Minor optimization for the direct GEMM kernel: don't ceil m and n unnecessarily high

1 files changed, 4 insertions, 4 deletions

diff --git a/src/routines/level3/xgemm.cpp b/src/routines/level3/xgemm.cpp
index 30e5999c..f4611aba 100644
--- a/src/routines/level3/xgemm.cpp
+++ b/src/routines/level3/xgemm.cpp
@@ -280,11 +280,11 @@ void Xgemm<T>::GemmDirect(const size_t m, const size_t n, const size_t k,
   kernel.SetArgument(16, static_cast<int>(b_conjugate));
 
   // Computes the global and local thread sizes
-  const auto m_ceiled = Ceil(m, db_["WGD"]);
-  const auto n_ceiled = Ceil(n, db_["WGD"]);
+  //const auto m_ceiled = Ceil(m, db_["WGD"]);
+  //const auto n_ceiled = Ceil(n, db_["WGD"]);
   const auto global = std::vector<size_t>{
-    (m_ceiled * db_["MDIMCD"]) / db_["WGD"],
-    (n_ceiled * db_["NDIMCD"]) / db_["WGD"]
+    CeilDiv(m * db_["MDIMCD"], db_["WGD"]),
+    CeilDiv(n * db_["NDIMCD"], db_["WGD"])
   };
   const auto local = std::vector<size_t>{db_["MDIMCD"], db_["NDIMCD"]};