[WIP] Buffer strides are required by tensorize

[yzhliu]

If strides are not specified, the vload, vstore stuff will calculate buffer offset as https://github.com/dmlc/tvm/blob/master/src/lang/buffer.cc#L227-L239, but in tensorize, the tensor/buffer is passed from outside, with start address already includes the axes in the intrinsic, thus results the incorrect offset.

For example,

def intrinc(in, outs):
  data = tvm.placeholder((d2,d3), dtype='uint8', name='data')
  data.vload([i2, i3])

Input = tvm.placeholder((d1, d2, d3))
s[Input].tensorize(Input.axis[1], intrinc) 

then the start offset passed to intrinc is (i1 * d2 + 0) * d3 + 0, without strides, vload will (incorrectly) return offset ((i1 * d2 + 0) * d3 + i2) * d3 + i3

It is confusing to users to understand, in most cases, tensorize results will be incorrect without binding a buffer with strides.

Will come with test case and doc improvements.

[tqchen]

document this function

[tqchen]

The strides requirement really specifies what is the requirement of the buffer. If there is no strides, then the data is expected to be continuous, and tensorize should raise an error if the data being matched is not continuous

[tqchen]

@ZihengJiang can you also take a look at this?

[ZihengJiang]

ye, I always feel confused about this too

[tqchen]

Hmm, maybe we should add a detailed document page about the differences

[tqchen]

some followup discussion here as well https://discuss.tvm.ai/t/how-to-use-tensorize/424/6

[yzhliu]

After offline discussing with Tianqi, this is problem with elem_offset calculation itself. Ref: #1762