Non-contiguous Tensor

Currently, most operator APIs support "non-contiguous tensor" as the input aclTensor, meaning that a tensor can be represented by (shape, strides, offset).

NOTE An aclTensor can be created through "Common APIs > aclCreateTensor" in Operator Library API.

Example 1

For example, a tensor with shape=(6, 5), strides=(10, 1), and offset=22 has the following memory layout:

a_0,0 , a_0,1 , a_0,2 , a_0,3 , a_0,4 , a_0,5 , a_0,6 , a_0,7 , a_0,8 , a_0,9 a_1,0 , a_1,1 , a_1,2 , a_1,3 , a_1,4 , a_1,5 , a_1,6 , a_1,7 , a_1,8 , a_1,9 a_2,0 , a_2,1 , a_2,2 , a_2,3 , a_2,4 , a_2,5 , a_2,6 , a_2,7 , a_2,8 , a_2,9 a_3,0 , a_3,1 , a_3,2 , a_3,3 , a_3,4 , a_3,5 , a_3,6 , a_3,7 , a_3,8 , a_3,9 a_4,0 , a_4,1 , a_4,2 , a_4,3 , a_4,4 , a_4,5 , a_4,6 , a_4,7 , a_4,8 , a_4,9 a_5,0 , a_5,1 , a_5,2 , a_5,3 , a_5,4 , a_5,5 , a_5,6 , a_5,7 , a_5,8 , a_5,9 a_6,0 , a_6,1 , a_6,2 , a_6,3 , a_6,4 , a_6,5 , a_6,6 , a_6,7 , a_6,8 , a_6,9 a_7,0 , a_7,1 , a_7,2 , a_7,3 , a_7,4 , a_7,5 , a_7,6 , a_7,7 , a_7,8 , a_7,9 a_8,0 , a_8,1 , a_8,2 , a_8,3 , a_8,4 , a_8,5 , a_8,6 , a_8,7 , a_8,8 , a_8,9 a_9,0 , a_9,1 , a_9,2 , a_9,3 , a_9,4 , a_9,5 , a_9,6 , a_9,7 , a_9,8 , a_9,9

That is, the tensor is laid out as shown in the dark-colored positions above. This complete tensor is non-contiguous in its memory layout. Strides describe the interval between two adjacent elements along a tensor dimension. If the stride along dimension 1 is 1, that dimension is contiguous; if the stride along dimension 0 is 10, then adjacent elements are separated by an interval of 10 elements, meaning the dimension is non-contiguous. Offset indicates the displacement of the first element of this tensor relative to addr.

Example 2

For example, consider a tensor with shape=(4, 3), strides=(20, 2), and offset=22. Its memory layout is as follows:

a_0,0 , a_0,1 , a_0,2 , a_0,3 , a_0,4 , a_0,5 , a_0,6 , a_0,7 , a_0,8 , a_0,9 a_1,0 , a_1,1 , a_1,2 , a_1,3 , a_1,4 , a_1,5 , a_1,6 , a_1,7 , a_1,8 , a_1,9 a_2,0 , a_2,1 , a_2,2 , a_2,3 , a_2,4 , a_2,5 , a_2,6 , a_2,7 , a_2,8 , a_2,9 a_3,0 , a_3,1 , a_3,2 , a_3,3 , a_3,4 , a_3,5 , a_3,6 , a_3,7 , a_3,8 , a_3,9 a_4,0 , a_4,1 , a_4,2 , a_4,3 , a_4,4 , a_4,5 , a_4,6 , a_4,7 , a_4,8 , a_4,9 a_5,0 , a_5,1 , a_5,2 , a_5,3 , a_5,4 , a_5,5 , a_5,6 , a_5,7 , a_5,8 , a_5,9 a_6,0 , a_6,1 , a_6,2 , a_6,3 , a_6,4 , a_6,5 , a_6,6 , a_6,7 , a_6,8 , a_6,9 a_7,0 , a_7,1 , a_7,2 , a_7,3 , a_7,4 , a_7,5 , a_7,6 , a_7,7 , a_7,8 , a_7,9 a_8,0 , a_8,1 , a_8,2 , a_8,3 , a_8,4 , a_8,5 , a_8,6 , a_8,7 , a_8,8 , a_8,9 a_9,0 , a_9,1 , a_9,2 , a_9,3 , a_9,4 , a_9,5 , a_9,6 , a_9,7 , a_9,8 , a_9,9

That is, the tensor is laid out as shown in the dark-colored positions above. This complete tensor is non-contiguous in memory layout. Strides describe the spacing between adjacent elements along a tensor dimension. If the stride along dimension 1 is 2, there is one element between adjacent elements along that dimension. If the stride along dimension 0 is 20, adjacent elements are spaced 20 elements apart, meaning the dimension is non-contiguous. The offset indicates the displacement of the first element of this tensor relative to addr.