XSIMD backend produces bad sounding plugin on Dwarf platform, EIGEN is not

[MaxPayne86]

This can be replicated with https://github.com/moddevices/mod-plugin-builder/blob/f73345bc16e64b048aa06df7bac3218ba9094782/plugins/package/aidadsp-lv2/aidadsp-lv2.mk#L7

-AIDADSP_LV2_VERSION = 0bac2f3feb811115d696b11ec2293f5f03a535ac
+AIDADSP_LV2_VERSION = 6e54495ce95bcbaf2432ef4a3931299492b14671
...
-AIDADSP_LV2_CONF_OPTS = -DRTNEURAL_EIGEN=ON
+AIDADSP_LV2_CONF_OPTS = -DRTNEURAL_XSIMD=ON

this is not happening on other build systems like yocto and this could help in identifying the issue.

The issue should be related on how MPB builds xsimd backend in RTNeural submodule and not on plugin source code.

@falkTX

[falkTX]

Your own yocto file has a very interesting line:

-funsafe-loop-optimizations causes the plugin to sound really bad, excluding

Maybe it is this simple...

[falkTX]

Yep, that seems to be the case. Fixed in https://github.com/moddevices/mod-plugin-builder/commit/a41eac68c0a62fa2a1280369489a864ff2c9884b Took the chance to enable LTO as well, for a little extra performance benefit.

[falkTX]

Might have spoken too soon. behaviour improves but issue still there.

[MaxPayne86]

@falkTX being a distortion effect, discussing about sound distorted may became an issue: the problem creates a corrupted sound similar to random noise and crackling mixed with the distorted sound can you confirm? Thanks a lot

[falkTX]

There is a high-pitch crackling noise still present. Still investigating...

[falkTX]

Something relevant to xsimd vs eigen.. we actually have better performance when using eigen backend. So perhaps the whole deal is a bit pointless

[MaxPayne86]

I agree to stay with eigen for now, the only replicable solution to me could be

• if in RTNeural we have a benchmark test for backends using static models

need to check