Closed marcomastri closed 8 years ago
Hi Marco,
Hmmm. Interesting. What happens if you do:
luajit -l cltorch -e 'print(torch.ClStorage())'
?
On my machine I get:
Using NVIDIA Corporation , OpenCL platform: NVIDIA CUDA
Using OpenCL device: GeForce 940M
[torch.ClStorage of size 0]
This is what I get:
Using Apple , OpenCL platform: Apple
Using OpenCL device: Iris Pro
[torch.ClStorage of size 0]
Hmmm, well, that looks correct... The line that is failing in your stack trace is:
tester:asserteq('[torch.ClStorage of size 0]\n', tostring(torch.ClStorage()))
It looks simple enough. So, the inner torch.ClStorage()
bit is giving the output it should. What if you do:
luajit -l cltorch -e 'print(tostring(torch.ClStorage()))'
? On my system, this gives:
$ luajit -l cltorch -e 'print(tostring(torch.ClStorage()))'
Using NVIDIA Corporation , OpenCL platform: NVIDIA CUDA
Using OpenCL device: GeForce 940M
[torch.ClStorage of size 0]
1
Hmmm, I wonder if we are using different versions of Tester.lua? It's from 'torch' pakage. Here is some information on my own Tester.lua:
ubuntu:~/torch/pkg/torch$ git log -n 5 Tester.lua
commit 6fc2c95707d238f5106501c2c98f726a6f737d41
Author: Adam Lerer <alerer@fb.com>
Date: Mon Jul 27 00:11:39 2015 -0700
torch Tester: run tests in alphabetical order, and support interrupt handler
commit 14f5b23527cc738f9f9e0dfb61f117daed367be7
Author: George van den Driessche <georgevdd@google.com>
Date: Thu Apr 9 12:19:59 2015 +0100
Separated choice of tests to run from their actual running.
commit cfa0b1cfd183cc4553dbe420bba596eda27b283d
Author: GeorgOstrovski <ostrovski@google.com>
Date: Thu Mar 19 16:26:26 2015 +0000
Improve assertTensorEq to work with precision 0
This change allows to use assertTensorEq to check for exact equality (previously setting precision = 0 would have made the condition unsatisfiable)
commit f6c1b74a3d4205f5347a7f3c799bdc4b959717ec
Author: GeorgOstrovski <ostrovski@google.com>
Date: Tue Mar 17 17:52:28 2015 +0000
fix assertTensorEq for case of empty tensors
The case of empty tensors was handled wrongly: leading to exception. In the test of torch.maskedSelect and similar the case of an empty tensor can occur, leading to occasional test failures.
commit eb1f2345aeea145c5fb275755e8904e0f3c99987
Author: soumith <soumith@fb.com>
Date: Mon Nov 17 13:23:59 2014 -0800
removing redundant uninitialized variable access
ubuntu:~/torch/pkg/torch$ md5sum Tester.lua
7e5ec5a6cbdb5cc90da88ceba811665f Tester.lua
ubuntu:~/torch/pkg/torch$ md5sum ~/torch/install/share/lua/5.1/torch/Tester.lua 7e5ec5a6cbdb5cc90da88ceba811665f /home/ubuntu/torch/install/share/lua/5.1/torch/Tester.lua
The output from the previous command looks fine:
Using Apple , OpenCL platform: Apple
Using OpenCL device: Iris Pro
[torch.ClStorage of size 0]
1
Here’s my info for Tester.lua:
commit 9a41915d490dd6d6530c963911f56a81d64f850e
Author: soumith <soumith@gmail.com>
Date: Mon Mar 7 21:42:34 2016 -0500
making sys optional
commit d8ff64c5707f716199e7782d630e44e2cf402a54
Author: David Saxton <saxton@google.com>
Date: Wed Feb 24 17:54:10 2016 +0000
Replace torch.Tester with totem.Tester + extra stuff.
This should bring a lot of benefit to code that uses torch.Tester (totem will
eventually become deprecated). Note that torch.Tester and totem.Tester once
shared the same code - this change brings it full circle.
At a glance, extra functionality includes:
- A general equality checker that accepts many different objects.
- Deep table comparison with precision checking.
- Stricter argument checking in using the test functions.
- Better output.
- torch.Storage comparison.
- Extra features for fine-grained control of testing.
commit 6fc2c95707d238f5106501c2c98f726a6f737d41
Author: Adam Lerer <alerer@fb.com>
Date: Mon Jul 27 00:11:39 2015 -0700
torch Tester: run tests in alphabetical order, and support interrupt handler
commit 14f5b23527cc738f9f9e0dfb61f117daed367be7
Author: George van den Driessche <georgevdd@google.com>
Date: Thu Apr 9 12:19:59 2015 +0100
Separated choice of tests to run from their actual running.
commit cfa0b1cfd183cc4553dbe420bba596eda27b283d
Author: GeorgOstrovski <ostrovski@google.com>
Date: Thu Mar 19 16:26:26 2015 +0000
Improve assertTensorEq to work with precision 0
This change allows to use assertTensorEq to check for exact equality (previously setting precision = 0 would have made the condition unsatisfiable)
ah-hah. You have a newer version, of torch
module. I updated my torch
module, and have the same issue now. Ok, I'll take a look.
ok. Do you want to pull down latest cltorch, rebuild, and try again? (addressed in 61bb912 )
Ok, now everything looks fine, thank you very much!
Ok, cool :-)
Hello, I tried to install cltorch on my macbook. The installation itself went fine, but I got an error when running the test:
This is the output from
th -l cltorch -e 'cltorch.about()'
:As per the subject my machine is a MacBook Pro Retina 2014 (2,2 GHz Intel Core i7, Intel Iris Pro 1536 MB), running El Capitan 10.11.1.
Thank you very much, Marco