1# Copyright 2020-2021 Huawei Technologies Co., Ltd 2# 3# Licensed under the Apache License, Version 2.0 (the "License"); 4# you may not use this file except in compliance with the License. 5# You may obtain a copy of the License at 6# 7# http://www.apache.org/licenses/LICENSE-2.0 8# 9# Unless required by applicable law or agreed to in writing, software 10# distributed under the License is distributed on an "AS IS" BASIS, 11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12# See the License for the specific language governing permissions and 13# limitations under the License. 14# ============================================================================ 15"""thor_ops""" 16import math 17 18from mindspore.ops.primitive import prim_attr_register, PrimitiveWithInfer 19from mindspore.common import dtype as mstype 20from mindspore import _checkparam as validator 21from mindspore.ops.operations.nn_ops import _check_positive_int_or_tuple 22 23__all__ = ["CusBatchMatMul", 24 "CusCholeskyTrsm", 25 "CusFusedAbsMax1", 26 "CusImg2Col", 27 "CusMatMulCubeDenseLeft", 28 "CusMatMulCubeFraczRightMul", 29 "CusMatMulCube", 30 "CusMatrixCombine", 31 "CusTranspose02314", 32 "CusMatMulCubeDenseRight", 33 "CusMatMulCubeFraczLeftCast", 34 "LoadIm2Col" 35 ] 36 37 38class CusBatchMatMul(PrimitiveWithInfer): 39 """ 40 Multiplies matrix `a` by matrix `b` in batch. 41 42 The rank of input tensors must be `3`. 43 44 Inputs: 45 - **input_x** (Tensor) - The first tensor to be multiplied. The shape of the tensor is :math:`(N, D, D)`. 46 - **input_y** (Tensor) - The second tensor to be multiplied. The shape of the tensor is :math:`(N, D, D)`. If 47 `transpose_b` is True. 48 49 Outputs: 50 Tensor, the shape of the output tensor is :math:`(N, D, D)`. 51 52 Examples: 53 >>> input_x = Tensor(np.ones(shape=[2, 128, 128]), mindspore.float32) 54 >>> input_y = Tensor(np.ones(shape=[2, 128, 128]), mindspore.float32) 55 >>> cus_batch_matmul = ops.CusBatchMatMul() 56 >>> output = cus_batch_matmul(input_x, input_y) 57 """ 58 59 @prim_attr_register 60 def __init__(self): 61 """Initialize CusBatchMatMul""" 62 self.init_prim_io_names(inputs=['x1', 'x2'], outputs=['y']) 63 from mindspore.ops._op_impl._custom_op.batch_matmul_impl import cus_batch_matmul 64 65 def infer_shape(self, data1_shape, data2_shape): 66 return data1_shape 67 68 def infer_dtype(self, data1_dtype, data2_dtype): 69 return data1_dtype 70 71 72class CusCholeskyTrsm(PrimitiveWithInfer): 73 r""" 74 L * LT = A. 75 LT * (LT)^-1 = I. 76 return (LT)^-1. 77 Only compute the res of the diag part of input matrix with dim 128. 78 The rank of input tensors must be `2`. 79 80 Inputs: 81 - **input_x** (Tensor) - The first tensor to be multiplied. The shape of the tensor is :math:`(N, N)`. 82 83 Outputs: 84 Tensor, the shape of the output tensor is :math:`(N // Split\_dim, Split\_dim, Split\_dim)`. 85 86 Examples: 87 >>> input_x = Tensor(np.ones(shape=[256, 256]), mindspore.float32) 88 >>> cus_choleskytrsm = ops.CusCholeskyTrsm() 89 >>> output = cus_choleskytrsm(input_x) 90 """ 91 92 @prim_attr_register 93 def __init__(self): 94 """Initialize CusCholeskyTrsm""" 95 self.init_prim_io_names(inputs=['x1'], outputs=['y']) 96 from mindspore.ops._op_impl._custom_op.cholesky_trsm_impl import cus_cholesky_trsm 97 98 def infer_shape(self, data1_shape): 99 ll = [] 100 m, _ = data1_shape 101 if m >= 128: 102 ll = [m // 128, 128, 128] 103 else: 104 ll = [1, 64, 64] 105 return ll 106 107 def infer_dtype(self, data1_dtype): 108 return data1_dtype 109 110 111class CusFusedAbsMax1(PrimitiveWithInfer): 112 """ 113 Computes the abs max of Tensor input. 