CUDAFunctions_inl.h 30 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598
  1. #pragma once
  2. // @generated by torchgen/gen.py from DispatchKeyFunctions_inl.h
  3. // NB: The implementing C++ file is RegisterDispatchKey.cpp
  4. // The only #includes we need are for custom classes that have defaults in the C++ API
  5. #include <c10/core/MemoryFormat.h>
  6. #include <c10/core/Scalar.h>
  7. #include <ATen/core/Reduction.h>
  8. #if defined(AT_PER_OPERATOR_HEADERS) && defined(TORCH_ASSERT_ONLY_METHOD_OPERATORS)
  9. #error This change adds a dependency on all pytorch operators, meaning the \
  10. file will need to be re-compiled every time an operator is changed or added. \
  11. Consider including a specific operator from \
  12. <ATen/ops/{my_operator}_cuda_dispatch.h>. \
  13. See NOTE [TORCH_ASSERT_ONLY_METHOD_OPERATORS].
  14. #endif
  15. #include <ATen/ops/_adaptive_avg_pool2d_cuda_dispatch.h>
  16. #include <ATen/ops/_adaptive_avg_pool2d_backward_cuda_dispatch.h>
  17. #include <ATen/ops/_adaptive_avg_pool3d_cuda_dispatch.h>
  18. #include <ATen/ops/_adaptive_avg_pool3d_backward_cuda_dispatch.h>
  19. #include <ATen/ops/_addmm_activation_cuda_dispatch.h>
  20. #include <ATen/ops/_aminmax_cuda_dispatch.h>
  21. #include <ATen/ops/_amp_foreach_non_finite_check_and_unscale_cuda_dispatch.h>
  22. #include <ATen/ops/_amp_update_scale_cuda_dispatch.h>
  23. #include <ATen/ops/_assert_async_cuda_dispatch.h>
  24. #include <ATen/ops/_cdist_backward_cuda_dispatch.h>
  25. #include <ATen/ops/_cdist_forward_cuda_dispatch.h>
  26. #include <ATen/ops/_cholesky_solve_helper_cuda_dispatch.h>
  27. #include <ATen/ops/_chunk_grad_outputs_efficient_attention_cuda_dispatch.h>
  28. #include <ATen/ops/_compute_linear_combination_cuda_dispatch.h>
  29. #include <ATen/ops/_conv_depthwise2d_cuda_dispatch.h>
  30. #include <ATen/ops/_convert_indices_from_coo_to_csr_cuda_dispatch.h>
  31. #include <ATen/ops/_convert_indices_from_csr_to_coo_cuda_dispatch.h>
  32. #include <ATen/ops/_ctc_loss_cuda_dispatch.h>
  33. #include <ATen/ops/_ctc_loss_backward_cuda_dispatch.h>
  34. #include <ATen/ops/_cudnn_ctc_loss_cuda_dispatch.h>
  35. #include <ATen/ops/_cudnn_init_dropout_state_cuda_dispatch.h>
  36. #include <ATen/ops/_cudnn_rnn_cuda_dispatch.h>
  37. #include <ATen/ops/_cudnn_rnn_backward_cuda_dispatch.h>
  38. #include <ATen/ops/_cudnn_rnn_flatten_weight_cuda_dispatch.h>
  39. #include <ATen/ops/_cummax_helper_cuda_dispatch.h>
  40. #include <ATen/ops/_cummin_helper_cuda_dispatch.h>
  41. #include <ATen/ops/_dirichlet_grad_cuda_dispatch.h>
  42. #include <ATen/ops/_efficient_attention_backward_cuda_dispatch.h>
  43. #include <ATen/ops/_efficient_attention_forward_cuda_dispatch.h>
  44. #include <ATen/ops/_efficientzerotensor_cuda_dispatch.h>
  45. #include <ATen/ops/_embedding_bag_cuda_dispatch.h>
  46. #include <ATen/ops/_embedding_bag_dense_backward_cuda_dispatch.h>
  47. #include <ATen/ops/_embedding_bag_forward_only_cuda_dispatch.h>
  48. #include <ATen/ops/_embedding_bag_per_sample_weights_backward_cuda_dispatch.h>
  49. #include <ATen/ops/_fake_quantize_learnable_per_channel_affine_cuda_dispatch.h>
  50. #include <ATen/ops/_fake_quantize_learnable_per_channel_affine_backward_cuda_dispatch.h>
  51. #include <ATen/ops/_fake_quantize_learnable_per_tensor_affine_cuda_dispatch.h>
  52. #include <ATen/ops/_fake_quantize_learnable_per_tensor_affine_backward_cuda_dispatch.h>
  53. #include <ATen/ops/_fake_quantize_per_tensor_affine_cachemask_tensor_qparams_cuda_dispatch.h>
  54. #include <ATen/ops/_fft_c2c_cuda_dispatch.h>
  55. #include <ATen/ops/_fft_c2r_cuda_dispatch.h>
  56. #include <ATen/ops/_fft_r2c_cuda_dispatch.h>
  57. #include <ATen/ops/_flash_attention_backward_cuda_dispatch.h>
  58. #include <ATen/ops/_flash_attention_forward_cuda_dispatch.h>
  59. #include <ATen/ops/_foreach_abs_cuda_dispatch.h>
  60. #include <ATen/ops/_foreach_acos_cuda_dispatch.h>
  61. #include <ATen/ops/_foreach_add_cuda_dispatch.h>
  62. #include <ATen/ops/_foreach_addcdiv_cuda_dispatch.h>
  63. #include <ATen/ops/_foreach_addcmul_cuda_dispatch.h>
  64. #include <ATen/ops/_foreach_asin_cuda_dispatch.h>
  65. #include <ATen/ops/_foreach_atan_cuda_dispatch.h>
  66. #include <ATen/ops/_foreach_ceil_cuda_dispatch.h>
  67. #include <ATen/ops/_foreach_clamp_max_cuda_dispatch.h>
  68. #include <ATen/ops/_foreach_clamp_min_cuda_dispatch.h>
