[OpenMP] Default to single SDMA engine

The actual concurrent use of multiple SDMA engines requires more work in
the plugin and how signals/streams are used to chain together data
copies. Thus, for the time being, we re-enable the original
implementation that let's ROCr decide which engine to use.

Change-Id: I4ff78284a5bbbe4a851b465052ff017fa6bc8a7c

Author: jplehr

openmp/libomptarget/plugins-nextgen/amdgpu/src/rtl.cpp

@@ -2486,7 +2486,7 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
         OMPX_StreamBusyWait("LIBOMPTARGET_AMDGPU_STREAM_BUSYWAIT", 2000000),
         OMPX_UseMultipleSdmaEngines(
             // setting default to true here appears to solve random sdma problem
-            "LIBOMPTARGET_AMDGPU_USE_MULTIPLE_SDMA_ENGINES", true),
+            "LIBOMPTARGET_AMDGPU_USE_MULTIPLE_SDMA_ENGINES", false),
         OMPX_ApuMaps("OMPX_APU_MAPS", false),
         OMPX_DisableUsmMaps("OMPX_DISABLE_USM_MAPS", false),
         OMPX_NoMapChecks("OMPX_DISABLE_MAPS", true),