114 115 The rank of input tensors must be `4` or `2`. 116 Inputs: 117 - **input_x** (Tensor) - The first tensor to be multiplied. The shape of the tensor is :math:`(N0, M0, N1, M1)` 118 or math:`(32, 64)`. 119 Outputs: 120 Tensor, the shape of the output tensor is :math:`(32, 64)` or math:`(1, )`. 121 122 Examples: 123 >>> input_x = Tensor(np.ones(shape=[1, 3]), mindspore.float32) 124 >>> cus_fused_abs_max1 = ops.CusFusedAbsMax1() 125 >>> output = cus_fused_abs_max1(input_x) 126 """ 127 128 @prim_attr_register 129 def __init__(self, origin_shape=(-1, -1)): 130 """Initialize CusFusedAbsMax1""" 131 self.init_prim_io_names(inputs=['x1'], outputs=['y']) 132 self.origin_shape = origin_shape 133 from mindspore.ops._op_impl._custom_op.fused_abs_max1_impl import cus_fused_abs_max1 134 135 def infer_shape(self, data1_shape): 136 ll = [] 137 if len(data1_shape) == 2: 138 ll = [1] 139 else: 140 ll = [32, 64] 141 return ll 142 143 def infer_dtype(self, data1_dtype): 144 return data1_dtype 145 146 147class CusImg2Col(PrimitiveWithInfer): 148 """ 149 Img2cols the feature map and the result in reorganized in NC1HWC0. 150 151 Args: 152 - **strides** (listInt) - the stride of the ops. 153 - **ksizes** (listInt) - the kernel size of the ops. 154 Inputs: 155 - **input_x** (Tensor) - The shape of the tensor is :math:`(N, C, H, W)`. 156 Outputs: 157 Tensor, the shape of the output tensor is :math:`(N * H_O * W_O, C1 * K_W * K_H * C0)`. 158 Examples: 159 >>> input_x = Tensor(np.ones(shape=[32, 3, 224, 224]), mindspore.float16) 160 >>> cusimg2col = ops.CusImg2Col() 161 >>> output = cusimg2col(input_x) 162 """ 163 164 @prim_attr_register 165 def __init__(self, ksizes, strides, dilates=(1, 1, 1, 1), mode="NC1HWC0"): 166 """Initialize CusImg2Col""" 167 self.init_prim_io_names(inputs=['x1'], outputs=['y']) 168 self.ksizes = ksizes 169 self.strides = strides 170 self.dilates = dilates 171 self.mode = mode 172 from mindspore.ops._op_impl._custom_op.img2col_impl import cus_img2col 173 174 def infer_shape(self, data1_shape): 175 bs, c, h, w = data1_shape 176 _, stride_h, stride_w, _ = self.strides 177 _, k_w, k_h, _ = self.ksizes 178 c0 = 16 179 c1 = c // 16 180 if c1 == 0: 181 c1 = 1 182 shape = [bs * int(h // stride_h) * int(w // stride_w), k_w * k_h * c1 * c0] 183 return shape 184 185 def infer_dtype(self, data1_dtype): 186 return data1_dtype 187 188 189class CusMatMulCubeDenseLeft(PrimitiveWithInfer): 190 """ 191 Multiplies matrix `a` by matrix `b`. 192 193 The rank of input_x1 must be `4`, the fractal format of the normal matrix. 194 The rank of input_x2 must be `2`. 195 196 Inputs: 197 - **input_x1** (Tensor) - The first tensor to be multiplied. 198 The shape of the tensor is :math:`(N0, M0, N1, M1)`. 199 - **input_x2** (Tensor) - The second tensor to be multiplied. The shape of the tensor is :math:`(M, C)`. 200 Outputs: 201 Tensor, the shape of the output tensor is :math:`(N, C)`. 202 Examples: 203 >>> input_x = Tensor(np.ones(shape=[16, 16, 16, 16]), mindspore.float16) 204 >>> input_y = Tensor(np.ones(shape=[256, 256]), mindspore.float16) 205 >>> matmulcubedenseleft = ops.CusMatMulCubeDenseLeft() 206 >>> output = matmulcubedenseleft(input_x, input_y) 207 """ 208 209 @prim_attr_register 210 def __init__(self): 211 """Initialize CusMatMulCubeDenseLeft""" 212 self.init_prim_io_names(inputs=['x1', 'x2'], outputs=['y']) 213 from mindspore.ops._op_impl._custom_op.matmul_cube_dense_left_impl import cus_matmul_cube_dense_left 214 215 def infer_shape(self, data1_shape, data2_shape): 216 return data2_shape 217 218 def infer_dtype(self, data1_dtype, data2_dtype): 219 return mstype.float16 220 221 222class CusMatMulCubeFraczRightMul(PrimitiveWithInfer): 223 """ 224 Multiplies matrix `a` by matrix `b` and muls the result by scalar `c`. 