  69. #include <ATen/ops/_foreach_cos_cuda_dispatch.h>
  70. #include <ATen/ops/_foreach_cosh_cuda_dispatch.h>
  71. #include <ATen/ops/_foreach_div_cuda_dispatch.h>
  72. #include <ATen/ops/_foreach_erf_cuda_dispatch.h>
  73. #include <ATen/ops/_foreach_erfc_cuda_dispatch.h>
  74. #include <ATen/ops/_foreach_exp_cuda_dispatch.h>
  75. #include <ATen/ops/_foreach_expm1_cuda_dispatch.h>
  76. #include <ATen/ops/_foreach_floor_cuda_dispatch.h>
  77. #include <ATen/ops/_foreach_frac_cuda_dispatch.h>
  78. #include <ATen/ops/_foreach_lerp_cuda_dispatch.h>
  79. #include <ATen/ops/_foreach_lgamma_cuda_dispatch.h>
  80. #include <ATen/ops/_foreach_log_cuda_dispatch.h>
  81. #include <ATen/ops/_foreach_log10_cuda_dispatch.h>
  82. #include <ATen/ops/_foreach_log1p_cuda_dispatch.h>
  83. #include <ATen/ops/_foreach_log2_cuda_dispatch.h>
  84. #include <ATen/ops/_foreach_maximum_cuda_dispatch.h>
  85. #include <ATen/ops/_foreach_minimum_cuda_dispatch.h>
  86. #include <ATen/ops/_foreach_mul_cuda_dispatch.h>
  87. #include <ATen/ops/_foreach_neg_cuda_dispatch.h>
  88. #include <ATen/ops/_foreach_norm_cuda_dispatch.h>
  89. #include <ATen/ops/_foreach_reciprocal_cuda_dispatch.h>
  90. #include <ATen/ops/_foreach_round_cuda_dispatch.h>
  91. #include <ATen/ops/_foreach_sigmoid_cuda_dispatch.h>
  92. #include <ATen/ops/_foreach_sin_cuda_dispatch.h>
  93. #include <ATen/ops/_foreach_sinh_cuda_dispatch.h>
  94. #include <ATen/ops/_foreach_sqrt_cuda_dispatch.h>
  95. #include <ATen/ops/_foreach_sub_cuda_dispatch.h>
  96. #include <ATen/ops/_foreach_tan_cuda_dispatch.h>
  97. #include <ATen/ops/_foreach_tanh_cuda_dispatch.h>
  98. #include <ATen/ops/_foreach_trunc_cuda_dispatch.h>
  99. #include <ATen/ops/_foreach_zero_cuda_dispatch.h>
  100. #include <ATen/ops/_fused_adam_cuda_dispatch.h>
  101. #include <ATen/ops/_fused_adamw_cuda_dispatch.h>
  102. #include <ATen/ops/_fused_dropout_cuda_dispatch.h>
  103. #include <ATen/ops/_fused_moving_avg_obs_fq_helper_cuda_dispatch.h>
  104. #include <ATen/ops/_fused_sdp_choice_cuda_dispatch.h>
  105. #include <ATen/ops/_index_put_impl_cuda_dispatch.h>
  106. #include <ATen/ops/_linalg_det_cuda_dispatch.h>
  107. #include <ATen/ops/_linalg_eigh_cuda_dispatch.h>
  108. #include <ATen/ops/_linalg_slogdet_cuda_dispatch.h>
  109. #include <ATen/ops/_linalg_solve_ex_cuda_dispatch.h>
  110. #include <ATen/ops/_linalg_svd_cuda_dispatch.h>
  111. #include <ATen/ops/_local_scalar_dense_cuda_dispatch.h>
  112. #include <ATen/ops/_log_softmax_cuda_dispatch.h>
  113. #include <ATen/ops/_log_softmax_backward_data_cuda_dispatch.h>
  114. #include <ATen/ops/_logcumsumexp_cuda_dispatch.h>
  115. #include <ATen/ops/_make_per_channel_quantized_tensor_cuda_dispatch.h>
  116. #include <ATen/ops/_make_per_tensor_quantized_tensor_cuda_dispatch.h>
  117. #include <ATen/ops/_masked_scale_cuda_dispatch.h>
  118. #include <ATen/ops/_masked_softmax_cuda_dispatch.h>
  119. #include <ATen/ops/_masked_softmax_backward_cuda_dispatch.h>
  120. #include <ATen/ops/_native_batch_norm_legit_cuda_dispatch.h>
  121. #include <ATen/ops/_native_decoder_only_multi_head_attention_cuda_dispatch.h>
  122. #include <ATen/ops/_native_multi_head_attention_cuda_dispatch.h>
  123. #include <ATen/ops/_nested_from_padded_cuda_dispatch.h>
  124. #include <ATen/ops/_nested_tensor_from_mask_cuda_dispatch.h>
  125. #include <ATen/ops/_nested_tensor_from_mask_left_aligned_cuda_dispatch.h>
  126. #include <ATen/ops/_nested_view_from_buffer_cuda_dispatch.h>
  127. #include <ATen/ops/_pdist_backward_cuda_dispatch.h>
  128. #include <ATen/ops/_pdist_forward_cuda_dispatch.h>
  129. #include <ATen/ops/_pin_memory_cuda_dispatch.h>
  130. #include <ATen/ops/_prelu_kernel_cuda_dispatch.h>
  131. #include <ATen/ops/_prelu_kernel_backward_cuda_dispatch.h>
  132. #include <ATen/ops/_reshape_alias_cuda_dispatch.h>
  133. #include <ATen/ops/_sample_dirichlet_cuda_dispatch.h>
  134. #include <ATen/ops/_scaled_dot_product_efficient_attention_cuda_dispatch.h>
  135. #include <ATen/ops/_scaled_dot_product_efficient_attention_backward_cuda_dispatch.h>
  136. #include <ATen/ops/_scaled_dot_product_flash_attention_cuda_dispatch.h>
  137. #include <ATen/ops/_scaled_dot_product_flash_attention_backward_cuda_dispatch.h>
  138. #include <ATen/ops/_segment_reduce_backward_cuda_dispatch.h>
