non-contiguous copy still buggy

[soumith]

Test case:

require 'cutorch'

val = 1
ps = torch.LongStorage({4, 4, 4})
cube = torch.Tensor(ps):apply(function()
          val = val + 1
          return val
         end):cuda()

ps = torch.LongStorage({4, 12})
x = torch.CudaTensor(ps):fill(-1)

l = 2
h = 1
w = 2

print(cube)
print(cube[l][{{h,h+2},{w,w+2}}])

x[{{1},{1,9}}]:copy(cube[l][{{h,h+2},{w,w+2}}]:contiguous())
print(x)

x[{{1},{1,9}}]:copy(cube[l][{{h,h+2},{w,w+2}}])
print(x)

x[{{1,1},{1,9}}]:copy(cube[l][{{h,h+2},{w,w+2}}])
print(x)


x[{1,{1,9}}]:copy(cube[l][{{h,h+2},{w,w+2}}])
print(x)

Output:

(1,.,.) =
   2   3   4   5
   6   7   8   9
  10  11  12  13
  14  15  16  17

(2,.,.) =
  18  19  20  21
  22  23  24  25
  26  27  28  29
  30  31  32  33

(3,.,.) =
  34  35  36  37
  38  39  40  41
  42  43  44  45
  46  47  48  49

(4,.,.) =
  50  51  52  53
  54  55  56  57
  58  59  60  61
  62  63  64  65
[torch.CudaTensor of dimension 4x4x4]

 19  20  21
 23  24  25
 27  28  29
[torch.CudaTensor of dimension 3x3]

 19  20  21  23  24  25  27  28  29  -1  -1  -1
 -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1
 -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1
 -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1
[torch.CudaTensor of dimension 4x12]

 19  20  21  22  23  24  25  26  27  -1  -1  -1
 -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1
 -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1
 -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1
[torch.CudaTensor of dimension 4x12]

 19  20  21  22  23  24  25  26  27  -1  -1  -1
 -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1
 -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1
 -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1
[torch.CudaTensor of dimension 4x12]

 19  20  21  22  23  24  25  26  27  -1  -1  -1
 -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1
 -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1
 -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1  -1
[torch.CudaTensor of dimension 4x12]

[nicholas-leonard]

So the last 3 are wrong?

[soumith]

yea 👎

[soumith]

update: one of the engineers at FB is fixing this.

[nicholas-leonard]

Awesome.