I tried to run some tutorials in VTA, and I reached to the step to customize it.
meanwhile, I found a simple question how we handle float computation in VTA?
it already provides some custom data type like float32 or users needs to use integer as a fixed point, or quantized somehow?
@tkclimb currently we support mainly fixed point computation of 8bit ints or less because it’s a very efficient use of FPGA resources.
However, it’s possible to stamp out floating point functional units to do processing in the floating point domain (either single-precision or half precision). It should be pretty straightforward to do, but expect throughput to take a big hit due to lower effective bandwdwidht, storage and computation capacity.
If you want to run experiments on floating point VTA, I suggest we start a Github issue to add such support. I can help guide you through the process of adding hardware support and compiler support.
Finally do you have an FPGA platform in mind?
thierry, is your reply still valid or has there been any enhancements to VTA for float computation