Dear Qualcomm Support Team,
I am currently working with SNPE user buffers on the DSP platform, specifically on the device SM8550, and operating in the LOW_POWER_SAVER mode. I have encountered some unexpected behavior regarding the runtime performance of different user buffer encodings, and I would appreciate your insights on the following:
- Runtime Performance Across Different Buffer Encodings:
- Observation: I have noticed that the runtime for UserBufferEncodingTf8 is approximately equal to that of UserBufferEncodingTfN and UserBufferEncodingFloat. This is based on measurements focusing solely on the execute runtime, excluding any pre and post-processing times.
- Question: Given that I am using 8-bit quantized buffers (and have converted the data according to the quantization parameters myself), I expected a reduction in execution runtime with UserBufferEncodingTf8. Why is there no noticeable difference in runtime between the 8-bit quantized buffer and the float buffer? Is this expected behavior, or could it indicate an issue with my implementation or understanding of the system?
- Difference Between Buffer Encodings and Issues in Buffer Creation:
- Question: What are the key differences between UserBufferEncodingUnsigned8Bit and UserBufferEncodingTf8 in terms of their use and performance implications?
- Issue: I am unable to create a user buffer with UserBufferEncodingUnsigned8Bit using the same method (ubFactory.createUserBuffer) as I did for other encodings, as it returns a null pointer. Is there a specific procedure or consideration for creating buffers with this encoding?
Any guidance or advice you can provide on these matters would be greatly appreciated. Understanding these aspects is crucial for optimizing our use of SNPE on the DSP platform.
Thank you for your assistance.
Best regards,
Alona Rimon