backward.yaml 66.6 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12
- backward_api : abs_double_grad
  forward : abs_grad (Tensor x, Tensor grad_out) -> Tensor(grad_x)
  args : (Tensor x, Tensor grad_x_grad)
  output : Tensor(grad_out_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : abs_double_grad
  data_transform:
    skip_transform : grad_x_grad

13 14 15 16
- backward_api : abs_grad
  forward : abs (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
17
  infer_meta :
18
    func : UnchangedInferMeta
19
    param : [x]
20
  kernel :
21
    func : abs_grad
22 23
  data_transform:
    skip_transform : out_grad
24
  backward : abs_double_grad
25

26 27 28 29
- backward_api : acos_grad
  forward : acos (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
30
  infer_meta :
31 32
    func : UnchangedInferMeta
    param : [x]
33
  kernel :
34
    func : acos_grad
35

36 37 38
- backward_api : acosh_grad
  forward : acosh (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
39
  output : Tensor(x_grad)
40 41 42 43 44
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : acosh_grad
45

46 47 48 49 50 51 52 53 54 55 56 57
- backward_api : add_double_grad
  forward : add_grad (Tensor x, Tensor y, Tensor grad_out, int axis = -1) -> Tensor(grad_x), Tensor(grad_y)
  args : (Tensor y, Tensor grad_out, Tensor grad_x_grad, Tensor grad_y_grad, int axis = -1)
  output : Tensor(grad_out_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [grad_out]
  kernel :
    func : add_double_grad
  optional : grad_x_grad, grad_y_grad
  backward : add_triple_grad

H
hong 已提交
58 59
- backward_api : add_grad
  forward : add (Tensor x, Tensor y) -> Tensor(out)
H
hong 已提交
60
  args : (Tensor x, Tensor y, Tensor out_grad, int axis = -1)
H
hong 已提交
61 62 63 64 65 66
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : add_grad
67
  no_need_buffer : x, y
68
  backward : add_double_grad
H
hong 已提交
69

70 71 72
- backward_api : add_n_grad
  forward : add_n (Tensor[] x) -> Tensor(out)
  args : (Tensor[] x, Tensor out_grad)
73
  output : Tensor[](x_grad){x.size()}
74 75 76
  invoke : add_n_grad_impl(x, out_grad)
  no_need_buffer : x

77 78 79 80 81 82 83 84 85 86
- backward_api : add_triple_grad
  forward : add_double_grad (Tensor y, Tensor grad_out, Tensor grad_grad_x, Tensor grad_grad_y, int axis = -1) -> Tensor(grad_grad_out)
  args : (Tensor grad_grad_x, Tensor grad_grad_y, Tensor grad_grad_out_grad, int axis = -1)
  output : Tensor(grad_grad_x_grad), Tensor(grad_grad_y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [grad_grad_x, grad_grad_y]
  kernel :
    func : add_triple_grad

87
- backward_api : addmm_grad
H
hong 已提交
88
  forward : addmm (Tensor input, Tensor x, Tensor y, float alpha, float beta) -> Tensor(out)
89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105
  args : (Tensor input, Tensor x, Tensor y, Tensor out_grad, float alpha, float beta)
  output : Tensor(input_grad), Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [input, x, y]
  kernel :
    func : addmm_grad

- backward_api : argsort_grad
  forward : argsort (Tensor x, int axis, bool descending) -> Tensor(out), Tensor(indices)
  args : (Tensor indices, Tensor x, Tensor out_grad, int axis, bool descending)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : argsort_grad
H
hong 已提交
106
  no_need_buffer : x
107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127

- backward_api : asin_grad
  forward : asin (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : asin_grad

- backward_api : asinh_grad
  forward : asinh (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : asinh_grad

C
chentianyu03 已提交
128 129 130 131 132 133 134 135
- backward_api : assign_grad
  forward : assign (Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
136
    func : assign
C
chentianyu03 已提交
137

138
- backward_api : atan2_grad
139
  forward : atan2 (Tensor x, Tensor y) -> Tensor(out)
140
  args : (Tensor x, Tensor y, Tensor out_grad)
H
hong 已提交
141 142 143 144 145
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
146
    func : atan2_grad
H
hong 已提交
147

148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167
- backward_api : atan_grad
  forward : atan (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : atan_grad

- backward_api : atanh_grad
  forward : atanh (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : atanh_grad

168 169 170 171 172 173 174 175 176 177 178 179
- backward_api : batch_norm_double_grad
  forward : batch_norm_grad (Tensor x, Tensor scale, Tensor bias, Tensor out_mean, Tensor out_variance, Tensor saved_mean, Tensor saved_variance, Tensor reserve_space, Tensor grad_out, float momentum, float epsilon, str data_layout, bool is_test, bool use_global_stats, bool trainable_statistics, bool fuse_with_relu) -> Tensor(grad_x), Tensor(grad_scale), Tensor(grad_bias)
  args : (Tensor x, Tensor scale, Tensor out_mean, Tensor out_variance, Tensor saved_mean, Tensor saved_variance, Tensor grad_out,  Tensor grad_x_grad, Tensor grad_scale_grad, Tensor grad_bias_grad, float momentum, float epsilon, str data_layout, bool is_test, bool use_global_stats, bool trainable_statistics, bool fuse_with_relu)
  output : Tensor(x_grad), Tensor(scale_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x, scale, x]
  kernel :
    func : batch_norm_grad_grad
    data_type : x
  optional : out_mean, out_variance

H
hong 已提交
180 181 182 183 184 185 186 187 188 189 190
- backward_api : batch_norm_grad
  forward : batch_norm (Tensor x, Tensor scale, Tensor bias, Tensor mean, Tensor variance, float momentum, float epsilon, str data_layout, bool is_test, bool use_global_stats, bool trainable_statistics, bool fuse_with_relu) -> Tensor(out), Tensor(mean_out), Tensor(variance_out), Tensor(saved_mean), Tensor(saved_variance), Tensor(reserve_space)
  args : (Tensor x, Tensor scale, Tensor bias, Tensor mean_out, Tensor variance_out, Tensor saved_mean, Tensor saved_variance, Tensor reserve_space, Tensor out_grad, float momentum, float epsilon, str data_layout, bool is_test, bool use_global_stats, bool trainable_statistics, bool fuse_with_relu)
  output : Tensor(x_grad), Tensor(scale_grad), Tensor(bias_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x, scale, bias]
  kernel :
    func : batch_norm_grad
    data_type : out_grad
  optional : mean_out, variance_out, reserve_space
191
  backward : batch_norm_double_grad
H
hong 已提交
192

193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223
- backward_api : bce_loss_grad
  forward : bce_loss (Tensor input, Tensor label) -> Tensor(out)
  args : (Tensor input, Tensor label, Tensor out_grad)
  output : Tensor(input_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [input]
  kernel :
    func : bce_loss_grad

- backward_api : brelu_grad
  forward : brelu (Tensor x, float t_min, float t_max) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float t_min, float t_max)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : brelu_grad

- backward_api : cast_grad
  forward : cast (Tensor x, DataType out_dtype) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : cast_grad
    data_type : out_grad

224 225 226 227 228 229 230 231 232 233
- backward_api : ceil_grad
  forward : ceil(Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [out_grad]
  kernel :
    func : ceil_grad

234 235 236 237 238 239 240 241 242 243 244
- backward_api : cholesky_grad
  forward : cholesky (Tensor x, bool upper) -> Tensor(out)
  args : (Tensor out, Tensor out_grad, bool upper)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : cholesky_grad

- backward_api : cholesky_solve_grad
245
  forward : cholesky_solve (Tensor x, Tensor y, bool upper) -> Tensor(out)
246
  args : (Tensor x, Tensor y, Tensor out, Tensor out_grad, bool upper)
H
hong 已提交
247 248 249 250 251
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
252 253
    func : cholesky_solve_grad

C
chentianyu03 已提交
254 255 256 257 258 259 260 261 262 263
- backward_api : clip_grad
  forward : clip (Tensor x, Scalar min, Scalar max) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, Scalar min = 0., Scalar max = 0.)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : clip_grad

264 265 266
- backward_api : concat_grad
  forward : concat (Tensor[] x, Scalar axis) -> Tensor(out)
  args : (Tensor[] x, Tensor out_grad, Scalar axis = 0)
267 268 269 270 271 272
  output : Tensor[](x_grad){x.size()}
  infer_meta :
    func : UnchangedMultiInferMeta
    param : [x]
  kernel :
    func : concat_grad
H
hong 已提交
273
  no_need_buffer : x
274

H
hong 已提交
275 276 277 278 279 280 281 282 283 284
- backward_api : conj_grad
  forward : conj (Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [out_grad]
  kernel :
    func : conj

