Results of ModelTest (GCC15.2 and MULTIFRAME_8X8_CONVOLUTION to “4”)
[root@moddwarf neural_amp_modeler.lv2]# ./ModelTest
Block size: 64 Quality Scale: 1
Loading models from: “/root/.lv2/neural_amp_modeler.lv2/Models”
WaveNet (A2 Full) Test
Model: “/root/.lv2/neural_amp_modeler.lv2/Models/BossWN-a2.nam”
Internal: 9.7822 (0.558293xRT)
NAM Core: 8.5402 (0.639486xRT)
NAM vs Internal RMS err: 9.74984e-08
Internal is: 0.873034x NAM
WaveNet (A2 Lite) Test
Model: “/root/.lv2/neural_amp_modeler.lv2/Models/BossWN-a2.nam”
Internal: 1.36605 (3.99789xRT)
NAM Core: 1.2163 (4.49013xRT)
NAM vs Internal RMS err: 8.3052e-08
Internal is: 0.890372x NAM
WaveNet (A1 Standard) Test
Model: “/root/.lv2/neural_amp_modeler.lv2/Models/BossWN-standard.nam”
Internal: 10.2996 (0.530246xRT)
NAM Core: 15.019 (0.363628xRT)
NAM vs Internal RMS err: 1.00481e-07
Internal is: 1.45821x NAM
LSTM (1x16) Test
Model: “/root/.lv2/neural_amp_modeler.lv2/Models/BossLSTM-1x16.nam”
Internal: 0.92932 (5.8767xRT)
NAM Core: 1.05275 (5.18767xRT)
NAM vs Internal RMS err: 5.24119e-07
Internal is: 1.13282x NAM
I don’t see many difference with MULTIFRAME_8X8_CONVOLUTION to “0”, or even with GCC9.4 . Only appreciable difference is with MULTIFRAME_8X8_CONVOLUTION to “8”, but looks still far from 1xRT
WaveNet (A2 Full) Test
Model: “/root/.lv2/neural_amp_modeler.lv2/Utils/Models/BossWN-a2.nam”
Internal: 7.661 (0.712875xRT)
Anyway for A2 LITE, the best results are with GCC15.2 + NamCore + Patch277, where the Dwarf CPU with just with the neural plugin goes from around 36,5% to 35,5%.