Explicitly allow openArray in Tensor [], []=, fix the CI (#666)

* explicitly allow `openArray` in `[]`, `[]=` for tensors

This was simply an oversight obviously

* fix CI by compiling tests with `-d:ssl`

* need a space, duh

* use AWS mirror from PyTorch for MNIST download

* fix regression caused by PR #659

By not resetting the offset here, operating on a Tensor view without
cloning could cause undefined behavior, because we would be accessing
elements outside the tensor buffer.

* add test case for regression of #659

arraymancer.nimble

@@ -143,9 +143,9 @@ proc test(name, switches = "", split = false, lang = "c") =
   if not dirExists "build":
     mkDir "build"
   if not split:
-    exec "nim " & lang & " -o:build/" & name & switches & " -r tests/" & name & ".nim"
+    exec "nim " & lang & " -d:ssl -o:build/" & name & switches & " -r tests/" & name & ".nim"
   else:
-    exec "nim " & lang & " -o:build/" & name & switches & " -r tests/_split_tests/" & name & ".nim"
+    exec "nim " & lang & " -d:ssl -o:build/" & name & switches & " -r tests/_split_tests/" & name & ".nim"
 
   # run tests that require old RNG for backward compat. reasos
   let rngTests = ["spatial/test_kdtree",

src/arraymancer/datasets/mnist.nim

@@ -71,7 +71,11 @@ const MNISTFilenames = [
 ]
 
 const
-  DefaultMnistUrl = "http://yann.lecun.com/exdb/mnist/"
+  ## NOTE: As of some time before 2024/09/20 the MNIST dataset cannot be downloaded from
+  ## Yann's website anylonger, due to 403 error.
+  ## The AWS mirror here comes from PyTorch Vision:
+  ## https://github.com/pytorch/vision/blob/6d7851bd5e2bedc294e40e90532f0e375fcfee04/torchvision/datasets/mnist.py#L39C10-L39C56
+  DefaultMnistUrl = "https://ossci-datasets.s3.amazonaws.com/mnist/" # "http://yann.lecun.com/exdb/mnist/"
   FashionMnistUrl = "http://fashion-mnist.s3-website.eu-central-1.amazonaws.com/"
 
 proc read_mnist_images(stream: Stream): Tensor[uint8] {.noinit.} =

src/arraymancer/tensor/private/p_accessors.nim

@@ -496,7 +496,7 @@ proc checkValidSliceType*(n: NimNode) =
   ## the `[]` and `[]=` macros. It will raise a CT error in case it is not.
   ##
   ## TODO: Do we / should we allow other integer types than `tyInt` / `int`?
-  const validTypes = {ntyInt, ntyObject, ntyArray, ntySequence, ntyGenericInst}
+  const validTypes = {ntyInt, ntyObject, ntyArray, ntySequence, ntyOpenArray, ntyGenericInst}
   # `ntyObject` requires to be `Span`, ...
   template raiseError(arg: untyped): untyped =
     let typ = arg.getTypeInst
@@ -508,7 +508,7 @@ proc checkValidSliceType*(n: NimNode) =
     of validTypes:
       if arg.typeKind in {ntyObject, ntyGenericInst} and not validObjectType(arg):
         raiseError(arg)
-      elif arg.typeKind in {ntyArray, ntySequence}:
+      elif arg.typeKind in {ntyArray, ntySequence, ntyOpenArray}:
         # Need to check inner type!
         checkValidSliceType(arg.getTypeInst()[^1])
         break

src/arraymancer/tensor/private/p_shapeshifting.nim

@@ -38,6 +38,7 @@ proc reshape_no_copy*(t: AnyTensor, new_shape: varargs[int]|Metadata|seq[int], r
 proc reshape_with_copy*[T](t: Tensor[T], new_shape: varargs[int]|Metadata|seq[int], result: var Tensor[T]) =
   contiguousImpl(t, rowMajor, result)
   reshape_no_copy(t, new_shape, result, rowMajor)
+  result.offset = 0 # Offset needs to be reset! Copy is a new tensor of `new_shape`
 
 proc infer_shape*(t: Tensor, new_shape: varargs[int]): seq[int] {.noinit.} =
   ## Replace the single -1 value on `new_shape` with the value that

tests/test_bugtracker.nim

@@ -64,5 +64,51 @@ proc main() =
           x.permute(1, 0) == expected
           x.permute(1, 0).reshape(2, 2) == expected
 
+    test "Test for PR #659 regression":
+      ## Data taken from `kdtree` test case, in which the node `split`
+      ## based on `data.percentile(50)` (i.e. `data.median`) suddenly
+      ## changed after PR:
+      ## https://github.com/mratsim/Arraymancer/pull/659/
+      let t = @[@[0.195513 ,       0.225253],
+                @[0.181441 ,       0.102758],
+                @[0.26576  ,       0.218074],
+                @[0.180852 ,     0.00262669],
+                @[0.219789 ,       0.191867],
+                @[0.160581 ,          0.131],
+                @[0.269926 ,       0.237261],
+                @[0.223423 ,       0.232116],
+                @[0.191391 ,       0.183001],
+                @[0.19654  ,      0.0809091],
+                @[0.191497 ,      0.0929182],
+                @[0.22709  ,       0.125705],
+                @[0.263181 ,       0.124787],
+                @[0.204926 ,     0.00688886],
+                @[0.151998 ,      0.0531739],
+                @[0.260266 ,      0.0583248],
+                @[0.214864 ,       0.110506],
+                @[0.247688 ,      0.0732228],
+                @[0.246916 ,       0.204899],
+                @[0.215206 ,       0.202225],
+                @[0.242059 ,       0.102491],
+                @[0.159926 ,       0.115765],
+                @[0.249105 ,       0.200658],
+                @[0.195783 ,       0.123984],
+                @[0.17145  ,      0.0506388],
+                @[0.258146 ,      0.0144846],
+                @[0.215311 ,       0.222503],
+                @[0.266231 ,       0.149363],
+                @[0.178909 ,       0.142174],
+                @[0.263406 ,      0.0867369],
+                @[0.264824 ,       0.221786]
+      ].toTensor
+
+      const exp = 0.124787
+      let arg = t[_, 1]
+      check arg.offset == 1 # offset of 1 due to slicing
+      # `reshape` copies here, because `arg` is a non contiguous tensor. Thus
+      # offset must be reset to 0
+      check arg.reshape([1, arg.size]).offset == 0
+      check t[_, 1].percentile(50) == exp
+
 main()
 GC_fullCollect()