WIP: Make Fortran tests fail with a non-zero exit code #4097
Conversation
|
@anagainaru Can you take a look at the Fortran MemorySpace failure? |
|
@eisenhauer When we change the default memory space to be Default regardless of GPU being enabled or not, I forgot to change the Fortran bindings to update this. I think it should pass now. |
|
Hmm, maybe it's more complicated, even though I don't understand how the changes we made could trigger this new error. I'll try to reproduce it locally (I don't see it on my laptop). |
anagainaru
force-pushed
the
FortranTests
branch
from
cbbd11d to
05caf52
Compare
|
Turns out the default space needed to be updated in two location (core and bindings). It should pass the tests now, let me know if you need me to clean the history in case you don't want to squash merge. |
I'll squash... Thank you! |
|
Odd. The dataman tests are failing now. I have to suspect that the change in the initialization of m_MemSpace has impacted this, but it's going to take some debugging to figure out how. I'm not going to tackle that now. |
… "Host" when GPU support is not enabled
eisenhauer
force-pushed
the
FortranTests
branch
from
0221798 to
a69c805
Compare
|
So, the deal is this: Prior to the change in VariableBase, m_MemSpace had value MemorySpace::Host if GPU support was not enabled. After the change, it was left at MemorySpace::Detect. The DataMan serializer was looking for Host or GPU values, but couldn't handle the Detect value. I changed DataMan to handle Detect as if it was Host. One could also have presumably replaced the Detect with Host at some point in the output process, but I'm not entirely clear where would be best. So I've left it at this for the moment. |
|
@eisenhauer by the time it reaches dataman there shouldn't be the option of Detect, let me check this before we merge. I did not catch this when I change the default since I am apparently checking this only in the fortran bindings. I might add a test for the default and set memory space when we test the BP engine. |
There was a problem hiding this comment.
The problem(s) should be fixed now @eisenhauer. Dataman had a bunch of issues due to changing the default memory space. Should we make a separate PR with the fixes and leave this one clean just with the Fortran CI fixes? Otherwise if we keep just one, take a look at the changes I made (I change the GPU part of your print function) and I think this can be merged (I'll approve so you have this choice).
|
I think I'll go ahead and merge to get the Fortran test fixes in there. We can sort DataMan as we need to... |
|
I'll add tests for it when I am back, I think it should work ok now. I tested a bunch of things on Frontier, but I did not have time to put them all in a test. |
* master: - Only add campaign store to file name if that is not absolute path - list command supports second argument as path flake8 fixes Update campaign manager script to handle config file, time in nanosecond format, and avoiding conflict when updating database dill 2024-03-12 (ebc98c4d) (ornladios#4091) Don't run derived test in MPI mode, it's not written for that (ornladios#4104) Fix static blosc2 build (ornladios#4093) ci: add ccache job summary (ornladios#4101) Fix typo in fortran.rst (ornladios#4102) WIP: Make Fortran tests fail with a non-zero exit code (ornladios#4097) Bison 3.8 Parser (ornladios#4062)
* Make Fortran tests fail with a non-zero exit code Co-authored-by: anagainaru <[email protected]>
* origin/adios-xrootd: (164 commits) Fixes for FreeBSD, including upstream (ornladios#4138) Bump version to v2.10.0 Setting the derived variable support OFF by default Add the CURL function to derived variables (ornladios#4114) Add -f file1 [file2...] option to process adios files from a list instead of a campaign recording DataPlane Configuration changes (ornladios#4121) update doc Add attribute support to campaign Add a random string to each database name to avoid collision when running multiple applications in the same directory at the same time. Fixes issues with CI that runs ctest in parallel Initialize ADIOS::m_UserOption before tentaively calling ProcessUserConfig() - set ACA version to 0.1 - remove debug prints - add doc on Campaign Management in Advanced section - change static struct of UserOptions to class member of ADIOS class to make it work with gcc 4.8 used size_t not int for map indexing to avoid type conversion fix remote server test, use the new binary name clang-format Use the name of the campaign in the cache path to avoid name collision rename remote_server to adios2_remote_server bug fix: the order of entries in bpdataset table is undefined but the campaign data reader relied on calculating the index as if it was sorted by the insertion order. Use a map instead to store the rowid and use that as index for the bpfile elements. Use yaml parser in campaign manager python script change a long variable to int64_t to avoid size confusion on windows do not include unistd.h Fix compiler error Remove extra file not needed Add support for user options in ~/.config/adios2/adios2.yaml Currently supported options: cmake: add sqlite3 and zlib dep in adios2 cmake pkg Different names for MPI and Serial tests (ornladios#4118) EVpath upstream to make NO_RDMA more robust (ornladios#4116) Warnings (ornladios#4113) Don't use assert() in tests (ornladios#4108) - Only add campaign store to file name if that is not absolute path - list command supports second argument as