Skip to content

Virtual machine (VM)link


The VM is an abstract machine that defines a type system of primitive and reference objects, module machinery, and a fairly involved mechanic for calls and dynamically binding extern funcs to their defs.

It comes with a bytecode module type, which implements the module interface and exposes a CFG based instruction set (sans pointers, as there are some security and portability to device scheduler goals in play) with an infinite number of registers. A VMFB, which the compiler produces by default is a serialization of this. The bytecode module has an interpreter for it.

VM moduleslink

There are multiple module types provided by default in the runtime:

  • HAL for interacting with devices
  • io_params for external data
  • check for testing

There are also Python bindings to create a module object dynamically and define exports in Python. Since the module interface is just natively defined in C, IREE also lets you load a .so dynamically which exports a function to create a new named module (this is what emitc produces for example).

When the tools are creating an iree_vm_context for execution, this is primarily about instantiating modules and adding them. Each module resolved its imports from the exports of the priors. In practice, an io_params module is added to access parameter providers, a hal module for devices, one bytecode module for each vmfb listed, and a native .so module for any .so listed.

That's all it is at the runtime. There's a lot of machinery on the compiler side for producing these modules and their interplay. The lowest level there to get a feel for what the compiler can emit, either a bytecode or a C based export, look at the vm dialect.

Lowering from the VM to Clink

There has been talk for years of having a direct lowering of VM to LLVM without going through C. While desirable in theory, it's just never become a priority... The C based export is what embedded folks want (because you never want to pin your system toolchain to a random LLVM like that). And the bytecode side has never been the bottleneck for others. It's also an annoying, maintenance prone bit of code to write and just never got done.

The high "it just works" quotient on the bytecode side has probably helped drive that too. "vmfb" has become a bit synonymous with "IREE" and teams using it think that is the main thing. But it is just one serialization of a VM module defining a program... But it has the best tools and debugging experience.

Call interfacelink

The VM call interface is modeled as a coroutine, and the bytecode interpreter supports multi task suspend/resume on a single thread. This is used for a lot of things (i.e. multi batch, async submissive, interfacing to io fabric, etc). Most of what people think of as "async" in the context of device interactions comes down to this and the cooperation of the hal module which provides device based synchronization and scheduling primitives.

The way it is structured, the VM was not designed to block, but it can suspend.