Google Git
Sign in
android / platform / external / pytorch / f23fb66e6e / . / caffe2 / operators
tree: da86dbbaf192ad0617e8fbdb6fac396c66eba593 [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. byte_weight_dequant_op.cc
  67. byte_weight_dequant_op.h
  68. cast_op.cc
  69. cast_op.cu
  70. cast_op.h
  71. cbrt_op.cc
  72. cbrt_op.cu
  73. cbrt_op.h
  74. cc_bmm_bg_op.cc
  75. cc_bmm_bg_op.h
  76. ceil_op.cc
  77. ceil_op.cu
  78. ceil_op.h
  79. channel_backprop_stats_op.cc
  80. channel_backprop_stats_op.cu
  81. channel_backprop_stats_op.h
  82. channel_shuffle_op.cc
  83. channel_shuffle_op.cu
  84. channel_shuffle_op.h
  85. channel_stats_op.cc
  86. channel_stats_op.cu
  87. channel_stats_op.h
  88. clip_op.cc
  89. clip_op.cu
  90. clip_op.h
  91. CMakeLists.txt
  92. collect_and_distribute_fpn_rpn_proposals_op.cc
  93. collect_and_distribute_fpn_rpn_proposals_op.h
  94. communicator_op.cc
  95. communicator_op_gpu.cc
  96. concat_split_op.cc
  97. concat_split_op.h
  98. concat_split_op_gpu.cc
  99. conditional_op.cc
  100. conditional_op.h
  101. conv_gradient_op.cc
  102. conv_op.cc
  103. conv_op.h
  104. conv_op_cache_cudnn.cc
  105. conv_op_cache_cudnn.h
  106. conv_op_cache_cudnn_test.cc
  107. conv_op_cudnn.cc
  108. conv_op_eigen.cc
  109. conv_op_gpu.cc
  110. conv_op_impl.h
  111. conv_op_shared.cc
  112. conv_op_shared.h
  113. conv_op_shared_gpu.cc
  114. conv_pool_op_base.h
  115. conv_transpose_gradient_op.cc
  116. conv_transpose_op.cc
  117. conv_transpose_op.h
  118. conv_transpose_op_cudnn.cc
  119. conv_transpose_op_gpu.cc
  120. conv_transpose_op_impl.h
  121. conv_transpose_op_mobile.cc
  122. conv_transpose_op_mobile.h
  123. conv_transpose_op_mobile_impl.h
  124. conv_transpose_op_mobile_test.cc
  125. conv_transpose_unpool_op_base.h
  126. copy_op.cc
  127. copy_op.cu
  128. copy_op.h
  129. cos_op.cc
  130. cos_op.cu
  131. cos_op.h
  132. cosh_op.cc
  133. cosh_op.cu
  134. cosh_op.h
  135. cosine_embedding_criterion_op.cc
  136. cosine_embedding_criterion_op.cu
  137. cosine_embedding_criterion_op.h
  138. counter_ops.cc
  139. counter_ops.h
  140. counter_ops_gpu.cc
  141. crash_op.cc
  142. create_scope_op.cc
  143. create_scope_op.h
  144. crf_viterbi_op.cc
  145. cross_entropy_op.cc
  146. cross_entropy_op.cu
  147. cross_entropy_op.h
  148. ctc_beam_search_decoder_op.cc
  149. ctc_beam_search_decoder_op.h
  150. ctc_greedy_decoder_op.cc
  151. ctc_greedy_decoder_op.h
  152. cube_op.cc
  153. cube_op.cu
  154. cube_op.h
  155. data_couple.cc
  156. data_couple.h
  157. data_couple_gpu.cu
  158. dataset_ops.cc
  159. dataset_ops.h
  160. deform_conv_gradient_op.cc
  161. deform_conv_op.cc
  162. deform_conv_op.cu
  163. deform_conv_op.h
  164. deform_conv_op_impl.h
  165. dense_vector_to_id_list_op.cc
