SWDEV-478624 - Use readback workaround to ensure kernel arg coherence

Use env var DEBUG_CLR_KERNARG_HDP_FLUSH_WA=1 to fall back to HDP flush
workaround. The default is 0

Change-Id: I7bdb9be61da60c30d15ac9991b7cd27351e1831c

rocclr/device/rocm/rocsettings.cpp

@@ -256,6 +256,8 @@ void Settings::setKernelArgImpl(const amd::Isa& isa, bool isXgmi, bool hasValidH
 
   auto kernelArgImpl = KernelArgImpl::HostKernelArgs;
 
+  hasValidHDPFlush &= DEBUG_CLR_KERNARG_HDP_FLUSH_WA;
+
   if (isXgmi) {
     // The XGMI-connected path does not require the manual memory ordering
     // workarounds that the PCIe connected path requires
@@ -281,6 +283,8 @@ void Settings::setKernelArgImpl(const amd::Isa& isa, bool isXgmi, bool hasValidH
   if (!flagIsDefault(HIP_FORCE_DEV_KERNARG)) {
     kernel_arg_impl_ = kernelArgImpl & (HIP_FORCE_DEV_KERNARG ? 0xF : 0x0);
   }
+
+  ClPrint(amd::LOG_INFO, amd::LOG_INIT, "Using dev kernel arg wa = %d", kernel_arg_impl_);
 }
 }  // namespace amd::roc
 

rocclr/utils/flags.hpp

@@ -258,7 +258,9 @@ release(bool, HIP_ALWAYS_USE_NEW_COMGR_UNBUNDLING_ACTION, false,              \
 release(bool, DEBUG_HIP_KERNARG_COPY_OPT, true,                               \
         "Enable/Disable multiple kern arg copies")                            \
 release(bool, DEBUG_CLR_USE_STDMUTEX_IN_AMD_MONITOR, false,                   \
-        "Use std::mutext in amd::monotor")                                    \
+        "Use std::mutex in amd::monotor")                                     \
+release(bool, DEBUG_CLR_KERNARG_HDP_FLUSH_WA, false,                          \
+        "Toggle kernel arg copy workaround")                                  \
 
 namespace amd {