Huawei_Technology
/
MindSpore-AKG
Not watched
1
0
Fork 0
Code
Releases 1 Wiki evaluate Activity
Issues 0 Pull Requests 0 Datasets Model Cloudbrain HPC
You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
Branch: master
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'master'
${ noResults }
MindSpore-AKG/python/akg/ops/nn/conv_backprop_input.py

408 lines
22 kB
Raw Permalink Blame History 

  1. #!/usr/bin/env python3

  2. # coding: utf-8

  3. # Copyright 2019 Huawei Technologies Co., Ltd

  4. #

  5. # Licensed under the Apache License, Version 2.0 (the "License");

  6. # you may not use this file except in compliance with the License.

  7. # You may obtain a copy of the License at

  8. #

  9. # http://www.apache.org/licenses/LICENSE-2.0

  10. #

  11. # Unless required by applicable law or agreed to in writing, software

  12. # distributed under the License is distributed on an "AS IS" BASIS,

  13. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.

  14. # See the License for the specific language governing permissions and

  15. # limitations under the License.

  16. 

  17. """operator dsl function: conv_backprop_input"""

  18. import logging

  19. import akg.tvm

  20. import akg

  21. import akg.lang.cce

  22. from akg import dim

  23. from akg.utils import validation_check as vc_util

  24. from akg.ops.math import cast

  25. 

  26. conv_backprop_input_tiling_args = {

  27.     str(((1, 1024, 14, 14), (2048, 1024, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1))): [2, 32, 64, 96, 128],

  28.     str(((1, 1024, 14, 14), (256, 1024, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [14, 256, 208, 64, 112],

  29.     str(((1, 1024, 14, 14), (512, 1024, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [4, 128, 48, 352, 16, 14],

  30.     str(((1, 1024, 14, 14), (512, 1024, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1))): [14, 512, 49, 32, 512],

  31.     str(((1, 128, 28, 28), (128, 128, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [28, 128, 128, 144, 128],

  32.     str(((1, 128, 28, 28), (512, 128, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [28, 128, 784, 16, 32],

  33.     str(((1, 128, 56, 56), (128, 128, 3, 3), (0, 1, 0, 1), (2, 2), (1, 1))): [6, 32, 112, 160, 32, 58],

  34.     str(((1, 16, 224, 224), (64, 16, 7, 7), (3, 3, 3, 3), (2, 2), (1, 1))): [10, 16, 16, 49 * 16, 16, 10],

  35.     str(((1, 2048, 7, 7), (512, 2048, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [7, 512, 49, 32, 512],

  36.     str(((1, 256, 13, 13), (384, 256, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [13, 64, 80, 48, 16, 15],

  37.     str(((1, 256, 14, 14), (1024, 256, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [14, 256, 112, 32, 240],

  38.     str(((1, 256, 14, 14), (256, 256, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [14, 128, 196, 144, 128],

  39.     str(((1, 256, 28, 28), (256, 256, 3, 3), (0, 1, 0, 1), (2, 2), (1, 1))): [9, 16, 48, 448, 16, 30],

  40.     str(((1, 256, 28, 28), (256, 256, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [14, 128, 196, 144, 128],

  41.     str(((1, 256, 56, 56), (128, 256, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [8, 128, 240, 128, 128, 56],

  42.     str(((1, 256, 56, 56), (128, 256, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1))): [7, 128, 252, 64, 128],

  43.     str(((1, 256, 56, 56), (512, 256, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1))): [3, 32, 32, 32, 32],

  44.     str(((1, 256, 56, 56), (64, 256, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [16, 64, 280, 16, 64],

  45.     str(((1, 3, 224, 224), (64, 3, 7, 7), (3, 3, 3, 3), (2, 2), (1, 1))): [10, 16, 16, 49 * 16, 16, 10],

  46.     str(((1, 384, 13, 13), (256, 384, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [12, 192, 16, 240, 96, 15],

  47.     str(((1, 384, 13, 13), (384, 384, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [9, 128, 96, 176, 80, 15],

  48.     str(((1, 512, 14, 14), (512, 512, 3, 3), (0, 1, 0, 1), (2, 2), (1, 1))): [10, 16, 80, 64, 16, 16],

