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_bn1.py

260 lines
15 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_bn1"""

  18. from functools import reduce

  19. import akg.topi

  20. import akg.tvm

  21. import akg

  22. import akg.lang.cce

  23. from akg.ops.math import cast

  24. from akg.ops.nn.conv import conv_core

  25. from akg.ops.nn.conv import conv_set_dim_func

  26. from akg.utils import validation_check as vc_util

  27. 

  28. conv_bn1_set_dim_map = {

  29.     str(((1, 1024, 14, 14), (2048, 1024, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1), False)):

  30.         ([14, 2048, 64, 96, 128], {"bypass": 1}),

  31.     str(((1, 1024, 14, 14), (256, 1024, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  32.         ([14, 256, 208, 64, 112], {"bypass": 1}),

  33.     str(((1, 1024, 14, 14), (512, 1024, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1), False)):

  34.         ([13, 144, 48, 48, 128, 13], {"bypass": 0}),

  35.     str(((1, 128, 28, 28), (128, 128, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1), False)):

  36.         ([30, 128, 240, 48, 64, 30], {"bypass": 0}),

  37.     str(((1, 128, 28, 28), (512, 128, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  38.         ([8, 256, 112, 16, 48, 28], {"bypass": 0}),

  39.     str(((1, 2048, 7, 7), (512, 2048, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  40.         ([7, 160, 48, 48, 96, 7], {"bypass": 0}),

  41.     str(((1, 256, 14, 14), (1024, 256, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  42.         ([14, 256, 48, 64, 256, 14], {"bypass": 0}),

  43.     str(((1, 256, 14, 14), (256, 256, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1), False)):

  44.         ([14, 192, 64, 128, 160, 16], {"bypass": 0}),

  45.     str(((1, 256, 56, 56), (128, 256, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1), False)):

  46.         ([7, 128, 112, 48, 16, 55], {"bypass": 0}),

  47.     str(((1, 256, 56, 56), (64, 256, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  48.         ([6, 64, 224, 32, 16, 56], {"bypass": 0}),

  49.     str(((1, 3, 224, 224), (64, 3, 7, 7), (2, 3, 2, 3), (2, 2), (1, 1), False)):

  50.         ([97, 64, 128, 128, 64, 229], {"bypass": 0}),

  51.     str(((1, 512, 28, 28), (128, 512, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  52.         ([28, 64, 48, 304, 32, 28], {"bypass": 0}),

  53.     str(((1, 512, 28, 28), (256, 512, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1), False)):

  54.         ([27, 192, 64, 48, 160, 27], {"bypass": 0}),

  55.     str(((1, 512, 7, 7), (2048, 512, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  56.         ([7, 128, 48, 176, 80, 7], {"bypass": 0}),

  57.     str(((1, 512, 7, 7), (512, 512, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1), False)):

  58.         ([9, 512, 64, 128, 96, 9], {"bypass": 1}),

  59.     str(((1, 64, 56, 56), (256, 64, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  60.         ([56, 256, 392, 16, 32], {"bypass": 1}),

  61.     str(((1, 64, 56, 56), (64, 64, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  62.         ([32, 64, 224, 32, 64, 56], {"bypass": 0}),

  63.     str(((1, 64, 56, 56), (64, 64, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1), False)):

  64.         ([10, 64, 224, 48, 48, 58], {"bypass": 1}),

  65.     str(((1, 256, 56, 56), (512, 256, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1), False)):

  66.         ([7, 320, 112, 160, 48, 55], {"bypass": 0}),

  67.     str(((1, 512, 28, 28), (1024, 512, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1), False)):

  68.         ([13, 496, 96, 176, 144, 27], {"bypass": 0}),

  69.     str(((1, 256, 56, 56), (128, 256, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  70.         ([12, 128, 112, 64, 128, 56], {"bypass": 0}),

  71.     str(((1, 512, 28, 28), (256, 512, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  72.         ([28, 176, 224, 112, 80, 28], {"bypass": 0}),

  73.     str(((1, 1024, 14, 14), (512, 1024, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  74.         ([7, 384, 96, 48, 224, 14], {"bypass": 0}),

  75.     str(((1, 128, 56, 56), (128, 128, 3, 3), (0, 1, 0, 1), (2, 2), (1, 1), False)):

  76.         ([37, 128, 224, 96, 96, 57], {"bypass": 0}),

  77.     str(((1, 256, 28, 28), (256, 256, 3, 3), (0, 1, 0, 1), (2, 2), (1, 1), False)):

  78.         ([29, 256, 80, 224, 144, 29], {"bypass": 1}),

  79.     str(((1, 512, 14, 14), (512, 512, 3, 3), (0, 1, 0, 1), (2, 2), (1, 1), False)):

  80.         ([15, 512, 64, 64, 272, 15], {"bypass": 1}),

  81. 

