Backout fp16 max/min HIP API change (#1771)

back out changes for rocm-5.5

Backout fp16 max/min HIP API change (#1771)
back out changes for rocm-5.5
42772fd6 · Umang Yadav · GitHub · 873d0473 · 42772fd6
Unverified Commit 42772fd6 authored May 23, 2023 by Umang Yadav Committed by GitHub May 23, 2023
Hide whitespace changes
Inline Side-by-side

Showing with 4 additions and 2 deletions

src/targets/gpu/kernels/include/migraphx/kernels/math.hpp src/targets/gpu/kernels/include/migraphx/kernels/math.hpp +4 -2

No files found.
--- a/src/targets/gpu/kernels/include/migraphx/kernels/math.hpp
+++ b/src/targets/gpu/kernels/include/migraphx/kernels/math.hpp
@@ -188,8 +188,10 @@ MIGRAPHX_DEVICE_MATH_BINARY_FOR(float, max, ::max)
 MIGRAPHX_DEVICE_MATH_BINARY_FOR(float, min, ::min)
 MIGRAPHX_DEVICE_MATH_BINARY_FOR(double, max, ::max)
 MIGRAPHX_DEVICE_MATH_BINARY_FOR(double, min, ::min)
-MIGRAPHX_DEVICE_MATH_BINARY_FOR(migraphx::half, max, ::__hmax)
-MIGRAPHX_DEVICE_MATH_BINARY_FOR(migraphx::half, min, ::__hmin)
+// Add overloads for half that calls the float version, this should use "hmax" and "hmin" once
+// perf CI docker is upgraded to rocm-5.5
+MIGRAPHX_DEVICE_MATH_BINARY_FOR(migraphx::half, max, ::fmaxf)
+MIGRAPHX_DEVICE_MATH_BINARY_FOR(migraphx::half, min, ::fminf)

 template <class T, MIGRAPHX_REQUIRES(not is_any_vec<T>())>
 constexpr auto max(const T& a, const T& b)