  49.     str(((1, 512, 28, 28), (1024, 512, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1))): [2, 64, 112, 32, 512],

  50.     str(((1, 512, 28, 28), (128, 512, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [14, 128, 448, 16, 64],

  51.     str(((1, 512, 28, 28), (256, 512, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [10, 256, 128, 32, 256, 28],

  52.     str(((1, 512, 28, 28), (256, 512, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1))): [7, 256, 98, 64, 256],

  53.     str(((1, 512, 7, 7), (2048, 512, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [7, 128, 49, 256, 128],

  54.     str(((1, 512, 7, 7), (512, 512, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [7, 64, 49, 144, 64],

  55.     str(((1, 6, 14, 14), (16, 6, 5, 5), (0, 0, 0, 0), (1, 1), (1, 1))): [18, 16, 64, 240, 16, 18],

  56.     str(((1, 64, 56, 56), (256, 64, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [14, 256, 784, 16, 32],

  57.     str(((1, 64, 56, 56), (64, 64, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [56, 64, 784, 16, 32],

  58.     str(((1, 64, 56, 56), (64, 64, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [56, 64, 128, 144, 64],

  59.     str(((1, 96, 28, 28), (256, 96, 5, 5), (2, 2, 2, 2), (1, 1), (1, 1))): [14, 48, 32, 384, 48, 32],

  60.     str(((32, 1024, 14, 14), (2048, 1024, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1))): [2, 224, 32, 32, 144, 14],

  61.     str(((32, 1024, 14, 14), (256, 1024, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [14, 224, 192, 64, 48, 14],

  62.     str(((32, 1024, 14, 14), (512, 1024, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [7, 352, 96, 80, 176, 14],

  63.     str(((32, 1024, 14, 14), (512, 1024, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1))): [14, 512, 49, 32, 512],

  64.     str(((32, 128, 28, 28), (128, 128, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [18, 64, 208, 144, 64, 30],

  65.     str(((32, 128, 28, 28), (512, 128, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [4, 384, 112, 16, 336, 28],

  66.     str(((32, 128, 56, 56), (128, 128, 3, 3), (0, 1, 0, 1), (2, 2), (1, 1))): [6, 112, 112, 144, 112, 58],

  67.     str(((32, 16, 224, 224), (64, 16, 7, 7), (3, 3, 3, 3), (2, 2), (1, 1))): [10, 16, 16, 49 * 16, 16, 10],

  68.     str(((32, 2048, 7, 7), (512, 2048, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [7, 32, 48, 272, 32, 7],

  69.     str(((32, 256, 13, 13), (384, 256, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [13, 64, 80, 48, 16, 15],

  70.     str(((32, 256, 14, 14), (1024, 256, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [2, 416, 32, 752, 16, 14],

  71.     str(((32, 256, 14, 14), (256, 256, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [16, 112, 112, 144, 112, 14],

  72.     str(((32, 256, 28, 28), (256, 256, 3, 3), (0, 1, 0, 1), (2, 2), (1, 1))): [6, 144, 112, 144, 112, 30],

  73.     str(((32, 256, 28, 28), (256, 256, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [14, 128, 196, 144, 128],

  74.     str(((32, 256, 56, 56), (128, 256, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [4, 128, 224, 64, 112, 56],

  75.     str(((32, 256, 56, 56), (128, 256, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1))): [8, 128, 224, 96, 48, 56],

  76.     str(((32, 256, 56, 56), (512, 256, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1))): [2, 288, 112, 144, 32, 56],

  77.     str(((32, 256, 56, 56), (64, 256, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [8, 64, 448, 64, 32, 56],

  78.     str(((32, 3, 224, 224), (64, 3, 7, 7), (2, 3, 2, 3), (2, 2), (1, 1))): [13, 16, 16, 49 * 16, 16, 13],

  79.     str(((32, 3, 224, 224), (64, 3, 7, 7), (3, 3, 3, 3), (2, 2), (1, 1))): [10, 16, 16, 49 * 16, 16, 10],

  80.     str(((32, 3, 32, 32), (6, 3, 5, 5), (0, 0, 0, 0), (1, 1), (1, 1))): [16, 16, 16, 16, 16, 16],

