search

doonny / PipeCNN

An OpenCL-based FPGA Accelerator for Convolutional Neural Networks
Apache License 2.0
1.23k stars 370 forks source link

difference between opencl for gpu and opencl for fpga #97

Open Thilanka97 opened 5 years ago

Thilanka97 commented 5 years ago

@doonny do you have an idea about the difference between, coding targeting fpga using opencl and coding targeting GPU and CPU using opencl ? How different are they ? what are the main differences. If you could help me with this, it would be a great help. please be kind enough to help me.

Thanks in advance!

doonny commented 5 years ago

OpenCL for FPGA is a new way for modeling macro architecture/digital circuit in high-level programming langurage OpenCL for GPU is still parallel programing on SIMD processor

Thilanka97 commented 5 years ago

@doonny Thank you very much for the reply! I am trying to implement yolo or tiny yolo on fpga using opencl, for that I am gonna use a yolo opencl implementation that was coded to run on gpu and cpu. But I dont have much of a idea about what changes do I need to make to be able to run on an fpga.

I am trying to read and understand your code also so that I will be able to get an idea of how to implement cnn on opencl. Do you have an idea about this ?

Thanks in advance!

doonny commented 5 years ago

Hi, you will find a proper way to implement YOLO based-on PipeCNN for sure, because we have verified that. The first thing to do is to fully understand the computation flow of YOLO and PipeCNN,.

Thilanka97 commented 5 years ago

@doonny thanks for your valuable advice! Now I am trying to understand the flow of yolo and pipecnn. from what I understand so far, I need to add more kernel functions to the kernel code other than the main 4 kernel functions you have for alexnet and vgg net. And also I need to change layer config file. And have to change the host file according to the changes I make in the kernel code right ? Am I correct ? Is there anything other than this that I need to do ? with these changes would I be able to run tiny yolo real time of fpga ? Please guide me.

Thanks in advance!

doonny commented 5 years ago

Basicly for YOLO, you only need to modify the convolution kernel to accelerator the algorithm, since the conv operations consumes most of the execution time. You don't really need new kernels.