The MLIR-AIE dialect builds a number of command utilities for compiling and transforming operations written in the MLIR-AIE dialect into other intermediate representations (IRs) as well as generating AIE program elfs and host executables to be run on the board. The two main output utilities that building the MLIR-AIE project gives you is aie-translate adn aie-opt which are used to transform MLIR-AIE dialect into other IRs. These utilities are then used by the convenience python utility aiecc.py to compile operations written in MLIR-AIE dialect into elf/host executables.
The basic way that we use aiecc.py to compile our MLIR-AIE written code (e.g. aie.mlir) into elf/host executable (core_*.elf, test.exe) is the following:
aiecc.py -j4 --sysroot=<platform sysroot> -host-target=aarch64-linux-gnu aie.mlir -I<runtime_lib> <runtime lib>/test_library.cpp ./test.cpp -o test.exe
This command compiles an MLIR-AIE dialect source file (e.g. aie.mlir) into AI Engine program elfs and generates the host API aie.mlir.prj/aie_inc.cpp which can be used to configure the AI Engine design. In addition, if provided with a host source file (e.g. test.cpp), aiecc.py will compile it to build a host executable (e.g. tutorial-1.exe). The AIE tile elfs are generated automatically for each AIE tile that needs to be programmed. Additionally, we usually pass a reference to<runtime lib>/test_library.cpp as it contains commonly used test functions.
You can see what arguments are supported with aiecc.py by calling it with the -h argument. Some useful arguments are listed below:
| Optional arguments | Description |
|---|---|
| --sysroot sysroot | sysroot for cross-compilation |
| -v | Trace commands as they are executed |
| --vectorize | Enable MLIR vectorization |
| --xbridge | Link using xbridge (default) |
| --xchesscc | Compile using xchesscc (default) |
| --compile | Enable compiling of AIE code (default) |
| --no-compile | Disable compiling of AIE code |
| --host-target HOST_TARGET | Target architecture of the host program (e.g. vck190 uses aarch64-linux-gnu) |
| --compile-host | Enable compiling of the host program (default) |
| --no-compile-host | Disable compiling of the host program |
| --link | Enable linking of AIE code (default) |
| --no-link | Disable linking of AIE code |
| -j NTHREADS | Compile with max n-threads in the machine (default is 1). An argument of zero corresponds to the maximum number of threads on the machine. |
| --profile | Profile commands to find the most expensive executions. |
| --unified | Compile all cores together in a single process (default) |
| --no-unified | Compile cores independently in separate processes |
| -n | Disable actually executing any commands. |
This is the primary utility does transforms/ optimizes source code from one representation to another within the MLIR-AIE defined dialect. This is primarily controlled via command line options and adding multiple options allows the utility to perform multiple transformations/ optimizations. A full description of the options can be found here, but some example options transforms from a logical description to a more physical one such as in between flow and switchboxes in tutorial-4
This utility is more geared toward translating a description into another format altogether. For example, generating accessory files like .bcf and .ldscript would be done with this utility. Again, the full description of options can be found here.
Rather than walk through the source code for aiecc.py, we will describe some of the main calls that aiecc.py makes to aie-translate and aie-opt and show the arguments used to provide some idea about what they do.
The main flow of aiecc.py is:
-
First set of optimizations
-
Translate file to count AI Engine cores in design
-
Second set of optimization
-
Translate in LLVM-IR
-
Compile invididual cores (e.g. xchesscc_wrapper)
-
Process ARM cross-compilation of host code
-
for loop over cores
- 1st set of core optimizations
- 2nd set of core optimizations
- Translate to generate .bcf or .ldscript
- Compile core (e.g. xchesscc_wrapper)
aie-opt
--lower-affine
--aie-canonicalize-device
--aie-assign-lock-ids
--aie-register-objectFifos
--aie-objectFifo-stateful-transform
--aie-lower-broadcast-packet
--aie-lower-multicast
--aie-assign-buffer-addresses
--convert-scf-to-cf
aie.mlir -o input_with_addresses.mlir
aie-translate --aie-generate-corelist input_with_addresses.mlir
aie-opt
--aie-localize-locks
--aie-standard-lowering
--aie-normalize-address-spaces
--canonicalize
--cse
--convert-vector-to-llvm
--expand-strided-metadata
--lower-affine
--convert-arith-to-llvm
--convert-memref-to-llvm
--convert-func-to-llvm=use-bare-ptr-memref-call-conv
--convert-cf-to-llvm
--canonicalize --cse
input_with_addresses.mlir -o input_opt_with_addresses.mlir
aie-translate
--opaque-pointers=0
--mlir-to-llvmir
input_opt_with_addreses.mlir -o input.ll
xchesscc_wrapper -c -d -f +P 4 file_llvmir_hacked -o input.o
This section does cross-compilation which includes calling clang with a specific target. However, aie-opt and aie-translate is called here to generate the aie.mlir.prj/aie_inc.cpp file.
aie-opt
--aie-create-pathfinder-flows
--aie-lower-broadcast-packet
--aie-lower-multicast
input_with_addresses.mlir -o input_physical.mlir
aie-translate --aie-generate-xaie --xaie-target=v2 input_physical.mlir -o aie_inc.cpp
aie-opt
--aie-localize-locks
--aie-standard-lowering=tilecol=COL tilerow=ROW % core[0:2]
input_with_addresses.mlir -o -o core_*.mlir
aie-opt
-aie-normalize-address-spaces
--canonicalize
--cse
--convert-vector-to-llvm
--expand-strided-metadata
--lower-affine
--convert-arith-to-llvm
--convert-memref-to-llvm
--convert-func-to-llvm=use-bare-ptr-memref-call-conv
--convert-cf-to-llvm
--canonicalize
--cse
core*.mlir -o opt.mlir
aie-translate input_with_addresses.mlir --aie-generate-bcf --tilecol=COL --tilerow=ROW -o core*.bcf
or
aie-translate input_with_addresses.mlir --aie-generate-ldscript --tilecol=COL --tilerow=ROW -o core*.ld.script
This section makes calls to xchesscc_wrapper to compile the final core*.elf.
xchesscc_wrapper -d -f core*.o link_with_obj +l core*bcf -o core*elf