* [BUILD][LLVM] Support LLVM mainline 5.0 6.0

* Reduce parallelism

* [NNPack] Support for threadpool

* fix lint

* fix lint

* Use static class function

* [Contrib] CuDNN v7 Support

* Add test

[TOPI] Add broadcast and reduce operators

* [WIP] C++ topi contributions

Summary:
This diff implements C++ topi contributions for:
  - relu with parametrix threshold
  - pad with generic padBefore / padAfter specification
  - matmult with transposes
  - conv2d_nchw, conv2d_hwcn with runtime constant padding and strides
  - depthwise_conv2d_nchw with runtime constant padding and strides
  - group_conv2d_ngchw with runtime constant padding and strides
  - broadcast_to a broadcastable shape
  - broadcast_bop where bop is an usual binary op (+ - * / %)

Convolution padding is implemented using the pad operation.
To avoid extra memory consumption, it is generally recommended to inline the padding with the autoinliner.
Unfortunately in its current form the elemwise checks are too restrictive to allow inlining.
So this diff also proposes an extension to LHS injective (i.e. no reduction axis in the current IR design)

Test Plan:
Tested in C++ testsuite in a separate repository, I am looking for suggestions to quickly spin up some tests for tvm.

Reviewers: tqchen

Subscribers:

Tasks:

Tags:

Blame Revision:

* Review + Lint + GSG C++

* [PASS][FIX] Fix LiftAttrScope with if

* [PASS] Fix on proc sync

* fix

* [TEST] Upgrade gpu docker to cudnn7

* fx

* [RUNTIME][PASS] Allow declare vector type array

* fix bcast

* [BUFFER] Enable vload/store function in buffer

* ok

* [NNPACK] Add nnpack.convolution

* Add instrinsic

* Fix lint

* [PASS] Refactor thread storage sync to a common visitor

* Fix the sync scope check behavior

* [SCHEDULE] Fix fuse node order

* Make fuse order consistent with split

* [GRAPH] Add GraphHandle

* Move to apps/graph_executor

* [SCHEDULE] Remap the cached bind_scope.

* more fix

* [STORAGE][BUFFER] Support access ptr for clear access pattern.

* fix lint

* [SCHEDULE] Detect useless bound early

* fix

* prefetch interface added

* prefetch python comments modified. prefetch info data structure maintained.

* start injecting prefetches. first step (domain touch) implemented.

* domain touch tested.

* Prefetch ir_mutator and ir_visitor dispatch registered.

* modify domain touched from passing a func_ref to passing a tensor

* modify domain touched from passing a func_ref to passing a tensor

* modify Tensor copy to Tensor ref

* temp commit for rebase

* debug info removed, typo fixed, ready to rebase

* prefetch flatten test add!

* roll back builtin functions to side effect functions

* lint error fixed!

* add cache line size to storage flatten argument

* forgot modifications add

* change code style to dmlc-like; get rid of can_prove, use manually compute instead

* python lint error fixed

* modify instrinsic name to pass tests

* [TEST] get rid of str(), replace them by accessing attributes

* change map to list comprehension

* redundant numpy import removed