  81.     str(((32, 384, 13, 13), (256, 384, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [12, 192, 16, 240, 96, 15],

  82.     str(((32, 384, 13, 13), (384, 384, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [9, 128, 96, 176, 80, 15],

  83.     str(((32, 512, 14, 14), (512, 512, 3, 3), (0, 1, 0, 1), (2, 2), (1, 1))): [10, 96, 112, 144, 48, 16],

  84.     str(((32, 512, 28, 28), (1024, 512, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1))): [2, 336, 64, 80, 208, 28],

  85.     str(((32, 512, 28, 28), (128, 512, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [4, 64, 112, 80, 64, 28],

  86.     str(((32, 512, 28, 28), (256, 512, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [8, 224, 112, 64, 96, 28],

  87.     str(((32, 512, 28, 28), (256, 512, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1))): [7, 192, 96, 48, 192, 28],

  88.     str(((32, 512, 7, 7), (2048, 512, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [7, 128, 49, 256, 128],

  89.     str(((32, 512, 7, 7), (512, 512, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [9, 80, 64, 144, 80, 9],

  90.     str(((32, 6, 14, 14), (16, 6, 5, 5), (0, 0, 0, 0), (1, 1), (1, 1))): [18, 16, 64, 240, 16, 18],

  91.     str(((32, 64, 56, 56), (256, 64, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [4, 224, 112, 224, 80, 56],

  92.     str(((32, 64, 56, 56), (64, 64, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [10, 64, 336, 16, 16, 56],

  93.     str(((32, 64, 56, 56), (64, 64, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [15, 64, 112, 144, 64, 58],

  94.     str(((32, 96, 27, 27), (256, 96, 5, 5), (2, 2, 2, 2), (1, 1), (1, 1))): [7, 32, 80, 48, 32, 31],

  95.     str(((32, 96, 28, 28), (256, 96, 5, 5), (2, 2, 2, 2), (1, 1), (1, 1))): [14, 48, 32, 384, 48, 32],

  96. }

  97. cast_tiling_args = {

  98.     str(((1, 1024, 14, 14), (2048, 1024, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1))): [2, 16, 64, 96, 128],

  99.     str(((1, 1024, 14, 14), (256, 1024, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [14, 256 // 2, 208, 64, 112],

  100.     str(((1, 1024, 14, 14), (512, 1024, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1))): [14, 16, 50, 32, 512],

  101.     str(((1, 128, 28, 28), (128, 128, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [28, 128 // 4, 128, 144, 128],

  102.     str(((1, 128, 28, 28), (512, 128, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [28, 128 // 2, 784, 16, 32],

  103.     str(((1, 2048, 7, 7), (512, 2048, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [7, 512 // 8, 49, 32, 512],

  104.     str(((1, 256, 14, 14), (1024, 256, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [14, 256 // 8, 112, 32, 240],

  105.     str(((1, 256, 14, 14), (256, 256, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [14, 128 // 8, 196, 144, 128],

  106.     str(((1, 256, 56, 56), (128, 256, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1))): [7, 128, 252, 64, 128],

  107.     str(((1, 256, 56, 56), (64, 256, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [16, 64, 280, 16, 64],

  108.     str(((1, 3, 224, 224), (64, 3, 7, 7), (3, 3, 3, 3), (2, 2), (1, 1))): [10, 16, 16, 49 * 16, 16, 10],

  109.     str(((1, 16, 224, 224), (64, 16, 7, 7), (3, 3, 3, 3), (2, 2), (1, 1))): [10, 16, 16, 49 * 16, 16, 10],

  110.     str(((1, 512, 28, 28), (128, 512, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [14, 128, 448, 16, 64],

  111.     str(((1, 512, 28, 28), (256, 512, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1))): [7, 256 // 4, 98, 64, 256],

  112.     str(((1, 512, 7, 7), (2048, 512, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [7, 128 // 8, 49, 256, 128],

  113.     str(((1, 512, 7, 7), (512, 512, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [7, 64 // 4, 49, 144, 64],

  114.     str(((1, 64, 56, 56), (256, 64, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [14, 256 // 8, 784, 16, 32],

  115.     str(((1, 64, 56, 56), (64, 64, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1))): [56, 64, 784, 16, 32],

  116.     str(((1, 64, 56, 56), (64, 64, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1))): [56, 64, 128, 144, 64],

  117.     str(((1, 256, 56, 56), (512, 256, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1))): [3, 16, 32, 32, 32],

  118.     str(((1, 512, 28, 28), (1024, 512, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1))): [2, 16, 112, 32, 512],

  119. }

  120. 

