You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
## NFC: Simplify some aspects of options management (OptionsContext)
- Adds a convenience 'OptionsContext::Option' class that simplifies how
options are declared.
- Closes a loophole where tuples of structs containing options can cause
crashes if they populate options in their constructor. Due to how the
external storage mechanism works, we can no longer use direct
`std::tuple`
of aggregate objects which invoke `addOption`. Instead, one must use
`unique_ptr` to wrap those types when used as elements of a
`std::tuple`.
- To help enforce this, we explicitly delete the move constructor of
`OptionsProvider`.
## [compiler|python] Update how cached pipelines/"Compiler Tasks" are
registered
This change updates how registration functions for "compilation tasks"
invoked. We now expose a C API method that can be invoked within the
Pybind11 module initializer. This decouples compiler task registration
from
pass or dialect registration.
This change also cleans up the C API function naming for pass/dialect
registration functions.
## [python] Add more robust CMake logic for fixing missing CAPI
dependency in core MLIR PyBind module
Adds CMake logic to ensure that the Core '_mlir' pybind extension has
the
correct CAPI dependencies declared until the upstream CMake declarations
can be
fixed.
## NFC: Remove unnecessary PyCapsule <-> CAPI casters in compiler and
runtime bindings
Removes unnecessary custom PyBind11 capsule -> C API object casters.
These cast functions are only required when it is desired to allow
PyBind11 to extract the C API object from the C++ python wrapper type
automatically.
## [tensorrt|compiler] Drop "layer metadata callback" utility from
TensorRT translation
This change removes the "layer metadata callback" feature from the
MLIR-to-TensorRT translation. It also removes the relevant APIs from the
MLIR-TensorRT compiler's C++ and Python APIs.
This capability was originally offered as a bridge for populating the
generated TensorRT ILayers with custom metadata. However, the mechanism
prevents caching of pass pipelines and therefore is too expensive to
use.
In the future, any metadata passed to TensorRT should be derived from
the MLIR operations' location information.
## NFC: update various uses of "Stablehlo" in class and function names
to have consistent capitalization
## NFC: Reorganize some directories
This change:
- Moves the top-level 'tools' to 'compiler/tools'
- Moves the top-level 'test' to 'compiler/test'
- Moves the 'mlir-tensorrt-tblgen' tool under 'tensorrt/tools'
since the 'tensorrt' project is supposed to be independent.
- Similarly move TensorRT-specific python definitions under
`tensorrt/python`.
## [executor]: Add a missing guard for builds without CUDA enabled.
Wrapping the makeCudaStringError function with MLIR_EXECUTOR_ENABLE_CUDA
fixes builds without CUDA enabled.
## [executor] Use Lua locals for block arguments
Previously, the Executor MLIR-to-Lua translator used Lua globals for
block arguments outside of the entry block since the values that
represent block arguments need to be passed between blocks. On the
other hand, the scope of Lua local variables are restricted to their
block. It is almost never a good idea to use Lua global variables in
our translation strategy, however -- for coroutine functions, a
translation that uses globals is obviously incorrect since all Lua
coroutines in a single Lua environment share the same set of globals.
This change declares all block arguments up front as locals in the
"entry block" and just sets them to `nil` initially. Since we don't
declare a block scope for the entry block, all the following Lua block
scopes will have these locals in scope. This allows us to retain the
use of locals for all block arguments.
GitOrigin-RevId: e9dd03c47eab6145e889ea8ff56fd1c71181f72a
0 commit comments