* Support vector operations for AMD (llvm IR)

* fix whitespace

* update comments, docstring

* inline AMD GPU functions

* feat(docs) add how_to for tvm install with nnpack support

* feat(docs) change python package paragraph

* feat(doc) remove unsure sentence

* add comments on nnpack usage vs TVM

* remove mxnet nnpack tips for nthread change

* Support vector operations for AMD (llvm IR)

* fix whitespace

* update comments, docstring

Change minimum 32-bit restriction for floating point types to 8-bit.
This change is to enable reduced precision types that may use vector operations underneath the hood (cases #lanes > 1 such as half4).

* enable popcount intrin

* fix lint

* add test

* fix python3

* removed fma dispatch

* added comments to explain why remove fma

* fix lint

* use fmuladd intrin for fma dispatch

* view llvm ir and gcn asm with module.get_source(...)

* fix lint

* [BUFFER] Smarter slice to detect compactness

* move simplify of begins early

* update topi/cuda schedules to use target.max_num_threads

* allow num_thread to be larger than cuda.max_num_threads

* remove get_max_num_threads and make it inline

* add friendly tips when not found cl and link

* fix lint

* added math function support

* bug fix extern func call in llvm based codegen

lint fix

fix build

bug fix extern func call in llvm based codegen

* moved rocm bitcodes detection to python

use `object.__eq__`(default object identity comparison) as default
implementation of same_as. This should be OK since `EqualOp` and
`NotEqualOp` are pure Python object, `object.__eq__` is sufficient.

* [PYTHON] Improve equal sugar

* fix comment

* [ARITH] More caninical simplfy

* [DEBUG] Use HalideIR with trace logging

* [FIX] Fix target warning

* [FIX] Deduplicate options

* Fix

* Fix

* [CODEGEN] Allow link additional module

* fix py3

* add register back

* [CODEGEN] Force not inline compute core for better debug

* also support llvm4

* [TVM] Introduce target generic dispatch system

* fix target warning

* [CODEGEN] Detect broadcast(cast(x)) pattern in FMA

* [CODEGEN] Improve

* [CODEGEN] Fix

* Add same_as to NodeBase

1. Most class inherited from NodeBase(Schedule, Stage, etc) still have
the convenience of using '==' for object identity. And this is the right
behavior for non-Expr classes.
2. subclasses of ExprOp now create EQ expression when '==' is used.

`__nonzero__` and `__bool__` in EQ and NE is a comprise that in some cases
object identity semantics is still useful, like in unit test. For instance:
````
assert a == b
````

"a == b" will create EQ expression, assert then calls `__nonzero__` of the
result expression. `Expr.__nonzero__` throws exception since it prohibits
evaluating IR expression.

More complex case like:
````
assert a in b # b is dict
````

it will call `__eq__` on a and all keys of b, then `__bool__` on the result
expression. This could not easily be done by same_as.

* Retain __hash__ from NodeBase in Python3

* added support for rocm gpu autodetect

* changed type casting from old style to static_cast

* fixed code to generate gfx specific code object

* fixed namespaces