  121. 

  122. def gen_key(fmap_shape, filter_shape, pad_, stride_, dilation_):

  123.     """generate key."""

  124.     key = str((tuple(fmap_shape), tuple(filter_shape), tuple(pad_), tuple(stride_), tuple(dilation_)))

  125.     return key

  126. 

  127. 

  128. def conv_backprop_input_compute(data, output_shape, filter_shape, input_shape, pad_, stride_,

  129.                                 block_size=16, attrs=None, key=None):

  130.     """core computation of conv_backprop_input."""

  131.     _, in_c, w_h, w_w = filter_shape

  132. 

  133.     # stride (stride_h, stride_w)

  134.     stride_h, stride_w = stride_

  135.     if stride_h != stride_w:

  136.         raise ValueError("stride_h must be equal to stride_w.")

  137. 

  138.     # output shape (NCHW -> NC1HWC0)

  139.     in_nn, in_cc, in_hh, in_ww = output_shape

  140.     if in_c % block_size != 0:

  141.         raise ValueError("in_c must be divided by block_size.")

  142.     input_shape_nc1hwc0 = (in_nn, in_cc // block_size, in_hh, in_ww, block_size)

  143.     in_nn, _, in_hh, in_ww, _ = input_shape_nc1hwc0

  144.     input_trans_shape_nc1hwc0 = (in_nn, in_cc // block_size, in_hh * stride_h, in_ww * stride_w, block_size)

  145.     in_n, in_c1, in_h, in_w, _ = input_trans_shape_nc1hwc0

  146. 

  147.     # kernel shape (NCHW -> NC1HWC0 -> Fractal)

  148.     k_n, k_c, k_h, k_w = filter_shape

  149.     if k_c % block_size != 0:

  150.         raise ValueError("k_c must be divided by block_size.")

  151.     kernel_shape_nc1hwc0 = (k_n, k_c // block_size, k_h, k_w, block_size)

  152.     k_n, k_c1, k_h, k_w, k_c0 = kernel_shape_nc1hwc0

  153.     kernel_shape_trans = (k_n // block_size * k_h * k_w, k_c // block_size, block_size, block_size)

  154.     k_c1 = k_n // block_size

  155.     k_n = k_c

  156. 

  157.     _, _, input_h, input_w = input_shape

  158. 

  159.     # padding ((padding_h, padding_w) -> (padding_top, padding_bottom, padding_left, padding_right))

  160.     padding = (pad_[0], pad_[1], pad_[2], pad_[3])

  161.     pad_t, pad_b, pad_l, pad_r = padding

  162. 

  163.     # padHT -> padHT'

  164.     p_top = k_h - pad_t - 1

  165.     # padHB -> padHB'

  166.     p_bottom = input_h + pad_t - stride_h * ((input_h + pad_t + pad_b - k_h) // stride_h + 1)

  167.     # padWL -> padWL'

  168.     p_left = k_w - pad_l - 1

  169.     # padWR -> padWR'

  170.     p_right = input_w + pad_l - stride_w * ((input_w + pad_l + pad_r - k_w) // stride_w + 1)

  171. 

  172.     s_h = 1

  173.     s_w = 1

  174. 

  175.     # NC1HWCO

  176.     a_value = data[0]

  177. 

  178.     if data[1].dtype == 'float32':

  179.         b_value = cast.cast(data[1], 'float16')

  180.         tiling_args = cast_tiling_args

  181.     else:

  182.         b_value = data[1]

  183.         tiling_args = conv_backprop_input_tiling_args

  184. 