  166. dense_vector_to_id_list_op.h
  167. depthwise_3x3_conv_op_cudnn.cu
  168. distance_op.cc
  169. distance_op.cu
  170. distance_op.h
  171. do_op.cc
  172. do_op.h
  173. do_op_gpu.cc
  174. dropout_op.cc
  175. dropout_op.cu
  176. dropout_op.h
  177. dropout_op_cudnn.cc
  178. elementwise_add_gradient_op.cc
  179. elementwise_add_op.cc
  180. elementwise_add_op.h
  181. elementwise_add_op_gpu.cc
  182. elementwise_div_gradient_op.cc
  183. elementwise_div_op.cc
  184. elementwise_div_op.cu
  185. elementwise_div_op.h
  186. elementwise_linear_op.cc
  187. elementwise_linear_op.cu
  188. elementwise_linear_op.h
  189. elementwise_logical_ops.cc
  190. elementwise_logical_ops.h
  191. elementwise_mul_gradient_op.cc
  192. elementwise_mul_op.cc
  193. elementwise_mul_op.cu
  194. elementwise_mul_op.h
  195. elementwise_op_gpu_test.cc
  196. elementwise_op_test.cc
  197. elementwise_op_test.h
  198. elementwise_ops.cc
  199. elementwise_ops.cu
  200. elementwise_ops.h
  201. elementwise_ops_schema.cc
  202. elementwise_ops_utils.cc
  203. elementwise_ops_utils.h
  204. elementwise_sub_gradient_op.cc
  205. elementwise_sub_op.cc
  206. elementwise_sub_op.h
  207. elementwise_sub_op_gpu.cc
  208. elementwise_sum_op.cc
  209. elu_op.cc
  210. elu_op.cu
  211. elu_op.h
  212. elu_op_cudnn.cc
  213. enforce_finite_op.cc
  214. enforce_finite_op.cu
  215. enforce_finite_op.h
  216. ensure_clipped_op.cc
  217. ensure_clipped_op.h
  218. ensure_cpu_output_op.cc
  219. ensure_cpu_output_op.cu
  220. ensure_cpu_output_op.h
  221. erf_op.cc
  222. erf_op.cu
  223. erf_op.h
  224. exp_op.cc
  225. exp_op.h
  226. exp_op_gpu.cc
  227. expand_op.cc
  228. expand_op.h
  229. expand_op_gpu.cc
  230. expand_squeeze_dims_op.cc
  231. expand_squeeze_dims_op.h
  232. expand_squeeze_dims_op_gpu.cc
  233. fc_inference.cc
  234. fc_inference.h
  235. feature_maps_ops.cc
  236. feature_maps_ops.h
  237. feed_blob_op.cc
  238. feed_blob_op.h
  239. filler_op.cc
  240. filler_op.cu
  241. filler_op.h
  242. find_duplicate_elements_op.cc
  243. find_duplicate_elements_op.h
  244. find_op.cc
  245. find_op.cu
  246. find_op.h
  247. flatten_op.cc
  248. flatten_op.h
  249. flexible_top_k.cc
  250. flexible_top_k.h
  251. floor_op.cc
  252. floor_op.cu
  253. floor_op.h
  254. free_op.cc
  255. free_op.h
  256. free_op_gpu.cc
  257. fully_connected_op.cc
  258. fully_connected_op.h
  259. fully_connected_op_gpu.cc
  260. fused_rowwise_8bit_conversion_ops.cc
  261. fused_rowwise_8bit_conversion_ops.h
  262. fused_rowwise_random_quantization_ops.cc
  263. fused_rowwise_random_quantization_ops.h
  264. gather_fused_8bit_rowwise_op.cc
  265. gather_fused_8bit_rowwise_op.h
  266. gather_op.cc
  267. gather_op.cu
  268. gather_op.cuh
  269. gather_op.h
  270. gather_ranges_to_dense_op.cc
