arm_biquad_cascade_df2T_f32 Neon version is not working (and scalar version is re-enabled in library)

[mw-kakkala]

 We are utilizing CMSIS-DSP version 1.14.3 to produce optimized code through the MATLAB workflow, targeting ARM Cortex-A
 family devices with CRL.  We've enabled the 'ARM_MATH_NEON' flag to leverage the NEON SIMD instruction set. 
 However, after generating the optimized code using this flag, the results do not align with the simulation outcomes. 
 Conversely, when we use the default C implementation of the 'arm_biquad_cascade_df2T_f32()' function from CMSIS-DSP,
  the results match the simulation. Please let me know if there are any updates regarding this issue.

Regards,
Kishor

[christophe0606]

@mw-kakkala Do you have a reproducer you can share ? Like some filter coefficients and a sample input ?

Note that in Neon mode, arm_biquad_cascade_df2T_init_f32 do not take the filter coefficients directly. You need to use arm_biquad_cascade_df2T_compute_coefs_f32 to generate a new array that can be passed to the init function

It is detailed in the documentation for arm_biquad_cascade_df2T_init_f32

[christophe0606]

I see latest version of CMSIS-DSP have disabled the Neon version of this kernel. So, it looks like it indeed has issues.