  185.     # Create reduction variables

  186.     kc1 = akg.tvm.reduce_axis((0, k_c1), name='kc1')

  187.     kh = akg.tvm.reduce_axis((0, k_h), name='kh')

  188.     kw = akg.tvm.reduce_axis((0, k_w), name='kw')

  189.     kc0 = akg.tvm.reduce_axis((0, k_c0), name='kc0')

  190.     use_auto_tiling = False

  191.     if attrs is not None and 'conv_tile' in attrs and len(attrs['conv_tile']) >= 5:

  192.         tile_value = attrs['conv_tile']

  193.     elif key in tiling_args:

  194.         tile_value = tiling_args[key]

  195.     else:

  196.         use_auto_tiling = True

  197. 

  198.     out_h = (in_h + p_top + p_bottom - k_h) // (s_h) + 1

  199.     out_w = (in_w + p_left + p_right - k_w) // (s_w) + 1

  200.     out_shape_nc1hwc0 = (in_n, k_n // block_size, out_h, out_w, block_size)

  201.     out_n, out_c1, out_h, out_w, out_c0 = out_shape_nc1hwc0

  202. 

  203.     # set dim

  204.     info = dim.Dim()

  205.     index_ = 0

  206. 

  207.     if not use_auto_tiling:

  208.         tile_hh = tile_value[0]

  209.         if tile_hh == input_h:

  210.             tile_hh += pad_t + pad_b

  211. 

  212.         tile_coco = tile_value[1]

  213.         tile_coco = (tile_coco + block_size - 1) // block_size * block_size

  214. 

  215.         tile_mm = tile_value[2]

  216.         tile_mm = (tile_mm + block_size - 1) // block_size * block_size

  217. 

  218.         tile_kk = tile_value[3]

  219.         if not tile_kk % (block_size * w_h * w_w) == 0:

  220.             logging.warning("Warning: tile_k must be a multiple of (block_size * w_h * w_w)")

  221.         tile_kk = (tile_kk + block_size * w_h * w_w - 1) // (block_size * w_h * w_w) * (block_size * w_h * w_w)

  222. 

  223.         tile_nn = tile_value[4]

  224.         tile_nn = (tile_nn + block_size - 1) // block_size * block_size

  225. 

  226.         tile_ww = input_w

  227.         if len(tile_value) >= 6 and tile_value[5] > 0:

  228.             tile_ww = tile_value[5]

  229.         if tile_ww == input_w:

  230.             tile_ww += pad_l + pad_r

  231. 

  232.         if tile_hh == in_h:

  233.             tile_hh += p_top + p_bottom

  234.         tile_out_h = (tile_hh - k_h) // s_h + 1

  235. 

  236.         if tile_ww == in_w:

  237.             tile_ww += p_left + p_right

  238.         tile_out_w = (tile_ww - k_w) // s_w + 1

  239. 

  240.         if tile_coco > 0:

  241.             c1_cut = tile_coco // block_size

  242.         else:

  243.             c1_cut = out_c1

  244. 

  245.         if out_n > 1:

  246.             info.setdim(index=index_, axis=0, tilel1=1, tilel0=0)  # n

  247.         if out_c1 > 1:

  248.             info.setdim(index=index_, axis=1, tilel1=c1_cut, tilel0=0)  # c1

  249.         if out_h > 1:

  250.             info.setdim(index=index_, axis="H", tilel1=tile_out_h, tilel0=0)  # h

  251.         if out_w > 1:

  252.             info.setdim(index=index_, axis="W", tilel1=tile_out_w, tilel0=0)  # w

  253.         if out_c0 > 1:

  254.             info.setdim(index=index_, axis=4, tilel1=out_c0, tilel0=0)  # c0

  255.         if in_c1 > 1:

  256.             info.setdim(index=index_, axis=5, tilel1=in_c1, tilel0=0)  # kc1

  257.         if k_h > 1:

  258.             info.setdim(index=index_, axis=5, tilel1=k_h, tilel0=0)  # kh

  259.         if k_w > 1:

  260.             info.setdim(index=index_, axis=5, tilel1=k_w, tilel0=0)  # kw

  261. 

  262.         info = str(info)

  263.     else:

  264.         info = ""

  265.     # Compute the convolution below

  266. 

  267.     output_name = "output0"

  268. 

