chocolate42

Added many luma variants and created a combination for level 0 that heavily-focuses on alpha. The new combination has no fastpath yet: ``` # Grand total for images decode_ms encode_ms...

- Reworked qoipcrunch to remove the iterator and instead use a curated list (one for rgb one for rgba) of opstrings, the idea being that opstrings higher up the list...

- Implemented luma232 bias from here in addition to the basic version that already existed: https://github.com/nigeltao/qoi2-bikeshed/issues/33 - Implemented delta from average (of previous and above pixel) from here: https://github.com/nigeltao/qoi2-bikeshed/issues/34 -...

I think you're completely right. The current state I have simplifies things by unifying qoipcrunch_rgb and qoipcrunch_rgba into a single list, so 3/4 channel input is no longer handled differently...

- Unified the crunch lists into one as described above. As well as making behaviour more predictable it'll make fastpath implementations more worthwhile to implement as they'll be used more...

The good thing about INDEX8 is that it only takes up one opcode making it hard to beat on efficiency. In theory 95% of the time INDEX8 is more beneficial...

Here's the result: ``` Adding 00030a0b0f1114 to make set size of 1 brings total size to 1618345376 Adding 000206090b0e0f1113 to make set size of 2 brings total size to 1582208112...

Early LZ4 entropy-coding results. LZ4 is not in the repo yet as support is only implemented for qoipbench but the results are right here so why not post them. Entropy...

> Sorry for not being clearer. I wasn't suggesting opstrings just without OP_INDEX8. I was suggesting opstrings without any OP_INDEXN for all of the Ns. Generation currently uses as-big-a 1-byte...

I looked into adding d28 ops but that was back when things were being fiddled with manually and not much benefit was noticed. Things are different now so they should...