225 226 The rank of input_x1 tensors must be `2`. 227 The rank of input_x2 tensors must be `4`. 228 229 Inputs: 230 - **input_x1** (Tensor) - The first tensor to be multiplied. The shape of the tensor is :math:`(N, C)`. 231 - **input_x2** (Tensor) - The second tensor to be multiplied. 232 The shape of the tensor is :math:`(C1, M1, C0, M0)`. 233 - **input_x3** (Tensor) - The third tensor to be multiplied. The shape of the tensor if :math`(1, )`. 234 Outputs: 235 Tensor, the shape of the output tensor is :math:`(N, M)`. 236 Examples: 237 >>> input_x1 = Tensor(np.ones(shape=[256, 256]), mindspore.float16) 238 >>> input_x2 = Tensor(np.ones(shape=[16, 16, 16, 16]), mindspore.float16) 239 >>> input_x3 = Tensor(np.ones(shape=[1, ]), mindspore.float16) 240 >>> cusmatmulfraczrightmul = ops.CusMatMulCubeFraczRightMul() 241 >>> output = cusmatmulfraczrightmul(input_x1, input_x2, input_x3) 242 """ 243 244 @prim_attr_register 245 def __init__(self): 246 """Initialize CusMatMulCubeFraczRightMul""" 247 self.init_prim_io_names(inputs=['x1', 'x2', 'x3'], outputs=['y']) 248 from mindspore.ops._op_impl._custom_op.matmul_cube_fracz_right_mul_impl import cus_matmul_cube_fraczrightmul 249 250 def infer_shape(self, data1_shape, data2_shape, data3_shape): 251 return data1_shape 252 253 def infer_dtype(self, data1_dtype, data2_dtype, data3_dtype): 254 return mstype.float32 255 256 257class CusMatMulCube(PrimitiveWithInfer): 258 """ 259 Multiplies matrix `a` by matrix `b`. 260 261 The rank of input tensors must be `2`. 262 263 Args: 264 transpose_a (bool): If true, `a` is transposed before multiplication. Default: ``False``. 265 transpose_b (bool): If true, `b` is transposed before multiplication. Default: ``False``. 266 267 Inputs: 268 - **input_x** (Tensor) - The first tensor to be multiplied. The shape of the tensor is :math:`(N, C)`. If 269 `transpose_a` is True, its shape must be :math:`(N, C)` after transposing. 270 - **input_y** (Tensor) - The second tensor to be multiplied. The shape of the tensor is :math:`(C, M)`. If 271 `transpose_b` is True, its shape must be :math:`(C, M)` after transpose. 272 273 Outputs: 274 Tensor, the shape of the output tensor is :math:`(N, M)`. 275 276 Examples: 277 >>> input_x = Tensor(np.ones(shape=[256, 256]), mindspore.float16) 278 >>> input_y = Tensor(np.ones(shape=[256, 256]), mindspore.float16) 279 >>> cusmatmulcube = ops.CusMatMulCube() 280 >>> output = matmul(input_x, input_y) 281 """ 282 283 @prim_attr_register 284 def __init__(self, transpose_a=False, transpose_b=False): 285 """Initialize CusMatMulCube""" 286 self.init_prim_io_names(inputs=['x1', 'x2'], outputs=['y']) 287 self.transpose_a = transpose_a 288 self.transpose_b = transpose_b 289 from mindspore.ops._op_impl._custom_op.matmul_cube_impl import cus_matmul_cube 290 291 def infer_shape(self, data1_shape, data2_shape): 292 if self.transpose_a: 293 _, m = data1_shape 294 else: 295 m, _ = data1_shape 296 if self.transpose_b: 297 n, _ = data2_shape 298 else: 299 _, n = data2_shape 300 shape = [m, n] 301 return shape 302 303 def infer_dtype(self, data1_dtype, data2_dtype): 304 return mstype.float32 305 306 307class CusMatrixCombine(PrimitiveWithInfer): 308 """ 309 move the batch matrix to result matrix diag part. 310 The rank of input tensors must be `3`. 