  269.     # weight_trans [ ko, no, ni, ki ]

  270.     # weight_trans [ co_1, kh, kw, ci_1, ci_0, co_0 ]

  271.     # kw = ko % k_w

  272.     # kh = ko // k_w % k_h

  273.     # co_1 = ko // k_w // k_h

  274.     # ci_1 = no

  275.     # -->

  276.     # weight [ ci_1, kh', kw', co_1, co_0, ci_0 ]

  277.     # weight [ no, k_h - ko // k_w % k_h - 1, k_w - ko % k_w - 1, ko // k_w // k_h, co_0, ci_0 ]

  278.     b_trans = akg.tvm.compute(kernel_shape_trans,

  279.                               lambda ko, no, ni, ki: b_value[((no * k_h + k_h - 1 - ko // k_w % k_h)

  280.                                                               * k_w + k_w - 1 - ko % k_w), ko // (k_h * k_w), ki, ni],

  281.                               name='B_trans')

  282. 

  283.     if ((stride_h > 1) or (stride_w > 1)):

  284.         @akg.tvm.hybrid.script

  285.         def data_trans_hybrid(output, inputs, const_zero):

  286.             """Implements data_trans ( B[n, c1, h * strideH, w * strideW, c0] = A[n, c1, h, w, c0] )."""

  287. 

  288.             stride_h = output.shape[2] // inputs.shape[2]

  289.             stride_w = output.shape[3] // inputs.shape[3]

  290. 

  291.             b = allocate(output.shape, output.dtype, 'local')

  292.             for n in range(output.shape[0]):

  293.                 for c1 in range(output.shape[1]):

  294.                     for h in range(output.shape[2]):

  295.                         for w in range(output.shape[3]):

  296.                             for c0 in range(output.shape[4]):

  297.                                 b[n, c1, h, w, c0] = const_zero

  298.                                 if h % stride_h == 0 and w % stride_w == 0:

  299.                                     b[n, c1, h, w, c0] = inputs[n, c1, h // stride_h, w // stride_w, c0]

  300. 

  301.             return b

  302. 

  303.         a_trans_init = akg.tvm.placeholder(input_trans_shape_nc1hwc0, dtype="float16", name='a_trans')

  304.         const_zero = akg.tvm.const(0, 'float16')

  305.         a_trans = data_trans_hybrid(a_trans_init, a_value, const_zero)

  306.     else:

  307.         a_trans = a_value

  308.     conv_attrs = {

  309.         "pragma_conv_kernel_n": k_n,

  310.         "pragma_conv_kernel_h": k_h,

  311.         "pragma_conv_kernel_w": k_w,

  312.         "pragma_conv_padding_top": p_top,

  313.         "pragma_conv_padding_bottom": p_bottom,

  314.         "pragma_conv_padding_left": p_left,

  315.         "pragma_conv_padding_right": p_right,

  316.         "pragma_conv_bypass_l1": 0,

  317.         "pragma_conv_backprop_input": 1,

  318.         "pragma_conv_stride_h": s_h,

  319.         "pragma_conv_stride_w": s_w,

  320.         "pragma_conv_dilation_h": 1,

  321.         "pragma_conv_dilation_w": 1,

  322.         "pragma_conv_fm_n": in_n,

  323.         "pragma_conv_fm_c": in_c,

  324.         "pragma_conv_fm_h": in_h,

  325.         "pragma_conv_fm_w": in_w,

  326.         "feature": a_trans.op.name,

  327.         "filter": b_value.op.name,

  328.         "bias": 'None',

  329.         "res": output_name}

  330.     if not use_auto_tiling:

  331.         conv_attrs["pragma_conv_h_cut"] = (tile_out_h - 1) * s_h + k_h

  332.         conv_attrs["pragma_conv_w_cut"] = (tile_out_w - 1) * s_w + k_w

  333.         conv_attrs["pragma_conv_co_cut"] = c1_cut * k_c0

  334.         conv_attrs["pragma_conv_m_cut"] = tile_mm

  335.         conv_attrs["pragma_conv_k_cut"] = tile_kk

  336.         conv_attrs["pragma_conv_n_cut"] = tile_nn

  337.     res_c = akg.tvm.compute(out_shape_nc1hwc0,

  338.                             lambda n, c1, h, w, c0: akg.lang.cce.mmad(

  339.                                 (akg.tvm.if_then_else(akg.tvm.any((h * s_h + kh) < p_top,

