cutorch failure in CUDA 8

[apvijay]

The below test code for image resize results in an error:

Specs:
Ubuntu 14.04
CUDA-8
OpenCV 3.2 / OpenCV 3.1 / OpenCV 3.1 (with CUDA 8 support)
Latest Torch-7and related packages
GPU: Titan X Pascal

Code:

require 'cutorch'
local cv = require 'cv'
require 'cv.cudawarping'
require 'cv.highgui'
require 'cv.imgcodecs'
require 'image'

cutorch.setDevice(3)
local img = cv.imread {arg[1] or 'lena.jpg', cv.IMREAD_COLOR}
print("Read image - done")
local imgCUDA = img:float():cuda() / 255
print("Cuda conversion - done")
for i=0,100,1
do
  print(i)
  local resized = cv.cuda.resize{imgCUDA, {1024, 768}}
  print("Resize - done")
  print(resized:size())
  print(resized:type())
  image.save('lenaresized.png',resized:transpose(1,3):transpose(2,3))
end

The output of the above test code is below:

Read image - done	
Cuda conversion - done	
0	
Resize - done	
  768
 1024
    3
[torch.LongStorage of size 3]

torch.CudaTensor	

.
.
.

15	
Resize - done	
  768
 1024
    3
[torch.LongStorage of size 3]

torch.CudaTensor	
THCudaCheck FAIL file=/tmp/luarocks_cutorch-scm-1-1283/cutorch/lib/THC/generic/THCStorage.c line=182 error=17 : invalid device pointer
/home/user/torch/install/bin/luajit: cuda runtime error (17) : invalid device pointer at /tmp/luarocks_cutorch-scm-1-1283/cutorch/lib/THC/generic/THCStorage.c:182

There is some memory conversion(?) issue after 15 times of looping. The corresponding line in THCStorage.c is

THCudaCheck(
        (*self->allocator->free)(self->allocatorContext, self->data));

This code runs in CUDA 6.5/OpenCV 3.1/ Torch 7. Does torch-opencv support CUDA-8?

Could someone look into this issue?

[shrubb]

Hi,

unfortunately, I still don't have a machine with GPU to test on :(

Can you try this code? What's the output?

require 'cutorch'
local cv = require 'cv'
require 'cv.cudawarping'
require 'cv.highgui'
require 'cv.imgcodecs'
require 'image'

cutorch.setDevice(3)
local img = cv.imread {arg[1] or 'lena.jpg', cv.IMREAD_COLOR}
print("Read image - done")
local imgCUDA = img:float():cuda() / 255
print("Cuda conversion - done")
local img2 = imgCUDA:clone(); img2 = nil; collectgarbage()
print("Deallocate image - done")
local resized = cv.cuda.resize{imgCUDA, {1024, 768}}
print("Resize - done")
print(resized:size())
print(resized:type())
resized = nil; collectgarbage()
print("Deallocate resized image - done")

[apvijay]

The error is the same. It does not execute the last statement print("Deallocate resized image - done").

Read image - done	
Cuda conversion - done	
Deallocate image - done	
Resize - done	
  768
 1024
    3
[torch.LongStorage of size 3]

torch.CudaTensor	
THCudaCheck FAIL file=/tmp/luarocks_cutorch-scm-1-1283/cutorch/lib/THC/generic/THCStorage.c line=182 error=17 : invalid device pointer
/home/ipcv/torch/install/bin/luajit: cuda runtime error (17) : invalid device pointer at /tmp/luarocks_cutorch-scm-1-1283/cutorch/lib/THC/generic/THCStorage.c:182

The code works with CUDA 6.5/OpenCV 3.1/ Torch 7 on GTX-980 (5.2 arch).

It shows the above error with CUDA 8.0 on Titan-X Pascal (6.1 arch).