H
hong 已提交
285 286 287 288 289
- backward_api : conv2d_grad
  forward : conv2d (Tensor input, Tensor filter, int[] strides, int[] paddings, str paddding_algorithm, int groups, int[] dilations, str data_format, bool use_addto, int workspace_size_MB, bool exhaustive_search) -> Tensor(out)
  args : (Tensor input, Tensor filter, Tensor out_grad,  int[] strides, int[] paddings, str paddding_algorithm, int groups, int[] dilations, str data_format, bool use_addto, int workspace_size_MB, bool exhaustive_search)
  output : Tensor(input_grad), Tensor(filter_grad)
  invoke : conv2d_grad_impl(input, filter, out_grad,  strides, paddings, paddding_algorithm, groups, dilations, data_format, use_addto, workspace_size_MB, exhaustive_search)
290 291 292 293 294 295 296 297 298 299 300
  backward : conv2d_grad_grad

- backward_api : conv2d_grad_grad
  forward : conv2d_grad (Tensor input, Tensor filter, Tensor grad_out,  int[] strides, int[] paddings, str paddding_algorithm, int groups, int[] dilations, str data_format, bool use_addto, int workspace_size_MB, bool exhaustive_search) -> Tensor(grad_input), Tensor(grad_filter)
  args : (Tensor input, Tensor filter, Tensor grad_out, Tensor grad_input_grad, Tensor grad_filter_grad, int[] strides, int[] paddings, str paddding_algorithm, int groups, int[] dilations, str data_format, bool use_addto, int workspace_size_MB, bool exhaustive_search)
  output : Tensor(input_grad), Tensor(filter_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param: [input, filter, grad_out]
  kernel :
    func : conv2d_grad_grad
301
    use_gpudnn : true
302
  optional : grad_input_grad, grad_filter_grad
H
hong 已提交
303

C
chentianyu03 已提交
304 305 306 307 308 309 310 311 312 313
- backward_api : conv2d_transpose_double_grad
  forward : conv2d_transpose_grad(Tensor x, Tensor filter, Tensor grad_out, int[] strides, int[] paddings, int[] output_padding, int[] output_size, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(grad_x), Tensor(grad_filter)
  args : (Tensor x, Tensor filter, Tensor grad_out, Tensor grad_x_grad, Tensor grad_filter_grad, int[] strides, int[] paddings, int[] output_padding, int[] output_size, str padding_algorithm, int groups, int[] dilations, str data_format)
  output : Tensor(x_grad), Tensor(filter_grad), Tensor(grad_out_grad)
  infer_meta :
    func : Conv2dTransposeDoubleGradInferMeta
  kernel :
    func : conv2d_transpose_grad_grad
    use_gpudnn : true

F
From00 已提交
314 315 316 317 318 319
- backward_api : conv2d_transpose_grad
  forward : conv2d_transpose(Tensor x, Tensor filter, int[] strides, int[] paddings, int[] output_padding, int[] output_size, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(out)
  args : (Tensor x, Tensor filter, Tensor out_grad, int[] strides, int[] paddings, int[] output_padding, int[] output_size, str padding_algorithm, int groups, int[] dilations, str data_format)
  output : Tensor(x_grad), Tensor(filter_grad)
  infer_meta :
    func : ConvTransposeGradInferMeta
320
  kernel :
F
From00 已提交
321
    func : conv2d_transpose_grad
322
    use_gpudnn : true
C
chentianyu03 已提交
323
  backward : conv2d_transpose_double_grad
F
From00 已提交
324 325 326 327 328 329 330 331 332

- backward_api : conv3d_transpose_grad
  forward : conv3d_transpose(Tensor x, Tensor filter, int[] strides, int[] paddings, int[] output_padding, int[] output_size, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(out)
  args : (Tensor x, Tensor filter, Tensor out_grad, int[] strides, int[] paddings, int[] output_padding, int[] output_size, str padding_algorithm, int groups, int[] dilations, str data_format)
  output : Tensor(x_grad), Tensor(filter_grad)
  infer_meta :
    func : ConvTransposeGradInferMeta
  kernel :
    func : conv3d_transpose_grad
333
    use_gpudnn : true
F
From00 已提交
334

335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354
- backward_api : cos_grad
  forward : cos (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : cos_grad

- backward_api : cosh_grad
  forward : cosh (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : cosh_grad

355 356 357 358 359 360 361 362 363 364
- backward_api : cross_entropy_with_softmax_grad
  forward : cross_entropy_with_softmax (Tensor input, Tensor label, bool soft_label, bool use_softmax, bool numeric_stable_mode, int ignore_index, int axis) -> Tensor(softmax), Tensor(loss)
  args : (Tensor label, Tensor softmax, Tensor loss_grad, bool soft_label, bool use_softmax, bool numeric_stable_mode, int ignore_index, int axis)
  output : Tensor(input_grad)
  infer_meta :
    func : CrossEntropyWithSoftmaxGradInferMeta
  kernel :
    func : cross_entropy_with_softmax_grad
    data_type : softmax

365 366 367 368 369 370 371 372 373 374
- backward_api : cross_grad
  forward : cross (Tensor x, Tensor y, int axis = 9) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : cross_grad

375 376 377 378 379 380 381 382 383 384
- backward_api : cumprod_grad
  forward : cumprod (Tensor x, int dim) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int dim)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : cumprod_grad

385 386 387 388 389 390 391 392 393
- backward_api : cumsum_grad
  forward : cumsum(Tensor x, int axis, bool flatten, bool exclusive, bool reverse) -> Tensor(out)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  args : (Tensor out_grad, int axis, bool flatten, bool exclusive, bool reverse)
  output : Tensor(x_grad)
  invoke : cumsum(out_grad, axis, flatten, exclusive, !reverse)

394 395 396 397 398 399 400 401
- backward_api : deformable_conv_grad
  forward : deformable_conv(Tensor x, Tensor offset, Tensor filter, Tensor mask, int[] strides, int[] paddings, int[] dilations, int deformable_groups, int groups, int im2col_step) -> Tensor(out)
  args : (Tensor x, Tensor offset, Tensor filter, Tensor mask, Tensor out_grad, int[] strides, int[] paddings, int[] dilations, int deformable_groups, int groups, int im2col_step)
  output : Tensor(x_grad), Tensor(offset_grad), Tensor(filter_grad), Tensor(mask_grad)
  infer_meta :
    func : DeformableConvGradInferMeta
  kernel :
    func : deformable_conv_grad
402
    data_type : x
403 404
  optional : mask

F
From00 已提交
405 406 407 408 409 410 411 412 413
- backward_api : depthwise_conv2d_transpose_grad
  forward : depthwise_conv2d_transpose(Tensor x, Tensor filter, int[] strides, int[] paddings, int[] output_padding, int[] output_size, str padding_algorithm, int groups, int[] dilations, str data_format) -> Tensor(out)
  args : (Tensor x, Tensor filter, Tensor out_grad, int[] strides, int[] paddings, int[] output_padding, int[] output_size, str padding_algorithm, int groups, int[] dilations, str data_format)
  output : Tensor(x_grad), Tensor(filter_grad)
  infer_meta :
    func : ConvTransposeGradInferMeta
  kernel :
    func : depthwise_conv2d_transpose_grad

C
chentianyu03 已提交
414 415 416 417 418 419 420 421
- backward_api : det_grad
  forward : det (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
422
    func : determinant_grad
C
chentianyu03 已提交
423

424 425 426 427 428 429 430 431 432
- backward_api : diagonal_grad
  forward : diagonal (Tensor x, int offset, int axis1, int axis2) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int offset = 0, int axis1 = 0, int axis2 = 1)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : diagonal_grad
H
hong 已提交
433
  no_need_buffer : x
434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453

- backward_api : digamma_grad
  forward : digamma (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : digamma_grad

- backward_api : dist_grad
  forward : dist (Tensor x, Tensor y, float p) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out, Tensor out_grad, float p)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : dist_grad
H
hong 已提交
454

455 456 457 458 459 460 461 462 463 464 465 466
- backward_api : divide_double_grad
  forward : divide_grad (Tensor x, Tensor y, Tensor out, Tensor grad_out, int axis = -1) -> Tensor(grad_x), Tensor(grad_y)
  args : (Tensor y, Tensor out, Tensor grad_x, Tensor grad_x_grad, Tensor grad_y_grad, int axis = -1)
  output : Tensor(y_grad), Tensor(out_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [y, grad_x, grad_x]
  kernel :
    func : divide_double_grad
    data_type : out
  optional : grad_x_grad, grad_y_grad

H
hong 已提交
467 468
- backward_api : divide_grad
  forward : divide (Tensor x, Tensor y) -> Tensor(out)
0
0x45f 已提交
469
  args : (Tensor x, Tensor y, Tensor out, Tensor out_grad, int axis = -1)
H
hong 已提交
470 471 472 473 474 475
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : divide_grad
476
  backward : divide_double_grad
H
hong 已提交
477

H
hong 已提交
478 479 480 481 482 483 484 485 486 487 488
- backward_api : dropout_grad
  forward : dropout (Tensor x, Tensor seed_tensor, float p, bool is_test, str mode, int seed, bool fix_seed) -> Tensor(out), Tensor(mask)
  args : (Tensor mask, Tensor out_grad, float p, bool is_test, str mode)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : dropout_grad
  optional : seed_tensor

