About compute, schedule, pattern

1

hi,

I am confused by compute, schedule and pattern, these three are all nnvm::Op. What’s difference between these three type ops?

Thanks

2

my question is how net/graph is mapping to pattern/compute/schedule?
for example net, params = from_mxnet(sym, arg_params, aux_params), how is net converted into pattern/compute/schedule in compiler.build? Because a lot of pattern have same name as schedule.

Thanks

3

First mxnet symbol is mapped to nnvm symbol (operator), each nnvm operator has its FCompute function registered. We can use TVM compute to define FCompute, e.g., for conv2d, it’s

conv = tvm.compute(oshape, lambda n, oc, oh, ow:
  tvm.sum(A[n, ic, oh*stride+kh, ow*stride+kw] * W[oc, ic, kh, kw], axis=[ic, kh, kw]))

TVM schedule is for optimizing the computing, e.g., reorder the loops, split axis, define cache read/write, etc.

TVM compute and schedule together defines how to calculate, in this case, conv2d operator.

Regarding the “pattern”, do you mean OpPattern? or something else?

4

Yes, pattern mean OpPattern. @yzhliu

5

OpPattern can be viewed as an attribute of an operator. Operators with same pattern can have same strategy of fusion, schedules, etc. For example, schedules element-wise operators on cpu can be pretty simple - just do parallel on the first several axes.