path flake8 fixes Update campaign manager script to handle config file, time in nanosecond format, and avoiding conflict when updating database dill 2024-03-12 (ebc98c4d) (ornladios#4091) Don't run derived test in MPI mode, it's not written for that (ornladios#4104) Fix static blosc2 build (ornladios#4093) ci: add ccache job summary (ornladios#4101) Fix typo in fortran.rst (ornladios#4102) WIP: Make Fortran tests fail with a non-zero exit code (ornladios#4097) Bison 3.8 Parser (ornladios#4062) Do not create adios-campaign/ unless there is something to record Add setup for Aurora (load adios2 as e4s package) Completely hide derived variables in C API if not enabled. Print warning inside Fortran F2C function. adios2_define_derived_variable C/Fortran API. C is compiled conditionally, like the C++ API. The Fortran function is always available, it will print a WARNING if derived variable is not supported. Added Fortran test for magnitude(). Fix links to tutorial materials (ornladios#4086) BlockIndex.Evaluate() made branch based on BP5 vs BP4. To support CampaignReader engine, decision is made based on whether MinBlocksInfo is supported by engine. Update documentation for 2.10 changes to the GPU-backend (ornladios#4083) Add test for single string attribute vs string array attribute with a single element - Python: fix for scalar reading. If a global value has 1 step (i.e. always in streaming), read returns a 0-dim numpy array (single value). If the variable has multiple steps (only in ReadRandomAccess mode), read returns a 1-dim numpy array even if the step selection is a single step. This way, read of a certain variable always results in the same type of array no matter the number of steps selected. - Python: fix for string attributes: return a string, not a list of one element which is a string, to be consistent with string global values and with other APIs. format more format Python: add the same treatment to attributes as to variables before: return scalars (0-dim ndarray) for single value attributes. Raise an exception if remote open fails (ornladios#4069) Fortran bindings for memory space related functions (ornladios#4077) consolidate (ornladios#4078) Making the Detect memory space available regardless of the backend used Testing code for the C bindings with memory space API Adding c bindings for getting the shape of a variable based on a memory space Adding c bindings for setting and getting the memory space Small typo fixes - Restructure python API doc in separate main topic, add working example to it. - What's new for 2.10 - Usage on DOE machines Fix Reord to use MinBlocksInfo where appropriate (ornladios#4071) Using the correct flag to detect the CUDA backend in ZFP clang-format fix fixed warning added support to read back from H5T_STRING VARIABLES it turns out strings written out through h5py are all variable strings format fixes to still be able to build with gcc 4.8.2. Needed for OLCF DTN nodes Add minmax and shape functions to CampaignReader, so that per-block info is complete when listing campaign archives do not flush io and adios in read mode. BP5 reader does not like it. Campaign engine is recognized by file extension. WIP. Changed the name of the campaign config file. Added reading of configuration parametyers from ./config/adios2 Use GetEstimatedSize in encryption operator plugin Allow plugin operators to take advantage of the estimated size API Setting the derived variable support OFF by default Add the CURL function to derived variables (ornladios#4114) Add -f file1 [file2...] option to process adios files from a list instead of a campaign recording DataPlane Configuration changes (ornladios#4121) update doc Add attribute support to campaign Add a random string to each database name to avoid collision when running multiple applications in the same directory at the same time. Fixes issues with CI that runs ctest in parallel Initialize ADIOS::m_UserOption before tentaively calling ProcessUserConfig() - set ACA version to 0.1 - remove debug prints - add doc on Campaign Management in Advanced section - change static struct of UserOptions to class member of ADIOS class to make it work with gcc 4.8 used size_t not int for map indexing to avoid type conversion fix remote server test, use the new binary name clang-format Use the name of the campaign in the cache path to avoid name collision rename remote_server to adios2_remote_server bug fix: the order of entries in bpdataset table is undefined but the campaign data reader relied on calculating the index as if it was sorted by the insertion order. Use a map instead to store the rowid and use that as index for the bpfile elements. Use yaml parser in campaign manager python script change a long variable to int64_t to avoid size confusion on windows do not include unistd.h Fix compiler error Remove extra file not needed Add support for user options in ~/.config/adios2/adios2.yaml Currently supported options: cmake: add sqlite3 and zlib dep in adios2 cmake pkg Different names for MPI and Serial tests (ornladios#4118) EVpath upstream to make NO_RDMA more robust (ornladios#4116) Warnings (ornladios#4113) Don't use assert() in tests (ornladios#4108) - Only add campaign store to file name if that is not absolute path - list command supports second argument as path flake8 fixes Update campaign manager script to handle config file, time in nanosecond format, and avoiding conflict when updating database ...
No description provided.