489 490 491 492 493 494 495 496 497
- backward_api : eigh_grad
  forward : eigh (Tensor x, str uplo) -> Tensor(out_w), Tensor(out_v)
  args : (Tensor out_w, Tensor out_v, Tensor out_w_grad, Tensor out_v_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_v]
  kernel :
    func : eigh_grad
498 499 500
    data_type : out_v
  data_transform:
    skip_transform : out_w, out_w_grad
H
hong 已提交
501

502 503 504 505 506 507 508 509 510 511
- backward_api : einsum_grad
  forward : einsum (Tensor[] x, str equation) -> Tensor(out)
  args : (Tensor[] x, Tensor out_grad, str equation)
  output : Tensor[](x_grad){x.size()}
  infer_meta :
    func : UnchangedMultiInferMeta
    param : [x]
  kernel :
    func : einsum_grad

512 513 514 515 516 517 518 519 520 521
- backward_api : elementwise_pow_grad
  forward : elementwise_pow(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis=-1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : elementwise_pow_grad

522 523 524 525 526 527 528 529 530 531
- backward_api : elu_double_grad
  forward : elu_grad (Tensor x, Tensor out, Tensor grad_out, float alpha)-> Tensor(grad_x)
  args : (Tensor x, Tensor grad_out, Tensor grad_x_grad, float alpha)
  output : Tensor(x_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, x]
  kernel :
    func : elu_double_grad

532 533 534 535 536 537 538 539 540
- backward_api : elu_grad
  forward : elu (Tensor x, float alpha) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, float alpha)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : elu_grad
541
  backward : elu_double_grad
542 543 544 545 546 547 548 549 550 551 552 553 554

- backward_api : erf_grad
  forward : erf (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : erf_grad
    data_type : out_grad

- backward_api : erfinv_grad
555
  forward : erfinv (Tensor x) -> Tensor(out)
556 557 558 559 560 561 562 563
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : erfinv_grad

C
chentianyu03 已提交
564 565 566 567 568 569 570 571 572 573
- backward_api : exp_grad
  forward : exp (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : exp_grad

H
hong 已提交
574 575 576 577 578 579 580 581 582
- backward_api : expand_as_grad
  forward : expand_as (Tensor x, Tensor y, int[] target_shape) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int[] target_shape)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : expand_as_grad
H
hong 已提交
583
  no_need_buffer : x
584

H
hong 已提交
585 586 587 588 589 590 591 592 593
- backward_api : expand_grad
  forward : expand (Tensor x, IntArray shape) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray shape)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : expand_grad
W
wanghuancoder 已提交
594
  no_need_buffer : x
H
hong 已提交
595

596 597 598 599 600 601 602 603 604 605
- backward_api : expm1_grad
  forward : expm1 (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : expm1_grad

606 607 608 609 610 611 612 613 614 615 616 617
- backward_api : flatten_grad
  forward : flatten(Tensor x, int start_axis, int stop_axis) -> Tensor(out), Tensor(xshape)
  args : (Tensor xshape, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func :  KernelWithXShapeInferMeta
    param : [xshape]
  kernel :
    func : flatten_grad
    data_type: out_grad
    backend: out_grad
    layout: out_grad
H
hong 已提交
618
  no_need_buffer : x
619

H
hong 已提交
620 621 622 623 624 625 626 627 628 629
- backward_api : flip_grad
  forward : flip (Tensor x, int[] axis) -> Tensor(out)
  args : (Tensor out_grad, int[] axis)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [out_grad]
  kernel :
    func : flip

630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659
- backward_api : floor_grad
  forward : floor(Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [out_grad]
  kernel :
    func : floor_grad

- backward_api : fmax_grad
  forward : fmax(Tensor x, Tensor y, int axis) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : fmax_grad

- backward_api : fmin_grad
  forward : fmin(Tensor x, Tensor y, int axis) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : fmin_grad

F
From00 已提交
660 661 662 663 664 665 666 667 668 669
- backward_api : frobenius_norm_grad
  forward : frobenius_norm(Tensor x, int64_t[] axis,  bool keep_dim,  bool reduce_all) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int64_t[] axis,  bool keep_dim,  bool reduce_all)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : frobenius_norm_grad

670 671 672 673 674 675 676 677 678 679
- backward_api : gather_grad
  forward : gather(Tensor x, Tensor index, Scalar axis=0) -> Tensor(out)
  args : (Tensor x, Tensor index, Tensor out_grad, Scalar axis=0, bool overwrite=false)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    data_type: x
    func : gather_grad
H
hong 已提交
680
  no_need_buffer : x
681

682 683 684 685 686 687 688 689 690
- backward_api : gather_nd_grad
  forward : gather_nd (Tensor x, Tensor index) -> Tensor(out)
  args : (Tensor x, Tensor index, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : gather_nd_grad
H
hong 已提交
691
  no_need_buffer : x
692

693 694 695 696 697 698 699 700 701 702
- backward_api : gelu_grad
  forward : gelu(Tensor x,  bool approximate) -> Tensor(out)
  args : (Tensor x, Tensor out_grad,  bool approximate)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : gelu_grad

703 704 705 706 707 708 709 710 711
- backward_api : graph_send_recv_grad
  forward : graph_send_recv (Tensor x, Tensor src_index, Tensor dst_index, str pool_type = "SUM", int64_t out_size = 0) -> Tensor(out), Tensor(dst_count)
  args : (Tensor x, Tensor src_index, Tensor dst_index, Tensor out, Tensor dst_count, Tensor out_grad, str pool_type = "SUM")
  output : Tensor(x_grad)
  infer_meta :
    func : GeneralUnaryGradInferMeta
    param : [x]
  kernel :
    func : graph_send_recv_grad
712
    data_type : out_grad
713 714
  optional: out, dst_count

H
hong 已提交
715 716 717 718 719 720 721 722 723 724
- backward_api : gumbel_softmax_grad
  forward : gumbel_softmax (Tensor x, float temperature, bool hard, int axis) -> Tensor(out)
  args : (Tensor out, Tensor out_grad, int axis)
  output : Tensor(x_grad)
  infer_meta :
    func : GumbelSoftmaxGradInferMeta
    param : [out, out_grad, axis]
  kernel :
    func : gumbel_softmax_grad

725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744
- backward_api : hard_shrink_grad
  forward : hard_shrink (Tensor x, float threshold) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float threshold)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : hard_shrink_grad

- backward_api : hard_sigmoid_grad
  forward : hard_sigmoid (Tensor x, float slope, float offset) -> Tensor(out)
  args : (Tensor out, Tensor out_grad, float slope, float offset)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : hard_sigmoid_grad

745 746 747 748 749 750 751 752 753 754
- backward_api : hard_swish_grad
  forward : hard_swish (Tensor x, float threshold = 6.0, float scale = 6.0, float offset = 3.0) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float threshold, float scale, float offset)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : hard_swish_grad

755 756 757 758 759 760 761 762 763 764
- backward_api : huber_loss_grad
  forward : huber_loss (Tensor input, Tensor label, float delta) -> Tensor(out), Tensor(residual)
  args : (Tensor residual, Tensor out_grad, float delta)
  output : Tensor(input_grad), Tensor(label_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [residual, residual]
  kernel :
    func : huber_loss_grad

Z
zyfncg 已提交
765 766 767 768 769 770
- backward_api : imag_grad
  forward : imag (Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  invoke : imag_grad_impl(out_grad)

771 772 773 774 775 776 777 778 779 780
- backward_api : index_sample_grad
  forward : index_sample (Tensor x, Tensor index) -> Tensor(out)
  args : (Tensor x, Tensor index, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : index_sample_grad
    data_type : out_grad
H
hong 已提交
781
  no_need_buffer : x
782

F
From00 已提交
783 784 785 786 787 788 789 790 791 792
- backward_api : index_select_grad
  forward : index_select(Tensor x, Tensor index,  int dim) -> Tensor(out)
  args : (Tensor x, Tensor index, Tensor out_grad,  int dim)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : index_select_grad
    data_type : x
H
hong 已提交
793
  no_need_buffer : x
F
From00 已提交
794

795 796 797 798 799 800 801 802 803
- backward_api : kldiv_loss_grad
  forward : kldiv_loss(Tensor x, Tensor label, str reduction) -> Tensor(out)
  args : (Tensor x, Tensor label, Tensor out_grad, str reduction)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : kldiv_loss_grad
H
hong 已提交
804
  no_need_buffer : x
805

806 807 808 809 810 811 812 813 814 815 816
- backward_api : kron_grad
  forward : kron (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : kron_grad
    data_type : out_grad

817 818 819 820 821 822 823 824 825 826
- backward_api : kthvalue_grad
  forward : kthvalue(Tensor x, int k, int axis, bool keepdim) -> Tensor(out), Tensor(indices)
  args : (Tensor x, Tensor indices, Tensor out_grad, int k, int axis, bool keepdim)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : kthvalue_grad

