Expose llvm.nearbyint intrinsic. This is a faster alternate to rounding.

[kimishpatel]

In the quantized gemm implementation we are working on, we need to quantize input data. During this step we first apply scale and zero point to the input data. Then we do rounding and casting to int8.

tvm::round gets lowered by llvm into roundf function call which make the op slower. I instead exposed llvm.nearbyint via tvm and was able to recover the lost performance.

So this PR is just upstreaming that change.

[kimishpatel]

@tqchen, sorry have to tag you again on this PR as it seems that you may be the right person. Please feel free to suggest anyone else, but this is a quite small PR.

[tqchen]

Please add a python wrapper, a reference to the docs in the https://github.com/dmlc/tvm/tree/master/docs/api/python, and a testcase

[tqchen]

As per https://docs.tvm.ai/contribute/code_review.html#ensure-test-coverage

Thanks @kimishpatel

[kimishpatel]

Thanks @tqchen, I have added test case and python binding. Thanks for your feedback.

[anijain2305]

@kimishpatel Thanks for the PR. You can also look at qnn.op.quantize. This is a wrapper that internally lowers to the sequence of relay ops you mentioned. This PR will benefit that wrapper as well.

[kimishpatel]

@anijain2305, thanks for the pointer. Wasn't quite aware of that.

[anijain2305]

@anijain2305, thanks for the pointer. Wasn't quite aware of that.

If you are looking at running pre quantized models, you might want to have a look at QNN dialect. We have added a number of QNN ops in there that deal with scale and zero points.

[kimishpatel]

@tqchen, sorry to bug again :). Seems like all checks have passed, so just nudging for the merge. Thanks a bunch.

[tqchen]

Thanks @anijain2305 @kimishpatel