sample_embedding_folder2/0766948.txt

Ticket Name: CCS/TDA2: EVE and DSP Utilization to run Network Layers in TDA2x/TdA3x

Query Text:
Part Number: TDA2 Tool/software: Code Composer Studio Hello Sir, I need to know, how to split the models across the EVEs and DSPs. Example how to utilize the 2 EVEs sequentially, utilizing 2 EVEs in parallel, EVE followed by DSP and other such configurations. I have deployed Caffe_Jacinto_SSD model on TDA2x by following TIDL_LIBRARY_USER_GUIDE.pdf to split Network layers through making some small changes in "LayergroupId" to run the model across EVE's and DSP's. Kindly suggest some important documents to understand all the configurations, their Implementations and gain Expertism in splitting Networks across EVE's and DSP's. Thanks and Regards, Vyom Mishra

Responses:
Hi Vyom, can you check FAQ #21 and #22 in TIDeepLearningLibrary_UserGuide.pdf and see if they can help you? Regards, Yordan

Hello Sir, I have mentioned above that, I have already deployed SSD by splitting it across the EVE and DSP, by following the FAQ 21 and 22 only. So, I am requesting you to provide more on it, If there are some more documents available, to explore. Thanks and Regards' Vyom Mishra

Hi Viyom, Can you clarify you question further. You mentioned that you could deploy network by setting layer group ID. Is your your question specific to uses case development for the network or optimally setting layer group ID

Hello Sir, Basic splitting of the SSD model across EVE and DSP is done by refering the User Guide(FAQ 21 and 22). Now I need your suggestions for below queries. How to utilize the 2 EVEs sequentially, utilizing 2 EVEs in parallel, EVE followed by DSP and other such configurations. Kindly suggest some important documents to understand all the configurations, their Implementations and gain Expertism in splitting Networks across EVE's and DSP's. Thanks and Regards, Vyom Mishra

Dear Sir, I have the following queries with respect to my question above. 1. If model is floating point, In TIDL is the model converted to fixed point for inference on DSP as well? 2. Can the intermediate output of DSP be given to EVE? May not be if DSP output is floating point. Pl confirm. ex: If I have model designed and grouped to run as EVE->DSP->EVE->DSP 3. Any limitations allocation of layerGroupId in import Config 4. Is there a standard way of optimally setting layerGroupId, other than what we know from used guide. Any other doc/experiments Thanks and Regards Vyom Mishra

1. yes. 2. Yse you can pass output of DSP to EVE. DSP also computes CNN layer in fixed point 3.No 4. Try to run all the convolution layers in EVE and Detectionoutput layer in DSP. Other layers you can either run DSP or EVE.