  271. gather_ranges_to_dense_op.h
  272. gelu_op.cc
  273. gelu_op.cu
  274. gelu_op.h
  275. generate_proposals_op.cc
  276. generate_proposals_op.cu
  277. generate_proposals_op.h
  278. generate_proposals_op_gpu_test.cc
  279. generate_proposals_op_test.cc
  280. generate_proposals_op_util_boxes.h
  281. generate_proposals_op_util_boxes_test.cc
  282. generate_proposals_op_util_nms.h
  283. generate_proposals_op_util_nms_gpu.cu
  284. generate_proposals_op_util_nms_gpu.h
  285. generate_proposals_op_util_nms_gpu_test.cc
  286. generate_proposals_op_util_nms_test.cc
  287. given_tensor_byte_string_to_uint8_fill_op.cc
  288. given_tensor_byte_string_to_uint8_fill_op.cu
  289. given_tensor_byte_string_to_uint8_fill_op.h
  290. given_tensor_fill_op.cc
  291. given_tensor_fill_op.cu
  292. given_tensor_fill_op.h
  293. glu_op.cc
  294. glu_op.cu
  295. glu_op.h
  296. group_norm_op.cc
  297. group_norm_op.cu
  298. group_norm_op.h
  299. gru_unit_op.cc
  300. gru_unit_op.h
  301. gru_unit_op_gpu.cu
  302. h_softmax_op.cc
  303. h_softmax_op.h
  304. half_float_ops.cc
  305. half_float_ops.cu
  306. half_float_ops.h
  307. half_float_ops_test.cc
  308. hard_sigmoid_op.cc
  309. hard_sigmoid_op.cu
  310. hard_sigmoid_op.h
  311. heatmap_max_keypoint_op.cc
  312. heatmap_max_keypoint_op.h
  313. if_op.cc
  314. if_op.h
  315. if_op_gpu.cc
  316. im2col_op.cc
  317. im2col_op.h
  318. im2col_op_gpu.cc
  319. index_hash_ops.cc
  320. index_hash_ops.h
  321. index_ops.cc
  322. index_ops.h
  323. inference_lstm_op.cc
  324. inference_lstm_op.h
  325. instance_norm_gradient_op.cc
  326. instance_norm_op.cc
  327. instance_norm_op.cu
  328. instance_norm_op.h
  329. integral_image_op.cc
  330. integral_image_op.cu
  331. integral_image_op.h
  332. is_empty_op.cc
  333. is_empty_op.h
  334. jsd_op.cc
  335. jsd_op.h
  336. key_split_ops.cc
  337. key_split_ops.h
  338. last_n_window_collector.cc
  339. layer_norm_op.cc
  340. layer_norm_op.cu
  341. layer_norm_op.h
  342. leaky_relu_op.cc
  343. leaky_relu_op.cu
  344. leaky_relu_op.h
  345. length_split_op.cc
  346. length_split_op.h
  347. lengths_pad_op.cc
  348. lengths_pad_op.cu
  349. lengths_pad_op.h
  350. lengths_reducer_fused_8bit_rowwise_ops.cc
  351. lengths_reducer_fused_8bit_rowwise_ops.h
  352. lengths_reducer_ops.cc
  353. lengths_reducer_ops.h
  354. lengths_reducer_rowwise_8bit_ops.cc
  355. lengths_reducer_rowwise_8bit_ops.h
  356. lengths_tile_op.cc
  357. lengths_tile_op.cu
  358. lengths_tile_op.h
  359. lengths_top_k_op.cc
  360. lengths_top_k_op.h
  361. listwise_l2r_op.cc
  362. listwise_l2r_op.h
  363. load_save_op.cc
  364. load_save_op.h
  365. load_save_op_gpu.cc
  366. local_response_normalization_op.cc
  367. local_response_normalization_op.cu
  368. local_response_normalization_op.h
