc
  218. enforce_finite_op.cc
  219. enforce_finite_op.cu
  220. enforce_finite_op.h
  221. ensure_clipped_op.cc
  222. ensure_clipped_op.h
  223. ensure_cpu_output_op.cc
  224. ensure_cpu_output_op.cu
  225. ensure_cpu_output_op.h
  226. erf_op.cc
  227. erf_op.cu
  228. erf_op.h
  229. exp_op.cc
  230. exp_op.h
  231. exp_op_gpu.cc
  232. expand_op.cc
  233. expand_op.h
  234. expand_op_gpu.cc
  235. expand_squeeze_dims_op.cc
  236. expand_squeeze_dims_op.h
  237. expand_squeeze_dims_op_gpu.cc
  238. fc_inference.cc
  239. fc_inference.h
  240. feature_maps_ops.cc
  241. feature_maps_ops.h
  242. feed_blob_op.cc
  243. feed_blob_op.h
  244. filler_op.cc
  245. filler_op.cu
  246. filler_op.h
  247. find_duplicate_elements_op.cc
  248. find_duplicate_elements_op.h
  249. find_op.cc
  250. find_op.cu
  251. find_op.h
  252. flatten_op.cc
  253. flatten_op.h
  254. flexible_top_k.cc
  255. flexible_top_k.h
  256. floor_op.cc
  257. floor_op.cu
  258. floor_op.h
  259. free_op.cc
  260. free_op.h
  261. free_op_gpu.cc
  262. fully_connected_op.cc
  263. fully_connected_op.h
  264. fully_connected_op_gpu.cc
  265. fused_rowwise_8bit_conversion_ops.cc
  266. fused_rowwise_8bit_conversion_ops.h
  267. fused_rowwise_random_quantization_ops.cc
  268. fused_rowwise_random_quantization_ops.h
  269. gather_fused_8bit_rowwise_op.cc
  270. gather_fused_8bit_rowwise_op.h
  271. gather_op.cc
  272. gather_op.cu
  273. gather_op.cuh
  274. gather_op.h
  275. gather_ranges_to_dense_op.cc
  276. gather_ranges_to_dense_op.h
  277. gelu_op.cc
  278. gelu_op.cu
  279. gelu_op.h
  280. generate_proposals_op.cc
  281. generate_proposals_op.cu
  282. generate_proposals_op.h
  283. generate_proposals_op_gpu_test.cc
  284. generate_proposals_op_test.cc
  285. generate_proposals_op_util_boxes.h
  286. generate_proposals_op_util_boxes_test.cc
  287. generate_proposals_op_util_nms.h
  288. generate_proposals_op_util_nms_gpu.cu
  289. generate_proposals_op_util_nms_gpu.h
  290. generate_proposals_op_util_nms_gpu_test.cc
  291. generate_proposals_op_util_nms_test.cc
  292. given_tensor_byte_string_to_uint8_fill_op.cc
  293. given_tensor_byte_string_to_uint8_fill_op.cu
  294. given_tensor_byte_string_to_uint8_fill_op.h
  295. given_tensor_fill_op.cc
  296. given_tensor_fill_op.cu
  297. given_tensor_fill_op.h
  298. glu_op.cc
  299. glu_op.cu
  300. glu_op.h
  301. group_norm_op.cc
  302. group_norm_op.cu
  303. group_norm_op.h
  304. gru_unit_op.cc
  305. gru_unit_op.h
  306. gru_unit_op_gpu.cu
  307. h_softmax_op.cc
  308. h_softmax_op.h
  309. half_float_ops.cc
  310. half_float_ops.cu
  311. half_float_ops.h
  312. half_float_ops_test.cc
  313. hard_sigmoid_op.cc
  314. hard_sigmoid_op.cu
  315. hard_sigmoid_op.h
  316. heatmap_max_keypoint_op.cc
  317. heatmap_max_keypoint_op.h
  318. if_op.cc
  319. if_op.h
  320. if_op_gpu.cc
  321. im2col_op.cc
  322. im2col_op.h
  323. im2col_op_gpu.cc
  324. index_hash_ops.cc
  325. index_hash_ops.h
  326. index_ops.cc
  327. index_ops.h
  328. inference_lstm_op.cc
  329. inference_lstm_op.h
  330. instance_norm_gradient_op.cc
  331. instance_norm_op.cc
  332. instance_norm_op.cu
  333. instance_norm_op.h
