Switching CI back to full_asm build#750
Conversation
|
Let us see how long this takes. Will this increase the test suite size? |
|
I assume we're going to get rid of asm_ci? Also @saadrahim it will increase the compile time as this adds more Tensile kernels than asm_ci and will increase the test suite size as well. We should add a gfx908 job for rocBLAS as well |
zaliu
left a comment
zaliu
left a comment
There was a problem hiding this comment.
As I mentioned previously, there is absolutely no reason to go for asm_full build on gfx900 and gfx906 CI machines. We want to do asm_full build only on gfx908 CI machines.
The memory footprint of a asm_full build is about 44GB. If a CI machine has 8 GPUs and 256GB physical memory, on average a fully loaded 8 CI runs, each gets only about 32GB, which might cause occasional spurious build failures.
Can @saadrahim or @amdkila please make the changes to build on the appropriate architectures? |
Memory should be ok as we have 64GB of RAM per GPU on our gfx908 node. One issue we have is related to packaging and the number of GPUs- in order to create rocBLAS debian and rpm packages with asm_full, we'd need 2 gfx908 jobs. We don't have the resources to do that right now. |
[rocblas] - matches hipblas OpenMP fix for gcc issue where clang headers used (#750) Additional cleanup: Remove unnecessary options added to COMMON_LINK_LIBS. If the openmp config is not found then OpenMP::OpenMP_CXX will be the fallback for both gcc and clang.
5 participants
This change ensures all kernels are being built and tested
Please inform if there are memory restrictions that need to be considered for a full rocBLAS build.