DDI extension support #285
Conversation
nrspruit
commented
Feb 22, 2025
•
nrspruit
commented
- Support in the L0 Loader for https://oneapi-src.github.io/level-zero-spec/level-zero/latest/core/EXT_Driver_DDIHandles.html#ze-extension-driver-ddi-handles
- To improve the speed for each call of L0 apis into the correct driver, the new support enables for the driver to allocate a header that stores that driver's ddi tables such that the Loader does not need to perform translation of handles to/from ze_object_t.
- Given a driver supports the new extension, the loader will no longer create the ze_object_t and instead use the packed ddi tables in each handle_t.
- The code has been refactored such that the legacy path will continue to work for the driver that has not converted over without interfering with a driver which now uses the "fast" path.
- This new code path also removes the need for handle translation by the user if the drivers all support the new extension.
nrspruit
force-pushed
the
ddi_extension_support
branch
4 times, most recently
from
c2f7ff5 to
b1a43bf
Compare
| pDriver = drivers[driver]; | ||
| pDevice = findDevice( pDriver, type ); | ||
| if( pDevice ) |
There was a problem hiding this comment.
So clearly, we're going to find all devices now, and not break out at the first one.... However down at line 127 does this logic still hold, if we skip to the end of the driver list via this loop? Maybe instead we insert a temp variable and check pDevice for valid value, increment a counter, and below )(line 128 here, or 127 in new version) replace if (!pDevice) with a if( found_count == 0) ?
| } | ||
| auto ${th.make_pfn_name(n, tags, obj)} = dditable->${th.get_table_name(n, tags, obj)}->${th.make_pfn_name(n, tags, obj)}; | ||
| if( nullptr == ${th.make_pfn_name(n, tags, obj)} ) | ||
| return ${X}_RESULT_ERROR_UNINITIALIZED; | ||
| %endif |
There was a problem hiding this comment.
Do we have to many %endif here
| @@ -97,6 +97,7 @@ zelLoaderTranslateHandleInternal( | |||
| } | |||
|
|
|||
| bool validHandle = false; | |||
| *handleOut = handleIn; | |||
There was a problem hiding this comment.
don't we need nullptr check before dereferencing the pointer?
There was a problem hiding this comment.
I can add that at the start of this api.
scripts/templates/helper.py
Outdated
| return "ZE_API_VERSION_1_11" | ||
| if version == "1.12": | ||
| return "ZE_API_VERSION_1_12" | ||
| return "ZE_API_VERSION_1_0" |
There was a problem hiding this comment.
I believe default path should break somehow, instead of just silently passing 1.0. If new api version is added and not handled here it shouldn't compile
source/loader/ze_loader_api.cpp
Outdated
| @@ -96,48 +96,91 @@ zelLoaderTranslateHandleInternal( | |||
| return ZE_RESULT_SUCCESS; | |||
| } | |||
|
|
|||
| bool validHandle = false; | |||
There was a problem hiding this comment.
what about handling invalid handle after the entire switch?
There was a problem hiding this comment.
we check for a valid handle for each handle type. This variable is set here such that we are not creating a new variable for each switch case.
The valid check is for if the handle was seen by the loader, I can change the name of the variable to be more self explanatory.
There was a problem hiding this comment.
I am removing the variable because it is not needed and I can simplify the check.
Signed-off-by: Neil R. Spruit <[email protected]>
Signed-off-by: Neil R. Spruit <[email protected]>
Signed-off-by: Neil R. Spruit <[email protected]>
Signed-off-by: Neil R. Spruit <[email protected]>
Signed-off-by: Neil R. Spruit <[email protected]>
Signed-off-by: Neil R. Spruit <[email protected]>
Signed-off-by: Neil R. Spruit <[email protected]>
Signed-off-by: Neil R. Spruit <[email protected]>
Signed-off-by: Neil R. Spruit <[email protected]>
Signed-off-by: Neil R. Spruit <[email protected]>
Signed-off-by: Neil R. Spruit <[email protected]>
Signed-off-by: Neil R. Spruit <[email protected]>
- Fixed get_version code gen script to fail if the api version for the api is not set or is invalid. - fixed the zelLoaderTranslateHandle to fail given nullptrs and remove the extra vairable and simplify the hasInstance handle check. Signed-off-by: Neil R. Spruit <[email protected]>
nrspruit
force-pushed
the
ddi_extension_support
branch
from
bde6a4a to
de519a7
Compare
