[Object] Introduce POD-C Compliant tvm::Map

[junrushao]

Related discussion: https://discuss.tvm.ai/t/discuss-runtime-array-containers-array-adt-string/4582.

This PR introduces tvm::Map and tvm::MapNode:

1. MapNode is a POD-C compliant alternative to `std::unordered_map<ObjectRef, ObjectRef>;
2. Map is the shared reference to the container which implements copy-on-write semantics.

Features

• Removes dependency to std::unordered_map
• The new impl is POD-C compatible in terms of Itanium C++ ABI
• Copy-on-write semantic is preserved in tvm::Map
• tvm::MapNode is kept to be mutable and iterable

Breaking changes

• MapNode no longer has the field std::unordered_map<ObjectRef, ObjectRef> data, and all its methods currently used are implemented to MapNode itself
• Iterating through MapNode is natively supported directly via interfaces like MapNode::begin()

Performance

To show that the introduction of tvm::Map won't hurt performance, I also did some rough benchmark on my laptop on the success insertion performance.

Hardware: MacBook Pro (16-inch, 2019)
Memory: 8 GB * 2, DDR4, 2667 MHz
CPU: Intel i7-9750H CPU @ 2.60GHz
OS: macOS Catalina 10.15.5
Compiler: Apple clang version 11.0.3 (clang-1103.0.32.62)

Number of Operations	Avg Total Time (ms) tvm::MapNode	Avg Total Time (ms) std::unordered_map
10	0.000309454	0.000996719
100	0.00309778	0.00949672
1000	0.0183351	0.0847953
10000	0.405679	0.926011
100000	4.29819	10.6617
1000000	57.6163	255.175
10000000	825.093	3743.89

Scripts are available on gist. Note that the measured time may depend on operation, randomness, cache line, compiler, etc.

Credits

We use [1] for placing an item on a table of size of power-of-2. We learned that triangle numbers that could iterate through the power-of-2-sized table through [2]. The idea that linked list can be placed in the same table is learned from [3].

[1] Fibonacci Hashing: https://programmingpraxis.com/2018/06/19/fibonacci-hash/
[2] Quadratic probing with triangle numbers: https://fgiesen.wordpress.com/2015/02/22/triangular-numbers-mod-2n/
[3] Flat hash map: https://github.com/skarupke/flat_hash_map

Thank you @mbrookhart for the valuable discussion!

Also CC @tqchen @icemelon9 @yzhliu @zhiics @comaniac @kevinthesun @jwfromm @jroesch if you guys are interested.

[junrushao]

Maybe we should get rid of __builtin_clz as it is the slow path anyway

[zhiics]

@junrushao1994 +1 MoveFrom should be way more expensive than it though.

[junrushao]

Some actionable items per offline discussion with @tqchen:

A1. We move things back to include/tvm/node, instead of runtime. This is mainly due to the concern of keeping tvm runtime minimal. Once we find a strong use case that really need HashMap to exist in tvm runtime, we can bring it back.
A2. We will implement a specialization for small maps due to the concern of memory overhead.
A3. We will have unordered_map as a backup option, which can be turned on using macro or something like that.
A4. We will add a lot of comments on how this hash map works.

So it requires some amount of work. I will turn it back to draft until the work is done :-)

[junrushao]

A1 and A2 is done over the commits above

[junrushao]

A4 is done with the last commit

[junrushao]

A3 is done over the last commits

[junrushao]

@zhiics @icemelon9 could you guys take another look? much appreciated!

[junrushao]

Summary of the change:

• Remove most of the short names (i, n, m, d, b) and use more informative ones instead.
• Added comments between important code blocks (insert vs rehash)
• Add test cases to cover the switch between SmallMapNode and DenseMapNode

[junrushao]

@tqchen Just responded to all the review comments, added test cases and code comments, could you take another look?

[tqchen]

@zhiics please take another look and https://tvm.apache.org/docs/contribute/code_review.html#approve-and-request-changes-explicitly

cc @wweic @spectrometerHBH @comaniac @antinucleon

[zhiics]

Only some nitpicks. Otherwise, LGTM.

[junrushao]

@zhiics Could you take another look? Thanks!

[tqchen]

THanks @junrushao1994 @zhiics !