  340.                                                                   (h * s_h + kh) > (in_h + p_top - 1),

  341.                                                                   (w * s_w + kw) < p_left,

  342.                                                                   (w * s_w + kw) > (in_w + p_left - 1)),

  343.                                                       akg.tvm.const(0.0, 'float16'),

  344.                                                       a_trans[n, kc1, (h * s_h + kh - p_top),

  345.                                                               (w * s_w + kw - p_left), kc0])

  346.                                  * b_trans[(kc1 * k_h + kh) * k_w + kw, c1, c0, kc0]).astype("float32"),

  347.                                 axis=[kc1, kh, kw, kc0]), name=output_name,

  348.                             attrs=conv_attrs)

  349. 

  350.     res_c = cast.cast(res_c, "float16")

  351. 

  352.     return res_c, {"dim": info, "pragma_reschedule": 1, "pragma_rmselfdep": 0}

  353. 

  354. @vc_util.check_input_type((list, tuple), (list, tuple), (list, tuple), (list, tuple), (list, tuple), (list, tuple),

  355.                           (dict, type(None)))

  356. def conv_backprop_input(data, fmap_shape, filter_shape, pad_, stride_, dilation_, attrs=None):

  357.     """

  358.     Computes dx according "conv forward".

  359. 

  360.     Args:

  361.         data (list[tvm.tensor.Tensor]): a list with length 2.

  362.               data[0](consider as dy) Tensor of type float16 ,shape 5D(out_n, out_c//C0, out_h, out_w,C0)

  363.               data[1](consider as w)  Tensor of type float16 ,shape 4D(wC//C0*wH*wW, wN//C0, C0,C0)

  364.         fmap_shape (list[int]): [fN, fC, fH, fW]

  365.         filter_shape (list[int]): [wN, wC, wH, wW]

  366.         pad_ (list[int]): [pad_left, pad_right, pad_top, pad_bottom]

  367.         stride_ (list[int]): [stride_h, stride_w]

  368.         dilation_ (list[int]): [dilation_h, dilation_w]

  369.         attrs (dict): a dict with keys like conv_tile,bypass and etc.

  370. 

  371.     Returns:

  372.         tvm.tensor.Tensor.

  373.         configs.

  374.     """

  375. 

  376.     if len(data) != 2:

  377.         raise IndexError("data contains output tensor and filter tensor")

  378. 

  379.     vc_util.convolution_format_check(fmap_shape, filter_shape, pad_, stride_, dilation_)

  380. 

  381.     block_size = 16

  382.     in_n, in_c, in_h, in_w = fmap_shape

  383.     cout, _, w_h, w_w = filter_shape

  384. 

  385.     in_c = (in_c + block_size - 1) // block_size * block_size

  386.     cout = (cout + block_size - 1) // block_size * block_size

  387. 

  388.     pad_top, pad_bottom, pad_left, pad_right = pad_

  389.     stride_h, stride_w = stride_

  390. 

  391.     dilation_h, dilation_w = dilation_

  392.     if dilation_h != 1 or dilation_w != 1:

  393.         raise ValueError("The value od elements in dilation_ must be 1.")

  394. 

  395.     out_n = in_n

  396.     out_c = cout

  397.     out_h = (in_h + pad_top + pad_bottom - w_h) // stride_h + 1

  398.     out_w = (in_w + pad_left + pad_right - w_w) // stride_w + 1

  399. 

  400.     x_shape = (out_n, out_c, out_h, out_w)

  401.     w_shape = (cout, in_c, w_h, w_w)

  402. 

  403.     key = gen_key(fmap_shape, filter_shape, pad_, stride_, dilation_)

  404.     res_c, configs = conv_backprop_input_compute(data, x_shape, w_shape, fmap_shape, pad_, stride_,

  405.                                                  block_size=block_size, attrs=attrs, key=key)

  406. 

  407.     return res_c, configs