Add all-in-one driver for compile-and-execute #97
Conversation
I agree with you. The code was copied from legacy. Now I have updated the coding style to CamelCase. |
| } | ||
| compute = *expectCompute; | ||
| } | ||
| llvm::ArrayRef<uint64_t> foldBufferIds; |
There was a problem hiding this comment.
Better to unify the element dtypes here with the Globals in MLIR module. "__num_orig_num_args", "__fold_args" and "__compute_args" are i32, and "__fold_buffer_ids" is i64.
There was a problem hiding this comment.
__fold_buffer_ids is the UID of a buffer, so we use i64. __num_orig_num_args and others are i32 because we won't meet too many args in the same function if it is reasonable. We use i32 for __fold_args and __compute_args for performance because it save half of space and cache (!!) at each execution time.
unittests/CMakeLists.txt
Outdated
There was a problem hiding this comment.
Do we need another one tests folder?
There was a problem hiding this comment.
The folder tests is for MLIR-related FileCheck-based tests and unittests is for MLIR-related C++ gtests. The convention of tests and unittests is borrowed from MLIR upstream - they use these folders too.
I have several pending PRs to put the C++-based gtests in unittests folder.
…ijie/mainfunc_wrapper
|
Hi, all! I have removed the constant cache related code. This PR now only contains all-in-one The constant cache related code is moved to branch |
…ijie/mainfunc_wrapper
| compute(realargs.data()); | ||
| } | ||
|
|
||
| // directly call compute(). args should be an array of void*. args[i] should |
For usage see unittest.
Also add a wrapper for
mlir::ExecutionEnginefor further developing of constant cache.