  334. integral_image_op.cc
  335. integral_image_op.cu
  336. integral_image_op.h
  337. is_empty_op.cc
  338. is_empty_op.h
  339. jsd_op.cc
  340. jsd_op.h
  341. key_split_ops.cc
  342. key_split_ops.h
  343. last_n_window_collector.cc
  344. layer_norm_op.cc
  345. layer_norm_op.cu
  346. layer_norm_op.h
  347. leaky_relu_op.cc
  348. leaky_relu_op.cu
  349. leaky_relu_op.h
  350. length_split_op.cc
  351. length_split_op.h
  352. lengths_pad_op.cc
  353. lengths_pad_op.cu
  354. lengths_pad_op.h
  355. lengths_reducer_fused_8bit_rowwise_ops.cc
  356. lengths_reducer_fused_8bit_rowwise_ops.h
  357. lengths_reducer_ops.cc
  358. lengths_reducer_ops.h
  359. lengths_reducer_rowwise_8bit_ops.cc
  360. lengths_reducer_rowwise_8bit_ops.h
  361. lengths_tile_op.cc
  362. lengths_tile_op.cu
  363. lengths_tile_op.h
  364. lengths_top_k_op.cc
  365. lengths_top_k_op.h
  366. listwise_l2r_op.cc
  367. listwise_l2r_op.h
  368. load_save_op.cc
  369. load_save_op.h
  370. load_save_op_gpu.cc
  371. local_response_normalization_op.cc
  372. local_response_normalization_op.cu
  373. local_response_normalization_op.h
  374. local_response_normalization_op_cudnn.cc
  375. locally_connected_op.cc
  376. locally_connected_op.h
  377. locally_connected_op_gpu.cc
  378. locally_connected_op_impl.h
  379. locally_connected_op_util.cc
  380. locally_connected_op_util.h
  381. log_op.cc
  382. log_op.h
  383. log_op_gpu.cc
  384. logit_op.cc
  385. logit_op.cu
  386. logit_op.h
  387. loss_op.cc
  388. loss_op.cu
  389. loss_op.h
  390. lp_pool_op.cc
  391. lp_pool_op.cu
  392. lpnorm_op.cc
  393. lpnorm_op.h
  394. lstm_unit_op.cc
  395. lstm_unit_op.h
  396. lstm_unit_op_gpu.cu
  397. lstm_utils.h
  398. map_ops.cc
  399. map_ops.h
  400. margin_ranking_criterion_op.cc
  401. margin_ranking_criterion_op.cu
  402. margin_ranking_criterion_op.h
  403. matmul_op.cc
  404. matmul_op.h
  405. matmul_op_gpu.cc
  406. max_pool_with_index.cu
  407. max_pool_with_index_gpu.h
  408. mean_op.cc
  409. mean_op.cu
  410. mean_op.h
  411. mem_query_op.cu
  412. merge_id_lists_op.cc
  413. merge_id_lists_op.h
  414. minmax_gradient_ops.cc
  415. minmax_ops.cc
  416. minmax_ops.cu
  417. minmax_ops.h
  418. mod_op.cc
  419. mod_op.h
  420. moments_op.cc
  421. moments_op.cu
  422. moments_op.h
  423. multi_class_accuracy_op.cc
  424. multi_class_accuracy_op.cu
  425. multi_class_accuracy_op.h
  426. negate_gradient_op.cc
  427. negate_gradient_op.h
  428. negate_gradient_op_gpu.cc
  429. negative_op.cc
  430. negative_op.h
  431. negative_op_gpu.cc
  432. ngram_ops.cc
  433. ngram_ops.h
  434. no_default_engine_op.h
  435. norm_planar_yuv_op.cc
  436. normalize_l1_op.cc
  437. normalize_l1_op.h
  438. normalize_op.cc
  439. normalize_op.h
  440. normalize_ops.cu
  441. numpy_tile_op.cc
  442. numpy_tile_op.h
  443. one_hot_ops.cc
  444. one_hot_ops.cu
  445. one_hot_ops.h
  446. onnx_while_op.cc
  447. onnx_while_op.h
  448. op_utils_cudnn.h
  449. operator_fallback_gpu.h
  450. operator_fallback_gpu_test.cc
  451. order_switch_ops.cc
  452. order_switch_ops.h
  453. order_switch_ops_cudnn.cc
  454. order_switch_ops_gpu.cc
  455. pack_rnn_sequence_op.cc
