The decoder time-comsuming is different. about lyra HOT 5 OPEN

Thank you for your jobs.
The readme.md show the decoder time-consuming is
combined_model_and_conditioning stats for generating 2000 frames of audio, max: 13173 us, min: 9463 us, mean: 9629 us, stdev: 270.788.
which means one frame(40ms) consume 9.463ms. But the log printed in my android app is
combined_model_and_conditioning stats for generating 2000 frames of audio, max: 65081 us, min: 53070 us, mean: 54055 us, stdev: 872.812.
which means one frame(40ms) consume 54.055ms.
The result make me confused. Why such a big difference?

from lyra.

What type of device are you using? As the doc says the numbers reported are for a Pixel 4.

from lyra.

I test it in xiaomi max3, android 10 ,the cpu is arm64-v8a and Linux, both of them consume about 50ms per frame.

from lyra.

I test it in xiaomi max3, android 10 ,the cpu is arm64-v8a and Linux, both of them consume about 50ms per frame.

Thank you for your report. We had not tested on Mi Max3. The lowest-end phone that we tested internally is Pixel1, which is still capable of decoding in real time (<40ms time to decode a 40ms frame). Unfortunately Mi Max3's Qualcomm Snapdragon 636 processor is slower than Pixel1's Snapdragon 821.

However, there is still chance for Max3 to be able to decode in real time, by using the fixed point representation (which could give you at most a 2X speedup). This is done by supplying --copt=-DUSE_FIXED16 in the building command. For example for the android app,

bazel build android_example:lyra_android_example --config=android_arm64 --copt=-DBENCHMARK --copt=-DUSE_FIXED16

Let us know if that works for you, and it would be great if you can share the benchmark results.

from lyra.

The new Lyra 1.2.0 release should give another 5x speed-up.

from lyra.