  139. #include <ATen/ops/_slow_conv2d_backward_cuda_dispatch.h>
  140. #include <ATen/ops/_slow_conv2d_forward_cuda_dispatch.h>
  141. #include <ATen/ops/_softmax_cuda_dispatch.h>
  142. #include <ATen/ops/_softmax_backward_data_cuda_dispatch.h>
  143. #include <ATen/ops/_standard_gamma_cuda_dispatch.h>
  144. #include <ATen/ops/_standard_gamma_grad_cuda_dispatch.h>
  145. #include <ATen/ops/_thnn_fused_gru_cell_cuda_dispatch.h>
  146. #include <ATen/ops/_thnn_fused_gru_cell_backward_cuda_dispatch.h>
  147. #include <ATen/ops/_thnn_fused_lstm_cell_cuda_dispatch.h>
  148. #include <ATen/ops/_thnn_fused_lstm_cell_backward_impl_cuda_dispatch.h>
  149. #include <ATen/ops/_transform_bias_rescale_qkv_cuda_dispatch.h>
  150. #include <ATen/ops/_transformer_decoder_only_layer_fwd_cuda_dispatch.h>
  151. #include <ATen/ops/_transformer_encoder_layer_fwd_cuda_dispatch.h>
  152. #include <ATen/ops/_triton_multi_head_attention_cuda_dispatch.h>
  153. #include <ATen/ops/_triton_scaled_dot_attention_cuda_dispatch.h>
  154. #include <ATen/ops/_unique_cuda_dispatch.h>
  155. #include <ATen/ops/_unique2_cuda_dispatch.h>
  156. #include <ATen/ops/_upsample_bicubic2d_aa_cuda_dispatch.h>
  157. #include <ATen/ops/_upsample_bicubic2d_aa_backward_cuda_dispatch.h>
  158. #include <ATen/ops/_upsample_bilinear2d_aa_cuda_dispatch.h>
  159. #include <ATen/ops/_upsample_bilinear2d_aa_backward_cuda_dispatch.h>
  160. #include <ATen/ops/_upsample_nearest_exact1d_cuda_dispatch.h>
  161. #include <ATen/ops/_upsample_nearest_exact1d_backward_cuda_dispatch.h>
  162. #include <ATen/ops/_upsample_nearest_exact2d_cuda_dispatch.h>
  163. #include <ATen/ops/_upsample_nearest_exact2d_backward_cuda_dispatch.h>
  164. #include <ATen/ops/_upsample_nearest_exact3d_cuda_dispatch.h>
  165. #include <ATen/ops/_upsample_nearest_exact3d_backward_cuda_dispatch.h>
  166. #include <ATen/ops/_use_cudnn_ctc_loss_cuda_dispatch.h>
  167. #include <ATen/ops/_validate_compressed_sparse_indices_cuda_dispatch.h>
  168. #include <ATen/ops/_weight_norm_interface_cuda_dispatch.h>
  169. #include <ATen/ops/_weight_norm_interface_backward_cuda_dispatch.h>
  170. #include <ATen/ops/abs_cuda_dispatch.h>
  171. #include <ATen/ops/acos_cuda_dispatch.h>
  172. #include <ATen/ops/acosh_cuda_dispatch.h>
  173. #include <ATen/ops/adaptive_avg_pool2d_cuda_dispatch.h>
  174. #include <ATen/ops/adaptive_avg_pool3d_cuda_dispatch.h>
  175. #include <ATen/ops/adaptive_avg_pool3d_backward_cuda_dispatch.h>
  176. #include <ATen/ops/adaptive_max_pool2d_cuda_dispatch.h>
  177. #include <ATen/ops/adaptive_max_pool2d_backward_cuda_dispatch.h>
  178. #include <ATen/ops/adaptive_max_pool3d_cuda_dispatch.h>
  179. #include <ATen/ops/adaptive_max_pool3d_backward_cuda_dispatch.h>
  180. #include <ATen/ops/add_cuda_dispatch.h>
  181. #include <ATen/ops/addbmm_cuda_dispatch.h>
  182. #include <ATen/ops/addcdiv_cuda_dispatch.h>
  183. #include <ATen/ops/addcmul_cuda_dispatch.h>
  184. #include <ATen/ops/addmm_cuda_dispatch.h>
  185. #include <ATen/ops/addmv_cuda_dispatch.h>
  186. #include <ATen/ops/addr_cuda_dispatch.h>
  187. #include <ATen/ops/all_cuda_dispatch.h>
  188. #include <ATen/ops/amax_cuda_dispatch.h>
  189. #include <ATen/ops/amin_cuda_dispatch.h>
  190. #include <ATen/ops/aminmax_cuda_dispatch.h>
  191. #include <ATen/ops/angle_cuda_dispatch.h>
  192. #include <ATen/ops/any_cuda_dispatch.h>
  193. #include <ATen/ops/arange_cuda_dispatch.h>
  194. #include <ATen/ops/argmax_cuda_dispatch.h>
  195. #include <ATen/ops/argmin_cuda_dispatch.h>
  196. #include <ATen/ops/argsort_cuda_dispatch.h>
  197. #include <ATen/ops/as_strided_cuda_dispatch.h>
  198. #include <ATen/ops/asin_cuda_dispatch.h>
  199. #include <ATen/ops/asinh_cuda_dispatch.h>
  200. #include <ATen/ops/atan_cuda_dispatch.h>
  201. #include <ATen/ops/atan2_cuda_dispatch.h>
  202. #include <ATen/ops/atanh_cuda_dispatch.h>
  203. #include <ATen/ops/avg_pool2d_cuda_dispatch.h>
  204. #include <ATen/ops/avg_pool2d_backward_cuda_dispatch.h>
  205. #include <ATen/ops/avg_pool3d_cuda_dispatch.h>
  206. #include <ATen/ops/avg_pool3d_backward_cuda_dispatch.h>
  207. #include <ATen/ops/baddbmm_cuda_dispatch.h>
  208. #include <ATen/ops/batch_norm_backward_elemt_cuda_dispatch.h>
  209. #include <ATen/ops/batch_norm_backward_reduce_cuda_dispatch.h>
  210. #include <ATen/ops/batch_norm_elemt_cuda_dispatch.h>