311 312 Inputs: 313 - **input_x** (Tensor) - The shape of the tensor is :math:`(N, D, D)`. 314 315 Outputs: 316 Tensor, the shape of the output tensor is :math:`(N * D, N * D)`. 317 318 Examples: 319 >>> input_x = Tensor(np.ones(shape=[2, 128, 128]), mindspore.float32) 320 >>> cusmatrixcombine = ops.CusMatrixCombine() 321 >>> output = cusmatrixcombine(input_x) 322 """ 323 324 @prim_attr_register 325 def __init__(self): 326 """Initialize CusMatrixCombine""" 327 self.init_prim_io_names(inputs=['x'], outputs=['y']) 328 from mindspore.ops._op_impl._custom_op.matrix_combine_impl import cus_matrix_combine 329 330 def infer_shape(self, data_shape): 331 a, b, c = data_shape 332 shape = [a * b, a * c] 333 334 return shape 335 336 def infer_dtype(self, data_dtype): 337 return data_dtype 338 339 340class CusTranspose02314(PrimitiveWithInfer): 341 """ 342 Permute input tensor with perm (0, 2, 3, 1, 4) 343 344 The rank of input tensors must be `5` with format NC1HWC0. 345 346 Inputs: 347 - **input_x** (Tensor) - The shape of the tensor is :math:`(N, C1, H, W, C0)`. 348 349 Outputs: 350 Tensor, the shape of the output tensor is :math:`(N, H, W, C1, C0)`. 351 352 Examples: 353 >>> input_x = Tensor(np.ones(shape=[32, 1, 224, 224, 16]), mindspore.float16) 354 >>> custranspose02314 = ops.CusTranspose02314() 355 >>> output = custranspose02314(input_x) 356 """ 357 358 @prim_attr_register 359 def __init__(self): 360 """Initialize CusTranspose02314""" 361 self.init_prim_io_names(inputs=['x1'], outputs=['y']) 362 from mindspore.ops._op_impl._custom_op.transpose02314_impl import cus_transpose02314 363 364 def get_bprop(self): 365 """Get backprop for CusTranspose02314.""" 366 367 def bprop(x, out, dout): 368 return (C.zeros_like(x),) 369 370 return bprop 371 372 def infer_shape(self, data1_shape): 373 n, c, h, w = data1_shape 374 c0 = 16 375 c1 = c // 16 376 shape = (n * h * w, c1 * c0) 377 return shape 378 379 def infer_dtype(self, data1_dtype): 380 return data1_dtype 381 382 383class CusMatMulCubeDenseRight(PrimitiveWithInfer): 384 """ 385 Multiplies matrix `a` by matrix `b`. 386 387 The rank of input_x1 tensor must be `2`. 388 The rank of input_x2 tensor must be `4`. 389 390 Inputs: 391 - **input_x** (Tensor) - The first tensor to be multiplied. The shape of the tensor is :math:`(N, C)`. 392 - **input_y** (Tensor) - The second tensor to be multiplied. 393 The shape of the tensor is :math:`(C1, M1, M0, C0)`. 394 395 Outputs: 396 Tensor, the shape of the output tensor is :math:`(N, M)`. 397 398 Examples: 399 >>> input_x = Tensor(np.ones(shape=[256, 256]), mindspore.float16) 400 >>> input_y = Tensor(np.ones(shape=[16, 16, 16, 16]), mindspore.float16) 401 >>> cusmatmulcubedenseright = ops.CusMatMulCubeDenseRight() 402 >>> output = cusmatmulcubedenseright(input_x, input_y) 403 """ 404 405 @prim_attr_register 406 def __init__(self): 407 """Initialize CusMatMulCubeDenseRight""" 408 self.init_prim_io_names(inputs=['x1', 'x2', 'x3'], outputs=['y']) 409 from mindspore.ops._op_impl._custom_op.matmul_cube_dense_right_impl import cus_matmul_cube_dense_right 410 411 def infer_shape(self, data1_shape, data2_shape, data3_shape): 412 return data1_shape 413 414 def infer_dtype(self, data1_dtype, data2_dtype, data3_dtype): 415 return mstype.float32 416 417 418class CusMatMulCubeFraczLeftCast(PrimitiveWithInfer): 419 """ 420 Multiplies matrix `a` by matrix `b`. 