Fixed a bug in omniprobe when rocminfo is a symlink - the was the case in the AMD dev cloud and it broke omniprobe. Also modifying to use the latest kerneldb

Author: Keith Lowery

external/kerneldb

@@ -222,6 +222,8 @@ def get_default_cache_location():
 
 def get_rocm_path():
     tmp = shutil.which("rocminfo")
+    if tmp:
+        tmp = os.path.realpath(tmp)
     nodes = tmp.split("/")
     if len(nodes) > 2:
         return "/".join(nodes[:len(nodes) - 2])

omniprobe/omniprobe

@@ -2,5 +2,5 @@
 echo $PATH
 #AMDCGN_INSTRUMENTATION_FUNCTIONS_FILE=/work1/amd/klowery/instrument-amdgpu-kernels/dh_comms/build/lib/hip/dh_comms_dev.gfx90a.bc hipcc -fpass-plugin=/work1/amd/klowery/instrument-amdgpu-kernels/build/hip-static/libAMDGCNMemTrace.so -I/opt/rocm/include -I/opt/rocm/include/roctracer  -O3 quicktest.cc -L/opt/rocm/roctracer/lib  -v -o quicktest
 #hipcc -L/home1/klowery/.local/lib -ldh_comms -Rpass-analysis=kernel-resource-usage -fpass-plugin=/work1/amd/klowery/logduration/omniprobe/lib/libAMDGCNMemTraceHIP.so -I/home1/klowery/.local/include/dh_comms -fgpu-rdc --offload-arch=gfx90a  -std=c++17 -ggdb -O3  quicktest.cc  -v  -o quicktest 
-hipcc -fpass-plugin=/work1/amd/klowery/logduration/build/external/instrument-amdgpu-kernels-rocm/build/lib/libAMDGCNSubmitBBStart-rocm.so -fgpu-rdc --offload-arch=gfx90a  -std=c++17 -ggdb  quicktest.cc -o quicktest
+hipcc -fpass-plugin=/work/klowery/omniprobe/build/external/instrument-amdgpu-kernels-rocm/build/lib/libAMDGCNSubmitBBStart-rocm.so -fgpu-rdc --offload-arch=gfx90a  -std=c++17 -ggdb  quicktest.cc -o quicktest
 #hipcc  -I/opt/rocm/include -I/opt/rocm/include/roctracer -g quicktest.cc -L/opt/rocm/roctracer/lib  -v -o quicktest