Fixes for the CUDA API

CMakeLists.txt

@@ -236,7 +236,6 @@ set(SOURCES
   src/routine.cpp
   src/routines/levelx/xinvert.cpp  # only source, don't include it as a test
   src/tuning/configurations.cpp
-  src/tuning/tuning_api.cpp
 )
 set(HEADERS  # such that they can be discovered by IDEs such as CLion and Visual Studio
   include/clblast_half.h
@@ -265,7 +264,7 @@ set(HEADERS  # such that they can be discovered by IDEs such as CLion and Visual
   src/tuning/routines/routine_tuner.hpp
 )
 if(OPENCL)
-  set(SOURCES ${SOURCES} src/clblast.cpp src/clblast_c.cpp)
+  set(SOURCES ${SOURCES} src/clblast.cpp src/clblast_c.cpp src/tuning/tuning_api.cpp)
   set(HEADERS ${HEADERS} include/clblast.h include/clblast_c.h src/clpp11.hpp)
   if(NETLIB)
     set(SOURCES ${SOURCES} src/clblast_netlib_c.cpp)

src/clblast_cuda.cpp

@@ -2593,7 +2593,8 @@ StatusCode GemmTempBufferSize(const Layout layout, const Transpose a_transpose,
     else {
       temp_buffer_size = Xgemm<T>::GetTempSize(layout, a_transpose, b_transpose, m, n, k,
                                                a_offset, a_ld, b_offset, b_ld, c_offset, c_ld,
-                                               db["MWG"], db["NWG"], db["KWG"]);
+                                               db["MWG"], db["NWG"], db["KWG"] * db["KREG"],
+                                               db["GEMMK"]);
     }
     temp_buffer_size *= sizeof(T); // translate from num-elements to bytes
     return StatusCode::kSuccess;