421 422 The rank of input_x1 tensor must be `4`. 423 The rank of input_x2 tensors must be `2`. 424 425 Inputs: 426 - **input_x1** (Tensor) - The first tensor to be multiplied. 427 The shape of the tensor is :math:`(C1, N1, N0, C0)`. 428 - **input_x2** (Tensor) - The second tensor to be multiplied. The shape of the tensor is :math:`(C, M)`. 429 430 Outputs: 431 Tensor, the shape of the output tensor is :math:`(N, M)`. 432 433 Examples: 434 >>> input_x = Tensor(np.ones(shape=[16, 16, 16, 16]), mindspore.float16) 435 >>> input_y = Tensor(np.ones(shape=[256, 256]), mindspore.float16) 436 >>> cusmatmulcubefraczleftcast = ops.CusMatMulCubeFraczLeftCast() 437 >>> output = cusmatmulcubefraczleftcast(input_x, input_y) 438 """ 439 440 @prim_attr_register 441 def __init__(self): 442 """Initialize CusMatMulCubeFraczLeftCast""" 443 self.init_prim_io_names(inputs=['x1', 'x2'], outputs=['y']) 444 from mindspore.ops._op_impl._custom_op.matmul_cube_fracz_left_cast_impl import cus_matmul_cube_fraczleftcast 445 446 def infer_shape(self, data1_shape, data2_shape): 447 return data2_shape 448 449 def infer_dtype(self, data1_dtype, data2_dtype): 450 return mstype.float16 451 452 453class ThorIm2Col(PrimitiveWithInfer): 454 """ 455 extracts image paths from image. 456 457 The rank of input_x1 must be `4`, data_format is "NCHW". 458 459 Inputs: 460 - **input_x1** (Tensor) - The feature map. 461 The shape of the tensor is :math:`(N, C, H, W)`. 462 Outputs: 463 Tensor. 464 Examples: 465 >>> input_x = Tensor(np.random.rand(32, 3, 224, 224).astype(np.float16)) 466 >>> img2col = ops.CusMatMulCubeDenseLeft(kernel_size=7, pad=3, stride=2) 467 >>> output = img2col(input_x) 468 """ 469 470 @prim_attr_register 471 def __init__(self, 472 kernel_size, 473 pad_mode="valid", 474 pad=0, 475 stride=1, 476 dilation=1): 477 """Initialize ThorIm2Col""" 478 self.init_prim_io_names(inputs=['x'], outputs=['output']) 479 self.kernel_size = _check_positive_int_or_tuple('kernel_size', kernel_size, self.name) 480 self.add_prim_attr('kernel_size', self.kernel_size) 481 self.stride = _check_positive_int_or_tuple('stride', stride, self.name, allow_four=True, ret_four=True) 482 self.add_prim_attr('stride', self.stride) 483 self.dilation = _check_positive_int_or_tuple('dilation', dilation, self.name, allow_four=True, ret_four=True) 484 self.add_prim_attr('dilation', self.dilation) 485 validator.check_value_type('pad', pad, (int,), self.name) 486 self.pad_mode = validator.check_string(pad_mode, ['valid', 'same', 'pad'], 'pad_mode', self.name) 487 self.pad = validator.check_pad_value_by_mode(pad_mode, pad, self.name) 488 if self.pad_mode == 'pad': 489 validator.check_non_negative_int(self.pad, 'pad', self.name) 490 self.add_prim_attr('data_format', "NCHW") 491 492 def infer_shape(self, x_shape): 493 validator.check_equal_int(len(x_shape), 4, "x rank", self.name) 494 kernel_size_h = self.kernel_size[0] 495 kernel_size_w = self.kernel_size[1] 496 stride_h = self.stride[2] 497 stride_w = self.stride[3] 498 dilation_h = self.dilation[2] 499 dilation_w = self.dilation[3] 500 if self.pad_mode == "valid": 501 h_out = math.ceil((x_shape[2] - dilation_h * (kernel_size_h - 1)) / stride_h) 502 w_out = math.ceil((x_shape[3] - dilation_w * (kernel_size_w - 1)) / stride_w) 503 pad_top, pad_bottom, pad_left, pad_right = 0, 0, 0, 0 504 elif self.pad_mode == "same": 505 h_out = math.ceil(x_shape[2] / stride_h) 