  369. local_response_normalization_op_cudnn.cc
  370. locally_connected_op.cc
  371. locally_connected_op.h
  372. locally_connected_op_gpu.cc
  373. locally_connected_op_impl.h
  374. locally_connected_op_util.cc
  375. locally_connected_op_util.h
  376. log_op.cc
  377. log_op.h
  378. log_op_gpu.cc
  379. logit_op.cc
  380. logit_op.cu
  381. logit_op.h
  382. loss_op.cc
  383. loss_op.cu
  384. loss_op.h
  385. lp_pool_op.cc
  386. lp_pool_op.cu
  387. lpnorm_op.cc
  388. lpnorm_op.h
  389. lstm_unit_op.cc
  390. lstm_unit_op.h
  391. lstm_unit_op_gpu.cu
  392. lstm_utils.h
  393. map_ops.cc
  394. map_ops.h
  395. margin_ranking_criterion_op.cc
  396. margin_ranking_criterion_op.cu
  397. margin_ranking_criterion_op.h
  398. matmul_op.cc
  399. matmul_op.h
  400. matmul_op_gpu.cc
  401. max_pool_with_index.cu
  402. max_pool_with_index_gpu.h
  403. mean_op.cc
  404. mean_op.cu
  405. mean_op.h
  406. mem_query_op.cu
  407. merge_id_lists_op.cc
  408. merge_id_lists_op.h
  409. minmax_gradient_ops.cc
  410. minmax_ops.cc
  411. minmax_ops.cu
  412. minmax_ops.h
  413. mod_op.cc
  414. mod_op.h
  415. moments_op.cc
  416. moments_op.cu
  417. moments_op.h
  418. multi_class_accuracy_op.cc
  419. multi_class_accuracy_op.cu
  420. multi_class_accuracy_op.h
  421. negate_gradient_op.cc
  422. negate_gradient_op.h
  423. negate_gradient_op_gpu.cc
  424. negative_op.cc
  425. negative_op.h
  426. negative_op_gpu.cc
  427. ngram_ops.cc
  428. ngram_ops.h
  429. no_default_engine_op.h
  430. norm_planar_yuv_op.cc
  431. normalize_l1_op.cc
  432. normalize_l1_op.h
  433. normalize_op.cc
  434. normalize_op.h
  435. normalize_ops.cu
  436. numpy_tile_op.cc
  437. numpy_tile_op.h
  438. one_hot_ops.cc
  439. one_hot_ops.cu
  440. one_hot_ops.h
  441. onnx_while_op.cc
  442. onnx_while_op.h
  443. onnxifi_op.cc
  444. onnxifi_op.h
  445. op_utils_cudnn.h
  446. operator_fallback_gpu.h
  447. operator_fallback_gpu_test.cc
  448. order_switch_ops.cc
  449. order_switch_ops.h
  450. order_switch_ops_cudnn.cc
  451. order_switch_ops_gpu.cc
  452. pack_rnn_sequence_op.cc
  453. pack_rnn_sequence_op.h
  454. pack_segments.cc
  455. pack_segments.cu
  456. pack_segments.h
  457. pad_op.cc
  458. pad_op.h
  459. pad_op_gpu.cu
  460. partition_ops.cc
  461. partition_ops.h
  462. percentile_op.cc
  463. percentile_op.h
  464. perplexity_op.cc
  465. perplexity_op.cu
  466. perplexity_op.h
  467. piecewise_linear_transform_op.cc
  468. piecewise_linear_transform_op.cu
  469. piecewise_linear_transform_op.h
  470. pool_gradient_op.cc
  471. pool_op.cc
  472. pool_op.cu
  473. pool_op.h
  474. pool_op_cudnn.cc
  475. pool_op_util.cc
  476. pool_op_util.h
  477. pow_op.cc
  478. pow_op.cu
  479. pow_op.h
  480. prefetch_op.h
  481. prelu_op.cc
  482. prelu_op.cu
  483. prelu_op.h
