What is default rounding mode in TVM?

Questions
#1

Hi,

I’m new at TVM and working with TVM to compile and run onnx models. I have some questions about the rounding mode for floating-point arithmetic in TVM: Can one change the rounding mode in TVM? Also, what is the default rounding mode in TVM?

Thanks in advance

#2

I run into this issue recently. ONNX and numpy.round is round to even, TVM doesn’t define round mode, it depends on intrinsic on each platform. CUDA/LLVM uses round to integer by default.

https://github.com/apache/tvm/pull/11446 solves the onnx issue. But TVM is still different from numpy

import numpy
print(numpy.round(0.5))

is 0.0

relay.round will get 1.

#3

tir::round is round to nearest int, which is the same as most backend intrinsics

#4

I happened to run into the same issue. IMO, if you really need roundeven, it is required to add the corresponding new intrinsic from backends like llvm.roundeven.

#5

If it’s cuda, it will be consistent with cuda’s default rounding mode