  82.     str(((2, 1024, 14, 14), (2048, 1024, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1), False)):

  83.         ([13, 112, 48, 176, 80, 13], {"bypass": 0}),

  84.     str(((2, 1024, 14, 14), (256, 1024, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  85.         ([14, 128, 48, 48, 64, 14], {"bypass": 0}),

  86.     str(((2, 1024, 14, 14), (512, 1024, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1), False)):

  87.         ([13, 144, 48, 48, 128, 13], {"bypass": 0}),

  88.     str(((2, 128, 28, 28), (128, 128, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1), False)):

  89.         ([30, 128, 240, 48, 64, 30], {"bypass": 0}),

  90.     str(((2, 128, 28, 28), (512, 128, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  91.         ([8, 256, 112, 16, 48, 28], {"bypass": 0}),

  92.     str(((2, 2048, 7, 7), (512, 2048, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  93.         ([7, 160, 48, 48, 96, 7], {"bypass": 0}),

  94.     str(((2, 256, 14, 14), (1024, 256, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  95.         ([14, 256, 48, 64, 256, 14], {"bypass": 0}),

  96.     str(((2, 256, 14, 14), (256, 256, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1), False)):

  97.         ([14, 192, 64, 128, 160, 16], {"bypass": 0}),

  98.     str(((2, 256, 56, 56), (128, 256, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1), False)):

  99.         ([7, 128, 112, 48, 16, 55], {"bypass": 0}),

  100.     str(((2, 256, 56, 56), (64, 256, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  101.         ([6, 64, 224, 32, 16, 56], {"bypass": 0}),

  102.     str(((2, 3, 224, 224), (64, 3, 7, 7), (2, 3, 2, 3), (2, 2), (1, 1), False)):

  103.         ([97, 64, 128, 128, 64, 229], {"bypass": 0}),

  104.     str(((2, 512, 28, 28), (128, 512, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  105.         ([28, 64, 48, 304, 32, 28], {"bypass": 0}),

  106.     str(((2, 512, 28, 28), (256, 512, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1), False)):

  107.         ([27, 192, 64, 48, 160, 27], {"bypass": 0}),

  108.     str(((2, 512, 7, 7), (2048, 512, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  109.         ([7, 128, 48, 176, 80, 7], {"bypass": 0}),

  110.     str(((2, 512, 7, 7), (512, 512, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1), False)):

  111.         ([9, 512, 64, 128, 96, 9], {"bypass": 1}),

  112.     str(((2, 64, 56, 56), (256, 64, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  113.         ([56, 256, 784, 16, 32], {"bypass": 1}),

  114.     str(((2, 64, 56, 56), (64, 64, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  115.         ([32, 64, 224, 32, 64, 56], {"bypass": 0}),

  116.     str(((2, 64, 56, 56), (64, 64, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1), False)):

  117.         ([10, 64, 224, 48, 48, 58], {"bypass": 1}),

  118.     str(((2, 256, 56, 56), (512, 256, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1), False)):

  119.         ([7, 320, 112, 160, 48, 55], {"bypass": 0}),

  120.     str(((2, 512, 28, 28), (1024, 512, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1), False)):

  121.         ([13, 496, 96, 176, 144, 27], {"bypass": 0}),

  122.     str(((2, 256, 56, 56), (128, 256, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  123.         ([12, 128, 224, 64, 128, 56], {"bypass": 0}),

  124.     str(((2, 512, 28, 28), (256, 512, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  125.         ([28, 176, 224, 112, 80, 28], {"bypass": 0}),

  126.     str(((2, 1024, 14, 14), (512, 1024, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  127.         ([7, 384, 96, 48, 224, 14], {"bypass": 0}),

  128.     str(((2, 128, 56, 56), (128, 128, 3, 3), (0, 1, 0, 1), (2, 2), (1, 1), False)):

  129.         ([37, 128, 224, 96, 96, 57], {"bypass": 0}),

  130.     str(((2, 256, 28, 28), (256, 256, 3, 3), (0, 1, 0, 1), (2, 2), (1, 1), False)):

  131.         ([29, 256, 80, 224, 144, 29], {"bypass": 1}),

  132.     str(((2, 512, 14, 14), (512, 512, 3, 3), (0, 1, 0, 1), (2, 2), (1, 1), False)):

  133.         ([15, 512, 64, 64, 272, 15], {"bypass": 1}),

  134. 