  484. prepend_dim_op.cc
  485. prepend_dim_op.h
  486. prepend_dim_op_gpu.cc
  487. quant_decode_op.cc
  488. quant_decode_op.h
  489. rank_loss_op.cc
  490. rank_loss_op.h
  491. reciprocal_gradient_op.cc
  492. reciprocal_op.cc
  493. reciprocal_op.cu
  494. reciprocal_op.h
  495. reduce_front_back_max_ops.cc
  496. reduce_front_back_max_ops.cu
  497. reduce_front_back_max_ops.h
  498. reduce_front_back_mean_ops.cc
  499. reduce_front_back_sum_mean_ops.cu
  500. reduce_front_back_sum_mean_ops.h
  501. reduce_front_back_sum_ops.cc
  502. reduce_ops.cc
  503. reduce_ops.cu
  504. reduce_ops.h
  505. reducer_functors.h
  506. reduction_ops.cc
  507. reduction_ops.cu
  508. reduction_ops.h
  509. relu_n_op.cc
  510. relu_n_op.cu
  511. relu_n_op.h
  512. relu_op.cc
  513. relu_op.cu
  514. relu_op.h
  515. remove_data_blocks_op.cc
  516. remove_data_blocks_op.h
  517. replace_nan_op.cc
  518. replace_nan_op.cu
  519. replace_nan_op.h
  520. reservoir_sampling.cc
  521. reshape_op.cc
  522. reshape_op.h
  523. reshape_op_gpu.cc
  524. reshape_op_gpu_test.cc
  525. resize_op.cc
  526. resize_op.cu
  527. resize_op.h
  528. reverse_packed_segs_op.cc
  529. reverse_packed_segs_op.cu
  530. reverse_packed_segs_op.h
  531. rmac_regions_op.cc
  532. rmac_regions_op.cu
  533. rmac_regions_op.h
  534. roi_align_gradient_op.cc
  535. roi_align_gradient_op.cu
  536. roi_align_gradient_op.h
  537. roi_align_op.cc
  538. roi_align_op.cu
  539. roi_align_op.h
  540. roi_align_op_gpu_test.cc
  541. roi_align_rotated_gradient_op.cc
  542. roi_align_rotated_gradient_op.cu
  543. roi_align_rotated_gradient_op.h
  544. roi_align_rotated_op.cc
  545. roi_align_rotated_op.cu
  546. roi_align_rotated_op.h
  547. roi_pool_op.cc
  548. roi_pool_op.cu
  549. roi_pool_op.h
  550. rowmul_op.cc
  551. rowmul_op.h
  552. rsqrt_op.cc
  553. rsqrt_op.cu
  554. rsqrt_op.h
  555. scale_blobs_op.cc
  556. scale_blobs_op.cu
  557. scale_blobs_op.h
  558. scale_op.cc
  559. scale_op.h
  560. scale_op_gpu.cc
  561. segment_reduction_op.cc
  562. segment_reduction_op.h
  563. segment_reduction_op_gpu.cu
  564. selu_op.cc
  565. selu_op.cu
  566. selu_op.h
  567. sequence_ops.cc
  568. sequence_ops.cu
  569. sequence_ops.h
  570. shape_op.cc
  571. shape_op.h
  572. shape_op_gpu.cc
  573. sigmoid_gradient_op.cc
  574. sigmoid_op.cc
  575. sigmoid_op.cu
  576. sigmoid_op.h
  577. sigmoid_op_cudnn.cc
  578. sin_op.cc
  579. sin_op.cu
  580. sin_op.h
  581. sinh_op.cc
  582. sinh_op.cu
  583. sinh_op.h
  584. sinusoid_position_encoding_op.cc
  585. sinusoid_position_encoding_op.h
  586. slice_op.cc
  587. slice_op.cu
  588. slice_op.h
  589. softmax_op.cc
  590. softmax_op.h
  591. softmax_op_cudnn.cc
  592. softmax_ops.cu
  593. softmax_utils.cc
  594. softmax_utils.h
  595. softmax_with_loss_op.cc
