Latency experiments on Qualcomm Snapdragon 865

[xzz777]

您好，我使用您源码转onnx，之后转mnn，在Galaxy S20 Ultra （Qualcomm Snapdragon 865）上，线程设置为1，精度为fp32，进行测速，其中有些算子不支持，无法完成测速。
另外，我使用同样的流程测试TopFormer-B-Full的延迟，比文中汇报的大了很多（1414ms vs 749 ms）。经过逐层分析，主要是resize上采样8倍之后的softmax操作和argmax操作延迟占比很大（约1000ms）。请问在您测速的过程中是否对这部分有所调整和优化？
顺便，请问论文的延迟是在什么机型上测试的？同cpu不同机型对延迟的影响大吗？
期待您的回复，祝好。
Hello, I use your source code to convert onnx and then mnn. I measure the latency on a Galaxy S20 Ultra (Qualcomm Snapdragon 865), the thread is set to 1, and the precision is fp32. There are some unsupported operators , so the speed measurement cannot be completed.
In addition, I used the same process to test the delay of TopFormer-B-Full, which is much larger than that reported in the article(1414ms vs 749 ms). After layer-by-layer analysis, it is mainly because the delay of the softmax operation and argmax operation after resize upsampling by 8 times accounts for a large proportion (about 1000ms). May I ask if you have adjusted and optimized this part during your speed measurement?
By the way, what type of mobile phone is the delay of the paper tested on? Do different phone types with the same CPU have a big impact on latency?
Looking forward to your reply, best wishes.