Arm architecture provides optional FP16 floating point support in two alternative formats, IEEE and an an alternative Arm format.

The ACLE (Arm C Language Extension) defined preprocessor symbol __ARM_FP16_FORMAT_IEEE can be used to distinguish between implementations providing IEEE and the Arm alternative format, but cannot, on its own, be used to determined if FP16 HW support is actually present.

Testing this preprocessor symbol can lead to undefined __floatdihf at runtime on an aarch64 target where no FP16 HW is present.

The relevant preprocessor symbol to determine whether FP16 HW support is present in the target is __ARM_FEATURE_FP16_SCALAR_ARITHMETIC, this symbol implies  __ARM_FP16_FORMAT_IEEE.

The relevant preprocessor symbols are defined by the ACLE standard, section 5.5.21 16-bit floating-point data processing operations, https://static.docs.arm.com/101028/0008/Q2-ACLE_2019Q2_release-0008.pdf

The existing sequence of pip install commands fetches and installs
torch==1.0.1.post2 then fetches an unpinned version of torchvision,
recent torchvision packages hardwire the specific torch version they
depend on, the overall effect is that we install a pinned torch
version then replace it with whatever version the torchvision package
depends on.

The most recent torchvision==0.4.1 package results in some test case
failures.

This patch pins torchvision back to 0.4.0, the most recent version
that the test suite worked.  Removing the explicit torch install
because it is implied and pinned as dependency of torchvision.

Change-Id: Ib30bf6aed79ff130ea15ef5134fefb0508790574

* [TOPI][x86] Cascade lake support.

* Jenkins test debug 1.

* Testing cascade lake alone.

* Improve build error when no lowered funcs

* Switch from fatal to warning

* [RUNTIME] Refactor object python FFI to new protocol.

This is a pre-req to bring the Node system under object protocol.
Most of the code reflects the current code in the Node system.

- Use new instead of init so subclass can define their own constructors
- Allow register via name, besides type idnex
- Introduce necessary runtime C API functions
- Refactored Tensor and Datatype to directly use constructor.

* address review comments

* add and fix gradients

* fix linter issues

* Fix infer type of kernel in dense.

* - Moving the check of weight being nullptr up as it is needed in both the branches now.
- Adding test case for validating that data dtype and kernel dtypes can be different.

* - Fix the dtype check for weight. If the weight is not present then we will use the data dtype.

* [Relay][AlterOpLayout] NHWC to NCHWc pad operator.

* Fixing culprit.

* Flaky test 1.

* Flaky test 2.

* [RUNTIME] Introduce new object protocol.

This PR introduces a new object protocol to unify the node and object.
We also updated the existing runtime::vm code to make use of the new system.

Update to the node will be done in a follow up PR.

Other changes:

- Remove object related code in json serializer as that code logic was not complete
  and we have a separate serializer for VM, can revisit later.

* address review  comment

* Fix the child slot logic

This implementation provides cast to limited number of dtypes
that tflite currently supports for placeholder op. Add INT64 in the
possible dtypes as it appears to be supported accrording to tlfite schema.

* overload half operators for cuda codegen

* add float16 te test_op_level1

* fix test_op_level1.py

* fix lint

* disable fp16 test if gpu does not support

* disable fp16 test if gpu does not support

* bypass float16 test if gpu does not support float16

* Fix layer norm for fp16

* [Fix] Fix arange for fp16

* [Fix] Fix mxnet frontend for fp16

* [Fix] Fix arange for fp16

* remove comments

* x

* fix nnvm

* [tvm][any] broadcast with values other than 1

* Add test for incompatible runtime values

* Remove hybrid script compact buffer binding

* retrigger ci

* update build_gcn.py tutorial

updates
* support bias in GCN layer
* download pretrained gcn model
* verify model accuracy
* use time_evaluator to measure runtime

* fix adding bias in gcn layer

* remove printing output

* fix small bug

* add DGL-PyTorch comparison into the build_gcn tutorial

* add accuracy testing

* adjust import order

* handle different dgl versions

* update number for dgl version checking

* Add FIFO buffer op to enable explicit computation re-use in convolution

* Add a test

* Add end-to-end test with 1D convolution

* Add a stub in MXNet frontend

* Address reviewer comments

* Add back stub for MXNet frontend

* app init push

* fix on readme

* change name, add bit serial explanantion

* rm serialLoadMM, change doc

* syntax change for readme

* add parallel test functionality

* fix readme

* add python doc

* syntax

- Adding support for Mxnet flavored dequantization for both default and using MKLDNN. User can choose between the two at runtime. (#3945)

- Added tests for new methods added.

* move the number of nodes constraint in op fusion up to the dom tree level

* add test case of limiting the max number of ops to be fused

* uncomment other test cases

The torch package depends on the future package but the torch wheel
does not expose that dependency resulting in an inconsitent install.

Ideally the wheel should declare all of its dependencies, I'm not sure
why the packagers have choosen not to do this, for now the simple work
around is to explicitly install the future package.

Change-Id: Ic9f0f4bb4c78ab65706fc1b20c1b4fd287856a9e

* [Relay][VM] Fix constant folding issue in VM compiler

1. allow pass params when compile a module
2. enhance profiler robustness

* remove dead code

* fix lint

* add get_params

* fix test

* don't pass params back

* remove get_params

* docs

* move compile function to api

* compile clashes with builtin name

* fix compilation error

* remove dead code

* add valid autotune

* fix pylint

The current bounds checking infrastructure inserts checks like:

```
for (i, 0, bounds[n]) {
  if (likely(i < bounds[n]) {
     ...
  }
}
```

into the TVM IR which is currently not removed by simplification infrastructure.
This is a little unclean, as these are trivially true since for a loop var `i`
with a given min and extent, we are guaranteed that `i >= min` and `i < min +
extent`. Thus, we can insert these checks into the IR and use them to eliminate
trivial bounds checks early on.