  596. softmax_with_loss_op.h
  597. softplus_op.cc
  598. softplus_op.cu
  599. softplus_op.h
  600. softsign_op.cc
  601. softsign_op.cu
  602. softsign_op.h
  603. space_batch_op.cc
  604. space_batch_op.h
  605. space_batch_op_gpu.cu
  606. sparse_normalize_op.cc
  607. sparse_normalize_op.h
  608. sparse_normalize_op_gpu.cu
  609. sparse_to_dense_mask_op.cc
  610. sparse_to_dense_mask_op.h
  611. sparse_to_dense_op.cc
  612. sparse_to_dense_op.cu
  613. sparse_to_dense_op.h
  614. spatial_batch_norm_gradient_op.cc
  615. spatial_batch_norm_op.cc
  616. spatial_batch_norm_op.cu
  617. spatial_batch_norm_op.h
  618. spatial_batch_norm_op_cudnn.cu
  619. spatial_batch_norm_op_impl.cuh
  620. spatial_softmax_with_loss_op.cc
  621. spatial_softmax_with_loss_op.h
  622. sqr_op.cc
  623. sqr_op.h
  624. sqr_op_gpu.cc
  625. sqrt_op.cc
  626. sqrt_op.h
  627. sqrt_op_gpu.cc
  628. square_root_divide_op.cc
  629. square_root_divide_op.h
  630. stats_ops.cc
  631. stats_put_ops.cc
  632. stats_put_ops.h
  633. stop_gradient.cc
  634. stop_gradient.h
  635. stop_gradient_gpu.cc
  636. string_ops.cc
  637. string_ops.h
  638. string_ops_test.cc
  639. stump_func_op.cc
  640. stump_func_op.cu
  641. stump_func_op.h
  642. stylizer_ops.cc
  643. summarize_op.cc
  644. summarize_op.cu
  645. summarize_op.h
  646. swish_op.cc
  647. swish_op.cu
  648. swish_op.h
  649. tan_op.cc
  650. tan_op.cu
  651. tan_op.h
  652. tanh_gradient_op.cc
  653. tanh_op.cc
  654. tanh_op.cu
  655. tanh_op.h
  656. tanh_op_cudnn.cc
  657. tensor_protos_db_input.cc
  658. tensor_protos_db_input.h
  659. tensor_protos_db_input_gpu.cc
  660. text_file_reader.cc
  661. text_file_reader_utils.cc
  662. text_file_reader_utils.h
  663. text_file_reader_utils_test.cc
  664. thresholded_relu_op.cc
  665. thresholded_relu_op.cu
  666. thresholded_relu_op.h
  667. tile_op.cc
  668. tile_op.cu
  669. tile_op.h
  670. top_k.cc
  671. top_k.cu
  672. top_k.h
  673. top_k_heap_selection.cuh
  674. top_k_radix_selection.cuh
  675. transpose_op.cc
  676. transpose_op.cu
  677. transpose_op.h
  678. transpose_op_cudnn.cc
  679. tt_linear_op.cc
  680. tt_linear_op.h
  681. unique_ops.cc
  682. unique_ops.cu
  683. unique_ops.h
  684. upsample_op.cc
  685. upsample_op.cu
  686. upsample_op.h
  687. utility_ops.cc
  688. utility_ops.cu
  689. utility_ops.h
  690. utility_ops_gpu_test.cc
  691. utility_ops_test.cc
  692. variable_length_sequence_padding.cc
  693. variable_length_sequence_padding.h
  694. weighted_multi_sampling_op.cc
  695. weighted_multi_sampling_op.h
  696. weighted_sample_op.cc
  697. weighted_sample_op.cu
  698. weighted_sample_op.h
  699. while_op.cc
  700. while_op.h
  701. while_op_gpu.cc
  702. workspace_ops.cc
  703. zero_gradient_op.cc
  704. zero_gradient_op.h
  705. zero_gradient_op_gpu.cc