Hello,
I am currently using fastcv on Samsung S7 cellphone which is based on Qualcomm 820 Soc.
I am trying to use fastcv to speed up the gaussion filter with fcvFilterGaussian11x11u8 API.
Firstly, I call the fcvFilterGaussian11x11u8 API which runs on ARM CPU by using fcvSetOperationMode(FASTCV_OP_CPU_PERFORMANCE) and get the following result of execution time: 18.8ms.
Then, I switch the mode to Hexagon DSP by using fcvSetOperationMode(FASTCV_OP_LOW_POWER) and get the following result of execution: 52.5ms !
I guess the difference between ARM CPU and Hexagon DSP is too huge to understand. Did anyone have any experience with this?
B.T.W, I am using Java JNI to call the native C function where fcvFilterGaussian11x11u8 is called, and the image size is 1280x1024 with 8bit gray value.
BR
Tao
Yes. I have similar experience. I run an audio filter on Hexagon DSP. It is very slow. By profiling I find out that most time is consumed on FastRPC audio data transfer.
FastCV uses FastRPC data transfer too.
Tduan