[GitHub] [incubator-mxnet] chinakook commented on issue #17907: Depthwise in windows is 10 times slower than linux on gpu

[GitBox <gi...@apache.org>]

chinakook commented on issue #17907: Depthwise in windows is 10 times slower than linux on gpu
URL: https://github.com/apache/incubator-mxnet/issues/17907#issuecomment-603938898
 
 
   There is another version in windows, pytorch takes 2.0s while mxnet takes 10.2s. I think this is a bug for a long time.
   MXNET version:
   ```python
   import os
   os.environ['MXNET_CUDNN_AUTOTUNE_DEFAULT'] = '0'
   import time
   import mxnet as mx
   from mxnet import gluon
   from mxnet.gluon import nn
   from gluoncv.model_zoo import get_model
   
   ctx = mx.gpu()
   net = get_model('mobilenetv2_1.0', norm_layer=gluon.nn.BatchNorm)
   net.initialize()
   
   net.collect_params().reset_ctx(ctx)
   
   s = time.time()
   for i in range(50):
       x = mx.nd.random.uniform(shape=(1,3,512,512), ctx=ctx)
   
       t = time.time()
       y = net(x)
       mx.nd.waitall()
       print(time.time() - t)
   
   print('TOTAL TIME: ', time.time() - s)
   ```
   ```
   0.2889983654022217
   0.15599989891052246
   0.14120268821716309
   0.1549980640411377
   0.14800024032592773
   0.1549973487854004
   0.1419973373413086
   0.16100192070007324
   0.15399909019470215
   0.14299798011779785
   0.1490001678466797
   0.17000079154968262
   0.1530005931854248
   0.14499974250793457
   0.1569969654083252
   0.15002942085266113
   0.14699625968933105
   0.14600133895874023
   0.143998384475708
   0.15400242805480957
   0.1439976692199707
   0.14451003074645996
   0.16103625297546387
   0.15851068496704102
   0.15300440788269043
   0.15399932861328125
   0.15399956703186035
   0.14400243759155273
   0.15401935577392578
   0.14500117301940918
   0.14951753616333008
   0.14799976348876953
   0.14800000190734863
   0.15600085258483887
   0.1529989242553711
   0.14699888229370117
   0.14899921417236328
   0.1512279510498047
   0.1525120735168457
   0.1549992561340332
   0.16200017929077148
   0.1529998779296875
   0.1510009765625
   0.14804387092590332
   0.14800000190734863
   0.15600061416625977
   0.15230464935302734
   0.15199899673461914
   0.14699792861938477
   0.1289997100830078
   TOTAL TIME:  10.248228788375854
   ```
   
   PYTORCH version:
   ```python
   import time
   import torch
   import torchvision
   
   torch.backends.cudnn.benchmark=False
   
   net = torchvision.models.mobilenet_v2()
   net.cuda()
   net.eval()
   
   s = time.time()
   for i in range(50):
       t = time.time()
       x = torch.rand([1,3,512,512]).cuda()
       y = net(x)
       print(time.time() - t)
   
   print('TOTAL TIME: ', time.time() - s)
   ```
   
   ```
   0.9051487445831299
   0.04097485542297363
   0.019997835159301758
   0.018999099731445312
   0.023026704788208008
   0.021998167037963867
   0.020003795623779297
   0.0209958553314209
   0.020031213760375977
   0.020966291427612305
   0.019999980926513672
   0.022031784057617188
   0.019968032836914062
   0.023028850555419922
   0.020004987716674805
   0.01996612548828125
   0.022998332977294922
   0.020999431610107422
   0.02000117301940918
   0.019997119903564453
   0.02300119400024414
   0.02200031280517578
   0.01899886131286621
   0.01999974250793457
   0.021999835968017578
   0.02000284194946289
   0.02000141143798828
   0.02000117301940918
   0.02099919319152832
   0.020000457763671875
   0.021001338958740234
   0.020998477935791016
   0.020000219345092773
   0.020998477935791016
   0.022002458572387695
   0.02502727508544922
   0.02000284194946289
   0.021997690200805664
   0.021001100540161133
   0.024999141693115234
   0.0299990177154541
   0.02599787712097168
   0.029999256134033203
   0.029999256134033203
   0.02700185775756836
   0.02520275115966797
   0.02800154685974121
   0.032999277114868164
   0.02400040626525879
   0.02900218963623047
   TOTAL TIME:  2.065340518951416
   ```

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services