libjxl
more speedups for fjxl decoding
Same as https://github.com/libjxl/libjxl/pull/1149 but now without the change to ClampedGradient, which wasn't helping much anyway and it apparently wasn't safe when the numbers use the full int32_t range, like in the lossless pfm conformance test case.
Hm, strange, seems like compared to current git main branch, this doesn't really give much of a speedup at all anymore:
Before: 3072 x 2048, geomean: 39.96 MP/s [29.81, 41.69], 30 reps, 0 threads. 3072 x 2048, geomean: 68.28 MP/s [45.65, 72.18], 30 reps, 2 threads. 3072 x 2048, geomean: 102.01 MP/s [58.48, 109.59], 30 reps, 4 threads.
After: 3072 x 2048, geomean: 40.17 MP/s [27.99, 41.99], 30 reps, 0 threads. 3072 x 2048, geomean: 69.07 MP/s [48.04, 73.16], 30 reps, 2 threads. 3072 x 2048, geomean: 102.10 MP/s [64.97, 109.45], 30 reps, 4 threads.
Hm, strange, seems like compared to current git main branch, this doesn't really give much of a speedup at all anymore:
Before: 3072 x 2048, geomean: 39.96 MP/s [29.81, 41.69], 30 reps, 0 threads. 3072 x 2048, geomean: 68.28 MP/s [45.65, 72.18], 30 reps, 2 threads. 3072 x 2048, geomean: 102.01 MP/s [58.48, 109.59], 30 reps, 4 threads.
After: 3072 x 2048, geomean: 40.17 MP/s [27.99, 41.99], 30 reps, 0 threads. 3072 x 2048, geomean: 69.07 MP/s [48.04, 73.16], 30 reps, 2 threads. 3072 x 2048, geomean: 102.10 MP/s [64.97, 109.45], 30 reps, 4 threads.
I'd consider not doing it then :)
Hm, strange, seems like compared to current git main branch, this doesn't really give much of a speedup at all anymore: Before: 3072 x 2048, geomean: 39.96 MP/s [29.81, 41.69], 30 reps, 0 threads. 3072 x 2048, geomean: 68.28 MP/s [45.65, 72.18], 30 reps, 2 threads. 3072 x 2048, geomean: 102.01 MP/s [58.48, 109.59], 30 reps, 4 threads. After: 3072 x 2048, geomean: 40.17 MP/s [27.99, 41.99], 30 reps, 0 threads. 3072 x 2048, geomean: 69.07 MP/s [48.04, 73.16], 30 reps, 2 threads. 3072 x 2048, geomean: 102.10 MP/s [64.97, 109.45], 30 reps, 4 threads.
I'd consider not doing it then :)
Agreed, doesn't make much sense to merge this if there is no real speedup.
I wonder why I was seeing more substantial speed improvements before though (see the numbers in the previous PR), so maybe leave this PR open for a while to remind me to investigate what happened there.