827 828 829 830 831 832 833 834 835 836 837
- backward_api : label_smooth_grad
  forward : label_smooth (Tensor label, Tensor prior_dist, float epsilon) -> Tensor(out)
  args : (Tensor out_grad, float epsilon)
  output : Tensor(label_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : label_smooth_grad
  optional : prior_dist

H
hong 已提交
838 839 840 841 842 843 844 845 846 847
- backward_api : layer_norm_grad
  forward : layer_norm (Tensor x, Tensor scale, Tensor bias, float epsilon, int begin_norm_axis, bool is_test) -> Tensor(out), Tensor(mean), Tensor(variance)
  args : (Tensor x,  Tensor scale, Tensor bias, Tensor mean, Tensor variance, Tensor out_grad, float epsilon, int begin_norm_axis, bool is_test)
  output : Tensor(x_grad), Tensor(scale_grad), Tensor(bias_grad)
  infer_meta :
    func : LayerNormGradInferMeta
    param : [x, scale, bias]
  kernel :
    func : layer_norm_grad
    data_type : out_grad
W
wanghuancoder 已提交
848
  no_need_buffer : bias
H
hong 已提交
849 850
  optional : scale, bias

851 852 853 854 855 856 857 858 859 860
- backward_api : leaky_relu_double_grad
  forward : leaky_relu_grad (Tensor x, Tensor grad_out, float alpha) -> Tensor(grad_x)
  args : (Tensor x, Tensor grad_x_grad, float alpha)
  output : Tensor(grad_out_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [grad_x_grad]
  kernel :
    func : leaky_relu_double_grad

861 862 863 864 865 866 867 868 869
- backward_api : leaky_relu_grad
  forward : leaky_relu (Tensor x, float alpha) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float alpha)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : leaky_relu_grad
870
  backward : leaky_relu_double_grad
871 872

- backward_api : lerp_grad
873
  forward : lerp (Tensor x, Tensor y, Tensor weight) -> Tensor(out)
874 875 876 877 878 879 880 881
  args : (Tensor x, Tensor y, Tensor weight, Tensor out, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : lerp_grad

882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921
- backward_api : lgamma_grad
  forward : lgamma(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : lgamma_grad

- backward_api : log10_grad
  forward : log10 (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : log10_grad

- backward_api : log1p_grad
  forward : log1p (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : log1p_grad

- backward_api : log2_grad
  forward : log2 (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : log2_grad

922 923 924 925 926 927 928 929 930 931
- backward_api : log_double_grad
  forward : log_grad (Tensor x, Tensor grad_out) -> Tensor(grad_x)
  args : (Tensor x, Tensor grad_out, Tensor grad_x_grad)
  output : Tensor(x_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, x]
  kernel :
    func : log_double_grad

932 933 934 935 936 937 938 939 940
- backward_api : log_grad
  forward : log (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : log_grad
941
  backward : log_double_grad
942

943 944 945 946 947 948 949 950 951 952
- backward_api : log_loss_grad
  forward : log_loss (Tensor input, Tensor label, float epsilon) -> Tensor(out)
  args : (Tensor input, Tensor label, Tensor out_grad, float epsilon)
  output : Tensor(input_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [input]
  kernel :
    func : log_loss_grad

953 954 955 956 957 958 959 960 961 962
- backward_api : log_softmax_grad
  forward : log_softmax(Tensor x,  int axis) -> Tensor(out)
  args : (Tensor out, Tensor out_grad,  int axis)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [out]
  kernel :
    func : log_softmax_grad

963 964 965 966 967 968 969 970 971 972
- backward_api : logit_grad
  forward : logit (Tensor x, float eps = 1e-6f) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float eps)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : logit_grad

973 974
- backward_api : logsigmoid_grad
  forward : logsigmoid (Tensor x) -> Tensor(out)
H
hong 已提交
975 976 977
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
978 979 980 981 982
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : logsigmoid_grad

983 984 985 986 987 988 989 990 991 992
- backward_api : logsumexp_grad
  forward : logsumexp(Tensor x, int64_t[] axis,  bool keepdim,  bool reduce_all) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int64_t[] axis,  bool keepdim,  bool reduce_all)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : logsumexp_grad

993 994 995 996 997 998 999 1000 1001 1002
- backward_api : masked_select_grad
  forward : masked_select (Tensor x, Tensor mask) -> Tensor(out)
  args : (Tensor x, Tensor mask, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : masked_select_grad
    data_type : x
H
hong 已提交
1003
  no_need_buffer : x
1004 1005

- backward_api : matmul_double_grad
1006 1007 1008
  forward : matmul_grad (Tensor x, Tensor y, Tensor grad_out, bool transpose_x=false, bool transpose_y=false) -> Tensor(grad_x), Tensor(grad_y)
  args : (Tensor x, Tensor y, Tensor grad_out, Tensor grad_x_grad, Tensor grad_y_grad, bool transpose_x=false, bool transpose_y=false)
  output : Tensor(x_grad), Tensor(y_grad), Tensor(grad_out_grad)
1009 1010
  infer_meta :
    func : GeneralTernaryGradInferMeta
1011
    param : [x, y, grad_out]
1012 1013
  kernel :
    func : matmul_double_grad
1014
  backward : matmul_triple_grad
1015
  optional : grad_x_grad, grad_y_grad
1016 1017 1018 1019 1020 1021 1022 1023 1024 1025

- backward_api : matmul_grad
  forward : matmul (Tensor x, Tensor y, bool transpose_x=false, bool transpose_y=false) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, bool transpose_x=false, bool transpose_y=false)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : matmul_grad
1026
  backward : matmul_double_grad
1027

1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038
- backward_api : matmul_triple_grad
  forward : matmul_double_grad (Tensor x, Tensor y, Tensor fwd_grad_out, Tensor fwd_grad_grad_x, Tensor fwd_grad_grad_y, bool transpose_x=false, bool transpose_y=false) -> Tensor(grad_x), Tensor(grad_y), Tensor(grad_grad_out)
  args : (Tensor x, Tensor y, Tensor fwd_grad_out, Tensor fwd_grad_grad_x, Tensor fwd_grad_grad_y, Tensor grad_x_grad, Tensor grad_y_grad, Tensor grad_grad_out_grad, bool transpose_x=false, bool transpose_y=false)
  output : Tensor(x_grad), Tensor(y_grad), Tensor(fwd_grad_out_grad), Tensor(fwd_grad_grad_x_grad), Tensor(fwd_grad_grad_y_grad)
  infer_meta :
    func : GeneralQuinaryGradInferMeta
    param : [x, y, fwd_grad_out, fwd_grad_grad_x, fwd_grad_grad_y]
  kernel :
    func : matmul_triple_grad
  optional : grad_x_grad, grad_y_grad, grad_grad_out_grad

1039 1040 1041 1042 1043 1044 1045 1046 1047 1048
- backward_api : matrix_power_grad
  forward : matrix_power (Tensor x, int n) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int n)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : matrix_power_grad

1049 1050 1051 1052 1053 1054 1055 1056 1057 1058
- backward_api : max_grad
  forward: max (Tensor x,  int64_t[] dims={},  bool keep_dim=false) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int64_t[] dims={},  bool keep_dim=false, bool reduce_all=false)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : max_grad

F
From00 已提交
1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076
- backward_api : max_pool2d_with_index_grad
  forward : max_pool2d_with_index(Tensor x, int[] kernel_size, int[] strides, int[] paddings, bool global_pooling, bool adaptive) -> Tensor(out), Tensor(mask)
  args : (Tensor x, Tensor mask, Tensor out_grad, int[] kernel_size, int[] strides, int[] paddings, bool global_pooling, bool adaptive)
  output : Tensor(x_grad)
  infer_meta :
    func : MaxPoolWithIndexGradInferMeta
  kernel :
    func : max_pool2d_with_index_grad

- backward_api : max_pool3d_with_index_grad
  forward : max_pool3d_with_index(Tensor x, int[] kernel_size, int[] strides, int[] paddings, bool global_pooling, bool adaptive) -> Tensor(out), Tensor(mask)
  args : (Tensor x, Tensor mask, Tensor out_grad, int[] kernel_size, int[] strides, int[] paddings, bool global_pooling, bool adaptive)
  output : Tensor(x_grad)
  infer_meta :
    func : MaxPoolWithIndexGradInferMeta
  kernel :
    func : max_pool3d_with_index_grad

1077 1078 1079 1080 1081 1082 1083 1084 1085 1086
- backward_api : maximum_grad
  forward : maximum(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis=-1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : maximum_grad

1087 1088 1089 1090 1091 1092 1093 1094 1095 1096
- backward_api : maxout_grad
  forward : maxout(Tensor x, int groups, int axis) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int groups, int axis)
  output : Tensor(x_grad)
  infer_meta :
    func : GeneralUnaryGradInferMeta
    param: [x]
  kernel :
    func : maxout_grad

1097 1098 1099 1100 1101 1102 1103 1104 1105 1106
- backward_api : mean_all_grad
  forward : mean_all(Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : mean_all_grad

