do two TODOs

[jonsneyers]

Does two TODOs, one small one in tf.rs, and a bigger one to use 16-bit HF coeff buffers (only relevant in case of progressive passes) instead of 32-bit ones.

Gives only an almost immeasurable speedup on my laptop, but a substantial memory reduction in case of multiple HF passes:

Before:

Decoded 333921030 pixels in 6.153 seconds: 54.271 MP/s
        6.39s real              6.35s user              0.02s sys
           236355584  maximum resident set size

After:

Decoded 333921030 pixels in 6.147 seconds: 54.322 MP/s
        6.38s real              6.35s user              0.01s sys
           154320896  maximum resident set size

[veluca93]

Approving to get benchmark results as usual ;)

[veluca93]

If we're doing a match anyway, why not make HfCoefficients have a None / I16(...) / I32(...) enum? Also, any reason not to make the tuple into an array of size 3?

[veluca93]

Alternatively, have you considered using three OwnedRawImages and a boolean flag to remember whether it is i32 or i16? This should have a bit less code.

[jonsneyers]

Did the enum thing. With OwnedRawImage it seems hard to avoid unsafe reinterpret casts, no?

[veluca93]

It shouldn't be, there is a safe way to get a typed image view out of a OwnedRawImage ;)
See

jxl-rs/jxl/src/image/typed.rs

Line 113 in 36c9c3b

pub fn from_raw(raw: OwnedRawImage) -> Self {

[veluca93]

and

jxl-rs/jxl/src/image/typed.rs

Line 265 in 36c9c3b

pub fn from_raw(raw: RawImageRectMut<'a>) -> Self {

for the rect version

[veluca93]

Performance Summary (Commit d4958fd)

Machine	Threading	Base MP/s	PR MP/s	Avg Improvement
desktop	Single	79.50	79.36	-0.43% ± 0.33%
desktop	Multi	79.11	79.11	-0.23% ± 0.39%
framework-desktop	Single	92.68	96.39	-0.32% ± 0.30%
framework-desktop	Multi	93.50	97.07	-0.32% ± 0.20%
pixel7a	Single (Fast)	28.57	28.29	+0.13% ± 0.33%
pixel7a	Single (Mid)	20.37	20.86	+1.00% ± 0.37%
pixel7a	Multi	28.47	28.67	+0.51% ± 0.54%

Detailed per-image results

[jonsneyers]

Performance Summary (Commit 0aabc58)

[...]

Probably this is just noise since there's no image where this should make a difference (requires multiple HF passes).
Let's take another look after adding a multi-pass test image.

[veluca93]

Performance Summary (Commit 0aabc58)

[...]

Probably this is just noise since there's no image where this should make a difference (requires multiple HF passes). Let's take another look after adding a multi-pass test image.

It seems to help a bit with progressive images.