Compiling AscendNPU IR

The bishengir-compile tool can compile AscendNPU IR into a binary file that can be executed on the AI processor.

Compile Options

Common compile options are described as follows.

Option

Required

Description

--help

No

Displays the help information.

-o <file>

No

Specifies the name and location of the output file.

--enable-triton-kernel-compile

No

Enables Triton-related compile optimization. The input IR must be the MLIR converted using the triton-shared tool.

--enable-hfusion-compile

No

Enables HFusion dialect-level compile optimization.

--enable-hivm-compile

Yes

Enables HIVM dialect-level compile optimization. It is enabled by default.

--enable-auto-multi-buffer

No

Enables automatic ping-pong multi-buffer optimization for the AI processor on-chip buffer.

Compile Command

During compile, the input.mlir file is compiled and optimized using the HFusion and HIVM dialects to generate the binary file outputFile.o.

The compile command is as follows:

1
bishengir-compile --enable-hfusion-compile=true --enable-hivm-compile=true input.mlir -o outputFile