  135.     str(((32, 1024, 14, 14), (2048, 1024, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1), False)):

  136.         ([13, 64, 48, 128, 64, 13], {"bypass": 0}),

  137.     str(((32, 1024, 14, 14), (256, 1024, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  138.         ([14, 128, 48, 48, 64, 14], {"bypass": 0}),

  139.     str(((32, 1024, 14, 14), (512, 1024, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1), False)):

  140.         ([13, 144, 48, 48, 128, 13], {"bypass": 0}),

  141.     str(((32, 128, 28, 28), (128, 128, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1), False)):

  142.         ([30, 128, 240, 48, 64, 30], {"bypass": 0}),

  143.     str(((32, 128, 28, 28), (512, 128, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  144.         ([8, 256, 224, 80, 48, 28], {"bypass": 0}),

  145.     str(((32, 2048, 7, 7), (512, 2048, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  146.         ([7, 128, 48, 112, 112, 7], {"bypass": 0}),

  147.     str(((32, 256, 14, 14), (1024, 256, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  148.         ([14, 256, 48, 64, 256, 14], {"bypass": 0}),

  149.     str(((32, 256, 14, 14), (256, 256, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1), False)):

  150.         ([16, 96, 80, 96, 96, 16], {"bypass": 0}),

  151.     str(((32, 256, 56, 56), (128, 256, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1), False)):

  152.         ([23, 128, 112, 240, 48, 55], {"bypass": 1}),

  153.     str(((32, 256, 56, 56), (64, 256, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  154.         ([26, 64, 448, 32, 16, 56], {"bypass": 0}),

  155.     str(((32, 3, 224, 224), (64, 3, 7, 7), (2, 3, 2, 3), (2, 2), (1, 1), False)):

  156.         ([61, 64, 224, 48, 64, 229], {"bypass": 0}),

  157.     str(((32, 512, 28, 28), (128, 512, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  158.         ([8, 96, 224, 48, 16, 28], {"bypass": 0}),

  159.     str(((32, 512, 28, 28), (256, 512, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1), False)):

  160.         ([27, 48, 48, 160, 48, 27], {"bypass": 0}),

  161.     str(((32, 512, 7, 7), (2048, 512, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  162.         ([7, 864, 48, 288, 16, 7], {"bypass": 0}),

  163.     str(((32, 512, 7, 7), (512, 512, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1), False)):

  164.         ([9, 512, 64, 128, 96, 9], {"bypass": 1}),

  165.     str(((32, 64, 56, 56), (256, 64, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  166.         ([8, 64, 448, 64, 32, 56], {"bypass": 0}),

  167.     str(((32, 64, 56, 56), (64, 64, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  168.         ([14, 64, 112, 64, 16, 56], {"bypass": 0}),

  169.     str(((32, 64, 56, 56), (64, 64, 3, 3), (1, 1, 1, 1), (1, 1), (1, 1), False)):

  170.         ([6, 64, 224, 64, 64, 58], {"bypass": 0}),

  171.     str(((32, 256, 56, 56), (512, 256, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1), False)):

  172.         ([7, 320, 112, 160, 48, 55], {"bypass": 0}),

  173.     str(((32, 512, 28, 28), (1024, 512, 1, 1), (0, 0, 0, 0), (2, 2), (1, 1), False)):

  174.         ([13, 496, 96, 176, 144, 27], {"bypass": 0}),

  175.     str(((32, 256, 56, 56), (128, 256, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  176.         ([2, 96, 112, 48, 96, 56], {"bypass": 0}),

  177.     str(((32, 512, 28, 28), (256, 512, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  178.         ([20, 128, 16, 448, 64, 28], {"bypass": 0}),

  179.     str(((32, 1024, 14, 14), (512, 1024, 1, 1), (0, 0, 0, 0), (1, 1), (1, 1), False)):

  180.         ([7, 384, 48, 48, 224, 14], {"bypass": 0}),

  181.     str(((32, 128, 56, 56), (128, 128, 3, 3), (0, 1, 0, 1), (2, 2), (1, 1), False)):

  182.         ([9, 128, 64, 64, 128, 57], {"bypass": 0}),

  183.     str(((32, 256, 28, 28), (256, 256, 3, 3), (0, 1, 0, 1), (2, 2), (1, 1), False)):

