Big count support for datatypes #13460
Conversation
There was a problem hiding this comment.
Pull Request Overview
This PR implements bigcount support for MPI datatypes by introducing type-punned count and displacement arrays that can store either 32-bit integers or 64-bit size_t/ptrdiff_t values. The implementation uses the least significant bit of pointers as a flag to distinguish between the two representations, allowing the datatype framework to handle both standard MPI and bigcount MPI_Count variants transparently.
Key changes:
- Introduced
opal_count_array_tandopal_disp_array_twrapper types with helper functions for type-punned storage - Updated datatype creation functions to accept these wrapper types instead of raw int/ptrdiff_t arrays
- Modified datatype argument storage to track both 32-bit and 64-bit count arrays separately
- Removed conditional bigcount compilation blocks and temporary array allocations throughout the codebase
Reviewed Changes
Copilot reviewed 54 out of 54 changed files in this pull request and generated 9 comments.
Show a summary per file
| File | Description |
|---|---|
| opal/util/count_disp_array.h | New header defining type-punned array wrapper types and manipulation functions |
| opal/util/Makefile.am | Added new header to build system |
| ompi/util/count_disp_array.h | Refactored to delegate to opal implementation |
| ompi/datatype/ompi_datatype.h | Updated function signatures to use count_array_t types and size_t for counts |
| ompi/datatype/ompi_datatype_args.c | Rewrote argument packing/unpacking to handle both int and size_t arrays |
| ompi/datatype/ompi_datatype_create*.c | Updated to use count_array_t wrappers and size_t |
| ompi/mpi/c/type_*.c.in | Removed bigcount overflow checks and temp arrays, use wrapper types |
| ompi/mca/*/**.c | Updated datatype creation calls to use wrapper macros |
| opal/datatype/opal_datatype_*.c | Changed parameter types from int32_t to ssize_t/size_t |
Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.
|
okay this is pretty big! i will probably not be able to get to most of this till Monday. |
devreal
force-pushed
the
bigcount-datatypes
branch
from
306687d to
5f3af4e
Compare
|
CI currently fails for NVIDIA because hcoll was not adapted but I figure I'll just wait for #13467 to land. Also, I plan to port the datatype tests for bigcount, ideally without having to duplicate the whole code. |
type_get_envelop.c.in? |
It didn't seem like it was being compiled (it had broken code and no compiled complained). The bigcount version has an additional parameter (number of counts) so it cannot just be generated. |
|
ah yes. i forgot in the ABI work i enhanced the python to handle the case of no-big-count vs big-count and could use the template. so no worries for now. |
Utilize the count and disp arrays for type-punning of int and MPI_Count arguments to datatype creation functions. Pack integer or MPI_Count depending on what is needed when packing a datatype. Adjust places where bigcount support for datatpyes was missing. Signed-off-by: Joseph Schuchart <[email protected]>
The signatures for MPI_Type_get_envelope and MPI_Type_get_envelope_c are different so we cannot generate them. This file seems to be unused. Signed-off-by: Joseph Schuchart <[email protected]>
Signed-off-by: Joseph Schuchart <[email protected]>
Signed-off-by: Joseph Schuchart <[email protected]>
Signed-off-by: Joseph Schuchart <[email protected]>
Signed-off-by: Joseph Schuchart <[email protected]>
devreal
force-pushed
the
bigcount-datatypes
branch
from
f03693e to
9090b44
Compare
For struct, hindex, and hindex_block MPI_Aint arguments turn into MPI_Count arguments for bigcount. This needs special handling. Signed-off-by: Joseph Schuchart <[email protected]>
This was a fragment of a previous change that got only partially rolled back. Signed-off-by: Joseph Schuchart <[email protected]>
That codepath wasn't compiled so it needed some adjustment. Signed-off-by: Joseph Schuchart <[email protected]>
Signed-off-by: Joseph Schuchart <[email protected]>
3 participants
Utilize the count and disp arrays for type-punning of int and MPI_Count arguments to datatype creation functions. Store and pack integer or size_t depending on what is needed. Adjust places where bigcount support for datatypes was missing.
This is work in progress and not yet properly tested, other than compiling through it. I wanted to put it out to get some eyes on it to see whether people agree with the design and that it's the right direction, before putting more effort into testing.