  211. #include <ATen/ops/batch_norm_gather_stats_cuda_dispatch.h>
  212. #include <ATen/ops/batch_norm_gather_stats_with_counts_cuda_dispatch.h>
  213. #include <ATen/ops/batch_norm_stats_cuda_dispatch.h>
  214. #include <ATen/ops/batch_norm_update_stats_cuda_dispatch.h>
  215. #include <ATen/ops/bernoulli_cuda_dispatch.h>
  216. #include <ATen/ops/binary_cross_entropy_cuda_dispatch.h>
  217. #include <ATen/ops/binary_cross_entropy_backward_cuda_dispatch.h>
  218. #include <ATen/ops/bincount_cuda_dispatch.h>
  219. #include <ATen/ops/binomial_cuda_dispatch.h>
  220. #include <ATen/ops/bitwise_and_cuda_dispatch.h>
  221. #include <ATen/ops/bitwise_left_shift_cuda_dispatch.h>
  222. #include <ATen/ops/bitwise_not_cuda_dispatch.h>
  223. #include <ATen/ops/bitwise_or_cuda_dispatch.h>
  224. #include <ATen/ops/bitwise_right_shift_cuda_dispatch.h>
  225. #include <ATen/ops/bitwise_xor_cuda_dispatch.h>
  226. #include <ATen/ops/bmm_cuda_dispatch.h>
  227. #include <ATen/ops/bucketize_cuda_dispatch.h>
  228. #include <ATen/ops/cat_cuda_dispatch.h>
  229. #include <ATen/ops/cauchy_cuda_dispatch.h>
  230. #include <ATen/ops/ceil_cuda_dispatch.h>
  231. #include <ATen/ops/cholesky_cuda_dispatch.h>
  232. #include <ATen/ops/cholesky_inverse_cuda_dispatch.h>
  233. #include <ATen/ops/clamp_cuda_dispatch.h>
  234. #include <ATen/ops/clamp_max_cuda_dispatch.h>
  235. #include <ATen/ops/clamp_min_cuda_dispatch.h>
  236. #include <ATen/ops/col2im_cuda_dispatch.h>
  237. #include <ATen/ops/complex_cuda_dispatch.h>
  238. #include <ATen/ops/conj_physical_cuda_dispatch.h>
  239. #include <ATen/ops/conv_depthwise3d_cuda_dispatch.h>
  240. #include <ATen/ops/convolution_backward_cuda_dispatch.h>
  241. #include <ATen/ops/copysign_cuda_dispatch.h>
  242. #include <ATen/ops/cos_cuda_dispatch.h>
  243. #include <ATen/ops/cosh_cuda_dispatch.h>
  244. #include <ATen/ops/count_nonzero_cuda_dispatch.h>
  245. #include <ATen/ops/cudnn_affine_grid_generator_cuda_dispatch.h>
  246. #include <ATen/ops/cudnn_affine_grid_generator_backward_cuda_dispatch.h>
  247. #include <ATen/ops/cudnn_batch_norm_cuda_dispatch.h>
  248. #include <ATen/ops/cudnn_batch_norm_backward_cuda_dispatch.h>
  249. #include <ATen/ops/cudnn_convolution_cuda_dispatch.h>
  250. #include <ATen/ops/cudnn_convolution_add_relu_cuda_dispatch.h>
  251. #include <ATen/ops/cudnn_convolution_relu_cuda_dispatch.h>
  252. #include <ATen/ops/cudnn_convolution_transpose_cuda_dispatch.h>
  253. #include <ATen/ops/cudnn_grid_sampler_cuda_dispatch.h>
  254. #include <ATen/ops/cudnn_grid_sampler_backward_cuda_dispatch.h>
  255. #include <ATen/ops/cumprod_cuda_dispatch.h>
  256. #include <ATen/ops/cumsum_cuda_dispatch.h>
  257. #include <ATen/ops/dense_dim_cuda_dispatch.h>
  258. #include <ATen/ops/dequantize_cuda_dispatch.h>
  259. #include <ATen/ops/digamma_cuda_dispatch.h>
  260. #include <ATen/ops/div_cuda_dispatch.h>
  261. #include <ATen/ops/dot_cuda_dispatch.h>
  262. #include <ATen/ops/elu_cuda_dispatch.h>
  263. #include <ATen/ops/elu_backward_cuda_dispatch.h>
  264. #include <ATen/ops/embedding_dense_backward_cuda_dispatch.h>
  265. #include <ATen/ops/embedding_renorm_cuda_dispatch.h>
  266. #include <ATen/ops/empty_cuda_dispatch.h>
  267. #include <ATen/ops/empty_strided_cuda_dispatch.h>
  268. #include <ATen/ops/eq_cuda_dispatch.h>
  269. #include <ATen/ops/equal_cuda_dispatch.h>
  270. #include <ATen/ops/erf_cuda_dispatch.h>
  271. #include <ATen/ops/erfc_cuda_dispatch.h>
  272. #include <ATen/ops/erfinv_cuda_dispatch.h>
  273. #include <ATen/ops/exp_cuda_dispatch.h>
  274. #include <ATen/ops/exp2_cuda_dispatch.h>
  275. #include <ATen/ops/expm1_cuda_dispatch.h>
  276. #include <ATen/ops/exponential_cuda_dispatch.h>
  277. #include <ATen/ops/eye_cuda_dispatch.h>
  278. #include <ATen/ops/fake_quantize_per_channel_affine_cachemask_cuda_dispatch.h>
  279. #include <ATen/ops/fake_quantize_per_tensor_affine_cachemask_cuda_dispatch.h>
  280. #include <ATen/ops/fill_cuda_dispatch.h>
  281. #include <ATen/ops/flip_cuda_dispatch.h>
  282. #include <ATen/ops/floor_cuda_dispatch.h>
  283. #include <ATen/ops/floor_divide_cuda_dispatch.h>
  284. #include <ATen/ops/fmax_cuda_dispatch.h>
  285. #include <ATen/ops/fmin_cuda_dispatch.h>
  286. #include <ATen/ops/fmod_cuda_dispatch.h>
  287. #include <ATen/ops/frac_cuda_dispatch.h>
  288. #include <ATen/ops/fractional_max_pool2d_cuda_dispatch.h>