  456. pack_rnn_sequence_op.h
  457. pack_segments.cc
  458. pack_segments.cu
  459. pack_segments.h
  460. pad_op.cc
  461. pad_op.h
  462. pad_op_gpu.cu
  463. partition_ops.cc
  464. partition_ops.h
  465. percentile_op.cc
  466. percentile_op.h
  467. perplexity_op.cc
  468. perplexity_op.cu
  469. perplexity_op.h
  470. piecewise_linear_transform_op.cc
  471. piecewise_linear_transform_op.cu
  472. piecewise_linear_transform_op.h
  473. pool_gradient_op.cc
  474. pool_op.cc
  475. pool_op.cu
  476. pool_op.h
  477. pool_op_cudnn.cc
  478. pool_op_util.cc
  479. pool_op_util.h
  480. pow_op.cc
  481. pow_op.cu
  482. pow_op.h
  483. prefetch_op.h
  484. prelu_op.cc
  485. prelu_op.cu
  486. prelu_op.h
  487. prepend_dim_op.cc
  488. prepend_dim_op.h
  489. prepend_dim_op_gpu.cc
  490. quant_decode_op.cc
  491. quant_decode_op.h
  492. rank_loss_op.cc
  493. rank_loss_op.h
  494. reciprocal_gradient_op.cc
  495. reciprocal_op.cc
  496. reciprocal_op.cu
  497. reciprocal_op.h
  498. reduce_front_back_max_ops.cc
  499. reduce_front_back_max_ops.cu
  500. reduce_front_back_max_ops.h
  501. reduce_front_back_mean_ops.cc
  502. reduce_front_back_sum_mean_ops.cu
  503. reduce_front_back_sum_mean_ops.h
  504. reduce_front_back_sum_ops.cc
  505. reduce_ops.cc
  506. reduce_ops.cu
  507. reduce_ops.h
  508. reducer_functors.h
  509. reduction_ops.cc
  510. reduction_ops.cu
  511. reduction_ops.h
  512. relu_n_op.cc
  513. relu_n_op.cu
  514. relu_n_op.h
  515. relu_op.cc
  516. relu_op.cu
  517. relu_op.h
  518. remove_data_blocks_op.cc
  519. remove_data_blocks_op.h
  520. replace_nan_op.cc
  521. replace_nan_op.cu
  522. replace_nan_op.h
  523. reservoir_sampling.cc
  524. reshape_op.cc
  525. reshape_op.h
  526. reshape_op_gpu.cc
  527. reshape_op_gpu_test.cc
  528. resize_op.cc
  529. resize_op.cu
  530. resize_op.h
  531. reverse_packed_segs_op.cc
  532. reverse_packed_segs_op.cu
  533. reverse_packed_segs_op.h
  534. rmac_regions_op.cc
  535. rmac_regions_op.cu
  536. rmac_regions_op.h
  537. roi_align_gradient_op.cc
  538. roi_align_gradient_op.cu
  539. roi_align_gradient_op.h
  540. roi_align_op.cc
  541. roi_align_op.cu
  542. roi_align_op.h
  543. roi_align_op_gpu_test.cc
  544. roi_align_rotated_gradient_op.cc
  545. roi_align_rotated_gradient_op.cu
  546. roi_align_rotated_gradient_op.h
  547. roi_align_rotated_op.cc
  548. roi_align_rotated_op.cu
  549. roi_align_rotated_op.h
  550. roi_pool_op.cc
  551. roi_pool_op.cu
  552. roi_pool_op.h
  553. rowmul_op.cc
  554. rowmul_op.h
  555. rsqrt_op.cc
  556. rsqrt_op.cu
  557. rsqrt_op.h
  558. scale_blobs_op.cc
  559. scale_blobs_op.cu
  560. scale_blobs_op.h
  561. scale_op.cc
  562. scale_op.h
  563. scale_op_gpu.cc
  564. segment_reduction_op.cc
  565. segment_reduction_op.h
  566. segment_reduction_op_gpu.cu
  567. selu_op.cc
  568. selu_op.cu
  569. selu_op.h
  570. sequence_ops.cc
  571. sequence_ops.cu
  572. sequence_ops.h
  573. shape_op.cc
  574. shape_op.h
  575. shape_op_gpu.cc
  576. sigmoid_gradient_op.cc
  577. sigmoid_op.cc
  578. sigmoid_op.cu
  579. sigmoid_op.h
  580. sigmoid_op_cudnn.cc
  581. sin_op.cc
  582. sin_op.cu
  583. sin_op.h