1107 1108 1109 1110 1111 1112
- backward_api : mean_double_grad
  forward: mean_grad (Tensor x, Tensor grad_out, int64_t[] dims={},  bool keep_dim=false, bool reduce_all = false) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, int64_t[] dims={},  bool keep_dim=false, bool reduce_all=false)
  output : Tensor(grad_out_grad)
  invoke : mean(grad_x_grad, dims, keep_dim)

1113 1114 1115 1116 1117 1118 1119 1120 1121
- backward_api : mean_grad
  forward: mean (Tensor x,  int64_t[] dims={},  bool keep_dim=false) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int64_t[] dims={},  bool keep_dim=false, bool reduce_all=false)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : mean_grad
1122
  backward : mean_double_grad
H
hong 已提交
1123
  no_need_buffer : x
1124

Y
YuanRisheng 已提交
1125 1126 1127
- backward_api : meshgrid_grad
  forward : meshgrid (Tensor[] inputs) -> Tensor[](outputs)
  args : (Tensor[] inputs, Tensor[] outputs_grad)
1128 1129 1130 1131 1132
  output : Tensor[](inputs_grad){inputs.size()}
  infer_meta :
    func : MeshgridGradInferMeta
  kernel :
    func : meshgrid_grad
Y
YuanRisheng 已提交
1133

1134 1135 1136 1137 1138 1139 1140 1141 1142 1143
- backward_api : min_grad
  forward: min (Tensor x,  int64_t[] dims={},  bool keep_dim=false) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int64_t[] dims={},  bool keep_dim=false, bool reduce_all=false)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : min_grad

1144 1145 1146 1147 1148 1149 1150 1151 1152 1153
- backward_api : minimum_grad
  forward : minimum(Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis=-1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, y]
  kernel :
    func : minimum_grad

1154 1155 1156 1157 1158 1159 1160 1161 1162 1163
- backward_api : mish_grad
  forward : mish (Tensor x, float threshold) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float threshold)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : mish_grad

1164 1165 1166 1167 1168 1169 1170 1171 1172 1173
- backward_api : mode_grad
  forward : mode(Tensor x,  int axis,  bool keepdim) -> Tensor(out), Tensor(indices)
  args : (Tensor x, Tensor indices, Tensor out_grad,  int axis,  bool keepdim)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : mode_grad

1174
- backward_api : modulo_grad
1175
  forward : modulo (Tensor x, Tensor y) -> Tensor(out)
1176 1177 1178 1179 1180 1181 1182 1183 1184
  args : (Tensor x, Tensor y, Tensor out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : modulo_grad
  no_need_buffer : x, y

1185 1186 1187
- backward_api : multi_dot_grad
  forward : multi_dot (Tensor[] x) -> Tensor(out)
  args : (Tensor[] x, Tensor out_grad)
1188 1189 1190 1191 1192
  output : Tensor[](x_grad) {x.size()}
  infer_meta :
    func : MultiDotGradInferMeta
  kernel :
    func : multi_dot_grad
1193 1194 1195 1196

- backward_api : multiplex_grad
  forward : multiplex (Tensor[] ins, Tensor ids) -> Tensor(out)
  args : (Tensor[] ins, Tensor ids, Tensor out_grad)
1197 1198 1199 1200 1201 1202 1203
  output : Tensor[](ins_grad){ins.size()}
  infer_meta :
    func : MultiplexGradInferMeta
    param : [ids, out_grad]
  kernel :
    func : multiplex_grad
    param : [ids, out_grad]
1204

1205 1206 1207 1208 1209 1210 1211 1212 1213 1214
- backward_api : multiply_double_grad
  forward : multiply_grad (Tensor x, Tensor y, Tensor grad_out, int axis = -1) -> Tensor(grad_x), Tensor(grad_y)
  args : (Tensor x, Tensor y, Tensor grad_out, Tensor grad_x_grad, Tensor grad_y_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [x, y, grad_out]
  kernel :
    func : multiply_double_grad
  optional : grad_x_grad, grad_y_grad
1215
  backward : multiply_triple_grad
1216

1217 1218 1219 1220 1221 1222 1223 1224 1225
- backward_api : multiply_grad
  forward : multiply (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : multiply_grad
1226
  backward : multiply_double_grad
1227

1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238
- backward_api : multiply_triple_grad
  forward : multiply_double_grad (Tensor x, Tensor y, Tensor fwd_grad_out, Tensor fwd_grad_grad_x, Tensor fwd_grad_grad_y, int aixs = -1) -> Tensor(grad_x), Tensor(grad_y), Tensor(grad_grad_out)
  args : (Tensor x, Tensor y, Tensor fwd_grad_out, Tensor fwd_grad_grad_x, Tensor fwd_grad_grad_y, Tensor grad_x_grad, Tensor grad_y_grad, Tensor grad_grad_out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad), Tensor(fwd_grad_out_grad), Tensor(fwd_grad_grad_x_grad), Tensor(fwd_grad_grad_y_grad)
  infer_meta :
    func : GeneralQuinaryGradInferMeta
    param : [x, y, fwd_grad_out, x, y]
  kernel :
    func : multiply_triple_grad
  optional : fwd_grad_grad_x, fwd_grad_grad_y, grad_grad_out_grad

1239 1240 1241 1242 1243 1244 1245
- backward_api : mv_grad
  forward : mv (Tensor x, Tensor vec) -> Tensor(out)
  args : (Tensor x, Tensor vec, Tensor out_grad)
  output : Tensor(x_grad), Tensor(vec_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, vec]
H
hong 已提交
1246
  kernel :
1247
    func : mv_grad
H
hong 已提交
1248

1249
- backward_api : nll_loss_grad
Z
zyfncg 已提交
1250 1251 1252
  forward : nll_loss (Tensor input, Tensor label, Tensor weight, int64_t ignore_index, str reduction) -> Tensor(out), Tensor(total_weight)
  args : (Tensor input, Tensor label, Tensor weight, Tensor total_weight, Tensor out_grad, int64_t ignore_index, str reduction)
  output : Tensor(input_grad)
H
hong 已提交
1253
  infer_meta :
Z
zyfncg 已提交
1254
    func : NllLossGradInferMeta
H
hong 已提交
1255
  kernel :
1256
    func : nll_loss_grad
Z
zyfncg 已提交
1257
    data_type : input
1258
  optional : weight
H
hong 已提交
1259

H
hong 已提交
1260 1261 1262 1263 1264 1265 1266 1267 1268 1269
- backward_api : norm_grad
  forward : norm (Tensor x, int axis, float epsilon, bool is_test) -> Tensor(out), Tensor(norm)
  args : (Tensor x, Tensor norm, Tensor out_grad, int axis, float epsilon, bool is_test)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : norm_grad

1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288
- backward_api : p_norm_grad
  forward : p_norm(Tensor x,  float porder,  int axis,  float epsilon,  bool keepdim,  bool asvector=false) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad,  float porder,  int axis,  float epsilon,  bool keepdim,  bool asvector)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : p_norm_grad

- backward_api : pad3d_grad
  forward : pad3d(Tensor x, IntArray paddings, str mode,  float pad_value, str data_format) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray paddings, str mode,  float pad_value, str data_format)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : pad3d_grad
W
wanghuancoder 已提交
1289
  no_need_buffer : x
1290

H
hong 已提交
1291 1292 1293 1294 1295 1296 1297 1298 1299
- backward_api : pixel_shuffle_grad
  forward : pixel_shuffle (Tensor x, int upscale_factor, str data_format) -> Tensor(out)
  args : (Tensor out_grad, int upscale_factor, str data_format)
  output : Tensor(x_grad)
  infer_meta :
    func : PixelShuffleGradInferMeta
  kernel :
    func : pixel_shuffle_grad

H
hong 已提交
1300 1301 1302 1303 1304 1305 1306 1307 1308 1309
- backward_api : poisson_grad
  forward : poisson (Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : poisson_grad

1310 1311 1312 1313 1314 1315 1316 1317 1318 1319
- backward_api : pool2d_double_grad
  forward : pool2d_grad(Tensor x, Tensor out, Tensor grad_out, int[] kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, int[] kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm)
  output : Tensor(grad_out_grad)
  infer_meta :
    func : PoolInferMeta
  kernel :
    func : pool2d_double_grad
    use_gpudnn : true

F
From00 已提交
1320 1321 1322 1323 1324 1325 1326 1327
- backward_api : pool2d_grad
  forward : pool2d(Tensor x, int[] kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int[] kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm)
  output : Tensor(x_grad)
  infer_meta :
    func : PoolGradInferMeta
  kernel :
    func : pool2d_grad
1328
    use_gpudnn : true
1329
  backward : pool2d_double_grad
1330 1331 1332 1333 1334 1335 1336 1337 1338 1339