  289. #include <ATen/ops/fractional_max_pool2d_backward_cuda_dispatch.h>
  290. #include <ATen/ops/fractional_max_pool3d_cuda_dispatch.h>
  291. #include <ATen/ops/fractional_max_pool3d_backward_cuda_dispatch.h>
  292. #include <ATen/ops/frexp_cuda_dispatch.h>
  293. #include <ATen/ops/gather_cuda_dispatch.h>
  294. #include <ATen/ops/gcd_cuda_dispatch.h>
  295. #include <ATen/ops/ge_cuda_dispatch.h>
  296. #include <ATen/ops/gelu_cuda_dispatch.h>
  297. #include <ATen/ops/gelu_backward_cuda_dispatch.h>
  298. #include <ATen/ops/geometric_cuda_dispatch.h>
  299. #include <ATen/ops/geqrf_cuda_dispatch.h>
  300. #include <ATen/ops/glu_cuda_dispatch.h>
  301. #include <ATen/ops/glu_backward_cuda_dispatch.h>
  302. #include <ATen/ops/glu_backward_jvp_cuda_dispatch.h>
  303. #include <ATen/ops/glu_jvp_cuda_dispatch.h>
  304. #include <ATen/ops/grid_sampler_2d_cuda_dispatch.h>
  305. #include <ATen/ops/grid_sampler_2d_backward_cuda_dispatch.h>
  306. #include <ATen/ops/grid_sampler_3d_cuda_dispatch.h>
  307. #include <ATen/ops/grid_sampler_3d_backward_cuda_dispatch.h>
  308. #include <ATen/ops/gt_cuda_dispatch.h>
  309. #include <ATen/ops/hardshrink_cuda_dispatch.h>
  310. #include <ATen/ops/hardshrink_backward_cuda_dispatch.h>
  311. #include <ATen/ops/hardsigmoid_cuda_dispatch.h>
  312. #include <ATen/ops/hardsigmoid_backward_cuda_dispatch.h>
  313. #include <ATen/ops/hardswish_cuda_dispatch.h>
  314. #include <ATen/ops/hardswish_backward_cuda_dispatch.h>
  315. #include <ATen/ops/hardtanh_cuda_dispatch.h>
  316. #include <ATen/ops/hardtanh_backward_cuda_dispatch.h>
  317. #include <ATen/ops/heaviside_cuda_dispatch.h>
  318. #include <ATen/ops/histc_cuda_dispatch.h>
  319. #include <ATen/ops/huber_loss_cuda_dispatch.h>
  320. #include <ATen/ops/huber_loss_backward_cuda_dispatch.h>
  321. #include <ATen/ops/hypot_cuda_dispatch.h>
  322. #include <ATen/ops/i0_cuda_dispatch.h>
  323. #include <ATen/ops/igamma_cuda_dispatch.h>
  324. #include <ATen/ops/igammac_cuda_dispatch.h>
  325. #include <ATen/ops/im2col_cuda_dispatch.h>
  326. #include <ATen/ops/index_cuda_dispatch.h>
  327. #include <ATen/ops/index_add_cuda_dispatch.h>
  328. #include <ATen/ops/index_copy_cuda_dispatch.h>
  329. #include <ATen/ops/index_fill_cuda_dispatch.h>
  330. #include <ATen/ops/index_reduce_cuda_dispatch.h>
  331. #include <ATen/ops/index_select_cuda_dispatch.h>
  332. #include <ATen/ops/is_pinned_cuda_dispatch.h>
  333. #include <ATen/ops/is_set_to_cuda_dispatch.h>
  334. #include <ATen/ops/isin_cuda_dispatch.h>
  335. #include <ATen/ops/isnan_cuda_dispatch.h>
  336. #include <ATen/ops/isneginf_cuda_dispatch.h>
  337. #include <ATen/ops/isposinf_cuda_dispatch.h>
  338. #include <ATen/ops/kthvalue_cuda_dispatch.h>
  339. #include <ATen/ops/lcm_cuda_dispatch.h>
  340. #include <ATen/ops/le_cuda_dispatch.h>
  341. #include <ATen/ops/leaky_relu_cuda_dispatch.h>
  342. #include <ATen/ops/leaky_relu_backward_cuda_dispatch.h>
  343. #include <ATen/ops/lerp_cuda_dispatch.h>
  344. #include <ATen/ops/lgamma_cuda_dispatch.h>
  345. #include <ATen/ops/linalg_cholesky_ex_cuda_dispatch.h>
  346. #include <ATen/ops/linalg_cross_cuda_dispatch.h>
  347. #include <ATen/ops/linalg_eig_cuda_dispatch.h>
  348. #include <ATen/ops/linalg_householder_product_cuda_dispatch.h>
  349. #include <ATen/ops/linalg_inv_ex_cuda_dispatch.h>
  350. #include <ATen/ops/linalg_ldl_factor_ex_cuda_dispatch.h>
  351. #include <ATen/ops/linalg_ldl_solve_cuda_dispatch.h>
  352. #include <ATen/ops/linalg_lstsq_cuda_dispatch.h>
  353. #include <ATen/ops/linalg_lu_cuda_dispatch.h>
  354. #include <ATen/ops/linalg_lu_factor_ex_cuda_dispatch.h>
  355. #include <ATen/ops/linalg_lu_solve_cuda_dispatch.h>
  356. #include <ATen/ops/linalg_matrix_exp_cuda_dispatch.h>
  357. #include <ATen/ops/linalg_qr_cuda_dispatch.h>
  358. #include <ATen/ops/linalg_solve_triangular_cuda_dispatch.h>
  359. #include <ATen/ops/linalg_vector_norm_cuda_dispatch.h>
  360. #include <ATen/ops/linspace_cuda_dispatch.h>
  361. #include <ATen/ops/log_cuda_dispatch.h>
  362. #include <ATen/ops/log10_cuda_dispatch.h>
  363. #include <ATen/ops/log1p_cuda_dispatch.h>
  364. #include <ATen/ops/log2_cuda_dispatch.h>
  365. #include <ATen/ops/log_normal_cuda_dispatch.h>
  366. #include <ATen/ops/log_sigmoid_backward_cuda_dispatch.h>
  367. #include <ATen/ops/log_sigmoid_forward_cuda_dispatch.h>