506 w_out = math.ceil(x_shape[3] / stride_w) 507 pad_needed_h = max(0, (h_out - 1) * stride_h + dilation_h * (kernel_size_h - 1) + 1 - x_shape[2]) 508 pad_top = math.floor(pad_needed_h / 2) 509 pad_bottom = pad_needed_h - pad_top 510 pad_needed_w = max(0, (w_out - 1) * stride_w + dilation_w * (kernel_size_w - 1) + 1 - x_shape[3]) 511 pad_left = math.floor(pad_needed_w / 2) 512 pad_right = pad_needed_w - pad_left 513 elif self.pad_mode == 'pad': 514 pad_top, pad_bottom, pad_left, pad_right = self.pad, self.pad, self.pad, self.pad 515 h_out = 1 + (x_shape[2] + 2 * self.pad - kernel_size_h - (kernel_size_h - 1) * (dilation_h - 1)) / stride_h 516 w_out = 1 + (x_shape[3] + 2 * self.pad - kernel_size_w - (kernel_size_w - 1) * (dilation_w - 1)) / stride_w 517 h_out = math.floor(h_out) 518 w_out = math.floor(w_out) 519 self.pad_list = [pad_top, pad_bottom, pad_left, pad_right] 520 self.add_prim_attr('pad_list', (pad_top, pad_bottom, pad_left, pad_right)) 521 batch_size = x_shape[0] 522 channel = x_shape[1] 523 k_h = kernel_size_h 524 k_w = kernel_size_w 525 out_shape = [channel, k_h, k_w, batch_size, h_out, w_out] 526 return out_shape 527 528 def infer_dtype(self, x_dtype): 529 valid_dtypes = [mstype.float16, mstype.float32] 530 validator.check_tensor_dtype_valid('x', x_dtype, valid_dtypes, self.name) 531 return x_dtype 532 533 534class NewIm2Col(PrimitiveWithInfer): 535 """ 536 extracts image paths from image by using TBE. 537 538 The rank of input_x1 must be `4`, data_format is "NCHW". 539 540 Inputs: 541 - **input_x1** (Tensor) - The feature map. 542 The shape of the tensor is :math:`(N, C, H, W)`. 543 Outputs: 544 Tensor. The shape of the tensor is :math:`(N, H, W, C)`. 545 546 Examples: 547 >>> input_x = Tensor(np.random.rand(32, 3, 224, 224).astype(np.float16)) 548 >>> im2col = ops.NewIm2Col(ksizes=(7,7), strides=2) 549 >>> output = im2col(input_x) 550 """ 551 552 @prim_attr_register 553 def __init__(self, 554 ksizes, 555 padding_mode="SAME", 556 strides=1, 557 dilations=1, 558 pads=0): 559 """Initialize NewIm2Col""" 560 self.init_prim_io_names(inputs=['x'], outputs=['output']) 561 self.ksizes = ksizes 562 self.strides = strides 563 self.add_prim_attr('ksizes', self.ksizes) 564 self.add_prim_attr('strides', self.strides) 565 self.dilations = dilations 566 self.add_prim_attr('dilations', self.dilations) 567 self.padding_mode = validator.check_string(padding_mode, ['VALID', 'SAME'], 'padding_mode', self.name) 568 self.add_prim_attr('data_format', "NCHW") 569 self.pads = pads 570 571 def infer_shape(self, x_shape): 572 "infer shape" 573 validator.check_equal_int(len(x_shape), 4, "x rank", self.name) 574 kernel_size_h = self.ksizes[0] 575 kernel_size_w = self.ksizes[1] 576 stride_h = self.strides 577 stride_w = self.strides 578 dilation_h = self.dilations 579 dilation_w = self.dilations 580 if self.padding_mode == "VALID": 581 h_out = math.ceil((x_shape[2] - dilation_h * (kernel_size_h - 1)) / stride_h) 582 w_out = math.ceil((x_shape[3] - dilation_w * (kernel_size_w - 1)) / stride_w) 583 pad_top, pad_bottom, pad_left, pad_right = 0, 0, 0, 0 584 elif self.padding_mode == "SAME": 585 h_out = math.ceil(x_shape[2] / stride_h) 586 w_out = math.ceil(x_shape[3] / stride_w) 587 pad_needed_h = max(0, (h_out - 1) * stride_h + dilation_h * (kernel_size_h - 1) + 1 - x_shape[2]) 588 pad_top = math.floor(pad_needed_h / 2) 589 pad_bottom = pad_needed_h - pad_top 590 pad_needed_w = max(0, (w_out - 1) * stride_w + dilation_w * (kernel_size_w - 1) + 