Improve sexp_buffer_from_stream performance

[ChiaMineJP]

Summary

When I develop clvm-js as a JavaScript implementation of clvm, I was suffered in poor performance in test_very_deep_tree at tests/serialize_test.

Later I found out that the root cause was huge number of memory allocation/copy invoked upon byte deserialization of sexp instance.

Looking back to clvm of original Python implementation, I noticed that one of the optimizations I applied to my JavaScript implementation can also be applied to the Python's.

So here I am proposing the pure performance enhancement of sexp_buffer_from_stream.

Performance measurement

Test environment

-	-
OS	Windows10 Pro
CPU	AMD Ryzen9 5900X
MEM	64GB@3200MHz
Storage	NVMe m.2 SSD Corsair MP600
Python	3.7.8

# Powershell
cd clvm/tests
for($i=0; $i -lt 10; $i++){ python -m unittest serialize_test.SerializeTest.test_very_deep_tree }

Result

Roughly 10-20% faster than before.

test_very_deep_tree

n	Before the PR	After the PR
1	5.023s	4.014s
2	4.696s	3.996s
3	4.882s	4.021s
4	4.403s	3.993s
5	4.596s	3.997s
6	4.561s	3.983s
7	4.453s	3.987s
8	4.442s	3.998s
9	4.493s	3.996s
10	4.468s	3.990s
AVG	4.602s	3.998s

P.S.

If you're interested, I also compared performance between clvm of Python and clvm of JavaScript.

Test execution

Python

# Powershell
cd <directory>
git clone https://github.com/Chia-Network/clvm
cd clvm/tests
for($i=0; $i -lt 10; $i++){ python -m unittest serialize_test.SerializeTest.test_very_deep_tree }
for($i=0; $i -lt 10; $i++){ python -m unittest serialize_test.SerializeTest.test_very_long_blobs }

JS

# Powershell
cd <directory>
git clone https://github.com/Chia-Mine/clvm-js
cd clvm-js
git checkout v0.0.19
yarn # Or npm install
for($i=0; $i -lt 10; $i++){ yarn test serialize_test --testNamePattern test_very_deep_tree } 
for($i=0; $i -lt 10; $i++){ yarn test serialize_test --testNamePattern test_very_long_blobs } 

Result

test_very_deep_tree

	Python	JS
1	4.014s	2.29s
2	3.996s	2.3s
3	4.021s	2.288s
4	3.993s	2.251s
5	3.997s	2.254s
6	3.983s	2.289s
7	3.987s	2.253s
8	3.998s	2.26s
9	3.996s	2.301s
10s	3.99s	2.282s
AVG	3.998s	2.277s

test_very_long_blobs

	Python	JS
1	0.454s	0.791s
2	0.455s	0.741s
3	0.478s	0.77s
4	0.466s	0.798s
5	0.474s	0.799s
6	0.479s	0.807s
7	0.467s	0.818s
8	0.488s	0.751s
9	0.423s	0.758s
10s	0.47s	0.743s
AVG	0.4654s	0.7778s

[arvidn]

the python implementation is not currently used by the full node itself, so its performance isn't critical (the full node uses clvm_rs). However, it is used by our tooling, and this seems like a straight-forward improvement.

[richardkiss]

Thank you!