From d8705a2b781342c0cfa93e224aefa3404367262b Mon Sep 17 00:00:00 2001
From: Nikita Grigorian <nikita.grigorian@intel.com>
Date: Wed, 29 May 2024 13:20:21 -0700
Subject: [PATCH] Use `sycl::rint` rather than `std::rint` to resolve failed
 tests on Nvidia hardware

When compiled for CUDA, `std::rint` would incorrectly round values halfway between two integers toward 0, rather than to the nearest even number as required per array API.

`sycl::rint` avoids such issues by not relying on the current rounding mode
---
 .../libtensor/include/kernels/elementwise_functions/round.hpp   | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/dpctl/tensor/libtensor/include/kernels/elementwise_functions/round.hpp b/dpctl/tensor/libtensor/include/kernels/elementwise_functions/round.hpp
index cdc16180f8..f129799ff8 100644
--- a/dpctl/tensor/libtensor/include/kernels/elementwise_functions/round.hpp
+++ b/dpctl/tensor/libtensor/include/kernels/elementwise_functions/round.hpp
@@ -81,7 +81,7 @@ template <typename argT, typename resT> struct RoundFunctor
 private:
     template <typename T> T round_func(const T &input) const
     {
-        return std::rint(input);
+        return sycl::rint(input);
     }
 };