- backward_api : pool2d_grad_gpudnn_unused
  forward : pool2d_gpudnn_unused(Tensor x, int[] kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int[] kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm)
  output : Tensor(x_grad)
  infer_meta :
    func : PoolGradInferMeta
  kernel :
    func : pool2d_grad
    use_gpudnn : false
F
From00 已提交
1340 1341 1342 1343 1344 1345 1346 1347 1348

- backward_api : pool3d_grad
  forward : pool3d(Tensor x, int[] kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int[] kernel_size, int[] strides, int[] paddings, bool ceil_mode, bool exclusive, str data_format, str pooling_type, bool global_pooling, bool adaptive, str padding_algorithm)
  output : Tensor(x_grad)
  infer_meta :
    func : PoolGradInferMeta
  kernel :
    func : pool3d_grad
1349
    use_gpudnn : true
F
From00 已提交
1350

1351 1352 1353 1354 1355 1356 1357 1358 1359 1360
- backward_api : pow_grad
  forward : pow(Tensor x, Scalar s) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, Scalar s=-1)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [x]
  kernel :
    func : pow_grad

1361 1362 1363 1364 1365 1366 1367 1368 1369 1370
- backward_api : prelu_grad
  forward : prelu(Tensor x, Tensor alpha, str data_format, str mode) -> Tensor(out)
  args : (Tensor x, Tensor alpha, Tensor out_grad, str data_format, str mode)
  output : Tensor(x_grad), Tensor(alpha_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param: [x, alpha]
  kernel :
    func : prelu_grad

1371
- backward_api : psroi_pool_grad
Z
zyfncg 已提交
1372 1373
  forward : psroi_pool (Tensor x, Tensor boxes, Tensor boxes_num, int pooled_height, int pooled_width, int output_channels, float spatial_scale) -> Tensor(out)
  args : (Tensor x, Tensor boxes, Tensor boxes_num, Tensor out_grad, int pooled_height, int pooled_width, int output_channels, float spatial_scale)
1374 1375
  output : Tensor(x_grad)
  infer_meta :
Z
zyfncg 已提交
1376
    func : GeneralUnaryGradInferMeta
1377 1378
    param : [x]
  kernel :
1379
    func : psroi_pool_grad
1380
    data_type : x
Z
zyfncg 已提交
1381
  optional : boxes_num
1382 1383 1384 1385 1386 1387

# output is optional
- backward_api : put_along_axis_grad
  forward : put_along_axis (Tensor x, Tensor index, Tensor value, int axis, str reduce) -> Tensor(out)
  args : (Tensor x, Tensor index, Tensor out_grad, int axis, str reduce)
  output : Tensor(x_grad), Tensor(value_grad)
H
hong 已提交
1388
  infer_meta :
1389 1390
    func : GeneralBinaryGradInferMeta
    param : [x, index]
H
hong 已提交
1391
  kernel :
1392
    func : put_along_axis_grad
H
hong 已提交
1393

Z
zyfncg 已提交
1394 1395 1396 1397 1398 1399
- backward_api : real_grad
  forward : real (Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  invoke : real_grad_impl(out_grad)

1400 1401 1402 1403 1404 1405 1406 1407 1408 1409
- backward_api : reciprocal_grad
  forward : reciprocal (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : reciprocal_grad

H
hong 已提交
1410 1411 1412 1413 1414 1415 1416 1417
- backward_api : reduce_prod_grad
  forward : reduce_prod (Tensor x, int64_t[] dims, bool keep_dim, bool reduce_all) -> Tensor(out)
  args : (Tensor x, Tensor out, Tensor out_grad, int64_t[] dims,  bool keep_dim, bool reduce_all)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
H
hong 已提交
1418
    func : prod_grad
H
hong 已提交
1419

1420 1421 1422
- backward_api : relu_double_grad
  forward : relu_grad (Tensor out, Tensor grad_out) -> Tensor(grad_x)
  args : (Tensor out, Tensor grad_x_grad)
1423
  output : Tensor(grad_out_grad)
1424
  infer_meta :
1425 1426
    func : UnchangedInferMeta
    param : [out]
1427 1428 1429
  kernel :
    func : relu_double_grad

1430 1431 1432
- backward_api : relu_grad
  forward : relu (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
H
hong 已提交
1433 1434 1435
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
1436
    param : [out]
H
hong 已提交
1437
  kernel :
1438
    func : relu_grad
1439
  backward: relu_double_grad
H
hong 已提交
1440

1441 1442 1443 1444 1445 1446 1447 1448 1449
- backward_api : reshape_double_grad
  forward : reshape_grad (Tensor xshape, Tensor grad_out) -> Tensor(grad_x)
  args : (Tensor grad_out, Tensor grad_x_grad)
  output : Tensor(grad_out_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [grad_out]
  kernel :
    func : reshape_double_grad
W
wanghuancoder 已提交
1450
  no_need_buffer : grad_out
1451

1452
- backward_api : reshape_grad
1453
  forward : reshape_with_xshape (Tensor x, IntArray shape) -> Tensor(out), Tensor(xshape)
1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464
  args : (Tensor xshape, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : KernelWithXShapeInferMeta
    param : [xshape]
  kernel :
    func : reshape_grad
    param : [out_grad]
    data_type: out_grad
    backend: out_grad
    layout: out_grad
1465
  backward : reshape_double_grad
1466

1467 1468 1469 1470 1471 1472 1473 1474 1475
- backward_api : roi_align_grad
  forward : roi_align (Tensor x, Tensor boxes, Tensor boxes_num, int pooled_height, int pooled_width, float spatial_scale, int sampling_ratio, bool aligned) -> Tensor(out)
  args : (Tensor x, Tensor boxes, Tensor boxes_num, Tensor out_grad, int pooled_height, int pooled_width, float spatial_scale, int sampling_ratio, bool aligned)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : roi_align_grad
1476
    data_type : boxes
W
wanghuancoder 已提交
1477
  no_need_buffer : x
1478 1479
  optional : boxes_num

Z
zyfncg 已提交
1480 1481 1482 1483 1484 1485 1486 1487 1488
- backward_api : roi_pool_grad
  forward : roi_pool (Tensor x, Tensor boxes, Tensor boxes_num, int pooled_height, int pooled_width, float spatial_scale) -> Tensor(out), Tensor(arg_max)
  args : (Tensor x, Tensor boxes, Tensor boxes_num, Tensor arg_max, Tensor out_grad, int pooled_height, int pooled_width, float spatial_scale)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : roi_pool_grad
1489
    data_type : x
Z
zyfncg 已提交
1490 1491
  optional : boxes_num

F
From00 已提交
1492 1493 1494 1495 1496 1497 1498 1499 1500 1501
- backward_api : roll_grad
  forward : roll(Tensor x, IntArray shifts, int64_t[] axis) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray shifts, int64_t[] axis)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : roll_grad
    data_type : x
H
hong 已提交
1502
  no_need_buffer : x
F
From00 已提交
1503

1504 1505 1506 1507 1508 1509 1510 1511 1512 1513
- backward_api : round_grad
  forward : round(Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param: [out_grad]
  kernel :
    func : round_grad

Z
zyfncg 已提交
1514 1515 1516 1517 1518 1519 1520 1521 1522 1523
- backward_api : rsqrt_grad
  forward : rsqrt (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : rsqrt_grad

1524 1525 1526 1527 1528 1529 1530
- backward_api : scale_double_grad
  forward : scale_grad (Tensor grad_out, Scalar scale, float bias, bool bias_after_scale) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, Scalar scale=1.0, float bias=0.0, bool bias_after_scale=true)
  output : Tensor(grad_out_grad)
  invoke : scale(grad_x_grad, scale, 0.0, bias_after_scale)
  backward : scale_triple_grad

1531 1532
- backward_api : scale_grad
  forward : scale (Tensor x, Scalar scale, float bias, bool bias_after_scale) -> Tensor(out)
1533
  args : (Tensor out_grad, Scalar scale=1.0, float bias=0.0, bool bias_after_scale=true)
H
hong 已提交
1534
  output : Tensor(x_grad)
1535
  invoke : scale(out_grad, scale, 0.0, bias_after_scale)
1536 1537 1538 1539 1540 1541 1542
  backward : scale_double_grad

- backward_api : scale_triple_grad
  forward : scale_double_grad (Tensor grad_grad_x, Scalar scale, float bias, bool bias_after_scale) -> Tensor(grad_grad_out)
  args : (Tensor grad_grad_out_grad, Scalar scale=1.0, float bias=0.0, bool bias_after_scale=true)
  output : Tensor(grad_grad_x_grad)
  invoke : scale(grad_grad_out_grad, scale, 0.0, bias_after_scale)
H
hong 已提交
1543 1544 1545 1546 1547 1548 1549 1550 1551 1552

- backward_api : scatter_grad
  forward : scatter (Tensor x, Tensor index, Tensor updates, bool overwrite) -> Tensor(out)
  args : (Tensor index, Tensor updates, Tensor out_grad, bool overwrite)
  output : Tensor(x_grad), Tensor(updates_grad)
  infer_meta :
    func : ScatterGradInferMeta
    param : [index, updates, out_grad, overwrite]
  kernel :
    func : scatter_grad
H
hong 已提交
1553
  no_need_buffer : updates
H
hong 已提交
1554 1555