  184.         ([29, 256, 80, 224, 144, 29], {"bypass": 1}),

  185.     str(((32, 512, 14, 14), (512, 512, 3, 3), (0, 1, 0, 1), (2, 2), (1, 1), False)):

  186.         ([15, 512, 64, 64, 272, 15], {"bypass": 1}),

  187.     # alexnet

  188.     str(((32, 3, 227, 227), (96, 3, 11, 11), (0, 0, 0, 0), (4, 4), (1, 1), False)):

  189.         ([63, 96, 208, 32, 96, 227], {"bypass": 0}),

  190.     str(((32, 96, 27, 27), (256, 96, 5, 5), (2, 2, 2, 2), (1, 1), (1, 1), False)):

  191.         ([21, 160, 176, 32, 96, 31], {"bypass": 0})

  192. }

  193. 

  194. 

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

  196.                           (bool, type(None)), (dict, type(None)))

  197. def conv_bn1(data, fmap_shape, filter_shape, pad, stride, dilation, use_bias=False, attrs=None):

  198.     """

  199.     Computes sums of 5-D convolutions and use convolution's fp32 result to compute first part of Fused_batch_norm.

  200. 

  201.     Fused_batch_norm's first part:

  202. 

  203.     \f[

  204.      m = N \times H \times W \\

  205.      \\mu_{tmp} = \\sum_{n, h, w}{\frac{x}{m}} \\

  206.      \\sigma^2_{tmp} = \\sum_{n, h, w}{\frac{x^2}{m}}

  207.     \f]

  208. 

  209.     Args:

  210.         data (list[tvm.tensor.Tensor]): the size is 3 if use_bias else the size is 2;

  211.               data[0] Tensor of type float16 ,shape 5D (fN, fC // C0, C0, fH, fW)

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

  213.               data[2] Tensor of type float16 ,shape 5D (1, wN // C0, 1, 1, 16)

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

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

  216.         pad (list[int]): [pad_top, pad_bottom, pad_left, pad_right]

  217.         stride (list[int]): [stride_h, stride_w]

  218.         dilation (list[int]): [dilation_h, dilation_w]

  219.         use_bias (bool): bool var.

  220.         attrs (dict): dict with keys for example: conv_tile,bypass

  221. 

  222.     Returns:

  223.         tvm.tensor.Tensor of same type as data, shape is 5D(oN, oC // C0, oH, oW, C0)

  224.     """

  225. 

  226.     if use_bias:

  227.         raise ValueError("do not support bias yet !!!")

  228. 

  229.     block_size = 16

  230.     dim_info, conv_tile, bypass, _ = conv_set_dim_func(fmap_shape, filter_shape, pad, stride, dilation, use_bias,

  231.                                                        block_size, attrs, conv_bn1_set_dim_map)

  232.     if attrs is None:

  233.         attrs = {"conv_tile": conv_tile, "bypass": bypass}

  234.     else:

  235.         attrs['conv_tile'] = conv_tile

  236.         attrs['bypass'] = bypass

  237. 

  238.     conv_res_32 = conv_core(data, fmap_shape, filter_shape, pad, stride, dilation, use_bias, attrs)

  239. 

  240.     conv_res_16 = cast.cast(conv_res_32, "float16")

  241. 

  242.     axes = [3, 2, 0]

  243.     conv_res_32_shape = [x.value for x in conv_res_32.shape]

  244.     num = reduce(lambda i, j: i * j, [conv_res_32_shape[i] for i in axes])

  245.     avg_num = round(float(1) / float(num), 12)

  246. 

  247.     res_sum = akg.topi.sum(conv_res_32, axes, keepdims=True)

  248.     mean = akg.lang.cce.vmuls(res_sum, avg_num)

  249. 

  250.     res_square = akg.tvm.compute(conv_res_32.shape, lambda *i: conv_res_32[i] * conv_res_32[i], name="res_square")

  251.     square_sum = akg.topi.sum(res_square, axes, keepdims=True)

  252.     var_part = akg.lang.cce.vmuls(square_sum, avg_num)

  253. 

  254.     # need pragma_force_rmselfdep to enable multicore using atomic add

  255.     # because default pragma_rmselfdep=1 will disable multicore of reduce axes

  256.     attrs = {"dim": dim_info, "pragma_reschedule": 1, "enable_bisect_optimize": 0,

  257.              "pragma_rmselfdep": 0, "pragma_force_rmselfdep": 1}

  258. 

  259.     return conv_res_16, var_part, mean, attrs