Command line¶

hexir¶

hexir [-emit=<stage>] [-o <file>] [-opt] [-placement=<op>=<device>,...] [input.mlir]

With no input file, the compiler builds a small program in C++ (compiler/Support/Builder.cpp) and compiles that instead.

-emit¶

Value	Output
`mlir`	the `hexir` graph, as written
`mlir-tir`	each compute op as a `hextir` kernel
`hxb`	a loadable module (see `-o`)
`mlir-affine`	accepted, but no affine pass is wired up
`mlir-linalg`	after lowering to linalg on tensors
`mlir-hetero`	linalg ops carrying their `device` attribute
`mlir-gpu`	CUDA-placed ops as `gpu.launch`
`mlir-llvm`	the LLVM dialect
`llvm`	translated LLVM IR
`jit`	compile and run in this process

Other options¶

-o <file>: Output path for -emit=hxb. Defaults to out.hxb.
-opt: Run LLVM’s O3 pipeline before executing or printing.
-placement=<op>=<device>[,...]: Override where an operation runs. Keys are graph-level op names (hexir.linear, hexir.add, hexir.relu); devices are cpu, cuda, or gpu as an alias for cuda. An unsupported pair is rejected before any pass runs.

MLIR’s own pass-manager flags work too, including --print-ir-after-all and --mlir-print-ir-after-failure.

Examples¶

hexir -emit=jit                                       # built-in program, CPU
hexir -emit=mlir-tir -placement=hexir.linear=cuda     # matmul as a GPU kernel
hexir -emit=hxb -o model.hxb mine.mlir                # ship a file
hexir -emit=mlir-linalg --print-ir-after-all mine.mlir # watch every pass

hexir-run¶

hexir-run <module.hxb> [--device=cpu|cuda] [--entry=name] [--quiet]
hexir-run --selftest

--device=: Which device to run on. Defaults to cpu. cuda is not implemented yet. A module is refused if it was compiled for a different device.
--entry=: Entry point name. Defaults to main.
--quiet: Suppress the module description and print only program output.
--selftest: Check the HAL without needing a module.

Exit codes¶

0 success · 1 load or execution failed · 2 bad arguments