Fixed a bug in the xGEMM routine related to the event incorrectly set

CHANGELOG

@@ -1,6 +1,7 @@
 
 Development version (next release)
 - Improved performance of large power-of-2 xGEMM kernels for AMD GPUs
+- Fixed a bug in the xGEMM routine related to the event incorrectly set
 
 Version 0.7.0
 - Added exports to be able to create a DLL on Windows (thanks to Marco Hutter)

src/routines/level3/xgemm.cc

@@ -184,12 +184,13 @@ StatusCode Xgemm<T>::DoGemm(const Layout layout,
 
       // Launches the kernel
       auto eventKernel = Event();
-      status = RunKernel(kernel, global, local, eventKernel.pointer(), eventWaitList);
+      auto eventPointer = (!c_no_temp) ? eventKernel.pointer() : event_;
+      status = RunKernel(kernel, global, local, eventPointer, eventWaitList);
       if (ErrorIn(status)) { return status; }
-      eventWaitList.push_back(eventKernel);
 
       // Runs the post-processing kernel if needed
       if (!c_no_temp) {
+        eventWaitList.push_back(eventKernel);
         status = PadCopyTransposeMatrix(event_, eventWaitList,
                                         m_ceiled, n_ceiled, m_ceiled, 0, c_temp,
                                         c_one, c_two, c_ld, c_offset, c_buffer,

test/correctness/tester.cc

@@ -334,7 +334,7 @@ bool TestSimilarity(const T val1, const T val2) {
 
   // Set the allowed error margin for floating-point comparisons
   constexpr auto kErrorMarginRelative = T(0.025);
-  constexpr auto kErrorMarginAbsolute = T(1.0e-4);
+  constexpr auto kErrorMarginAbsolute = T(1.0e-3);
 
   // Shortcut, handles infinities
   if (val1 == val2) {