Fix issue 639 (#640)

* Make newTensorUninit[T]() create a rank-1 empty tensor

Similar to same change done to `newTensor` on the previous commit.

* Make newTensor[T]() create a rank-1 empty tensor

Up until now calling newTensor without arguments would create a rank-0 tensor which does not work well (e.g. it reports its size as size 0)! Instead we now create a rank-1 empty tensor when no shape is provided.

It is still possible to explicitly create a rank-0 tensor by explicitly passing an empty shape (i.e. `[]`) to newTensor (e.g. `newTensor[float]([])`). This can be useful to create "sentinel" values for procedures that take tensors as arguments.

* Fix issue #639 (`size` returns 1 for rank-0 tensors)

This fixes #639.

While this adds an extra check to `size` which might be called frequently, I have not seen a major difference on several of the benchmarks.

src/arraymancer/laser/dynamic_stack_arrays.nim

@@ -91,6 +91,8 @@ func `$`*(a: DynamicStackArray): string =
   result.add("]")
 
 func product*[T:SomeNumber](a: DynamicStackArray[T]): T =
+  if unlikely(a.len == 0):
+    return 0
   result = 1
   for value in items(a):
     result *= value

src/arraymancer/laser/tensor/initialization.nim

@@ -184,7 +184,14 @@ proc setZero*[T](t: var Tensor[T], check_contiguous: static bool = true) =
         chunk_size * sizeof(T)
       )
 
-proc newTensor*[T](shape: varargs[int]): Tensor[T] =
+proc newTensor*[T](shape: varargs[int] = [0]): Tensor[T] =
+  ## Create a new tensor of type T with the given shape.
+  ##
+  ## If no shape is provided, we create an empty rank-1 tensor.
+  ## To create a rank-0 tensor, explicitly pass and empty shape `[]`.
+  ##
+  ## Note that in general it is not a good idea to use rank-0 tensors.
+  ## However, they can be used as "sentinel" values for Tensor arguments.
   var size: int
   initTensorMetadata(result, size, shape)
   allocCpuStorage(result.storage, size)
@@ -193,6 +200,7 @@ proc newTensor*[T](shape: varargs[int]): Tensor[T] =
     setZero(result, check_contiguous = false)
 
 proc newTensor*[T](shape: Metadata): Tensor[T] =
+  ## Create a new tensor of type T with the given shape.
   var size: int
   initTensorMetadata(result, size, shape)
   allocCpuStorage(result.storage, size)

src/arraymancer/tensor/init_cpu.nim

@@ -23,15 +23,17 @@ import std / [random, math]
 
 export initialization
 
-proc newTensorUninit*[T](shape: varargs[int]): Tensor[T] {.noinit, inline.} =
-  ## Creates a new Tensor on Cpu backend
+proc newTensorUninit*[T](shape: varargs[int] = [0]): Tensor[T] {.noinit, inline.} =
+  ## Creates a new uninitialized Tensor of type `T` on the Cpu backend
   ## Input:
-  ##      - Shape of the Tensor
-  ##      - Type of its elements
+  ##      - Shape of the Tensor (defaults to an empty rank-1 tensor)
   ## Result:
   ##      - A Tensor of the proper shape with NO initialization
-  ## Warning ⚠
-  ##   Tensor data is uninitialized and contains garbage.
+  ## Warnings ⚠:
+  ##      - Tensor data is uninitialized and contains garbage.
+  ##      - If no shape is provided, a 1D tensor of size 0 is created.
+  ##        It is possible to create a rank-0 tensor by explicitly
+  ##        providing an empty shape `[]` (e.g. `newTensorUninit[float]([])`).
   var size: int
   initTensorMetadata(result, size, shape)
   allocCpuStorage(result.storage, size)
@@ -42,14 +44,13 @@ proc newTensorUninit*[T](size: int): Tensor[T] {.noinit, inline.} =
   result = newTensorUninit[T]([size])
 
 proc newTensorUninit*[T](shape: Metadata): Tensor[T] {.noinit, inline.} =
-  ## Creates a new Tensor on Cpu backend
+  ## Creates a new uninitialized Tensor of type `T` on the Cpu backend
   ## Input:
   ##      - Shape of the Tensor
-  ##      - Type of its elements
   ## Result:
   ##      - A Tensor of the proper shape with NO initialization
-  ## Warning ⚠
-  ##   Tensor data is uninitialized and contains garbage.
+  ## Warning ⚠:
+  ##      - Tensor data is uninitialized and contains garbage.
   var size: int
   initTensorMetadata(result, size, shape)
   allocCpuStorage(result.storage, size)