Tensor stride calculation

daunclestone · January 21, 2020, 5:25am

I have a newbie question about tensor stride calculation. I am having problem using the stride values returned by vxMapTensorPatch().

I see that in ownInitTensor() the stride values are calculated as

         tensor->stride[i] = tensor->stride[i - 1] * tensor->dimensions[i - 1];

but in vxMapTensorPatch() the stride values are calculated as

         stride[i] = stride[i - 1] * (view_end[i] - view_start[i]);

this causes the stride values returned by vxMapTensorPatch() to be different from that in tensor->stride[]. Why is the calculation in vxMapTensorPatch() not as follows

         stride[i] = stride[i - 1] * (view_end[i-1] - view_start[i-1]);

Shouldn’t the stride calculation be the same in both places?

thx

daunclestone · June 24, 2020, 7:52pm

dvorotnev (Dmity Vorotnev) · GitHub also spotted this issue and fixed it. see link below.

github.com/KhronosGroup/OpenVX-sample-impl

Incorrect strides in vxMapTensorPatch

opened 01:23PM - 19 May 20 UTC

dvorotnev

bug

Hello! In the function `vxMapTensorPatch` from OpenVX v1.3 calculations for the …strides seem incorrect. For example, for a tensor with data type `VX_TYPE_INT16` `dimensions = {3, 1, 2, 2}` and the function params `view_start = {0, 0, 0, 0}`, `view_end = {3, 1, 2, 2}` it calculates `strides = {2, 2, 4, 8}`: ``` for (vx_uint32 i = 1; i < number_of_dims; i++) { stride[i] = stride[i - 1] * (view_end[i] - view_start[i]); } ``` I think that the strides should be calculated as: ``` for (vx_uint32 i = 1; i < number_of_dims; i++) { stride[i] = stride[i - 1] * (view_end[i - 1] - view_start[i - 1]); } ``` and there are `strides = {2, 6, 6, 12}` for the given example.