1 - x_shape[3]) 591 pad_left = math.floor(pad_needed_w / 2) 592 pad_right = pad_needed_w - pad_left 593 self.pad_list = [pad_top, pad_bottom, pad_left, pad_right] 594 self.add_prim_attr('pad_list', (pad_top, pad_bottom, pad_left, pad_right)) 595 batch_size = x_shape[0] 596 channel = x_shape[1] 597 k_h = kernel_size_h 598 k_w = kernel_size_w 599 out_shape = [batch_size, h_out, w_out, channel * k_h * k_w] 600 return out_shape 601 602 def infer_dtype(self, x_dtype): 603 "infer dtype" 604 valid_dtypes = [mstype.float16, mstype.int8] 605 validator.check_tensor_dtype_valid('x', x_dtype, valid_dtypes, self.name) 606 return x_dtype 607 608 609class LoadIm2Col(PrimitiveWithInfer): 610 """ 611 extracts image patches from image. 612 613 The rank of input_x1 must be `4`, data_format is "NCHW". 614 Only supports when C is divisible by 16. 615 616 Inputs: 617 - **input_x1** (Tensor) - The feature map. 618 The shape of the tensor is :math:`(N, C, H, W)`. 619 Outputs: 620 Tensor. 621 Examples: 622 >>> input_x = Tensor(np.random.rand(32, 16, 224, 224).astype(np.float16)) 623 >>> img2col = ops.LoadIm2Col(kernel_size=(7,7), stride=(2,2)) 624 >>> output = img2col(input_x) 625 """ 626 627 @prim_attr_register 628 def __init__(self, 629 ksizes, 630 strides, 631 pad_mode="same", 632 dilates=(1, 1, 1, 1)): 633 """Initialize LoadIm2Col""" 634 635 self.init_prim_io_names(inputs=['x1'], outputs=['y']) 636 self.ksizes = ksizes 637 self.strides = strides 638 self.pad_mode = validator.check_string(pad_mode, ['same'], 'pad_mode', self.name) 639 self.dilation = dilates 640 641 def infer_shape(self, data1_shape): 642 bs, c, h, w = data1_shape 643 stride_h, stride_w = self.strides 644 k_w, k_h = self.ksizes 645 h_out = math.ceil(h / stride_h) 646 w_out = math.ceil(w / stride_w) 647 m = h_out * w_out 648 if m % 16 != 0: 649 shape = [(bs * m) // 16, (c * k_h * k_w) // 16, 16, 16] 650 else: 651 shape = [bs, m // 16, (c * k_h * k_w) // 16, 16, 16] 652 return shape 653 654 def infer_dtype(self, data1_dtype): 655 return data1_dtype 656 657 658class UpdateThorGradient(PrimitiveWithInfer): 659 """ 660 Updates Thor Gradient with Approximate Fisher info matrix(for GPU backend). 661 662 The rank of input_x1 must be `3`, which indicates the A matrix. 663 The rank of input_x2 must be `2`, which indicates the 1st-order gradient. 664 The rank of input_x3 must be `4`, which indicates the G matrix. 665 666 Inputs: 667 - **input_x1** (Tensor) - The first input is the diag part of the cov matrix of feature map. 668 Supported dtype [float32]. 669 - **input_x2** (Tensor) - The second input is the corresponding 1st-order grad. Supported dtype [float32]. 670 - **input_x3** (Tensor) - The third input is the diag part of the cov matrix of dout. 671 Supported dtype [float32]. 672 673 Outputs: 674 Tensor, the shape is the same as the shape of input_x2, it will be used to update the weights. 675 676 Examples: 677 >>> input_x1 = Tensor(np.random.rand(16, 128, 128).astype(np.float32)) 678 >>> input_x2 = Tensor(np.random.rand(2048, 1024).astype(np.float32)) 679 >>> temp_x3 = np.random.rand(8, 128, 128).astype(np.float32) 680 >>> input_x3 = np.zeros(16,8,128,128).astype(np.float32) 681 >>> for i in range(16): 682 ... input_x3[i,:,:,:] = temp_x3 683 >>> input_x3 = Tensor(input_x3) 684 >>> update_thor_gradient = ops.UpdateThorGradient(split_dim=128) 685 >>> output = update_thor_gradient(input_x1, input_x2, input_x3) 686 """ 687 688 @prim_attr_register 689 def __init__(self, split_dim=1): 690 """Initialize UpdateThorGradient""" 691 self.init_prim_io_names(inputs=['x1', 'x2', 'x3'], outputs=['y']) 692 self.split_dim = split_dim 693 self.add_prim_attr('split_dim', self.split_dim) 694 695 def infer_shape(self, x1_shape, x2_shape, x3_shape): 696 return x2_shape 697 698 def infer_dtype(self, x1_dtype, x2_dtype, x3_dtype): 699 validator.check_tensors_dtypes_same_and_valid( 700 {'x1_dtype': x1_dtype, 'x2_dtype': x2_dtype, 'x3_dtype': x3_dtype}, 701 [mstype.float32], self.name) 702 return x2_dtype 703 704 705class _Cholesky(PrimitiveWithInfer): 706 """ 707 Inner API for _Cholesky base class. 708 """ 709 710 @prim_attr_register 711 def __init__(self, lower=False, clean=True, split_dim=0): 712 self.init_prim_io_names(inputs=['x1'], outputs=['y']) 713 self.lower = validator.check_value_type("lower", lower, [bool], self.lower) 714 self.clean = validator.check_value_type("clean", clean, [bool], self.clean) 715 self.lower = lower 716 self.add_prim_attr('lower', self.lower) 717 self.clean = clean 718 self.add_prim_attr('clean', self.clean) 719 self.split_dim = split_dim 720 self.add_prim_attr('split_dim', self.split_dim) 721 722 def infer_shape(self, x1_shape): 723 if self.split_dim != 0: 724 height = x1_shape[0] 725 width = x1_shape[1] 726 if height <= self.split_dim: 727 out_shape = [1, height, width] 728 else: 729 batch = height // self.split_dim 730 if height != batch * self.split_dim: 731 batch += 1 732 out_shape = [batch, self.split_dim, self.split_dim] 733 else: 734 out_shape = x1_shape 735 return out_shape 736 737 def infer_dtype(self, x1_dtype): 738 validator.check_tensor_dtype_valid('x1', x1_dtype, [mstype.float32, mstype.float64], self.name) 739 return x1_dtype 740 741 742class Cholesky(_Cholesky): 743 """ 744 Inner API for positive-definite matrix Cholesky decomposition GPU backend. 745 """ 746 747 748class CholeskyTrsm(_Cholesky): 749 """ 750 Inner API for resnet50 THOR GPU backend. 751 """ 752 753 754class DetTriangle(PrimitiveWithInfer): 755 """ 756 Calculate the determinant of triangle matrices. 757 758 Args: 759 fill_mode (tuple): The target shape to broadcast. 760 761 Inputs: 762 - **input_x** (Tensor) - The input tensor. 763 764 Outputs: 765 Tensor, with the given `shape` and the same data type as `input_x`. 766 767 Examples: 768 >>> shape = (2, 3) 769 >>> input_x = Tensor(np.array([1, 2, 3]).astype(np.float32)) 770 >>> broadcast_to = P.BroadcastTo(shape) 771 >>> broadcast_to(input_x) 772 [[1.0, 2.0, 3.0], [1.0, 2.0, 3.0]] 773 """ 774 775 @prim_attr_register 776 def __init__(self, fill_mode=0): 777 self.init_prim_io_names(inputs=['x1'], outputs=['y']) 778 self.fill_mode = fill_mode 779 self.add_prim_attr('fill_mode', self.fill_mode) 780 781 def infer_shape(self, x1_shape): 782 out_shape = x1_shape 783 del out_shape[-2:] 784 return out_shape 785 786 def infer_dtype(self, x1_dtype): 787 validator.check_tensor_dtype_valid('x1', x1_dtype, [mstype.float32], self.name) 788 return x1_dtype 789 790 791class ProdForceSeA(PrimitiveWithInfer): 792 """ 793 ProdForceSeA. 794 """ 795 796 @prim_attr_register 797 def __init__(self, natoms=192): 798 self.init_prim_io_names(inputs=['net_deriv_tensor', "in_deriv_tensor", "nlist_tensor"], outputs=['y']) 799 self.natoms = natoms 800 self.add_prim_attr('natoms', self.natoms) 801 802 def infer_shape(self, x1_shape, x2_shape, x3_shape): 803 out_shape = [x3_shape[0], x3_shape[1], 3] 804 return out_shape 805 806 def infer_dtype(self, x1_dtype, x2_dtype, x3_dtype): 807 return x1_dtype 808