  368. #include <ATen/ops/logaddexp_cuda_dispatch.h>
  369. #include <ATen/ops/logaddexp2_cuda_dispatch.h>
  370. #include <ATen/ops/logical_and_cuda_dispatch.h>
  371. #include <ATen/ops/logical_not_cuda_dispatch.h>
  372. #include <ATen/ops/logical_or_cuda_dispatch.h>
  373. #include <ATen/ops/logical_xor_cuda_dispatch.h>
  374. #include <ATen/ops/logit_cuda_dispatch.h>
  375. #include <ATen/ops/logit_backward_cuda_dispatch.h>
  376. #include <ATen/ops/logspace_cuda_dispatch.h>
  377. #include <ATen/ops/lshift_cuda_dispatch.h>
  378. #include <ATen/ops/lt_cuda_dispatch.h>
  379. #include <ATen/ops/lu_unpack_cuda_dispatch.h>
  380. #include <ATen/ops/masked_fill_cuda_dispatch.h>
  381. #include <ATen/ops/masked_scatter_cuda_dispatch.h>
  382. #include <ATen/ops/masked_select_cuda_dispatch.h>
  383. #include <ATen/ops/max_cuda_dispatch.h>
  384. #include <ATen/ops/max_pool2d_with_indices_cuda_dispatch.h>
  385. #include <ATen/ops/max_pool2d_with_indices_backward_cuda_dispatch.h>
  386. #include <ATen/ops/max_pool3d_with_indices_cuda_dispatch.h>
  387. #include <ATen/ops/max_pool3d_with_indices_backward_cuda_dispatch.h>
  388. #include <ATen/ops/max_unpool2d_cuda_dispatch.h>
  389. #include <ATen/ops/max_unpool3d_cuda_dispatch.h>
  390. #include <ATen/ops/maximum_cuda_dispatch.h>
  391. #include <ATen/ops/mean_cuda_dispatch.h>
  392. #include <ATen/ops/median_cuda_dispatch.h>
  393. #include <ATen/ops/min_cuda_dispatch.h>
  394. #include <ATen/ops/minimum_cuda_dispatch.h>
  395. #include <ATen/ops/miopen_batch_norm_cuda_dispatch.h>
  396. #include <ATen/ops/miopen_batch_norm_backward_cuda_dispatch.h>
  397. #include <ATen/ops/miopen_convolution_cuda_dispatch.h>
  398. #include <ATen/ops/miopen_convolution_add_relu_cuda_dispatch.h>
  399. #include <ATen/ops/miopen_convolution_relu_cuda_dispatch.h>
  400. #include <ATen/ops/miopen_convolution_transpose_cuda_dispatch.h>
  401. #include <ATen/ops/miopen_depthwise_convolution_cuda_dispatch.h>
  402. #include <ATen/ops/miopen_rnn_cuda_dispatch.h>
  403. #include <ATen/ops/miopen_rnn_backward_cuda_dispatch.h>
  404. #include <ATen/ops/mish_cuda_dispatch.h>
  405. #include <ATen/ops/mish_backward_cuda_dispatch.h>
  406. #include <ATen/ops/mm_cuda_dispatch.h>
  407. #include <ATen/ops/mode_cuda_dispatch.h>
  408. #include <ATen/ops/mse_loss_cuda_dispatch.h>
  409. #include <ATen/ops/mse_loss_backward_cuda_dispatch.h>
  410. #include <ATen/ops/mul_cuda_dispatch.h>
  411. #include <ATen/ops/multi_margin_loss_cuda_dispatch.h>
  412. #include <ATen/ops/multi_margin_loss_backward_cuda_dispatch.h>
  413. #include <ATen/ops/multilabel_margin_loss_backward_cuda_dispatch.h>
  414. #include <ATen/ops/multilabel_margin_loss_forward_cuda_dispatch.h>
  415. #include <ATen/ops/multinomial_cuda_dispatch.h>
  416. #include <ATen/ops/mvlgamma_cuda_dispatch.h>
  417. #include <ATen/ops/nan_to_num_cuda_dispatch.h>
  418. #include <ATen/ops/nanmedian_cuda_dispatch.h>
  419. #include <ATen/ops/nansum_cuda_dispatch.h>
  420. #include <ATen/ops/native_batch_norm_cuda_dispatch.h>
  421. #include <ATen/ops/native_batch_norm_backward_cuda_dispatch.h>
  422. #include <ATen/ops/native_dropout_cuda_dispatch.h>
  423. #include <ATen/ops/native_dropout_backward_cuda_dispatch.h>
  424. #include <ATen/ops/native_group_norm_cuda_dispatch.h>
  425. #include <ATen/ops/native_group_norm_backward_cuda_dispatch.h>
  426. #include <ATen/ops/native_layer_norm_cuda_dispatch.h>
  427. #include <ATen/ops/native_layer_norm_backward_cuda_dispatch.h>
  428. #include <ATen/ops/ne_cuda_dispatch.h>
  429. #include <ATen/ops/neg_cuda_dispatch.h>
  430. #include <ATen/ops/nextafter_cuda_dispatch.h>
  431. #include <ATen/ops/nll_loss2d_backward_cuda_dispatch.h>
  432. #include <ATen/ops/nll_loss2d_forward_cuda_dispatch.h>
  433. #include <ATen/ops/nll_loss_backward_cuda_dispatch.h>
  434. #include <ATen/ops/nll_loss_forward_cuda_dispatch.h>
  435. #include <ATen/ops/nonzero_cuda_dispatch.h>
  436. #include <ATen/ops/norm_cuda_dispatch.h>
  437. #include <ATen/ops/normal_cuda_dispatch.h>
  438. #include <ATen/ops/ormqr_cuda_dispatch.h>
  439. #include <ATen/ops/poisson_cuda_dispatch.h>
  440. #include <ATen/ops/polar_cuda_dispatch.h>
  441. #include <ATen/ops/polygamma_cuda_dispatch.h>
  442. #include <ATen/ops/pow_cuda_dispatch.h>
  443. #include <ATen/ops/prod_cuda_dispatch.h>