- backward_api : scatter_nd_add_grad
1556
  forward : scatter_nd_add (Tensor x, Tensor index, Tensor updates) -> Tensor(out)
H
hong 已提交
1557 1558 1559 1560 1561 1562
  args : (Tensor index, Tensor updates, Tensor out_grad)
  output : Tensor(x_grad), Tensor(updates_grad)
  infer_meta :
    func : ScatterNdAddGradInferMeta
    param : [index, updates, out_grad]
  kernel :
1563
    func : scatter_nd_add_grad
H
hong 已提交
1564
  no_need_buffer : updates
H
hong 已提交
1565

1566 1567 1568 1569
- backward_api : segment_pool_grad
  forward : segment_pool (Tensor x, Tensor segment_ids, str pooltype) -> Tensor(out), Tensor(summed_ids)
  args : (Tensor x, Tensor segment_ids, Tensor out, Tensor summed_ids, Tensor out_grad, str pooltype)
  output : Tensor(x_grad)
H
hong 已提交
1570
  infer_meta :
1571 1572
    func : UnchangedInferMeta
    param : [x]
H
hong 已提交
1573
  kernel :
1574
    func : segment_pool_grad
1575
    data_type : x
H
hong 已提交
1576
  optional : summed_ids
H
hong 已提交
1577

1578 1579 1580 1581
- backward_api : selu_grad
  forward : selu (Tensor x, float scale, float alpha) -> Tensor(out)
  args : (Tensor out, Tensor out_grad, float scale, float alpha)
  output : Tensor(x_grad)
H
hong 已提交
1582
  infer_meta :
1583 1584
    func : UnchangedInferMeta
    param : [out]
H
hong 已提交
1585
  kernel :
1586
    func : selu_grad
H
hong 已提交
1587

1588 1589 1590 1591 1592 1593 1594 1595
- backward_api : sigmoid_cross_entropy_with_logits_grad
  forward : sigmoid_cross_entropy_with_logits (Tensor x, Tensor label, bool normalize, int ignore_index) -> Tensor(out)
  args : (Tensor x, Tensor label, Tensor out_grad, bool normalize, int ignore_index)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
1596
    func : sigmoid_cross_entropy_with_logits_grad
H
hong 已提交
1597

1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608
- backward_api : sigmoid_double_grad
  forward : sigmoid_grad (Tensor out, Tensor fwd_grad_out) -> Tensor(grad_x)
  args : (Tensor out, Tensor fwd_grad_out, Tensor grad_x_grad)
  output : Tensor(out_grad), Tensor(fwd_grad_out_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [out, fwd_grad_out]
  kernel :
    func : sigmoid_double_grad
  backward : sigmoid_triple_grad

1609 1610 1611 1612 1613 1614 1615 1616 1617
- backward_api : sigmoid_grad
  forward : sigmoid (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : sigmoid_grad
1618 1619 1620 1621 1622 1623 1624 1625 1626 1627
  backward : sigmoid_double_grad

- backward_api : sigmoid_triple_grad
  forward : sigmoid_double_grad (Tensor out, Tensor fwd_grad_out, Tensor grad_grad_x) -> Tensor(grad_out), Tensor(grad_grad_out)
  args : (Tensor out, Tensor fwd_grad_out, Tensor grad_grad_x, Tensor grad_out_grad, Tensor grad_grad_out_grad)
  output : Tensor(out_grad), Tensor(fwd_grad_out_grad), Tensor(grad_grad_x_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [out, fwd_grad_out, grad_grad_x]
  kernel :
1628
    func : sigmoid_triple_grad
H
hong 已提交
1629

1630 1631 1632
- backward_api : silu_grad
  forward : silu (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
H
hong 已提交
1633 1634 1635 1636 1637
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
1638
    func : silu_grad
H
hong 已提交
1639

1640 1641 1642 1643
- backward_api : sin_grad
  forward : sin (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
H
hong 已提交
1644
  infer_meta :
1645 1646
    func : UnchangedInferMeta
    param : [x]
H
hong 已提交
1647
  kernel :
1648
    func : sin_grad
H
hong 已提交
1649

1650 1651 1652 1653
- backward_api : sinh_grad
  forward : sinh (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
H
hong 已提交
1654
  infer_meta :
1655 1656
    func : UnchangedInferMeta
    param : [x]
H
hong 已提交
1657
  kernel :
1658
    func : sinh_grad
H
hong 已提交
1659

H
hong 已提交
1660 1661 1662 1663 1664 1665 1666 1667 1668
- backward_api : slice_grad
  forward : slice (Tensor input, int64_t[] axes, IntArray starts, IntArray ends, int64_t[] infer_flags, int64_t[] decrease_axis) -> Tensor(out)
  args : (Tensor input, Tensor out_grad, int64_t[] axes, IntArray starts, IntArray ends, int64_t[] infer_flags, int64_t[] decrease_axis)
  output : Tensor(input_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [input]
  kernel :
    func : slice_grad
H
hong 已提交
1669
  no_need_buffer : input
H
hong 已提交
1670

1671 1672 1673 1674
- backward_api : soft_shrink_grad
  forward : soft_shrink (Tensor x, float lambda) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float lambda)
  output : Tensor(x_grad)
H
hong 已提交
1675 1676
  infer_meta :
    func : UnchangedInferMeta
1677
    param : [x]
H
hong 已提交
1678
  kernel :
1679
    func : soft_shrink_grad
H
hong 已提交
1680

1681 1682 1683 1684 1685 1686 1687 1688 1689
- backward_api : softmax_grad
  forward : softmax (Tensor x, int axis) -> Tensor(out)
  args : (Tensor out, Tensor out_grad, int axis)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : softmax_grad
1690
    use_gpudnn : true
H
hong 已提交
1691

1692
- backward_api : split_grad
1693
  forward : split (Tensor x, IntArray num_or_sections, Scalar axis) -> Tensor[](out)
H
hong 已提交
1694
  args : (Tensor[] out_grad, Scalar axis = -1)
1695 1696 1697
  output : Tensor(x_grad)
  invoke : concat( out_grad, axis)
# TODO(zhangyunfei) The config of double grad and triple grad will be supported in the future.
H
hong 已提交
1698

1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718
- backward_api : sqrt_grad
  forward : sqrt (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out]
  kernel :
    func : sqrt_grad

- backward_api : square_grad
  forward : square (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : square_grad

1719 1720 1721 1722 1723 1724 1725 1726 1727 1728
- backward_api : squeeze_grad
  forward : squeeze(Tensor x, int[] axes) -> Tensor(xshape), Tensor(out)
  args : (Tensor xshape, Tensor out_grad, int[] axes)
  output : Tensor(x_grad)
  infer_meta :
    func : KernelWithXShapeInferMeta
    param: [xshape]
  kernel :
    func : squeeze_grad

1729 1730 1731
- backward_api : stack_grad
  forward : stack (Tensor[] x, int axis) -> Tensor(out)
  args : (Tensor[] x, Tensor out_grad, int axis)
1732 1733 1734 1735 1736 1737 1738
  output : Tensor[](x_grad){x.size()}
  infer_meta :
    func : StackGradInferMeta
    param: [out_grad, axis]
  kernel :
    func : stack_grad
    param : [out_grad, axis]
1739 1740
  no_need_buffer : x

1741 1742 1743 1744 1745 1746 1747 1748 1749
- backward_api : strided_slice_grad
  forward : strided_slice (Tensor x, int[] axes, IntArray starts, IntArray ends, IntArray strides) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int[] axes, IntArray starts, IntArray ends, IntArray strides)
  output : Tensor(x_grad)
  infer_meta :
    func : GeneralUnaryGradInferMeta
    param : [x]
  kernel :
    func : strided_slice_grad
H
hong 已提交
1750
  no_need_buffer : x
1751

1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763
- backward_api : subtract_double_grad
  forward : subtract_grad (Tensor x, Tensor y, Tensor grad_out, int axis = -1) -> Tensor(grad_x), Tensor(grad_y)
  args : (Tensor y, Tensor grad_out, Tensor grad_x_grad, Tensor grad_y_grad, int axis = -1)
  output : Tensor(grad_out_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [grad_out]
  kernel :
    func : subtract_double_grad
  optional : grad_x_grad, grad_y_grad
  no_need_buffer : y, grad_out

1764 1765 1766 1767 1768 1769 1770 1771 1772
- backward_api : subtract_grad
  forward : subtract (Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out_grad, int axis = -1)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : subtract_grad
H
hong 已提交
1773
  no_need_buffer : x, y
1774
  backward : subtract_double_grad
H
hong 已提交
1775

1776 1777 1778 1779 1780 1781 1782
- backward_api : sum_double_grad
  forward : sum_grad (Tensor x, Tensor grad_out, int64_t[] dims, bool keep_dim, bool reduce_all=false) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, int64_t[] dims={}, bool keep_dim=false)
  output : Tensor(grad_out_grad)
  invoke : sum(grad_x_grad, dims, grad_x_grad.dtype(), keep_dim)
  backward : sum_triple_grad

