Final AI Engine Algorithm Performance - 2025.1 English - XD100

The impact of all of the algorithm adaptations outlined above was assessed by integrating Vitis Functional Simulation models of four prototype AI Engine graphs into the baseline MATLAB model as follows:

• The IFFT size change was implemented directly in MATLAB code.

• AI Engine prototype graphs were constructed for the cos() and sin() instantiations of the func_approx() Vitis DSP Library IPs.

• An AI Engine prototype graph was constructed for the fmod_floor() function.

• An AI Engine prototype graph was constructed for the sqrt() instantiation of the func_approx() library element. MATLAB code was written to mimic the required pre-scaling and post-scaling around it.

Based on these behavioral AI Engine algorithm models, the performance of the AI Engine implementation was run on the same GOTCHA data set as the MATLAB baseline. The SSIM metric between the original MATLAB baseline and the AI Engine model was evaluated using the built-in ssim() MATLAB function. The results are shown in the following figure. There is good agreement between the baseline model and the AI Engine implementation. The algorithm adaptations introduce a minor discrepancy into the output images which are very difficult to see visually and lead to an SSIM of 0.9966 which is deemed to be an acceptable level of degradation.