  444. #include <ATen/ops/put_cuda_dispatch.h>
  445. #include <ATen/ops/quantize_per_channel_cuda_dispatch.h>
  446. #include <ATen/ops/quantize_per_tensor_cuda_dispatch.h>
  447. #include <ATen/ops/quantize_per_tensor_dynamic_cuda_dispatch.h>
  448. #include <ATen/ops/random_cuda_dispatch.h>
  449. #include <ATen/ops/randperm_cuda_dispatch.h>
  450. #include <ATen/ops/range_cuda_dispatch.h>
  451. #include <ATen/ops/reciprocal_cuda_dispatch.h>
  452. #include <ATen/ops/record_stream_cuda_dispatch.h>
  453. #include <ATen/ops/reflection_pad1d_cuda_dispatch.h>
  454. #include <ATen/ops/reflection_pad1d_backward_cuda_dispatch.h>
  455. #include <ATen/ops/reflection_pad2d_cuda_dispatch.h>
  456. #include <ATen/ops/reflection_pad2d_backward_cuda_dispatch.h>
  457. #include <ATen/ops/reflection_pad3d_cuda_dispatch.h>
  458. #include <ATen/ops/reflection_pad3d_backward_cuda_dispatch.h>
  459. #include <ATen/ops/relu_cuda_dispatch.h>
  460. #include <ATen/ops/remainder_cuda_dispatch.h>
  461. #include <ATen/ops/renorm_cuda_dispatch.h>
  462. #include <ATen/ops/repeat_interleave_cuda_dispatch.h>
  463. #include <ATen/ops/replication_pad1d_cuda_dispatch.h>
  464. #include <ATen/ops/replication_pad1d_backward_cuda_dispatch.h>
  465. #include <ATen/ops/replication_pad2d_cuda_dispatch.h>
  466. #include <ATen/ops/replication_pad2d_backward_cuda_dispatch.h>
  467. #include <ATen/ops/replication_pad3d_cuda_dispatch.h>
  468. #include <ATen/ops/replication_pad3d_backward_cuda_dispatch.h>
  469. #include <ATen/ops/resize_cuda_dispatch.h>
  470. #include <ATen/ops/roll_cuda_dispatch.h>
  471. #include <ATen/ops/round_cuda_dispatch.h>
  472. #include <ATen/ops/rrelu_with_noise_cuda_dispatch.h>
  473. #include <ATen/ops/rshift_cuda_dispatch.h>
  474. #include <ATen/ops/rsqrt_cuda_dispatch.h>
  475. #include <ATen/ops/rsub_cuda_dispatch.h>
  476. #include <ATen/ops/scatter_cuda_dispatch.h>
  477. #include <ATen/ops/scatter_add_cuda_dispatch.h>
  478. #include <ATen/ops/scatter_reduce_cuda_dispatch.h>
  479. #include <ATen/ops/searchsorted_cuda_dispatch.h>
  480. #include <ATen/ops/segment_reduce_cuda_dispatch.h>
  481. #include <ATen/ops/set_cuda_dispatch.h>
  482. #include <ATen/ops/sgn_cuda_dispatch.h>
  483. #include <ATen/ops/sigmoid_cuda_dispatch.h>
  484. #include <ATen/ops/sigmoid_backward_cuda_dispatch.h>
  485. #include <ATen/ops/sign_cuda_dispatch.h>
  486. #include <ATen/ops/signbit_cuda_dispatch.h>
  487. #include <ATen/ops/silu_cuda_dispatch.h>
  488. #include <ATen/ops/silu_backward_cuda_dispatch.h>
  489. #include <ATen/ops/sin_cuda_dispatch.h>
  490. #include <ATen/ops/sinc_cuda_dispatch.h>
  491. #include <ATen/ops/sinh_cuda_dispatch.h>
  492. #include <ATen/ops/slow_conv_dilated2d_cuda_dispatch.h>
  493. #include <ATen/ops/slow_conv_dilated3d_cuda_dispatch.h>
  494. #include <ATen/ops/slow_conv_transpose2d_cuda_dispatch.h>
  495. #include <ATen/ops/slow_conv_transpose3d_cuda_dispatch.h>
  496. #include <ATen/ops/smooth_l1_loss_cuda_dispatch.h>
  497. #include <ATen/ops/smooth_l1_loss_backward_cuda_dispatch.h>
  498. #include <ATen/ops/softplus_cuda_dispatch.h>
  499. #include <ATen/ops/softplus_backward_cuda_dispatch.h>
  500. #include <ATen/ops/softshrink_cuda_dispatch.h>
  501. #include <ATen/ops/softshrink_backward_cuda_dispatch.h>
  502. #include <ATen/ops/sort_cuda_dispatch.h>
  503. #include <ATen/ops/sparse_dim_cuda_dispatch.h>
  504. #include <ATen/ops/special_airy_ai_cuda_dispatch.h>
  505. #include <ATen/ops/special_bessel_j0_cuda_dispatch.h>
  506. #include <ATen/ops/special_bessel_j1_cuda_dispatch.h>
  507. #include <ATen/ops/special_bessel_y0_cuda_dispatch.h>
  508. #include <ATen/ops/special_bessel_y1_cuda_dispatch.h>
  509. #include <ATen/ops/special_chebyshev_polynomial_t_cuda_dispatch.h>
  510. #include <ATen/ops/special_chebyshev_polynomial_u_cuda_dispatch.h>
  511. #include <ATen/ops/special_chebyshev_polynomial_v_cuda_dispatch.h>
  512. #include <ATen/ops/special_chebyshev_polynomial_w_cuda_dispatch.h>
  513. #include <ATen/ops/special_entr_cuda_dispatch.h>
  514. #include <ATen/ops/special_erfcx_cuda_dispatch.h>
  515. #include <ATen/ops/special_hermite_polynomial_h_cuda_dispatch.h>
  516. #include <ATen/ops/special_hermite_polynomial_he_cuda_dispatch.h>
  517. #include <ATen/ops/special_i0e_cuda_dispatch.h>
  518. #include <ATen/ops/special_i1_cuda_dispatch.h>
