Improve ROCm support. (#780)

- updates to support ROCm 5.2
- workarounds in tests where NV tools were used unconditionally
- implemented `get_num_blocks()` and `add_memfence()` for AMD GPU
- backported from history some atomics
- added bf16 support
- minor warnings cleanup
- added dockerfile to run on a ROCm enabled machine

Co-authored-by: B1tway <andrew.shukshov@gmail.com>
Co-authored-by: Andrey Shukshov <36711069+B1tway@users.noreply.github.com>

lib/driver/dispatch.cc

@@ -212,6 +212,7 @@ bool dispatch::hipinit(){
   return res;
 }
 
+#define HIP_DEFINE0(ret, fname) DEFINE0(hipinit, hip_, ret, fname)
 #define HIP_DEFINE1(ret, fname, t1) DEFINE1(hipinit, hip_, ret, fname, t1)
 #define HIP_DEFINE2(ret, fname, t1, t2) DEFINE2(hipinit, hip_, ret, fname, t1, t2)
 #define HIP_DEFINE3(ret, fname, t1, t2, t3) DEFINE3(hipinit, hip_, ret, fname, t1, t2, t3)
@@ -268,7 +269,8 @@ HIP_DEFINE2(hipError_t, hipEventCreate, hipEvent_t *, unsigned int)
 HIP_DEFINE3(hipError_t, hipEventElapsedTime, float *, hipEvent_t, hipEvent_t)
 HIP_DEFINE2(hipError_t, hipEventRecord, hipEvent_t, hipStream_t)
 HIP_DEFINE1(hipError_t, hipEventDestroy, hipEvent_t)
-
+// error handling
+HIP_DEFINE0(hipError_t, hipGetLastError)
 
 /* ------------------- *
  * COMMON

lib/driver/llvm.cc

@@ -268,7 +268,7 @@ std::string llir_to_amdgpu(llvm::Module* module, const std::string& _proc) {
   std::string triple = "amdgcn-amd-amdhsa";
   std::string layout = "";
   std::string features="+sramecc,-xnack";
-  std::string proc = "gfx908";
+  std::string proc = STRINGIFY(MI_GPU_ARCH);
   // name kernel
   auto in_time_t = std::chrono::system_clock::to_time_t(std::chrono::system_clock::now());
   std::stringstream cur_time;