Thierry Moreau comments

Results 35 comments of


                                            Thierry Moreau

[ONNX][Python][VM] Generation and saving ONNX model hash for VirtualMachine

CC @adelbertc for now

[ONNX][Python][VM] Generation and saving ONNX model hash for VirtualMachine

CC @vanessahlyan as well

[Hardware][Xilinx] explicitly specify acc dep distance to avoid hidden pitfall

Thanks @zhanghaohit, for catching this error. I agree with the fact that the dependence should not be hardcoded. However in order to not to add too many parameters in the...

[Hardware][Xilinx] explicitly specify acc dep distance to avoid hidden pitfall

Thank you for the explanation. Do you mind breaking down what happens in the 32x32 configuration that causes correctness to fail (since the dependence analysis waive is unbounded)? I agree...

[Hardware][Xilinx] explicitly specify acc dep distance to avoid hidden pitfall

@remotego I just read your comment that must have been posted as I wrote my reply. Your explanation makes sense, ultimately the critical path in the accumulation is READ ->...

[Hardware][Xilinx] explicitly specify acc dep distance to avoid hidden pitfall

I assume that if you have an FPGA that needs to be clocked at a high frequency, you'll likely end up with more cycles.

[Hardware][Xilinx] explicitly specify acc dep distance to avoid hidden pitfall

So to conclude, I agree with you that we may want this parameter to be larger than 2 cycles if the hardware pipeline becomes more complicated. So the big question...

[Hardware][Xilinx] explicitly specify acc dep distance to avoid hidden pitfall

Perhaps one way to look at this is to start with DISTANCE=3 by default. And if the II>1 for the GEMM, we issue a warning telling the user to increase...

[Hardware][Xilinx] explicitly specify acc dep distance to avoid hidden pitfall

I'll think some more about this problem overnight, thanks for bringing attention to this issue @zhanghaohit @remotego

[Hardware][Xilinx] explicitly specify acc dep distance to avoid hidden pitfall

@zhanghaohit I think that the warning approach is a reasonable compromise between user-defined flexibility and making sure that they set the value correctly. To echo @liangfu , it would be...