perf(fromArrayBuffer): use less memory for large buffers

[raphinesse]

Platforms affected

All

Motivation and Context

This improves performance for converting large files to Base64 strings (see apache/cordova-plugin-file#364 for more info, especially the extensive performance analysis by @LightMind).

Another win here, is that we reduce the code size considerably. A downside is that we actually loose a bit of performance for small instances, but I do not think that this would be noticeable.

I explored different approaches here, like converting the original algorithm to use ArrayBuffers. The problem here is that we still need to convert the end result to a string which then becomes a performance bottleneck if you do not have support for TextDecoder (which we cannot assume with our current ES5 target).

Fixes #241

Description

base64.fromArrayBuffer now uses btoa to convert bytes to a base64 encoded string. We already use its counterpart atob in base64.toArrayBuffer. Since btoa unfortunately operates on binary strings instead of buffers, we first need to convert the raw bytes to a binary string. This is the main performance bottleneck here, but applying String.fromCharCode to large chunks of data works reasonably well.

Testing

I added a test that should hopefully preventing people from making stupid changes in the future. However, its reliance on an absolute expected runtime might lead to problems in the future.

I also did some performance comparisons (ops/s) between the new and old version in my local Chrome browser:

bytes	old	new	new/old
1K	61409	33029	54%
32K	1569	913	58%
1M	11	24	218%
32M	.3	.9	300%

[codecov-commenter]

Codecov Report

Merging #242 (691ccc1) into master (af904b7) will decrease coverage by 0.57%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #242      +/-   ##
==========================================
- Coverage   84.23%   83.65%   -0.58%     
==========================================
  Files          14       14              
  Lines         539      520      -19     
==========================================
- Hits          454      435      -19     
  Misses         85       85              

Impacted Files	Coverage Δ
src/common/base64.js	100.00% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update af904b7...691ccc1. Read the comment docs.

[erisu]

Looks OK to me

[raphinesse]

For future reference: replacing bytesToBinaryString(array) with new TextDecoder('utf-16').decode(Uint16Array.from(array)) doubles the throughput for large files in my experiments. In ten years we might be able to use it 😉

[ath0mas]

Hello, should it be fine and time to switch to new TextDecoder('utf-16').decode(Uint16Array.from(array)) as you suggest now?

Shipping it in next major release seems right given:

• v7 already started a year ago (feat: bump version 7.0.0-dev #258)
• TextDecoder browser compatibility
• mobile target sdks upgrading every year
• cordova cli and android major updates already containing Breaking Changes and updating min sdk meaning renewed browser compatibility