F
From00 已提交
1783 1784 1785 1786 1787 1788 1789 1790 1791
- backward_api : sum_grad
  forward : sum (Tensor x, int64_t[] dims={}, DataType out_dtype=paddle::experimental::DataType::UNDEFINED, bool keep_dim=false) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int64_t[] dims, bool keep_dim, bool reduce_all=false)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : sum_grad
W
wanghuancoder 已提交
1792
  no_need_buffer : x
1793 1794 1795 1796 1797 1798 1799
  backward : sum_double_grad

- backward_api : sum_triple_grad
  forward : sum_double_grad (Tensor grad_grad_x, int64_t[] dims={}, bool keep_dim=false) -> Tensor(grad_grad_out)
  args : (Tensor grad_grad_x, Tensor grad_grad_out_grad, int64_t[] dims={}, bool keep_dim=false, bool reduce_all=false)
  output : Tensor(grad_grad_x_grad)
  invoke : sum_grad(grad_grad_x, grad_grad_out_grad, dims, keep_dim, reduce_all)
H
hong 已提交
1800
  no_need_buffer : x
F
From00 已提交
1801

1802 1803 1804 1805 1806 1807 1808 1809 1810 1811
- backward_api : swish_grad
  forward : swish (Tensor x, float beta=1.0) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float bete=1.0)
  output : Tensor(x_grad)
  infer_meta :
    func : GeneralUnaryGradInferMeta
    param : [x]
  kernel :
    func : swish_grad

1812 1813 1814 1815 1816 1817 1818 1819 1820
- backward_api : take_along_axis_grad
  forward : take_along_axis (Tensor x, Tensor index, int axis) -> Tensor(out)
  args : (Tensor x, Tensor index, Tensor out_grad, int axis)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : take_along_axis_grad
H
hong 已提交
1821

1822 1823 1824
- backward_api : tan_grad
  forward : tan (Tensor x) -> Tensor(out)
  args : (Tensor x, Tensor out_grad)
H
hong 已提交
1825 1826 1827 1828 1829
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
1830
    func : tan_grad
H
hong 已提交
1831

1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842
- backward_api : tanh_double_grad
  forward : tanh_grad (Tensor out, Tensor grad_out) -> Tensor(grad_x)
  args : (Tensor out, Tensor grad_out, Tensor grad_x_grad)
  output : Tensor(out_grad), Tensor(grad_out_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [out, out]
  kernel :
    func : tanh_double_grad
  backward : tanh_triple_grad

1843 1844 1845 1846
- backward_api : tanh_grad
  forward : tanh (Tensor x) -> Tensor(out)
  args : (Tensor out, Tensor out_grad)
  output : Tensor(x_grad)
H
hong 已提交
1847
  infer_meta :
1848 1849
    func : UnchangedInferMeta
    param : [out]
H
hong 已提交
1850
  kernel :
1851
    func : tanh_grad
1852
  backward : tanh_double_grad
H
hong 已提交
1853

1854 1855
- backward_api : tanh_shrink_grad
  forward : tanh_shrink (Tensor x) -> Tensor(out)
Z
zhangbo9674 已提交
1856 1857 1858 1859 1860 1861
  args : (Tensor x, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
1862
    func : tanh_shrink_grad
H
hong 已提交
1863

1864 1865 1866 1867 1868 1869 1870 1871 1872 1873
- backward_api : tanh_triple_grad
  forward : tanh_double_grad (Tensor out, Tensor grad_out_forward, Tensor grad_x_grad_forward) -> Tensor(grad_out_new), Tensor(grad_out_grad)
  args : (Tensor out, Tensor grad_out_forward, Tensor grad_x_grad_forward, Tensor grad_out_new_grad, Tensor grad_out_grad_grad)
  output : Tensor(out_grad), Tensor(grad_out_forward_grad), Tensor(grad_x_grad_forward_grad)
  infer_meta :
    func : GeneralTernaryGradInferMeta
    param : [out, out, grad_x_grad_forward]
  kernel :
    func : tanh_triple_grad

1874 1875 1876 1877 1878 1879 1880 1881 1882
- backward_api : thresholded_relu_grad
  forward : thresholded_relu (Tensor x, float threshold) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, float threshold)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : thresholded_relu_grad
H
hong 已提交
1883

1884
- backward_api : tile_grad
1885 1886
  forward : tile (Tensor x, IntArray repeat_times) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, IntArray repeat_times)
1887 1888 1889 1890 1891 1892
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : tile_grad
H
hong 已提交
1893
  no_need_buffer : x
H
hong 已提交
1894

1895 1896 1897 1898 1899 1900 1901 1902 1903 1904
- backward_api : top_k_grad
  forward : top_k (Tensor x, Scalar k, int axis = -1, bool largest = true, bool sorted = true) -> Tensor(out), Tensor(indices)
  args : (Tensor x, Tensor indices, Tensor out_grad, Scalar k = -1, int axis = -1, bool largest = true, bool sorted = true)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : top_k_grad

1905 1906 1907 1908 1909 1910 1911 1912 1913
- backward_api : trace_grad
  forward : trace (Tensor x, int offset, int axis1, int axis2) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int offset, int axis1, int axis2)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : trace_grad
H
hong 已提交
1914
  no_need_buffer : x
H
hong 已提交
1915

1916 1917 1918 1919 1920 1921
- backward_api : transpose_double_grad
  forward : transpose_grad (Tensor grad_out, int[] axis) -> Tensor(grad_x)
  args : (Tensor grad_x_grad, int[] axis)
  output : Tensor(grad_out_grad)
  invoke : transpose(grad_x_grad, axis)

1922 1923 1924 1925 1926 1927 1928 1929 1930
- backward_api : transpose_grad
  forward : transpose (Tensor x, int[] axis) -> Tensor(out)
  args : (Tensor out_grad, int[] axis)
  output : Tensor(x_grad)
  infer_meta :
    func : TransposeGradInferMeta
    param : [out_grad, axis]
  kernel :
    func : transpose_grad
1931
  backward : transpose_double_grad
H
hong 已提交
1932

H
hong 已提交
1933 1934 1935 1936 1937 1938 1939 1940 1941 1942
- backward_api : triangular_solve_grad
  forward : triangular_solve (Tensor x, Tensor y, bool upper, bool tranpose, bool unitriangular) -> Tensor(out)
  args : (Tensor x, Tensor y, Tensor out, Tensor out_grad, bool upper, bool tranpose, bool unitriangular)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : triangular_solve_grad

F
From00 已提交
1943 1944 1945 1946 1947 1948 1949 1950 1951 1952
- backward_api : tril_triu_grad
  forward : tril_triu(Tensor x,  int diagonal,  bool lower) -> Tensor(out)
  args : (Tensor out_grad,  int diagonal,  bool lower)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : tril_triu_grad

1953 1954 1955 1956 1957 1958 1959 1960 1961
- backward_api : trunc_grad
  forward : trunc (Tensor x) -> Tensor(out)
  args : (Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [out_grad]
  kernel :
    func : trunc_grad
H
hong 已提交
1962

1963 1964 1965 1966 1967 1968
- backward_api : unbind_grad
  forward : unbind (Tensor input, int axis) -> Tensor[](out)
  args : (Tensor[] out_grad, int axis)
  output : Tensor(input_grad)
  invoke : stack(out_grad, axis)

1969 1970 1971 1972 1973 1974 1975 1976 1977
- backward_api : unfold_grad
  forward : unfold (Tensor x, int[] kernel_sizes, int[] strides, int[] paddings, int[] dilations) -> Tensor(out)
  args : (Tensor x, Tensor out_grad, int[] kernel_sizes, int[] strides, int[] paddings, int[] dilations)
  output : Tensor(x_grad)
  infer_meta :
    func : UnchangedInferMeta
    param : [x]
  kernel :
    func : unfold_grad
H
hong 已提交
1978
  no_need_buffer : x
H
hong 已提交
1979

1980 1981 1982 1983 1984 1985 1986 1987 1988 1989
- backward_api : unsqueeze_grad
  forward : unsqueeze(Tensor x, IntArray axes) -> Tensor(xshape), Tensor(out)
  args : (Tensor xshape, Tensor out_grad)
  output : Tensor(x_grad)
  infer_meta :
    func : KernelWithXShapeInferMeta
    param: [xshape]
  kernel :
    func : unsqueeze_grad

1990 1991 1992 1993 1994 1995 1996 1997 1998
- backward_api : where_grad
  forward : where (Tensor condition, Tensor x, Tensor y) -> Tensor(out)
  args : (Tensor condition, Tensor x, Tensor y, Tensor out_grad)
  output : Tensor(x_grad), Tensor(y_grad)
  infer_meta :
    func : GeneralBinaryGradInferMeta
    param : [x, y]
  kernel :
    func : where_grad
H
hong 已提交
1999
  no_need_buffer : x, y