Open ib00 opened 6 months ago
Thank you for the detailed explanation. We will address them.
Regarding the longer run times, with --bitcode --precompile
options only the user code should be compiled and linked with the HIPRT trace kernels in runtime. Without --bitcode --precompile
, all kernels are compiled for the first time and cached for the next runs.
Compilation (on Linux and for CUDA) is still not without problems. I am following instructions in README.
This is however harmless as compilation continues.
Another harmless error:
If you use "--precompile" flag the above with hipcc crashes kernel compilation:
Traceback (most recent call last): File "/tmp/HIPRT/scripts/bitcodes/compile.py", line 164, in
compileAmd()
File "/tmp/HIPRT/scripts/bitcodes/compile.py", line 83, in compileAmd
result = subprocess.check_output(cmd, shell=True)
compiling hiprt02003_nv_lib.fatbin... ../../hiprt/hiprt_common.h(127): error: identifier "uint8_t" is undefined static_assert( ( sizeof( uint8_t ) == 1 ), "" ); ^
This can be fixed by including:
include
in "hiprt/hiprt_common.h"
../../hiprt/impl/PlocBuilderKernels.h(142): warning #186-D: pointless comparison of unsigned integer with zero if ( index >= 0 && index < numberOfClusters ) ^
Compilation SUCCEEDED. compiling oro_compiled_kernels.fatbin... Compilation SUCCEEDED. export built programs ... cp: target '../../dist/bin/Debug/': No such file or directory cp: target '../../dist/bin/Release/': No such file or directory
../dist/bin/Release/unittest64 -p
tests run, but it takes a long time for tests to start. Are the kernels being recompiled?