CodeDiffs

Compare code and display the difference in the terminal side-by-side, based on DeepDiffs.jl. Supports syntax highlighting.

The @code_diff macro is the main entry point. If possible, the code type will be detected automatically, otherwise add e.g. type=:native for native assembly comparison:

Syntax highlighting for Julia AST is also supported:

The @code_for macro is a convinence macro which will give you only one side of @code_diff's output, therefore it behaves like all @code_native/@code_** macros but with seamless support for GPU and additional cleanup functionalities.

Supported languages

• :native native CPU assembly (output of @code_native)
• :llvm native LLVM IR (output of @code_llvm)
• :typed Typed Julia IR (output of @code_typed)
• :ast Julia AST (any Expr, relies on Revise.jl)

Their equivalents for each GPU package is also supported:

Code type	CUDA.jl	AMDGPU.jl	oneAPI.jl	Metal.jl
Julia Typed	:cuda_typed	:rocm_typed	:one_typed	:mtl_typed
LLVM IR	:cuda_llvm	:rocm_llvm	:one_llvm	:mtl_llvm
Native	:cuda_native/:ptx	:rocm_native/:gcn	:one_native/:spirv	:mtl_native/:agx

Additionally, SASS assembly from CUDA is also supported with :sass.

Each output should match the output @device_code_typed, @device_code_llvm and @device_code_native defined by their respective GPU package, as well as accepting the same options.

Calls to kernels from KernelAbstractions.jl will give the code of the actual underlying kernel seamlessly.