  584. sinh_op.cc
  585. sinh_op.cu
  586. sinh_op.h
  587. sinusoid_position_encoding_op.cc
  588. sinusoid_position_encoding_op.h
  589. slice_op.cc
  590. slice_op.cu
  591. slice_op.h
  592. softmax_op.cc
  593. softmax_op.h
  594. softmax_op_cudnn.cc
  595. softmax_ops.cu
  596. softmax_utils.cc
  597. softmax_utils.h
  598. softmax_with_loss_op.cc
  599. softmax_with_loss_op.h
  600. softplus_op.cc
  601. softplus_op.cu
  602. softplus_op.h
  603. softsign_op.cc
  604. softsign_op.cu
  605. softsign_op.h
  606. space_batch_op.cc
  607. space_batch_op.h
  608. space_batch_op_gpu.cu
  609. sparse_normalize_op.cc
  610. sparse_normalize_op.h
  611. sparse_normalize_op_gpu.cu
  612. sparse_to_dense_mask_op.cc
  613. sparse_to_dense_mask_op.h
  614. sparse_to_dense_op.cc
  615. sparse_to_dense_op.cu
  616. sparse_to_dense_op.h
  617. spatial_batch_norm_gradient_op.cc
  618. spatial_batch_norm_op.cc
  619. spatial_batch_norm_op.cu
  620. spatial_batch_norm_op.h
  621. spatial_batch_norm_op_cudnn.cu
  622. spatial_batch_norm_op_impl.cuh
  623. spatial_softmax_with_loss_op.cc
  624. spatial_softmax_with_loss_op.h
  625. sqr_op.cc
  626. sqr_op.h
  627. sqr_op_gpu.cc
  628. sqrt_op.cc
  629. sqrt_op.h
  630. sqrt_op_gpu.cc
  631. square_root_divide_op.cc
  632. square_root_divide_op.h
  633. stats_ops.cc
  634. stats_put_ops.cc
  635. stats_put_ops.h
  636. stop_gradient.cc
  637. stop_gradient.h
  638. stop_gradient_gpu.cc
  639. string_ops.cc
  640. string_ops.h
  641. string_ops_test.cc
  642. stump_func_op.cc
  643. stump_func_op.cu
  644. stump_func_op.h
  645. stylizer_ops.cc
  646. summarize_op.cc
  647. summarize_op.cu
  648. summarize_op.h
  649. swish_op.cc
  650. swish_op.cu
  651. swish_op.h
  652. tan_op.cc
  653. tan_op.cu
  654. tan_op.h
  655. tanh_gradient_op.cc
  656. tanh_op.cc
  657. tanh_op.cu
  658. tanh_op.h
  659. tanh_op_cudnn.cc
  660. tensor_protos_db_input.cc
  661. tensor_protos_db_input.h
  662. tensor_protos_db_input_gpu.cc
  663. text_file_reader.cc
  664. text_file_reader_utils.cc
  665. text_file_reader_utils.h
  666. text_file_reader_utils_test.cc
  667. thresholded_relu_op.cc
  668. thresholded_relu_op.cu
  669. thresholded_relu_op.h
  670. tile_op.cc
  671. tile_op.cu
  672. tile_op.h
  673. top_k.cc
  674. top_k.cu
  675. top_k.h
  676. top_k_heap_selection.cuh
  677. top_k_radix_selection.cuh
  678. transpose_op.cc
  679. transpose_op.cu
  680. transpose_op.h
  681. transpose_op_cudnn.cc
  682. tt_linear_op.cc
  683. tt_linear_op.h
  684. unique_ops.cc
  685. unique_ops.cu
  686. unique_ops.h
  687. upsample_op.cc
  688. upsample_op.cu
  689. upsample_op.h
  690. utility_ops.cc
  691. utility_ops.cu
  692. utility_ops.h
  693. utility_ops_gpu_test.cc
  694. utility_ops_test.cc
  695. variable_length_sequence_padding.cc
  696. variable_length_sequence_padding.h
  697. weighted_multi_sampling_op.cc
  698. weighted_multi_sampling_op.h
  699. weighted_sample_op.cc
  700. weighted_sample_op.cu
  701. weighted_sample_op.h
  702. while_op.cc
  703. while_op.h
  704. while_op_gpu.cc
  705. workspace_ops.cc
  706. zero_gradient_op.cc
  707. zero_gradient_op.h
  708. zero_gradient_op_gpu.cc