  519. #include <ATen/ops/special_i1e_cuda_dispatch.h>
  520. #include <ATen/ops/special_laguerre_polynomial_l_cuda_dispatch.h>
  521. #include <ATen/ops/special_legendre_polynomial_p_cuda_dispatch.h>
  522. #include <ATen/ops/special_log_ndtr_cuda_dispatch.h>
  523. #include <ATen/ops/special_modified_bessel_i0_cuda_dispatch.h>
  524. #include <ATen/ops/special_modified_bessel_i1_cuda_dispatch.h>
  525. #include <ATen/ops/special_modified_bessel_k0_cuda_dispatch.h>
  526. #include <ATen/ops/special_modified_bessel_k1_cuda_dispatch.h>
  527. #include <ATen/ops/special_ndtri_cuda_dispatch.h>
  528. #include <ATen/ops/special_scaled_modified_bessel_k0_cuda_dispatch.h>
  529. #include <ATen/ops/special_scaled_modified_bessel_k1_cuda_dispatch.h>
  530. #include <ATen/ops/special_shifted_chebyshev_polynomial_t_cuda_dispatch.h>
  531. #include <ATen/ops/special_shifted_chebyshev_polynomial_u_cuda_dispatch.h>
  532. #include <ATen/ops/special_shifted_chebyshev_polynomial_v_cuda_dispatch.h>
  533. #include <ATen/ops/special_shifted_chebyshev_polynomial_w_cuda_dispatch.h>
  534. #include <ATen/ops/special_spherical_bessel_j0_cuda_dispatch.h>
  535. #include <ATen/ops/special_xlog1py_cuda_dispatch.h>
  536. #include <ATen/ops/special_zeta_cuda_dispatch.h>
  537. #include <ATen/ops/sqrt_cuda_dispatch.h>
  538. #include <ATen/ops/sspaddmm_cuda_dispatch.h>
  539. #include <ATen/ops/std_cuda_dispatch.h>
  540. #include <ATen/ops/std_mean_cuda_dispatch.h>
  541. #include <ATen/ops/sub_cuda_dispatch.h>
  542. #include <ATen/ops/sum_cuda_dispatch.h>
  543. #include <ATen/ops/take_cuda_dispatch.h>
  544. #include <ATen/ops/tan_cuda_dispatch.h>
  545. #include <ATen/ops/tanh_cuda_dispatch.h>
  546. #include <ATen/ops/tanh_backward_cuda_dispatch.h>
  547. #include <ATen/ops/tensordot_cuda_dispatch.h>
  548. #include <ATen/ops/threshold_cuda_dispatch.h>
  549. #include <ATen/ops/threshold_backward_cuda_dispatch.h>
  550. #include <ATen/ops/to_sparse_cuda_dispatch.h>
  551. #include <ATen/ops/to_sparse_bsc_cuda_dispatch.h>
  552. #include <ATen/ops/to_sparse_bsr_cuda_dispatch.h>
  553. #include <ATen/ops/to_sparse_csc_cuda_dispatch.h>
  554. #include <ATen/ops/to_sparse_csr_cuda_dispatch.h>
  555. #include <ATen/ops/topk_cuda_dispatch.h>
  556. #include <ATen/ops/trace_cuda_dispatch.h>
  557. #include <ATen/ops/triangular_solve_cuda_dispatch.h>
  558. #include <ATen/ops/tril_cuda_dispatch.h>
  559. #include <ATen/ops/tril_indices_cuda_dispatch.h>
  560. #include <ATen/ops/triu_cuda_dispatch.h>
  561. #include <ATen/ops/triu_indices_cuda_dispatch.h>
  562. #include <ATen/ops/trunc_cuda_dispatch.h>
  563. #include <ATen/ops/unfold_cuda_dispatch.h>
  564. #include <ATen/ops/unfold_backward_cuda_dispatch.h>
  565. #include <ATen/ops/uniform_cuda_dispatch.h>
  566. #include <ATen/ops/unique_consecutive_cuda_dispatch.h>
  567. #include <ATen/ops/unique_dim_cuda_dispatch.h>
  568. #include <ATen/ops/unique_dim_consecutive_cuda_dispatch.h>
  569. #include <ATen/ops/upsample_bicubic2d_cuda_dispatch.h>
  570. #include <ATen/ops/upsample_bicubic2d_backward_cuda_dispatch.h>
  571. #include <ATen/ops/upsample_bilinear2d_cuda_dispatch.h>
  572. #include <ATen/ops/upsample_bilinear2d_backward_cuda_dispatch.h>
  573. #include <ATen/ops/upsample_linear1d_cuda_dispatch.h>
  574. #include <ATen/ops/upsample_linear1d_backward_cuda_dispatch.h>
  575. #include <ATen/ops/upsample_nearest1d_cuda_dispatch.h>
  576. #include <ATen/ops/upsample_nearest1d_backward_cuda_dispatch.h>
  577. #include <ATen/ops/upsample_nearest2d_cuda_dispatch.h>
  578. #include <ATen/ops/upsample_nearest2d_backward_cuda_dispatch.h>
  579. #include <ATen/ops/upsample_nearest3d_cuda_dispatch.h>
  580. #include <ATen/ops/upsample_nearest3d_backward_cuda_dispatch.h>
  581. #include <ATen/ops/upsample_trilinear3d_cuda_dispatch.h>
  582. #include <ATen/ops/upsample_trilinear3d_backward_cuda_dispatch.h>
  583. #include <ATen/ops/var_cuda_dispatch.h>
  584. #include <ATen/ops/var_mean_cuda_dispatch.h>
  585. #include <ATen/ops/vdot_cuda_dispatch.h>
  586. #include <ATen/ops/view_cuda_dispatch.h>
  587. #include <ATen/ops/view_as_complex_cuda_dispatch.h>
  588. #include <ATen/ops/view_as_real_cuda_dispatch.h>
  589. #include <ATen/ops/where_cuda_dispatch.h>
  590. #include <ATen/ops/xlogy_cuda_dispatch.h>
  591. #include <ATen/ops/zero_cuda_dispatch.h>